Open Access

Location-based distributed sleeping cell detection and root cause analysis for 5G ultra-dense networks

EURASIP Journal on Wireless Communications and Networking20162016:149

https://doi.org/10.1186/s13638-016-0649-6

Received: 1 August 2015

Accepted: 23 May 2016

Published: 21 June 2016

Abstract

The sleeping cell problem is one of the most critical issues for cellular deployments, consisting in the outage of a cellular station, which, conversely, works properly from the point of view of the monitoring system. This problem is often not detectable by the operators, and it could lead to severe degradations in the service provision in the long term. This issue has been commonly managed by the centralized analysis of network performance indicators. However, those solutions are unsuitable for the new ultra-dense small cell scenarios that will characterize 5G deployments. New approaches are required to cope with the high level of cell overlapping as well as the huge number of sites to be managed. In this context, a novel mechanism to detect sleeping cell issues is proposed, which takes advantage of the recent advances in indoor localization as well as monitoring data obtained by user equipments. In addition, a root cause analysis of the cell failure is presented. The capabilities of the proposed approach are evaluated in a realistic key scenario, showing the feasibility and usefulness of the proposed location-based approach.

Keywords

Self-organizing networks (SON)5GDenseNetsSmall cellLocalizationSelf-healingSleeping cellTroubleshootingDetection

1 Introduction

Current studies on 5G focus on three main objectives: converged fiber-wireless, super-efficient, and super-fast mobile networks [1]. To achieve the last two objectives, 5G networks are expected to hugely increase network densification. For these, the ultra-dense deployment of small cells or DenseNets (with a few tens of meters inter-site distance) will be one of the main approaches used to reach the upcoming coverage and throughput requirements [2]. Small cells consist on low-powered, low-range base stations (BSs) used to cover especially difficult radio locations, as well as to increase network capacity for specific hotspots. Small cells are already started to become a common approach to provide coverage indoors (e.g., malls or offices) or to shadowed locations. They are also employed to serve high demanding spots (e.g., airports, stadiums, etc.). Different small cell models have been defined for indoor scenarios, such as picocells, i.e., cells with up to 200-m coverage, and femtocells, i.e., cells with coverage in the range of tens of meters and making use of broadband non-dedicated infrastructure, e.g., digital subscriber line (DSL), to connect with the operator’s network [3].

The large number of small cells, together with the growing coexistence of multiple radio access technologies (RATs) like global system for mobile communications (GSM), universal mobile telecommunications system (UMTS), long-term evolution (LTE), etc., leads to an increasing complexity in the operations, administration, and management (OAM) of cellular networks. In order to overcome this complexity, the Next Generation Mobile Networks (NGMN) Alliance and the 3rd Generation Partnership Project (3GPP) introduced the concept of self-organizing networks (SON), which aims to automate the OAM procedures [4], reducing costs and increasing the network performance and quality of service. SON encompasses three main categories: self-configuration, which is related to the capability of automatically deploying new elements in a network and self-optimization, on the ability of the network to adapt to changing service requirements. Finally, self-healing covers the tasks related to failure management and prevention.

This work is focused on the self-healing functionalities for ultra-dense small cell networks. Self-healing is key for network OAM automation and performance, since network failures can lead to service degradations that might highly impact the brand image and the long-term revenue of operators. Failure management is divided in four main subtasks [5]. Firstly, detection consists in the discovery of network problems, i.e., identifying cells with degradations in the provided service. Secondly, diagnosis, also called root cause analysis, aims at identifying the specific cause or fault producing the degradation. Finally, once a problem has been detected, different actions can take place to compensate its effects until recovery actions restore the network to its full functionality. In classic failure management approaches, these are very time consuming and signaling generating tasks. This makes the automation of these functions a field attracting an increasing attention, where the implications of its application for 5G scenarios have been only scarcely considered.

Moreover, one of the most common problems in small cell scenarios is the sleeping cell issue, which is the situation where a base station is not able to properly serve users, and this is not directly reflected in the OAM monitoring indicators [6]. The causes behind this problem range from BS unplugging, failures in the cell hardware, or incorrect configuration. Since many of these fault causes behind the sleeping cell issue could be quickly compensated and/or recovered with automatic actions (like restarting the BS or updating its software), fast mechanisms for failure detection/diagnosis are essential to automatically trigger those tasks. Therefore, detection and diagnosis should, ideally, identify the failure and its causes in the range of minutes/seconds.

However, small cells can be particularly prone to failures (due to its more accessible hardware, use of non-dedicated backhaul, etc.) and have limited reporting capabilities and reduced coverage areas which very variable level of use. Therefore, it is common that not alarm or clear performance degradation might be reported to the OAM system in these situations.

Instead, user equipment (UE) positioning information can serve as an additional input for troubleshooting of these issues. For example, the knowledge on UEs location is becoming growlingly available in both outdoor and indoor scenarios [7]. Localization mechanisms based on the small cell deployments [8] open the door to using location to support the OAM tasks. Even if its use for self-healing in indoor scenarios has been just recently proposed [9], its usefulness has been previously demonstrated for macrocell scenarios and other OAM tasks (e.g., for coverage optimization [10]).

Also, another important limitation of classical approaches is the high signaling costs of transmitting the monitoring data to the centralized OAM/SON systems, which might overload the operator’s network. Additionally, 5G services and the operation of dense small cell deployments will also impose very demanding response time and computational requirements when performed in a centralized manner. In order to avoid these problems, SON mechanisms should be as automatic and distributed as possible, avoiding the saturation of the network and the centralized OAM elements.

Taking all this into account, the present work defines a novel fully distributed and automated location-based mechanism for sleeping cell detection and cause diagnosis in ultra-dense scenarios based on the deployment of small cells. Here, the required UE locations are assumed to be provided by external localization sources. The details of the particular method used for UE localization are considered outside the scope of the detection/diagnosis algorithm, making it agnostic to the use of any localization source. Hence, this paper is organized as follows: Section 2 discusses the challenges and state of the art in location-based mechanisms and detection and diagnosis of sleeping cell issues for the considered small cells scenario. Section 3 summarizes the characteristics and assumptions for location-based processing of monitoring information. The proposed mechanisms for detection and cause analysis are detailed in sections 4 and 5, respectively. The combined distributed scheme is then described in Section 6. The defined system is evaluated in Section 6 and the conclusions of this study are finally presented in Section 7.

2 Related work

Classic self-healing macrocell solutions [11] are unsuitable for small cell deployments due to multiple reasons. Firstly, small cells provide reduced monitoring functions in comparison to their macrocell counterparts, as a result of their limited computing capabilities and in order to avoid backhaul saturation. Secondly, general network performance indicators might not be highly affected by the cell failure, making the issues to remain undetected for long periods of time. This is mainly due to the occasional low number of UEs in the affected area. Also, because most of the UEs that should be served by the faulty cell might be covered by neighboring BSs (given the high level of coverage overlapping between cells in dense small cell environments). Only when there are users in spots that are poorly covered by other cells or if the service demand is high (because there are a large number of users or they require high capacity), the failure of one BS may be detected by classic network performance indicators (e.g., call dropping ratio).

Instead of cell-level performance indicators, other classic solutions may rely on direct reporting produced by the UEs, such as drive tests, UE app monitoring, or analysis of the control plane messages. In order to overcome limitations of drive tests (classically performed by field engineers with especial terminals), recent standardization work has established mechanisms for minimization of drive test (MDT). In MDT, measurements automatically gathered from common terminals are used for network monitoring. In this way, the work presented in [6] is based on the use of MDT information for sleeping cell detection. However, MDT is typically applied in outdoor environments where the terminals’ reports are combined with geographical information obtained from satellite or cellular-based localization (e.g., based on observed time difference of arrival (OTDoA)). Thus, those systems have very low performance indoors due to the lack of satellite coverage and the short distance between cells.

Following a different approach, the work in [12] proposed a system that performs particle filtering for detection of sleeping cells. This is applied in a distributed manner based on UE measurements. However, this method required the use of large databases of previous measurements. Also, positioning information was not used and no diagnosis analysis or the conditions of implementing the distributed algorithm were addressed.

UE localization systems are increasingly available in indoor scenarios, based on multiple technologies under development or already commercialized. These are often based on the analysis of the received signal characteristics (e.g., time of arrival, received power, etc.) from different systems such as radio-frequency identification (RFID) [13], WiFi [14], cellular [8], and ultra-wideband (UWB) [15]. Although the use of UE localization has already attracted interest in other OAM tasks, e.g., for the presented MDT cases, it has been mainly neglected in previous self-healing works and particularly for indoor scenarios.

Therefore, up to our knowledge, none of the previous works presented a consistent fast and distributed detection/diagnosis mechanism for sleeping cell issues in dense scenarios. Equally, previous developments did not take advantage of automatically obtained location information for self-healing.

In this respect, the present work presents three main contributions. Firstly, it defines a location-based mechanism for sleeping small cell detection. Secondly, it establishes the general architecture, procedures and restrictions for a distributed implementation of the mechanism in DenseNets. Thirdly, it also proposes a system for the distributed analysis of the root cause behind the sleeping cell issue.

3 Location and radio information application

3.1 UE radio measurements

In the proposed approach, UE measurements are established as the main source of information for detection, instead of the classical approach based on centralized data coming from the OAM system. In particular, received signal strength (RSS) measurements are chosen as the input for the mechanisms. RSS is defined as the received power level from the base station downlink reference signals. The use of RSS values has multiple advantages over other indicators. Firstly, they are not event oriented, i.e., they can be obtained by the terminal at any time/position, not only in connected mode but also in idle mode. Secondly, in comparison to signal quality indicators such as the call blocking ratio (CQI), RSS values are not affected by the variable interference conditions, which are also very dependent on the network load (especially for the LTE case). Thirdly, regarding practical implementation aspects, radio analysis Android apps (e.g., G-NetTrack [16]) provide the RSS measured values, whereas they do not present quality indicators for most commercial terminals [16]. This is especially important not only from a prototype point of view but also for the possible over-the-top user-level implementations of the proposed system, which could be adopted in some scenarios (e.g., if the monitoring function is combined with a navigation application installed in the smartphones).

Even though the integration between user-level apps and the cellular management plane could be not straightforward, recent works indicate a trend towards an increasingly tighter integration between them. In this respect, the work in [9] proposed the interfaces and architecture required to integrate app-level information into the 3GPP management plane. Meanwhile, the work in [17] developed a real testbed prototype of such UE-level app based network troubleshooting. Also, end-to-end analysis and UE-level service-performance measurements are becoming common for network operators [18]. Moreover, the integration user-level apps with the cellular management plane would become even more achievable if particular apps (such as localization applications based on the cellular signal analysis) or application programming interfaces (APIs) are provided/commercialized by the operators themselves.

Fourthly, RSS-based indicators are defined for any existent RAT and they will also be established for any future 5G standard. For example in UMTS, RSS values correspond to the common pilot channel (CPICH) received signal code power (RSCP) [19], meanwhile in LTE/LTE-A, the equivalent measurement is the reference signal received power (RSRP) [20].

RSS values can be sometimes measured concurrently for both the serving and the neighboring cells. However, that is not always the case in live scenarios because the reception of neighboring cells RSS values is conditioned by multiple issues:
  • Incomplete neighbor cell list: most of the time, the UE served by a cell might only report the cells included in the serving neighbor cell list. That list might be incomplete, especially in UMTS. In that technology, the neighbor list is typically manually configured. Also the serving cell may require to directly receive power from each neighbor in order to keep it in the list.

  • Unavailability to obtain neighbor cell received power from the monitoring application: if the application used to obtain the RSS values is a user-level app (e.g., Android or iOS), the majority of commercial terminals only report the values for the serving cell [16].

  • Randomness on the neighboring cell UE reporting period: even at control and radio-link monitoring level, the neighboring cells are often measured with reduced periodicity or only associated with certain events.

For these reasons, the availability of neighboring cell RSS is not guaranteed. Therefore, and in order to develop widely applicable mechanisms, only serving cell RSS values are considered available and used as inputs for the algorithm.

In addition to the UE measurements, it is assumed that UE positions are available to the SON system as part of the information provided by indoor positioning systems. The required architectural coordination of different localization sources with SON mechanisms has been presented in [9]. In this way, UEs positioning information can be obtained indistinctively from the UEs or external sources, such as surveillance camera localization.

3.2 Location-based measurements processing

In order to support efficient methods for the detection and diagnosis of sleeping cells, the combination of the presented UE RSS measurements with positioning data is proposed. Traditionally, key performance indicators (KPIs) were calculated based on statistical analysis of UE measurements and events. With that aim, the UE RSS-based indicators are computed for each cell by generating statistics of the set of multiple measurements M j [t], captured from a specific set of terminals (e.g., typically those served by a specific cell j ) during a particular period of time t. The calculation is commonly performed periodically, e.g., one KPI value generated every hour based on the measurements gathered during that period.

Conversely, in the location-based proposed method, the individual RSS measurements gathered from the terminals are processed to include UE position information to generate the KPIs. Multiple solutions are possible for the integration of the positioning data with the UE RSS measurements, adapting those previously applied in other fields, especially on image processing, e.g., spatial correlation, pattern-based analysis, etc. For example, the work in [17] presented a system based on the correlation between each individual UE received levels and the expected values given their position. Such kind of approaches can be valid in femtocell environments where only a few UEs are analyzed. However, the need to pursue fast and computationally low cost methods where large number of UEs is present leads to the choice of sample weights [21] for the generation of the indicators involved in the diagnosis. Although sample weights are commonly used in fields like social polling, up to our knowledge, it has only been applied in cellular networks for centralized diagnosis [22], and not for detection.

Using sample weights, the statistical relevance of each sample is potentiated depending on the expected impact of the failure in its measurement spot. Each sth RSS value is therefore assigned with a different weight, \( {w}_{AOI}^{cell_i}\left({\gamma}_{xyz}\left[s\right]\right) \), based on the value of the weight function \( {w}_{AOI}^{cell_i}\left({\gamma}_{xyz}\right) \) for the position γ xyz [s] where it was measured. In the proposed approach, \( {w}_{AOI}^{cel{l}_i}\left({\gamma}_{xyz}\right) \) is defined for a certain area of interest (AOI) of the analyzed cell i , e.g., its expected coverage area.

For any particular deployment, a set of multiple areas of interest, AOI can be established, where AOI AOI. Each of the AOIs implies a particular \( {w}_{AOI}^{cel{l}_i}\left({\gamma}_{xyz}\right) \) and therefore different location-based KPIs. These are used as the inputs for the detection of issues in different cells. The specific definition of these AOIs and weights is further analyzed in subsections 3.A and 3.C, respectively.

It has to be noticed that previous uses of the term AOI refer to a delimited geographical of a cellular deployment defined by engineers to characterize during a performance study. These often have an extension of several kilometers. In the proposed approac,h however, the AOIs imply the different set of areas from where the UE measurements are considered in the statistics calculations used by the sleeping cell detection algorithm, being automatically generated for each small cell, and having a much reduce size (a few meters).

In the classic non-location-based solutions, two main statistics are commonly applied for OAM purposes: the RSS mean and the RSS 5th percentile. Their classic expressions are therefore adapted in this approach in order to use of location-based weight samples, as described below:
  • Location-based RSS_mean, \( {\overline{RSS}}_{AOI}^{cel{l}_i}\left(cel{l}_j,t\right), \) defined as:
    $$ {\overline{RSS}}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\mathrm{cell}}_j,t\right)=\frac{1}{E_W}{\displaystyle \sum_{s=1}^{\left|{M}_j\left[t\right]\right|}}{w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{xyz}\left[s\right]\right){m}_{\mathrm{RSS}}\left[s\right], $$
    (1)
    where M j [t] is the set of RSS values m RSS[s] measured from the serving cell (or set of cells) cell j during the period t by the UEs located in the AOI of cell i . |M j [t]| indicates the number of values in the set. \( {w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{xyz}\left[s\right]\right) \) represents the individual weight applied to one measurement m RSS[s], depending on the coordinates where the sample was measured γ xyz [s]. Finally, E W is the sum of all the applied weights, \( {E}_{\mathrm{W}}={\displaystyle {\sum}_{s=1}^{\left|{M}_j\left[t\right]\right|}{w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{xyz}\left[s\right]\right)} \).

    The serving cell j can be equal to cell i or a different cell of the scenario. When different, it means that the generated indicator includes information on the UEs served by cell j but located in an AOI of cell i , as presented in Fig. 1. This is one of the main characteristics of the proposed approach, as it allows the monitoring of a possible faulty cell i by its neighbors. In this way, the detection of the sleeping cell is done from the measurements obtained from its neighboring cells served UEs.

  • Location-based RSS_5th percentile, calculated as the RSS value below which the 5 % of the lowest collected RSS values are. For non-location approaches, this is a common indicator of the values gathered in the edges and/or at far distance of the cell as well as from low covered (shadow) spots. If a cell is in outage, the classic RSS_5th percentile of their neighboring cells would especially reflect the RSS received by the UEs more poorly served in the area originally covered by the faulty cell.

Fig. 1

AOIs, sample gathering, and connection between cells

For the location-based approach and similarly to the RSS_mean, the fifth percentile indicator is based on the M j [t] measurements located in a specific AOI of cell i . To compute it, the common procedure for percentile calculation [23] is adapted to the use of weight samples. In this way, M j [t], the set of RSS samples in a period, is sorted from minimum value to maximum. Then, each sample in the position n of the ordered list \( {M}_j^{\mathrm{ord}}\left[t\right] \) is assigned with the ordinal rank r n calculated as:
$$ {r}_n=\frac{100}{E_{\left|{M}_j^{\mathrm{ord}}\left[t\right]\right|}}\left({E}_n-\frac{w_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{xyz}^{\mathrm{ord}}\left[n\right]\right)}{2}\right), $$
(2)
where \( {w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{\mathrm{xyz}}^{\mathrm{ord}}\left[n\right]\right) \) is the weight assigned to the RSS value in the position n of the ordered list, where \( n\le \left|{M}_j^{\mathrm{ord}}\left[t\right]\right|=\left|{M}_j\left[t\right]\right| \). E n is the partial sum of all weights applied to the ordered samples 1 to n, \( {E}_n={\displaystyle {\sum}_{s=1}^n{w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{xyz}^{\mathrm{ord}}\left[s\right]\right)} \). The fifth percentile would be the sample with r n  = 5. If no measurement has r n  = 5, the ordered samples with directly inferior r k and directly superior r k + 1 nearest rank are selected, in a way that r k  < 5 < r k + 1. The correspondent RSS samples, \( {m}_{\mathrm{RSS}}^{\mathrm{ord}}\left[k\right] \) and \( {m}_{\mathrm{RSS}}^{\mathrm{ord}}\left[k+1\right] \) are then used to obtain the fifth percentile by linear interpolation:
$$ RSS\_5t{h}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\mathrm{cell}}_j,t\right)={m}_{\mathrm{RSS}}^{\mathrm{ord}}\left[k\right]+\frac{5-{r}_k}{r_{k+1}-{r}_k}\left({m}_{\mathrm{RSS}}^{\mathrm{ord}}\left[k\right]-{m}_{\mathrm{RSS}}^{\mathrm{ord}}\left[k+1\right]\right), $$
(3)

3.3 Sample weights and AOIs for the detection of sleeping cells

In the definition of the AOIs and the sample weights function \( {w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{xyz}\right) \) , different approaches can be applied to reflect the relevance of a particular location, γ xyz , in respect to a possible failure in any cell i . For the sleeping cell issue, it is expected that the areas most impacted by the possible fault would be those contained in the coverage area of cell i , as it would be populated by UEs most affected by its failure.

The AOIs are therefore geographically defined in order to restrict the samples considered in each location-based indicator. Two different AOIs per cell are envisaged as alternative approaches (with different level of filtering) for the detection of the sleeping cell issue:
  • Expected coverage area (ECov) of the possible sleeping cell. The UEs in this area are most likely to be impacted by the cell issue. However, depending on the range of overlapping between cells, this might be compensated by the coverage coming from neighbors BSs

  • Expected center area (ECent) of the cell refers to locations in the core of the coverage area of a cell. In these, the signal of the cell is clearly predominant in respect to its neighbors.

Additionally, to avoid the effect generated by the macrocell and the UEs outside the area of the indoor scenario, only the terminals inside the analyzed indoor scenario (e.g., building) are considered as inside the AOIs.

Due to the difficulty of obtaining reliable information about the specific height of the terminals and the complexity that would introduce a continuous vertical dimension, coverage areas are commonly defined as multiple two-dimensional layers (e.g., one per building floor). In this way, γ xyz  = (x, y, z = ζ) represents a point in the two-dimensional Euclidean space E 2 of the analyzed plane/floor ζ. Following the same approach, the AOIs are defined as two-dimensional, (x, y) E 2, for each floor of the deployment. Given this and for simplicity, the following AOI definitions would be described

These expected coverage areas can be defined in terms of signal quality (e.g., CQI-based). However, pure RSS-based definitions are preferred as justified in Section 3.A. Based on the expected received power \( \widehat{\mathrm{RSS}}\left(\left(x,y,z\right),{\mathrm{cell}}_i\right) \), the AOI ECent(cell i ) is defined as:
$$ \mathrm{ECent}\left({\mathrm{cell}}_i\right)=\left\{\left(x,y\right)\in \left.{E}^2\right|\widehat{\mathrm{RSS}}\left(\left(x,y,z\right),{\mathrm{cell}}_i\right)>\widehat{\mathrm{RSS}}\left(\left(x,y,z\right),{\mathrm{cell}}_j\right)+\varDelta {\mathrm{RSS}}_{\mathrm{cent}},\forall i\ne j\right\} $$
(4)
where ΔRSScent is the additional power (in dB) to be received in the point (x, y, z) from cell i in comparison with any other cell j , to consider the point as part of the center of the cell. If ΔRSScent = 0, the expression results in the expected coverage area ECov(cell i ).

3.4 AOIs calculation considerations

Different approaches can be adopted to define the AOIs by approximating Eq. (4) for any given (x, y, z). Particularly, test campaigns, site-specific detailed-scenario propagation models (considering specific obstacles, walls, etc.) or simple log-distance path loss (assuming all points as in line-of sight to the BSs) can be adopted. The application of one or another technique for the AOI calculation is dependent on the available information about the scenario.

Table 1 summarizes the information required for the location-based indicators depending on the approach used for the AOIs estimation. Here, the transmitted cell power can be obtained from the normal configuration/monitoring of the BSs. However, the availability of the rest of the information has to be analyzed.
Table 1

Required information depending on the AOI calculation approach

AOI calculation approach

Required network information

Required localization/scenario information

Classic (no location/AOIs)

UE measurements

Test campaign

UE measurements, radiomap database

UE positions, positions during radiomap gathering

Site-specific detailed propagation

UE measurements

UE positions, BSs positions, transmitted power, obstacles, walls, propagation conditions

Simple log-distance path loss

UE measurements

UE positions, BSs positions, transmitted power

On the one hand, in common situations, test campaigns or detailed-scenario models (including obstacles, walls, etc.) are not available, making not possible to apply detailed-scenario propagation models for AOI estimation. Additionally, the details of such information can become rapidly obsolete if the scenario changes.

On the other hand, simple log-distance path loss calculation (not considering site-specific walls and obstacles) only requires the positions and relative power of the base stations (BSs). The knowledge about the BSs positions might be not available if the deployment is unplanned, as it can be especially the case home femtocells. However, the BSs position is typically known for the type of large indoor scenarios (malls, large office, airports) considered. In these environments, BSs (picocells or enterprise femtocells) are mainly installed in fixed positions by professional field engineers, allowing the registry of their locations. For these scenarios, careful planning of the deployment is also recommended [24]. Additionally, if the cellular network supports UE positioning, the location of the BSs will be also required for such service.

Although not considering site-specific details can led to inaccuracies, it has to be noted that this assumption is only used to estimate geographical AOIs, not to obtain detailed RSS values. Also, many indoor large scenarios are semi-unobstructed or containing symmetric obstacles (e.g., walls) for the different BSs, minimizing also the importance of a very detailed propagation approach to calculate the AOIs.

Therefore, taking into account the issues associated with more complex approaches (particularly the lack of the necessary information), simple log-distance path loss would commonly be the most feasible solution to estimate the AOIs in many real deployments scenarios. In this case, for the AOIs definition, a similar reasoning to the one followed in [25] for coverage probability estimation is adopted. For each point (x, y, z), the expected received power from cell i , value \( \widehat{\mathrm{RSS}}\left(\left(x,y,z\right),{\mathrm{cell}}_i\right) \), consists of the transmitted power of the cell, P tx (cell i ) minus the path loss to that point PL((x, y, z), cell i ). Without including fading effects, the path loss is inversely proportional to the distance between the point and the cell: \( \widehat{\mathrm{RSS}}\left(\left(x,y,z\right),{\mathrm{cell}}_i\right)\sim 10* \log \left({P}_{tx}\left({\mathrm{cell}}_i\right)/{d}^{\propto}\left(\left(x,y,z\right),cel{l}_i\right)\right) \) in dB, where d((x, y, z), cell i ) is the Euclidean distance between (x, y, z) and the cell i BS. Moreover, is the path loss exponent. This exponent has commonly a value close to 2 for line-of-sight propagation in indoor environments (e.g.,  = 1.87 for Winner II indoor-indoor A1 model [26]).

Fast-fading effects in the propagation are not included. This is consistent with their reduced values in the final coverage due to the terminal interleaver and the existent time and power margins for mobility and cell reselection.

If, as expected, the obstacles of the scenario are unknown and the cells transmit roughly in the same band (as it is typically the case), the expression of ECov based on Eq. (4) should be approximated as:
$$ \mathrm{ECov}\left({\mathrm{cell}}_i\right)=\left\{\left(x,y\right)\in {E}^2\left|\frac{d\left(\left(x,y,z\right),{\mathrm{cell}}_i\right)}{{P_{tx}}^{1/\propto}\left({\mathrm{cell}}_{\mathrm{i}}\right)}<\frac{d\left(\left(x,y,z\right),{\mathrm{cell}}_j\right)}{{P_{tx}}^{1/\propto}\left({\mathrm{cell}}_j\right)},\forall i\ne j\right.\right\} $$
(5)
This expression is equivalent to a multiplicatively weighted Voronoi tessellation [25]. If all the small cells transmit with the same power, this is equivalent to the classic polygonal Voronoi tessellation:
$$ \mathrm{ECov}\left({\mathrm{cell}}_i\right)=\left\{\left(x,y\right)\in {E}^2\left|d\left(\left(x,y,z\right),{\mathrm{cell}}_i\right)<d\left(\left(x,y,z\right),{\mathrm{cell}}_j\right),\forall i\ne j\right.\right\} $$
(6)
However, if the BSs transmitted powers are not equal, coverage areas might be non-convex, non-polygonal and they can include “holes”. This highly increases the complexity of calculating whether a point belongs or not to the AOI. To avoid these issues, the work in [25] proposed a circular approximation to the coverage:
$$ {R}_{\mathrm{cov}}\left({\mathrm{cell}}_i\right)=\begin{array}{c}\hfill \min \hfill \\ {}\hfill \forall i\ne j\hfill \end{array}\left(\frac{d\left({\mathrm{cell}}_i,{\mathrm{cell}}_j\right)}{1+{\left(\frac{P_{tx}\left({\mathrm{cell}}_j\right)}{P_{tx}\left({\mathrm{cell}}_i\right)}\right)}^{\frac{1}{\propto }}}\right) $$
(7)
Following the same approach, ECent(cell i ) can be calculated as a circle centered in the BS with a radius equivalent to a configurable portion, k R  [0, 1], of the minimum distance between the BS and the border of its coverage: R cent(cell i ) = k R  * R cov(cell i ). In this way, the margin ΔRSScent as defined in Eq. (4) would be equivalent to:
$$ \varDelta {\mathrm{RSS}}_{\mathrm{cent}}=10 \log \left(\frac{k_R}{1+\left(1-{k}_R\right){\left(\frac{P_{tx}\left({\mathrm{cell}}_j\right)}{P_{tx}\left({\mathrm{cell}}_i\right)}\right)}^{1/\propto }}\right)\ \left(\mathrm{dB}\right) $$
(8)

These expressions allow to compute the AOIs prior to the online detection, storing them with a very reduced cost, whereas they also allow direct AOI reconfiguration if there are changes in the transmitted power.

For the sample weights, an increasing relevance should be given to those positions better covered by cell i and less covered by its neighbors in normal operation. Assuming also a simple approach for their definition, the proposed expression for the weight is:
$$ {w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left(x,y\right)=\left\{\begin{array}{c}\hfill \frac{P_{tx}\left({\mathrm{cell}}_i\right)/{P}_{tx}\left({\mathrm{cell}}_{\mathrm{dom}}\right)}{{\left(d\left(\left(x,y\right),{\mathrm{cell}}_i\right)/d\left(\left(x,y\right),{\mathrm{cell}}_{\mathrm{dom}}\right)\right)}^{\propto }}\kern1em \mathrm{if}\ \left(x,y\right)\ \mathrm{inside}\ AOI\hfill \\ {}\hfill 0\kern5em \mathrm{otherwise}\hfill \end{array}\right., $$
(9)
where celldom is the dominant adjacent cell which highest estimated received power on (x, y), celldom : {P tx(celldom)/d((x, y), celldom) > P tx(cell j )/d((x, y), cell j )| j ≠ i, j ≠ dom}. In this way, and with ECov and ECent following the same propagation approach, the computation of \( {w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i}\left({\gamma}_{xyz}\right) \) consists in a simple point-in-polygon (PIP) [27] and/or point-in-circle calculation.

4 Detection algorithm

Based on any available indicator F, typical detection mechanisms are based on calculating which of its samples crosses a predefined detection threshold, considering then that the network is under failure [28]. The presented indicators (RSS_mean or RSS_5th percentile, location-based or not) could be straightforwardly used for this process.

This detection approach has important advantages in comparison with more complex mechanisms, like the ones presented in reference [6], as it allows the immediate detection of degradations based on a unique indicator and in a fast manner. However, it implies the definition of thresholds, typically by human expertise, being each considered indicator associated with a particular threshold. Threshold definition is commonly a costly task that implies large times and a deep knowledge of the network behavior that may be not available. Additionally, the thresholds may vary depending on the specific network or BS, complicating their definition.

To avoid this issue, a baseline method based on a naive Bayes classifier is proposed. For its description, uppercase letters will refer to variables/indicators, lowercase to specific values of such variables and bold letters to vectors of multiple variables or values. The Bayes classifier method would make use of any set of input indicators F = {F 1, F 2, …} and their current values, e.g., f[t] = {F 1 = f 1[t], F 2 = f 2[t], …} in order to estimate the posterior probability \( \widehat{p}\left({C}_i={c}_i\left|\boldsymbol{F}\right.=\boldsymbol{f}\left[t\right]\right) \) of a cell i of being in a certain cell status C i  = c i , where c i  {Normal i , Sleeping i }.

This process is based on two main stages: a training phase and an online phase that should be performed (in parallel) for any cell i considered for analysis.

4.1 Training phase

During this initial stage, the conditional probability density functions (PDF) of all FF indicators given both normal and failure case has to be estimated: \( \widehat{p}\left(\left.F\right|{\mathrm{Normal}}_i\right) \) and \( \widehat{p}\left(\left.F\right|{\mathrm{Sleeping}}_i\right) \), as well as the prior probabilities, \( \widehat{p}\left({\mathrm{Normal}}_i\right) \), and \( \widehat{p}\left({\mathrm{Sleeping}}_i\right) \) of the cell status.

The PDFs are calculated from the relative frequency of the values in the training set. Then, the kernel smoothing (KS) density approximation [29] is proposed for estimating the PDFs. KS probabilistic model is generated based on the superposition of multiple Gaussian functions. Conversely, it is commonly assumed that RSS-based indicators values can be modeled by a unique Gaussian or Beta distributions. However, this assumption is only correct for uniform users’ distributions and non-limited areas of analysis. As it would be shown in the evaluation section, this is not the case for data obtained in more irregular scenarios with complex user mobility. The KS solution is non-parametric and more computationally expensive than other mechanisms. However, this problem can be mitigated for real-time detection as the model is fitted during the training phase, and it is already constructed in the online phase. To save storage space, the KS distribution is calculated only in the range of discrete RSS values that can be reported by the UE, which depends on the particular standard.

Therefore, the stored conditional PDF for each possible status C i (normal or sleeping) is a discrete function such as \( {\mathrm{pdf}}_F\left({r}_F\right)=\left\{\widehat{p}\left({r}_F\left|{C}_i\right.\right)\ \left|\forall {r}_F\epsilon \mathrm{R}\mathrm{G}\right.\left({\mathrm{pdf}}_F\right)\right\} \), where r F is any of the discrete values of the range RG(pdf F (r F )) where the distribution is defined. For example, in LTE and considering RSRP values, RG(pdf F ) is equal to the set of integers in the interval [0 … 97], which is a direct mapping of the receive power measurements from −140 to −44 dBm with a 1-dB resolution [20].

However, during the online phase, the mean or percentile obtained from the set of multiple of RSS samples can result in values not included in RG(pdf F ). In that case, KS smoothing or simpler linear approximation between the immediately lower and higher integer values can be applied depending on the available computational capacity.

One issue for this approach would be obtaining the values for the training set. For the normal status, this is easy to obtain, as the indicator values can be gathered under the normal behavior of the network. For the sleeping cell status, if real fault-labeled cases are not available (which is the most common situation), two different options are envisaged:
  • ON/OFF calibration period: by a simple procedure of disconnecting and connecting the cells in an alternative manner (which can be performed automatically), the system can obtain the needed sleeping cell training set. These may however alter the normal operation of the network.

  • Neighboring cells measurements analysis: if the terminals are able to measure and report RSS values from neighboring cells, this information can be used to approximate their expected serving cell values if one cell is disconnected. This process would have the advantage of not disrupting the cell service provision, and it could be performed continuously.

Independently, of the used method, the validity of the calculated models can be jeopardized by changes in the network: in the BSs positions, in the scenario characteristics, obstacles, etc. In such cases, mechanisms for automatic update of the models shall be applied. Here, reference [30] provided a solution for identifying the need of updating the models as well as online updating. Such mechanisms can be straightforwardly applied to any network indicator, not requiring therefore any particular modification for the proposed location-based approach.

4.2 Online phase

This stage is performed in the already working network. Here, the estimated status of cell i is calculated by a proposed feature-weighted version of the naive Bayes classifier. The followed concept of weighted naive Bayes classifier was proposed in [31] for general centralized data mining. In the present work, however, the applied weights, expressions, and application have been completely redefined for its use in cellular distributed failure detection.

In this way, the Bayes classifier is used to combine the values of f[t]. For any C i  = c i of the two possible status of cell i , C i  = {Normal i , Sleeping i }, \( \widehat{p}\left({C}_i={c}_i\left|\boldsymbol{F}\right.=\boldsymbol{f}\left[t\right]\right) \) is therefore estimated as:
$$ \widehat{p}\left({C}_i={c}_i\left|\boldsymbol{F}\right.=\boldsymbol{f}\left[t\right]\right)=\frac{\widehat{p}\left({C}_i={c}_i\right){\displaystyle {\prod}_{\forall F\in \boldsymbol{F}}\widehat{p}{\left(F=f\left[t\right]\left|{C}_i={c}_i\right.\right)}^{\Omega_F\left[t\right]}}}{\widehat{p}\left(\boldsymbol{F}=\boldsymbol{f}\left[t\right]\right)}, $$
(10)
where F = f[t] is the set of current values of the indicators. \( \widehat{p}\left({C}_i={c}_i\right) \) is the estimated prior probability of each status value; this means, the total likelihood of the status if no data is known. \( \widehat{p}\left(\boldsymbol{F}=\boldsymbol{f}\left[t\right]\right) \) is the estimated likelihood of the evidence. \( \widehat{p}\left(F=f\left[t\right]\left|{C}_i={c}_i\right.\right) \) is the conditional probability for a particular input indicator value F = f[t] given the status c i of cell i and calculated as described from the pdf F (r F ). Ω F [t] is the confidence level of the f[t] value, and it is used to modulate the contribution of each value of \( {F}_{ij}^{\mathrm{AOI}} \) in the weighted classifier in each instant.
The consideration of independence between the variables in F is disputable; however, it is commonly assumed for cellular indicators [11]. Also, naive Bayes classifiers have demonstrated good results even when that condition is not fulfilled [32]. Based on this, the estimated status of cell i , ĉ i is obtained by choosing the status with the highest posterior probability. As only two possible statuses have been defined C i  = {Normal i , Sleeping i } and being the evidence equal to both, the classifier expression can be simplified to:
$$ \frac{\widehat{p}\left({\mathrm{Normal}}_i\right)}{\widehat{p}\left({\mathrm{Sleeping}}_i\right)}{\displaystyle \prod_{\forall F\in \boldsymbol{F}}{\left(\frac{\widehat{p}\left(\left.f\left[t\right]\ \right|{\mathrm{Normal}}_i\right)}{\widehat{p}\left(\left.f\left[t\right]\ \right|{\mathrm{Sleeping}}_i\right)}\right)}^{\Omega_F\left[t\right]}}<1\to \begin{array}{c}\hfill {\mathrm{cell}}_i\ \mathrm{is}\ \hfill \\ {}\hfill \mathrm{faulty}\hfill \end{array} $$
(11)

4.3 Confidence level definition

Ω F [t] is a main parameter in order to establish the importance of each indicator in the classification. To do so, a particular definition of this parameter is established, defined as a function of the number of samples used for the calculation of f[t] (|M F [t]|), the weight relevance factor (φ F [t]) of those measurements (if F is a location-based indicator), as well as the statistical difference (Ψ(F)) between F normal and sleeping PDFs.

Firstly, |M F [t]| is used to define a minimum number of samples, |M| th , to consider the indicator significant. Secondly, in the case where F is a location-based indicator, φ F [t] is calculated by the normalized sum of the sample weights applied to the RSS measurements, giving higher weight to those f[t] based on larger number of samples and/or more relevant ones:
$$ {\varphi}_F\left[t\right]=\frac{1}{\overline{{\boldsymbol{\varphi}}_{\boldsymbol{F}}}\left[t\right]}{\displaystyle \sum_{s=1}^{\left|{M}_F\left[t\right]\right|}}{w}_F\left({\gamma}_{xyz}\left[s\right]\right), $$
(12)
where \( \overline{{\boldsymbol{\varphi}}_{\boldsymbol{F}}}\left[t\right] \) represents the normalization factor, being the average of the relevant factors applied in t to all the indicators in F.
Thirdly, Ψ(F) is calculated following the Hellinger score [33], serving as a metric of the level of overlapping between the normal and sleeping PDFs of F. This is where it is applied to the discrete conditional PDFs:
$$ \Psi \left({\mathrm{F}}_{ij}^{AOI}\right)=\mathrm{H}\left(\widehat{p}\left(\left.\mathrm{F}\right|{\mathrm{Normal}}_i\right),\widehat{p}\left(\left.\mathrm{F}\right|{\mathrm{Sleeping}}_i\right)\right)=\frac{1}{\sqrt{2}}\sqrt{{\displaystyle \sum_{\forall {r}_F\epsilon \mathrm{R}\mathrm{G}\left({\mathrm{pdf}}_F\right)}{\left(\sqrt{\widehat{p}\left(\left.{r}_F\right|{\mathrm{Normal}}_i\right)}-\sqrt{\widehat{p}\left(\left.{r}_F\right|{\mathrm{Sleeping}}_i\right)}\right)}^2}} $$
(13)

In comparison to other statistical distances (like the Kullback–Leibler divergence proposed in [31]), the Hellinger metric has the advantages of being symmetric (H(P, Q) = H(Q, P) for any Q, P distributions) and satisfying the expression 0 ≤ H ≤ 1, where “0” indicates complete similarity while “1” means complete independency, being a parameter of easy incorporation to any decision algorithm.

With these three parameters, the expression of Ω F [t] is finally defined as:
$$ {\Omega}_F\left[t\right]=\left(\left|{M}_F\left[t\right]\right|<{\left|M\right|}_{th}\right)\ast {\varphi}_F\left[t\right]\ast \Psi (F) $$
(14)

5 Diagnosis of sleeping cell causes

Once the sleeping cell problem is detected, compensation and recovery mechanisms can be supported by knowledge on the root cause behind the problem. In that respect, Table 2 shows the main causes of failures in small cell networks. Here, causes 1, 2, and 3 represent the most common roots for the catatonic sleeping cell problem in small cell networks. Four and 5 are failures in the OAM system that could lead to the erroneous identification of the cell issue in classic approaches.
Table 2

Sleeping small cell causes and reachability

No.

Failure

Description

Indicator

UE RSS values

Small cell NETACC

Backhaul NETACC (general access to the backhaul)

1

Small cell booting process

Small cell starting logon procedure into the operator’s network fails. The small cell stays active but without transmitting.

Affected

Unaffected

Unaffected

2

Small cell disconnection

Small cell is disconnected from power supply and/or backhaul connection. It stops transmitting.

Affected

Affected

Unaffected

3

Backhaul NETACC

Failure of the backhaul connection of the network.

Affected

Affected

Affected

4

Checking entity NETACC

The entity checking the cells is not able to connect with them.

Unaffected

Affected

Unaffected

5

SON system NETACC

The SON system is unable to connect specifically with the router due to wrong IP config.

Unaffected

Affected

Affected

While the presented detection mechanism has been centered on UE RSS analysis, other inputs can be used for the assessment of cellular degradations. These come from the analysis of the network accessibility (NETACC) of the cellular system elements from the point of view of their non-cellular links: LAN, optical fiber, DSL, etc. This can be checked by centralized SON entities as well as between small cells. An element in charge of that task is named as NETACC checking entity. This element or elements can perform the checking periodically or following an event/demand [9]. This check can be performed for a specific small cell element or for the complete deployment backhaul, e.g., by IP-PING style messages. Such information can be modeled as a binary random variable whose value is “1” if the element analyzed is reachable and “0” otherwise.

Combining both the network accessibility indicators and the analysis performed on the UE RSS values, it is possible to univocally identify each cause. Therefore, simple rules are proposed for the diagnosis of the different cases, as shown in Table 2.

6 Distributed Self-healing scheme

The implementation of SON mechanisms in current heterogeneous networks brings several issues to the classic centralized approach for OAM in cellular systems. Firstly, one of the main challenges to achieve fast-response failure management in cellular networks is the associated delay and signaling cost of the communications between the central OAM systems and the network elements (UEs, BSs). The network backhaul can be easily overloaded by the signaling costs associated with network monitoring. This is especially the case for femtocells, which make use of non-dedicated consumer-oriented infrastructure. Secondly, the growing number of cells and the complexity of the network might even lead to saturation of the operator’s backbone and central OAM elements, as they are in charge of monitoring and operating huge number of network elements.

In order to avoid these issues, the elements of each deployment should be able to manage themselves. Thus, a distributed scheme is deemed indispensable for the proper performance of the network. In this field, previous work [9] aimed to reduce signaling costs in femtocell deployments by establishing a local OAM element, allowing hybrid algorithms and minimizing backhaul use. Additionally, since the proposed on-site OAM element was only responsible for a local small cell network, the number of cells to be operated by such entity is much reduced, significantly limiting computational costs. While such scheme could be adopted for the mechanisms proposed in the present work, it is still vulnerable to failure in that centralized entity. Also the small cells may not be part of the same LAN, or their interconnection capacity may be low (e.g., if it is based on WiFi-LAN). Conversely, the use of a fully distributed algorithm would increase the resiliency of the system.

6.1 Distributed scheme

Given the presented detection and diagnosis mechanisms, a procedure is defined for its distributed application. The proposed approach is based on the collaborative classification of any cell i status, C i by itself and its neighbor cells. This implies that, for the analysis of possible failures in cell i , each \( {\mathrm{cell}}_j\in {\mathbf{cells}}_{\boldsymbol{i}}^{\mathbf{imp}} \) would be involved in the process, where \( {\mathbf{cells}}_{\boldsymbol{i}}^{\mathbf{imp}} \) is the set of sites that would be impacted by the possible failure: typically, cell i itself and its adjacent ones (as graphically represented in Fig. 1).

The simplified detection rule presented in Eq. (11) has as input a set of indicators calculated from the \( {\mathbf{cells}}_{\boldsymbol{i}}^{\mathbf{imp}} \). In this scheme, each cell j is in charge of computing an indicator \( {F}_{ij}^{\mathrm{AOI}}\in {\boldsymbol{F}}_{\boldsymbol{i}}^{\mathrm{AOI}} \). This is calculated from the measurements \( {M}_{ij}^{\mathrm{AOI}}\left[t\right] \), consisting on those measurements served by cell j in a certain AOI (e.g., ECov or ECent) of cell i . For example, the RSS_5th percentile of the UEs served by the cell j in ECov(cell i ). In this way, the samples are locally gathered and aggregated, minimizing signaling costs.

The complete self-healing distributed procedure is presented from the perspective of any cell j participating in the detection of a possible faulty cell i . cell j will apply the same process for any cell i such that \( {\mathrm{cell}}_j\in {\mathbf{cells}}_{\boldsymbol{i}}^{\mathbf{imp}} \). In the description of the different phases, the expressions and variables already described would be particularized for the distributed case:
  1. 1.
    Definition of \( \mathbf{cell}{\mathbf{s}}_{\boldsymbol{i}}^{\mathbf{imp}} \) and AOIs
    • In order to define the cells likely to be affected by a failure in a particular cell i , the approach is to automatically include in \( {\mathbf{cells}}_{\boldsymbol{i}}^{\mathbf{imp}} \) the cell i itself and its adjacent neighbors. The adjacent cells can be defined from the estimated coverage area maps, selecting the BSs whose coverage areas are in contact. \( {\mathbf{cells}}_{\boldsymbol{i}}^{\mathbf{imp}} \) set can also be updated based on the neighbor cell list of each cell, as they are automatically updated during network operation [31]. All the cells in the deployment should have knowledge of the different \( {\mathbf{cells}}_{\boldsymbol{i}}^{\mathbf{imp}} \) sets and their relative position in order to participate in the detection of problems of those sets where they are part of.

      Based on \( \mathbf{cell}{\mathbf{s}}_{\boldsymbol{i}}^{\mathbf{imp}} \) and their relative positions, the AOIs of cell i can be calculated (in a centralized or distributed way) and stored.

     
  2. 2.
    Training phase
    • As described in Section 4, cell j participating in the detection of a failure in cell i needs to be in possession of the prior likelihood of each status and the conditional PDFs: \( {\mathrm{pdf}}_{ij}^{\mathrm{AOI},\;{\mathrm{Normal}}_i}\left({r}_{ij}\right) \) and \( {\mathrm{pdf}}_{ij}^{\mathrm{AOI},\kern0.5em {\mathrm{Sleeping}}_i}\left({r}_{ij}\right) \) (where its previously presented nomenclature has been particularized for the indicator \( {F}_{ij}^{AOI} \)). During the training phase, the PDFs can be constructed and stored directly by cell j as \( {F}_{ij}^{\mathrm{AOI}} \) is locally generated by the BS. From these, the \( \Psi \left({\mathrm{F}}_{ij}^{\mathrm{AOI}}\right) \) parameter can be also calculated and stored. The prior likelihoods are assigned with a default or configured value.

     
  3. 3.
    Online phase
    • During the operational life of the network, the process is divided in different stages of computation and information sharing between the cells.

     
  4. 3.a
    Individual stage
    • This stage is described in Fig. 2. Firstly, cell j gathers the RSS samples reported by its served UEs. Secondly, the localization associated with each measurement is obtained directly from localization sources that can be the UEs themselves, a cellular-based positioning system or an external localization service [9]. Thirdly, with this information and the stored AOIs, the values of the \( {w}_{\mathrm{AOI}}^{{\mathrm{cell}}_i} \) are calculated for all the RSS samples and then the location-based indicator value \( {F}_{ij}^{\mathrm{AOI}}={f}_{ij}^{\mathrm{AOI}}\left[t\right] \) is obtained.
      Fig. 2

      Individual stage

      Fourthly, based on this and the conditional PDFs, the likelihoods of the current value given the status of cell i , \( \widehat{p}\left({F}_{ij}^{\mathrm{AOI}}={f}_{ij}^{\mathrm{AOI}}\left[t\right]\ \left|\mathrm{Normal}\right.\right) \) and \( \widehat{p}\left({F}_{ij}^{\mathrm{AOI}}={f}_{ij}^{\mathrm{AOI}}\left[t\right]\ \left|\mathrm{Sleeping}\right.\right) \) are calculated as well as \( {\varphi}_{ij}^{\mathrm{AOI}}\left[t\right] \) (see Eq. (13)). Also, the number of samples inside the AOI and used to generate the indicator are propagated to the following distribution stage.

     
  5. 3.b
     Distribution stage
    • Afterwards, each cell of \( {\mathrm{cells}}_i^{\mathrm{imp}} \) shares their estimated conditional probabilities with the rest of the cells of the set. This and the next stages are presented in Fig. 3.
      Fig. 3

      Distribution, computation, consensus and diagnosis stages

      The message from a cell might be not received due to incorrect timing, connection losses, or failure in the cell, which cannot be considered a univocal consequence of a sleeping cell failure (as described in Section 5). Therefore, the conditional probabilities for the indicator of that cell are assumed “1” for both status, which means that such input is not considered in the classifier.

     
  6. 3.c
     Computation stage
    • Having the conditional probabilities from the other cells of \( {\mathrm{cells}}_i^{\mathrm{imp}} \), the status of cell i can be calculated by any of them based on the naive Bayes classifier detection rule presented in Eq. (15), providing its estimated status.

     
  7. 3.d
     Diagnosis stage
    • If a neighboring cell is detected as sleeping, any other BS can check its NETACC in order to determine the particular cause behind the problem. This, together with the estimated status is used to specify the particular cause by means of binary logic following Table 2.

     
  8. 3.e
     Consensus
    • If the information has been properly received and computed by each BS, the results in terms of the estimated status shall be equivalent to all of them. However, this may not be the case if any distributed message is lost. Also, if the NETACC check provides different results (e.g., due to congestion) for different BSs.

      Therefore, consensus techniques can be applied for the results achieved by all of the BSs. In order to achieve a common diagnosis from the possible different results of each node, multiple mechanisms have been developed for the general field of distributed computation. For example, the selection of a master/coordinator cell dedicated to perform the final posterior probability calculations and then share them with the other cells can keep consensus as well as reduce the computational costs by freeing some of the BSs from the need of performing the classification [31]. However, the system becomes then more vulnerable to failures in such master cell. To avoid so, the use of strong consistency as presented in [34] is recommended. This is based on making the independent diagnosis performed by each cell consistent by sharing and checking their mutual results.

     
  9. 4.
    Compensation/recovery activity
    • Once the cell status is detected and diagnosed, compensation and recovery mechanisms can be triggered. For instance, readjusting cell powers to compensate a neighboring sleeping cell, rebooting automatically themselves if found faulty or alert the operator’s OAM system about the issue.

     

6.2 Implementation

The proposed distributed system would benefit for the direct interconnection of the cells, preferably by local high performance connections such as Ethernet LAN. That is the case for femtocell deployments, where the cells are often connected to the same router and/or access point to the Internet backhaul. There, local IP access (LIPA) [35] protocols can be used for over-the-top implementations of the proposed distributed approach. In a more standardized manner, the proposed communications can be integrated through the LTE/LTE-A defined X2 interface between the cells.

The distribution phase could also make use of multicast-type protocols, where the deployed cells can be grouped to reduce the need of addressing the messages to particular base stations.

Moreover, the proposed approach requires a level of synchronization between the cells in order to share the same monitoring and message exchange periods. Such synchronization is already present as part of the coordination between the cellular network elements, and the proposed system does not introduce any additional requirement.

7 Evaluation

For the evaluation of the proposed mechanisms, the system level simulator detailed in [36] was extended to integrate the proposed mechanisms as well as cell faults. The defined simulation parameters are included in Table 3. As baseline RAT, LTE is chosen. The radio environment is simulated by Winner II propagation model [26], where fast-fading is modeled by the Extended Indoor A model [37]. RSS-type information is provided by the LTE RSRP metric. The simulated scenario follows an expected typical 5G deployment comprising 12 small cells with a reduced average inter-site distance of 50 meters. Fully variable user distributions are emulated in a realistic airport model: Málaga airport, IATA code: AGP (ranked the 33rd busiest airport in Europe with 12.5 million passengers per year) is selected as a key scenario for the deployment (see Fig. 4). This indoor location is placed in a large external area (3 × 2.6 km2), where a macrocell station is located at 500 m from the airport.
Table 3

Simulation parameters

Propagation model

Indoor-indoor

Winner II A1

 

Indoor-outdoor

Winner II A2

 

Outdoor-outdoor

Winner II C2

 

Outdoor-indoor

Winner II C4

Base station model

EIRP

3 (small cells)/43 (macro) dBm

 

Directivity

Omni (small)/tri-sector (macro)

 

Access

Open (small)/open (macro)

Mobile station model

Noise figure

9 dB

 

Noise density

−174dBm/Hz

Traffic model

Calls

Poisson (avg. 0.43calls/user·h)

 

Duration

Exponential (avg. 100 s)

Mobility model

Outdoor

3 km/h, random direction, and wrap-around

 

Indoor

Random waypoint

Service model

Voice over IP

16 kbps

 

Full buffer

 

RRM model

Bandwidth

1.4 MHz (6 PRBs)

 

Access control

Directed retry (threshold = −44dBm)

 

Cell reselection

Criteria S, R

 

Handover

Events A3, A5

 

Scheduler

Voice: round-robin best channel

Full buffer: proportional fair

Time resolution

 

100 ms

Load balancing algorithm

Epoch time

60 s

Fig. 4

Evaluation scenario

In the simulation, it is assumed that UEs report RSRP values and locations once per second. This is consistent with the signaling constraints of possible over-the-top implementations [9], whereas control plane solutions may allow higher frequency reporting (as 0.1 s considered in [12]).

Catatonic sleeping cells are modeled in the simulation in the following way: for each possible faulty cell 400 continuous monitoring periods of 1 m are simulated: first 200 periods corresponding to a normal case situation (where the cell works properly), while the following 200 periods model a sleeping cell failure of one BS that stops transmitting. During these periods, the movement of users along the airport is modeled based on random waypoint including realistic users’ concentration in security checkpoints and boarding gates areas.

7.1 Impact of sleeping cell case in classic performance indicators

For the presented set-up, Fig. 5 shows the call blocking ratio (CBR); this means, the percentage of calls not able to access the cellular network. CBR is commonly used as an indicator of network accessibility and applied on the analysis of possible network failures. The figure reflects the common case where a sleeping cell failure (in this case, cell 11) cannot be spotted by purely classical performance indicators.
Fig. 5

Call blocking ratio time evolution for normal and failure periods and different user concentrations

On the one hand, Fig. 5, top, presents a situation with a high number of average users per cell (around 12 active users simultaneously). In this case, the CBR of the neighboring cell 12 highly increases after the failure of cell 11. However, it is observed that there are some peaks of high CBR even in normal status, as well as quite few periods with low CBR under failure condition.

On the other hand, Fig. 5, bottom, presents a situation with a reduced number of users (an average of 7 UEs per cell). For this case, the CBR is not impacted at all by the failure, making impossible its detection based on this metric.

7.2 RSS indicators and AOIs

Hereafter, the indicators proposed for the RSS-based detection are assessed with different restrictions in terms of the area selected and the distributed/centralized approach, particularized for the case where cell 11 is faulty (\( {\mathrm{cells}}_{11}^{\mathrm{imp}}=\left\{9,10,11,12\right\} \)).

Figure 6 compares the classic RSS 5th percentile values for the UE served by the cells in \( {\mathrm{cells}}_{11}^{\mathrm{imp}} \) with the ones obtained applying ECov and ECent (with k R  = 0.75) calculated following the simple log-distance propagation approach and the sample weights defined in Eq. (9) for any serving cell. The results for both ECov and ECent are assessed and compared as alternative AOI definitions, where ECent represents a far more restricted area than ECov.
Fig. 6

Location-based fifth percentile RSRP indicators for different AOIs

It can be seen how the classic indicator do not show any clear impact due to the failure in most of the loops. However, the variation is evident for the ones using location ECent thanks to the selection of the most relevant samples and their weighting. Also it could be observed how some periods (e.g., 106, 114, and 254) have no values for the ECent indicator due to the lack of UEs in the area. This indicates how reducing the AOI increases the visibility of the failure impact but also increments the possibility of not having enough samples to generate the metric.

7.2.1 Detection performance

The proposed detection Bayes scheme is applied for location and non-location indicators as well as centralized (not filtering the RSS measurements by serving cell) and distributed approaches. In the evaluation, the training phase is performed using as calibration set 50 periods under each cell status, where these periods are considered as previously labeled. As commented, in real deployments the training set can be obtained from real failure recorded cases, by an operator defined ON/OFF calibration phase or from neighboring cells measurements analysis as indicated in Section 4.A. The conditional PDFs are therefore obtained from this training set. A value of |M|th = 10 is also established. RSS 5th percentile statistics are selected as inputs for the classifier, as they show a higher variation between the normal and sleeping cases than the indicators based on the RSS mean.

Common figures of merit for detection are the false normal rate (FN), being the ratio of faulty periods identified as normal; false alarm rate (FA), the percentage of normal periods identified as faulty; and the inconclusive rate (IN), or the percentage of periods where there is no UE RSS measurements in the considered AOI and no detection is performed.

Figure 7 shows the FA, FN, and IN values for the different combinations of detection options and for the case where cell 11 is analyzed. Non-location local refers to the classic detection performed based uniquely on the 5th percentile of the UE served by the most affected neighbor (cell 12 in this case). Non-location centralized refers to the classic approach of using the 5th percentile of the RSRP values gathered by all UEs served by \( {\mathrm{cells}}_{11}^{\mathrm{imp}} \). Non-location distributed follows the proposed distribution algorithm but applied over-the-classic RSRP indicators from the UE served by each cell. Centralized ECov and ECent make use of the proposed location-based approach generating the indicators based on the samples gathered in the AOIs without distinguishing their serving cell. Finally, distributed ECov and ECent follows the complete proposed location-based distributed approach presented in Section 6.
Fig. 7

Detection error rates for cell 11 failure

For non-location-based approaches, it can be observed how the use of the distributed approach highly increases the accuracy of the detection, achieving FA = 0 % with respect to the centralized and local approaches. However, the similarity between the classic RSRP values for both normal a faulty status causes a high FN = 21 %. The use of localization highly improves this aspect: while the use of ECov does not introduce significant improvement in the detection due to its wide area, ECent achieves FN = 1 %, highly outperforming previous approaches.

The study is extended to the rest of the central cells of the scenario {9, 10, 11, 12}, where the failure of each of them is modeled. The resultant detection performances in percentage are shown in Table 4. It is observed how the results of the distributed location-based proposed approach highly outperformed those of the classic mechanisms except for the cell 12 case. This low performance is due to the complex wall environment surrounding such station, which makes that even in normal operation, cell 11 serves a high amount of users in the ECent of cell 12. This indicates how more complex definitions of the AOIs (such as those based on the described detailed propagation models) might be required where these situations are present in the deployment.
Table 4

Summary of detection performance for different cells

Cell

Cell 9

Cell 10

Cell 11

Cell 12

Indicator type\figure of merit

FN

FA

IN

FN

FA

IN

FN

FA

IN

FN

FA

IN

Non-location local

69.0

25.6

0.0

13.5

35.2

0.0

17.5

44.7

0.0

39.5

19.1

0.0

Non-location centralised

46.0

45.7

0.0

19.0

29.1

0.0

33.5

28.1

0.0

48.5

9.5

0.0

Non-location distributed

22.0

0.0

0.0

14.0

0.0

0.0

21.0

0.0

0.0

35.0

0.0

0.0

Centralised ECov

13.0

15.1

10.5

3.5

0.5

2.8

31.5

37.7

0.0

35.0

21.1

0.0

Centralised ECent

7.0

0.0

40.1

2.5

0.5

17.8

21.5

0.5

3.3

63.5

0.0

4.3

Distributed Ecov

3.5

0.5

10.8

0.0

0.0

2.8

21.5

0.0

0.0

57.0

0.0

0.8

Distributed Ecent

0.0

0.0

40.6

0.0

0.0

18.0

1.0

0.0

3.5

53.5

0.0

5.0

For the rest of the cells, the distributed approaches highly improve the results of their equivalent centralized or local methods. Moreover, the use of location-based indicators provides far better performance than classic ones. Also, the use of the narrower ECent instead of ECov allows reduced FA and FN at the cost of increasing IN, due to the minor number of available measurements in the AOI.

7.3 Impact of UE localization error

The previous evaluation demonstrates the capabilities of the location-based approach to improve sleeping cell detection mechanisms. However, one of the main characteristics to take into account for its applicability is its robustness to localization inaccuracies. These are modeled in the simulator by an added Zero-mean Gaussian noise to the real location of each terminal before feeding the self-healing algorithms with this data, following the expression:
where e u x and e u y are the added error distance for each coordinate, and μ and σ are respectively the mean and deviation of the normal distribution Ɲ, being consistent to the most common approach for error modeling of positioning methods [38]. In this way, the system is tested for different values of the location error standard deviation σ = {0, 0.5, 1, 2, 3, 5, 7, 10}, which is equivalent to the root mean squared position error (RMSE) of the positioning technique. Such error levels cover from very precise indoor positioning solutions as UWB-based (e.g., reference [15] proposed technique achieves around 1-m mean accuracy in real experiments) to just an extremely rough estimation, including more common positioning mechanisms as WiFi-based (e.g., the system developed in [14] showed mean errors of 7 m in real scenarios).
The average results for cell 11 given those errors are presented in Fig. 8 for the ECent distributed non-location and location-based techniques. FA is not included in the graph as it is always 0. The figure shows the robustness of the location-based approach even for large inaccuracies, keeping a much better performance than non-location ones even for very high-positioning errors as 7 m.
Fig. 8

Detection performance given different location error

8 Conclusions

This work has assessed the use of classic network indicators for sleeping cell failure detection in dense networks, showing the reduced impact that they might reflect. This would commonly lead to delay or wrong detection of network issues based on those indicators.

To overcome this issue, the use of location-supported metrics has been proposed. Such metrics highly increase the capability to detect cell failures. In this way, the proposed location-based indicators support the design of troubleshooting systems with improved failure detection than previous schemes. These have been integrated in a newly defined cell distributed detection and diagnosis algorithm, where the required associated architecture and methodology have been defined.

The proposed system has been assessed with a system level simulator that models a representative DenseNet indoor scenario. The evaluation has shown how the proposed approach highly outperforms the results achieved by classic mechanisms and by centralized implementations. The behavior of the algorithm in the presence of inaccuracies in the localization information has also been analyzed, showing high robustness for the normal levels of inaccuracies found in existing positioning systems.

Future works will analyze the extension of the proposed approach to scenarios including additional heterogeneity, combining different cellular and non-cellular systems. Also, the impact of additional 5G radio conditions, such as higher frequency radio links, will be further explored.

9 Abbreviations

3GPP, Third Generation Partnership Project; CBR, call blocking ratio; CQI, channel quality indicator; DenseNet, Dense Network; ECent, expected center area; ECov, expected coverage area; EIRP, equivalent isotropically radiated power; FA, false alarm rate; FN, false normal rate; IATA, International Air Transport Association; IN, inconclusive rate; KPI, key performance indicators; LTE, long-term evolution; NGMN, next-generation mobile networks; PDF, probability density function; PRB, physical resource block; RAT, radio access technology; RFID, radio-frequency identification; RRM, radio resource management; RSRP, reference signal received power; SINR, signal to interference-plus-noise; SON, self-organizing networks; UE, user equipment; UWB, ultra-wideBand; VoIP, voice over IP

Declarations

Acknowledgements

This work has been funded by Junta de Andalucía (Proyecto de Investigación de Excelencia P12-TIC-2905), and Spanish Ministry of Economy and Competitiveness / FEDER (TEC2015-69982-R and Network of Excellence Red ARCO 5G, TEC2014-56469-REDT).

Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Authors’ Affiliations

(1)
Universidad de Málaga, Andalucía Tech, Departamento de Ingeniería de Comunicaciones

References

  1. “What is 5G? 5G visions”. GSM history: history of GSM, mobile networks, vintage mobiles. Online: http://www.gsmhistory.com/5g/. Accessed 14 June 2016
  2. D Lopez-Perez, M Ding, H Claussen, AH Jafari, Towards 1 Gbps/UE in Cellular Systems: Understanding Ultra-Dense Small Cell Deployments. IEEE Commun. Surv. Tutorials 17(4):2078–2101 (2015). doi:10.1109/COMST.2015.2439636 View ArticleGoogle Scholar
  3. “Small cells—what’s the big idea? Femtocells are expanding beyond the home,” White Paper, Small Cell Forum, 2012. http://smallcellforum.org/smallcellforum/Files/File/SCF-Small_Cells_White_Paper.pdf
  4. 3GPP, Telecommunication management; principles and high level requirements, TS 32.101, 3rd Generation Partnership Project (3GPP) (2015). http://www.3gpp.org/ftp/Specs/html-info/32101.htm
  5. 3GPP, “Telecommunication management; self-organizing networks (SON); self-healing concepts and requirements,” 3rd Generation Partnership Project (3GPP), TS 32.541, 2016. http://www.3gpp.org/DynaReport/32541.htm
  6. F Chernogorov, T Ristaniemi, K Brigatti, S Chernov, “N-gram analysis for sleeping cell detection in LTE networks,” 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC. 2013, pp. 4439–4443. doi: 10.1109/ICASSP.2013.6638499
  7. T Perry, The Indoor Navigation Battle Heats Up IEEE Spectrum magazine, (IEEE, New York, US, 2012). http://spectrum.ieee.org/tech-talk/consumer-electronics/portable-devices/the-indoornavigation-battle-heats-up Google Scholar
  8. M Molina-García, J Calle-Sánchez, JI Alonso, A Fernández-Durán, FB Barba, Enhanced in-building fingerprint positioning using femtocell networks. Bell Labs Tech. J. 18(2), 195–211 (2013)View ArticleGoogle Scholar
  9. S Fortes, A Aguilar-García, R Barco, F Barba, J Fernández-luque, A Fernández-Durán, Management architecture for location-aware self-organizing LTE/LTE-a small cell networks. IEEE Commun. Mag. 53(1), 294–302 (2015)View ArticleGoogle Scholar
  10. A Aguilar-Garcia, S Fortes, M Molina-García, J Calle-Sánchez, JI Alonso, A Garrido, A Fernández-Durán, R Barco, Location-aware self-organizing methods in femtocell networks. Comput. Netw. 93(1), 125–140 (2015)View ArticleGoogle Scholar
  11. R Barco, P Lazaro, P Munoz, A unified framework for self-healing in wireless networks. IEEE Commun. Mag. 50(12), 134–142 (2012)View ArticleGoogle Scholar
  12. W Wei, L Qing, Z Qian, COD: a cooperative cell outage detection architecture for self-organizing femtocell networks. IEEE Trans. Wirel. Commun. 13(11), 6007–6014 (2014)View ArticleGoogle Scholar
  13. K Bouchard, D Fortin-Simard, S Gaboury, B Bouchard, A Bouzouane, Accurate trilateration for passive RFID localization in smart homes. Int. J. Wireless Inf. Networks 21, 32–47 (2014)View ArticleGoogle Scholar
  14. M Raitoharju, H Nurminen, R Piché, Kalman filter with a linear state model for PDR+WLAN positioning and its application to assisting a particle filter. EURASIP J. Adv. Signal Process. 2015, 1–13 (2015). doi:10.1186/s13634-015-0216-z View ArticleGoogle Scholar
  15. P Muller, H Wymeersch, R Piche, UWB positioning with generalized gaussian mixture filters. IEEE Trans. Mob. Comput. 13(10), 2406–2414 (2014). doi:10.1109/TMC.2014.2307301 View ArticleGoogle Scholar
  16. G-NetTrack phone measurement capabilities. [Online]. Available: http://www.gyokovsolutions.com/survey/surveyresults.php. Accessed 14 June 2016
  17. S Fortes, AA Garcia, JA Fernandez-Luque et al., Context-aware self-healing: user equipment as the main source of information for small-cell indoor networks. IEEE Veh Technol Mag 11, 76–85 (2016). doi:10.1109/MVT.2015.2479657 View ArticleGoogle Scholar
  18. ERICSSON, “App Experience Optimization,” Webpage. [Online]. Available: http://www.ericsson.com/ourportfolio/services/app-experience-optimization?nav=fgb_101_108|fgb_101_592. Accessed 14 June 2016
  19. 3GPP, “Universal Mobile Telecommunications System (UMTS); Physical layer; Measurements (TDD),” 3rd Generation Partnership Project (3GPP), TS 25.225, 2015. http://www.3gpp.org/DynaReport/25215.htm
  20. 3GPP, “LTE; Evolved Universal Terrestrial Radio Access (E-UTRA); Requirements for Support of Radio Resource Management,” 3rd Generation Partnership Project (3GPP), TS 36.133, 2014.Google Scholar
  21. R Groves, F Fowler, M Couper, J Lepkowski, E Singer, R Tourangeau, Survey Methodology, ser. Wiley Series in Survey Methodology (Wiley, United States, 2009).Google Scholar
  22. S Fortes, R Barco, A Aguilar-García, P Muñoz, Contextualized indicators for online failure diagnosis in cellular networks. Comput. Networks 82, 96–113 (2015)View ArticleGoogle Scholar
  23. E Langford, Quartiles in elementary statistics. J. Stat. Educ. 14, 3 (2006)Google Scholar
  24. Small Cell Forum, Enterprise femtocell deployment guidelines, 2013Google Scholar
  25. B Tianyang, RW Heath, Location-specific coverage in heterogeneous networks. IEEE Signal Process Lett. 20(9), 873–876 (2013)View ArticleGoogle Scholar
  26. WINNER II IST project, “D1.1.2. WINNER II channel models. part II. radio channel measurement and analysis results. v1.0,” WINNER II IST project, Tech. Rep., 2007. [Online]. Available: www.ist-winner.org/. Accessed 14 June 2016
  27. K Hormann, A Agathos, The point in polygon problem for arbitrary polygons. Comput. Geom. 20(3), 131–144 (2001)MathSciNetView ArticleMATHGoogle Scholar
  28. S Hämäläinen, H Sanneck, C Sartori, LTE Self-Organising Networks (SON): Network Management Automation for Operational Efficiency (Wiley, United States, 2011)Google Scholar
  29. AW Bowman, A Azzalini, Applied smoothing techniques for data analysis (Oxford University Press Inc., New York, 1997)MATHGoogle Scholar
  30. WP Gajewski, Adaptive naïve Bayesian anti-spam engine. Int. J. Inf. Technol. 3(CERN-OPEN-2007-011), 153–159 (2006)Google Scholar
  31. CH Lee, F Gutierrez, D Dou, “Calculating feature weights in naive Bayes with Kullback-Leibler measure,” 2011 IEEE 11th International Conference on Data Mining, Vancouver,BC, pp.1146-1151, 2011. doi: 10.1109/ICDM.2011.29
  32. H Zhang, The optimality of naive bayes, in FLAIRS Conference, 2004, pp. 562–567Google Scholar
  33. MS Nikulin, Hellinger Distance, in Encyclopedia of Mathematics, Springer, ed. by M Hazewinkel, 2001. ISBN 978-1-55608-010-4Google Scholar
  34. NL Tran, Q Dugauthier, S Skhiri, "A Distributed Data Mining Framework Accelerated with Graphics Processing Units," Cloud Computing and Big Data (CloudCom-Asia), 2013 International Conference on, Fuzhou, 2013, pp. 366–372. doi: 10.1109/CLOUDCOM-ASIA.2013.17
  35. 3GPP, “3rd Generation Partnership Project; Technical Specification Group Services and System Aspects; Local IP Access and Selected IP Traffic Offload (LIPA-SIPTO) (Release 10),” 3rd Generation Partnership Project(3GPP), TS 23.289, 2011. http://www.3gpp.org/DynaReport/23829.htm
  36. JM Ruiz-Avilés, S Luna-Ramírez, M Toril et al., Design of a computationally efficient dynamic system-level simulator for enterprise LTE femtocell scenarios. J. Electr. Comput. Eng. 2012, Article ID 802606 (2012)MathSciNetGoogle Scholar
  37. T Sorensen, P Mogensen, F Frederiksen, Extension of the ITU channel models for wideband (OFDM) systems. IEEE Veh. Technol. Conf. 1, 392–396 (2005). DallasGoogle Scholar
  38. T Hillebrandt, H Will, M Kyas, in Progress in Location-Based Services, ed. by MJ Krisp (Springer Berlin Heidelberg, Berlin, 2013), pp. 173–194View ArticleGoogle Scholar

Copyright

© Fortes et al. 2016