THE exponential growth of wireless data services driven by

Transcription

1 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 67, NO. 12, DECEMBER Machine Learning Aided Context-Aware Self-Healing Manageent for Ultra Dense Networks ith QoS Provisions Meng Qin, Qinghai Yang, Meber, IEEE, Nan Cheng, Meber, IEEE, Haibo Zhou, Senior Meber, IEEE, Raesh R. Rao, Fellow, IEEE, and Xuein Shen, Fellow, IEEE Abstract The self-organizing network is envisioned as a key technology to future wireless networks, especially for densely deployed sall cell scenarios. Self-healing (SH) is an essential functionality to allow the networks to autoatically detect and copensate for cell outages, which typically occur when unexpected network failures arise. In this paper, reaping the benefits of achine learning, we propose a novel SH fraework in ultra dense sall cell networks for eeting the deands of low-cost and fast network operation, quality of service (QoS), and energy efficiency. The proposed SH schee coprises sall cell outage detection (SCOD) and sall cell outage copensation (SCOC) to enable self-healing in ultra dense sall cell networks. Based on the context inforation of the partial key perforance indicator (KPI) statistics, we propose a novel SCOD algorith to detect the outage by applying support vector data description (SVDD) approach. The SCOD algorith detects a sall cell outage efficiently considering two situations: KPIs available situation and non-kpis available situation. Furtherore, in order to copensate the sall cell outage, SCOC is forulated as a network utility axiization proble to optially copensate for the outaged zone in sall cell network. A distributed copensation algorith with low coputational coplexity is developed to balance the load of sall cell networks, considering the QoS provision for users. Siulation results deonstrate that the proposed SH schee can detect the sall cell outage efficiently and can achieve an optiized QoS perforance when copensating for the detected sall cell outage. Manuscript received May 16, 2018; revised Septeber 12, 2018; accepted Septeber 27, Date of publication October 24, 2018; date of current version Deceber 14, This work was supported in part by National Natural Science Foundation of China under Grant , in part by 111 Project (B08038), in part by the National Research Foundation of Korea (MSIP) under Grant NRF-2014K1A3A1A , and in part by the Natural Sciences and Engineering Research Council (NSERC), Canada. The review of this paper was coordinated by Dr. A.-C. Pang. (Corresponding author: Qinghai Yang.) M. Qin and Q. Yang are with State Key Laboratory of ISN, School of Telecounications Engineering, and also with Collaborative Innovation Center of Inforation Sensing and Understanding, Xidian University, No.2 Taibainanlu, Xi an, , Shaanxi, China (e-ail:, engqin@stu.xidian. edu.cn; qhyang@xidian.edu.cn). H. Zhou is with the School of Electronic Science and Engineering, Nanjing University, Nanjing , China (e-ail:,haibozhou@nju.edu.cn). R. R. Rao is with the California Institute for Telecounications and Inforation Technology (CALIT2), University of California at San Diego, La Jolla, CA USA (e-ail:,rrao@ucsd.edu). N. Cheng and X. Shen are with the Departent of Electrical and Coputer Engineering, University of aterloo, aterloo, ON N2L 3G1, Canada (e-ail:, n5cheng@uwaterloo.ca; sshen@uwaterloo.ca). Color versions of one or ore of the figures in this paper are available online at Digital Object Identifier /TVT Index Ters Self-organizing network (SON), self-healing (SH), cell outage, achine learning, key perforance indicator (KPI), load balancing, ultra dense networks. I. INTRODUCTION THE exponential growth of wireless data services driven by obile Internet and sart devices, which leads to dense deployent of sall cell networks, has triggered the investigation of the network planning, the anageent and the perforance optiization for a better user quality of service (QoS) in future wireless networks, especially for 5G cellular networks [1], [2]. It is notable that the coplexity of operations, capital expenditure (CAPEX) and operational expenditure (OPEX), is the ajor challenge for operators of future cellular networks [3], [4]. Particularly, with increasing scale of 5G networks especially for ultra dense sall cell scenarios, new approaches of autoatic detection and copensation with high efficiency are required to cope with the risks of outage owing to various kinds of hardware or software failures in wireless networks and to reduce the operation cost due to a ass of configuration operation paraeters, which is a great challenge to anage 5G network efficiently. Self-organizing network (SON) has recently been recognized as an attractive paradig for the 5G networks, which enables autonoic features, including self-configuration, selfoptiization and self-healing (SH) [5]. The ain task of SH functionality is twofold, i.e., autonoous cell outage detection and cell outage copensation. Cell outage detection ais to autonoously detect outaged cells, and cell outage copensation adjusts the paraeters of nearby cells to recover the service of users in the outage cells. The ain advantage of SH functionality is that SH significantly reduces the tie to detect and copensate the cell outage autoatically, which could take hours for anual operations. This is especially iportant with the ultra dense sall cell deployent in 5G networks, where anual operations are highly costly due to the huge aount of unplanned network deployent [6] [9]. There are extensive research works focusing on SH in wireless networks [10] [18]. The cell fault identification algoriths are designed based on onitoring of Reference Signal Received Power (RSRP) and Channel Quality Indicator (CQI) easureents in [10], [11]. Miniization of drive test (MDT) easureents are used to both profile network behavior and IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See standards/publications/rights/index.htl for ore inforation.

2 12340 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 67, NO. 12, DECEMBER 2018 detect sleeping cells in [12], [13], which have been studied in Release 9 by the 3GPP in order to reduce the operation cost for drive test. Owing to the key advantages of artificial intelligence, researchers have applied ethods fro the achine learning such as clustering algoriths and Bayesian networks to autoate the detection of faulty cell behavior [14] [18]. A genetic algorith is proposed for learning the outage rules in [14]. An autoated cell outage detection echanis based on dynaic affinity propagation clustering algorith is introduced in [15]. Furtherore, the cell outage copensation is perfored during the tie period when the network fault happens. The proble of cell outage copensation is studied in [16] [18]. A cell outage copensation ethodology is proposed by adapting different outage copensation strategies to different cell outage situations. In [17], a fraework for itigating the adverse ipact of dynaic outages is proposed while taking into account both the channel characteristics and enodeb s residual resources. In [18], a novel pre-planned reactive cell outage copensation approach is presented to itigate the effects of fronthaul failure. Furtherore, contextual inforation (e.g., user location and network state inforation) is introduced for iproving anageent of wireless networks [19]. Although SH is extensively studied in the literature, there are still soe challenges, especially for the ultra dense sall cell networks. First, cell outage detection and cell outage copensation algoriths have been independently presented. However, it is necessary to jointly consider the cell outage detection and cell outage copensation processes for quick and flexible response to the network operation anageent. Such a coplete SH fraework is still issing. Second, the evolving 5G networks will be featured by ultra densely deployed sall cells, with the benefits of covering holes, offloading traffic, and increasing energy efficiency [20], [21]. Sall cells are particularly prone to failures due to its unplanned assive deployent, ore hardware, and plenty of configuration paraeters. Thereby, typical SH solutions in acrocell cannot be directly applied in ultra dense sall cell scenario due to that sall cell networks have reduced onitoring functions and liited coputing capabilities. Moreover, it is very coon in SH anageent that no KPI inforation (e.g., alar or perforance degradation inforation) ay be reported to network operation and anageent (OAM) syste in ultra dense sall cell networks, due to unplanned deployent and liited reporting capability. The difficulty that knowing the effects of each outage cause when the KPIs are not available, not only ipacts the design of SH schee but also liits the outage detection efficiency under realistic conditions. Hence, an efficient SH schee is urgently needed to deal with the outage situation that no KPIs are available. In addition, the outage copensation algoriths in the literature, were developed for optiizing the capacity and coverage of the identified outage cell zone but ignoring the optial load balance and different users QoS requireents, which is of vital iportance to sall cell network perforance. Therefore, the SH echanis with partial KPI inforation needs further investigation, considering the load balancing and users QoS requireents. There have been several research works that study the SH schee in sall cell networks. The authors investigated the proble in sall cell networks without considering the context of MDT reports in [10] [18]. Research works in [22] [24] ainly focus on outage detection, which require fairly large network easureent data of users, leading to high coputational cost. However, to the best of our knowledge, there has not been a coprehensive SH schee that is designed for the ultra dense sall cell networks with low coputational coplexity, and can deal with the partial KPI situation. In this paper, we investigate the SH proble in SON-based ultra dense sall cell networks, where KPI inforation of soe sall cells ay not be available. e propose a coprehensive SH schee including both sall cell outage detection echanis and sall cell outage copensation echanis, and is capable of dealing with partial KPI situations. Specifically, the proposed SH schee includes SCOD stage and SCOC stage, where in SCOD stage the outage is detected, followed by the SCOC stage to copensate the users in the outage zones. For the SCOD stage, we propose a SCOD algorith by applying support vector data description approach (SVDD), considering partial KPIs statistics. An effective SVDD algorith with low coputational coplexity is developed to detect and locate the outaged sall cells with the context of KPIs and user position inforation. The proposed SVDD algorith has the tie coplexity and space coplexity of O(k 3 ) and O(k 2 ), respectively, with k being the saple nuber. For the SCOC stage, to copensate the users in detected outage zone, we propose a novel distributed resource allocation algorith that ais to optiize load balancing of the detected outage sall cells area. The SCOC schee is designed to allocate resources of neighboring sall cells to the outage users, considering the dynaics and density of sall cell environents. Specifically, we forulate resource allocation as a ixed integer optiization proble, solved by Lagrangian dual theory to guarantee both load balancing and QoS requireents of users. The ain contributions of this paper are suarized as follows. e propose a novel self-healing fraework for ultra dense sall cell networks based on achine learning approach that jointly considers outage detection and copensation, even when only partial KPI inforation is available. e develop a low-coplexity SCOD algorith to detect and locate the outage sall cells with the context inforation of KPIs and user position, in which both isconfigurations and sleeping cells can be accurately detected. e propose a distributed copensation algorith for outaged sall cells guaranteeing load balancing in sall cell networks, which is very practical and can be applied in real systes. The reainder of this paper is organized as follows. Section II presents the syste odel. In Section III, the proposed SCOD algorith based on SVDD approach is illustrated, and we analyze its effectiveness. Section IV presents a novel resource allocation algorith for outage copensation. Section V presents the siulation results. Finally, Section VI concludes the paper. II. SYSTEM MODEL e consider the ultra dense sall cell networks scenario with partial KPI inforation. In this section, we present the syste

3 QIN et al.: MACHINE LEARNING AIDED CONTEXT-AARE SH MANAGEMENT FOR ULTRA DENSE NETORKS ITH QoS PROVISIONS Fig. 2. The self-healing fraework of SON-based sall cells. Fig. 1. SON-based sall cell syste odel. odel of the paper, including the self-healing architecture, network odel, and channel odel. A. Coprehensive SH Architecture e consider a scenario with a set of sall cells, and focus on SH in SON-based sall cell networks, where sall cells are connected to a centralized unit (for OAM) as shown in Fig. 1. The sall cell base station (SCBS) which ay experience cell outage with a certain probability in the process of operation, can share inforation for cooperation through OAM. The SCBS transits reference signals periodically on downlink. The reference signals, which facilitate users channel easureents (e.g., the RSRP and reference signal received quality (RSRQ) easureents), are sent back to the SCBS as feedback essages [23]. Particularly, the outaged SCBS cannot transit or receive any signals. A coprehensive SH architecture including both SCOD and SCOC is proposed, as shown in 2, i) In the SCOD stage, the OAM collects MDT reports fro cells to build a MDT database, and the SCOD algorith onitors KPIs profiled by MDT reports to deterine if the sall cells experience probles or failures. Then, the probles can be located based on the contextual inforation of sall cell locations if there are failures occurred. ii) In the SCOC stage, the SCOC algorith is executed to copensate the affected users by noral sall cells and produce optial copensation policy for the sall cell networks. Based on this architecture, the SCOD phase and the SCOC phase are elaborated in detail in Section III and Section IV, respectively. B. Network Model e consider a set of sall cells S = {1, 2,...,S} and a set of users N = {1, 2,...,N} in sall cell networks. e assue that perfect CQI is acquired via OAM and the identities of users previously served by each SCBS are recorded in OAM. The assuption is reasonable since the overhead of reporting CQI and user inforation is necessary to guarantee the network perforance [16] [18]. Therefore, the users can be identified for copensation when the serving SCBS becoes faulty. e also assue that users whose services are disconnected should find the preable or pilot of a neighboring SCBS. In this situation, they can obtain the inforation about cooperative resource allocation, and can be served continuously by noral SCBS. C. Channel Model In the sall cell network scenario, the channel gain of user u to SCBS s is deterined based on the odel described in [26]: ( ) q d0 g u,s = e X u,s e Y u,s, (1) d u,s where d 0 is the reference distance, d u,s is the distance between the SCBS and user u, and q denotes the path loss exponent. e X u,s and e Y u,s are the shadow fading factor and ulti-path fading factor, respectively. The shadowing fading follows a Gaussian distribution described by X u,s N(0,δ). The ulti-path fading is odeled by Rayleigh fading with zero ean 1 [23]. III. THE SCOD SCHEME ITH CONTEXT INFORMATION OF MDT MEASUREMENTS In this section, we propose a sall cell outage detection fraework considering the partial KPIs 2 easured on every SCBS by neighboring sall cells, using the MDT report acquired fro a fault-free operating scenario to profile the behavior of the sall cell network. Particularly, each MDT report is tagged with the context of the location and the tie inforation, which is regarded as contextual inforation. The goal of the proposed outage detection is to detect the cells outage accurately and efficiently. To this end, two stages are involved: a trigger stage with no inter-cell counication, and a cooperative detection stage with high accuracy and low delay. In the trigger stage, each sall cell collects the MDT easureents reported by users, and the OAM server initiates the detection stage to ake a final decision. In the detection stage, the detection algorith is executed to detect outages in sall cells. The procedure is illustrated in Fig As shown in [23], shadowing fading effects are assued to be independent of each other over tie. ith this assuption, the RSRP statistics of a user are independent rando variables and can be characterized by (1). 2 The OAM onitors the values of different KPIs for each cell, including the situation that the lack of KPIs for a certain cell, so we collect partial KPIs.

4 12342 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 67, NO. 12, DECEMBER 2018 TABLE I KEY PERFORMANCE INDICATORS finding a sphere with inial volue to contain all target data (the noral KPIs) [28], in which the sphere (the sphere eans the doain consisting of the noral KPIs in sall cell area) described as the objective function is characterized by center ϖ and radius Φ. The outaged cell detection proble is transfored to find an optial sphere by iniizing Φ 2 based on the collected KPI data saples as follows. in Φ 2 A. Sall Cell Outage Profiling and Triggering Users can report easureents like RSRP and RSRQ to sall cells, which could priely reflect the syste perforance. All KPIs are collected by virtue of the MDT reporting schee, which has been defined in LTE Release 10 [27]. The MDT configuration report procedure consists of configuration, easureents, report and storing phase, and the schee flow has shown in [25]. In the outage profiling stage, we obtain KPIs inforation (as shown in Table I) fro the reference sall cell scenarios to build the database, which is used to learn the sall cell network s profile (the perforance of sall cell networks). In addition, for the profiling phase, the tracing KPI database is processed to extract the feature vector D KPI, by storing the ebedded easureents that represent the noral operation of the sall cell networks, which is expressed in (2), shown at the botto of this page. In (2), subscript s indicates the serving sall cell, and subscript c1 cn denotes n neighboring cells of the serving cells. This reference database is used in the cell outage detection algorith to learn the noral sall cell network profile. The goal of the SCOD algorith is to define an anoaly detection rule that can differentiate between noral and abnoral MDT easureents by coputing a threshold. At first, the KPI data set is pre-processed to be noralized, and then it is taken as input paraeters for cell outage detection algorith. The algorith naed SCOD is proposed on the basis of a achine learning technique, called SVDD algorith, which is inspired by Support Vector Machine (SVM). SVM is a powerful data-driven approach for fault detection and diagnosis, which can use a hypothesis space of linear functions in a high diensional feature space and can be trained with a learning algorith fro optiization theory [28]. The SVDD algorith is able to for a decision boundary around the learned KPI data doain with very little (or even zero) inforation fro outside the boundary, which is considered as outliers. The SVDD ethod originates fro the idea of s.t. x i ϖ Φ 2,i= 1,...n (3) where x i,i= 1, 2,...,nis the i-th exaple in the MDT easured KPI data set D KPI, and the constraint in (3) indicates that all of the sapled KPI data should be contained by the sphere. To ensure the possibility of outliers in the training set based on the collected KPIs data saples, the distance fro x i to center ϖ should not be strictly saller than Φ 2, and a larger distance for D KPI should be penalized. Hence, we introduce the slack variables ξ i 0 to rewrite the proble as in f(φ,ϖ,ξ)=φ 2 + C n ξ i, s.t. x i ϖ Φ 2 + ξ i, x i S, ξ i 0, (4) where C is a particular constant paraeter, which ais to control the trade-off between the volue of the sphere and the errors that naely the fraction false positives (outliers accepted) and the fraction false negatives (targets rejected). Hence, the outaged sall cell detection proble is forulated as a convex quadratic optiization proble with convex constraints, and we introduce Lagrange ultipliers in (5), shown at the botto of this page. α i 0 and γ i 0 are Lagrange ultipliers. Then, setting partial derivatives to zero gives the constraints as L Φ = 2Φ α i 2Φ =0 α i = 1, i L ϖ = α i (2x i 2ϖ) =0 ϖ = i α ix i i α = α i x i, i i C α i γ i = 0, i = 1,...,N, 0 α i C. (6) D KPI =[RSRP s,rsrp c1,...rsrp cn,rsrq s,rsrq c1,...rsrq cn, Load s,load c1,...load cn,cbr s,cbr c1 CBR cn,cdr s,cdr c1 CDR cn,cqi] (2) n n n L(Φ,ϖ,ξ,α,γ)=Φ 2 + C ξ i α i [Φ 2 + ξ i ( x i 2 2ϖ x i + ϖ 2 )] γ i ξ i (5)

5 QIN et al.: MACHINE LEARNING AIDED CONTEXT-AARE SH MANAGEMENT FOR ULTRA DENSE NETORKS ITH QoS PROVISIONS Then, we can generalize the Lagrange dual function as n n L(Φ,ϖ,ξ,α,γ)= α i (x i x i ) α i α j (x i x j ).,j=1 By solving (7), we obtain the optial solution α = {αi,i= 1,...N}, and the vectors with coefficient αi 0 are called as support vectors (SVs). 3 The values of ϖ and Φ 2 depend on the support vectors illustrated as n n Φ 2 =(x k,x k ) 2 α i (x i,x k )+ α i α j (x i,x j ).,j=1 (8) Then, we build an optial sphere for the outaged cell detection based on the collected noral KPIs data saple, by solving the convex quadratic optiization proble. e define the cell outage detection rule whereby the new collected KPI data saple ν in sall cell network can be treated as noral when it satisfies the following: n ν ϖ 2 =(ν ν) 2 α i (ν x i ) + n,j=1 (7) α i α j (x i,x j ) Φ 2. (9) Support objects with α i = C will occur when C<1 is satisfied. These objects are outside of the sphere which are considered outliers (which eans the abnoral KPIs in this paper). The rest of the training data is within the description (which eans the noral KPIs in this paper). Then, the sall cell outage can be detected by using the cell outage detection rule of KPIs data doain description. B. Sall Cell Outage Detection and Localization In the detection and outage localization phase, each MDT report is tagged with location and tie inforation. e can obtain the noral area of the sall network based on the SVDD algorith and the new MDT KPIs are then tested by checking if they fall in the optial noral area. If not, the sall cell outage algorith is triggered, and we can locate the outaged sall cell based on the location inforation. However, there are also other situations in which the cell is sleeping or switched off by the operator for aintenance tasks or due to energy saving reasons. Hence, we introduce the concept of incoing handovers (Ξ in ), which are easured on a per sall cell basis by neighboring sall cells. Specifically, the sall cell onitors the nuber of Ξ in during easureent tie period T, to deterine the tie period between two executions of the outage detection algorith. If this nuber becoes zero for a certain sall cell, that cell should be considered as an outaged sall cell, and no KPIs can be easured in this situation as discussed in [29]. e define Ξ a as the nuber of Ξ in in the last period for T and define Ξ b 3 Generally, when the training set involves a large aount points, SVs is a sall proportion, ost of the training points are non-support vectors. That is to say, ost coponents of the Lagrange Dual solution are zero [28]. Algorith 1: The Proposed SCOD Algorith. Input: MDT easured KPIs: RSRP, RSRQ, CDR, CBR, Load, Ξ in and T Output: sall cell outage detection For each sall cell take the following steps: 1: Collect the KPIs fro SCBS, users and OAM periodically during easureent tie T ; 2: Calculate Ξ a and Ξ b ; 3: If Ξ a == 0, Ξ b > 0, and KPIs are available, then, follow these steps: 3.1: KPI data preparation to obtain a easureent of N diensions space, x it = {x it (1),x it (2), x it (N)} (the KPI data that are collected in i-th T period); 3.2: Calculate the values of ϖ and Φ for the noral KPIs objective based on (6) and (8) by using the collected KPI data; collect the new KPI saple fro SCBS, and set the label of the new KPIs as ν = x + 1; 3.3: If ν ϖ 2 Φ 2, then we set F = 1, which indicates that KPIs of the sall cell are noral, and we also set the outage counter L = 0; else, we set F = 0 which represents that the KPIs are abnoral, and the outage counter L ++; 3.4: If the outage counter satisfies L>M, the sall cell is selected as an outaged sall cell, in which M is introduced as an outage threshold value; Else The sall cell is selected as an outaged sall cell; 4: Generate the outage warning; End; as the nuber of Ξ in in the previous period for T. The easureent period tie T is a configurable paraeter, which has an effect on the execution tie of outage detection algorith. The saller the easureent period T is, the faster the outage detection is. During this period T, the KPIs are collected statistically and updated the KPIs in the OSS periodically. Then, we propose a novel SCOD algorith that considers KPIs available situation and non KPIs available situation based on SVDD, taking advantage of localization inforation as well as onitoring KPI data obtained by user equipents. The proposed SCOD algorith is shown in Algorith 1. IV. DISTRIBUTED SMALL CELL SELF-HEALING SCHEME Once the outaged cell proble is detected by the proposed SCOD algorith, a novel copensation echanis is required to copensate the users in the outage cells to iprove the overall network perforance. e propose a novel resource allocation schee for selfhealing, which ais to copensate for an abrupt cell outage in sall cell networks. In particular, we consider a scenario with a set of sall cells, and soe of these sall cells have already been detected as outages by the proposed the SCOD algorith. Then, we optiize the resource allocation to balance the load in the identified outage zone whilst guaranteeing coverage and users QoS requireents. To ipleent this approach, we propose a decentralized algorith for sall cell outage copensation.

6 12344 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 67, NO. 12, DECEMBER 2018 e assue that the whole bandwidth of the sall cell network is split into two parts: noral bandwidth and SH bandwidth. Particularly, noral bandwidth is used to transit data for noral users, and SH bandwidth is used to serve users in the outaged sall cells [30]. The whole SH bandwidth is allocated to outaged sall cells for copensation. Let U F N be the set of users previously served by the outaged sall cells, naely copensated users, who should be served by a set of noral cells and denote M = {1, 2,...M} S as the set of noral SCBS. The binary assignent indicator a u, {0, 1} takes the value 1 if user u U F is assigned to SCBS M, otherwise it is 0. A user of the outaged sall cell can be assigned to exactly one noral sall cell, and the assignent indicator in the assignent atrix A = {a u, } M UF satisfies M a u, = 1, M, u U F. (10) =1 The bandwidth that SCBS allocates to user u is denoted by w u,. Since all SCBSs share the sae SH bandwidth, intra-cell interference arises when two or ore neighboring SCBSs use the sae SH bandwidth. Hence, the rate of a copensated user u U F when served by SCBS Mis expressed as ) M r u = log 2 (1 + a u,pw u, g u 2 σ 2 + M (1 a u,)pw u, g u 2, (11) where p is the fixed value per unit of frequency and easured in [Joule/sec/Hz], pw u, is the transit power on bandwidth w u, allocated by SCBS. g u is the channel gain related to SCBS and copensated user u U F and σ 2 is the corresponding noise power. For a unit of bandwidth, the signal to interference plus noise ratio (SINR) [31] of copensated user u in (11) can be rewritten as SINR u, = p g u 2 p g u 2 ( UF j = 1 w j,s ) + σ 2. (12) Particularly, the traditional spectral efficiency for SCBS is defined as L = N n=1 a u, r w u,, which represents SCBS load and the value of spectral efficiency becoes unbounded when w u, 0. Additionally, we siplify the spectral efficiency function. For SCBS to a linear equation defined as N L = (a u, r u κw u, ),, u U F, (13) n=1 where we use the context of inforation about the iniu rate requireent ru req instead of the actual rate, and κ is a tuning paraeter (the larger the value of κ, the higher the cost of the bandwidth resource). A. Proble Stateent e consider a scenario where soe sall cells have already been detected as outaged through SCOD approach. Under this scenario, our objective is to find an optial user association and resource allocation policy to copensate the users in the outaged sall cells. e forulate the copensation proble to axiize the utility U (L ), which reflects the level of load balancing and users satisfaction, considering the diverse users. As shown in [32], a logarithic utility function U = log (L ) can naturally achieve load balancing and a certain level of fairness aong users. Therefore, we use a logarithic utility function as an objective utility. To optially copensate users while balancing the load under the sall cell outage scenario, we forulate the aggregate utility function axiization proble P1as rate requireent r req u P1 : ax L,A,w M U (L ) =1 s.t. C1 : C2 : C3 : M a u = 1, =1 U F u=1 w u,, 0 w u, a u, in { ru κ, }, C4 : w u, log 2 (1 + SINR u, ) a u, ru req,, u, C5 : a u, {0, 1},, u U F, (14) where L = {L 1, L 2,...L } is the load vector of SCBSs in sall cells network, A denotes the association atrix between users and SCBSs, and w = {w 1,w 1,...,w u F } denotes the allocated bandwidth vector for outaged users. Constraints C1 and C5 guarantee that the outaged user u U F can be assigned to only and exactly to one noral SCBS, C2 denotes the liits of the bandwidth allocation for all the SCBSs, and C3 specifies the users QoS requireents. C4 restricts the axiu bandwidth for outaged user u U F, which is a nonlinear constraint, leading to great difficulty solving this proble. Hence, we propose Theore 1 to facilitate solving our proble. Reark 1: The logarith function U = log (L ) is concave and widely eployed in other works [32, 35]. The property of this function can encourage load balancing efficiently. This is consistent with the resource allocation philosophy in real systes, where allocating ore resources for a well-served user is considered low priority, whereas providing ore resources to users with low rates is considered desirable [32]. Theore 1: Constraint C4 is equivalent to (15) in solving proble P1, p g u, 2 w u, r req u ( UF p g u,s 2 j=1 w ) j,s + σ 2. (15) Proof: Please refer to Appendix A. According to Theore 1, we replace constraint C4 with (15), and thus proble P1 can be reforulated into a proble with linear constraint set, which is easy to handle.

7 QIN et al.: MACHINE LEARNING AIDED CONTEXT-AARE SH MANAGEMENT FOR ULTRA DENSE NETORKS ITH QoS PROVISIONS Furtherore, constraint C3 actually states that w u, = 0 when a u, = 0, otherwise { ru } 0 w u, in κ,. (16) For better understanding, we replace C3 with (16), which is equivalent to C3 in the context of P1. Proble P1 is a ixed integer prograing proble. By linearly relaxing the binary user assignent indictors into a continuous value within [0, 1], proble P1 becoes a convex proble with a concave objective function and can be solved by known techniques, although with high coputational coplexity [33], based on Theore 1 and (16). To directly solve the optiization proble P1 with the aid of convex prograing [33], [34], global network inforation is required, which necessitates a centralized controller for user association and coordination. Additional issues with centralized echaniss include excessive coputational coplexity and low reliability, as any disabling of the centralized control operation will disrupt load balancing [35] [37]. In sall cell networks, it is usually difficult to coordinate sall cells, soe of which (e.g. a fetocell) are deployed by either operators or users [37]. Therefore, a low-coplexity distributed algorith without coordination is desirable. In this section, we propose a distributed algorith via Lagrangian dual decoposition approach. e proceed by relaxing with each equality a real Lagrange ultiplier, λ R, and with each inequality a real non-negative Lagrange ultiplier, μ u, R +. Adding the to the objective function, we reforulate proble P1 toprob- le D in (17), shown at the botto of this page, in which UF j = 1 w j,s ). ψ = p g u,s 2 ( By further analyzing proble D, we find that there is no ters coupling aong L, A, and w in the objective for constraints. Hence, in order to further reduce the coputational coplexity, the dual decoposition ethod is eployed to solve proble D. It is based on the decoposition of the Lagrangian dual proble [38], and it can decopose the original large proble into distributively solvable subprobles. Then, they are easy to handle and can be solved separately on the users side and SCBSs side respectively [31]. ith the dual analysis, we observe that the solution for the assignent variables will not be influenced by solving D fro (17). Hence, the sub-probles are presented and solved independently as follows: Utility Distribution: the optial utility per sall cell is given by solving (18) over l, ax [U (L ) λ L ]. (18) L User Association Distribution: the optial user association distribution is derived by axiizing the overall network utility and is given by solving the proble below [ ax λ r u a u, + μ u, (1 a u, )r u A [ ]] ψu,s ax + σu 2 s.t. M a u, = 1,, u U F. (19) =1 Bandwidth Allocation: the optial bandwidth allocation is derived by solving ) ax (w u, (μ u, p g u, 2 κ u λ w s.t. u, μ u, r u ψu,s ax + σu 2 U F u=1 w u,, 0 w u, in { ru κ, }, (20) where the constraint actually states that when the user is not connected to sall cell, then necessarily, bandwidth w u, = 0, otherwise 0 w u, in{ r u κ,}.esiplify the constraints by assuing that the total bandwidth available is sufficiently large 1, so that the constraint can be always satisfied with strict inequality. In particular, given Lagrange ultipliers (λ,μ), called fro now on as price, we have the following solutions by solving probles (18) (20). Theore 2: a) Given Lagrange ultipliers for the relaxed probles (18) (20), the optial utility value of SCBS is the solution to λ = du (L ) dl. (21) D :D(λ,μ) = ax L [U (L ) λ L ] + ax A + ax w λ r u a u, + μ u, (1 a u, )r u ψ ax + σu 2 μ u, ph u, w u, κ u λ w u, μ u, r u ψ + σu 2 (17) u,

8 12346 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 67, NO. 12, DECEMBER 2018 b) Furtherore, each outaged user u is assigned to SCBS satisfying u = arg ax λ μ u, ψ ax + σu 2. (22) c) The outaged users in sall cells can find an optial copensation sall cell based on (22). The optial bandwidth allocation for outaged user u is constrained by the following conditions w u, (0, in( γ req u κ ),). UF u=1 w u,. u u, p g u, 2 λ κ + j u j,sγ j p g j, 2. (23) Proof: Please refer to Appendix B. Note that there ay be a certain inconsistency between the assignent of a user to a single SCBS whilst satisfying (23) and the allocation of positive bandwidth to possibly ore than one station. The reason for this is the change of the constraint fro C3 to (16). In the following, we will conclude that the optial solution derived fro C3 is the sae as that fro (16). Theore 3: Constraints C3 and (16) are equivalent in ters of solving proble P1. Moreover, if a u, = 1, then wu, ( 0, in{ r uκ,} ),elsea u, = 0, then wu, = 0. Proof: Theore 3 is derived directly according to the copleentary slackness condition related to (15) and Theore 2. The detail derivation is oitted due to lack of space, and interested readers can refer to [31] for details. Fro Theore 3, once the user assignent variables are acquired, we can deterine the bandwidth allocation for outaged users. Therefore, we derive Corollary 1 to acquire the user assignent schee. Particularly, since we ai at providing an algorith possible to ipleent in sall cell networks, in the following sections, we will explain how the users in the outaged SCBS are copensated by noral SCBS. Corollary 1: Choosing an appropriate SCBS for the outaged users is based on the following rules with the load price and interference cost. e choose the sall cell with the axiu utility equal to the load price and the iniu interference cost towards the neighboring SCBSs that can satisfy = arg ax λ ζ }{{} u p g j, 2 j,sr j j/ N }{{, } Load price Interference cost (24) where ζ is a tuning factor giving higher or lower weight on the interference cost. Based on the above, we present the algorith for the optial load balancing aong SCBSs of sall cell networks. Based on Theores 1 3 and Corollary 1, we develop a greedy ascend algorith to acquire the user assignent. Considering the inforation of the outaged users set U F and easured MDT data of SCBSs, the outline of the SCOC schee procedures is detailed in Algorith 2. Algorith 2: The Proposed SCOC Algorith. Input: The outaged SCBS user association and resource allocation vector π 0 = {L, a, w}, where all users can gain inforation over the channel through RSRP easureents. Afterwards they counicate their channel quality vectors G n =[g u,1,...g u,m ]. Set K as the nuber of algorith iteration. Output: the adequate reconfiguration of SCBS user association and resource allocation For :K 1: Each SCBS calculates the current load L i using (13), the current load price λ i using (21), and the interference cost I i using (24); 2: SCBSs exchange the current values of load and interference cost with their neigbors which are listed in its neighbors list. 3: Based on inforation over the other load and interference cost of SCBSs, each sall cell can decide whether it s a target sall cell for its neighborhood to load balancing. 4: The outaged sall cell users U F are defined as candidate users, then calculate the possible change in load of the other cells and get the utility; 5: The user set which axiize the utility is chosen; 6: Update π i+1 π i ; 7: hen λ i = λ i 1 and I i = I i 1, stop; End; B. Coplexity Analysis The proposed SH anageent algorith for SON-based sall cell networks consists of two ain phases: the SCOD phase and the SCOC phase. In SCOD phase, the coputational coplexity of cell outage detection process is deterined by the SVDD algorith, and the tie and space coplexity of obtaining the decision boundary using SVDD are O(k 3 ) and O(k 2 ), respectively (k is the saple nuber of KPIs). In particular, the ain idea of SVDD algorith is transitted to solve a quadratic prograing proble, which needs to copute a sphere around the data in the input space, and it is stated copletely in ters of inner products between vectors as shown in (7). Based on the characteristics of coputer dealing with linear and quadratic optiization probles, we can achieve the coplexity of the detection algorith in the SCOD phase. Furtherore, in SCOC phase, the outaged users are properly served based on resource allocation. In each iteration of this phase, the coplexity of the proposed distributed SCOC algorith is O( U F M ), where U F denotes the nuber of copensation users and M is the nuber of SCBSs. Thus, the proposed SH schee can reduce the coputational coplexity copared with the optial solution which has exponential coputational coplexity. V. SIMULATION RESULTS A. Siulation Setup In the siulation, we consider a sall cell network with 10 sall cells and 50 users uniforly distributed within a 50

9 QIN et al.: MACHINE LEARNING AIDED CONTEXT-AARE SH MANAGEMENT FOR ULTRA DENSE NETORKS ITH QoS PROVISIONS TABLE II SIMULATION PARAMETERS Fig. 4. The outage detection of SCOD based on abnoral KPIs. Fig. 3. The outage detection of SCOD based on noral KPIs. 50 area. The propagation odel is deterined based on the ITU and COST231 odel described in [23]. The standard deviation of shadowing factor δ db = 8 db, where δ db = 10 δ/ ln(10). The siulation paraeters are given in Table II. B. Siulation Perforance In this section, siulation results are provided to validate our theoretical perforance analysis. Different faults which lead to the eergence of cell outages can be detected separately as we described in Section II. In the siulation, different antenna gain reductions are configured to represent different degrees of failures for cell outages. As detailed in Section III, the preprocessed MDT easureents data set is used as the input data for SH schee. e take 100 available KPIs (MDT easureent reports, which are all noral data), which contain 80 training saples and 20 test saples to verify the perforance of the proposed algorith. Fig. 3(a) shows that the proposed SCOD algorith obtains a spherically shaped boundary for sall cell outage detection (This boundary can cover the class of objects (e.g., noral KPIs) represented by the training set, and ideally should reject all other possible objects (e.g., abnoral KPIs) in the object space.). The easured KPIs data that are covered in the boundary indicates that the sall cell operates norally, and the KPIs data outside the boundary denotes that the sall cell is outaged, requiring copensation for this sall cell. e further use the collected RSRP values to detect the sall cell outage in Fig. 3(b). The horizontal line in Fig. 3(b) represents the radius of the sphere (sphere thr represents the optial threshold value of the optial sphere), and we see that an outage (the distance of data which is larger than the radius) has occurred in a sall cell. Furtherore, we also take 100 KPIs, which contain noral KPIs and abnoral KPIs to verify the perforance of the SCOD algorith. Fig. 4(a) and 4(b) show that the algorith can also detect the sall cell outage efficiently. Particularly, we detect the situation that the sall cell is also outaged, when there is lack of availability of KPIs based on SCOD algorith, which is also treated as a sall cell outage. Furtherore, we take SVM-aided cell outage schee as the benchark to verify the perforance of SCOD schee fro the two aspects of runtie and accuracy, and the siulation results in Table III show that the proposed SCOD schee has uch lower coplexity with high detection accuracy. Thus, the proposed SCOD algorith can be very practical and robust in real syste ipleentations. Fig. 5 shows the sall cell network perforance of the proposed SCOC algorith when changing tuning paraeter κ (the higher the value of κ, the higher the cost of the bandwidth

10 12348 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 67, NO. 12, DECEMBER 2018 TABLE III RUNTIME AND ACCURACY SIMULATION RESULTS Fig. 7. Network utility vs. nuber of SCBS. Fig. 5. Fig. 6. Network utility vs. nuber of iterations. Network utility vs. nuber of users. resource.). This is reasonable because ore resources in each sall cell will be occupied by copensation users in order to iprove the rate to satisfy their QoS requireents according to (12) and (13). Fig. 5 also shows that the proposed SCOC algorith converges quickly in just a few iterations, because the algorith operates in a distributed anner, and sall cells can ake decisions locally without counication with operation centers copared with traditional centralized algoriths. Fig. 6 shows the network utility of sall cell network perforance in ters of the nuber of users, copared to the case where non-scoc schee is ipleented. The coalition gae based resource allocation self-healing algorith (GA-based SH) is used as the benchark to verify the perforance of SCOC schee enabling self-healing and copensating abrupt cell outage in sall cell networks. e can see that the perforance of the SCOC schee is uch better than that of non-scoc option in outaged sall cell networks. Moreover, we can see that higher network utility will be achieved by the proposed SCOC schee copared with the GA-based schee. In particular, fro Fig. 6, it can be concluded that the SCOC schee can also achieve uch better network utility under different users QoS requireents. Most iportantly, the proposed SH schee can operate in an autoatic anner, which is expected in the future network anageent with a SON functionality. Fig. 7 shows the network utility of sall cell networks perforance in ters of the nuber of SCBSs and investigates the ipact of the density of sall cells. As expected, it shows that the proposed SCOC schee provides better results than GAbased schee in outaged dense sall cell networks. In addition, the proposed schee also achieves better perforance than the non-scoc schee in sall cell networks. The reason is that we take the interference cost into consideration, which is an iportant factor in dense sall cell networks. Fig. 8 further shows the load distribution in dense sall cell networks. e see that the load of the sall cells changes with the nuber of iterations and converges to a balanced load distribution, which denotes that the sall cell network can obtain efficient iproveent on load balancing with a QoS provision. Furtherore, as users in outaged sall cells becoe copensated users, which are served by noral sall cells, the algorith converges quickly and will be very practical and robust in real wireless network s syste ipleentations.

11 QIN et al.: MACHINE LEARNING AIDED CONTEXT-AARE SH MANAGEMENT FOR ULTRA DENSE NETORKS ITH QoS PROVISIONS Fig. 8. Load distribution of SCBS vs. nuber of iterations. Then, we transfor (26) to the for that p g u, 2 w u, a u, ru req ( UF p g u, 2 j=1 w ) j,s + σ 2 (27) and (28) is the expansion of (27). (1 a u, )ru req [ ( UF p g u, 2 j=1 w )] ax j,s + σ 2 + p g u, 2 w u, ru req ( UF p g u, 2 j=1 w ) j,s + σ 2 (28) Furtherore, we transfor the constraint set into a set of linear inequalities, and get the inequalities as shown in (15). This copletes the proof of Constraint C4. VI. CONCLUSIONS AND FUTURE ORKS In this paper, a self-healing schee to deal with cell outages in SON-based sall cell networks has been investigated. e have proposed a SCOD algorith using achine learning approach based on partial KPI statistics that are a large scale collection of MDT reports. e also have proposed a novel SCOC algorith to fairly allocate resources to outaged users for copensation, considering the dynaic and dense deployent of sall cells environent. Siulation results deonstrate that the proposed SH schee can detect the sall cell outage efficiently and can achieve an optiized QoS perforance when copensating for the detected sall cell outage. In addition, the coplexity of operation and anageent can becoe the biggest challenge in 5G, a coprehensive fraework for epowering SONs with big data to address the anageent requireents of 5G is urgently needed. e will explore fro the doain of achine learning to create self-organized end-to-end intelligence of 5G networks. APPENDIX A PROOF OF THEOREM 1 In this paper, we set a iniu rate requireent for each user u, denoted by ru req. The constraint w u, log 2 (1 + SINR u, ) a u, ru req can be satisfied when guaranteeing different users QoS requireents based on Shannon capacity forulation. e ake an approxiation log 2 (1 + SINR u, ) SINR u, for siplicity in sall cell networks, which was proved valid in [31]. So the constraint can be rewritten as w u, SINR u, a u, r req u. (25) e substitue the SINR u, expression fro (12) into (25) and have p g u, 2 w u, ( UF ) a u, r p g u, 2 u req. (26) j = 1 w j,s + σ 2 APPENDIX B PROOF OF THEOREM 2 Given Lagrange ultipliers for the relaxed proble P1, we can find the optial values for utility distribution, user assignent and bandwidth allocation by solving each one of the subprobles respectively. For utility distribution in the proble, the optial solution is given by solving the utility distribution sub-proble, which satisfies the expression ax l [U (L ) λ L ] (29) and the solution is L = λ 1. Clearly, the variable a is decoupled in the constraints set of (10), and we can separate the optiization in (17) into sub-probles eploying the priary decoposition approach. At the lower level, the sub-probles, one for each user u, can be written as [ ax λ r u a u, + μ u, (1 a u, )r u A [ ]] ψu,s ax + σ 2 s.t. M a u, = 1. (30) =1 Furtherore, we observe that each outaged user u is assigned to SCBS in user association proble satisfying u = arg ax λ μ u, ψ ax + σu 2. (31) The outaged users in sall cells can find an optial copensation sall cell based on (31). Considering the bandwidth allocation proble, we siplify the constraints by assuing that the total available

12 12350 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 67, NO. 12, DECEMBER 2018 bandwidth is sufficiently large. e can obtain the optial resource allocation by solving over w as ) ax (w u, (μ u, p g u 2 κ u λ w s.t. u, μ u, r u ψu,s ax + σu 2 U F u=1 w u,, 0 w u, in { ru κ, }, (32) where a u, and w u, are related, the constraint actually states that when the user is not connected to sall cell, bandwidth w u, = 0, otherwise 0 w u. in{ r u κ,}. And we siplify the constraints by assuing that total bandwidth available is sufficiently large such as >>1, so that the constraint can always be satisfied with strict inequality. Then, the optial resource allocation is equal to ( { ru }) w u, 0, in κ,, (33) when ensuring that each SCBS with channel quality above the threshold u u, p g u, 2 λ κ + j u j,sγ j p g j, 2. The optial value of proble P1 is deterined by solving the dual proble of (17). e can find the optial values on load, user association and resource allocation by solving each one of the above subprobles respectively. Given a Lagrange ultipliers (λ,μ), we use a line search ethod to find the optial solution for proble P1, which can increase the value of the objective by searching fro a feasible direction. REFERENCES [1] N. Zhang, N. Cheng, A. T. Gaage, K. Zhang, J.. Mark, and X. Shen, Cloud assisted HetNets toward 5G wireless networks, IEEE Coun. Mag., vol. 53, no. 6, pp , Jun [2] N. Cheng et al., Big data driven vehicular netw., IEEE Network, vol. 32, no. 6, pp , Nov./Dec [3] A. Iran, A. Zoha, and A. A. Dayya, Challenges in 5G: How to epower SON with big data for enabling 5G, IEEE Trans. Signal Process., vol.62, no. 14, pp , Jul [4]. Quan, Y. Liu, H. Zhang, and S. Yu, Enhancing crowd collaborations for software defined vehicular networks, IEEE Coun. Mag., vol. 55, no. 8, pp , Aug [5] S. Fortes, R. Barco, and A. A. Garcia, Location-based distributed sleeping cell detection and root cause analysis for 5G ultra-dense networks, Eurasip J. ireless Coun. Netw., vol. 1, no. 149, pp. 1 18, [6] 3rd Generation Partnership Project; Techinical Specification Group Services and Syste Aspects; Telecounications Manageent; Self- Organizing Networks (SON); Self-Healing Concepts and Reqiureents (Release 11), 3GPP TS 2.541, v Sep [7] A. G. Andrades, P. Munoz, I. Serrano, and R. Barco, Autoatic root cause analysis for LTE networks based on unsupervised techniques, IEEE Trans. Veh. Technol., vol. 65, no. 4, pp , Apr [8] M. Alias, N. Saxena, and A. Roy, Efficient cell outage detection in 5G HetNets using hidden Markov odel, IEEE Coun. Lett., vol.20,no.3, pp , Mar [9] S. Fortes, A. Garcia, J. Luque, A. Garrido, and R, Barco, Context-Aware Self-Healing: User equipent as the ain source of inforation for sallcell indoor networks, IEEE Veh. Technol. Mag.,vol.11,no.1,pp.76 85, Mar [10] A. G. Andrades, P. Munoz, E. J. Khatib, I. Bandera, I. Serrano, and R. Barco, Methodology for the design and evaluation of self-healing LTE networks, IEEE Trans. Veh. Technol. vol. 65, no. 8, pp , Aug [11] A. Ebrahi and E. Alsusa, Interference and resource anageent through sleep ode selection in heterogeneous networks, IEEE Trans. Coun., vol. 65, no. 1, pp , Jan [12] A. Zoha, A. Saeed, A. Iran, M. A. Iran, and A. A. Dayya, A SON solution for sleeping cell detection using low-diensional ebedding of MDT easureents, in Proc. IEEE PMIRC, Sep. 2014, pp [13] A. G. Andrades, R. Barco, P. Munoz, and I. Serrano, Data analytics for diagnosing the RF condition in self-organizing networks, IEEE Trans. Mobile Coput., vol. 16, no. 6, pp , Jun [14] E. J. Khatib, R. Barco, A. Andrades, and I. Serrano, Diagnosis based on genetic fuzzy algoriths for LTE self-healing, IEEE Trans. Veh. Technol. vol. 65, no. 3, pp , Mar [15] Y. Ma, M. Peng,. Xue, and X. D. Ji, A dynaic affinity propagation clustering algorith for cell outage detection in self-healing networks, in Proc. IEEE ireless Coun. Netw. Conf., Apr. 2013, pp [16] I. D. L. Bandera, P. M. Luengo, I. Serrano, and R. Barco, Adaptive cell outage copensation in self-organizing networks, IEEE Trans. Veh. Technol., vol. 67, no. 6, pp , Jun [17] R. Sivaraj, I. Broustis, N. K. Shankaranarayanan, V. Aggarwal, and P. Mohapatra, Mitigating acro-cell outage in LTE-advanced deployents, in Proc. IEEE INFOCOM, Apr. 2015, pp [18] M. Seli, A. E. Kaal, K. Elsayed, H. M. A. Atty, and M. Alnue, Fronthaul cell outage copensation for 5G networks, IEEE Coun. Mag., vol. 54, no. 8, pp , Aug [19] P. Makris, D. N. Skoutas, and C. Skianis, A survey on context-aware obile and wireless networking: On networking and coputing environents Integration, IEEE Coun. Surveys Tuts., vol.15,no.1,pp , Jan. Mar [20] L. Qian, Y. u, J. Zheng, H. Zhou, and X. Shen, Green-Oriented traffic offloading through dual-connectivity in future heterogeneous sall-cell networks, IEEE Coun. Mag., vol. 56, no. 5, pp , May [21] N. Zhang, S. Zhang, J. Zheng, X. Fang, J.. Mark, and X. Shen, QoE driven decentralized spectru sharing in 5G Networks: potential gae approach, IEEE Trans. Veh. Technol., vol. 66, no. 9, pp , Sep [22]. ang and Q. Zhang, Local cooperation architecture for self-healing fetocell networks, IEEE ireless Coun., vol. 21, no. 2, pp , Apr [23]. ang, Q. Liao, and Q. Zhang, COD: A cooperative cell outage detection architecture for self-organizing fetocell networks, IEEE Trans. ireless Coun., vol. 13, no. 11, pp , Nov [24] S. Fan and H. Tian, Cooperative resource allocation for self-healing in sall cell networks, IEEE Coun. Lett.,vol.19,no.7,pp , Jul [25] O. Onireti et al., A cell outage anageent fraework for dense heterogeneous networks, IEEE Trans. Veh. Technol. vol. 65, no. 4, pp , Apr [26] V. Erceg et al., An epirically based path loss odel for wireless channels in suburban environents, in Proc. IEEE GLOBECOM, Nov. 1998, pp [27] 3GPP, Universal Mobile Telecounications Syste (UMTS); LTE; Universal Terrestrial Radio Access (UTRA) and Evolved Universal Terrestrial Radio Access (E-UTRA); Radio easureent Collection for Miniization of Drive Tests (MDT); Overall Description; Stage 2, 3GPP TS , , v Release 10. [28] K. Y. Lee, D.. Ki, and K. H. Lee, Density-induced support vector data description, IEEE Trans. Neural Netw., vol. 18, no. 1, pp , Jan [29] L. B. I. De, R. Barco, P. Munoz, and I. Serrano, Cell outage detection based on handover statistics, IEEE Coun. Lett., vol. 19, no. 7, pp , Jul [30] K. Lee, H. Lee, and D. H. Cho, On the low-coplexity resource allocation for self-healing with reduced essage passing in indoor wireless counication systes, IEEE Trans. ireless Coun., vol. 15, no. 3, pp , Mar [31] A. Giovanidis, Q. Liao, and S. Stanczaky, A distributed interferenceaware load balancing algorith for LTE ulti-cell networks, in Proc. IEEE Sart Antennas, Mar. 2012, pp [32] Q. Ye, B. Rong, Y. Chen, M. Shalash, C. Caraanis, and J. G. Andrews, User association for load balancing in heterogeneous cellular networks, IEEE Trans. ireless Coun., vol. 12, no. 6, pp , Jun [33] S. Boyd and L. Vandenberghe, Convex Optiization. Cabridge, U.K.: Cabridge Univ. Press, 2004.

13 QIN et al.: MACHINE LEARNING AIDED CONTEXT-AARE SH MANAGEMENT FOR ULTRA DENSE NETORKS ITH QoS PROVISIONS [34] A. Ben-Tal, Lectures on Modern Convex Optiization: Analysis, Algoriths, and Engineering Applications. Philadelphia, PA, USA: SIAM, [35] N. Cheng, N. Zhang, N. Lu, X. Shen, J. Mark, and F. Liu, Opportunistic spectru access for CR-VANETs: A gae-theoretic approach, IEEE Trans. Veh. Technol., vol. 63, no. 1, pp , Jan [36] S. Zhang, N. Zhang, S. Zhou, J. Gong, Z. Niu, and X. Shen, Energy-aware traffic offloading for green heterogeneous networks, IEEE J. Sel. Areas Coun., vol. 34, no. 5, pp , May [37] L. Qian, Y. u, H. Zhou, and X. Shen, Joint uplink base station association and power control for sall-cell networks with non-orthogonal ultiple access, IEEE Trans. ireless Coun.,vol.16,no.9,pp , Mar [38] Y. Shi, J. Zhang, B. Donoghue, and K. B. Letaief, Large-Scale convex optiization for dense wireless cooperative networks, IEEE Trans. Signal Process., vol. 63, no. 18, pp , Sep and vehicular networks. Haibo Zhou (M 14 SM 18) received the Ph.D. degree in inforation and counication engineering fro Shanghai Jiao Tong University, Shanghai, China, in Fro 2014 to 2017, he was a Postdoctoral Fellow with the Broadband Counications Research Group, Electrical and Coputer Engineering Departent, University of aterloo. He is currently an Associate Professor with the School of Electronic Science and Engineering, Nanjing University. His research interests include resource anageent and protocol design in cognitive radio networks in wireless networks. Meng Qin received the B.S. degree in counication engineering fro the Taiyuan University of Technology, Taiyuan, China, in 2012 and the M.S. degree in inforation and counication systes fro Xidian University, Xi an, China, in He is currently working toward the Ph.D. degree in counication and inforation systes at Xidian University. His research interests include wireless network operation and anageent, achine learning, self-organized network, statistical quality of service (QoS) provisioning, and applications of stochastic optiization Raesh R. Rao (M 85 SM 90 F 10) received the bachelor s degree fro the University of Madras (the National Institute of Technology), Tiruchirapalli, India, in 1980, and the M.S. and Ph.D. degrees in electrical engineering fro the University of Maryland, College Park, MD, USA, in 1982 and 1984, respectively. He has been a Faculty Meber with the UC San Diego (UCSD) since 1984, and the Director of the Qualco Institute, UCSD division, California Institute for Telecounications and Inforation Technology (Calit2), since He holds the Qualco Endowed Chair in Telecounications and Inforation Technologies in the Jacobs School of Engineering, UCSD, and is a eber of the school s Electrical and Coputer Engineering Departent. Previously, he was the Director of the Center for ireless Counications, UCSD. He is a Senior Fellow of the California Council on Science and Technology. Qinghai Yang received the B.S. degree in counication engineering fro the Shandong University of Technology, Shanghai, China, in 1998, the M.S. degree in inforation and counication systes fro Xidian University, Xi an, China, in 2001, and the Ph.D. degree in counication engineering fro Inha University, Incheon, South Korea in 2007 with the University-President Award. Fro 2007 to 2008, he was a Research Fellow with UB-ITRC, South Korea. Since 2008, he is with Xidian University. His current research interest lies in the fields of autonoic counication, content delivery networks, and LTE-A techniques. Nan Cheng (S 12 M 16) received the B.E. and M.S. degrees fro Tongji University, Shanghai, China, in 2009 and 2012, respectively, and the Ph.D. degree fro the University of aterloo, aterloo, ON, Canada, in He is currently a Postdoctoral Fellow with the Departent of Electrical and Coputer Engineering, University of Toronto and the Departent of Electrical and Coputer Engineering, University of aterloo, under the supervision of Prof. Ben Liang and Prof. Sheran (Xuein) Shen. His current research focuses on big data in vehicular networks and self-driving syste. His research interests also include perforance analysis, MAC, opportunistic counication, and application of AI for vehicular networks. Xuein (Sheran) Shen (M 97 SM 02 F 09) received the B.Sc. degree fro Dalian Maritie University, Dalian, China, in 1982 and the M.Sc. and Ph.D. degrees fro Rutgers University, New Brunswick, NJ, USA, in 1987 and 1990, respectively, all in electrical engineering. He is currently a University Professor and the Associate Chair for Graduate Studies, Departent of Electrical and Coputer Engineering, University of aterloo, Canada. His research focuses on resource anageent, wireless network security, social networks, sart grid, and vehicular ad hoc and sensor networks. He was the Technical Progra Coittee Chair/Co-Chair for IEEE Globeco 16, Infoco 14, IEEE VTC 10 Fall, and Globeco 07, the Syposia Chair for IEEE ICC 10, the Tutorial Chair for IEEE VTC 11 Spring and IEEE ICC 08, the General Co-Chair for ACM Mobihoc 15, Chinaco 07, and the Chair for IEEE Counications Society Technical Coittee on ireless Counications. He is also or was the Editor-in-Chief for the IEEE INTERNET OF THINGS JOURNAL, IEEE NETORK, Peer-to-Peer Networking and Application, andiet Counications; a Founding Area Editor for the IEEE TRANSACTIONS ON IRELESS COMMUNICATIONS; an Associate Editor for the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, Coputer Networks, and ACM/ireless Networks, etc.; and the Guest Editor for the IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, IEEE IRELESS COMMUNICATIONS, and IEEE COMMUNICATIONS MAGAZINE, etc. He received the Excellent Graduate Supervision Award in 2006, and the Preiers Research Excellence Award (PREA) in 2003 fro the Province of Ontario, Canada. He is a registered Professional Engineer of Ontario, Canada, an Engineering Institute of Canada Fellow, a Canadian Acadey of Engineering Fellow, a Royal Society of Canada Fellow, and a Distinguished Lecturer of IEEE Vehicular Technology Society and Counications Society.