An Ensemble Learning Approach for Fault Diagnosis in Self-Organizing Heterogeneous Networks

Yueping Wang, Kun Zhu, Mengyun Sun, Yueyu DengIEEE Access 2019 · OpenReview

Abstract

Self-organizing networks (SON) aim to offer high quality services while reducing both capital and operational expenditures by enabling three main functions: self-configuration, self-optimization, and self-healing. Though there exits only few studies on self-healing, it plays an important role in intelligent network maintaining. In this paper, we firstly propose an ensemble learning based fault diagnosis system for self-organizing heterogeneous networks. Specifically, the base learner is strengthened in each iteration and the final diagnosis result is obtained from the combination of all base classifications for performance improvement. Moreover, traditional classification algorithms are designed with the premise of balanced data set. However, when they are applied to fault diagnosis in cellular network which has imbalanced data, the classification accuracy for minority classes is not satisfying. To address this issue, synthetic minority over-sampling technique (SMOTE) is applied to handle the data-imbalance. Furthermore, considering the fact that misclassification is unavoidable, and most existing schemes aim to achieve a low detection error rate, while ignoring the fact that different type of misclassification errors can cause different economic losses to the operators. We consider the cost sensitivity and use rescaling method to help the classifier differentiate the importance of different samples, so that a minimal total loss can be achieved. Also, for handling the sparse data and dense deployment issues in small cells of a heterogeneous network, we provide a distributed diagnose system for lowering the communications cost. Extensive simulations are performed and the results show the effectiveness of the proposed system.

Introduction

As cellular networks evolve toward 5G and 6G, they are becoming increasingly heterogeneous, characterized by a dense integration of macrocells and numerous small cells. Manually diagnosing faults in such complex environments is no longer feasible. While Self-Organizing Networks (SON) aim to automate maintenance through "self-healing," several critical challenges persist:

  1. Data Imbalance: In operational networks, "normal" status data is abundant, whereas fault data is rare. Standard machine learning models often become biased toward the majority class, leading to poor detection of critical but infrequent faults.
  2. Unequal Error Costs: Misclassification errors carry different economic weights. For instance, failing to detect a service-impacting fault is significantly more costly than a false alarm.
  3. Communication Overhead: Centralized diagnosis for thousands of small cells creates significant data bottlenecks.

Methodology

The proposed system addresses these challenges through a multi-layered ensemble learning framework designed for heterogeneous environments.

System Architecture

The framework integrates fault detection, diagnosis, and recovery. It utilizes historical data and real-time Key Performance Indicators (KPIs) to feed a diagnosis algorithm that operates locally or centrally depending on the cell type.

System architecture for automated fault diagnosis in SON.

Ensemble Learning and Data Balancing

The core of the system is an AdaBoost-based ensemble learner using Decision Stumps. To handle the inherent data imbalance, the researchers employ the Synthetic Minority Over-sampling Technique (SMOTE). Unlike simple duplication, SMOTE generates new, plausible data points for minority fault classes, enhancing the model's ability to distinguish rare faults from normal operation.

Illustration of the SMOTE technique for generating synthetic minority samples.

Cost-Sensitive Rescaling

To minimize economic impact, a cost-sensitive rescaling method is introduced. By assigning higher weights to expensive errors (e.g., missed faults) in the training phase, the classifier is incentivized to prioritize minimizing total operational loss over raw mathematical accuracy. This ensures that the most damaging errors are avoided.

Distributed Diagnosis

For dense small cell deployments, a distributed architecture is proposed. The Macrocell Base Station (MBS) serves as a local aggregator, performing diagnosis for its cluster of small cells. This reduces communication costs and mitigates the "sparse data" problem by pooling information locally before transmission.

Experimental Results

The system was evaluated using a real-world dataset from a Spanish mobile operator, featuring KPIs such as handover success rate (HOSR), RSRP, RSRQ, and throughput. The dataset included seven classes: five specific fault types (e.g., Antenna Downtilt, Coverage Hole), a "Normal" state, and a general failure state.

Key findings include:

  • Performance: The Ensemble Learning (EML) approach consistently outperformed Support Vector Machines (SVM), K-Nearest Neighbor (KNN), and Back-Propagation (BP) networks across various training set sizes.
  • Accuracy: Using a Gaussian kernel SVM within the distributed framework, the system achieved a peak classification accuracy of 93.6%.
  • Economic Impact: The cost-sensitive rescaling method significantly reduced total economic losses compared to cost-blind methods as the penalty for missed faults increased.

Conclusion

This research demonstrates that a robust fault diagnosis system must account for the practical realities of cellular networks: imbalanced data and varying economic costs of errors. By combining ensemble learning with SMOTE and cost-sensitive analysis, the proposed framework offers a scalable solution for self-healing in heterogeneous networks. Future work will focus on dynamic cost modeling and automated feature extraction to further enhance system resilience in dynamic 5G environments.

Summary updated

Abstract

Self-organizing networks (SON) aim to offer high quality services while reducing both capital and operational expenditures by enabling three main functions: self-configuration, self-optimization, and self-healing. Though there exits only few studies on self-healing, it plays an important role in intelligent network maintaining. In this paper, we firstly propose an ensemble learning based fault diagnosis system for self-organizing heterogeneous networks. Specifically, the base learner is strengthened in each iteration and the final diagnosis result is obtained from the combination of all base classifications for performance improvement. Moreover, traditional classification algorithms are designed with the premise of balanced data set. However, when they are applied to fault diagnosis in cellular network which has imbalanced data, the classification accuracy for minority classes is not satisfying. To address this issue, synthetic minority over-sampling technique (SMOTE) is applied to handle the data-imbalance. Furthermore, considering the fact that misclassification is unavoidable, and most existing schemes aim to achieve a low detection error rate, while ignoring the fact that different type of misclassification errors can cause different economic losses to the operators. We consider the cost sensitivity and use rescaling method to help the classifier differentiate the importance of different samples, so that a minimal total loss can be achieved. Also, for handling the sparse data and dense deployment issues in small cells of a heterogeneous network, we provide a distributed diagnose system for lowering the communications cost. Extensive simulations are performed and the results show the effectiveness of the proposed system.
BibTeX
@inproceedings{wang2019ensemble,
  author = {Yueping Wang and Kun Zhu and Mengyun Sun and Yueyu Deng},
  title = {{An Ensemble Learning Approach for Fault Diagnosis in Self-Organizing Heterogeneous Networks}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=ilzjU2365z},
}
Figure