The Method of Communication System Fault Diagnosis Based on Deep Belief Net

Juan Li, Bin ChenInternational Journal of Circuits Systems and Signal Processing 2021 · DOI

Abstract

To meet the need of fault diagnosis for military communication systems, an effective method based on Deep Belief Networks (DBN) is proposed. The model extracts fault features from input signals and outputs classification results via a Softmax layer. To address computational intensity, a parallel solution using the MapReduce framework is implemented. Experimental verification using a laboratory-built communication system demonstrates high accuracy in fault diagnosis compared to traditional methods, providing a robust technical foundation for equipment health management.

Introduction

Modern military communication systems are increasingly complex, rendering traditional manual or statistical fault diagnosis methods inadequate. Fault signals are typically non-linear and noisy, complicating manual feature extraction. Furthermore, the deep learning models required to process such data are computationally expensive, necessitating scalable training solutions for practical deployment. This paper addresses these challenges by introducing a parallelized DBN framework for precise fault localization and identification.

Methodology: Deep Belief Networks and Parallelization

The proposed diagnostic model utilizes a Deep Belief Network (DBN), which consists of a stack of Restricted Boltzmann Machines (RBMs) followed by a Softmax classification layer.

Architecture of the Deep Belief Network (DBN) used for fault classification.

The training process involves two distinct stages:

  1. Unsupervised Pre-training: A greedy layer-wise approach is used where each RBM layer learns to represent features from the preceding layer. This initializes the network weights effectively without requiring labeled data.
  2. Supervised Fine-tuning: The Backpropagation (BP) algorithm refines the entire network using labeled fault data to ensure high-level features translate into accurate diagnostic outputs.

To handle large-scale datasets and deep architectures, the authors implemented these algorithms within the MapReduce framework. This parallelization covers both Gibbs sampling during RBM pre-training and weight updates during the BP stage, significantly reducing total training time.

Experimental Setup and Datasets

The research utilized a physical communication experiment system to generate realistic data. Output signals from both the transmitter and receiver were measured, with each sample comprising 1024 data points. A total of 1,500 samples were collected for several operational scenarios.

Experimental scenarios defining fault locations and phenomena.

As shown in the table above, the scenarios ranged from normal operation (Scene A) to specific hardware failures, such as receiver frequency drift (Scene B) and low transmitting power due to antenna or power amplifier module issues (Scenes C and D). The model was specifically tested on its ability to distinguish between four identical power amplifier units (Scenes D-1 through D-4).

Results and Discussion

The DBN model demonstrated superior performance over traditional diagnostic methods:

  • High Precision: The model achieved 99.6% accuracy in detecting receiver frequency drift and an average of 92.7% when distinguishing between antenna and power amplifier failures.
  • Granular Localization: Remarkably, the model could distinguish between failures in four identical power amplifier modules with an average accuracy of 81.6%, a task significantly harder than general fault detection.
  • Automated Feature Engineering: Unlike traditional methods requiring expert-defined frequency domain features, the DBN automatically identifies essential data patterns through its hidden layers.
  • Scalability: Increasing the DBN depth (up to 8 layers) improved reconstruction accuracy, with the MapReduce framework ensuring that training remained computationally viable.

Conclusion

The proposed DBN-based method offers a simple yet powerful approach to military communication system health management. While the current laboratory setup provides a proof of concept, future work will focus on accounting for real-world signal degradation (e.g., noise and fading) and moving toward real-time, online diagnostic capabilities. This research provides a critical software foundation for maintaining the operational readiness of advanced communication equipment.

Summary updated

Abstract

To meet the need of fault diagnosis for military communication system, an effective method based on deep belief (DBN) net is proposed. During the fault diagnosis, the bottom layer of DBN model is used to receive the input fault signals to extract the fault features and the fault classification results will be outputted after softmax classified. Accordingly, algorithms for DBN model and training and RBM parameter learning have been designed. To reduce the running time, parallel solutions based on MapReduce framework have been provided. In order to test and verify the effect of DBN fault diagnosis, the communication experiment system is built in the laboratory which the output signals of the transmitter and the receiver are measured and collected as the original data for further learning and training. Compared with the traditional fault diagnosis methods, it can be found that DBN method has high accuracy in fault diagnosis and the process is simple and friendly. It is impossible to realize real-time diagnosis and online diagnosis for the communication system. The research can be applicated to the health management of communication equipment, and it will provide advanced technical support and software program for the health of communication equipment
BibTeX
@article{li2021method,
  author = {Juan Li and Bin Chen},
  title = {{The Method of Communication System Fault Diagnosis Based on Deep Belief Net}},
  journal = {International Journal of Circuits Systems and Signal Processing},
  year = {2021},
  doi = {10.46300/9106.2021.15.105},
}
Figure