What question did this study set out to answer?

The aim is to improve fault diagnosis in rail transit systems using deep learning techniques despite limited data.

February 8, 2026

Transfer Learning and Multi-Task Framework for Fault Diagnosis in Rail Transit Systems Using Deep Neural Networks

Key Points

The aim is to improve fault diagnosis in rail transit systems using deep learning techniques despite limited data.
Utilized a pre-trained ResNet-50 model for transfer learning.
Frozen the convolution layer and fine-tuned the fully connected layer.
Employed a 1D CNN to extract time series features from sensor signals.
Implemented a multi-task learning framework to classify different fault types.
Achieved diagnostic accuracies of 92.5% for power system faults.
Achieved 91.3% accuracy for brake system faults.
Achieved 94.1% accuracy for sensor faults.
Achieved 89.7% accuracy for signal distortion faults.
Achieved 88.4% accuracy for noise interference faults.

Abstract

This study aims to address the issues of insufficient fault data and weak model generalization ability in rail transit train fault diagnosis, which are critical challenges affecting safety and efficiency in practical applications. This paper adopts a method that combines transfer learning (TL) and multi-task learning (MTL) to solve this problem. First, by using the pre-trained ResNet-50 model, the knowledge of the source domain is transferred to the target domain; the convolution layer is frozen; the fully connected layer is fine-tuned, thereby reducing the dependence on the labeled data in the target domain. Then, a 1D convolutional neural network (1D-CNN) is utilized to extract the time series features of rail transit trains, and combined with multiple layers of convolution and pooling layers, the time domain and frequency domain features of the sensor signals are recorded. The MTL framework shares the feature extraction layer, and independent tasks are classified through a dedicated output layer. Multiple fault diagnosis tasks are jointly trained to enhance the model’s diagnostic ability for different types of faults. Experimental results show that the accuracy of the proposed model in the diagnosis of power system fault, brake system fault, sensor fault, signal distortion fault, and noise interference fault is 92.5%, 91.3%, 94.1%, 89.7%, and 88.4%, respectively. Through these technologies, the proposed method has great application potential and effectively enhances the efficiency and accuracy of rail transit train fault diagnosis.

Perguntar à IA

Bookmark