A Novel Fault Diagnosis Method for ZPW-2000A Track Circuit Applied to Small Samples

The occurrence of faults in the ZPW-2000A track circuit exhibits characteristics of diversity and sporadicity. Diversity implies a wide variety and complexity of fault types, while sporadicity results in limited fault data being available, presenting the features of high-dimensionality and small sample size. To enhance diagnostic performance in multi-task scenarios with small sample datasets, this paper proposes a novel fault diagnosis approach. Specifically, the model first employs a Long Short-Term Memory Network combined with an attention mechanism to extract complex information from temporal data. Subsequently, it utilizes Linear Discriminant Analysis for dimensionality reduction of the features, eliminating redundant information. Finally, a multi-task fault classification model is constructed, which can mine data distribution information and manifold structures. Experimental results demonstrate that our approach achieves higher accuracy and robustness compared to traditional machine learning methods, such as Logistic Regression, Convolutional Neural Network, K-Nearest Neighbors, Decision Trees, Random Forests, and Naive Bayes. It improves the average accuracy by 0.27%, 0.21%, 2.20%, 0.44%, and 0.50% respectively across five small sample datasets, while also exhibiting significant advantages in computational performance.