Research On Data-driven Based Anomaly Detection And Fault Diagnosis Of DK-2 Brake System

The DK-2 brake system is a widely equipped electro-pneumatic brake system in all kinds of locomotives,whose health status is crucial for train operation safety.Due to the coupled components,complex structure and multiple working modes,it is difficult to capture and analyze the characteristics of operation faults.Besides,the faults of the DK-2 brake system are nonlinear and intermittent,which makes it hard to develop feasible anomaly detection and fault diagnosis.To address this issue,the fault characteristics are obtained by combining correlation analysis and stage division.The Gaussian mixture model is constructed to detect anomalies of the system and fault diagnosis incorporated with semi-supervised and ensemble learning.The proposed method improves the accuracy of anomaly detection and fault diagnosis,and the main contribution of this thesis can be summarized as follows:Firstly,in order to detect running abnormalities,an anomaly detection strategy based on the Gaussian mixture model is proposed for the DK-2 brake system.By analyzing the system structure and working principle of the DK-2 brake system,the key components for the brake system are selected.The zero-normalized cross-correlation algorithm is utilized to calculate the correlation coefficient of the pressure data of the cascade and adjacent key components.Then the joint indices in the time domain and frequency domain are established based on the correlation sequence.Based on the aforementioned joint indices,a Gaussian mixture model representing the brake system’s working state is trained.The expectation-maximization algorithm is applied to solve the optimal parameters of the model,which is used to detect the abnormal running state.In the end,the accurate anomaly detection of the DK-2 brake system is achieved.Secondly,an improved semi-supervised random forest model is proposed to diagnose the multiple and coupling faults of the DK-2 brake system.The detected abnormal data was divided into stages based on the charging and discharging process to construct fault features.Then the features are filtered according to their fault sensitivity.The random forest model is trained to establish the relation between features and faults of the braking system.To solve the problem of unbalanced labels of fault samples in the running data set of the braking system,a semi-supervised methodology is introduced to expand the training data set.By combining clustering and ensemble learning,the accuracy of fault diagnosis for the DK-2 brake system is improved to ensure the safety of locomotive operation.Finally,the anomaly detection and fault diagnosis platform of the DK2 brake system is constructed.The availability and accuracy of the datadriven based anomaly detection and fault diagnosis method are verified by conducting the experiments on the platform.