Research On Fault Diagnosis And Health Assessment Of Distributed Brake

The new generation of locomotive brake generally adopts distributed structure and intelligent modular design,which greatly improves the system operation,safety redundancy and control response,and improves the reliability of locomotive operation.However,the complexity of distributed structure and the coupling functions between modules bring great challenges to the accurate fault location and health evaluation of the brake system.In this thesis,the running mode and working mechanism of the recent distributed brake system are analyzed.The fault diagnosis and health evaluation of the distributed brake system are investigated by nonlinear aerodynamic theory,intelligent diagnosis theory,reliability theory and machine learning method.The main work is as follows:Based on the characteristics of modularization,intelligence and networking of distributed brake system,the working mechanism of the brake system is analyzed from three aspects.The three aspects include air pipeline,electrical interface and network transmission with respect to automatic braking function,individual braking function and emergency braking function.According to the structural level of ”component-module-function-system”and considering the coupling relationship between components under different working modes,the distributed brake system is modeled based on multistage pressure propagation mechanism.The model includes the pressure transfer model from pre-controlled air cylinder to storage air cylinder with different volumes,and the multi-pipeline joint control model of twoway valve.The key issues of fault diagnosis and health evaluation of the distributed brake system are investigated deeply,and the overall framework of intelligent fault diagnosis and health evaluation of the distributed brake machine is presented.In order to address the fault detection challenge brought by distributed brake operating mode diversity,a mode self-adapting diagnosis method is proposed based on physical-data fusion.According to the pressure data of the key components of the brake system,the features with high dimension are extracted to train the Gaussian mixture model,and the adaptive recognition of the working mode of the distributed brake system is realized.The physical model under different running modes is constructed with the identified parameters.Meanwhile,the data-driven model of the function module is constructed based on gradient boosting decision tree.The adaptive neural fuzzy inference system is utilized to fuse the output of the physical model and the data-driven model.Then the accurate pressure response curve of the system is calculated.By analyzing the residual sequence of the pressure response curve and the pressure measurement value,the fault detection of modules in the brake system is achieved accurately and effectively under different working modes.To address the challenge of fault location caused by the running status switch of distributed modules during the operation,a transfer learning based fault diagnosis method is proposed to diagnose the fault of key components.The off-line testing data of components is utilized to replenish the locomotive operation data,and a data-driven model is trained to accurately diagnose faults.The source domain data is generated by off-line test for components.According to the information divergence,the source domain data closest to the actual operation data is selected as the input characteristics of transfer.Then the advanced features of source domain data are extracted by using the convolution neural network.The model parameters are updated by adversarial training of fault classifier and domain classifier and the back propagation of diagnosing error.Further,the source domain extension strategy is designed based on model Transfer.By utilizing the above methods,the fault features and diagnosis knowledge are transferred.The fault diagnosis model for identifying critical components is obtained,and the fault accuracy of critical components is improved.As for the issues of imbalanced health-fault data,few labeled fault samples,and heavy labeling workload of unlabeled samples,which makes it difficult to evaluate the health status of key components of brake system,a component-level and system-level health evaluation method of brake system based on semi-supervised learning is proposed.For component-level health assessment,according to the importance and representativeness of the samples,two homogeneous regression learners are used for co-training to evaluate the accuracy of sample labeling with confidence,expanding the available training data set,and improving the accuracy of health assessment through continuous iteration.Furthermore,based on the reliability analysis theory,a system level health assessment method based on improved Bayesian network is proposed.The method synthesizes the health states of multiple components under different braking modes from the node failure rate,node importance and the connection coupling relationship between modules.The health index on system level is constructed,and the accurate multilevel health assessment is realized for the components and entire system of distributed brake systems.In this thesis,a fault diagnosis and health assessment platform is designed and implemented for the distributed brake system.The locomotive command,train states and process data of the brake system are collected through the train bus network and the internal bus network of the brake system.The test platform of the brake system and its key components are constructed.It realizes the data preprocessing,the working mode identification,the module fault detection,the fault diagnosis of key components and the health status evaluation of the brake system.The effectiveness of the proposed fault diagnosis and health evaluation method is verified by the practical experiments on the platform.