Data-Driven Fault Diagnosis Methodologies Of Inverters In Traction Converters

With the further expansion of high-speed railway network,it is extremely important to establish a reliable fault diagnosis system to guarantee the safety of high-speed railways.Traction inverter is a key component of a high-speed railway traction drive system,in which open-circuit fault is prone to occur in Insulated Gate Bipolar Transistor(IGBT),resulting in a high failure rate of IGBT.IGBT open-circuit faults will cause electromagnetic torque ripple of a traction motor,even burn down the motor.The research of open-circuit faults diagnosis algorithm of traction inverters can provide decision support for fault-tolerant control of the traction drive system and subsequent maintenance of maintenance personnel.In the paper,online intelligent diagnosis algorithms of open-circuit faults of traction inverters based on data-driven were studied.The convolutional neural network was applied to the diagnosis of IGBT open-circuit faults of traction inverters.For different fault diagnosis scenarios,the paper proposed several diagnostic algorithms with good performance.Firstly,based on MATLAB/Simulink and the actual parameters of a certain type of Electric Multiple Units(EMU),the traction drive system simulation model of the EMU was constructed.The paper used the model to supplement the experimental data.Then the paper classified the IGBT open-circuit faults of traction inverters.On this basis,the influence of each type of open-circuit faults on the stator current of traction motors was studied.Finally,fault features of the current when open-circuit faults occurred were summarized.Since convolution neural networks can automatically extract fault features,the paper first proposed a fault diagnosis plan based on Alex Net.The plan preprocessed data by using data enhancement techniques such as angle domain synchronous resampling and sliding window sampling,and then the paper transformed the stator current signals into images.The experimental result shows that when the test set and training set are in the same working condition,only using 5% of the total sample set to fine-tune Alex Net,the plan can achieve 100% diagnostic accuracy.Alex Net has a large number of parameters,and there are many blank areas in the images generated by the current signal,which results in a waste of resources.To solve these problems and further enhance the domain adaptability of diagnosis algorithms,a method to transform the three-phase stator current into gray-scale images and a convolutional neural network named CI-CNN based on extracting the comprehensive information of three-phase current were proposed.Compared with Alex Net,CI-CNN has a lighter structure,better domain adaptability,and stronger anti-noise ability.After trained using the data of different working conditions,CI-CNN can achieve more than 99% diagnostic accuracy in working conditions outside the training set.At the same time,the paper proposed an anti-noise method based on segment fitting,which can further improve the anti-noise ability of the diagnosis algorithm without changing the parameters of CI-CNN and is suitable for online diagnosis.To obtain better domain adaptability,CI-CNN needs data of various working conditions for training.To solve the problem,the paper has done two jobs,the first is to propose an integral based waveform repair method,the waveform repair method can effectively eliminate the current distortion caused by open-circuit faults of traction inverters,enlarge the common features of the same type open-circuit faults,reduce the private features of current under different working conditions,at the same time,it can avoid the influence of the difference between current phase of different samples on the output of convolutional neural network.The second is to propose a fault diagnosis algorithm which combines data-driven and signal-based,the fault diagnosis algorithm mainly contains two parts: a convolutional neural network named T-CNN for single-phase current open-circuit faults diagnosis and five fault conflict resolution rules.After trained only using the data of one working condition,the diagnosis algorithm can achieve more than 98% diagnostic accuracy in other working conditions,and it is also suitable for other high-speed railways with two-level inverters.The diagnosis algorithm has a wide application space.