Fault Diagnosis of Traction Rectifier Based on t-Distributed Stochastic Neighbor Embedding Fusion Optimal Multi-Band Box Dimension
In the traction converters of electric locomotives,the traction rectifier has a very high fault rate due to its complicated control and operating conditions.Fault diagnosis of rectifier can improve the operation and maintenance ability of electric locomotives and ensure the safety and stability of the train.The traditional rectifier fault diagnosis technology is limited in system modeling and signal analysis under variable load and strong noise environments.The data-driven rectifier fault diagnosis method is independent of the precise system model and prior knowledge of signal mode.However,redundant information and feature conflicts exist in the feature extraction process.The extracted fault feature dimension is large,which increases the difficulty of fault identification and diagnosis accuracy.This paper proposes a rectifier fault diagnosis method based on fractal box features fused by manifold learning.The fractal box dimension feature is extracted from the fault signal,the effective fault feature is obtained by manifold learning fusion,and the fault component of the traction rectifier is accurately diagnosed.Firstly,signals of the traction rectifier under different working conditions and fault modes are decomposed using nine Daubechies wavelet functions,and the information of 9 frequency bands for each fault mode is obtained.According to the larger energy entropy ratio of each frequency band coefficient,the optimal frequency band is selected.Secondly,the fractal box dimension of the optimal frequency band is calculated.The t-distributed random neighborhood embedding(t-SNE)algorithm fuses the fault features.Finally,a support vector machine is used for pattern recognition to realize fault diagnosis of the traction rectifier.The optimal frequency band decomposes signal decomposition with different wavelet functions.The fractal box dimension quantitatively describes nonlinear and non-stationary signals.The t-SNE algorithm lowers the conflict and redundancy among features,reducing the difficulty of fault identification and the computational burden of the classifier.Traction rectifier results show that when the SNR is 10 dB and the preset ratio is 3∶2,the average diagnostic accuracy is 95.333%.Increasing the SNR and preset ratio improves the diagnostic results,indicating the proposed method is robust regarding noise and preset ratio.The influence of voltage fluctuation and load change shows that the remaining diagnosis results are more than 97.9%,and the diagnosis result is 89.791 7%under the no-load state.The proposed method can still obtain simple and effective fault characteristics when the voltage fluctuation and load change on the network side.Comparisons with feature fusion algorithms in fusion time,training and diagnosis time,and computational complexity verify the superiority of the proposed method.The following conclusions can be drawn.(1)The proposed method only needs to extract and fuse the fault features of the signal and finally locate the fault component through pattern recognition.Compared with the model-based and signal-based methods that rely on the system model and signal analysis,it is practical for the fault diagnosis of traction rectifiers under complex working conditions.(2)The optimal band box dimension feature can quantify the dynamic changes of system fault signals,reflect different operating states,and distinguish different faults.(3)The proposed fusion algorithm effectively solves the redundancy and conflict problems among features,reduces the cost of pattern recognition,and improves the diagnosis rate.
Traction rectifierenergy entropy ratiofractal box dimensiont-distributed random neighborhood embeddingfeature fusion
国家科技期刊平台
NSTL
万方数据
