Research On Stator Fault Diagnosis Method Of Asynchronous Motor Based On Machine Learning Model

Stator faults are one of the most common types of faults in asynchronous motors.Most of the phase-to-phase short circuit and ground short-circuit are caused by the development of undetected inter-turn short-circuits.Therefore,it is of great significance to make effective early diagnosis of stator inter-turn short-circuit faults.In recent years,the rise of the wave of big data and artificial intelligence has brought new ideas and methods to the field of scientific research.The development of the industrial internet has also put forward new requirements for the data-driven intelligent diagnosis of electrical equipment.This paper studies the method of asynchronous motor stator failure diagnosis based on machine learning models.Based on the exploratory analysis of the motor data,this paper first studies the fault diagnosis method based on the shallow machine learning models,the sequence components of the stator fundamental wave current and apparent impedance and the three-phase current phase difference are comprehensively used as features.Random forest and support vector machine were used to diagnose stator inter-turn faults and verified by experiments,achieved 97.4% and 95.3% accuracy on the test set respectively.Next,a diagnosis method using a deep feedforward neural network model is proposed,which also relies on feature engineering and is verified on the Tensor Flow platform,and finally achieves a test set accuracy rate of 97.1%.The results show that negative-sequence current and negative-sequence apparent impedance are good features for stator turn-to-turn short-circuit fault diagnosis,and the random forest model has short training time and high accuracy,and has good interpretability and robustness.Finally,two methods based on convolutional neural networks are proposed.The first one uses the three-phase current sampling signals of the stator to construct 3-channel input data to train the convolutional neural network,and 97.1% accuracy were obtained on the test sets.In the second method,the current fundamental wave in phase A is filtered,and 5-layer wavelet packet decomposition on the filtered signal is performed to extract the first 8 low-frequency decomposition signals to form an 8-channel signal as the input of the convolutional neural network,and the accuracy of the test set is 98.6%.The results show that under the condition of a large data set,the convolutional neural network model can learn the failure pattern in the time series signal and has a high generalization performance.