Research On Fault Diagnosis And RUL Prediction Method Based On Data Generation Model

Predictive maintenance of mechanical equipment in the intelligent manufacturing process can realize the transition from regular maintenance to conditional maintenance.Real-time accurate early fault diagnosis and accurate remaining useful life prediction are two research hotspots in predictive maintenance research.As an effective data feature extraction tool,deep learning has received widespread attention in data-driven fault diagnosis and remaining useful life prediction,but its effectiveness depends on the number and quality of available samples.The problem of unbalanced data in early slowly changing fault diagnosis will make the results of early fault diagnosis unreliable,which will lead to inaccurate remaining useful life prediction.On the other hand,massive life-cycle data is a key factor restricting the accuracy of deep learning-based remaining useful life prediction models.In the case that the precise mechanism model of the equipment cannot be obtained,this paper combines data generation models to study the problems caused by the imbalanced data and the lack of full life cycle data when the deep learning methods such as DNN and LSTM are used in the predictive maintenance of electromechanical equipment.The main innovations of the paper are as follows:(1)This paper proposes a deep learning fault diagnosis method based on global optimization GAN for unbalanced data to overcome the inherent shortcomings of the two-stage method of training GAN first and then training the DNN diagnostic model when traditional GAN is used for fault diagnosis of unbalanced data.This method first designs a new data generation model.The features of the original data are used in the training process of the generator to make full use of the useful information in the finite imbalance fault samples.At the same time,the fault diagnosis model is used as a discriminator of GAN,and the error of fault diagnosis is used to guide the training of the generator.Then,a global optimization training mechanism is designed to realize the joint optimization of the fault diagnosis model and the data generation model,thereby establishing an end-to-end unbalanced data fault diagnosis model with higher accuracy.Therefore,the purpose of improving the diagnostic accuracy of unbalanced faults on the basis of ensuring the overall diagnostic accuracy is achieved.(2)In the case of little or no full life cycle data,an accurate RUL prediction method based on non full life cycle data is proposed.First,the collected partial non-full life cycle fault evolution data is used to train a rough data prediction model for predicting subsequent fault evolution data,so as to obtain the estimated value of the full life cycle data.Then use the estimated full life cycle data to train the LGAN-LSTM.During the training process,the prediction error of the RUL prediction model and the prediction error of the data prediction model are used to guide the training of the LGAN generator.The global optimization mechanism is used to alternately train the generator,discriminator,and LSTM-based prediction model to improve the quality of the generator’s full-life cycle data and the online prediction ability based on the LSTM prediction model.