Health State Identification And Prediction For Rotating Machinery Under Complex Conditions Based On Deep Learning

With the development of "Industry 4.0",rotating machineries are developing in the direction of complexity,informatization and intelligence.At the same time,higher requirements are put forward for the intelligence,accuracy,adaptability,and engineering practicability of health state identification and prediction methods.Thanks to the powerful adaptive feature extraction capability of deep learning models,the intelligent machinery health monitoring method based on deep learning has a good development prospect.Existing health state identification models and remaining useful life(RUL)prediction models based on deep learning have shown excellent performance under simple monitoring conditions such as constant operating condition and single vibration signal.However,it is difficult to achieve accurate health state identification and RUL prediction under complex industrial site monitoring conditions.On the one hand,due to the influence of multi-point vibration monitoring,complex and variable operating conditions,and time-varying degradation rate,the information components contained in the monitoring signals are very complex,so it is difficult to extract effective health state features and degradation trend features.On the other hand,under complex monitoring conditions such as variable operating conditions,time-varying degradation rates,and the condition that it is difficult to obtain historical fault data,the existing deep learning models cannot meet the two restrictive assumptions of ―independent and identically distributed‖ and ―sufficient training samples‖.Based on the above background,this paper studies the health state identification and RUL prediction methods of rotating machinery under complex conditions based on deep learning theory.Improved health state identification models and RUL prediction models suitable for complex monitoring conditions are proposed according to the characteristics of monitoring signals under the conditions of multi-point vibration monitoring,time-varying operating,insufficient training samples,and time-varying degradation rate,respectively.This paper solves the two problems of "poor feature extraction ability" and "not satisfying the two basic assumptions of deep learning" existing in the existing deep learning based health state identification and RUL prediction methods under complex monitoring conditions.The detailed research works are as follows:In order to solve the problem that it is difficult to fully extract health state features from raw multi-point vibration signals under the multi-point vibration monitoring condition,the Time-Distributed Convolutional Long Short-Term Memory(TDConv LSTM)model is proposed.According to the characteristics of non-linear spatiotemporal correlation and high sampling rate of multi-point vibration signals,the proposed model extracts the relevant features of multi-point vibration signals in both spatial and temporal dimensions.A time-distributed local spatiotemporal feature extraction module and a global spatiotemporal feature extraction module are successively stacked to fully extract deep spatiotemporal fusion features,which improve the feature extraction ability and the identification accuracy of the model under the multi-point vibration monitoring condition.Through the gearbox health state identification experiment under the multi-point vibration monitoring condition and the feature visualization analysis,it is proved that by fully extracting spatiotemporal fusion features,the accuracy of health state identification under the multi-point vibration monitoring condition can be improved.Under variable operating conditions,existing health state identification models based on deep learning are difficult to extract health state features that are insensitive to operating conditions and cannot meet the IID assumption,resulting in poor identification performance under unknown operating conditions.In order to solve this problem,a novel end-to-end deep learning model named adaptive weighted multiscale convolutional neural network(AWMSCNN)is proposed for adaptive health state under variable operating conditions.A multi-scale feature extraction module and a multi-scale feature adaptive weighting module are designed to adaptively modulate multi-scale features according to changes of operating condition,which can enhance the features that are sensitive to health state and suppress the features that are sensitive to operating conditions and other useless features.This model can finally extract health state features that are not affected by the operating conditions.Through the bearing health state identification experiments under variable operating conditions,visualization analysis of the weight vector,and visualization analysis of features,it is proved that the AWMSCNN model has the adaptive ability to noise,speed changes and load changes,and can improve the health state identification accuracy under variable operating conditions.In order to solve the problem of insufficient model training samples when the AWMSCNN model is applied in actual projects,this paper combines transfer learning with the AWMSCNN model,and proposes a model training method in the case of few training samples based on across-machine transfer of model parameters.The proposed method improves the identification accuracy of the AWMSCNN model in the case of few training samples.The proposed method is verified by the train wheelset bearing health state identification experiment under the compound scenario of few training samples and variable working condition.In addition,the transferability of each module of the AWMSCNN model and the retraining strategy of the target model are also been studied,which provides a basis for the selection of the retraining strategy.In order to solve the problem that existing methods are difficult to dynamically and accurately perceive the degradation trend of rotating machinery and accurately predict the RUL under the condition of time-varying degradation rate only by extracting fixed-scale degradation trend features,the Adaptive Weighted Multi-Scale LSTM(AWMSLSTM)model is proposed for RUL prediction.In the AWMSLSTM model,a multi-scale degradation trend features extraction module and a multi-scale degradation trend features adaptive weighting module are designed to make the model have the ability to adaptively adjust the contribution rate of degradation trend features in different time scales to RUL prediction.Through bearing RUL prediction experiments under the condition of time-varying degradation rate and the visualization analysis of the weight vector,it is proved that the AWMSLSTM model can dynamically extract the degradation trend features of rotating machinery according to the time-varying degradation rate and can improve the RUL prediction performance.