Research On Key Problems Of Fault Prediction And Maintenance Based On Improved Deep Learning

Predictive maintenance of critical equipment in the intelligent manufacturing process can transform the equipment fault diagnosis process from planned maintenance to conditional maintenance,which is essential for safe and stable operation of the system.Real-time early fault diagnosis and accurate life prediction model construction are two key aspects of predictive maintenance.As an advanced data feature extraction method,deep learning has received extensive attention in fault diagnosis and residual life prediction.However,the traditional deep learning model is incapable of real-time accurate diagnosis of mechanical faults that only exhibit significant signs in the frequency domain.On the other hand,the existing full-connection of the LSTM(Long Short-Term Memory)-based deep learning residual life prediction model is available.Network structure has high computational complexity and cannot guarantee the accuracy of remaining life prediction.In the case that the accurate fault mechanism model of the equipment cannot be obtained,this paper uses DNN(Deep Neural Networks),LSTM and other deep learning methods as feature extraction tools to carry out research on key problems of fault prediction and maintenance based on equipment operating state big data,and solve traditional deep learning methods.Problems in improving the real-time of fault diagnosis and improving the accuracy of remaining life prediction.The main innovations of this paper are as follows:(1)A deep learning fault diagnosis method based on fusion of differential geometric features is proposed.The traditional deep learning method can only accurately diagnose the amplitude anomaly faults in real time when classifying faults in the time domain,and has limited effect on the frequency fault diagnosis of amplitude-invariant frequency anomalies.The fault classification result obtained by deep learning of spectral data obtained after Fourier transform is accurate but cannot guarantee the real-time of fault diagnosis,and real-time accurate fault diagnosis is the premise of predictive maintenance.To this end,this paper introduces the differential geometry of slope,curvature and other dynamic trend information to characterize the local frequency anomaly of the data in the time domain,and design the feature fusion layer in the depth learning framework,by validating the different features extracted by the three DNNs.The time domain features with local frequency information are obtained by fusion,and fault classification is performed.It not only ensures the real-time performance of frequency-based fault diagnosis but also the accuracy.(2)Using the sparse idea of Highway network to design a sparse denoising LSTM network that suppresses redundant neurons to achieve more accurate residual life(RUL)prediction.Different from the idea that the traditional Highway network is sparse in time direction,this paper transforms the traditional LSTM network by designing sparse gates,and suppresses those neurons that have contributed little to the next layer in the previous layer,and highlights those nerves that contribute more.The role of the element,thereby achieving the goal of sparseness and " denoising " at the same time.When the time series is long,the prediction accuracy of RUL prediction using the sparse denoising LSTM network(Sparse Denoising LSTM,SD-LSTM)is high,and the sparse gate structure can also reduce the computational complexity to a certain extent.(3)The online update mechanism of the sparse denoising LSTM prediction model is designed,and the new online data generated by the in-service equipment is fully utilized to improve the prediction accuracy of the LSTM-based small sample data RUL prediction model.In general,the same life cycle data sample size of the same type of equipment is small enough to establish an accurate RUL prediction model.On the other hand,new sample data is collected as the equipment is in service.For this reason,by using the original LSTM network parameters as the initial value,the iterative learning is continued by using the newly acquired online sample data,and the LSTM network parameters are gradually optimized to obtain a more accurate RUL prediction model.