Research On Remaining Useful Life Prediction Method Of Machinery Based On Deep Recurrent Prediction Network

Predicting the remaining useful life(RUL)of machinery and determining the appropriate maintenance timing is critical to the safety and reliability of machinery.How to effectively use the massive monitoring data to obtain accurate RUL prediction results has become an urgent problem.Deep recurrent neural network can effectively mine the degradation information in historical monitoring data,and thus it has become a powerful tool to solve the above problems.Aiming at the problem that most of the existing prediction methods lack consideration,i.e.,how to effectively fuse multi-sensor monitoring data for RUL prediction,a multi-source data prediction method based on multi-channel attention bidirectional long short-term memory neural network(LSTM)is proposed.In the proposed method,sensitive features are firstly extracted and selected from the monitoring data of each sensor individually.Then,multi-channel bidirectional LSTM is constructed with the sensitive features as the inputs.Meanwhile,time attention and channel attention are utilized to realize the adaptive information fusion.Finally,the effectiveness of the proposed method is verified by the multi-sensor monitoring data from life testing of milling cutters.Aiming at the problem that traditional data-driven remaining life prediction methods can only obtain prediction results as point estimation,a prediction uncertainty quantification method based on residual convolutional LSTM network is proposed.In the proposed method,a new Res Net-based convolution LSTM layer is stacked with convolution LSTM layer to extract degradation representations from monitoring data.Then,based on the presumption that RUL prediction results generally follow a normal distribution,an appropriate output layer is constructed to quantify the uncertainty of the forecast result.And the loss function that comprehensively considers the prediction accuracy and interval coverage effect is constructed to further optimize the performance of the network.Finally,the effectiveness of the proposed method is verified by the accelerated life testing data of rolling bearings.Aiming at the problem of remaining life prediction under variable operating conditions,a bidirectional convolutional recurrent attention neural network based on residual input is proposed to predict the RUL under variable operating conditions.The method first constructs a bidirectional Res Net-based convolutional LSTM layer to directly obtain degradation representations from original monitoring data.Then,time attention mechanism is introduced to obtain time attention weight,and working condition information refining unit is constructed to obtain the working condition influence factor.Next,the weight and the factor are used to weight the degradation representation to alleviate the influence caused by the variance of working condition on prediction results.Finally,the effectiveness of the proposed method is verified by the testing data of rolling bearings accelerated life with variable speed.The proposed prognostic method for multi-sensor data is applied in the RUL prediction of Stirling refrigerator.First of all,the refrigerator condition monitoring and RUL prediction system is designed and developed,which integrates the functions of condition monitoring and RUL prediction.And the proposed method is developed for RUL prediction of refrigerator.Finally,the proposed method is verified by the real-case degradation data of refrigerators.