Method for predicting remaining useful life of railway train bearing based on CAN-LSTM

A method for predicting remaining useful life of railway train bearings based on a convolutional attention network and a long short-term memory network (CAN-LSTM), comprising:

• constructing a degradation feature parameter set, comprising extracting parameters of time domain features, frequency domain features, and time-frequency domain features from bearing lifecycle vibration data, where both the time domain features and the frequency domain features comprise root mean square, kurtosis, peak-to-peak value, skewness, mean square value, and mean square error, and perform normalization processing;
• performing three-layer wavelet packet decomposition on vibration signals by a db5 wavelet packet to generate an energy ratio of 8 frequency subbands as the time-frequency domain feature;
• taking the time domain features, the frequency domain features, and the time-frequency domain features as a first feature sequence;
• defining a degradation time point label, taking a period from a bearing degradation point to a complete failure time as remaining useful life (RUL) of bearings, standardizing a period on the RUL to [0,1], and taking the period on the RUL as a label for training a convolutional attention network (CAN) model and a long short-term memory network (LSTM) model;
• y
• =
• i
• -
• k
• n
• -
• k
• -
• 1
• ,
• where i is a current moment value, n is a bearing life value, and k is a degradation start moment;
• CAN training, comprising adding attention mechanism to a convolutional neural network (CNN) to obtain the CAN model, extracting deep degradation features of a channel and a temporal attention in the vibration signals, and performing feature extraction on the first feature sequence by the CAN model to obtain a second feature sequence;
• RUL prediction, comprising inputting the second feature sequence into a LSTM, training the LSTM, where a cyclic network structure of the LSTM is capable of processing time sequences between different degradation states, reserving a model having a minimum error in a training process, and predicting the RUL of the bearings by the model having the minimum error to obtain a feature quantization value and finally obtain a RUL prediction value of the bearings;
• and
• prediction result evaluation, comprising smoothing the feature quantization value by a five-point sliding smoothing method for reducing influence of oscillation on the RUL prediction value, obtaining the RUL prediction value, and evaluating prediction results by an evaluation function;
• and
• predicting remaining useful life of railway train bearings based on CAN-LSTM in order to solve the problem that a conventional method for predicting the RUL of the railway train bearings lacks specific learning mechanism and affects prediction precision.