Research On Remaining Life Prediction Method Of Rotating Machinery Based On Convolutional Neural Network And LSTM

Rotating machinery is one of the most important equipment in today’s industrialized era.Due to its complex system,poor operating conditions and long working cycle,key components are very prone to performance degradation or even serious failure,thus forcing equipment downtime and causing economic losses.Therefore,the Remaining Useful Life(RUL)prediction for critical components of rotating machinery is essential.With the rapid development of artificial intelligence and industrial internet technology,data-driven deep learning research methods are a recent research hotspot in the field of RUL prediction.However,there are still some problems in the current research methods.In this paper,we focus on the key components of rotating machinery as the research object,and use deep learning technology to develop RUL prediction method research,the main research content is as follows:(1)In order to solve the problem that the original data is disturbed by noise and contains single degraded information.In this paper,feature information is extracted from the time domain,frequency domain and timefrequency domain for the bearing vibration signals collected by homogeneous sensors,which can describe the bearing degradation process more comprehensively.The extracted features are selected by constructing a comprehensive evaluation index.Then,the deeper abstract features are extracted by multi-layer convolutional neural networks and long and short-term memory networks to fully obtain the hidden information in the degradation process.Finally,the overall prediction accuracy of the model is improved by fusing with the partially preferred sensitive features directly through the fully connected layer.The superiority of the proposed method in homogeneous sensor signal prediction is verified by two experimental cases.(2)In order to address the problem that the signals collected by heterogeneous sensors at different time scales or frequencies are independent of each other and inconsistent in magnitude,traditional models have low prediction accuracy and cannot accurately account for the diversity and complexity of time-varying multivariate signals.In this paper,we propose a multi-unit neural network structured prediction model based on self-attention mechanism by combining convolutional neural network to extract local feature information and bidirectional long-and short-term memory network to make full use of the dependency relationship between current moment data with historical data and future data.The design of independent network units enhances the feature extraction capability of the model by focusing only on the sensor signals of separate channels.The addition of the self-attention mechanism allows the model to assign higher weights to important feature information,which improves the prediction accuracy of the model.The superiority of the proposed method was verified by a large number of comparative experiments.(3)For supervised learning methods,sufficient labeled data is required for training,while the data collected in practical applications lack labels.In this paper,we propose a RUL prediction method based on semi-supervised learning Firstly,the variational autoencoder is used to pretrain the data with partial labels.The reduced-dimensional feature data are fed into a parallel network structure of temporal convolutional network and bidirectional long and short-term memory network to mine the hidden information of the data from the feature dimension and temporal dimension.The improved self-attention mechanism is introduced to assign different weights to the contribution of features at different times.The experimental study shows that the proposed model has better robustness and accuracy when the training data lack labels.