Gear as the transmission device of mechanical equipment, its health has a great impact on the stability and life of the equipment. With the improvement of the integration of modern mechanical equipment, the accuracy of gear is also higher and higher. In the long-term operation, it is easy to wear, resulting in tooth fracture. Monitoring the health status of gear is an important means to ensure the reliable operation of mechanical equipment. Based on reliability and economy, gear fault prediction and health management (PHM) has been paid more and more attention, and the remaining useful life (rul) prediction is the core research content. With the development of information sensing technology, the number of monitoring points and signal sampling frequency of the equipment increase, and the equipment can obtain a large amount of data in the process of operation, which promotes PHM into the era of big data. How to use massive data to mine the degradation rule of equipment state, so as to predict the remaining life of equipment, has become a big challenge for prediction and health management.
The gear is usually packaged in the, and the information received by the sensor reflects the degradation state of the equipment. The data-driven residual life prediction method is based on the theory of statistics and artificial intelligence. It uses sensors to obtain the data that characterizes the degradation of equipment status, so as to realize real-time residual life prediction. Liu Ying et al. Used time series ARMA (auto regressive moving average) prediction model to diagnose steam turbine fault according to simulated pseudo signal, but ARMA prediction model is suitable for short-term prediction of time series and processing stable data with certain regularity. The first mock exam is to predict the faults of the gearbox by using WANG vector and machines regression (SVR) and X Support (spectrum). The results of the model are better than those of the single model. But the structure of the model is complex, when the sample data is large, the prediction results are time-consuming. Davies TM and others used artificial neural network (ANN) to predict the faults of diesel engine turbocharger and automobile engine, and used genetic algorithm to optimize the network structure parameters of ANN. Neural network plays a key role in Ann. The complexity of calculation depends on the number of neurons and the number of hidden layers. If there are too many samples, it is easy to cause “dimension disaster” or “over learning problem”.
In recent years, deep learning has been widely concerned. Deep learning constructs a neural network model with many hidden layers through massive training data, transforms the feature representation of samples in the original space into a new feature space, and extracts the potential mapping relationship in the input and output samples layer by layer. Recurrent neural networks (RNN) can memorize the previous sequence input and apply it to the later calculation, but it can not solve the long-term dependence problem. Long short term memory (LSTM) model can avoid the problem of gradient vanishing through model parameters and gating unit system controlling information flow, which makes time series information prediction more accurate. Haitao Zhao and others applied long-term and short-term memory network in fault diagnosis to classify faults, but empirical mode decomposition is needed for data. If the monitoring signal is interrupted, it will cause mode aliasing, which is not conducive to fault diagnosis. Zhao Jianpeng et al. Applied LSTM to single step state prediction of rotating machinery, and achieved better results than support vector regression. Wang Xin applied LSTM to the prediction of aircraft fault time series. Compared with many time series prediction models, LSTM has strong applicability and higher accuracy.
Aiming at the large-scale sequence prediction, a new real-time residual life prediction model based on memory mechanism recurrent neural network is established, and the accuracy of network model prediction is improved from the construction of network structure, learning rate and window setting. The random gradient descent algorithm with nesterov momentum is used to prevent the model from falling into local optimum, and the rmsprops algorithm is used to modify the adaptive rate of the model. Finally, the gear bending fatigue test data are used to verify, and compared with the time series prediction model, the test results show that the new memory mechanism recurrent neural network real-time residual life prediction model has good advantages in residual life prediction.
By studying the characteristics of gear vibration data, an improved real-time residual life prediction model based on memory mechanism recurrent neural network is established, and the parameters of the prediction model are optimized. The prediction model is applied to the research of gear residual life prediction. The results show that the memory mechanism recurrent neural network model has higher accuracy in predicting real-time residual life. The next step is to fuse the data of multiple monitoring points and study the residual life prediction model.