Research On Remaining Useful Life Prediction Method Of Rolling Bearing Based On Deep Learning

As an essential component of rotating machinery,the health status of rolling bearings will directly affect the safety and stability of the entire mechanical equipment.In order to effectively monitor the operation status of equipment and ensure its regular operation,the technology of bearing fault prediction and health management(PHM)has been proposed and widely used.Remaining useful life(RUL)prediction technology,as an essential part of PHM,can effectively identify the degree of the bearing failure,complete the assessment of its remaining useful life,providing the possibility for timely maintenance of equipment and ultimately ensuring its safe and stable operation.Depth learning model can effectively capture the spatiotemporal correlation in condition monitoring signals,thus it has been widely applied in RUL prediction of bearings.However,most prediction models based on deep learning only a single predictor structure,making it challenging to integrate the degradation features of bearings effectively.In addition,the prediction models are easily affected by information redundancy,leading to poor robustness of the model.The key to constructing a prediction model is the selection of hyperparameters.Most deep learning models use deep learning optimization algorithms to tune parameters,which often leads to inaccurate parameter selection,resulting in poor final prediction results.At the same time,an inaccurate setting of the prediction threshold results in low prediction accuracy of the model.To address these issues,this study conducted research on the RUL prediction of rolling bearings based on the XJTU-SY bearing dataset from several aspects.(1)A hybrid prediction model for predicting RUL of rolling bearing,CA-LSTM,was proposed to solve the problems of information redundancy,insufficient feature expression capability of the multi-layer perceptron(MLP)structure in the long and short-term memory(LSTM)unit,and the impact of manual threshold setting on model prediction accuracy.This model constructs a feature extraction module by combining one-dimensional convolution and convolutional block attention mechanism(CBAM)to solve the problem of model susceptibility to information redundancy.It replaces the MLP structure at the input end of the LSTM unit to address the problem of insufficient expression of fault features at the input end of LSTM.The model uses the remaining life percentage as the input label to avoid the impact of artificial threshold setting on network prediction accuracy.The validation results on the XJTU-SY bearing dataset demonstrate that the proposed method has certain advantages in bearing life prediction accuracy.(2)In the fault prediction task,the recurrent structure and convolution structure have defects in extracting the global dependence of bearing degradation features;when Transformer model is directly used for feature extraction of time series data,it has low performance on local features,thus the dilated and multi scale-feature fusion-Transformer(DMSF-Transformer)prediction model is proposed for bearing RUL prediction in this thesis.This method introduces a Transformer encoder structure that completely abandons recurrent and convolution structures to improve the characterization of the global dependency relationship of fault features in bearing degradation.The model introduces a dilated convolution module in the Transformer encoder structure to enhance the model’s learning ability for local features of bearing degradation and proposes a multi-feature fusion predictor structure to improve the model’s integration capability of fault features.The validation results demonstrate that the proposed improved model has a higher cumulative relative accuracy(CRA)evaluation value.(3)Aiming at the problem that most deep learning model parameters need to be adjusted manually,which leads to the complexity of model parameter selection and the low prediction accuracy of the model,an improved local search based competitive particle swarm optimization(L-CSO)is proposed to optimize the parameters of DMSF-Transformer prediction model.In this algorithm,competition and exploration mechanism is introduced to find the optimal particle through a new local search,and the problem of particles falling into the local optimal solution is effectively avoided by using the competition between the particles outside the population and the optimal particles inside the population.In order to verify the effectiveness of the algorithm,it is compared with other seven classical algorithms on 17 test problems,and the results show that the algorithm can achieve more accurate optimal solutions on multiple test problems.Finally,the improved algorithm is applied to the parameter optimization of the prediction model to construct the L-CSO-DMSF-Transformer prediction model.The experimental results show that the prediction accuracy of the model is improved obviously.