Deep Learning Network For Rotating Component Fault Diagnosis And Life Prediction Improvement Method

Rotating machinery and equipment in the era of intelligent manufacturing are highly automated,fast and networked,which requires accurate and timely understanding of their operating status,health management and fault diagnosis to maintain the reliable performance of the equipment,product integrity and production safety.In this thesis,the main area of research is in the field of Prognostics and Health Management(PHM),using deep learning technology to conduct fault diagnosis and life prediction for rotating machinery and equipment.The research includes the enhancement of the generalisability of deep learning network models,the feature learning capability of fault diagnosis models,the improvement of the prediction accuracy of remaining life prediction models,the solution of small samples and data imbalance,and the development of examples of intelligent diagnosis models,which are validated by the application of bearing and hydraulic pump experimental bench data and the actual monitoring data of slurry circulation pumps of power plant equipment.The main points of this thesis are as follows:(1)In order to solve the problems that traditional mechanical equipment fault diagnosis algorithms rely too much on manual feature extraction and feature extraction is unstable and lacks self-adaptability in the face of massive data.The One-Dimensional Deep Residual Network(1D-DRN)fault model is proposed,and the applicability of Convolution Neural Network(CNN)in the fault diagnosis task of rotating machinery is investigated.The effect of network depth and convolutional kernel size on the extraction effect of CNN feature extraction is analysed;the hidden layer in CNN is visualised and the learning process of the network during feature extraction is explored.The proposed method is experimentally validated in a rolling bearing,axial piston pump failure simulation dataset.(2)In order to solve the existing difficulties of modeling the degradation state of rotary mechanical equipment based on feature extraction and the unfavorable life prediction effect based on big data information,a deep learning-based Remaining Useful Life(RUL)prediction scheme is proposed.Improved One-Dimensional Deep Convolution Autoencoder(1D-DCAE)was used to model the degradation state of rotary machinery and obtain device health factors.For accurate data labelling,health factors are used as labels for the training data in supervised learning models.A multi-level Bidirectional Long Short-Term Memory(Bi-LSTM)network combined with a one-dimensional convolutional time-series pattern attention mechanism is proposed for life prediction of the full life cycle data of rotating mechanical equipment as a time series.Finally,the full-life experimental data of rolling bearings and gear pumps are used for analysis and validation.(3)In order to solve the problems of poor universality of current deep learning-based fault diagnosis models,difficulty in training deep models,high cost of repeated training,and difficulty in applying across data domains.Combining deep learning models with migration learning ideas,a model training strategy of pre-training + parameter fine-tuning is proposed and applied to rotating machinery fault diagnosis to achieve fast training and model migration of deep network structures.The model training method based on incremental learning is also proposed to establish typical samples to improve the recognition accuracy of fault diagnosis models in the face of new data.Finally,analysis and experimental validation are carried out using rolling bearing and hydraulic pump fault datasets.(4)To solve the problem of small sample data caused by the accident of rotating machinery and the difficulty of sample collection of typical faults in practical engineering.A One-Dimensional Deep Convolutional Generative Adversarial Networks(1D-DCGAN)model is proposed to generate small sample data using 1D-DCGAN for the purpose of balancing data sets,improving the quality of training data for fault diagnosis models to improve the fault diagnosis accuracy of rotating machinery and enriching the feature sample base of typical rotating machinery.(5)A distributed deep learning scheme based on big data technology is proposed to analyse the problem that it is difficult to effectively monitor and quickly process and calculate the huge amount of data brought by industrial big data.In view of the gap and limitations between experimental training and practical application of deep learning models,anomaly detection based on Support Vector Data Description(SVDD)is introduced in the practical application of deep learning models,and the strategy of first anomaly detection and then fault diagnosis is proposed in the fault diagnosis process.A set of deep learning development examples in PHM systems for rotating machinery and equipment is also designed.The overall solution is applied and validated in the actual working process of slurry circulation pump of power plant equipment.