Health Assessment For Key Components Of Machine Tools Based On Machine Learning

With the rapid development of Industry 4.0,Computerized Numerical Control(CNC)machine tools are playing an increasing role in machining.On the one hand,as the CNC machine tools are becoming more refined and intelligent,the requirements for machining accuracy are becoming higher.On the other hand,as the machining process progresses,the wear of various key parts of the machine tool will cause inaccuracy,which may increase machining errors and even cause component failures and accidents.Based on the above background,there is an urgent need for online condition monitoring of machine tools,and developing corresponding methods for accurate health assessment of key components.In order to establish a complete equipment monitoring and health assessment system for domestic CNC machine tools,this thesis focuses on the evaluation of the health status of the ball screw,the analysis of spindle tool wear,and the prediction of bearing remaining life.A non-linear mapping between sensor data and equipment health status was established,so as to achieve the safety maintenance of machine tools.The main content of this thesis includes the following four aspects:First,for discriminant analysis of the health of the ball screw,an extreme gradient lifting cluster model method based on the feature optimization of a stacked denoising autoencoder is proposed.For monitoring of ball screw condition,vibration,torque,and current signal fusion feature extraction methods are used.Wavelet packet decomposition is used to extract the main features.Stacked denoising autoencoders are used to perform feature encoding.Finally,the state of coded vectors is evaluated by extreme gradient boosting which avoids the difficulties of model generalization caused by the introduction of complex models.Secondly,for the classification of tool wear,few-shot training method is used.Due to the multiple working conditions and small number of collected samples,few-shot meta-learning was used,and the priori of the unknown model on different tasks was used to quickly generalize to the meta-model.The model uses the data of acceleration and force of tool wear to extract features in different domains,which has rapid generalizability,multiple operating conditions adaptability and good robustness.Third,for the prediction of the remaining life of bearing,a multi-scale frozen convolution and activated memory neural network was developed.This thesis proposes to directly use the original multi-channel time-domain signals and use a step-by-step model training strategy for model deployment.First,an adaptive multi-scale convolutional neural network is used to extract features from the original time-domain signal.Then convolutional memory network is connected at the frozen fully connected layer of the first model.Finally,the whole network is fine tuned in the activated process.Finally,by analyzing the monitoring needs of CNC machine tools,a device monitoring and health evaluation system platform for CNC machine tool clusters was proposed and built.The multi-sensor system is used to collect signals from the key equipment of the machine tool,modules of equipment monitoring,data analysis,fault diagnosis,artificial intelligence discrimination,and system management and so on are developed and integrated,mainly based on the browser service system with neural network model embedded.The proposed software provides theoretical,experimental,and system verification for equipment monitoring and health evaluation of CNC machine tools.