Research On Tool Wear State Recognition And Prediction Based On Data Enhancement Strategy

Developments in the internet technology,internet of things,cloud computing,big data,and artificial intelligence in recent years have brought significant advancements in manufacturing,among which manufacturing digitalization contributes greatly to the everincreasing productivity.As the fundamental device used for execution in smart manufacturing systems,the intelligent machine tool plays a significant role in the overall production system.The cutting tool is the tooth of the intelligent machine tool and it directly affects workpiece surface quality as well as machining tool performance.Real-time and accurate assessment of tool wear status can not only reduce production costs,but also effectively improve machining tool utilization rates.Aiming at accelerating the pace of equipment intelligence to tool intelligence,the main research focus of this thesis as follows.Firstly,by the mechanism of tool wear,the concrete form of tool wear,the process of tool wear and the standard for tool wear blunt are described.In addition,the tool wear test scheme for turning and the tool wear test scheme for milling are introduced in detail,including the equipment,instruments and acquisition methods adopted.The tool failure index and machining parameters are also confirmed,and the original signal collected have been analyzed briefly.Afterwards,three signal features for tool wear monitoring are analyzed in detail: the time-domain features,frequency-domain features,and wavelet-domain features.According to the respective characteristics of vibration signal,spindle current signal and cutting force signal,the original acquisition signals are processed and extracted from TD,FD and TFD.Based on the feature analysis in time and frequency domain and wavelet packet analysis,a wavelet scale image extraction method with high temporal and frequency resolutions is proposed.Next,Morlet scale diagram of wavelet with high time and frequency resolutions are selected as the input features,deep transfer convolutional neural network model is proposed to research the classification performance of tool wear labels for small sample data sets and enhancement data sets.Furthermore,the classification performance of tool wear condition based on image feature fusion of multisensory signal in turning and milling is also studied,The results show that the deep transfer learning model based on Alexnet and Vgg NET-16 is a better choice to realize the two-dimensional time-frequency image feature classification task.The multisensory signal image feature fusion can effectively improve the prediction accuracy of tool wear condition and is helpful for deep combination and mining of different tool state information.Lastly,the model of combining stacked automatic coding network(SAEN)and generative adversarial network(GAN)is presented,which is used to investigate the classification performance on tool wear condition of generated sample data in handling small samples and unbalanced label samples.The research results show that the proposed model can generate the data which is like the real data collected by the sensor,which effectively improve the overall prediction accuracy of the unbalance label samples,and meet the requirements of prediction and classification of tool wear state.Taking the sample data generated by the proposed model as input,the PSO-LSSVM model of tool wear value is used for regression prediction.The results show that the evaluation results of the generated sample data are almost close to the real sample data,the effectiveness of the proposed model in predicting tool wear is verified again,it is regarded as a new idea for tool wear on-line monitoring.