Research On Adaptive Monitoring Of Tool Wear Based On Deep Learning

With the continuous advancement of science and technology and the continuous advancement of Industry 4.0,emerging technologies are gradually applied to manufacturing and related industries,and intelligent manufacturing has entered a booming era.As an important component of the machine tool in the intelligent manufacturing workshop,the working state of the tool directly affects the accuracy and quality of the entire processing device,and even affects the state of the machine tool where it is located.Traditional tool condition monitoring relies on rich manual experience and a large amount of runtime data,resulting in obvious bottlenecks in its overall monitoring cost and intelligence.Therefore,monitoring the tool effectively is an important research topic.The focus of this thesis is on the wear state of the tool during the tool processing.The current data of the rotary axis and the precession axis of the tool spindle are used as monitoring indicators to build a high-performance data acquisition platform to collect real-time operation data in the actual processing and production process.According to the data analysis results,the tool status is judged.The main work of this thesis is as follows:(1)A self-adaptive extraction scheme of machining stage data is proposed,and the tool state is classified by feature extraction combined with traditional classifiers.Aiming at the problem that the data cannot be directly interpolated due to the lack of timestamp information,the yolov3 network and dynamic time rounding algorithm are used to find out the mapping relationship,and interpolate according to the mapping relationship.In the actual production process,the data is missing due to network fluctuations and other reasons,which affects the accuracy of data extraction.Under the condition of timestamp information,this paper performs different interpolation operations on the data and selects the optimal interpolation method; at the same time,this paper further considers the scenario where the data collection process lacks timestamp information and cannot directly interpolate the data.In order to solve this scenario,first use the serial number of the sequence as the x-axis coordinate,and the corresponding current load data value as the y-axis coordinate,use a two-dimensional graph to represent the time series data,and combine the yolov3 network in the target detection to make candidate frames in the data select.Then,after the candidate frame is determined,the dynamic time rounding algorithm is used to find the mapping relationship between the candidate frame and the template,and the sequence to be interpolated is generated according to the mapping relationship,and then the data in the candidate frame is interpolated using the linear interpolation method.Finally,the time-domain feature analysis is performed on the extracted data,and the tool wear status is analyzed in combination with the classification algorithm in traditional machine learning.Aiming at the problem of insufficient expression of sample features,this paper integrates the wide&deep architecture and deep residual shrinkage network,so that the network model can comprehensively consider the high-latitude features and lowlatitude features of the sample data in the feature extraction stage,and further strengthen the memory of the model ability and anti-interference ability.Aiming at the problem of small data sample size,on the basis of integrating the wide&deep architecture and deep residual shrinkage network,the deep residual shrinkage twin network based on wide&deep,the deep residual shrinkage prototype network and the deep residual shrinkage network are further designed.Poor contraction of relational networks.In order to avoid the single small sample model from falling into the local optimum dilemma in complex calculations,the above models are fused and analyzed to improve the accuracy of the tool wear state monitoring model.