Research And Application In CNC System Fault Diagnosis Technology Based On Nonparametric Active Learning

Nowadays,with the remarkable development of computer science and technol-ogy,especially the vigorous development of artificial intelligence,big data and other key technologies,more and more institutions begin to study the fault diagnosis tech-nology of CNC system based on artificial intelligence automatic diagnosis.At the same time,there are more and more equipment in the large-scale CNC production line,and its physical and mechanical structure is more and more complex,which is difficult to detect manually in real time,which puts forward a new adjustment to the maintenance work The traditional mode of "on-time maintenance" and "post-accident maintenance" has been far from meeting the requirements of modern large-scale production.This further highlights that once high-risk faults occur suddenly in the production process,they will not only greatly affect the production equipment and the established efficiency,but also pose a certain threat to the life safety of on-site employees.This intrinsically requires us to develop fault diagnosis methods of CNC system based on AI technology.Real-time detection method based on AI can work out suitable maintenance plan in advance,real-time detection system risk,not only can significantly reduce the influence of fault on production,also can be targeted to maintenance of equipment,thereby greatly improve the effective service life of the equipment and produce the whole cycle life expectancy of production safety,further savings can prevent the overhead for the enterprise.There-fore,how to efficiently realize the fault prediction of CNC system,so as to provide the corresponding guarantee for a series of follow-up work,this is our key research problemBut all the above methods need to manually annotate and analyze all the data gen-erated by the CNC system.In fact,in large-scale projects,the amount of data produced by numerical control system is huge,and the vast majority of cases are the normal op-eration of the data,which has produced a large number of redundant data key problem At the same time,the storage and tagging of all the data also poses a great challenge to the practicability of the system.Obviously,in the high precision CNC system with high real-time requirements,finding representative data analysis and eliminating the in-terference of redundant data are the inherent requirements of fault diagnosis technology for the future high precision CNC system with rapid deployment of small amount of data to improve real-time performance of the system,reduce the cost of data annotation and reduce the storage capacity of redundant dataBased on this,we systematically analyze the redundancy,data representativeness and data selection bottleneck problems of large-scale industrial data sets.This paper systematically studies the theoretical methods for selecting typical representatives from large-scale data.Based on the ideas of active learning,Dirichlet stochastic process,computational geometry and Monte Carlo approximation,we firstly propose a novel non-parametric active learning(NAL)method and prove that NAL acquisition function is a tightly upper-bound of naive form.We validate our method on TCN(Temporal Convo-lutional Network)and achieve the state of the art performance on CWRU benchmark,providing mighty data effectiveness enhancement on industrial fieldThe specific contributions of this paper are as follows:1,NAL method,which is a non-parametric active learning method,can be from the mass of unmarked data to pick out the most valuable data for marking,greatly improve the efficiency of data,significantly reduce the deployment time caused by marking,analysis of data2.This paper theoretically proved that NAL is the tightly upper bound of the naive form of BALD,and it is applicable to the latest neural network architecture,including fully connected neural network,convolutional neural network,recurrent neural network,full convolutional neural network and other AI architectures.This is because NAL does not rely on architecture-specific decisions,only on feedback from the AI architecture3.The best effect with the minimum data consumption and the highest accuracy has been achieved in CWRU Benchmark,the industry benchmark database about fault diagnosis.This provides a powerful data-efficient approach to the industry field.