Tagging Of High Frequency Time Series Signals In Machining Process Based On Sequence Labeling Model

A large number of high-frequency time sequence signals are generated in the process of machining,which contains rich mechanical equipment status information,and can be used in the fields of mechanical equipment status monitoring,fault diagnosis and predictive maintenance.However,unlabeled signals have low value density and are difficult to be used effectively.In actual analysis,signals need to be intercepted and tagged according to different scenes.At present,it is manually tagged and intercepted,which is time-consuming and costly.In order to solve the problems of low value density of high-frequency time series signal and difficulty in automatic tagging,this paper proposes a new method based on hidden semi Markov model(HSMM)and bi-directional long short term memory(BiLSTM)+ conditional random field(CRF)in different scenarios.In this paper,the spindle vibration signal of milling process is taken as the research object to realize the tagging and interception of high-frequency time sequence signal,so as to reduce the labeling cost,improve the tagging automation,and to provide data support for related scenes.According to the spindle state and state transition,the signal label set is defined,the milling experiments of simple parameter and complex variable parameter processing scene are designed,and the spindle vibration data acquisition platform is built.In the simple parameter processing scene,according to the characteristics of the signal,the Mel cepstrum coefficient and gamma tone cepstrum coefficient features of the spindle vibration signal are extracted,and linear discriminant analysis(LDA)is used to reduce the dimension of features.The signal automatic tagging model based on HSMM is established,which the state durations can be explicitly modeled.The effectiveness of the HSMM is finally verified in the simple parameter processing scene.In the complex parameter processing scene,the short-time Fourier transform(STFT)of spindle vibration signal is extracted as the low-level feature,and the dependence of time series data is captured by BiLSTM to realize the deep feature extraction of time series signal.CRF is introduced to explicitly model the dependence between labels,and the classification probability of boundary frame is made to the nth power to balance the class imbalance between boundary frame and non boundary frame The over segmentation problem was relieved,and the final labeling F1 score reached 91.99%.This paper proposes an accurate location method of signal boundary based on matrix profile.Short time Fourier transform and wavelet synchronous compression transform are used to analyze the signal in time-frequency domain.A distance calculation method based on time-frequency feature is defined to reform the matrix profile,which can effectively reduce the average location error of signal correlation boundary to within 0.11 s.The validity of the signal labeling model based on HSMM in the single parameter scenario,the signal marking model based on BiLSTM-CRF in the complex parameter scenario and the accurate boundary location method based on matrix profile are proved by the practical application in the tool residual life experiment and Machine Tool Dynamics Research scenario.