CN-116127289-B - Machining process monitoring signal noise reduction method based on causal back door path

Abstract

A processing process monitoring signal noise reduction method based on a causal back gate path is characterized by establishing a structural causal model taking noise as a confounding factor aiming at non-homologous noise generated by different noise sources in monitoring signals, introducing a new back gate path through an intermediate variable, and constructing residual regression models of two groups of monitoring signals on the condition of the intermediate variable so as to reconstruct the monitoring signals to realize the elimination of the non-homologous noise. The invention solves the difficult problems of noise reduction of non-homologous noise signals and reconstruction of timing signals.

Inventors

• LI YINGGUANG
• LIU CHANGQING
• HUA JIAQI
• LIU XU
• HAO XIAOZHONG

Assignees

• 南京航空航天大学

Dates

Publication Date

20260508

Application Date

20221216

Claims (3)

1. 1. A noise reduction method for processing process monitoring signals based on causal back gate paths is characterized by dividing the noise of two groups of monitoring signals X t and Y t acquired by the same sensor into two homologous and nonhomologous parts, wherein the real to-be-measured values corresponding to X t and Y t are R t and Q t , establishing a structural causal model with homologous noise N as a confounding factor aiming at nonhomologous noise N X and N Y generated by different noise sources in the monitoring signals, introducing a new back gate path through an intermediate variable C t , and establishing a residual regression model of the two groups of monitoring signals on the condition of the intermediate variable The structural causal model construction method is characterized in that causal relations among real to-be-measured R t and Q t , monitoring signals X t and Y t , homologous noise N and non-homologous noise N X and N Y are represented by the following formulas: (1) (2) Wherein f 1 is the causal effect of Q t on Y t , f 2 is the causal effect of R t on X t , g 1 is the causal effect of N on Y t , g 2 is the causal effect of N Y on Y t , g 3 is the causal effect of N on X t , and g 4 is the causal effect of N X on X t ; The reconstruction method of the monitoring signal in the non-homologous noise comprises the steps of predicting Y t from X t under the condition of C t , extracting the influence of the non-homologous noise on Y t , removing the influence, and reserving the influence of C t on the non-homologous noise to obtain the estimation of Q t so as to achieve the purpose of reducing the non-homologous noise: (3) In the middle of Is a regression prediction model.
2. 2. The method for reducing noise of machining process monitoring signals based on a causal back door path according to claim 1, wherein the back door path construction method is characterized in that intermediate variables meeting non-homologous noise common-cause assumption are introduced, and information association between non-homologous noises is established.
3. 3. The method for reducing noise of machining process monitoring signals based on a causal backdoor path according to claim 1, wherein the method for constructing a residual regression model is characterized in that Y t is taken as input, X t is taken as output, a mapping relation from Y t to X t is established, and the selected regression model is an attractor manifold reconstruction residual regression model.

Description

Machining process monitoring signal noise reduction method based on causal back door path Technical Field The invention relates to the field of numerical control machining and artificial intelligence, in particular to a machining process-oriented monitoring signal noise reduction method, and specifically relates to a machining process monitoring signal noise reduction method based on a causal backdoor path. Background Noise reduction of monitoring data in the processing process is an important research problem, and the quality of the data is particularly important because the working condition of the aerospace part processing process is changed continuously, the data acquisition difficulty is high, the cost is high, and the data volume capable of being used for data driving modeling is small. The traditional statistical-based data denoising method is limited in that random noise and system noise statistics priori knowledge is needed, the statistical characteristics of the noise are not clear and are difficult to acquire in advance in the variable working condition machining process, so that the traditional statistical-based denoising method is difficult to be suitable for the traditional statistical-based denoising method, the numerical control machining is a complex process, particularly the variable working condition machining process, the formation mechanism of measurement noise is very complex, the mechanism modeling method can only carry out a series of assumptions, approximations and simplifications, such as continuity assumptions, linear simplification of a model and the like, so that accurate mechanism modeling and denoising of the measurement noise are difficult to realize, the data-driven denoising method is difficult to acquire noise labels, the method is only suitable for a simulation environment, modeling errors exist due to the inherent characteristics of the data-driven method, and the modeling errors of the data-driven method are larger in the variable working condition machining process of which data does not meet independent identical distribution, so that the denoising effect of the data in the machining process is seriously affected. The influence of unobservable noise clutter factors is removed by introducing causal knowledge, which is an effective means for realizing data noise reduction, and the existing causal reasoning-based half-sibling regression method utilizes a given causal structure to introduce data influenced by the same noise source, remove the influence of unobservable noise clutter factors in a system, and realize noise reduction in a prediction task under the condition of not modeling noise. However, in the processing process, noise has the characteristic of non-homology due to the change of noise caused by the change of working conditions, namely, observable information is influenced by different noise sources besides the same noise source, and the existing causal reasoning-based semi-sibling regression method is difficult to be suitable for the problem of noise reduction of non-homologous noise. The data noise reduction in the processing process belongs to the reconstruction problem of the cause variable in the causal reasoning research, is a common basic problem, and widely exists in the related application fields of causal reasoning such as image recognition, data noise reduction and the like. Reconstruction of a causal variable refers to deriving the causal variable from the resulting variable produced by the causal variable in the event that the causal variable is not observable. The problem of noise reduction of non-homologous data in the processing process belongs to the problem of reconstruction of the causal variables with a plurality of unobservable nodes, and the difficulty and challenge of the problem are that the information of the unobservable nodes is difficult to obtain, namely the information of independent noise sources is difficult to obtain, so that the noise generated by the independent noise sources is difficult to remove. The invention provides a processing process monitoring signal noise reduction method based on a causal back door path, which introduces a new back door path through a processing working condition intermediate variable to provide conditions for causal intervention, thereby realizing noise reduction of the processing process monitoring signal. Disclosure of Invention The invention aims at solving the problem of non-homologous noise data noise reduction in the processing process, provides a processing process monitoring signal noise reduction method based on a causal back door path, builds a structural causal model by taking noise as a confounding factor, and introducing a new back gate path through the intermediate variable, and further constructing a residual regression model of two groups of monitoring signals on the condition of the intermediate variable so as to reconstruct the monitoring signals, thereby realizing the eliminat