A method for processing process monitoring signal denoising based on causal backdoor path

Abstract

The application discloses a processing procedure monitoring signal denoising method based on a causal backdoor path, and is characterized in that a structural causal model with noise as a mixed factor is established for nonhomologous noise generated by different noise sources in a monitoring signal; a new backdoor path is introduced through an intermediate variable, and two sets of residual regression models of the monitoring signal are constructed with the intermediate variable as a condition to reconstruct the monitoring signal, so that the nonhomologous noise is eliminated. The application solves the denoising of nonhomologous noise signals and the reconstruction difficulty of time series signals.

Description

Technical Field

[0001] This invention relates to the fields of CNC machining and artificial intelligence, and in particular to a method for noise reduction of monitoring signals for machining processes, specifically a method for noise reduction of monitoring signals for machining processes based on causal backdoor paths. Background Technology

[0002] Noise reduction of monitoring data during the manufacturing process is an important research problem. Due to the constantly changing working conditions in the manufacturing process of aerospace parts, data acquisition is difficult and costly, and the amount of data available for data-driven modeling is limited. Therefore, data quality is particularly important.

[0003] Traditional statistical data denoising methods are limited by the need for prior statistical knowledge of random and system noise. In variable-condition machining processes, the statistical characteristics of noise are unclear and difficult to obtain in advance, making traditional statistical denoising methods unsuitable. Traditional mechanism-based denoising methods are also limited by the complexity of CNC machining, especially in variable-condition machining, where the formation mechanism of measurement noise is highly complex. Mechanism-based modeling methods can only make a series of assumptions, approximations, and simplifications, such as continuity assumptions and linear model simplifications, making it difficult to achieve accurate mechanism modeling and denoising of measurement noise. Data-driven denoising methods struggle to obtain noise labels and are only applicable to simulation environments. Furthermore, due to the inherent characteristics of data-driven methods, modeling errors exist, especially in variable-condition machining processes where data does not satisfy independent and identically distributed rules. These modeling errors are even greater, severely impacting the denoising effect of machining process data. Introducing causal knowledge to remove the influence of unobservable confounding factors is an effective means of data denoising. Existing half-sibling regression methods based on causal reasoning utilize a given causal structure to introduce data affected by the same noise source, removing the influence of unobservable noise confounding factors in the system, and achieving denoising in prediction tasks without modeling the noise. However, during processing, noise changes due to variations in operating conditions, and noise has the characteristic of being non-homogeneous. That is, observable information is affected not only by the same noise source but also by other different noise sources. Existing half-sibling regression methods based on causal reasoning are difficult to apply to the denoising problem of non-homogeneous noise.

[0004] Data denoising in processing falls under the category of causal variable reconstruction in causal reasoning research. It is a fundamental and common problem widely applied in causal reasoning fields such as image recognition and data denoising. Causal variable reconstruction refers to deriving the causal variable from the consequent variable produced by the causal variable when the causal variable is unobservable. The denoising problem of non-homogeneous data during processing belongs to a class of causal variable reconstruction problems with multiple unobservable nodes. The difficulty and challenge of this type of problem lies in the difficulty of obtaining information about these unobservable nodes, i.e., the difficulty in obtaining information about independent noise sources, making it difficult to remove the noise generated by these independent noise sources. This invention proposes a denoising method for processing monitoring signals based on a causal backdoor path. By introducing a new backdoor path through intermediate variables of the processing conditions, conditions for causal intervention are provided, thereby achieving denoising of the processing monitoring signals. Summary of the Invention

[0005] The purpose of this invention is to address the problem of noise reduction for non-homogeneous noise data in the processing process. It proposes a noise reduction method for processing monitoring signals based on causal backdoor paths. The method constructs a structural causal model with noise as a confounding factor, introduces a new backdoor path through intermediate variables, and then constructs a residual regression model of two sets of monitoring signals with the intermediate variables as conditions to reconstruct the monitoring signals, thereby achieving the elimination of non-homogeneous noise.

[0006] The technical solution of this invention is:

[0007] A method for denoising process monitoring signals based on causal backdoor paths, characterized in that: two sets of monitoring signals X acquired by the same sensor are denoised... t and Y t The noise is divided into two parts: homogeneous and heterogeneous. X t and Y t The corresponding actual measurand is R. t and Q t For non-homogeneous noise N X and N Y A structural causal model is established with the same source noise N as the confounding factor, and the intermediate variable C is used to establish the model. t A new backdoor path was introduced, and a residual regression model for two sets of monitoring signals was constructed with intermediate variables as conditions. Thus reconstructing Q t This is to eliminate non-homogeneous noise.

[0008] Furthermore, the method for distinguishing between homogeneous and heterogeneous noise defines noise generated from the same noise source as homogeneous noise, such as noise from a CNC machining system; and defines noise generated from different noise sources as heterogeneous noise, such as noise from different directions of a piezoelectric vibration sensor.

[0009] Furthermore, the aforementioned structural causal model construction method uses a causal graph to represent the real measurand R. t and Q t Monitoring signal X t and Y t Noise from the same source N, noise from different sources N X and N Y The causal relationship between them is expressed by the following formula:

[0010] Y t ∶=f1(Q t )+g1(N)+g2(N Y (1)

[0011] X t ∶=f2(R t )+g3(N)+g4(N X (2)

[0012] In the formula: f1 is Q t For Y t The causal effect, f2 is R t For X t The causal effect, g1 is N for Y t The causal effect, g2 is N Y For Y t The causal effect, g3 is N pairs of X t The causal effect, g4 is N X For X t The causal effect.

[0013] Furthermore, the backdoor path construction method introduces intermediate variables that satisfy the common cause assumption of non-homogeneous noise, such as the working conditions during the processing, to establish information correlation between non-homogeneous noise.

[0014] Furthermore, the method for reconstructing the monitoring signal in non-homogeneous noise uses C t For conditions from X t Y t Extracting non-source noise pairs for Y t The influence of [the virus] is then removed, and C is retained. t The effect of non-source noise is obtained by Q. t The estimation is used to reduce non-source noise.

[0015]

[0016] In the formula This is a regression prediction model.

[0017] Furthermore, the residual regression model construction method uses Y t For input, X tFor output, establish Y t To X t The mapping relationship can be determined by any suitable regression model, with the attractor manifold reconstruction residual regression model being preferred.

[0018] The beneficial effects of this invention are:

[0019] 1. This invention improves the structural causal model of half-sib regression, solving the problem of reconstructing causal variables with multiple unobservable nodes.

[0020] 2. This invention introduces a new backdoor path by using intermediate variables to reconstruct non-same-source noise, thus solving the noise reduction problem of non-same-source noise signals.

[0021] 3. This invention solves the problem of time series signal reconstruction by modeling the residual regression model through attractor manifold reconstruction. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the structural causal model for noise reduction of the monitoring signal in this invention. In the diagram, X and Y represent two sets of monitoring signals from the same sensor, R and Q represent the actual measurands corresponding to the two sets of monitoring signals, and N represents the noise from the same source. X and N Y These represent the non-homogeneous noise of monitoring signals X and Y, respectively.

[0023] Figure 2 This is a schematic diagram of the structural causal model of this invention, using vibration signals as an example. X in the diagram... t and Y t Let X represent the monitoring signals of the vibration sensor in the x and y directions at time t, respectively. t-1 and Y t-1 R represents the monitoring signals of the vibration sensor in the x and y directions at time t-1, respectively. t and Q t R represents the actual measurement in the x and y directions at time t, respectively. t-1 and Q t-1 N represents the actual measurement in the x and y directions at time t-1, respectively. t and N t-1 Let N represent the noise from the same source at time t and time t-1, respectively. X t and N Y t Let N represent the non-homogeneous noise of X and Y at time t, respectively. X t-1 and N Y t-1 Let C represent the non-homogeneous noise of X and Y at time t-1, respectively. t and C t-1These represent the processing conditions at time t and time t-1, respectively. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and examples, but the present invention is not limited to these embodiments.

[0025] like Figure 1-2 As shown.

[0026] A method for denoising process monitoring signals based on causal backdoor paths includes the following steps:

[0027] 1. Taking noise reduction of vibration signals during processing as an example, the noise from the same source is random noise and noise from the CNC machining system, while the noise from different sources is the monitoring signal noise generated by the crystal elements in the x and y directions of the vibration sensor.

[0028] 2. First, a structural causal model is constructed by introducing measurement data at different time points and physical components in different directions at the same processing position. Since the processing conditions simultaneously affect the vibration physical quantity and the noise of the vibration signal in different directions, and the processing conditions are observable, this invention introduces a new backdoor path through the processing conditions. The structural causal model for monitoring signal noise reduction is constructed as follows: Figure 1 As shown in the figure. X and Y represent two sets of monitoring signals from the same sensor, R and Q represent the actual measurands corresponding to the two sets of monitoring signals, and N represents the noise from the same source. X and N Y These represent the non-homogeneous noise of monitoring signals X and Y, respectively.

[0029] 3. This invention uses a nonlinear additive noise model to construct a structural causal model of the monitoring signal, such as... Figure 2 As shown. Where Q t and R t Let Q be the actual vibrations in the y and x directions at time t, respectively, and Q t and R t Influenced only by the vibration of the previous moment and the known processing conditions, it satisfies the causal Markov effect, C t For machining conditions, specifically referring to known cutting parameters such as feed rate and speed at the current moment, Y t and X t Let N be the vibration monitoring signals in the y and x directions measured by the vibration sensor at time t, and be affected by the actual vibration and noise at the current time. t For the random noise of the vibration signal and the noise of the processing system, N X t and N Y t This refers to the noise from different directional components and different channels of the vibration sensor.

[0030] 4. Given C, Q and R are independent of each other. For noise from the same source, X t This includes noise N pairs of Y t Related information regarding the impact. This applies to non-homogeneous noise N. X t and N Y t Because they only have a common cause C t Therefore X t and Y t The noise N comes from non-same source. X t and N Y t The information is contained in C t , attempting to use C t For conditions from X t Y t It can extract N Y t For Y t The influence of [the factor] is then removed to obtain Q. t An idealized estimate of Q is used to reduce non-source noise. t Reconstructed as Y t From X t and C t The residuals after regression are to preserve operating condition C. t Regarding the impact of non-homogeneous noise, even if the noise N contains a temporal structure, it will not affect the reconstruction result:

[0031]

[0032] In the formula This is a regression prediction model.

[0033] 5. Based on the characteristic that multi-directional monitoring signals in the processing process have information correlation in the time and spatial domains, this invention uses attractors containing spatiotemporal information in the monitoring signals of different directions, and employs attractor manifold reconstruction to model the residual regression model. For a dynamic system with a certain degree of determinism, such as the sensor monitoring signals in the processing process, it can be represented by an attractor manifold, that is, the trajectory of a point in a high-dimensional space as time progresses. In the triaxial vibration monitoring system, the vibration signals in the x, y, and z directions belong to the same dynamic system; therefore, the time series X... t Y t and Z t Sharing a manifold M = (X t ,Y t Z t ), and X t Y t and Z tThe sequence of these three dimensions constitutes the attractor of the three-dimensional space of the vibration monitoring signal, that is, the set or trajectory of points in the space of the vibration monitoring system.

[0034] For an attractor with a vibration signal in a specific direction (such as the x-direction), it is possible to construct an attractor manifold M in that specific direction using the vibration signal at time t and its adjacent times. x =(X t-1 ,X t ,X t+1 ), and X t-1 X t and X t+1 These three-dimensional sequences constitute the shadow attractor of the vibration monitoring signal in the x-direction, and they share an attractor manifold M. x Similarly, the vibration monitoring signal time series Y in the y-direction is used. t Construct an attractor manifold M y Using the vibration monitoring signal time series Z in the z-direction t Construct an attractor manifold M z .

[0035] For an attractor manifold M with triaxial vibration monitoring signals, let its x-direction vibration time series be X(t) = {X(1), ..., X(L)}, where L is the length of the time series. According to Taken's theorem, by selecting an appropriate embedding dimension E and a delay τ, the original manifold space can be reconstructed from the delayed sequence of X(t):

[0036]

[0037] Where M x Given an E*L matrix, in the vibration monitoring signal, taking E=3, then:

[0038] M x = x (t)=(X(t),X(t-τ),X(t-2τ)) (6)

[0039] Similarly, an M-shaped structure can be constructed for vibration monitoring signals in the y-direction. y :

[0040] M y =y(t)=(Y(t),Y(t-τ),Y(t-2τ)) (7)

[0041] For M x Every point on x (t), calculate the distance between any two points, and extract the E+1 nearest neighbors with the smallest distance as the nearest neighbor NN. x(t), this invention uses Manhattan distance (MD) as a suitable distance metric for attracting subspaces, then M x Distance metric matrix for:

[0042]

[0043] Where d is the Manhattan distance Then for point x (t1)∈M x When E=3, there will be four nearest neighbor points:

[0044] NN x (t1)={s1,s2,s3,s4} (9)

[0045] Therefore, the E+1 nearest neighbors of the vibration monitoring signal time series in the x-direction are:

[0046]

[0047] Similarly, M can be constructed. y Distance metric matrix D y and its nearest neighbor NN y (t).

[0048] 6. According to formula (3), at the current time t, it can be obtained through M x The vibration monitoring signal in the y-direction is reconstructed using the nearest spatiotemporal neighborhood, and a regression model is constructed. Since the vibration monitoring signal in the y-direction is a one-dimensional vector, while M x The nearest spatiotemporal neighborhood is a 4-dimensional vector, so a fully connected neural network is used to parameterize the regression model, ultimately achieving noise reduction of the vibration monitoring signal in the y-direction:

[0049]

[0050] The parts not covered in this invention are the same as or can be implemented using existing technologies.

Claims

1. A method for denoising process monitoring signals based on causal backdoor paths, characterized in that: Two sets of monitoring signals X acquired by the same sensor t and Y t The noise is divided into two parts: homogeneous and heterogeneous. X t and Y t The corresponding actual measurand is R. t and Q t ; For non-homogeneous noise N generated by different noise sources in the monitoring signal X and N Y A structural causal model is established with the same source noise N as the confounding factor; through the intermediate variable C t A new backdoor path is introduced, and a residual regression model for two sets of monitoring signals is constructed with intermediate variables as conditions. Thus reconstructing Q t This achieves the elimination of non-homogeneous noise; the structural causal model construction method involves establishing a causal graph to characterize the real measurand R. t and Q t Monitoring signal X t and Y t Noise from the same source N, noise from different sources N X and N Y The causal relationship between them is expressed by the following formula: （1） （2） In the formula: f1 is Q t For Y t The causal effect, f2 is R t For X t The causal effect, g1 is N for Y t The causal effect, g2 is N Y For Y t The causal effect, g3 is N pairs of X t The causal effect, g4 is N X For X t The causal effect; The method for reconstructing the monitoring signal in non-homogeneous noise is based on C. t For conditions from X t Y t Extracting non-source noise pairs for Y t The influence of [the virus] is then removed, and C is retained. t The effect of non-source noise is obtained by Q. t The estimation is used to reduce non-source noise. （3） In the formula This is a regression prediction model.

2. The method for denoising processing monitoring signals based on causal backdoor paths according to claim 1, characterized in that: The backdoor path construction method described above is to introduce intermediate variables that satisfy the common cause assumption of non-same-source noise and establish information correlation between non-same-source noise.

3. The method for denoising processing monitoring signals based on causal backdoor paths according to claim 1, characterized in that: The residual regression model construction method is based on Y t For input, X t For output, establish Y t To X t The mapping relationship was determined, and the selected regression model was the attractor manifold reconstruction residual regression model.

Images