A cause mechanism independent estimation based causal cross-domain learning method

Abstract

A cause-effect cross-domain learning method based on cause mechanism independence estimation, characterized in that it comprises: for data-driven prediction, decision cross-domain learning problem, establishing a causal diagram to determine the cause-effect relationship between variables; considering the correlation between the distribution of cause variables and the mechanism of cause-effect, establishing an estimation method of cause mechanism independence based on the structural integral of the product of the distribution of cause variables and the derivative of the mechanism; on this basis, a cross-domain learning model fine-tuning method based on the independence estimation value is constructed, forming a cause-effect cross-domain learning modeling method based on cause mechanism independence estimation. The cause-effect cross-domain learning modeling method has the advantages of high model precision, good stability and small target domain data requirement, and is suitable for cross-domain learning scenes such as simulation-actual processing deformation prediction.

Description

Technical Field

[0001] This invention relates to a numerical control manufacturing technology, and more particularly to a data-driven prediction and decision-making cross-domain learning field. Specifically, it is a causal cross-domain learning method that considers the correlation between the distribution of causal variables and the mechanism by which causes lead to results. Background Technology

[0002] With the development of sensing and information technology, data-driven methods have become an important means of solving various production and daily life problems. Cross-domain learning is a major research hotspot in data-driven methods. By addressing the differences in data distribution across different domains, it enables data-driven models to apply the knowledge learned in one domain to other domains, thereby expanding the applicability of the models.

[0003] Taking manufacturing as an example, cross-domain learning can be used to address the problem of insufficient data in manufacturing scenarios. Existing cross-domain learning applications in manufacturing primarily rely on data-driven methods. These methods achieve cross-domain learning by learning the statistical correlations between variables, and are mainly suitable for scenarios where all variables are observable. When unobservable variables introduce bias during data generation, the input-output mapping may not be bijective, thus affecting the accuracy of the cross-domain learning model. For instance, in the typical cross-domain prediction problem of machining deformation in manufacturing, since residual stress cannot be directly monitored during processing, deformation can only be predicted cross-domain by monitoring deformation forces and other physical quantities that can be used to perceive the deformation state. However, residual stress, an unobservable variable, is highly susceptible to introducing bias during data generation, affecting prediction accuracy.

[0004] Despite the existence of unobservable variables, the causal relationships between variables are clear and invariant across domains. By integrating causal knowledge into cross-domain learning models, the influence of introduced biases can be effectively reduced. Compared with purely data-driven methods, existing causal cross-domain learning methods have achieved better results in handling classification problems. These methods all follow the assumption of causal-mechanism independence, meaning that changing the distribution of causes does not affect the corresponding mechanisms, but this assumption is not strictly valid. Taking force-induced material deformation as an example, when the magnitude of the force is within the material's yield limit, the mechanism of force-induced material deformation is elastic deformation. When the distribution of force magnitude is changed, exceeding the material's yield limit, the mechanism of force-induced material deformation will transform into plastic deformation, meaning that the distribution of force magnitude and the mechanism of force-induced material deformation are not independent. If the independence of cause and mechanism is assumed in cross-domain learning, the accuracy of causal cross-domain learning models will be limited, making it difficult to meet the accuracy requirements of regression problems such as cross-domain deformation prediction.

[0005] To this end, this patent considers the correlation between cause and mechanism, breaks the assumption of independence between cause and mechanism, and proposes a new causal cross-domain learning modeling method based on the estimation of the independence of cause and mechanism. Summary of the Invention

[0006] To address the above problems, this invention proposes a causal cross-domain learning modeling method based on the estimation of causal mechanism independence. A structural integral is established between the variable distribution of the cause and its mechanism derivative. Based on this, a method for quantifying causal mechanism independence is developed, and a fine-tuning method for the cross-domain learning model based on the independence estimate is constructed, thus forming a causal cross-domain learning method based on the estimation of causal mechanism independence.

[0007] The technical solution of this invention is:

[0008] A causal cross-domain learning method based on cause-mechanism independence estimation includes: establishing a causal graph for the cross-domain learning problem; establishing a causal cross-domain learning model based on this graph; pre-training the model using source domain data; measuring the independence of cause and mechanism using the structural integral ST of the cause distribution P(c) and the mechanism derivative h′(c) of the cause; obtaining the cause-mechanism independence estimate estimation(c); and then fine-tuning the model using target domain data based on the cause-mechanism independence estimate estimation(c) to achieve causal cross-domain learning based on cause-mechanism independence estimation.

[0009] Furthermore, the method for calculating the cause-mechanism independence estimate, estimation(c), is as follows:

[0010] estimation(c) = |ST s -ST t |

[0011] Where c is the causal variable, ST s To calculate the structural integral value in the target domain using source domain data, ST T This is the structural integral value in the target domain obtained by approximating the target domain data.

[0012] Furthermore, the structural integral ST of the causal distribution P(c) and the causal mechanism derivative h′(c) is calculated as follows:

[0013] ST=∫h'(c)P(c)dc

[0014] Using the reference density U of C C Characterize P(c), and fit h′(c) to h′(c) using source domain data. s The integral of the structured integral over the target domain is calculated as follows:

[0015] ST s =∫h'(c) s U C DC

[0016] Where the reference density UC It is a probability density function of the causal variable C, with reference density U. C Preferably, the distribution is Gaussian or uniform, h′(c) s It is based on fitting the source domain data through a preset function form, preferably a polynomial function; the integral value of the approximate structure integral in the target domain using the target domain data is:

[0017]

[0018] Where m is the data volume of the target domain, e i Let c be the value of the i-th variable E, where E is the result variable of cause C. i Let C be the value of the i-th variable.

[0019] Furthermore, the model is fine-tuned using the target domain data based on the cause-mechanism independence estimate (c). This is achieved by adjusting the learning rate of the cross-domain learning model during the fine-tuning phase. The calculation method for adjusting the learning rate is as follows:

[0020] lr ft (c)=α·estimation(c)·lr pt

[0021] lr ft and lr pt α represents the learning rate during the model fine-tuning stage and the pre-training stage, respectively, and α is the optimization coefficient.

[0022] The beneficial effects of this invention are:

[0023] 1. This invention considers the correlation between cause and mechanism, breaks the assumption of independence between cause and mechanism, and establishes a method for quantifying the independence of cause and mechanism, providing a more reliable basis for decision-making.

[0024] 2. Considering the structural integral of the distribution of causal variables and the mechanism derivative, this invention has strong stability against changes in data distribution and is suitable for diverse and complex data scenarios.

[0025] 3. Compared with traditional methods, this invention introduces the estimation of causal mechanism independence, which improves the accuracy of cross-domain learning models. Attached Figure Description

[0026] Figure 1 This is an illustration of the effect of the application of the present invention. Figure 1 In the causal graphical model, X1, X2, ..., X m Let X1, ..., X be the causal variable for Y. When performing cross-domain learning, the causal relationships between the variables in the graph remain unchanged. Assume X1, ..., X... n The corresponding mechanisms f1,…,f nChanges and X n+1 ,…,X m The corresponding mechanism f n+1 ,…,f m Unchanged. When this cross-domain learning problem is not solved using the method of this invention, the mechanism f1,…,f remains unchanged. m All will be updated (represented as) However, this method can introduce significant errors when solving for Y. In contrast, the method employed in this invention estimates the causes X1,…,X… m With mechanism f1,…,f m To optimize the model by maintaining independence between domains, the mechanism f1,…,f changes only due to cross-domain issues. n It will be updated (represented as) The cross-domain invariant mechanism f n+1 ,…,f m It will not be updated (it will still be represented as f). n+1 ,…,f m This ensures the accurate solution for Y.

[0027] Figure 2 This diagram illustrates the problem of solving the deformation during the machining of aircraft structural components, as exemplified by this invention. In the diagram, C represents residual stress, D represents the domain, F represents deformation force, G represents part geometry, and Y represents deformation amount.

[0028] Figure 3 This is a neural network structure diagram of an example of the present invention, wherein g C The mechanism by which residual stress C affects deformation Y is g. G This is the mechanism by which the part geometry G affects the deformation amount Y. Detailed Implementation

[0029] 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.

[0030] like Figure 1-3 As shown.

[0031] Let's take aircraft structural components as an example to illustrate the problem of controlling machining deformation. Aircraft structural components are a typical problem in the manufacturing field, and they are quite representative. Machining deformation is a key quality indicator in the manufacturing process of aircraft structural components.

[0032] A causal cross-domain learning method based on causal mechanism independence estimation includes the following steps:

[0033] 1. For cross-domain learning problems involving processing deformation, where input state variables C and G are used to output target variable Y, a cause-effect graph needs to be drawn, as shown below. Figure 2 As shown. The deformation force F and the part geometry G are known quantities, the residual stress C and the domain D are latent variables, and the deformation Y is the variable to be solved.

[0034] 2. Based on the deformation force F and part geometry G from the source domain data, a characterization of residual stress C is established. Based on the pre-trained model using the source domain data, the mechanism g of the effect of residual stress C on deformation Y is obtained. C The mechanism by which the geometry G of the part affects the deformation Y G .

[0035] 3. Following the approach of estimating the independence of the causal mechanism as follows, solve for the residual stress C and its mechanism g. C The independence estimate, estimation(C), part geometry G and its mechanism g G The independence estimate, estimation(G), is obtained by using the structural integral ∫h′(c)P(c)dc of the distribution of the causal variable P(c) and the mechanism derivative of the causal variable h′(c), calculating the integral value of the structural integral in the target domain, approximating the integral value of the structural integral in the target domain using data from the target domain, and then solving for the independence estimate of the causal mechanism.

[0036] 4. Using the obtained estimation(C) and estimation(G), adjust the learning rate of the cross-domain learning model during the fine-tuning phase. Then, based on the adjusted learning rate, fine-tune the cross-domain learning model using target domain data to obtain the fine-tuned mechanism. and And solve for the deformation Y.

[0037] The parts not covered in this invention are the same as or can be implemented using existing technologies.

Claims

1. A causal cross-domain learning method based on causal mechanism independence estimation, comprising the following steps: 1) In the problem of solving machining deformation, it is necessary to draw a cause-effect graph based on the input state variables C and G and the output target variable Y. The deformation force F and the part geometry G are known quantities, the residual stress C and domain D are latent variables, and the deformation amount Y is the variable to be solved. 2) Based on the deformation force F and part geometry G from the source domain data, a characterization of residual stress C is established. Based on the pre-trained model from the source domain data, the mechanism of residual stress C on deformation Y is obtained. The mechanism of the effect of part geometry G on deformation Y ; 3) Following the approach of estimating the independence of causal mechanisms, solve for the residual stress C and its mechanism. Independence estimates Part geometry G and its mechanism Independence estimates Using the distribution of causal variables Mechanism derivative of cause Both structural integrals The integral value of the structure integral in the target domain is calculated, and the integral value of the structure integral in the target domain is approximated using the data in the target domain, thereby solving for the independence estimate of the causal mechanism; 4) Using the required... and The learning rate of the cross-domain learning model is adjusted during the fine-tuning phase. Based on the adjusted learning rate, the cross-domain learning model is then fine-tuned using target domain data to obtain the fine-tuned mechanism. and And solve for the deformation Y; The distribution of causal variables Mechanism derivative of cause Structural integrals of both The calculation method is as follows: use Reference density Characterization Through source domain data Fit to The integral of the structured integral over the target domain is calculated as follows: Reference density It is a probability density function of the causal variable c, and the reference density is... It is either a Gaussian distribution or a uniform distribution. It is based on fitting the source domain data using a pre-defined function form, which is a polynomial function; the integral value of the approximate structure integral in the target domain using the target domain data is: in For the amount of data in the target domain, For the first Variables The value, The result variable is the cause variable c. For the first The value of the causal variable c.

2. The method according to claim 1, characterized in that: The method for establishing the cross-domain learning model is as follows: First, establish a causal graph for the cross-domain learning problem; second, establish a causal cross-domain learning model based on this graph, using source domain data to pre-train the model, and then applying the model according to the causal distribution. Mechanism derivative of cause Structural integrals of both Measure the independence of cause and mechanism, and obtain an estimate of cause-mechanism independence. Finally, based on the cause-mechanism independence estimate... By fine-tuning the model using target domain data, causal cross-domain learning based on the estimation of independence of causal mechanisms can be achieved.

3. The method according to claim 2, characterized in that: The cause-mechanism independence estimate The calculation method is as follows: in As the cause variable, To calculate the structured integral value in the target domain using source domain data, This is the structural integral value in the target domain obtained by approximating the target domain data.

4. The method according to claim 2, characterized in that: The method utilizes target domain data to estimate the cause-mechanism independence value. Fine-tuning the model is achieved by adjusting the learning rate of the cross-domain learning model during the fine-tuning phase. The calculation method for adjusting the learning rate is as follows: in and These are the learning rates for the model fine-tuning stage and the pre-training stage, respectively. To optimize the coefficients.

Images