Intelligent reasoning method for machine tool machining precision driven by iterative physical embedded graph network

Abstract

The invention discloses a machine tool machining precision intelligent reasoning method driven by an iterative physical embedded graph network, and relates to the technical field of machining precision reasoning, and the method comprises the steps: solving a physical field of each axis of a machine tool through a physical field solving model; then, based on the physical field of each shaft of the machine tool, the thermal-force error of each shaft of the machine tool is determined, finally, a forward transmission model of the thermal-force error of the machine tool is obtained by combining a forward kinematic chain model of the machine tool, and intelligent reasoning of the machining precision of the machine tool is achieved. And when the solving error of the physical field solving model is greater than a preset error threshold value, the iteration unit reserves part of network parameters based on a weight sharing mechanism, and the iteration formula independently learns and finely tunes the model parameters according to real-time data. According to the scheme, heat and stress mechanism equations are introduced to participate in the training process of the model, efficient solving of a physical field with interpretability is achieved, and intelligent reasoning on the machining precision of the machine tool under the dynamic complex working condition is guaranteed based on an iterative updating mechanism.

Description

Technical Field

[0001] The present application relates to the technical field of machining accuracy reasoning, and in particular to an intelligent reasoning method for machine tool machining accuracy driven by an iterative physical embedded graph network. Background Art

[0002] With the rapid development of the manufacturing industry, more stringent requirements have been placed on the machining accuracy of CNC machine tools. Factors such as bearing heating, friction, and unilateral or bilateral fixed constraints on the axes of the machine tool lead to thermal-mechanical deformation of the axes of the machine tool. Statistics show that errors caused by thermal-mechanical factors can account for 40%-70% of the total error of the machine tool, which has a significant impact on the machining accuracy of the machine tool. Therefore, accurately predicting the real-time thermal-mechanical error of the machine tool under different working conditions and performing precision reasoning are crucial for adjusting machining parameters and optimizing compensation strategies. It can ensure high precision and stability during the machine tool processing process.

[0003] At present, in order to infer the machining accuracy of machine tools under the action of thermal-mechanical errors, domestic and foreign scholars mainly use two types of methods: mechanism-driven and data-driven. The mechanism-driven method usually takes complex boundary conditions such as bearing heat generation, heat conduction, and mechanical constraints as the calculation basis, and simplifies the analysis of each axis of the machine tool through methods such as concentrated mass method and finite element analysis to solve its deformation field. This type of method fully integrates thermal and mechanical mechanisms and often has high error prediction accuracy under specific working conditions. However, since its calculation accuracy mostly depends on the fineness of grid division, and with the increase of calculation accuracy requirements, the calculation time increases exponentially; at the same time, this type of method has poor adaptability to changes in working conditions and is difficult to predict thermal-mechanical errors in real time in complex and changeable processing environments.

[0004] The data-driven method is mainly based on machine learning and other technologies, using sensor data such as temperature and stress at the finite nodes of each axis of the machine tool as input and thermal-mechanical error as output to establish an error reasoning model. Compared with the mechanism-driven method, the data-driven error reasoning method can make full use of the real-time data of the actual machining process of the machine tool, and usually has better real-time performance and accuracy. However, this type of method generally lacks thermal-mechanical mechanism analysis, which makes the internal working principle between the "input-output" of the data-driven model lack of explainability, and the prediction accuracy depends on the existing training samples. For unknown working conditions that cannot be covered by the training samples, the reasoning effect is difficult to guarantee. Summary of the invention

[0005] The purpose of this application is to provide an iterative physical embedded graph network driven intelligent reasoning method for machine tool processing accuracy, which can realize intelligent reasoning of machine tool processing accuracy under dynamic and complex working conditions.

[0006] To achieve the above objectives, this application provides the following solutions: The present application provides an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy, comprising the following steps: The temperature sensing data and stress sensing data of each axis of the machine tool are obtained; the axes of the machine tool include the main axis and the feed axis.

[0007] Using the pre-trained physical field solution model, the stress field of each axis of the machine tool is solved according to the temperature sensor data and stress sensor data of each axis of the machine tool; the physical field solution model is a model based on the physical embedding graph network; the physical field solution model uses a multi-scale graph attention network to capture the sensor characteristics of each axis of the machine tool at different scales, and through the physical embedding loss function, the real-time temperature sensor data of each axis of the machine tool, the real-time stress sensor data of each axis of the machine tool and the mechanism equation are used to standardize the training process of the physical field solution model.

[0008] Based on the stress field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined separately.

[0009] According to the thermal-mechanical errors of each axis of the machine tool and the forward kinematic chain model of the machine tool, the forward transfer model of the thermal-mechanical error of the machine tool is determined; the forward transfer model of the thermal-mechanical error of the machine tool includes the thermal-mechanical errors of the spindle and each feed axis from the spindle of the machine tool to the tool at the end of the machine tool, realizing intelligent reasoning of the machine tool processing accuracy.

[0010] When the solution error of the physical field solution model is greater than a preset error threshold, the parameters of the physical field solution model are locally iteratively adjusted based on the weight sharing mechanism according to the real-time sensor data through the iteration unit until the solution error of the physical field solution model is less than the preset error threshold; the solution error of the physical field solution model is the root mean square error value between the real-time stress sensor data of any axis of the machine tool and the value at the corresponding node of the stress field of the axis obtained by solving the physical field solution model.

[0011] Optionally, the parameters of the physical field solution model are locally iteratively adjusted based on the weight sharing mechanism according to the real-time sensor data by the iteration unit, specifically including the following steps: Based on the weight sharing mechanism, some network parameters in the physical field solution model are retained, and the network connection relationship between the physical field solution model is reshaped through iterative units.

[0012] Based on real-time temperature sensing data or stress sensing data, the parameters of the physical field solution model are independently learned and fine-tuned to correct the errors of the physical field solution model under specific working conditions.

[0013] Optionally, a pre-trained physics solution model can be obtained through the following process: Acquire the temperature sensing data of each axis of the machine tool and the stress sensing data of each axis of the machine tool.

[0014] Based on the temperature sensing data of each axis of the machine tool, the stress sensing data of each axis of the machine tool and the mechanism equation, the actual stress field of each axis of the machine tool is solved.

[0015] The temperature sensor data of each axis of the machine tool and the stress sensor data of each axis of the machine tool are used as the input of the physical field solution model, and the actual stress field of each axis of the machine tool is used as the target output of the physical field solution model. The physical field solution model is trained to obtain a pre-trained physical field solution model.

[0016] Optionally, when training the physical field solution model, a dual-track traction mechanism of mechanism-data loss that can balance the data loss term and the mechanism loss term is used to train the physical field solution model; the data loss term is the mean square error between the real real-time sensor data and the sensor data at the corresponding node of the solved physical field; the mechanism loss term is the heat conduction equation, the thermal stress equation, the stress-strain equation and the stress transfer equation, which are used to constrain the training process of the physical field solution model. The physical field solution model is trained using the dual-track traction mechanism of mechanism-data loss, specifically: The data loss and mechanism loss are converted through the embedding layer to obtain the data loss vector and mechanism loss vector.

[0017] The dot product attention mechanism is used to calculate the correlation between the data loss vector and the mechanism loss vector, and the correlation score is normalized to obtain the weight of the data loss term and the weight of the mechanism loss term.

[0018] The total loss is calculated by combining the weighted loss of the data loss term, the weighted loss of the mechanism loss term, the residual penalty, and the penalty term for small weights; Dynamic factors are used to adjust the weights of the data loss term and the mechanism loss term according to changes in the total loss, balancing the contribution between the mechanism loss and data loss of the physics field solution model, thereby improving the stability and accuracy of the physics field solution model during training.

[0019] Optionally, based on the temperature sensing data of each axis of the machine tool, the stress sensing data of each axis of the machine tool and the mechanism equation, the actual stress field of each axis of the machine tool is obtained by solving the method, which specifically includes the following steps: Based on the temperature sensing data of each axis of the machine tool and the heat conduction equation, the temperature field of each axis of the machine tool is solved.

[0020] Based on the temperature field and temperature-thermal strain equation of each axis of the machine tool, the thermal strain load of each axis of the machine tool is calculated.

[0021] Based on the thermal strain load of each axis of the machine tool, stress sensing data, stress-strain equation and stress transfer equation, the stress field of each axis of the machine tool is solved.

[0022] Optionally, the multi-scale graph attention network gradually captures the sensing features of the sensing data of each axis of the machine tool at different scales from local to global through multi-layer graph convolution operations that are expanded layer by layer; after each layer of graph convolution operation, the multi-head attention mechanism is used to weightedly adjust the information propagation intensity between sensor nodes, and further strengthen the fusion of sensing features at different scales.

[0023] Optionally, based on the stress field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined respectively, which specifically includes the following steps: Based on the stress field of each axis of the machine tool, the strain field of each axis of the machine tool is determined respectively.

[0024] Based on the strain field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined respectively; the thermal-mechanical error includes the linear displacement of each axis of the machine tool and the angular deviation of each axis of the machine tool.

[0025] Optionally, according to the thermal-mechanical errors of each axis and the forward kinematic chain model of the machine tool, a forward transfer model of the thermal-mechanical errors of the machine tool is determined, which specifically includes the following steps: Based on the transmission relationship of each axis of the machine tool, the homogeneous transformation matrix is ​​used to establish the forward kinematic chain model of the machine tool.

[0026] Based on the differential motion theory, the thermal-mechanical errors of each axis of the machine tool are converted into corresponding differential expressions.

[0027] The differential expressions of the thermal-mechanical errors of each axis of the machine tool are integrated with the forward kinematic chain model of the machine tool to obtain the forward transfer model of the thermal-mechanical errors of the machine tool.

[0028] According to the specific embodiments provided in this application, this application discloses the following technical effects: The present application provides an iterative physical embedded graph network driven intelligent reasoning method for machine tool processing accuracy, in which the stress field of each axis of the machine tool is solved by using a pre-trained physical field solution model based on the temperature sensor data and stress sensor data of each axis of the machine tool; the physical field solution model in this step is a model based on the physical embedded graph network, which can use a multi-scale graph attention network to capture the physical field changes of different levels and different ranges of the sensor data of each axis of the machine tool, and in the training process, the physical embedding loss function is used to standardize the training process of the physical field solution model with the real-time temperature sensor data of each axis, the real-time stress sensor data of each axis and the mechanism equation; then, based on the stress field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined respectively, and finally, combined with the forward motion chain model of the machine tool, the forward transfer model of the thermal-mechanical error of the machine tool can be determined to realize the intelligent reasoning of the machine tool processing accuracy; when the solution error of the physical field solution model is greater than the preset error threshold, the iterative unit retains some network parameters based on the weight sharing mechanism, and iteratively learns and fine-tunes the model parameters independently according to the real-time data. The solution provided in this application fully captures the physical field change characteristics of different levels and ranges of sensor data through a multi-scale graph attention network, and introduces thermal and stress mechanism equations to participate in the model training process, thereby achieving efficient and interpretable solution of the physical field. When the solution error is greater than the threshold, the solution effectively overcomes the problem that traditional pre-trained models are difficult to respond to changes in working conditions in real time through online iterative updating methods, thereby achieving intelligent reasoning for machine tool processing accuracy under dynamic and complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0030] Figure 1 A flowchart of an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy provided in one embodiment of the present application.

[0031] Figure 2 A flowchart of a pre-trained physical field solution model obtained in an iterative physical embedded graph network driven intelligent reasoning method for machine tool processing accuracy provided in one embodiment of the present application.

[0032] Figure 3 A flowchart of step B2 in an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy provided in one embodiment of the present application.

[0033] Figure 4A flowchart of step B3 in an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy provided in one embodiment of the present application.

[0034] Figure 5 A flowchart of step A3 in an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy provided in one embodiment of the present application.

[0035] Figure 6 A flowchart of step A4 in an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy provided in one embodiment of the present application.

[0036] Figure 7 A flowchart of step A5 in an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy provided in one embodiment of the present application. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0038] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0039] The embodiment of the present application provides an iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy. In an exemplary embodiment, Figure 1 As shown, the following steps are included: A1. Obtain the temperature sensing data and stress sensing data of each axis of the machine tool; the axes of the machine tool include the main axis and the feed axis. Arrange temperature sensors and stress sensors near each axis of the machine tool to collect the temperature sensing data and stress sensing data in real time during the operation of the machine tool.

[0040] A2. Using the pre-trained physical field solution model, the stress field of each axis of the machine tool is obtained according to the temperature sensor data and stress sensor data of each axis of the machine tool; the physical field solution model is a model based on the physical embedded graph network.

[0041] The physical field solution model constructed in this embodiment uses a multi-scale graph attention network to capture the sensor features of each axis of the machine tool at different scales, and uses the real-time temperature sensor data of each axis of the machine tool, the real-time stress sensor data of each axis of the machine tool and the mechanism equation to standardize the training process of the physical field solution model.

[0042] Specifically, in this physical field solution model, the multi-scale graph attention network gradually captures the sensing features of different scales of the sensing data of each axis of the machine tool from local to global through multi-layer graph convolution operations that are expanded layer by layer; after each layer of graph convolution operation, the multi-head attention mechanism is used to weightedly adjust the information propagation intensity between sensor nodes, and further strengthen the fusion of sensing features at different scales.

[0043] The physical embedding loss function contains two parts: data loss and mechanism loss. The data loss is the mean square error between the real real-time sensor data and the sensor data at the corresponding nodes of the solved physical field; the mechanism loss is the mechanism equations such as the heat conduction equation, thermal stress equation, stress-strain equation and stress transfer equation, which are used to constrain the network training process and ensure that the network solution results conform to the real physical laws.

[0044] As an exemplary embodiment, a pre-trained physical field solution model is obtained through the following process: Figure 2 As shown, the following steps are included: B1. Obtain the temperature sensing data of each axis of the machine tool and the stress sensing data of each axis of the machine tool.

[0045] B2. Based on the temperature sensing data of each axis of the machine tool, the stress sensing data of each axis of the machine tool and the mechanism equation, the actual stress field of each axis of the machine tool is obtained. Figure 3 As shown, step B2 specifically includes the following steps: B21. Based on the temperature sensing data of each axis of the machine tool and the heat conduction equation, the temperature field of each axis of the machine tool is solved.

[0046] B22. Based on the temperature field and temperature-thermal strain equation of each axis of the machine tool, the thermal strain load of each axis of the machine tool is calculated.

[0047] B23, based on the thermal strain load, stress sensing data, stress-strain equation and stress transfer equation of each axis of the machine tool, the stress field of each axis of the machine tool is solved. Then, based on step B2, the stress field training model is solved, thereby realizing the training process of standardizing the physical field solution model through the mechanism equation, real-time temperature sensing data and stress sensing data.

[0048] B3. Using the temperature sensor data of each axis of the machine tool and the stress sensor data of each axis of the machine tool as the input of the physical field solution model, and using the actual stress field of each axis of the machine tool as the target output of the physical field solution model, the physical field solution model is trained to obtain a pre-trained physical field solution model.

[0049] Specifically in this embodiment, when training the physical field solution model, a mechanism-data loss dual-track traction mechanism that can balance the data loss term and the mechanism loss term is used to train the physical field solution model.

[0050] The dual-track traction mechanism of mechanism and data loss is used to train the physical field solution model, such as Figure 4 As shown, the specific process includes the following: B31. Convert the data loss and mechanism loss through the embedding layer to obtain the data loss vector and mechanism loss vector.

[0051] B32. Use the dot product attention mechanism to calculate the correlation between the data loss vector and the mechanism loss vector, and normalize the correlation score to obtain the weight of the data loss term and the weight of the mechanism loss term.

[0052] B33. Combine the weighted loss of the data loss term, the weighted loss of the mechanism loss term, the residual penalty, and the penalty term for small weights to calculate the total loss; B34. Use dynamic factors to adjust the weights of the data loss term and the mechanism loss term according to the change of the total loss. Balance the contribution between the mechanism loss and data loss of the physics field solution model, thereby improving the stability and accuracy of the physics field solution model training.

[0053] A3. Based on the stress field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined. In this embodiment, Figure 5 As shown, step A3 specifically includes the following steps: A31. Based on the stress field of each axis of the machine tool, the strain field of each axis of the machine tool is determined respectively.

[0054] A32. Based on the strain field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined respectively; the thermal-mechanical error includes the linear displacement of each axis of the machine tool and the angular deviation of each axis of the machine tool.

[0055] A4. According to the thermal-mechanical errors of each axis of the machine tool and the forward kinematic chain model of the machine tool, the forward transfer model of the thermal-mechanical errors of the machine tool is determined; the forward transfer model of the thermal-mechanical errors of the machine tool includes the thermal-mechanical errors of the spindle and each feed axis between the spindle of the machine tool and the tool at the end of the machine tool, and realizes the intelligent reasoning of the machining accuracy of the machine tool. In this embodiment, if Figure 6 As shown, step A4 specifically includes the following steps: A41. Based on the transmission relationship between the axes of the machine tool, the homogeneous transformation matrix is ​​used to establish the forward kinematic chain model of the machine tool.

[0056] A42. Based on the differential motion theory, the thermal-mechanical errors of each axis of the machine tool are converted into corresponding differential expressions; such as small displacements, small rotation changes, etc.

[0057] A43. The differential expression of the thermal-mechanical error of each axis of the machine tool is integrated with the forward kinematic chain model of the machine tool to obtain the forward transfer model of the thermal-mechanical error of the machine tool.

[0058] A5. When the solution error of the physical field solution model is greater than a preset error threshold, the parameters of the physical field solution model are locally iteratively adjusted based on the weight sharing mechanism according to the real-time sensor data through the iteration unit until the solution error of the physical field solution model is less than the preset error threshold. The solution error of the physical field solution model is the root mean square error value between the real-time stress sensor data of any axis of the machine tool and the value at the corresponding node of the stress field of the axis obtained by solving the physical field solution model.

[0059] In real-time reasoning, if the solution error of the physical field solution model is greater than a preset error threshold, the parameters of the physical field solution model are locally iteratively adjusted based on the weight sharing mechanism according to the real-time sensor data through the iteration unit, such as Figure 7 As shown, the process specifically includes the following steps: A51. Based on the weight sharing mechanism, some network parameters in the physical field solution model are retained, and the network connection relationship between the physical field solution model and the iteration unit is reshaped. The weight sharing mechanism allows the iteration unit and the original physical field solution model to use the same parameters at different locations or layers.

[0060] A52. Independently learn and fine-tune the parameters of the physical field solution model based on real-time temperature sensor data or stress sensor data to correct the errors of the physical field solution model under specific working conditions.

[0061] When the solution error of the physical field solution model is lower than the preset error threshold, the iterated physical field solution model is used to replace the original physical field solution model, ensuring that the physical field solution model can continuously improve its accuracy through local adjustment and feedback under real-time changing working conditions, while maintaining computing efficiency and training stability, and continuously improving machine tool processing accuracy.

[0062] The above-mentioned scheme provided in the present application uses a pre-trained physical field solution model to solve the stress field of each axis of the machine tool based on the temperature sensor data and stress sensor data of each axis of the machine tool. The physical field solution model in this step can use a multi-scale graph attention network to capture the sensor characteristics of each axis of the machine tool at different scales, and during the training process, use a physical embedding loss function to standardize the training process of the physical field solution model with the real-time temperature sensor data of each axis, the real-time stress sensor data of each axis and the mechanism equation; then, based on the stress field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined respectively, and finally, combined with the forward motion chain model of the machine tool, the forward transfer model of the thermal-mechanical error of the machine tool can be determined to realize intelligent reasoning of the machine tool processing accuracy; and when the solution error of the physical field solution model is greater than the preset error threshold, the iterative unit retains some network parameters based on the weight sharing mechanism, and iteratively learns and fine-tunes the model parameters independently according to the real-time data.

[0063] In the above scheme, this application fully captures the physical field change characteristics of different levels and ranges of sensor data through a multi-scale graph attention network, and introduces thermal and stress mechanism equations to participate in the model training process, thereby achieving efficient and interpretable solution of the physical field. When the solution error is greater than the threshold, through online iterative updating means, it effectively overcomes the problem that traditional pre-trained models are difficult to respond to changes in working conditions in real time, and realizes intelligent reasoning of machine tool processing accuracy under dynamic and complex working conditions.

[0064] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. An iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy, characterized in that: include: Acquire temperature sensing data and stress sensing data of each axis of the machine tool; the axes of the machine tool include the main axis and the feed axis; Using a pre-trained physical field solution model, the stress field of each axis of the machine tool is solved according to the temperature sensor data and stress sensor data of each axis of the machine tool; the physical field solution model is a model based on a physical embedding graph network; the physical field solution model uses a multi-scale graph attention network to capture the sensor features of different scales of each axis of the machine tool, and through a physical embedding loss function, the real-time temperature sensor data of each axis of the machine tool, the real-time stress sensor data of each axis of the machine tool and the mechanism equation are used to standardize the training process of the physical field solution model; Based on the stress field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined respectively; According to the thermal-mechanical errors of each axis of the machine tool and the forward kinematic chain model of the machine tool, a forward transfer model of the thermal-mechanical errors of the machine tool is determined; the forward transfer model of the thermal-mechanical errors of the machine tool includes the thermal-mechanical errors of the spindle and each feed axis between the spindle of the machine tool and the tool at the end of the machine tool, and realizes intelligent reasoning of the machining accuracy of the machine tool; When the solution error of the physical field solution model is greater than a preset error threshold, the parameters of the physical field solution model are locally iteratively adjusted based on a weight sharing mechanism according to real-time sensor data through an iteration unit until the solution error of the physical field solution model is less than the preset error threshold; the solution error of the physical field solution model is the root mean square error value between the real-time stress sensor data of any axis of the machine tool and the values ​​at the corresponding nodes of the stress field of the axis obtained by solving the physical field solution model.

2. The iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy according to claim 1 is characterized in that: The parameters of the physical field solution model are locally iteratively adjusted based on the weight sharing mechanism according to the real-time sensor data by the iteration unit, specifically including: Retaining some network parameters in the physical field solution model based on a weight sharing mechanism, and reshaping the network connection relationship between the physical field solution model and the physical field solution model through an iterative unit; Based on real-time temperature sensing data or stress sensing data, the parameters of the physical field solution model are independently learned and fine-tuned to correct the error of the physical field solution model under specific working conditions.

3. The iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy according to claim 1 is characterized in that: The pre-trained physics field solving model is obtained through the following process: Acquire temperature sensing data of each axis of the machine tool and stress sensing data of each axis of the machine tool; Based on the temperature sensing data of each axis of the machine tool, the stress sensing data of each axis of the machine tool and the mechanism equation, the actual stress field of each axis of the machine tool is solved; The temperature sensor data of each axis of the machine tool and the stress sensor data of each axis of the machine tool are used as the input of the physical field solution model, and the actual stress field of each axis of the machine tool is used as the target output of the physical field solution model. The physical field solution model is trained to obtain a pre-trained physical field solution model.

4. The iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy according to claim 3 is characterized in that: When training the physical field solution model, a mechanism-data loss dual-track traction mechanism capable of balancing data loss terms and mechanism loss terms is used to train the physical field solution model; the data loss term is the mean square error between the real real-time sensor data and the sensor data at the corresponding node of the solved physical field; the mechanism loss term is the heat conduction equation, the thermal stress equation, the stress-strain equation and the stress transfer equation, which are used to constrain the training process of the physical field solution model; the physical field solution model is trained using the mechanism-data loss dual-track traction mechanism, specifically: Converting the data loss and the mechanism loss through an embedding layer to obtain a data loss vector and a mechanism loss vector; The dot product attention mechanism is used to calculate the correlation between the data loss vector and the mechanism loss vector, and the correlation score is normalized to obtain the weight of the data loss term and the weight of the mechanism loss term; The total loss is calculated by combining the weighted loss of the data loss term, the weighted loss of the mechanism loss term, the residual penalty and the penalty term for small weights; A dynamic factor is used to adjust the weight of the data loss term and the weight of the mechanism loss term according to the change of the total loss, so as to balance the contribution between the mechanism loss and the data loss of the physical field solution model, thereby improving the stability and accuracy of the physical field solution model during training.

5. The iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy according to claim 3 is characterized in that: Based on the temperature sensor data of each axis of the machine tool, the stress sensor data of each axis of the machine tool and the mechanism equation, the actual stress field of each axis of the machine tool is solved, including: Based on the temperature sensor data of each axis of the machine tool and the heat conduction equation, the temperature field of each axis of the machine tool is solved; Based on the temperature field and temperature-thermal strain equation of each axis of the machine tool, the thermal strain load of each axis of the machine tool is calculated; Based on the thermal strain load of each axis of the machine tool, stress sensing data, stress-strain equation and stress transfer equation, the stress field of each axis of the machine tool is solved.

6. The iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy according to claim 1 is characterized in that: The multi-scale graph attention network gradually captures the sensor features of different scales of the sensor data of each axis of the machine tool from local to global through a multi-layer graph convolution operation that is expanded layer by layer; After each layer of graph convolution operation, the multi-head attention mechanism is used to weightedly adjust the information propagation intensity between sensor nodes and further strengthen the fusion of sensing features at different scales.

7. The iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy according to claim 1 is characterized in that: Based on the stress field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined separately, including: Based on the stress field of each axis of the machine tool, the strain field of each axis of the machine tool is determined respectively; Based on the strain field of each axis of the machine tool, the thermal-mechanical error of each axis of the machine tool is determined respectively; the thermal-mechanical error includes the linear displacement of each axis of the machine tool and the angular deviation of each axis of the machine tool.

8. The iterative physical embedded graph network driven intelligent reasoning method for machine tool machining accuracy according to claim 1 is characterized in that: According to the thermal-mechanical errors of each axis of the machine tool and the forward kinematic chain model of the machine tool, the forward transfer model of the thermal-mechanical error of the machine tool is determined, which specifically includes: Based on the transmission relationship of each axis of the machine tool, the forward kinematic chain model of the machine tool is established using the homogeneous transformation matrix; Based on the differential motion theory, the thermal-mechanical errors of each axis of the machine tool are converted into corresponding differential expressions; The differential expressions of the thermal-mechanical errors of each axis of the machine tool are integrated with the forward kinematic chain model of the machine tool to obtain the forward transfer model of the thermal-mechanical errors of the machine tool.

Images