Intelligent servo control method of glass edge grinding machine guided by causal priori

Abstract

The invention discloses a causal priori guided intelligent servo control method for a glass edge grinding machine, and relates to the technical field of servo control, the causal relationship between variables is learned from historical observation data of a servo control system of the glass edge grinding machine, a causal graph is constructed, a weighted adjacency matrix A is obtained, a neural network is used as a PID regulator, and a weighted adjacency matrix B is obtained; in the training process of the neural network, the weighted adjacency matrix A is used as priori knowledge, a matrix B capable of quantifying the function dependency relationship between input variables in the network is defined in the neural network, a causal regularization loss item is introduced, so that the matrix B approaches the matrix A, and the neural network is constrained to learn a mapping relationship consistent with a real causal mechanism. According to the method, false correlation caused by hybrid variables can be effectively avoided, and the generalization ability and robustness of a control strategy under variable working conditions are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of servo control technology, and in particular to a causal prior-guided intelligent servo control method for glass edging machines. Background Technology

[0002] In the glass edging industry, high-precision edging quality is crucial in determining product grade. Early glass edging machine servo control systems typically employed a fixed-parameter PID (Proportional-Integral-Derivative) control algorithm. This control method has relatively simple logic: the system presets an ideal target position or pressure value, and the servo drive, through feedback encoder signals, strives to make the motor's actual output (such as torque and speed) track this target quickly and accurately. Under ideal conditions of stable operation and constant load, this control method can meet basic requirements.

[0003] However, in actual industrial production environments, this traditional intelligent servo control method reveals its inherent limitations: it is highly susceptible to interference from various confounding factors, leading to persistent deviations in control accuracy. In the intelligent servo control of glass edging machines, the input information for the deep learning network mainly includes: the current deviation, the trend of deviation changes, the workload of the grinding head, the vibration of the equipment, and process settings and environmental parameters. However, it is precisely this input information, which should make control more precise, that is entangled by various confounding factors in actual operation, causing the neural network to struggle to make correct judgments.

[0004] First, multiple pieces of information interfere with each other, confusing causal relationships. A single source of the problem can simultaneously alter the behavior of all input information. For example, as the grinding wheel gradually wears down and becomes dull, the motor load will naturally increase to polish the glass, and the equipment's vibration characteristics will also change. The rotation speed may therefore fall behind the command, resulting in deviations. The neural network will learn the pattern that "increased load" requires "increased output." However, the problem is that when processing glass with slightly higher hardness, the motor load will also increase. From the input information alone, the control system has difficulty distinguishing whether the "increased load" is caused by the dulling of the grinding wheel or the hardening of the glass. If a misjudgment is made, such as mistakenly believing that the grinding wheel is dull and forcibly increasing the rotation speed to polish the hard glass, it can easily lead to the glass being chipped or developing defects.

[0005] Secondly, the measurement signals themselves are not pure, and the true state is masked by noise. The data returned by the sensors is not an ideal signal. The motor current readings may contain subtle fluctuations caused by the switching of electronic components; vibration signals are more easily interfered with by the operation of other machine parts (such as guide rails and conveyor belts). More insidiously, temperature changes may cause a slow "drift" in the sensor readings themselves, causing the measured "load" value to deviate from the true situation. If the neural network learns these noises and interferences as useful information, it will lead to hesitant control commands, frequent jitter, or subtle deviations that are difficult to eliminate.

[0006] Finally, the entire system is constantly changing, lacking a fixed pattern. The advantage of neural networks is learning patterns from historical data. However, the working environment of an edge grinding machine is non-static. The grinding wheel itself is a consumable that constantly changes from new to old, meaning the operating characteristics of the entire device are slowly altered every day. A perfect model trained last month with a new grinding wheel will, this month, face a worn-out old grinding wheel, and the "normal range" of all input information it receives will have changed. Its past learning may no longer be applicable, leading to decreased control precision and exhibiting a "getting less intelligent with use" phenomenon.

[0007] Therefore, the core bottleneck currently facing intelligent servo control can be summarized as a problem of causal identification. Due to the presence of multiple dynamic confounding factors (such as grinding wheel wear and glass material fluctuations), the observed multi-source input signals (such as load current and vibration spectrum) and the output variables to be controlled (such as grinding quality) exhibit a highly nonlinear correlation rather than a clear causal mechanism. If the control system only learns these superficial statistical correlations and cannot identify the root cause of changes in the driving process (for example, it cannot distinguish whether the increase in load is due to grinding wheel wear or changes in glass hardness), its decision-making will lose robustness. Therefore, the key to overcoming this bottleneck lies in developing new control algorithms that can identify and decouple key causal factors from strongly disturbed observation data, enabling the system to adaptively adjust according to the true causes of grinding wheel state and workpiece properties, thereby achieving a leap from correlation-driven to causal-driven control. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this application provides a causal prior-guided intelligent servo control method for glass edging machines, which solves the problem of insufficient adaptive capability of traditional PID controllers in glass edging machine servo control systems caused by mixed factors such as grinding wheel wear and changes in glass material.

[0009] To achieve the above objectives, this application adopts the following technical solution, including: A causal prior-guided intelligent servo control method for glass edging machines. By learning the causal relationships between variables from historical observation data of the servo control system of a glass edging machine, a causal graph is constructed, and the corresponding weighted adjacency matrix is ​​obtained. A Matrix elements Used to measure variables For variables The strength of direct causality; A neural network is used as a PID controller to predict the PID control parameters of a glass edging machine servo control system. The input to the neural network is the value of each variable, and the output is the PID control parameters. The neural network is trained using historical observation data from the glass edging machine servo control system. During the training process, a weighted adjacency matrix is ​​used... A As prior knowledge, a matrix is ​​defined in neural networks to quantify the functional dependencies between input variables within the network. B Matrix elements Used to measure variables in the internal representation of a neural network For variables The directional influence strength of information encoding; introducing a causal regularization loss term, making matrix elements With matrix elements Semantic alignment constrains the mapping relationship between neural network learning and the real causal mechanism.

[0010] Preferably, a causal discovery algorithm is used to learn the causal relationships between variables from the historical observation data of the servo control system of the glass edging machine and construct a causal graph.

[0011] Preferred, variable set , representing feed rate, motor load torque, grinding head vibration displacement, grinding head speed, and speed tracking error, respectively. The weighted adjacency matrix is: .

[0012] Preferably, matrix elements The calculation method is as follows: Let global features The activation value of a certain intermediate layer in a neural network is derived from global features. Extract the specific representation associated with each variable; introduce an importance weight vector for each variable. Weight vector Each element in the vector represents the importance of the neuron to the relevant information of the variable; weight vector Estimation can be performed using gradients, meaning the final output of the neural network represents the global features. The gradient of, therefore the gradient of, Neural representation of each variable for: ; in, This indicates element-wise multiplication; Matrix elements Defined as a variable Small perturbations on variables Sensitivity to representation in feature space: ; in, Indicates the distribution of data; The larger the value, the more significant the variable in the feature space of the network. Changes in variables The more significant the change.

[0013] Preferred, causal regularization loss term for: ; in, Represents the parameters in the neural network; the causal regularization loss term penalizes any deviation between the internal functional dependencies of the neural network and the true causal structure.

[0014] Preferably, the regression loss is calculated using the mean squared error, and the regression loss term... for: ; in, It is the size of the training batch. It is the first n Vector labels of the optimal PID control parameters for each sample.

[0015] Preferably, a differentiable objective function for neural network training : ; in, This is the proportionality coefficient.

[0016] This application also provides a readable storage medium, characterized in that it stores a computer program thereon, which, when executed, implements the aforementioned causal prior-guided intelligent servo control method for glass edging machines.

[0017] This application also provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned causal prior-guided intelligent servo control method for a glass edging machine.

[0018] This application also provides a computer program product comprising a computer program / instruction that, when executed by a processor, implements the aforementioned causal prior-guided intelligent servo control method for a glass edging machine.

[0019] The advantages of this application are: (1) To address the problem of insufficient adaptive capability of traditional PID controllers in the servo control system of glass edging machines due to confounding factors such as grinding wheel wear and changes in glass material, this application proposes a neural network PID parameter adjustment method based on causal regularization. This method utilizes a causal discovery algorithm to extract the causal structure between variables from historical system operating data (feed speed, motor load torque, grinding head vibration displacement, grinding head speed, and speed tracking error), encoding it as a causal adjacency matrix as prior knowledge. By constructing a neural influence matrix representing the functional dependencies between input variables within the neural network, and introducing a causal regularization loss term during neural network training, the mapping relationship between network learning and the true causal mechanism is constrained. Theoretical and experimental results show that this method can effectively avoid spurious correlations caused by confounding variables and significantly improve the generalization ability and robustness of the control strategy under varying operating conditions.

[0020] (2) This application embeds causal relationships as prior knowledge into the neural network, so that when the neural network learns the PID parameter adjustment mapping, it not only fits the correlation of data, but also follows the real causal mechanism between variables. This approach combines causal inference and deep learning, aiming to improve the robustness, interpretability and generalization ability of the system (especially when the system is disturbed or the data distribution changes). Attached Figure Description

[0021] Figure 1 This is a flowchart of a causal prior-guided intelligent servo control method for a glass edging machine according to this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] like Figure 1 As shown in the figure, the specific content of the causal prior-guided intelligent servo control method for glass edging machines of this application is as follows: Causal priors refer to prior knowledge about the causal relationships between variables within a system, existing before the machine learning model is built. It goes beyond simply knowing the variables. and variables It is relevant, but more importantly, it is clear. yes The "cause" yes The "effect" refers to the direction of causal flow. This prior knowledge is a structured constraint that reflects an understanding of the underlying operating mechanism of the system. A causal graph is the most intuitive and mathematical tool for representing causal priors. It is a directed graph where nodes represent system variables and directed edges (arrows) represent causal directions. For example, an arrow pointing from "grinding wheel wear" to "motor load" indicates that "grinding wheel wear is the cause of changes in motor load." Therefore, a causal graph can perfectly encapsulate and express our causal prior knowledge for a specific domain.

[0024] There are two main approaches to constructing causal graphs: those based on expert knowledge and those based on causal discovery algorithms. Expert knowledge-based methods involve domain experts manually drawing the graphs based on physical principles, mechanical structures, and long-term experience. Causal discovery algorithm-based methods utilize specialized algorithms to automatically learn the causal structure between variables from the system's historical observation data. In complex industrial processes like glass edging machines, combining causal discovery algorithms is particularly suitable. The grinding process involves the coupling of multiple physical fields, including mechanical, electrical, and fluid dynamics, resulting in intricate relationships between variables. It is difficult for experts to accurately depict all the causal paths based solely on experience. Furthermore, modern intelligent edging machines are typically equipped with a wealth of sensors that continuously record large amounts of operational data, providing a foundation for data-driven causal discovery.

[0025] Therefore, this application utilizes a causal discovery algorithm to extract the causal structure between variables from the system's historical operating data (feed speed, motor load torque, grinding head vibration displacement, grinding head speed, and speed tracking error), encoding it as a causal adjacency matrix as prior knowledge. By constructing a neural influence matrix representing the functional dependencies between input variables within the neural network, and introducing a causal regularization loss term during neural network training, the mapping relationship between the network's learning and the real causal mechanism is constrained.

[0026] Define the set of observed variables in historical data as , representing feed rate, motor load torque, grinding head vibration displacement, grinding head speed, and speed tracking error, respectively. The PID parameter tuning task aims to learn a mapping. , making It can minimize the overall performance index of the system. The causal prior of the system is represented by a weighted directed acyclic graph. This indicates that its corresponding weighted adjacency matrix is Matrix elements Quantified variables For variables The direct causal strength. This matrix can be learned from historical data using a score-based causal discovery algorithm.

[0027] To embed causal priors into neural networks, this application defines a matrix that can quantify the functional dependencies between input variables within the network, i.e., the matrix... Its core idea is: matrix elements The input variable should be measured in the internal representation of the neural network. For input variables The strength of the directional influence of information encoding. This definition is intended to be related to the causal adjacency matrix. (variable For variables The true causal effect is semantically aligned. Let This represents the activation value of a certain intermediate layer in the network. (This is for calculation...) First, from global features We extract the specific representation associated with each input variable. We then introduce an importance weight vector for each variable. Each element represents the importance of a neuron to information related to a variable. The weight vector can be estimated using the gradient, i.e., the weights of the final output of the neural network. The gradient of . Therefore, the gradient of the first . The neural representation of each variable is as follows: ; in, This indicates element-wise multiplication. (Matrix) elements Defined as input Small perturbations on variables Sensitivity to representation in feature space: ; in, It refers to the distribution of data. The larger the value, the higher the probability of occurrence of certain features in the network's feature space. Changes will lead to Significant changes have occurred. This suggests that the network has learned... right The dependency relationship reflects the causal relationship. .matrix A The semantics of derive from the concept of intervention in causal inference. Quantifying the effect of variables in an ideal randomized controlled experiment External intervention ( When ), the variable Changes in expected value reflect the causal mechanism between variables. (Matrix) B The semantics are designed as A Analog quantities in neural networks. Quantization is performed on the input within the neural network. When making minor interventions, representatives within the network Information feature representation The degree of change. If a neural network is a good estimate of the real data generation process, then the feature relationships within it should be able to capture the causal structure of the external world. Therefore, forced B Approaching A Essentially, it constrains the internal representation space of the neural network to the direction of the real causal graph.

[0028] Therefore, the causal regularization loss term can be expressed as: ; in, This represents the parameters in the neural network. This loss term penalizes the functional dependencies within the neural network. With true causal structure Any deviation between them.

[0029] Regression loss can be represented by mean squared error: ; in, It is the size of the training batch. It is the first n Vector labels of the optimal PID control parameters for each sample.

[0030] Combining the above two points, we obtain the differentiable final objective function for training the neural network: ; in, This is the proportionality coefficient.

[0031] The above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims

1. A method for causality-priori-guided intelligent servo control of a glass edger, characterized in that, Learn the causal relationship between variables from the historical observation data of the servo control system of the glass edging machine, construct a causal graph, and obtain the corresponding weighted adjacency matrix A , matrix elements for measuring variables direct causal strength of variables ; The neural network is used as a PID regulator to predict PID control parameters of a glass edging machine servo control system, the neural network has values of each variable as inputs and PID control parameters as outputs, and the neural network is trained by using historical observation data of the glass edging machine servo control system; in the training process of the neural network, a weighted adjacency matrix is added to the neural network A As prior knowledge, a matrix capable of quantifying functional dependency relationships between internal input variables in the neural network is defined in the neural network B The matrix elements are used to measure the directional influence intensity of the variable on the information coding of the variable A causal regularization loss term is introduced, so that the matrix elements are semantically aligned with the matrix elements , and the mapping relationship learned by the neural network is consistent with the real causal mechanism.

2. The intelligent servo control method of a cause and effect prior guided glass edger according to claim 1, wherein, a causality discovery algorithm is used to learn the causal relationships between variables from historical observation data of a servo control system of the glass edger, and a causal graph is constructed.

3. The intelligent servo control method of a cause and effect prior guided glass edger according to claim 1, wherein, Variable set , respectively, representing the feed speed, motor load torque, grinding head vibration displacement, grinding head speed, speed tracking error, and the weighted adjacency matrix is .

4. The intelligent servo control method of a cause and effect prior guided glass edger according to claim 1, wherein, Matrix elements are calculated as follows: Let the global feature be separated from the global feature for the activation value of a certain intermediate layer of the neural network Introduce an importance weight vector for each variable Each element in the weight vector represents the importance of the neuron to the variable-related information; the weight vector can be estimated by the gradient, i.e., the gradient of the final output of the neural network to the global feature , so the neural representation of the first variable is: wherein denotes element-wise multiplication; matrix elements defined as, a small perturbation to the variable sensitivity of the representation in feature space to the variable defined as, a small perturbation to the variable wherein, represents a distribution of data; The larger the value, the more significant changes in the variable will result from changes in the variable in the feature space in the network.

5. The causally-prior guided intelligent servo control method for a glass edger as claimed in claim 4, wherein, causal regularization loss term is: wherein, denote parameters in the neural network; the causal regularization loss term penalizes any deviation between the internal functional dependencies of the neural network and the true causal structure.

6. The causally-prior guided intelligent servo control method for a glass edger as claimed in claim 5, wherein, The regression loss is calculated using mean square error, and the regression loss term is is: wherein, is the size of the training batch, is the vector label of the optimal PID control parameters corresponding to the n th sample.

7. The causally-prior guided intelligent servo control method for a glass edger as claimed in claim 6, wherein, Differentiable objective function for neural network training : wherein is a proportionality factor.

8. A readable storage medium, characterized by, a computer program is stored thereon, and the computer program, when executed, implements the method for causality-priori-guided intelligent servo control of a glass edger according to any one of claims 1-7.

9. An electronic device, comprising: it comprises a processor, a memory, and a computer program stored on the memory and executable on the processor, and the processor, when executing the computer program, implements the method for causality-priori-guided intelligent servo control of a glass edger according to any one of claims 1-7.

10. A computer program product, characterised in that, it comprises computer programs / instructions, and the computer programs / instructions, when executed by a processor, implement the method for causality-priori-guided intelligent servo control of a glass edger according to any one of claims 1-7.

Images