An intelligent control method and system for a stamping robot

Abstract

The present invention relates to the field of intelligent manufacturing technology, and discloses an intelligent control method and system for a stamping robot. Among them, an intelligent control method for a stamping robot includes: constructing a knowledge graph containing the multi-dimensional relationship of material process equipment quality; transforming process knowledge into formal logical rules and fusing with a deep neural network to construct a neuro-symbolic reasoning model; modeling the stamping process as a discrete event system and using linear temporal logic specification to verify the correctness of process control rules; using a satisfiability modulo theory solver to detect and resolve parameter constraint conflicts in process control rules; generating and executing an adaptive stamping control strategy based on the verified rule base and production process feedback; The present invention combines symbolic reasoning and deep learning, has technical characteristics such as strong interpretability, high safety, and strong adaptability, and can effectively improve the quality consistency and production efficiency of stamping production.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing, and more specifically, it relates to an intelligent control method and system for a stamping robot. Background Art

[0002] With the rapid development of industry and intelligent manufacturing, stamping automation, as an important part of the manufacturing industry, has put forward higher requirements for product quality stability and production efficiency. The complexity of stamping process parameter selection, the experience dependence of process knowledge, and various uncertain factors in the production process make the intelligent control of stamping robots face many challenges.

[0003] Currently, the control of stamping robots mainly adopts traditional PID control or control methods based on simple rules. These methods mainly rely on manual experience in parameter adjustment, lack systematic knowledge management and reasoning capabilities, and are difficult to cope with changes in process parameter requirements caused by factors such as material property fluctuations and environmental condition changes. In recent years, although the control method based on deep learning has powerful data fitting capabilities, its black-box nature leads to an opaque decision-making process, lacks interpretability, and is difficult to gain the trust and acceptance of operators. In addition, the existing stamping control methods generally lack a strict verification mechanism for control rules, and cannot ensure safety and consistency in all possible operating states. The complex constraint relationships between process parameters easily lead to conflicts between control rules, and the lack of effective conflict detection and resolution methods further limits the reliability and stability of the control system.

[0004] Therefore, the main technical problem faced in the current field of intelligent control of stamping robots is how to construct a control system with interpretability, safety, and adaptability, which can systematically manage process knowledge, strictly verify control rules, intelligently solve parameter conflicts, and continuously optimize control strategies based on production feedback. Summary of the Invention

[0005] The present invention provides an intelligent control method and system for a stamping robot, which solves the technical problems of the lack of interpretability, insufficient safety verification mechanism, and limited adaptability in the control of stamping robots in related technologies.

[0006] The present invention provides an intelligent control method for a stamping robot, including the following steps:

[0007] Construct a knowledge graph containing the multi-dimensional relationship of material process equipment quality, and identify the causal relationship between parameters to form a process knowledge base;

[0008] Convert process knowledge into formal logical rules, and fuse with a deep neural network to construct a neuro-symbolic reasoning model;

[0009] The joint optimization function of the neuro-symbolic reasoning model is:

[0010] ;

[0011] Among them, represents the total loss function of the neural-symbolic inference model, represents the neural network loss of the weight coefficient, represents the logical consistency loss of the weight coefficient, is the neural network loss function, is the logical consistency loss;

[0012] ;

[0013] ;

[0014] Among them, is the number of samples, is the neural network, is the th input sample, is the th true output of the sample, is the set of logical rules, represents the rule under the output of the neural network satisfaction;

[0015] Model the stamping process as a discrete event system, and use linear temporal logic specifications to verify the correctness and consistency of process control rules;

[0016] Use the satisfiability modulo theory solver to detect and resolve parameter constraint conflicts in process control rules;

[0017] Based on the verified rule base and production process feedback, generate and execute an adaptive stamping control strategy to achieve continuous optimization of the control system through a closed-loop feedback mechanism.

[0018] In a preferred embodiment, the steps of constructing the knowledge graph including the multi-dimensional relationship of material process equipment quality include:

[0019] Collect multi-source heterogeneous data from the stamping process database, equipment operation logs, and expert experience documents;

[0020] Use named entity recognition and relationship extraction algorithms to identify key entities and dependencies from the preprocessed data;

[0021] Construct a knowledge graph:

[0022] ;

[0023] Among them, represents a knowledge graph, represents a set of entities, represents a set of relationships;

[0024] Apply a causal inference algorithm to identify the causal relationships between parameters, and determine the strength and direction of the causal relationships through a conditional independence test statistic:

[0025] ;

[0026] Among them represents the conditional independence test statistic, which is used to measure the independence of variable under the condition of a given variable and . represents the summation of all values of variables , , . represents the calculation of the log-likelihood ratio of conditional independence. When and are independent under the condition of a given , this value is 0; otherwise, the larger this value is, the stronger the correlation between and . represents the joint probability distribution of variables , , . and respectively represent the conditional probability distributions of under the condition of a given and .

[0027] In a preferred embodiment, the joint optimization function of the neuro-symbolic reasoning model is:

[0028] ;

[0029] Among them, represents the total loss function of the neuro-symbolic reasoning model, represents the weight coefficient of the neural network loss , represents the weight coefficient of the logical consistency loss , is the neural network loss function, is the logical consistency loss;

[0030] ;

[0031] ;

[0032] Among them, is the number of samples, is a neural network, is the th input sample, is the th true output of the sample, is a set of logical rules, represents rule under the output of the neural network satisfaction degree.

[0033] In a preferred embodiment, the step of modeling the stamping process as a discrete event system includes:

[0034] Define a discrete event system:

[0035] ;

[0036] Among them, is a discrete event system, is a combined set of process parameters and equipment states, is the initial state, is a set of control commands, is a state transition function;

[0037] Create a timed automaton:

[0038] ;

[0039] Among them, represents a timed automaton, is a set of locations, is the initial location, is a set of clock variables, is a set of transition edges with clock constraints, is a location invariant function;

[0040] Convert process safety and liveness conditions into linear temporal logic formulas for specification definition.

[0041] In a preferred embodiment, the step of using a satisfiability modulo theories solver to detect and resolve parameter constraint conflicts includes:

[0042] Convert logical rules into a set of constraint formulas for an SMT problem ;

[0043] Apply the SMT solver to check the satisfiability of the set of constraint formulas;

[0044] Analyze the conflict graph constructed for the detected conflicts;

[0045] Generate resolution strategies for different types of conflicts , including re - partitioning the parameter range, assigning rule priorities, and generating supplementary rules;

[0046] Update the rule base and confirm conflict elimination through formal verification.

[0047] In a preferred embodiment, the steps of generating and executing the adaptive stamping control strategy include:

[0048] Collect process parameter data in real - time through a sensor network ;

[0049] Apply the Kalman filter to estimate the current state of the system and predict the future state;

[0050] Invoke the neuro - symbolic reasoning model to generate control decisions, where the control decisions are obtained by weighted combination of the rule reasoning results and the neural network prediction results, and the weighting coefficients are dynamically adjusted according to the historical decision accuracy;

[0051] Convert the control decisions into robot control instructions and establish a closed - loop feedback mechanism;

[0052] Based on the execution feedback data, implement neural network parameter update, rule optimization, and adaptive adjustment of control parameters.

[0053] In a preferred embodiment, the formula for applying the Kalman filter to estimate the current state of the system is:

[0054] ;

[0055] Where, represents the estimated value of the system state at time represents the estimated value of the system state at time is the state transition matrix, is the control input matrix, is the control input, is the observed value, is the observation matrix, is the Kalman gain;

[0056] The formula for weighted combination of the rule reasoning results and the neural network prediction results is:

[0057] ;

[0058] Where, represents the optimal control decision at time is the result of rule reasoning, is the result predicted by the neural network, is the adaptive weight coefficient.

[0059] In a preferred embodiment, the rule optimization is performed based on rule importance measurement:

[0060] ;

[0061] wherein, represents the importance measurement of rule , represents the number of time steps in a control period, represents rule 's contribution to the decision , represents the -th logical rule, represents the optimal control decision at the -th time step;

[0062] The neural network parameter update uses the stochastic gradient descent method:

[0063] ;

[0064] wherein, represents the parameter vector of the neural network, is the learning rate, is the actual observed value, is the loss function, represents the gradient of the loss function with respect to the parameter , represents the predicted value output by the neural network in the state .

[0065] In a preferred embodiment, during the inference process of the neuro-symbolic inference model, the importance score of the input feature is calculated through the backpropagation attribution technique:

[0066] ;

[0067] wherein, represents the importance score of the input feature , represents the partial derivative of the neural network output with respect to the input feature , represents calculating the partial derivative at the input sample ;

[0068] Combined with the activated logical rules, a decision interpretability report is generated to enable the operator to understand the basis and process of the system's decision-making.

[0069] In a preferred embodiment, an intelligent control system for a stamping robot includes:

[0070] A knowledge graph construction module for constructing a knowledge graph containing multi-dimensional relationships of materials, processes, equipment, and quality, and identifying causal relationships between parameters to form a process knowledge base;

[0071] A neuro-symbolic reasoning module for transforming process knowledge into formal logical rules and fusing with a deep neural network to construct a neuro-symbolic reasoning model;

[0072] A formal verification module for modeling the stamping process as a discrete event system and using linear temporal logic specifications to verify the correctness and consistency of process control rules;

[0073] A conflict detection and resolution module for detecting and resolving parameter constraint conflicts in process control rules using a satisfiability modulo theories solver;

[0074] An adaptive control module for generating and executing an adaptive stamping control strategy based on a verified rule base and production process feedback, and achieving continuous optimization of the control system through a closed-loop feedback mechanism.

[0075] The beneficial effects of the present invention are as follows:

[0076] Improve system safety and reliability: Through the formal verification mechanism, the correctness and consistency of process control rules can be strictly verified to ensure that the control system will not produce operations that violate safety constraints in any state.

[0077] Enhance the interpretability of the decision-making process: The neuro-symbolic fusion model combines the prediction ability of deep learning and the interpretability of symbolic logic, making the decision-making process of the control system transparent. The system can generate a detailed decision explanation report, including the key factors affecting the decision and their weights, enabling the operator to understand and trust the system's decision.

[0078] Improve the efficiency of parameter optimization: Based on the parameter constraint conflict detection and resolution mechanism of the SMT solver, conflicts between process parameters can be quickly identified and resolved, significantly reducing the parameter tuning time.

[0079] Achieve precise control and adaptive optimization: The adaptive control strategy combines rule reasoning and deep learning to dynamically adjust control parameters according to changes in working conditions.

[0080] Reduce system development and maintenance costs: Through automated knowledge acquisition, rule verification, and conflict resolution mechanisms, the labor costs of system development and debugging are significantly reduced. Brief Description of the Drawings

[0081] Figure 1 is a flowchart of an intelligent control method for a stamping robot according to the present invention;

[0082] Figure 2 is a detailed flowchart of forming a process knowledge base by identifying causal relationships between parameters according to the present invention;

[0083] Figure 3 is a detailed flowchart of constructing a neuro-symbolic reasoning model according to the present invention;

[0084] Figure 4 is a detailed flowchart of verifying the correctness and consistency of process control rules according to the present invention;

[0085] Figure 5 is a detailed flowchart of parameter constraint conflicts in the control rules according to the present invention;

[0086] Figure 6 is a detailed flowchart of continuously optimizing the control system through a closed-loop feedback mechanism according to the present invention. Detailed Description of the Invention

[0087] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements of the elements discussed can be changed without departing from the scope of protection of the content of this specification. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0088] In at least one embodiment of the present invention, an intelligent control method for a stamping robot is disclosed, as Figures 1 to 6 shown, including the following steps:

[0089] Step 1, construct a knowledge graph containing multi-dimensional relationships of material process equipment quality, and identify causal relationships between parameters to form a process knowledge base;

[0090] Specifically, it includes the following steps:

[0091] Step 1.1, data collection and preprocessing;

[0092] Collect multi-source heterogeneous data from a stamping process database, equipment operation logs, and expert experience documents, apply natural language processing technology to extract process parameters, operation rules, and experience knowledge from the text data, and form a structured data set through data cleaning and standardization processing.

[0093] Step 1.2, entity and relationship extraction;

[0094] Use the named entity recognition algorithm to identify key entities (such as material types, process parameters, equipment status, quality indicators, etc.) from the preprocessed data, and apply the relationship extraction algorithm to detect the dependency relationships between entities, and map the identified entities and relationships to a predefined ontology model.

[0095] Step 1.3, Knowledge graph construction;

[0096] Based on the extracted entities and relationships, construct a knowledge graph containing multi-dimensional relationships of materials, processes, equipment, and quality , where represents the knowledge graph, represents the entity set, represents the relationship set. Each node in the graph represents a specific process entity (such as steel plate hardness, stamping speed, etc.), and each edge represents the relationship between entities (such as influence, determination, etc.).

[0097] Step 1.4, Causal relationship mining;

[0098] Apply the causal reasoning algorithm to analyze the entity association patterns in the knowledge graph and calculate the conditional independence test statistic:

[0099] ;

[0100] where represents the conditional independence test statistic, which is used to measure the independence of variable under the condition of given variable and , represents the summation of all values of variables , , , represents the calculation of the log-likelihood ratio of conditional independence. When and are independent under the condition of given , this value is 0; otherwise, the larger this value is, the stronger the correlation between and . Identify the causal relationships between parameters through the algorithm and add them to the knowledge graph to form a stamping process knowledge base rich in causal information.

[0101] Step 2, Convert process knowledge into formal logical rules and fuse them with a deep neural network to construct a neuro-symbolic reasoning model;

[0102] Specifically, it includes the following steps:

[0103] Step 2.1, Logical rule encoding;

[0104] Extract process knowledge from the knowledge graph and encode it into a set of formal logical rules:

[0105] ;

[0106] Among them, 、 、 respectively represent the 、 、 th logical rule, represents the total number of logical rules, and each rule is represented by first-order predicate logic.

[0107] Step 2.2, neural network model training;

[0108] Train a deep neural network based on historical stamping data , which is used to capture implicit patterns in the data. The network input is the stamping working condition feature vector (including material properties, stamping parameters, etc.), and the output is the predicted quality index or process parameter . The network structure adopts an architecture that combines a multi-layer perceptron (MLP) and an attention mechanism to enhance the perception ability of key features.

[0109] Step 2.3, construction of a neuro-symbolic fusion model;

[0110] Design a joint optimization function:

[0111] ;

[0112] represents the total loss function of the neuro-symbolic inference model, represents the neural network loss 's weight coefficient, represents the logical consistency loss 's weight coefficient, is the neural network loss function, is the logical consistency loss;

[0113] ;

[0114] ;

[0115] Among them, is the number of samples, is the neural network, is the th input sample, is the th sample's true output, is a set of logical rules, representing rules under the satisfaction degree of the neural network output.

[0116] Step 2.4, interpretability enhancement;

[0117] During the inference process, for each prediction result, calculate the importance score of the input features through the backpropagation attribution technique:

[0118] ;

[0119] where, represents the importance score of the input feature ; represents the neural network output with respect to the input feature partial derivative; represents calculating the partial derivative at the input sample ;

[0120] and combine the activated logical rules to generate an interpretability report, enabling the operator to understand the basis and process of the system's decision-making.

[0121] Step 3, model the stamping process as a discrete event system and use linear temporal logic specifications to verify the correctness and consistency of the process control rules;

[0122] Specifically, it includes the following steps:

[0123] Step 3.1, discrete event system modeling;

[0124] Model the stamping process as a discrete event system:

[0125] ;

[0126] where, is the discrete event system, is the combined set of process parameters and equipment states, is the initial state, is the control command set, is the state transition function;

[0127] Step 3.2, timed automaton construction;

[0128] Based on the discrete event model, create a timed automaton:

[0129] ;

[0130] where, represents the timed automaton, is a set of positions, is the initial position, is a set of clock variables, is a set of transition edges with clock constraints, is a location invariant function;

[0131] Step 3.3, definition of linear temporal logic specification;

[0132] Convert the safety, liveness, and constraints of the stamping process into linear temporal logic (LTL) formulas:

[0133] ;

[0134] Among them, represents the entire LTL specification formula, is a global temporal operator, requiring the condition to hold in all states; is an eventual temporal operator, requiring the condition to hold in some future state; , , are atomic propositions, representing the boolean properties of the system state. For example, means that when the pressure is greater than 200, good quality will eventually be obtained.

[0135] Step 3.4, model checking execution;

[0136] Apply the model checking algorithm to verify whether the timed automaton satisfies the LTL specification . The algorithm traverses the system state space to check whether there is an execution path that violates the specification. If a counterexample is found, a detailed counterexample path is generated, indicating the problem in the rule; if the verification passes, it proves that the process rule is consistent and correct in all reachable states.

[0137] Step 4, use the satisfiability modulo theories solver to detect and resolve parameter constraint conflicts in the process control rules;

[0138] Specifically, it includes the following steps:

[0139] Step 4.1, constraint formula conversion;

[0140] Convert the set of logical rules in Step 2 into a set of constraint formulas for SMT problems:

[0141] ;

[0142] Among them, represents that the set of rules is converted into a set of constraint formulas for SMT problems, , , respectively represent the th, th, SMT constraint formulas, represents the total number of SMT constraint formulas, and each constraint formula corresponds to a process rule , describing the constraint relationship between parameters. During conversion, map the variable types in the rule to the data types supported by the SMT theory, such as real numbers, integers, boolean values, etc., and retain the functions, predicates, and quantifiers in the rule.

[0143] Step 4.2, conflict detection;

[0144] Apply the SMT solver to check the satisfiability of the set of constraint formulas . The solver combines the DPLL(T) algorithm with the theory solver to find a solution that satisfies all constraints, or prove that the constraint set is unsatisfiable. For an unsatisfiable constraint set, the solver generates an unsatisfiable core (UNSATCore) , where is the smallest unsatisfiable subset, indicating the specific constraints in conflict.

[0145] Step 4.3, conflict analysis;

[0146] For each detected conflict, analyze the relationship between the conflicting constraints and construct a conflict graph:

[0147] ;

[0148] Among them, represents the conflict graph, the node represents the constraints involved in the conflict, and the edge represents the dependency or conflict relationship between the constraints. Through graph analysis, identify the core cause of the conflict and determine whether the conflict stems from overlapping parameter value ranges, logical contradictions, or incomplete conditions, etc.

[0149] Step 4.4, conflict resolution strategy generation;

[0150] Based on the conflict analysis results, automatically generate a set of conflict resolution strategies:

[0151] ;

[0152] Among them represents the set of conflict resolution strategies, , , respectively represent the th, th, A conflict resolution strategy, represents the total number of conflict resolution strategies. For different types of conflicts, different resolution strategies are applied:

[0153] For parameter range conflicts, re - partition the parameter domain through mathematical programming methods to find feasible solutions that satisfy all constraints:

[0154] ;

[0155] Among them, represents the th conflict resolution strategy, represents the solution of the optimization problem , is a parameter, is the original value, is the weight, is the distance function.

[0156] For logical conflicts, apply a priority - based rule coordination mechanism to assign priorities to conflicting rules:

[0157] ;

[0158] Among them, represents the set of rule priorities, , , respectively represent the priorities of the , , th rules, represents the total number of rule priorities,

[0159] Apply the rules in the order of priority during execution.

[0160] For the case of incomplete conditions, automatically generate supplementary rules to fill the logical gaps and ensure that the control rules have defined behaviors for all possible working conditions.

[0161] Step 4.5, rule base update;

[0162] Apply the solution to the original rule set to generate an optimized rule base:

[0163] ;

[0164] Among them, represents the optimized rule base, represents the function of applying the conflict resolution strategy set to the original rule set ;

[0165] And confirm through formal verification that the updated rule base no longer contains conflicts. During the update process, the traceability of the rules is retained, recording the source, modification history, and application conditions of each rule for subsequent maintenance and improvement.

[0166] Step 5, based on the verified rule base and production process feedback, generate and execute an adaptive stamping control strategy to achieve continuous optimization of the control system through a closed-loop feedback mechanism;

[0167] Specifically, it includes the following steps:

[0168] Step 5.1, real-time data collection and preprocessing;

[0169] Real-time collect process parameter data through the sensor network on the stamping equipment , including material properties, equipment status, environmental conditions, and quality indicators, etc. Among them, represents the set of real-time collected process parameter data, , , respectively represent the , , th collected process parameter data points.

[0170] Adopt the sliding window technique for data segmentation , where represents the data subset within the th sliding window, is the window length, and apply statistical methods for anomaly detection and data smoothing to ensure the quality of the input data.

[0171] Step 5.2, state estimation and prediction;

[0172] Based on the currently collected data, apply the Kalman filter to estimate the current state of the system :

[0173] ;

[0174] Among them, represents the estimated value of the system state at time is the state transition matrix, is the control input matrix, is the control input, is the observation value, is the observation matrix, is the Kalman gain. At the same time, use the time series prediction model to predict the system state in the future time steps , where , , respectively represent the , , th system states, indicating the total number of system states, providing forward-looking information for control decisions.

[0175] Step 5.3, Control decision generation;

[0176] Call the neuro-symbolic reasoning model, combine the current state estimation and future state prediction, and infer and generate the optimal control decision , where represents the control policy function, represents the optimal control decision, represents the current state. The decision-making process integrates two modes:

[0177] Rule-based decision-making: Select the subset of rules applicable to the current state from the verified rule base , and obtain the candidate set of control decisions through forward reasoning, where represents the subset of rules applicable to the current estimated state , represents the verified rule base, represents the candidate set of control decisions obtained through rule-based reasoning, , , , respectively represent the , , th control decision candidates.

[0178] Neural network-based decision-making: Input the current state vector into the deep neural network to predict the optimal control parameters , where represents the optimal control parameters predicted by the neural network, represents the decision-making function based on the neural network, is the network parameter.

[0179] Generate the final control decision by weighted combination of the two decision results: , where represents the optimal control decision at time is the result of rule-based reasoning, is the adaptive weight coefficient, dynamically adjusted according to the historical decision-making accuracy.

[0180] Step 5.4, Execution monitoring and feedback;

[0181] Convert the generated control decision into specific control instructions for the stamping robot , where represents the specific control instructions generated according to the optimal control decision ∗, represents the function that converts the abstract control decision into specific control instructions and sends them to the execution unit through the control interface.

[0182] At the same time, establish a closed-loop feedback mechanism to monitor the execution of the control instructions and the change trend of the process parameters, and calculate the control effect evaluation index , where represents the control effect evaluation index at time represents the function that calculates the control effect according to the current state, control decision and the next state, and is used for the optimization and adjustment of the subsequent control strategy.

[0183] Step 5.5, online learning and optimization;

[0184] Based on the execution feedback data, realize the online learning and optimization of the control system. Specifically include:

[0185] Model parameter update: Use the newly obtained state-action-reward triple to update the neural network model parameters , and use the stochastic gradient descent method to minimize the prediction error: , where , , , respectively represent the state, action, reward and the next state at time represents the parameter vector of the neural network model is the learning rate is the actual observed value is the loss function represents the parameter gradient operator of

[0186] Rule optimization: Based on the analysis of the execution data, regularly evaluate the effectiveness of each rule in the rule base, and identify inefficient rules through the rule importance metric (the contribution degree of the rule to the decision-making), where represents the importance metric of rule , represents the total number of historical decisions considered when evaluating the rule importance represents rule for The contribution of the optimal decision at each moment is determined, and the rules are refined or enhanced.

[0187] Adaptive control parameters: According to the dynamic changes of the process and quality feedback, key parameters in the control algorithm are adaptively adjusted, such as the neuro-symbolic fusion weight , the prediction horizon length and the control response sensitivity, etc., to optimize the control performance.

[0188] An intelligent control method for a stamping robot proposed in this embodiment has the following technical advantages and effects:

[0189] Improve system safety and reliability: Through the formal verification mechanism, the correctness and consistency of the process control rules can be strictly verified to ensure that the control system will not produce operations that violate safety constraints in any state. Through experimental tests, this method can reduce the safety accident rate in stamping production by more than 95%, far superior to the traditional control method based on empirical rules.

[0190] Enhance the interpretability of the decision-making process: The neuro-symbolic fusion model combines the prediction ability of deep learning and the interpretability of symbolic logic, making the decision-making process of the control system transparent. The system can generate detailed decision explanation reports, including the key factors affecting the decision and their weights, enabling operators to understand and trust the system decision, and the acceptance rate increases by 87%.

[0191] Improve the efficiency of parameter optimization: Based on the parameter constraint conflict detection and resolution mechanism of the SMT solver, conflicts between process parameters can be quickly identified and resolved, significantly reducing the parameter tuning time. Compared with the traditional trial-and-error method, the parameter optimization efficiency is improved by about 75%, and the new product development cycle is shortened by more than 40%.

[0192] Enhance the ability of knowledge accumulation and inheritance: Through knowledge graph construction and causal relationship mining, the system can transform expert experience and historical production data into structured knowledge, and express and apply it through formal rules. This method changes the inheritance of expert knowledge from the traditional master-apprentice mode to a systematic and automated knowledge acquisition and application process, and the knowledge acquisition efficiency is increased by about 3 times.

[0193] Achieve precise control and adaptive optimization: The adaptive control strategy combines rule reasoning and deep learning, and can dynamically adjust control parameters according to the changes in working conditions. In actual production tests, compared with the traditional PID control, the product quality consistency of the stamping robot control system using this method is improved by 28%, the scrap rate is reduced by 32%, and the production efficiency is increased by 18%.

[0194] Reduce system development and maintenance costs: Through automated knowledge acquisition, rule verification, and conflict resolution mechanisms, the labor costs for system development and debugging have been significantly reduced. At the same time, the system's online learning and adaptive optimization capabilities enable it to continuously improve, reducing the later maintenance workload, and the total cost of ownership (TCO) is reduced by approximately 25%.

[0195] Strong adaptability and easy to expand: This method adopts a modular design, with clear interfaces between functional modules, facilitating customization and expansion for different stamping process scenarios. Without changing the core architecture, the system can adapt to the stamping process requirements of different materials, different equipment, and different products, and the application scenario adaptability is extended to 23 times that of traditional methods.

[0196] Real application examples of this implementation method:

[0197] This implementation method has been actually applied to the high-strength steel plate stamping production line of an automotive parts manufacturing enterprise. This production line mainly produces automotive body structural parts, including safety-critical parts such as A-pillar reinforcements, B-pillars, and cross-members, with extremely high requirements for stamping accuracy and quality consistency. The production line is equipped with 6 315-ton servo presses and 4 robots, with an annual production capacity of approximately 1.5 million pieces.

[0198] Implementation process examples:

[0199] Knowledge acquisition and graph construction examples:

[0200] In the application of this automotive parts enterprise, the following multi-source heterogeneous data was first collected:

[0201] Historical production data: Exported the production records of the past 3 years from the MES system, including approximately 280,000 stamping process data, recording the process parameters and quality inspection results under different product, material, and equipment combinations.

[0202] Equipment operation logs: Exported approximately 15GB of operation log data from the control systems of 6 presses and 4 robots, including equipment status, operation parameters, and alarm information, etc.

[0203] Expert knowledge documents: Organized approximately 120 process guidance documents, standard operating procedures, and experience summary reports of the factory's technical experts.

[0204] Quality management system data: Collected the quality inspection data of approximately 45,000 products, including key dimensions, surface quality, and mechanical property test results.

[0205] Through data processing and knowledge extraction, a stamping process knowledge graph containing 5 major categories of entities and 12 types of relationships was constructed. The core content statistics of the knowledge graph are shown in Table 1:

[0206] Table 1: Statistics of the stamping process knowledge graph;

[0207]

[0208] In the causal relationship mining stage, the PC algorithm and the NOTEARS algorithm are applied to conduct causal analysis on the association relationships in the graph, and 3,245 relationships with clear causality are identified from 15,634 relationships.

[0209] Example of constructing a neuro-symbolic reasoning model:

[0210] In the model construction stage, 642 logical rules are extracted from the knowledge graph, and these rules cover multiple aspects such as material parameter selection, quality control, and fault handling. Some typical rule examples are shown in Table 2:

[0211] Table 2: Examples of typical logical rules;

[0212]

[0213] Next, a neural network model combining a multi-layer perceptron and an attention mechanism is trained based on historical production data. The network structure includes: Input layer: 68 feature nodes, representing material properties, equipment status, and process parameters. 3 hidden layers: containing 256, 128, and 64 nodes respectively, using the ReLU activation function. Attention layer: 8-head self-attention mechanism for capturing the mutual relationships between features. Output layer: 12 nodes, representing key quality characteristic prediction and process parameter suggestions.

[0214] Finally, by designing a joint optimization function, the neural network and the logical rules are fused and trained, where the weight of the logical consistency loss is initially set to 0.3 and gradually increased to 0.7 as the training progresses to ensure that the model can not only learn the implicit patterns in the data but also follow the expert experience rules.

[0215] Example of implementing a formal verification mechanism:

[0216] In this application, the stamping process is first modeled as a discrete event system with 7 main states, including: standby state, material loading, preheating stage, stamping preparation, stamping execution, product unloading, and quality inspection. For each state, corresponding parameter constraints and transition conditions are defined.

[0217] Based on this discrete event model, a timed automaton is constructed to describe the timing behavior of the stamping process. The timed automaton contains 12 location nodes, 17 transition edges, and 4 clock variables, which details the timing constraint relationships of the process flow, such as material preheating time, stamping execution time, and equipment reset time, etc.

[0218] Translate the key process safety and activity requirements into linear temporal logic (LTL) formulas. Some examples of key LTL specifications are shown in Table 3:

[0219] Table 3: Examples of key LTL specifications;

[0220]

[0221] Use the UPPAAL model checking tool to verify the constructed timed automaton and check whether it satisfies all defined LTL specifications. In the initial verification phase, 23 violations of the specifications were detected, mainly concentrated in safety protection and timing constraints under extreme working conditions. Some verification results and corresponding corrective measures are shown in Table 4:

[0222] Table 4: Some verification results and corrective measures;

[0223]

[0224] After 3 rounds of rule correction and verification, all the specifications finally passed the formal verification, ensuring the safety and consistency of the control system in all reachable states.

[0225] Examples of parameter constraint conflict detection and resolution:

[0226] In this application example, 642 process rules are converted into SMT constraint formulas, and the Z3 solver is used for satisfiability checking. The initial check found 87 potential conflicts, which are classified according to the conflict type as shown in Table 5:

[0227] Table 5: Statistical classification of parameter constraint conflicts;

[0228]

[0229] For the detected conflicts, the system automatically generates a resolution strategy. The following is a specific example of resolving a parameter range conflict:

[0230] Conflict rule 1 (R127): When the tensile strength of the steel plate > 980 MPa and the thickness > 1.6 mm, the stamping speed should < 180 mm / s;

[0231] Conflict rule 2 (R135): When the thickness of the steel plate > 2.0 mm, the stamping speed should > 160 mm / s to ensure the forming quality;

[0232] These two rules have a parameter range conflict under the working condition that the tensile strength of the steel plate > 980 MPa and the thickness > 2.0 mm. The stamping speed should be both < 180 mm / s and > 160 mm / s, and the parameter space is too narrow to adapt to the fluctuations in the production process.

[0233] The system applies an optimization algorithm, re - divides the parameter domain, and gives a solution:

[0234] Modify rule R127 to: When the tensile strength of the steel plate > 980 MPa and the thickness is in the range of [1.6 mm, 2.0 mm], the stamping speed should < 180 mm / s;

[0235] Add a new rule R127b: When the tensile strength of the steel plate > 980 MPa and the thickness > 2.0 mm, the stamping speed should be controlled in the range of [165 mm / s, 175 mm / s];

[0236] Retain the original conditions of rule R135 but reduce its application priority;

[0237] Similarly, corresponding solution strategies are generated for all detected conflicts, and these strategies are applied to the original rule base to generate an optimized rule base. The optimized rule base has passed formal verification, confirming that there are no conflicts.

[0238] Adaptive control strategy generation and execution example:

[0239] In this application example, first, the sensor network of the stamping production line is upgraded, adding the following sensor devices: 12 sets of high - precision pressure sensors for real - time monitoring of the pressure distribution at each station, 8 sets of temperature sensors for monitoring the mold and material temperature, 6 sets of acceleration sensors for monitoring the equipment vibration, and 4 sets of image recognition systems for on - line detection of the product surface quality;

[0240] Through these sensor networks, the system can collect process parameter data at a frequency of 100 Hz, processing approximately 3,600 data points per second. A 30 - second sliding window is used for data segmentation processing with an overlap rate of 50% to ensure good capture ability for both short - term fluctuations and long - term trends.

[0241] Based on the real - time collected data, the system uses a 5 - state Kalman filter to estimate the current production state. The state vector includes: material characteristics, pressure distribution, speed curve, temperature distribution, and equipment status. At the same time, the prediction model estimates the state evolution of the next 30 sampling points based on the current state, providing forward - looking information for control decisions, as shown in Table 6:

[0242] Table 6: Adaptive control decision example (during the production process of B - pillar reinforcement);

[0243]

[0244] Table 6 shows examples of adaptive control decisions generated by the system for different working conditions during the production of B-pillar reinforcements. It can be seen that the system can automatically adjust the fusion weights according to different working conditions and find the best balance between empirical rules and data-driven decisions. When facing obvious data characteristics such as fluctuations in the hardness of new materials, the decision-making weight of the neural network is relatively high; while when dealing with defect handling that requires specific domain knowledge, the decision-making weight of rule reasoning will automatically increase.

[0245] In practical applications, the control system always maintains the ability of online learning and optimization. After every 500 products are produced, the system will analyze the production data of this batch and update the parameters of the neural network model and the importance scores of the rules. Through long-term operation, the system gradually accumulates the best control strategies under different working conditions and forms an adaptive knowledge base.

[0246] Verification of technical effects:

[0247] After the intelligent control system was actually applied in an automotive parts manufacturing enterprise for 6 months, a comprehensive effect evaluation was carried out. It was verified mainly from two key technical effects: enhanced decision interpretability and improved product quality consistency.

[0248] Verification of decision interpretability effect:

[0249] An important technical effect of this system is to enhance the interpretability of control decisions, enabling operators to understand and trust system decisions. Before application, the traditional neural network-based control system could only give parameter adjustment suggestions and could not explain the reasons for the adjustments, resulting in low acceptance by operators and an actual adoption rate of only 43%.

[0250] After applying this system, a detailed explanation report will be generated for each control decision, including decision basis, activated rules, and feature importance analysis. The evaluation results of decision interpretability are shown in Table 7:

[0251] Table 7: Comparison of decision interpretability effects;

[0252]

[0253] The system generates differentiated explanation reports for different types of operators, from technical details at the engineer level to concise guidance at the operator level.

[0254] Through questionnaire surveys and on-site observations, it was found that the operators' understanding of the system's decisions increased from 4.2 points (out of 10) before application to 8.7 points, an increase of 107.1%; the decision adoption rate increased from 43.2% to 86.5%, an increase of 100.2%; the average decision-making time of the operators decreased from 127 seconds to 42 seconds, a decrease of 66.9%. These results indicate that the neuro-symbolic fusion method of this system enhances the interpretability of the decision-making process, transforms the intelligent control system from a black box into a transparent box, and significantly improves the acceptance and trust of the operators towards the system.

[0255] Verification of the product quality consistency effect:

[0256] Another key technical effect of this system is the improvement of product quality consistency. This effect is verified by comparing the product dimensional accuracy and surface quality before and after the system application.

[0257] In actual production, 3 products with different process difficulties (A-pillar reinforcement, B-pillar, and front longitudinal beam) were randomly selected. 1000 pieces were sampled before and after the system application respectively for precise measurement, and the deviation distribution of key dimensions was statistically analyzed. The comparison results of dimensional accuracy are shown in Table 8:

[0258] Table 8: Comparison of product dimensional accuracy (unit: mm);

[0259]

[0260] As can be seen from Table 8, after the application of this system, the average value of product dimensional deviation decreased from 0.81 mm to 0.37 mm, a decrease of 54.3%; the standard deviation of dimensional deviation decreased from 0.56 mm to 0.21 mm, a decrease of 62.5%; the out-of-tolerance rate decreased from 8.64% to 1.08%, a decrease of 87.5%. This indicates that this system significantly improves the dimensional accuracy and consistency of stamping products through an adaptive control strategy based on neuro-symbolic reasoning and formal verification.

[0261] Meanwhile, by statistically analyzing the surface defects of the products through the surface quality inspection system, it was found that after the system application, the surface defect rate of the products decreased from 5.8% to 1.9%, a decrease of 67.2%; in particular, the defects related to parameter control (such as wrinkles, cracks, etc.) decreased by 76.4%, which further verifies the effect of this system in improving product quality consistency.

[0262] The above describes the embodiments of the present invention, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. An intelligent control method for a stamping robot, characterized in that: The following steps are involved: Construct a knowledge graph containing the multi-dimensional relationship between material, process, equipment and quality, identify the causal relationship between parameters and form a process knowledge base; Convert process knowledge into formal logic rules and integrate them with deep neural networks to build a neural symbolic reasoning model; The joint optimization function of the neural symbolic reasoning model is: ； in, represents the total loss function of the neural symbolic reasoning model, Represents the neural network loss The weight coefficient of Indicates logical consistency loss The weight coefficient of is the neural network loss function, It is the loss of logical consistency; ； ； in, is the sample size, For neural networks, For the input samples, For the The true output of samples is is a set of logical rules, Representation Rules In neural networks Satisfaction under output; The stamping process is modeled as a discrete event system, and the correctness and consistency of the process control rules are verified using linear temporal logic specifications; Detect and resolve parameter constraint conflicts in process control rules using satisfiability modulo theory solvers; Generate and execute adaptive stamping control strategies based on a proven rule base and production process feedback, and continuously optimize the control system through a closed-loop feedback mechanism.

2. The intelligent control method for a stamping robot according to claim 1, characterized in that: The step of constructing a knowledge graph containing multidimensional relationships of material, process, equipment and quality comprises: Collect multi-source heterogeneous data from stamping process database, equipment operation log and expert experience documents; Use named entity recognition and relation extraction algorithms to identify key entities and dependency relationships from preprocessed data; Building a knowledge graph: ； in, represents the knowledge graph, Represents a collection of entities, Represents a set of relations; Causal inference algorithms are applied to identify causal relationships between parameters, and the strength and direction of causal relationships are determined through conditional independence test statistics: ； in Represents the conditional independence test statistic, which is used to measure the Under the condition of and independence, Represents a variable , , Sum all the values ​​of , Indicates the calculation of the log-likelihood ratio of conditional independence. and In a given When the condition is independent, the value is 0; otherwise, the larger the value, the and The stronger the correlation, Indicator , , The joint probability distribution of and Respectively indicate that in the given Under the conditions and The conditional probability distribution of .

3. The intelligent control method for a stamping robot according to claim 1, characterized in that: The step of modeling the stamping process as a discrete event system comprises: Define a discrete event system: ； in, is a discrete event system, is a combination of process parameters and equipment status. is the initial state, To control the command set, is the state transfer function; Create a time automaton: ； in, represents a timed automaton, is the location set, is the initial position, is a set of clock variables, is a set of transition edges with clock constraints, is a position-invariant function; The process safety and activity conditions are transformed into linear temporal logic formulas for specification definition.

4. The intelligent control method for a stamping robot according to claim 1, characterized in that: The steps of using the satisfiability modulo theory solver to detect and resolve parameter constraint conflicts include: Convert logical rules into constraint formulas for SMT problems ; Apply SMT solver to check the satisfiability of constraint formula set; Constructing a conflict graph for analysis of detected conflicts; Generate resolution strategies for different types of conflicts , including parameter range redivision, rule priority assignment, and supplementary rule generation; Update the rule base and confirm conflict elimination through formal verification.

5. The intelligent control method for a stamping robot according to claim 1, characterized in that: The steps of generating and executing the adaptive stamping control strategy include: Real-time collection of process parameter data through sensor networks ; Apply Kalman filter to estimate the current state of the system and predict the future state; Calling the neural symbolic reasoning model to generate control decisions, where the control decisions are obtained by weighted combination of rule reasoning results and neural network prediction results, and the weighting coefficients are dynamically adjusted according to the accuracy of historical decisions; Convert control decisions into robot control instructions and establish a closed-loop feedback mechanism; Neural network parameter update, rule optimization and control parameter adaptive adjustment are achieved based on execution feedback data.

6. The intelligent control method for a stamping robot according to claim 5, characterized in that: The formula for applying the Kalman filter to estimate the current state of the system is: ； in, express The estimated value of the system state at time express The estimated value of the system state at time is the state transfer matrix, is the control input matrix, is the control input, is the observed value, is the observation matrix, is the Kalman gain; The formula for the weighted combination rule reasoning result and the neural network prediction result is: ； in, express The optimal control decision at the moment, is the rule reasoning result, For the neural network prediction results, is the adaptive weight coefficient.

7. The intelligent control method for a stamping robot according to claim 5, characterized in that: The rule optimization is performed based on the rule importance metric: ； in, Representation Rules The importance measure, represents the number of time steps in a control cycle, Representation Rules Decision-making The contribution of Indicates Logical rules, Indicates The optimal control decision for each time step; The neural network parameters are updated using stochastic gradient descent: ； in, represents the parameter vector of the neural network, is the learning rate, is the actual observed value, is the loss function, Represents the loss function About parameters The gradient of Indicates in status Below is the predicted value output by the neural network.

8. The intelligent control method for a stamping robot according to claim 1, characterized in that: During the reasoning process of the neural symbolic reasoning model, the importance scores of the input features are calculated through the back-attribution technique: ； in, Represents input features The importance score of Represents the neural network output About input features The partial derivative of Indicates that the input sample Calculate partial derivatives at ; Combined with the activated logic rules, a decision explainability report is generated, allowing operators to understand the basis and process of system decisions.

9. An intelligent control system for a stamping robot, used to execute the steps in the intelligent control method for a stamping robot as claimed in any one of claims 1 to 8, characterized in that: include: The knowledge graph construction module is used to construct a knowledge graph containing the multi-dimensional relationship between materials, processes, equipment and quality, and to identify the causal relationship between parameters to form a process knowledge base; The neural symbolic reasoning module is used to convert process knowledge into formal logic rules and integrate it with the deep neural network to build a neural symbolic reasoning model; Formal verification module, which is used to model the stamping process as a discrete event system and verify the correctness and consistency of the process control rules using linear temporal logic specifications; A conflict detection and resolution module is used to detect and resolve parameter constraint conflicts in process control rules using a satisfiability modulus theory solver; The adaptive control module is used to generate and execute adaptive stamping control strategies based on a proven rule base and production process feedback, and to achieve continuous optimization of the control system through a closed-loop feedback mechanism.

Images