A digital control method and system for online operation of multiple punching machines

Abstract

The present invention relates to the field of intelligent control technology, specifically to a digital control method and system for the online operation of multiple punching machines, which obtains production line historical data and real-time collection of punching machine operation status data, combines product appearance images and quality inspection parameters, realizes product appearance defect detection through the YOLOv7 model, generates offset quality inspection parameters in combination with quality inspection standards; and generates process abnormality positioning results using the offset quality inspection parameters and appearance defect detection results. The process abnormality positioning results are fused with the punching machine operation status data to form multimodal comprehensive operation feature data for identifying abnormal processes. After the abnormal process is detected, the abnormal propagation path is analyzed through a dynamic working condition adaptive graph neural network model, the current working condition characteristics are extracted, and similarity matching is performed with historical data to extract the best historical reference parameters. The production process adjustment parameters are generated by dynamically correcting the historical reference parameters. The adaptive ability and flexible response ability of the production line to complex working conditions are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, in particular to a digital control method and system for online operation of multiple punching machines. Background Art

[0002] Punch press, also known as stamping machine tool, is a mechanical device that processes materials through a press and a mold. In large-scale production, multiple punch presses are usually required to operate online to form an automated stamping production line. Each punch press can perform different processes, such as punching, trimming and forming, so as to achieve continuous operation and form a complete processing production line. This online operation mode not only improves production efficiency, but also reduces production costs. It is an important development direction of modern manufacturing industry.

[0003] With the continuous development of online punching machine operation technology, digital control systems have made significant progress in accuracy and stability. However, in complex production lines with multiple punching machines operating in coordination, there are still problems with inaccurate abnormal positioning and lack of dynamic adaptability in process parameter adjustment. Traditional control methods usually rely on preset parameters and fixed logic, and can only generate static "optimal" parameters based on specific historical data, or predict and adjust parameters through machine learning models. These methods fail to fully consider the complex working conditions and real-time dynamic changes in the production process, such as the impact of equipment aging, fluctuations in material properties, or changes in environmental conditions on punching machine performance.

[0004] In actual production, this adjustment method that relies only on theoretical optimal parameters is often difficult to adapt to the nonlinear dynamic changes of working conditions, and is prone to adjustment lag or deviation, which can cause problems such as process backlogs, unbalanced rhythms or idle resources, significantly reducing production efficiency and product quality. In addition, due to the inability to dynamically respond to changes in actual working conditions, the effect of parameter adjustment is often limited and difficult to meet actual production needs, exposing the lack of adaptive and flexible response capabilities of the production line under complex working conditions, which needs to be improved urgently.

[0005] Therefore, a digital control method and system for online operation of multiple punching machines are proposed. Summary of the invention

[0006] The purpose of the present invention is to provide a digital control method and system for the online operation of multiple punching machines. By combining historical data, real-time operating status data and product appearance defect detection results, and using multimodal data fusion and dynamic modeling technology, the accurate positioning of abnormal processes and efficient analysis of associated propagation paths are achieved, and the abnormal propagation range is intelligently inferred based on the dynamic working condition adaptive graph neural network. By combining similarity calculation with historical data matching, the best historical reference parameters that match the current working conditions are extracted, and the production process adjustment parameters are dynamically corrected and generated to ensure that the adjustment parameters can adapt to changes in complex working conditions, environments and materials in real time. This design can significantly improve the accuracy of abnormal process identification and the efficiency of associated analysis, and optimize the process adjustment strategy of the stamping automation production line.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A digital control method for online operation of multiple punching machines, comprising:

[0009] Obtain historical data of a specified stamping automated production line; the automated production line includes N punching machines; obtain appearance images and quality inspection parameters of products produced by the stamping automated production line; input the preprocessed appearance images into a trained YOLOv7 model to obtain product appearance defect detection results; determine offset quality inspection parameters based on the quality inspection parameters and quality inspection standards; generate process anomaly positioning results based on the offset quality inspection parameters and the product appearance defect detection results;

[0010] The cumulative number of operation cycles, vibration signals, noise signals, punching forces, stroke speeds and mold surface temperatures of the N punching machines are collected respectively, and the punching machine equipment operation status data is obtained after being summarized; the process abnormality positioning result is used as an independent feature, and is integrated with the punching machine equipment operation status data to generate multi-modal comprehensive operation feature data;

[0011] Abnormal processes are identified according to the multimodal comprehensive operation characteristic data; when the abnormal process is identified, the associated abnormal process and the propagation path are obtained through the dynamic operating condition adaptive graph neural network model; the current operating condition characteristics are obtained by extracting the characteristics of the abnormal process, the associated abnormal process and the propagation path; in combination with similarity calculation, historical records matching the current operating condition characteristics in the historical data are identified, corresponding historical process parameters are extracted from the historical records, and optimal historical reference parameters are obtained; the optimal historical reference parameters are dynamically corrected to obtain production process adjustment parameters.

[0012] Furthermore, the historical data includes historical process parameters of each process from the specified start time to the current time, historical product indicators of each process, historical production rhythm of each process and historical task completion time of each process.

[0013] Furthermore, the quality inspection parameters include shape parameters, size parameters and function inspection parameters.

[0014] Furthermore, generating the process abnormality location result according to the offset quality inspection parameter and the product appearance defect detection result includes: establishing, according to expert knowledge and a process path diagram, an association mapping relationship between the offset quality inspection parameter and the process and an association mapping relationship between the product appearance defect detection result and the process, respectively, to obtain a quality inspection parameter-process mapping and an appearance defect-process mapping, respectively; calculating a comprehensive abnormality weight value for each process according to the quality inspection parameter-process mapping and the appearance defect-process mapping; determining a maximum comprehensive abnormality weight value from N comprehensive abnormality weight values; and selecting the process corresponding to the maximum comprehensive abnormality weight value as the process abnormality location result.

[0015] Furthermore, identifying the abnormal process according to the multimodal comprehensive operation feature data includes: summarizing the multimodal comprehensive operation feature data in units of each process to obtain a process-level comprehensive feature vector; inputting the process-level comprehensive feature vector into an isolation forest model, and calculating the abnormality score of each process by constructing a set of randomly segmented decision trees; when the abnormality score exceeds a preset threshold, determining that the process corresponding to the abnormal score is the abnormal process.

[0016] Furthermore, obtaining the associated abnormal process and the propagation path through the dynamic working condition adaptive graph neural network model includes: when the abnormal process is identified, marking the multimodal comprehensive operation feature data corresponding to the abnormal process as an abnormal seed node, constructing a propagation weight graph in combination with the process dependency, dynamically adjusting the edge weights of the propagation weight graph according to the process coupling degree and the historical propagation probability, dividing the comprehensive operation feature data corresponding to the remaining processes into potential abnormal nodes and normal nodes, and generating classified node feature data; inputting the classified node feature data into the initial graph neural network model, weighting the node features in combination with the real-time operation status through the dynamic feature enhancement module to obtain a dynamic fusion feature vector; and performing multiple rounds of aggregation and updating of the dynamic fusion feature vector using a message passing mechanism, and controlling the abnormal propagation range in combination with a time attenuation factor; calculating a node abnormality score; when the node abnormality score is higher than a set threshold, determining that the process corresponding to the node abnormality score is an associated abnormal process, and generating an abnormal propagation path from the abnormal seed node to the associated abnormal process according to the node abnormality score result and the edge weight information.

[0017] Furthermore, dynamically correcting the optimal historical reference parameters to obtain the production process adjustment parameters includes: acquiring real-time operating condition data, environmental status data and material status data; respectively inputting the real-time operating condition data, the environmental status data and the material status data into a trained multi-output nonlinear regression model, wherein the multi-output nonlinear regression model independently calculates the correction amount of each optimal historical reference parameter according to the nonlinear relationship corresponding to each other between the optimal historical reference parameters and the real-time operating condition data, the environmental status data and the material status data, wherein the correction amount includes an operating condition correction amount, an environmental correction amount and a material correction amount; for each optimal historical reference parameter, optimizing each correction amount through a dynamic weight allocation mechanism to obtain an optimized correction amount; adding the optimized correction amount to the corresponding optimal historical reference parameter to generate the production process adjustment parameter; the production process adjustment parameter is used for production control of the stamping automation production line.

[0018] A digital control system for the online operation of multiple punching machines, comprising:

[0019] A historical data management module, used to obtain historical data of a specified stamping automated production line; the automated production line includes N punching machines;

[0020] A data acquisition and processing module is used to respectively collect the cumulative number of operation cycles, vibration signals, noise signals, punching force, stroke speed and mold surface temperature of the N punching machines, and obtain the punching machine equipment operation status data after summarizing;

[0021] An appearance defect detection module is used to obtain an appearance image of a product produced by the stamping automation production line; input the preprocessed appearance image into a trained YOLOv7 model to obtain a product appearance defect detection result;

[0022] A quality inspection analysis module, used to obtain quality inspection parameters of products produced by the stamping automation production line; and determine offset quality inspection parameters based on the quality inspection parameters and quality inspection standards;

[0023] An abnormality positioning module, used to generate a process abnormality positioning result according to the offset quality inspection parameter and the product appearance defect detection result;

[0024] A feature fusion and abnormality identification module is used to use the abnormality location result of the process as an independent feature, fuse it with the operation status data of the punching machine equipment, generate multi-modal comprehensive operation feature data; and identify abnormal processes according to the multi-modal comprehensive operation feature data;

[0025] The abnormal propagation and working condition analysis module is used to obtain the associated abnormal process and the propagation path through the dynamic working condition adaptive graph neural network model when the abnormal process is identified; and obtain the current working condition characteristics by extracting the characteristics of the abnormal process, the associated abnormal process and the propagation path;

[0026] The historical matching and parameter optimization module is used to combine similarity calculation to identify the historical records in the historical data that match the current operating conditions, extract the corresponding historical process parameters from the historical records, and obtain the best historical reference parameters; dynamically correct the best historical reference parameters to obtain the production process adjustment parameters.

[0027] Furthermore, generating the process abnormality location result according to the offset quality inspection parameter and the product appearance defect detection result includes: establishing, according to expert knowledge and a process path diagram, an association mapping relationship between the offset quality inspection parameter and the process and an association mapping relationship between the product appearance defect detection result and the process, respectively, to obtain a quality inspection parameter-process mapping and an appearance defect-process mapping, respectively; calculating a comprehensive abnormality weight value for each process according to the quality inspection parameter-process mapping and the appearance defect-process mapping; determining a maximum comprehensive abnormality weight value from N comprehensive abnormality weight values; and selecting the process corresponding to the maximum comprehensive abnormality weight value as the process abnormality location result.

[0028] Furthermore, obtaining the associated abnormal process and the propagation path through the dynamic working condition adaptive graph neural network model includes: when the abnormal process is identified, marking the multimodal comprehensive operation feature data corresponding to the abnormal process as an abnormal seed node, constructing a propagation weight graph in combination with the process dependency, dynamically adjusting the edge weights of the propagation weight graph according to the process coupling degree and the historical propagation probability, dividing the comprehensive operation feature data corresponding to the remaining processes into potential abnormal nodes and normal nodes, and generating classified node feature data; inputting the classified node feature data into the initial graph neural network model, weighting the node features in combination with the real-time operation status through the dynamic feature enhancement module to obtain a dynamic fusion feature vector; and performing multiple rounds of aggregation and updating of the dynamic fusion feature vector using a message passing mechanism, and controlling the abnormal propagation range in combination with a time attenuation factor; calculating a node abnormality score; when the node abnormality score is higher than a set threshold, determining that the process corresponding to the node abnormality score is an associated abnormal process, and generating an abnormal propagation path from the abnormal seed node to the associated abnormal process according to the node abnormality score result and the edge weight information.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. By combining the offset quality inspection parameters and product appearance defect detection results, and using expert knowledge and process path diagrams to establish the correlation mapping relationship between parameters and processes, the impact of quality inspection parameters and appearance defects on each process can be effectively captured. Quality inspection parameters and appearance defects reflect the processing quality of the process and product appearance problems respectively. The fusion of the two forms the calculation basis for the comprehensive abnormality weight value, further ensuring that the positioning results take into account the multi-dimensional abnormal characteristics. Through the selection of the maximum abnormality weight value, the problem process is quickly and accurately located, avoiding the misjudgment or omission caused by relying on a single data source in traditional methods. Through the integration of multimodal data analysis and domain knowledge, technical problems such as insufficient positioning accuracy and delayed processing of process abnormalities in the production process are solved, laying a solid foundation for subsequent abnormal propagation evaluation and dynamic adjustment, which helps to improve the efficiency and accuracy of abnormal processing of production lines.

[0031] 2. The present invention proposes a dynamic working condition adaptive graph neural network model, which realizes the accurate analysis of the abnormal propagation range and the efficient identification of the associated abnormal processes. By marking the multimodal comprehensive operation feature data corresponding to the abnormal process as an abnormal seed node, combining the process dependency and the historical propagation mode to construct a propagation weight graph, and dynamically adjusting the edge weight, the dynamic adaptability and accuracy of the abnormal propagation range are ensured. At the same time, the node features are weighted by the dynamic feature enhancement module in combination with the real-time operation status to obtain a dynamic fusion feature vector; it can capture the impact of dynamic changes on abnormal propagation in real time, and improve the model's adaptability to complex working conditions. The multi-round aggregation update of the message passing mechanism effectively improves the inference accuracy of the graph model for the associated abnormal processes, and the introduction of the time decay factor further controls the abnormal propagation range and avoids the excessive diffusion of abnormal influences. The method dynamically combines static and real-time features, so that the system can quickly adapt to dynamic changes when processing complex working conditions, realizes accurate analysis of the abnormal propagation range and improves the flexible response capability, and helps to realize intelligent abnormal management and efficient operation of the production line.

[0032] 3. The present invention uses a dynamic correction mechanism to enable the production process adjustment parameters to adapt to complex and changeable production conditions, environmental conditions and material properties in real time, solving the technical problem that traditional static parameter adjustment methods are difficult to match actual production dynamics. By introducing a multi-output nonlinear regression model, the nonlinear relationship between the best historical reference parameters and real-time data is fully explored, and the correction amount of each parameter is independently calculated to ensure the pertinence and accuracy of the correction process. At the same time, the correction amount is optimized through a dynamic weight allocation mechanism, so that the adjustment of each parameter can be adaptively optimized according to its sensitivity to the process target, thereby effectively improving the accuracy of parameter adjustment. The production process adjustment parameters finally generated can respond to the impact of changes in equipment status, environmental conditions and material properties in real time, quickly match actual production dynamics, and maintain the balance of the production line and the stability of product quality in dynamic changes, significantly improving the adaptability and flexible response capabilities of process parameter adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of a digital control method for online operation of multiple punching machines provided by an embodiment of the present invention;

[0034] Figure 2 A structural diagram of a digital control system for online operation of multiple punching machines provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] See also Figure 1 to Figure 2 The present invention provides a digital control method and system for online operation of multiple punching machines, and the technical solution is as follows:

[0037] Embodiment 1:

[0038] A smart manufacturing company A is mainly engaged in the research and development, manufacturing and sales of metal forming machine tools, CNC machine tools and molds. In order to meet the requirements of modern manufacturing for high efficiency, high quality and low cost, the company deployed an intelligent stamping automation production line based on the online operation of multiple punching machines, mainly used for the mass production of hardware products. The digital control method, such as Figure 1 As shown, including: See Figure 1S1 in the figure obtains historical data of a specified stamping automated production line; the automated production line includes N punching machines; in this embodiment, the automated production line includes 6 punching machines, each punching machine corresponds to a stamping process, and the 6 punching machines are responsible for the production and processing of a complete automated production line through online collaborative operation.

[0039] See also Figure 1 S2 in which the appearance image and quality inspection parameters of the product currently produced by the stamping automation production line are obtained in real time; Figure 1 In S3, the pre-processed appearance image is input into the trained YOLOv7 model to obtain the product appearance defect detection result; see Figure 1 S4 in which the offset quality inspection parameter is determined according to the quality inspection parameter and the quality inspection standard; see Figure 1 S5, generating a process abnormality positioning result according to the offset quality inspection parameter and the product appearance defect detection result;

[0040] See also Figure 1 S6, respectively collecting the cumulative number of operation cycles, vibration signals, noise signals, punching force, stroke speed and mold surface temperature of the N punching machines in real time, and summarizing them to obtain the punching machine equipment operation status data;

[0041] See also Figure 1 In S7, the process abnormality positioning result is used as an independent feature and integrated with the punching equipment operation status data to generate multi-modal comprehensive operation feature data;

[0042] See also Figure 1 S8, identifying abnormal processes according to the multi-modal comprehensive operation characteristic data;

[0043] See also Figure 1 In S9, when the abnormal process is identified, the associated abnormal process and the propagation path are obtained through the dynamic working condition adaptive graph neural network model;

[0044] See also Figure 1 S10 in which the current working condition characteristics are obtained by extracting the characteristics of the abnormal process, the associated abnormal process and the propagation path; referring to Figure 1 In S11, combining similarity calculation, identifying the historical records in the historical data that match the current working condition characteristics, extracting the corresponding historical process parameters from the historical records, and obtaining the best historical reference parameters; referring to Figure 1 In S12, the optimal historical reference parameters are dynamically corrected to obtain production process adjustment parameters.

[0045] Specifically, in order to train the YOLOv7 model to achieve accurate detection of appearance defects of stamped products, it is first necessary to collect high-quality image data covering multiple defect types, including all defects that may occur in each process, such as cracks, dents, scratches, and deformations. At the same time, images of normal products must be collected as positive samples for model training. Use professional annotation tools to annotate the defective areas in the image and generate an annotation file that conforms to the YOLO format, which details the defect category, location, and bounding box information, and ensure that each defect type is associated with its corresponding process one by one to reflect the possibility and characteristics of the process causing the defect. After the annotation is completed, the data set is divided into training set, validation set, and test set, and it is ensured that the defect samples related to each process are evenly distributed in each part of the data. Subsequently, the sorted data set is loaded into the YOLOv7 model, and the pre-trained weights are used as the initial parameters. Combined with the requirements of stamped product inspection, the input size (such as 640 640), learning rate, batch size and other hyperparameters, complete transfer learning training within 50-100 training rounds to ensure that the model can learn and identify the characteristics of defects corresponding to each process. After training, the validation set is used to conduct a preliminary evaluation of the model performance, with the goal of achieving an mAP value of more than 90%, and the generalization ability of the model is further verified on the test set to ensure that it can accurately detect defect types in different processes. The final model should output the category, location and confidence of each defect. In actual applications, the trained YOLOv7 model is deployed to the digital control system and combined with the product appearance images collected in real time on the production line to quickly complete defect detection.

[0046] Furthermore, the historical data includes historical process parameters of each process from the specified start time to the current time, historical product indicators of each process, historical production rhythm of each process and historical task completion time of each process.

[0047] The historical data covers the historical process parameters, historical product indicators, historical production rhythm and historical task completion time of each process from the specified start time to the current time, and can provide an all-round and multi-dimensional operation data foundation. It can not only fully reflect the process operation rules of the production line and the dynamic relationship between processes, but also quantify the relationship between product quality and production efficiency, so as to have stronger background reference and decision support capabilities in abnormal analysis and dynamic adjustment. By combining historical process parameters with product indicators, the system can accurately capture the impact of the process on quality; through historical production rhythm and task completion time, it can analyze the collaborative relationship and bottleneck links between processes, and provide a basis for judging the propagation path of abnormal processes. Ultimately, the accuracy of matching historical data with current working conditions is significantly improved, ensuring that the extracted historical process parameters are more targeted and practical, and achieving more accurate process optimization and dynamic response capabilities.

[0048] Furthermore, the quality inspection parameters include shape parameters, size parameters and function inspection parameters.

[0049] Quality inspection parameters cover different dimensions and can fully reflect the overall picture of product processing quality. Shape parameters and size parameters can be used to quantify process accuracy, while functional detection parameters provide a basis for evaluating the actual performance of the product. The introduction of such multi-dimensional quality inspection parameters helps to comprehensively evaluate the quality status of the process links and accurately locate the causes of abnormalities to specific processes. At the same time, the diversity of quality inspection parameters provides more data support for subsequent analysis and parameter optimization, significantly improving the system's sensitivity to abnormalities and its adaptability to complex production scenarios, and ultimately achieving product quality stability and improved production efficiency.

[0050] Further, generating the process abnormality location result according to the offset quality inspection parameter and the product appearance defect detection result includes: establishing an association mapping relationship between the offset quality inspection parameter and the process according to expert knowledge and a process path diagram to obtain a quality inspection parameter-process mapping; the association degree between each offset quality inspection parameter and each process is expressed in the form of a weight, as shown in Table 1;

[0051] Table 1 Quality inspection parameters - process mapping example table

[0052]

[0053] Table 1 shows the correlation mapping relationship between offset quality inspection parameters and processes, which is established based on expert knowledge and process path diagram. Each offset quality inspection parameter in the table (such as length deviation, surface roughness change, etc.) clearly indicates its correlation degree with different processes through weight values. The weight value range is between 0 and 1, and the larger the value, the stronger the correlation degree. For example, the correlation weights of offset quality inspection parameter 002 with process 1 and process 2 are 0.3 and 0.7 respectively, indicating that this parameter is mainly generated by process 2, but process 1 also has a certain influence; the weight of offset quality inspection parameter 003 is only 1 in process 5, indicating that it is completely generated by process 5; and the weight of offset quality inspection parameter 004 is 0.5, corresponding to process 1 and process 2, respectively, indicating that this parameter is jointly affected by these two processes, and the degree of influence is equal. These weights clearly show the correlation strength between each offset quality inspection parameter and each process, providing a scientific basis for abnormal location.

[0054] According to the expert knowledge and the process path diagram, the association mapping relationship between the product appearance defect detection result generation process is established to obtain the appearance defect-process mapping; the association degree between each product appearance defect detection result and each process is expressed in the form of weight, as shown in Table 2;

[0055] Table 2 Appearance defect-process mapping example table

[0056]

[0057] According to the quality inspection parameter-process mapping and the appearance defect-process mapping, the comprehensive abnormality weight value of each process is calculated; the calculation formula of the comprehensive abnormality weight value is:

[0058] ;

[0059] in, is the index of the process; is the index of the offset quality inspection parameter; An index of the product appearance defect detection result; For process The comprehensive abnormal weight value of Offset the quality inspection parameters in the quality inspection parameter-process mapping Process The weight of The product appearance defect detection result in the appearance defect-process mapping Process The weight of

[0060] and are two global weights, representing the influence of offset quality inspection parameters and appearance defect detection results on the anomaly.

[0061] Determining a maximum comprehensive anomaly weight value from the N comprehensive anomaly weight values;

[0062] The process corresponding to the maximum comprehensive abnormality weight value is selected as the process abnormality location result.

[0063] Determining a maximum comprehensive anomaly weight value from the N comprehensive anomaly weight values;

[0064] The process corresponding to the maximum comprehensive abnormality weight value is selected as the process abnormality location result.

[0065] Specifically, calculate the comprehensive abnormal weight value of N processes: ;

[0066] Find the maximum value among N comprehensive anomaly weight values; ;

[0067] Determine the process with the largest comprehensive abnormality weight value ; If the comprehensive abnormality weight values ​​of multiple processes are the same and are all the maximum comprehensive abnormality weight values, the front process in the process path is selected as the process abnormality location result.

[0068] Through the mapping of quality inspection parameters to processes and appearance defects to processes, the correlation between product quality indicators and each process is quantified. Combined with the calculation formula of comprehensive abnormal weight values, the fusion analysis of multi-source data is realized, and the abnormal process is accurately located in a data-driven manner. By converting the complex abnormality location problem into a weight-based calculation problem, this method replaces the traditional method of relying on experience judgment with quantitative analysis, significantly improving the accuracy and efficiency of abnormality location. The weighted calculation of the comprehensive abnormality weight value enables the system to fully capture the correlation information between product quality and process operation status, adapt to the complex working conditions of different production lines, and support real-time calculation to quickly output abnormality location results. This method is based on weighted calculation and data-driven abnormality location analysis, combines the quality inspection parameter-process mapping and appearance defect-process mapping constructed by expert knowledge and historical data, and uses the method of multi-source data fusion and weight weighted analysis to achieve accurate identification of abnormal processes. It not only effectively improves the accuracy of abnormal location and the rapid response capability of the production line, but also provides a reliable data basis for subsequent abnormal propagation analysis and process optimization, ensuring the stability of production quality and the improvement of production efficiency.

[0069] Furthermore, identifying the abnormal process according to the multimodal comprehensive operation feature data includes: summarizing the multimodal comprehensive operation feature data in units of each process to obtain a process-level comprehensive feature vector; inputting the process-level comprehensive feature vector into an isolation forest model, and calculating the abnormality score of each process by constructing a set of randomly segmented decision trees; when the abnormality score exceeds a preset threshold, determining that the process corresponding to the abnormality score is the abnormal process, as shown in Table 3.

[0070] Specifically, the feature vector of each process ,in is the feature dimension, Is the process index.

[0071] Table 3 Multimodal comprehensive operation feature data and abnormal score data table

[0072]

[0073] Table 3 lists the relevant characteristic data of each process (such as vibration signal, noise signal and punching force, etc.) and its abnormality score. The abnormality score of each process is generated by algorithms such as the isolation forest model and compared with the preset threshold. In Table 3, the process abnormality location result is generated by the offset quality inspection parameters and the appearance defect detection results, indicating whether the process is a potential abnormal process (0 is normal, 1 is abnormal); in this embodiment, the preset threshold is 0.8, and process 3 and process 5 are judged as abnormal processes.

[0074] This method generates multimodal comprehensive operation feature data by fusing the process anomaly location results with the punching equipment operation status data, and summarizes the features for each process, which can fully reflect the operation status and abnormal characteristics of each process. Combined with the unsupervised anomaly detection capability of the isolation forest model, by constructing a set of randomly segmented decision trees, the abnormal behaviors of different processes in complex production processes can be accurately captured. Through the combination of multimodal data, the method can effectively avoid the problem of missed detection or misjudgment of anomalies caused by the one-sidedness of data in traditional single data source methods, and greatly improve the adaptability and comprehensiveness of detection. At the same time, through the threshold judgment mechanism of anomaly scoring, key abnormal processes can be quickly identified, which not only shortens the anomaly location time, but also lays a data foundation for subsequent anomaly propagation evaluation and parameter optimization, significantly improving the system's adaptability and real-time response capabilities to complex production environments.

[0075] Furthermore, when the abnormal process is identified, the associated abnormal process and the propagation path are obtained through the dynamic working condition adaptive graph neural network model, including: when the abnormal process is identified, the multimodal comprehensive operation feature data corresponding to the abnormal process is marked as an abnormal seed node, and a propagation weight graph is constructed in combination with the process dependency relationship, the edge weights of the propagation weight graph are dynamically adjusted according to the process coupling degree and the historical propagation probability, and the comprehensive operation feature data corresponding to the remaining processes are divided into potential abnormal nodes and normal nodes to generate classified node feature data; the classified node feature data is input into the initial graph neural network model, and the node features are weighted in combination with the real-time operation status through the dynamic feature enhancement module to obtain to a dynamic fusion feature vector; and use a message passing mechanism to perform multiple rounds of aggregation and update on the dynamic fusion feature vector, and combine the time attenuation factor to control the propagation range of the anomaly; calculate the node anomaly score; when the node anomaly score is higher than the set threshold, determine that the process corresponding to the node anomaly score is an associated abnormal process, and generate an abnormal propagation path from the abnormal seed node to the associated abnormal process according to the node anomaly score result and edge weight information; the dynamic feature enhancement module uses a dynamic feature mapping function to convert the real-time operating status data into a dynamic feature of the same dimension as the static feature, and weightedly fuses the static feature and the dynamic feature through a dynamic weighting coefficient, and finally forms the dynamic working condition adaptive graph neural network model.

[0076] The above node features refer to the multimodal comprehensive operation feature data corresponding to each process, including static features (such as fixed relationships between processes or historical propagation patterns) and dynamic features (such as real-time operation status and equipment status, etc.). The dynamic fusion feature vector is a multidimensional feature expression generated by weighted fusion of static features and real-time dynamic features, which can comprehensively reflect the working conditions and dynamic status of the node. The node anomaly score is calculated by inputting the feature vector into the scoring function (such as based on a multi-layer perceptron or a weighted sum model), which is used to quantify the degree of anomaly of each node.

[0077] The nonlinear scoring process based on the multi-layer perceptron (MLP) is as follows:

[0078] Build a simple multilayer perceptron that learns nonlinear relationships from the feature vector and outputs anomaly scores:

[0079] ;

[0080] Multilayer Perceptron (MLP) network structure:

[0081] Specifically, it includes an input layer, two hidden layers, and an output layer. The input layer receives the feature vector of the node , whose dimension is n, represents the multimodal features of the node (such as vibration signal, temperature and pressure, etc.). The first hidden layer is set to neurons, using the ReLU activation function to extract high-order combination information from the input features; the second hidden layer is set to In the output layer, a neuron is used to directly output the scalar anomaly score. , and the score is normalized to the range of [0,1] using the Sigmoid activation function.

[0082] Furthermore, dynamically correcting the optimal historical reference parameters to obtain the production process adjustment parameters includes: acquiring real-time operating condition data, environmental status data and material status data; respectively inputting the real-time operating condition data, the environmental status data and the material status data into a trained multi-output nonlinear regression model, wherein the multi-output nonlinear regression model independently calculates the correction amount of each optimal historical reference parameter according to the nonlinear relationship corresponding to each other between the optimal historical reference parameters and the real-time operating condition data, the environmental status data and the material status data, wherein the correction amount includes an operating condition correction amount, an environmental correction amount and a material correction amount; for each optimal historical reference parameter, optimizing each correction amount through a dynamic weight allocation mechanism to obtain an optimized correction amount; adding the optimized correction amount to the corresponding optimal historical reference parameter to generate the production process adjustment parameter; the production process adjustment parameter is used for production control of the stamping automation production line.

[0083] In order to realize the dynamic correction of the best historical reference parameters, a multi-output nonlinear regression model suitable for modeling complex nonlinear relationships is selected. Specifically, in this embodiment, the model adopts a deep neural network (DNN). To fully capture the nonlinear mapping relationship between real-time working condition data, environmental status data, material status data and target parameters. Through the training of historical data, the model independently outputs the correction amount of each best historical reference parameter (such as speed correction amount and pressure correction amount, etc.), and on this basis, further adjusts it through a dynamic weight optimization mechanism to ensure that the correction result can flexibly adapt to the dynamic changes of real-time working conditions, generate high-precision production process adjustment parameters, and finally achieve the adaptability and real-time performance of parameter adjustment.

[0084] This method introduces a multi-output nonlinear regression model, combines real-time working condition data, environmental status data and material status data, dynamically calculates the correction amount of each optimal historical reference parameter, and realizes the accurate generation of production process adjustment parameters. Compared with the traditional method that relies only on historical parameters, this method can effectively consider the complex real-time changing factors in the production process, such as the degradation of equipment performance, fluctuations in environmental conditions and differences in material properties, so that the adjusted process parameters are more in line with current production conditions. The correction amount is further optimized through the dynamic weight distribution mechanism to ensure the flexibility and accuracy of each parameter adjustment, which not only improves the response speed of production control, but also significantly improves the applicability of the adjustment parameters and the stability of the production line. This design overcomes the problem that static parameter adjustment is prone to failure in a dynamic environment, and provides higher intelligence and adaptability for stamping automation production lines.

[0085] A set of comparative experiments was conducted below. Experimental group A used a digital control method proposed in the present invention for the online operation of multiple punching machines; experimental group B used a traditional machine learning model (XGBoost model) to predict production operation parameters, and built a prediction model based on static data by learning the relationship between operation parameters and operation results in the historical database. In actual application, the model uses the current working conditions, equipment status and environmental conditions as inputs and directly outputs the predicted control parameters. For the implementation of experimental group C, please refer to Figure 1 From S1 to S11, the best historical reference parameters were directly used as control parameters. Experimental group D skipped product appearance inspection and quality inspection parameter analysis, and directly identified abnormal processes based on the punching equipment operation status data. Subsequently, the same dynamic correction method as experimental group A was used to generate process adjustment parameters.

[0086] Table 4 Comparative experimental results data table

[0087]

[0088] From the comparative experimental results in Table 4, it can be seen that experimental group A (using the method proposed by the present invention) performs well in two key dimensions: production line cycle time and product qualification rate. Its production line cycle time is 4.5 seconds, the shortest among all experimental groups, indicating that the method of the present invention can significantly improve production efficiency; at the same time, its product qualification rate reaches 98.5%, which is higher than other experimental groups, demonstrating the accuracy of the present invention in dynamic adjustment of process parameters and adaptation to complex working conditions. However, the number of parameter adjustments in experimental group A is 3 times. Although it is not the group with the least number of adjustments (experimental group C has 2 adjustments), it still shows higher stability than experimental groups B and D, and can achieve comprehensive optimization of production efficiency and product quality based on a small amount of adjustments.

[0089] In contrast, experimental group B (based on traditional machine learning models) failed to fully adapt to real-time dynamic conditions due to its reliance on static historical data to generate parameter predictions, resulting in its production line cycle time being extended to 5.2 seconds, the product qualification rate dropping to 94.3%, and the number of parameter adjustments being high, 5 times, indicating the shortcomings of traditional methods in terms of dynamics and stability. Although experimental group C (directly using the best historical reference parameters) only needed to adjust parameters twice, its cycle time was 6 seconds, significantly higher than other experimental groups, and its product qualification rate was the lowest, only 90.5%, showing the limitations of historical parameters that are difficult to adapt to the current production environment without correction. Experimental group D (skipping appearance inspection and quality inspection parameter analysis) performed better than experimental groups B and C in terms of cycle time and qualification rate, but slightly worse than experimental group A, and the number of parameter adjustments was 4 times, indicating that product appearance inspection and quality inspection parameter analysis play an important role in further improving the accuracy of process adjustment.

[0090] Overall, experimental group A achieved an optimal balance between cycle time, product qualification rate and parameter adjustment stability, and demonstrated high flexibility and adaptability, especially under dynamic and complex working conditions.

[0091] The data flow design of this method is centered on the gradual aggregation of multi-source data, and it advances from anomaly detection to propagation analysis and then to dynamic parameter correction. First, by collecting product appearance images and quality inspection parameters, the YOLOv7 model and offset quality inspection parameter analysis are used to quickly locate potential abnormal processes. In this way, the problem process can be directly locked through explicit data, significantly reducing the computational complexity of subsequent multi-source data fusion. Furthermore, the characteristics of the preliminary abnormal process are combined with the operating status data of the punching equipment (such as vibration signals, punching force, etc.) to generate comprehensive operating characteristic data. This integrates static process design information with real-time operating dynamics, which not only ensures the breadth of data, but also enhances the accuracy of problem location. After identifying the abnormal process, the associated abnormal process and propagation path are dynamically calculated based on the graph neural network model. This step dynamically evaluates the impact range of the anomaly by comprehensively considering the dependency relationship and propagation history between the processes, avoiding the limitations of isolated anomaly analysis. Furthermore, the current working condition characteristics are used to match the historical data for similarity, and combined with the dynamic correction method to generate the real-time optimal production process adjustment parameters. By considering the current working conditions, environment and material status, the problem of ignoring dynamic changes in traditional methods is solved, and true adaptive adjustment is achieved.

[0092] Embodiment 2:

[0093] Company B is a professional forging machinery manufacturer that has long been engaged in the research, design, manufacturing, marketing and service of mechanical presses. It is committed to providing customers with efficient and intelligent stamping solutions. In a modern production environment, especially in the fields of automotive parts and large household appliance casings, customers often require multiple punching machines to operate online in coordination to achieve efficient assembly line production. However, due to the large number of punching machines, complex processes, and huge operating data, the traditional fixed parameter control method has been unable to cope with complex and changing production needs. In order to solve this problem, Company B applied a digital control system for the online operation of multiple punching machines, such as Figure 2 As shown, it includes the following modules:

[0094] A historical data management module, used to obtain historical data of a specified stamping automated production line; the automated production line includes N punching machines;

[0095] A data acquisition and processing module is used to respectively collect the cumulative number of operation cycles, vibration signals, noise signals, punching force, stroke speed and mold surface temperature of the N punching machines, and obtain the punching machine equipment operation status data after summarizing;

[0096] An appearance defect detection module is used to obtain an appearance image of a product produced by the stamping automation production line; input the preprocessed appearance image into a trained YOLOv7 model to obtain a product appearance defect detection result;

[0097] A quality inspection analysis module, used to obtain quality inspection parameters of products produced by the stamping automation production line; and determine offset quality inspection parameters based on the quality inspection parameters and quality inspection standards;

[0098] An abnormality positioning module, used to generate a process abnormality positioning result according to the offset quality inspection parameter and the product appearance defect detection result;

[0099] A feature fusion and abnormality identification module is used to use the abnormality location result of the process as an independent feature, fuse it with the operation status data of the punching machine equipment, generate multi-modal comprehensive operation feature data; and identify abnormal processes according to the multi-modal comprehensive operation feature data;

[0100] The abnormal propagation and working condition analysis module is used to obtain the associated abnormal process and the propagation path through the dynamic working condition adaptive graph neural network model when the abnormal process is identified; and obtain the current working condition characteristics by extracting the characteristics of the abnormal process, the associated abnormal process and the propagation path;

[0101] The historical matching and parameter optimization module is used to combine similarity calculation to identify the historical records in the historical data that match the current operating conditions, extract the corresponding historical process parameters from the historical records, and obtain the best historical reference parameters; dynamically correct the best historical reference parameters to obtain the production process adjustment parameters.

[0102] Furthermore, the historical data includes historical process parameters of each process from the specified start time to the current time, historical product indicators of each process, historical production rhythm of each process and historical task completion time of each process.

[0103] Furthermore, the quality inspection parameters include shape parameters, size parameters and function inspection parameters.

[0104] Further, generating the process abnormality positioning result according to the offset quality inspection parameter and the product appearance defect detection result includes:

[0105] Establishing the association mapping relationship between the offset quality inspection parameters and the process according to the expert knowledge and the process path diagram to obtain the quality inspection parameter-process mapping; the association degree between each offset quality inspection parameter and each process is expressed in the form of weight;

[0106] Establishing the association mapping relationship between the product appearance defect detection result generation processes based on expert knowledge and the process path diagram to obtain the appearance defect-process mapping; the association degree between each product appearance defect detection result and each process is expressed in the form of weight;

[0107] According to the quality inspection parameter-process mapping and the appearance defect-process mapping, the comprehensive abnormality weight value of each process is calculated; the calculation formula of the comprehensive abnormality weight value is:

[0108] ;

[0109] in, is the index of the process; is the index of the offset quality inspection parameter; An index of the product appearance defect detection result; For process The comprehensive abnormal weight value of Offset the quality inspection parameters in the quality inspection parameter-process mapping Process The weight of The product appearance defect detection result in the appearance defect-process mapping Process The weight of

[0110] and are the first global weight and the second global weight, which respectively represent the influence of the offset quality inspection parameters and the appearance defect detection results on the anomaly.

[0111] Furthermore, identifying the abnormal process according to the multimodal comprehensive operation feature data includes: summarizing the multimodal comprehensive operation feature data in units of each process to obtain a process-level comprehensive feature vector; inputting the process-level comprehensive feature vector into an isolation forest model, and calculating the abnormality score of each process by constructing a set of randomly segmented decision trees; when the abnormality score exceeds a preset threshold, determining that the process corresponding to the abnormal score is the abnormal process.

[0112] Furthermore, when the abnormal process is identified, the associated abnormal process and the propagation path are obtained through the dynamic working condition adaptive graph neural network model, including: when the abnormal process is identified, the multimodal comprehensive operation feature data corresponding to the abnormal process is marked as an abnormal seed node, and a propagation weight graph is constructed in combination with the process dependency relationship, the edge weights of the propagation weight graph are dynamically adjusted according to the process coupling degree and the historical propagation probability, and the comprehensive operation feature data corresponding to the remaining processes are divided into potential abnormal nodes and normal nodes to generate classified node feature data; the classified node feature data is input into the initial graph neural network model, and the node features are weighted in combination with the real-time operation status through the dynamic feature enhancement module to obtain to a dynamic fusion feature vector; and use a message passing mechanism to perform multiple rounds of aggregation and update on the dynamic fusion feature vector, and combine the time attenuation factor to control the propagation range of the anomaly; calculate the node anomaly score; when the node anomaly score is higher than the set threshold, determine that the process corresponding to the node anomaly score is an associated abnormal process, and generate an abnormal propagation path from the abnormal seed node to the associated abnormal process according to the node anomaly score result and edge weight information; the dynamic feature enhancement module uses a dynamic feature mapping function to convert the real-time operating status data into a dynamic feature of the same dimension as the static feature, and weightedly fuses the static feature and the dynamic feature through a dynamic weighting coefficient, and finally forms the dynamic working condition adaptive graph neural network model.

[0113] Furthermore, dynamically correcting the optimal historical reference parameters to obtain the production process adjustment parameters includes: acquiring real-time operating condition data, environmental status data and material status data; respectively inputting the real-time operating condition data, the environmental status data and the material status data into a trained multi-output nonlinear regression model, wherein the multi-output nonlinear regression model independently calculates the correction amount of each optimal historical reference parameter according to the nonlinear relationship corresponding to each other between the optimal historical reference parameters and the real-time operating condition data, the environmental status data and the material status data, wherein the correction amount includes an operating condition correction amount, an environmental correction amount and a material correction amount; for each optimal historical reference parameter, optimizing each correction amount through a dynamic weight allocation mechanism to obtain an optimized correction amount; adding the optimized correction amount to the corresponding optimal historical reference parameter to generate the production process adjustment parameter; the production process adjustment parameter is used for production control of the stamping automation production line.

[0114] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A digital control method for online operation of multiple punching machines, characterized in that: include: Get historical data of specified stamping automation production line; The automated production line includes N punching machines; obtaining appearance images and quality inspection parameters of products produced by the automated stamping production line; Input the preprocessed appearance image into the trained YOLOv7 model to obtain the product appearance defect detection result; determine the offset quality inspection parameter according to the quality inspection parameter and quality inspection standard; Generate a process abnormality location result according to the offset quality inspection parameter and the product appearance defect detection result; including: according to expert knowledge and a process path diagram, respectively establish an association mapping relationship between the offset quality inspection parameter and the process and an association mapping relationship between the product appearance defect detection result and the process, and respectively obtain a quality inspection parameter-process mapping and an appearance defect-process mapping; calculate a comprehensive abnormality weight value of each process according to the quality inspection parameter-process mapping and the appearance defect-process mapping; determine a maximum comprehensive abnormality weight value from N comprehensive abnormality weight values; and select the process corresponding to the maximum comprehensive abnormality weight value as the process abnormality location result; The cumulative number of operation cycles, vibration signals, noise signals, punching forces, stroke speeds and mold surface temperatures of the N punching machines are collected respectively, and the punching machine equipment operation status data is obtained after being summarized; the process abnormality positioning result is used as an independent feature, and is integrated with the punching machine equipment operation status data to generate multi-modal comprehensive operation feature data; Identify abnormal processes according to the multimodal comprehensive operation feature data; including: aggregating the multimodal comprehensive operation feature data in units of each process to obtain a process-level comprehensive feature vector; inputting the process-level comprehensive feature vector into an isolation forest model, and calculating the abnormality score of each process by constructing a set of randomly segmented decision trees; when the abnormality score exceeds a preset threshold, determining that the process corresponding to the abnormal score is the abnormal process; When the abnormal process is identified, the associated abnormal process and the propagation path are obtained through the dynamic working condition adaptive graph neural network model; the current working condition feature is obtained by extracting the features of the abnormal process, the associated abnormal process and the propagation path; combined with similarity calculation, the historical records matching the current working condition feature in the historical data are identified, and the corresponding historical process parameters are extracted from the historical records to obtain the optimal historical reference parameters; Dynamically correct the optimal historical reference parameters to obtain production process adjustment parameters; including: obtaining real-time working condition data, environmental status data and material status data; respectively inputting the real-time working condition data, the environmental status data and the material status data into a trained multi-output nonlinear regression model, the multi-output nonlinear regression model independently calculates the correction amount of each optimal historical reference parameter according to the nonlinear relationship corresponding to each other between the optimal historical reference parameters and the real-time working condition data, the environmental status data and the material status data, the correction amount includes the working condition correction amount, the environmental correction amount and the material correction amount; for each optimal historical reference parameter, each correction amount is optimized by a dynamic weight allocation mechanism to obtain an optimized correction amount; the optimized correction amount is added to the corresponding optimal historical reference parameter to generate the production process adjustment parameter; the production process adjustment parameter is used for production control of the stamping automation production line.

2. A digital control method for online operation of multiple punching machines according to claim 1, characterized in that: The historical data includes historical process parameters of each process from the specified start time to the current time, historical product indicators of each process, historical production rhythm of each process and historical task completion time of each process.

3. A digital control method for online operation of multiple punching machines according to claim 1, characterized in that: The quality inspection parameters include shape parameters, size parameters and function inspection parameters.

4. A digital control method for online operation of multiple punching machines according to claim 1, characterized in that: Obtaining the associated abnormal process and the propagation path through the dynamic working condition adaptive graph neural network model includes: when the abnormal process is identified, marking the multimodal comprehensive operation feature data corresponding to the abnormal process as an abnormal seed node, building a propagation weight graph in combination with the process dependency, dynamically adjusting the edge weights of the propagation weight graph according to the process coupling degree and the historical propagation probability, dividing the comprehensive operation feature data corresponding to the remaining processes into potential abnormal nodes and normal nodes, and generating classified node feature data; inputting the classified node feature data into the initial graph neural network model, weighting the node features in combination with the real-time operation status through the dynamic feature enhancement module, and obtaining a dynamic fusion feature vector; and performing multiple rounds of aggregation and updating on the dynamic fusion feature vector by using a message passing mechanism, and controlling the abnormal propagation range in combination with a time decay factor; calculating a node abnormality score; when the node abnormality score is higher than a set threshold, determining that the process corresponding to the node abnormality score is an associated abnormal process, and generating an abnormal propagation path from the abnormal seed node to the associated abnormal process according to the node abnormality score result and the edge weight information.

5. A digital control system for the online operation of multiple punching machines, characterized in that: include: Historical data management module, used to obtain historical data of a specified stamping automation production line; The automated production line includes N punching machines; A data acquisition and processing module is used to respectively collect the cumulative number of operation cycles, vibration signals, noise signals, punching force, stroke speed and mold surface temperature of the N punching machines, and obtain the punching machine equipment operation status data after summarizing; An appearance defect detection module, used to obtain an appearance image of a product produced by the stamping automation production line; Input the preprocessed appearance image into the trained YOLOv7 model to obtain a product appearance defect detection result; A quality inspection analysis module, used to obtain quality inspection parameters of products produced by the stamping automation production line; and determine offset quality inspection parameters based on the quality inspection parameters and quality inspection standards; An abnormality locating module is used to generate a process abnormality locating result according to the offset quality inspection parameter and the product appearance defect detection result; including: establishing the association mapping relationship between the offset quality inspection parameter and the process and the association mapping relationship between the product appearance defect detection result and the process according to expert knowledge and a process path diagram, respectively, and obtaining the quality inspection parameter-process mapping and the appearance defect-process mapping respectively; calculating the comprehensive abnormality weight value of each process according to the quality inspection parameter-process mapping and the appearance defect-process mapping; determining the maximum comprehensive abnormality weight value from N comprehensive abnormality weight values; selecting the process corresponding to the maximum comprehensive abnormality weight value as the process abnormality locating result; The feature fusion and anomaly identification module is used to use the process abnormality positioning result as an independent feature, fuse it with the punching equipment operation status data, and generate multimodal comprehensive operation feature data; identify abnormal processes according to the multimodal comprehensive operation feature data; including: summarizing the multimodal comprehensive operation feature data in units of each process to obtain a process-level comprehensive feature vector; inputting the process-level comprehensive feature vector into the isolation forest model, and calculating the anomaly score of each process by constructing a set of randomly segmented decision trees; when the anomaly score exceeds a preset threshold, determining that the process corresponding to the anomaly score is the abnormal process; The abnormal propagation and working condition analysis module is used to obtain the associated abnormal process and the propagation path through the dynamic working condition adaptive graph neural network model when the abnormal process is identified; and obtain the current working condition characteristics by extracting the characteristics of the abnormal process, the associated abnormal process and the propagation path; A history matching and parameter optimization module is used to combine similarity calculation to identify the historical records in the historical data that match the current working condition characteristics, extract the corresponding historical process parameters from the historical records, and obtain the optimal historical reference parameters; dynamically correct the optimal historical reference parameters to obtain production process adjustment parameters; including: obtaining real-time working condition data, environmental status data and material status data; inputting the real-time working condition data, the environmental status data and the material status data into a trained multi-output nonlinear regression model respectively, and the multi-output nonlinear regression model independently calculates the correction amount of each of the optimal historical reference parameters according to the nonlinear relationship corresponding to the optimal historical reference parameters and the real-time working condition data, the environmental status data and the material status data, and the correction amount includes the working condition correction amount, the environmental correction amount and the material correction amount; for each of the optimal historical reference parameters, each correction amount is optimized through a dynamic weight allocation mechanism to obtain an optimized correction amount; the optimized correction amount is added to the corresponding optimal historical reference parameter to generate the production process adjustment parameter; the production process adjustment parameter is used for production control of the stamping automation production line.

6. A digital control system for online operation of multiple punching machines according to claim 5, characterized in that: Obtaining the associated abnormal process and the propagation path through the dynamic working condition adaptive graph neural network model includes: when the abnormal process is identified, marking the multimodal comprehensive operation feature data corresponding to the abnormal process as an abnormal seed node, building a propagation weight graph in combination with the process dependency, dynamically adjusting the edge weights of the propagation weight graph according to the process coupling degree and the historical propagation probability, dividing the comprehensive operation feature data corresponding to the remaining processes into potential abnormal nodes and normal nodes, and generating classified node feature data; inputting the classified node feature data into the initial graph neural network model, weighting the node features in combination with the real-time operation status through the dynamic feature enhancement module, and obtaining a dynamic fusion feature vector; and performing multiple rounds of aggregation and updating on the dynamic fusion feature vector by using a message passing mechanism, and controlling the abnormal propagation range in combination with a time decay factor; calculating a node abnormality score; when the node abnormality score is higher than a set threshold, determining that the process corresponding to the node abnormality score is an associated abnormal process, and generating an abnormal propagation path from the abnormal seed node to the associated abnormal process according to the node abnormality score result and the edge weight information.

Images