Aluminum veneer production control method and system based on large model

Abstract

The invention relates to the technical field of aluminum veneer production control, in particular to an aluminum veneer production control method and system based on a large model, and the method comprises the steps: constructing an aluminum veneer production whole-process knowledge graph, and receiving the natural language production task demands of a customer for the surface coating color, the plate thickness and the special-shaped cutting contour description of an aluminum veneer; performing semantic analysis and entity extraction on the demand by using a production instruction analysis engine, and converting the demand into a standardized process parameter query request capable of matching knowledge graph nodes; traversing the knowledge graph to retrieve a historical successful production batch with the highest association degree, and extracting a complete process parameter chain covering the whole process; and calling a parameter optimization model based on an attention mechanism, combining real-time equipment state data of the production line, performing dynamic fine adjustment on the parameter chain, and generating an optimal production control instruction set. According to the method, the precision and adaptability of production control are improved, and the personalized requirements of customers are met.

Description

Technical Field

[0001] This invention relates to the field of aluminum panel production control technology, and in particular to an aluminum panel production control method and system based on a large model. Background Technology

[0002] Aluminum panel production involves several closely related processes, including material preparation, forming, welding, grinding, and spraying. The rationality of the process parameters in each process directly determines the product quality. The production control process must take into account both the customer's personalized needs and the actual conditions of the production site. In the existing aluminum panel production control model, the production requirements proposed by customers are mostly presented in natural language, including key information such as the surface coating color of the aluminum panel, the thickness of the panel, and the irregular cutting contour. These requirements need to be manually analyzed and converted into executable production process parameters. Then, relying on the operator's experience, historical production parameters are selected as a reference to finally formulate production control instructions.

[0003] Manual interpretation of natural language requirements is prone to misunderstandings, leading to errors in process parameter settings. Furthermore, existing technologies lack a systematic integration of knowledge across the entire aluminum panel production process, making it impossible to quickly and accurately retrieve historically successful production batches and complete process parameter chains that highly match current customer needs. This results in inefficient parameter selection and a high degree of subjectivity. In addition, in current production control systems, process parameters remain fixed once determined, failing to dynamically adjust based on real-time equipment status on the production line. This makes it difficult to adapt to fluctuations in equipment operation under different production environments, leading to unstable production quality and an inability to meet customers' personalized and high-precision requirements for aluminum panel products. Therefore, a production control method is needed that can accurately interpret requirements, efficiently match parameters, and achieve dynamic adjustments. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a production control method and system for aluminum single-panel based on a large model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a production control method for aluminum single-panel based on a large model, comprising:

[0006] Construct a knowledge graph of the entire process of aluminum single-panel production, and receive natural language production task requirements input by users. The natural language production task requirements include the customer's description of the surface coating color, plate thickness and irregular cutting contour of the aluminum single-panel.

[0007] The production instruction parsing engine is used to perform semantic parsing and entity extraction on the natural language production task requirements. The information on the surface coating color, board thickness and irregular cutting contour of the aluminum single panel described by the customer is transformed into a standardized process parameter query request that can match the corresponding node in the full process knowledge graph.

[0008] Based on the standardized process parameter query request, the entire process knowledge graph is traversed to retrieve the historical successful production batches that have the highest correlation with each requirement in the standardized process parameter query request. The complete process parameter chain corresponding to the historical successful production batch is extracted. The complete process parameter chain covers the blanking process parameters, forming process parameters, welding process parameters, grinding process parameters, and spraying process parameters.

[0009] By invoking a parameter optimization model based on an attention mechanism and combining it with real-time equipment status data of the current production line, the parameters in the extracted complete process parameter chain are dynamically fine-tuned and adaptively corrected to generate an optimal production control instruction set suitable for the current production environment.

[0010] As a further aspect of the present invention, a production instruction parsing engine is used to perform semantic parsing and entity extraction on the natural language production task requirements, including:

[0011] Deploy a pre-trained large language model as a production instruction parsing engine;

[0012] The natural language production task requirements are input into the production instruction parsing engine, which uses its built-in word vector matrix to perform word segmentation and word embedding representation on the input text.

[0013] The word embedding sequence is semantically encoded using the multi-layer Transformer encoder structure of the production instruction parsing engine to generate a feature vector sequence containing global semantic information.

[0014] On the feature vector sequence, a classifier trained for the aluminum single-panel production field is applied to identify key semantic segments in the text that belong to the process requirement category;

[0015] Named entity recognition is performed on the identified key semantic fragments to specifically identify the color card code entity corresponding to the surface coating color of the aluminum single panel, the millimeter value entity corresponding to the thickness of the panel, and the two-dimensional contour point set description entity corresponding to the irregular cut contour.

[0016] The identified color card code entity, millimeter value entity, and two-dimensional contour point set description entity are encapsulated according to a preset process parameter data structure to form the standardized process parameter query request.

[0017] As a further aspect of the present invention, based on the standardized process parameter query request, the entire process knowledge graph is traversed to retrieve the historically successful production batches with the highest correlation to each requirement in the standardized process parameter query request, including:

[0018] The full-process knowledge graph includes raw material attribute data, equipment operating parameter data, historical process formula data, and finished product quality inspection data. The relationships between different categories of data are structured and stored in the form of nodes and edges in the full-process knowledge graph.

[0019] Using the color card code entity in the standardized process parameter query request as the starting query node, a graph traversal is performed in the coating material attribute subgraph of the full process knowledge graph to locate all historical production batch nodes that have used color card code coatings.

[0020] Among all the identified historical production batch nodes, the historical production batch nodes with mismatched plate thickness parameters are filtered out using the millimeter value entity as a constraint.

[0021] For the remaining historical production batch nodes, the geometric similarity between their associated contour description data and the two-dimensional contour point set description entity is further calculated, and a set of historical production batch nodes with a geometric similarity higher than a preset threshold is selected.

[0022] Calculate the comprehensive score of the historical finished product quality inspection data corresponding to each node in the historical production batch node set;

[0023] The historical production batch node with the highest comprehensive score is selected as the historical successful production batch with the highest correlation. All upstream process data nodes ending at the historical production batch node are extracted from the full process knowledge graph to form the complete process parameter chain.

[0024] As a further aspect of the present invention, the step of invoking a parameter optimization model based on an attention mechanism, combined with real-time equipment status data of the current production line, to dynamically fine-tune and adaptively correct each parameter in the extracted complete process parameter chain includes:

[0025] The real-time equipment status data of the current production line is obtained from the production monitoring system. The real-time equipment status data includes the tool wear coefficient of the CNC punch press, the atomization pressure value of the spray gun of the spraying robot, and the current oven temperature distribution data of the baking oven.

[0026] The original blanking process parameters, original forming process parameters, original welding process parameters, original grinding process parameters, and original spraying process parameters in the complete process parameter chain are input together with the real-time equipment status data into the parameter optimization model based on the attention mechanism.

[0027] The attention-based parameter optimization model first calculates the correlation weight between each parameter in the real-time equipment status data and each original process parameter in the complete process parameter chain;

[0028] Based on the calculated correlation weights, the original process parameters in the complete process parameter chain are weighted and corrected, wherein the original process parameters that are highly correlated with equipment status anomalies will receive a greater degree of correction.

[0029] After the weighted correction, the parameter optimization model based on the attention mechanism outputs adjusted adaptive blanking process parameters, adaptive molding process parameters, adaptive welding process parameters, adaptive grinding process parameters, and adaptive spraying process parameters, which together constitute the optimal production control instruction set.

[0030] As a further aspect of the present invention, it also includes the steps of executing the optimal production control instruction set and collecting production feedback data:

[0031] The adaptive blanking process parameters in the optimal production control instruction set are sent to the CNC cutting machine controller in the blanking workshop to drive the CNC cutting machine to blank aluminum plates according to the specified contour and size.

[0032] The adaptive forming process parameters in the optimal production control instruction set are sent to the CNC bending machine controller in the forming workshop to drive the CNC bending machine to form the sheet metal according to the specified angle and bending sequence.

[0033] While the material cutting and forming processes are underway, multi-dimensional feedback data is collected from the actual production process through sensor arrays deployed at key nodes of the production line. The multi-dimensional feedback data includes the smoothness measurement value of the aluminum plate cutting edge, the actual angle deviation value of the bending angle, and the timestamp record of the inter-process transmission.

[0034] The collected multi-dimensional feedback data is compared in real time with the corresponding expected parameters in the optimal production control instruction set to generate a process-level production status deviation report.

[0035] As a further aspect of the present invention, it also includes the step of iteratively updating the full-process knowledge graph based on production feedback data:

[0036] After a complete production batch is completed, the multi-dimensional feedback data collected from subsequent processes such as welding, grinding, and spraying are summarized to form a complete production execution trajectory data for a complete production batch.

[0037] The final quality inspection report of a complete production batch is associated with the production execution trajectory data, and the production batch is marked as a successful batch or a failed batch.

[0038] Using a graph structure update algorithm, the labeled production execution trajectory data is added as a new production instance node to the full-process knowledge graph;

[0039] Establish association edges between the new production instance node and the existing raw material attribute nodes, equipment parameter nodes, and process parameter nodes in the full-process knowledge graph. The weights of the association edges are initialized based on the quality of the production batch results.

[0040] The weights of all associated edges in the full-process knowledge graph are periodically recalculated and updated. The weight updates are based on the statistical information of recently added successful and failed production instances.

[0041] As a further aspect of the present invention, it also includes using the production instruction parsing engine to process production anomaly events:

[0042] When the online inspection device on the production line detects a product quality abnormality, it automatically generates a natural language alarm message containing a description of the abnormality characteristics.

[0043] The natural language alarm information is input into the production instruction parsing engine;

[0044] The production instruction parsing engine parses the natural language alarm information to identify the anomaly type, the process in which the anomaly occurred, and the severity level of the anomaly.

[0045] Based on the identified anomaly type and the process in which the anomaly occurred, the production instruction parsing engine performs a reverse tracing query in the full process knowledge graph to find production records of similar anomalies in the past and the corrective measures taken.

[0046] Based on the corrective action records retrieved and combined with the current real-time equipment status data, a targeted anomaly handling suggestion instruction is generated. The anomaly handling suggestion instruction includes equipment parameter adjustment, process step suspension, or material replacement scheme.

[0047] As a further aspect of the present invention, the calculation of the geometric similarity between the associated contour description data and the two-dimensional contour point set description entity includes:

[0048] The contour description data associated with the historical production batch nodes is extracted from the knowledge graph and parsed into a set of historical contour points.

[0049] The entity described by the two-dimensional contour point set is parsed into the target contour point set;

[0050] The historical contour point set and the target contour point set are subjected to coordinate normalization processing so that the historical contour point set and the target contour point set are in the same size and coordinate system;

[0051] The iterative nearest point algorithm is applied to calculate the minimum root mean square error required to align the historical contour point set to the target contour point set.

[0052] The calculated minimum root mean square error is inversely calculated and normalized to the range of zero to one. The resulting value is the geometric similarity.

[0053] As a further aspect of the present invention, the periodic recalculation and update of the weights of all associated edges in the full-process knowledge graph includes:

[0054] Set a fixed weight update cycle, and at the end of each weight update cycle, count all newly added successful production instances and failed production instances within the cycle;

[0055] For each associated edge in the full-process knowledge graph, find all historical production instance nodes that pass through the associated edge, and filter out the new instances that belong to the current update cycle;

[0056] Statistically analyze the number of successfully produced instances and the number of failed produced instances among the newly selected instances;

[0057] Based on the number of successful production instances and the number of failed production instances, the current confidence weight of the associated edge is recalculated according to a preset formula.

[0058] The calculated current confidence weight is used to replace the original weight value of the associated edge, thereby completing the dynamic update of the knowledge association strength in the full-process knowledge graph.

[0059] As a further aspect of the present invention, the present invention also includes a large-scale aluminum panel production control system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the large-scale aluminum panel production control method described above.

[0060] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0061] A knowledge graph representing the entire process of aluminum panel production is constructed. A production instruction parsing engine performs semantic analysis and entity extraction on natural language production task requirements, including descriptions of aluminum panel surface coating color, panel thickness, and irregular cutting contours. This is transformed into standardized process parameter query requests that match corresponding nodes in the full-process knowledge graph. Based on these requests, the entire knowledge graph is traversed to retrieve the most relevant historical successful production batches and extract the complete process parameter chain. This solution achieves accurate conversion of natural language requirements into standardized process parameter query requests, avoiding biases introduced by manual parsing. Simultaneously, the full-process knowledge graph enables rapid retrieval of historical successful process parameters, ensuring that the obtained process parameter chain covers the entire production process. This solves the problems of incomplete parameter matching, low retrieval efficiency, and strong subjectivity in conventional technologies, making process parameter settings more targeted and reducing production problems caused by missing or inaccurate parameters.

[0062] By employing an attention-based parameter optimization model and combining it with real-time equipment status data from the current production line, the system dynamically fine-tunes and adaptively corrects each parameter in the extracted complete process parameter chain, generating an optimal production control instruction set suitable for the current production environment. This solution accurately captures the impact of real-time equipment status on production parameters, enabling dynamic adaptation and adjustment of process parameters. It breaks the limitations of fixed parameters in conventional technologies, allowing production control instructions to align with the actual conditions of the current production environment. This reduces the impact of equipment status fluctuations on production quality, improves production process stability, ensures that key indicators such as aluminum panel surface coating, panel thickness, and irregular cutting contours meet customer requirements, and reduces the incidence of production failures. Attached Figure Description

[0063] Figure 1 This is a flowchart of a large-model-based aluminum single-panel production control method according to the present invention.

[0064] Figure 2 This is a flowchart of semantic parsing and entity extraction;

[0065] Figure 3 Flowchart for parameter optimization and instruction generation;

[0066] Figure 4 Radar chart comparing quality inspection indicators for production batches;

[0067] Figure 5 This is a comparison and analysis chart of process parameters. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0069] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0070] See Figure 1 The system receives natural language production task requirements input by the user, which include the customer's description of the aluminum panel surface coating color, panel thickness, and irregular cutting contour. Using a production instruction parsing engine, the system performs semantic parsing and entity extraction on these requirements, transforming the customer's description of the aluminum panel surface coating color, panel thickness, and irregular cutting contour into a standardized process parameter query request that matches the corresponding node in the full-process knowledge graph. Based on this standardized process parameter query request, the system traverses the full-process knowledge graph, retrieving the historically successful production batches with the highest correlation to each requirement in the query request. It then extracts the complete process parameter chain corresponding to this historically successful production batch, covering parameters for blanking, forming, welding, grinding, and spraying processes. Finally, an attention-based parameter optimization model is invoked, combined with real-time equipment status data from the current production line, to dynamically fine-tune and adaptively correct each parameter in the extracted complete process parameter chain, generating the optimal production control instruction set suitable for the current production environment.

[0071] See Figure 2In one embodiment of the present invention, a pre-trained large language model is deployed as a production instruction parsing engine. The production instruction parsing engine receives natural language production task requirements input by the user. An example of a natural language production task requirement is "I need to produce a batch of exterior wall panels, with a gray-white surface, a thickness of approximately 2.5 mm, and a shape of rounded rectangles." The production instruction parsing engine uses a built-in word vector matrix to perform word segmentation and word embedding representation on the input text, mapping words to distributed vectors. The word embedding sequence is then semantically encoded using a multi-layer Transformer encoder structure of the production instruction parsing engine. The multi-layer Transformer encoder consists of multiple identical layers stacked together, each layer containing a self-attention mechanism and a feedforward neural network. The encoding process generates a feature vector sequence containing global semantic information. On the feature vector sequence, a classifier trained for the aluminum single-panel production field is applied. The classifier identifies key semantic segments in the text that belong to the process requirement category. For example, it identifies "the surface is grayish-white", "the thickness is about 2.5 mm" and "the shape is a rectangle with rounded corners" as key semantic segments from the input. Data comparison shows that for the colloquial description "2.5 mm thick", the domain-trained classifier can accurately identify it as a thickness requirement, while the general classifier may ignore it or misclassify it.

[0072] In some embodiments, named entity recognition (NAME) is performed on the identified key semantic fragments, specifically identifying the color card code entity corresponding to the surface coating color of the aluminum panel, the millimeter numerical entity corresponding to the panel thickness, and the two-dimensional contour point set description entity corresponding to the irregular cut contour. From the key semantic fragment "the surface is grayish-white", the NAME recognition operation parses it into the color card code entity "RAL9002" according to the domain color card mapping table; from "the thickness is approximately 2.5 mm", the millimeter numerical entity "2.5" is identified; from "the shape is a rectangle with rounded corners", combined with the contour template library, the two-dimensional contour point set description entity is parsed out, which is expressed as a set of coordinate points defining the four corners of the rectangle as rounded corners. Optionally, the NAME recognition operation adopts a conditional random field model. The model is trained on the labeled aluminum panel production corpus and can distinguish color descriptions such as "grayish-white" from non-process descriptions such as "customer urgently needed". Data comparison shows that for the recognition of "light gray" and "dark gray", the domain-trained model has an accuracy of over 99%, while the general NAME recognition model often recognizes them as ordinary adjectives.

[0073] In practical implementation, the identified color card code entities, millimeter value entities, and two-dimensional contour point set description entities are encapsulated according to a preset process parameter data structure to form a standardized process parameter query request. The process parameter data structure is a dictionary containing fixed fields; the encapsulation process ensures that the same requirement from different expressions is normalized. It can be understood that the preset data structure is the fixed format required by the subsequent knowledge graph query interface. When performing contextual semantic encoding, the multi-layer Transformer encoder of the production instruction parsing engine uses its self-attention mechanism to calculate the association strength between each word and all other words. The association strength calculation can be expressed as:

[0074]

[0075] in: The query matrix is ​​obtained by linear transformation of the input word embedding vector sequence. The key matrix, also obtained by transforming the input sequence, is represented by... Represents a value matrix, It is the dimension of the key vector. It is a normalized exponential function. This represents the matrix transpose operation, which allows the vector representation of the word "rounded corners" to incorporate the contextual information of "rectangle" and "shape".

[0076] In one embodiment of the present invention, the full-process knowledge graph includes raw material attribute data, equipment operating parameter data, historical process formula data, and finished product quality inspection data. The relationships between different categories of data are structured and stored in the full-process knowledge graph in the form of nodes and edges. For example, a historical production batch node named "Batch P2025032001" is connected to the coating material attribute node "Color Card RAL9010" through the "Used Coating" edge, to the raw material attribute node "Thickness 3mm Aluminum Coil" through the "Used Raw Material" edge, to a set of specific process parameter nodes through the "Application Parameter" edge, and to a finished product quality inspection data node with an "Overall Score 95" through the "Output Finished Product" edge. Using the color card code entity "RAL9010" in the standardized process parameter query request as the starting query node, a graph traversal is performed in the coating material attribute subgraph of the full process knowledge graph to locate all historical production batch nodes that have used the color card code RAL9010 coating. The graph traversal may return multiple nodes including "batch P2025032001" and "batch P2023121505".

[0077] In some embodiments, among all historical production batch nodes using the RAL9010 coating located, filtering is performed using the millimeter value entity "3" in the normalized process parameter query request as a constraint. The raw material attribute data associated with each historical production batch node is traversed, and the "plate thickness" parameter is checked. Historical production batch nodes with mismatched plate thickness parameters are excluded; for example, "batch P2023121505" associated with "2.5mm thick aluminum coil" is excluded, while "batch P2025032001" associated with "3mm thick aluminum coil" is retained. For the remaining historical production batch nodes with matching plate thickness, the geometric similarity between the contour description data associated with each node and the two-dimensional contour point set description entity in the normalized process parameter query request is further calculated. The two-dimensional contour point set description entity is parsed into the target contour point set; for example, the user-described "arc wavy edge" is parsed into a sequence of 200 coordinate points. The contour description data associated with the historical production batch nodes is extracted from the knowledge graph and parsed into historical contour point sets; for example, the historical record "wavy panel contour" is parsed into another sequence of 180 coordinate points. The historical contour point set and the target contour point set are normalized by performing coordinate normalization. The two point sets are translated and scaled to be within a unit circle, so that the historical contour point set and the target contour point set are in the same size and coordinate system.

[0078] In practical implementation, the Iterative Closest Point Algorithm (RMSA) is applied to calculate the minimum root mean square error required to align the historical contour point set to the target contour point set. The IMSSA algorithm iteratively searches for the optimal rigid body transformation between two point sets to minimize the distance between corresponding points. After calculating the minimum RMSA, its reciprocal is normalized to the range of zero to one; the resulting value is the geometric similarity. The formula for calculating geometric similarity is:

[0079]

[0080] in: This represents the calculated geometric similarity. This represents the minimum root mean square error calculated using the iterative nearest point algorithm. This is a constant used to prevent the denominator from being zero and to control the shape of the normalization curve. A set of historical production batch nodes with geometric similarity higher than a preset threshold is selected. Data comparison shows that for the target contour "arc-shaped wavy edge," the calculated geometric similarity value with the historical contour "wavy" is 0.91, and the calculated geometric similarity value with the historical contour "rectangle" is 0.23. Therefore, the "wavy" historical batch is included in the set, while the "rectangle" historical batch is excluded.

[0081] Optionally, when calculating geometric similarity, if the number of points in the historical contour point set is inconsistent with the number in the target contour point set, a resampling method is used to ensure that the two point sets have the same number of points. For example, if the target contour point set has 200 points and a historical contour point set has 180 points, linear interpolation is performed on the historical contour point set to add 20 points, making the number of historical contour point sets also 200, before performing the iterative nearest-point algorithm calculation. It can be understood that coordinate normalization eliminates the influence of the actual size and position of the contour on the similarity calculation, allowing the algorithm to focus on comparing shape similarity. For example, a large wave contour with a side length of 10 meters and a small wave contour with a side length of 1 meter may have highly similar point set shapes after normalization, thus obtaining a high geometric similarity value. In some embodiments, the graph traversal of the full-process knowledge graph uses a breadth-first search algorithm. Starting from the "Color Card RAL9010" node, all "Use Coating" edges pointing to historical production batch nodes are traversed, collecting the first layer of nodes. Then, for each collected historical production batch node, the thickness value in the raw material attribute node connected by the "Used Raw Materials" edge is checked to see if it matches. Data comparison shows that for a query, the traversal process can quickly locate dozens of candidate batch nodes from thousands of nodes, and after filtering, a few nodes remain for geometric similarity calculation.

[0082] See Figure 3 In one embodiment of the present invention, real-time equipment status data of the current production line is obtained from the production monitoring system. This real-time equipment status data includes the tool wear coefficient of the CNC punch press, the atomization pressure value of the spray gun of the painting robot, and the current oven temperature distribution data of the baking oven. For example, the real-time equipment status data shows that the tool wear coefficient of the CNC punch press "equipment C001" is 0.85, the atomization pressure value of the spray gun of the painting robot "equipment S002" is 0.45 MPa, and the current oven temperature distribution data of the baking oven "equipment H003" shows an average temperature of 195 degrees Celsius on the left side of the oven and an average temperature of 205 degrees Celsius on the right side. The original blanking process parameters, original forming process parameters, original welding process parameters, original grinding process parameters, and original painting process parameters from the complete process parameter chain are input together with the real-time equipment status data into a parameter optimization model based on an attention mechanism. The original blanking process parameters may include "cutting speed 80 mm / s" and "cutting pressure 300 N," and the original painting process parameters may include "painting distance 200 mm" and "gun travel speed 500 mm / s."

[0083] The attention-based parameter optimization model first calculates the correlation weight between each parameter in the real-time equipment status data and each original process parameter in the complete process parameter chain. The formula for calculating the correlation weight is:

[0084]

[0085] in: This represents the correlation weight between the i-th equipment status parameter and the j-th original process parameter. This represents the standardized value of the i-th device status parameter. This represents the standardized value of the j-th original process parameter. This represents the Sigmoid activation function. This represents a trainable weight matrix. This represents the bias vector.

[0086] In some embodiments, the original process parameters in the complete process parameter chain are weighted and corrected based on the calculated correlation weights. The rule for weighted correction is that original process parameters with high correlation to equipment status anomalies will receive a greater correction magnitude. For example, an anomaly is detected in the "current oven temperature distribution data" where the left side is lower than the right side (195 degrees Celsius on the left and 205 degrees Celsius on the right, while the standard requirement is 200±5 degrees Celsius), and the correlation weight between this anomaly and the "baking time 300 seconds" parameter is calculated to be 0.90. Based on this high weight, the parameter optimization model makes a significant correction to the "baking time" parameter, dynamically adjusting the original baking time from 300 seconds to 315 seconds according to the oven temperature deviation model to compensate for insufficient coating curing that may be caused by the lower oven temperature on the left side. Data comparison shows that for parameter pairs with correlation weights below 0.2, such as "tool wear coefficient" and "welding current," the correction magnitude is less than 1% of the original value; for parameter pairs with correlation weights above 0.8, the correction magnitude can reach 5% to 10% of the original value.

[0087] After weighted correction, the attention-based parameter optimization model outputs adjusted adaptive parameters for the blanking, forming, welding, grinding, and spraying processes. These adaptive parameters together constitute the optimal production control instruction set. For example, the corrected adaptive blanking parameters are "cutting speed 75 mm / s," and the adaptive spraying parameters are "spraying distance 210 mm" and "gun speed 480 mm / s." Executing the optimal production control instruction set involves sending the adaptive blanking parameters to the CNC cutting machine controller in the blanking workshop, driving the CNC cutting machine to blank the aluminum sheet according to the specified contour and dimensions. Similarly, the adaptive forming parameters are sent to the CNC bending machine controller in the forming workshop, driving the CNC bending machine to form the sheet metal according to the specified angle and bending sequence.

[0088] In practice, during the material cutting and forming processes, multi-dimensional feedback data is collected from the actual production process through sensor arrays deployed at key nodes of the production line. The sensor arrays include a laser profilometer installed at the end of the cutting path, angle sensors installed on the bending machine mold, and photoelectric switches distributed at each workstation. The multi-dimensional feedback data includes measurements of the smoothness of the aluminum plate's cut edge, the actual angle deviation of the bending angle, and timestamp records of inter-process transmission. For example, the laser profilometer measures the smoothness Ra value of the first aluminum plate in the current batch to be 12.5 micrometers, the angle sensor measures the actual angle of the first bend to be 89.8 degrees (the command value is 90 degrees), and the photoelectric switches record the transmission time from material cutting to the forming station as 120 seconds.

[0089] The collected multi-dimensional feedback data is compared in real time with the corresponding expected parameters in the optimal production control instruction set to generate process-level production status deviation reports. For example, comparing the smoothness measurement value Ra=12.5 micrometers with the expected Ra≤10 micrometers in the instruction generates a deviation entry for "edge smoothness in the blanking process exceeds expectations". Comparing the actual bending angle of 89.8 degrees with the instruction value of 90 degrees generates a deviation entry for "angle deviation in the forming process -0.2 degrees". It can be understood that timestamp records are used to calculate the process cycle time and compared with the production plan cycle time to generate a deviation entry for "transmission time delay of 20 seconds". Data comparison shows that when producing similar products using historical fixed parameters, the average deviation of the first batch bending angle is ±0.5 degrees, while after applying the dynamically fine-tuned optimal production control instruction set, the average deviation of the first batch bending angle is reduced to ±0.2 degrees.

[0090] In one embodiment of the present invention, after a complete production batch is completed, multi-dimensional feedback data collected from subsequent processes such as welding, grinding, and spraying are summarized to form complete production execution trajectory data for a complete production batch. For example, when a production batch identified as "Batch N2026031501" is completed, the system summarizes the material edge smoothness sequence, actual angle sequence of each bend, welding current and voltage waveform data, grinding roughness measurement value sequence, film thickness data of each spraying pass, and transmission time between processes recorded during the production process of this batch. These multi-dimensional feedback data arranged by timestamps are packaged to form the production execution trajectory data for "Batch N2026031501". The final quality inspection report of a complete production batch is associated with the production execution trajectory data. The quality inspection report includes the test results and overall judgment conclusions for items such as finished product dimensional accuracy, coating adhesion, color difference, and appearance defects. The system binds the production execution trajectory data with the quality inspection report and marks the production batch as a successful batch or a failed batch based on the overall judgment conclusion of the quality inspection report.

[0091] Using a graph structure update algorithm, the labeled production execution trajectory data is added as a new production instance node to the full-process knowledge graph. For example, the production execution trajectory data of "batch N2026031501" is encapsulated as a new node. The node attributes include batch ID, production time, final quality status (success / failure), and a link to detailed production execution trajectory data. The graph structure update algorithm inserts this new node into the "production instance" node category of the full-process knowledge graph. Connection edges are established between the new production instance node and existing raw material attribute nodes, equipment parameter nodes, and process parameter nodes in the full-process knowledge graph. The weights of these connections are initialized based on the quality of the production batch's results. If "Batch N2026031501" is marked as a successful batch, then the initial weights of the "Use" edges pointing from the batch node to the actual used raw material attribute node of "3mm thick aluminum coil", the equipment parameter node of "Model RBT-200 spraying robot", and the process parameter node of "spraying distance 210mm" are all set to a higher base value; if it is a failed batch, the initial weights are set to a lower base value.

[0092] The weights of all associated edges in the entire process knowledge graph are periodically recalculated and updated. These weight updates are based on statistical information about recently added successful and failed production instances. A fixed weight update cycle is set, such as weekly. At the end of each cycle, the system compiles statistics on all newly added successful and failed production instances within that cycle. See Table 1 for a summary of new instance statistics within an example weight update cycle.

[0093] Table 1: Statistics of New Production Instances Added During the Weight Update Cycle in Week 15

[0094] Instance Batch ID Related raw materials Related key process parameters Related key equipment Production results N2026031501 3mm thick aluminum coil Spraying distance 210mm RBT-200 painting robot success N2026031601 3mm thick aluminum coil Spraying distance 200mm RBT-200 painting robot fail N2026031701 2.5mm thick aluminum coil Cutting speed 75mm / s CNC punch press C-1000 success

[0095] In some embodiments, for each associated edge in the full-process knowledge graph, all historical production instance nodes connected to the associated edge are searched, and new instances belonging to the current update cycle are selected. Taking the associated edge "Use-Spraying Robot RBT-200" as an example, all production instance nodes connected to the "Spraying Robot RBT-200" equipment parameter node through the "Use" edge are searched in the graph, and new instance nodes with batch IDs belonging to week 15 (i.e., N2026031501, N2026031601) are selected from these nodes. The number of successful production instances and the number of failed production instances among the selected new instances are counted. For the edge "Use-Spraying Robot RBT-200", among the new instances selected in the current cycle, the number of successful production instances is 1 (N2026031501), and the number of failed production instances is 1 (N2026031601).

[0096] Based on the number of successful and failed production instances, the current confidence weight of the associated edges is recalculated according to a preset formula. The confidence weight update formula is:

[0097]

[0098] in: This represents the calculated current confidence weight of the associated edge. This indicates the number of successfully generated instances that have passed through this associated edge within the current weight update period. This indicates the number of failed production instances that passed through this associated edge during the current weight update cycle. and To smooth hyperparameters, used to provide prior estimates when the amount of data is small. For example, setting... , For the side "Using - RBT-200 painting robot", substitute... , Calculations yielded The calculated current confidence weight replaces the original weight value of the associated edge, thus dynamically updating the knowledge association strength in the entire process knowledge graph. If the original weight of the edge was 0.7, it is updated to 0.5.

[0099] It is understandable that weight updates are based on local statistics of newly added instances within a period, and the weight updates of different associated edges are calculated independently and in parallel. For example, the associated edge "Use - 3mm thick aluminum coil" connects one successful instance (N2026031501) and one failed instance (N2026031601) in the same period of newly added instances, and its weight is recalculated and updated; while the associated edge "Use - Cutting speed 75mm / s" only connects to one successful instance (N2026031701), and its recalculated weight value will be higher. Data comparison shows that an edge "Use - Model X welding machine" that has been stably connected to successful batches for a long time maintains a weight above 0.85 after multiple updates; while a newly appearing edge "Use - New Model Y coating" that connects to two consecutive failed batches has a weight that drops rapidly from the initial 0.9 to 0.3 after one update.

[0100] See Figure 4This is a radar chart comparing the quality inspection indicators of production batches, showing the scores of successful batch (N2026031501) and failed batch (N2026031601) across six key quality dimensions. The successful batch is significantly better in dimensional control, coating adhesion is a prominent advantage, while color difference control is a clear weakness of the failed batch. The successful batch exhibits more consistent appearance quality, significant differences in bending process precision control, and relatively smaller quality differences in the cutting process. The successful batch scores significantly higher than the failed batch in all dimensions, forming a complete "high-quality closed loop." The failed batch scores lowest in color difference and bending angle precision, which are likely the main reasons for batch failure. Even the failed batch performs relatively well in cutting smoothness, indicating relatively mature process control in this step. For color difference and bending angle precision, the process parameters of the spraying and bending processes should be optimized.

[0101] In one embodiment of the present invention, when the online inspection device on the production line detects a product quality anomaly, it automatically generates a natural language alarm message containing a description of the anomaly characteristics. The online inspection device can be a machine vision inspection system installed at the end of the painting line. When the machine vision inspection system detects a "sagging" defect on the surface of an aluminum panel, the automatically generated natural language alarm message reads: "Vertical stripes of paint enrichment were found on the surface of panel number 15 in batch P2026032001, suspected to be a sagging defect, located in the lower right corner of the panel, with a moderate severity." The natural language alarm message is then input into the production instruction parsing engine, which parses the alarm message to identify the anomaly type, the process in which the anomaly occurred, and the severity level of the anomaly. The production instruction parsing engine processes the input text "Vertical stripes of paint enrichment were found on the surface of plate number 15 in batch P2026032001, suspected to be a sagging defect, located in the lower right corner of the plate, with a moderate severity." Through its internal named entity recognition model and classification model, it identifies the anomaly type as "sagging," the anomaly occurrence process as "spraying process," and the anomaly severity level as "moderate."

[0102] In some embodiments, based on the identified anomaly type and the process in which the anomaly occurred, the production instruction parsing engine performs a reverse tracing query in the full process knowledge graph. The reverse tracing query starts with the "spraying process" node and the "sagging" defect type node, searching the full process knowledge graph for historical production records of similar anomalies and the corrective actions taken. The query logic searches for historical production instance nodes that are connected to the "sagging" node via a "causing defect" edge and to the "spraying process" node via a "belonging process" edge. For example, the query finds that the node for the historical production batch "P2023120803" is connected to the "sagging" node via a "causing defect" edge, and the corrective action record node associated with this batch node records past actions such as "adjusting the spraying robot's spray gun atomization pressure from 0.4 MPa to 0.45 MPa" and "reducing the baking oven heating rate from 5 degrees Celsius / minute to 3 degrees Celsius / minute."

[0103] Based on the retrieved corrective action records and combined with the current real-time equipment status data, targeted anomaly handling suggestions are generated. For example, combining the retrieved historical corrective actions and reading the current real-time equipment status data, it shows that the atomization pressure of the spray gun of the painting robot is 0.38 MPa and the current heating rate of the baking oven is 6 degrees Celsius / minute. Based on this, the generated anomaly handling suggestions include specific equipment parameter adjustments, process step suspension, or material replacement solutions. For the above example scenario, the generated anomaly handling suggestions might be: "Immediately execute: 1. Adjust the atomization pressure of the spray gun of the RBT-200 painting robot to 0.43 MPa; 2. Lower the heating rate setting of the H-3 baking oven to 3.5 degrees Celsius / minute; 3. Suspend the production line for 2 minutes and check the paint viscosity." Data comparison shows that for the same "sagging" alarm, if the current real-time equipment status data (current atomization pressure 0.38 MPa) is not considered, the historical value of 0.45 MPa might be directly suggested. However, after considering the current status, the suggested value is corrected to 0.43 MPa, making the suggestion more adaptable.

[0104] In practical implementation, the production instruction parsing engine identifies the severity level of anomalies by analyzing the keyword strength in the alarm message text. For example, when words like "severe" or "large area" appear in the alarm message, the severity level is identified as "high"; when words like "minor" or "local" appear, the severity level is identified as "low"; and when words like "moderate" or "partial" appear, the severity level is identified as "medium". The severity level is used to filter and sort the historical corrective action records retrieved. For anomalies with a severity level of "high", records containing measures such as "production line suspension" or "immediate material change" are prioritized. In reverse tracing queries, the association strength between historical corrective action records and historical production instance nodes is represented by the weight of a "take action" edge. The weight value is dynamically updated through the effectiveness evaluation of historical measures, and corrective action records with high association weights are prioritized during queries.

[0105] It is understandable that the generated anomaly handling suggestion instructions are formatted as structured control instructions or work orders, directly issued to the production execution system or displayed to operators. The structured instructions include the specific execution object, target parameter adjustment values, and optional action sequences. For example, "Object: Spray painting robot RBT-200, Parameter: Atomization pressure, Target value: 0.43MPa, Action: Adjust immediately"; "Object: Baking oven H-3, Parameter: Heating rate, Target value: 3.5°C / min, Action: Effective in the next batch". Optionally, when generating handling suggestion instructions, the synergistic effect of multiple historical corrective measures will be evaluated. For example, historically, there are records of separately "adjusting atomization pressure" and separately "reducing the heating rate" for "sagging," as well as a record of adjusting both parameters simultaneously. The system will calculate the expected effect score of these measures combined under the current real-time equipment state and select the combination with the highest score to generate the suggestion. The formula for calculating the expected effect score is:

[0106]

[0107] in: This represents the expected effect score of measure combination c. This represents the average effective weight of the m-th individual measure in the combination across its historical application records. This indicates the degree of matching between the current equipment status parameters and the historical equipment status parameters targeted by this measure. The higher the degree of matching, the better. The closer the value is to 1, the better. Data comparison shows that, for the same historical measure library, under the same abnormal and real-time conditions, the recommended instructions generated after synergy effect assessment have a higher degree of consistency with expert decisions in simulation verification compared to randomly selecting a historical measure.

[0108] In some embodiments, the natural language alarm information automatically generated by the online inspection device follows a predefined template. The template includes fields such as anomaly location, defect description, inspection time, and equipment number. These fields are filled in to form a complete statement. For example, the template might be "[Defect description] was found on the surface of [part number] of [batch number], located in [location area], with a severity of [level]". The production instruction parsing engine achieves higher parsing accuracy for templated information than for descriptions of completely free text because the template ensures the existence and structure of key entities. It can be understood that reverse tracing queries not only find completely identical anomaly types but also similar anomaly types. For example, for an "orange peel" defect, the query will also search for historical records of surface appearance defects such as "poor leveling" and "uneven gloss," and recommend results based on the similarity of the defect feature vectors.

[0109] See Figure 5 This is a comparative analysis chart of process parameters, showing the numerical comparison of standard parameters, actual parameters, and optimized parameters in five core processes throughout the entire aluminum panel production process. It visually reflects the effectiveness of the parameter optimization model. The actual parameters for blanking are lower than the standard, while the optimized parameters are slightly higher, achieving a small improvement. The actual parameters for forming show a significant gap from the standard, and the optimized parameters show the largest improvement, demonstrating a significant effect. The actual parameters for welding are slightly lower than the standard, but the optimized parameters effectively compensate for this gap. The actual parameters for grinding are the lowest among all processes, but the optimized parameters restore them to a reasonable level. The actual parameters for spraying differ from the standard, but the optimized parameters reach the highest level in the entire process. In all processes, the actual parameters are lower than the standard parameters, while the optimized parameters are higher than the standard parameters. The largest improvement indicates that this process is most sensitive to parameter changes and is a key link in the optimization model's effectiveness. The lowest actual parameters represent a weak point in production control, requiring close attention to equipment status and process adjustments.

[0110] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for aluminum single-panel production control based on a large model, characterized in that, The method includes: Construct a knowledge graph of the entire process of aluminum single-panel production, and receive natural language production task requirements input by users. The natural language production task requirements include the customer's description of the aluminum single-panel surface coating color, panel thickness and irregular cutting contour. The production instruction parsing engine is used to perform semantic parsing and entity extraction on the natural language production task requirements. The information on the surface coating color, board thickness and irregular cutting contour of the aluminum single panel described by the customer is transformed into a standardized process parameter query request that can match the corresponding node in the full process knowledge graph. Based on the standardized process parameter query request, the entire process knowledge graph is traversed to retrieve the historical successful production batches that have the highest correlation with each requirement in the standardized process parameter query request. The complete process parameter chain corresponding to the historical successful production batch is extracted. The complete process parameter chain covers the blanking process parameters, forming process parameters, welding process parameters, grinding process parameters, and spraying process parameters. By invoking a parameter optimization model based on an attention mechanism and combining it with real-time equipment status data of the current production line, the parameters in the extracted complete process parameter chain are dynamically fine-tuned and adaptively corrected to generate an optimal production control instruction set suitable for the current production environment.

2. The aluminum single-panel production control method based on a large model according to claim 1, characterized in that, The production instruction parsing engine is used to perform semantic parsing and entity extraction on the natural language production task requirements, including: Deploy a pre-trained large language model as a production instruction parsing engine; The natural language production task requirements are input into the production instruction parsing engine, which uses its built-in word vector matrix to perform word segmentation and word embedding representation on the input text. The word embedding sequence is semantically encoded using the multi-layer Transformer encoder structure of the production instruction parsing engine to generate a feature vector sequence containing global semantic information. On the feature vector sequence, a classifier trained for the aluminum single-panel production field is applied to identify key semantic segments in the text that belong to the process requirement category; Named entity recognition is performed on the identified key semantic fragments to specifically identify the color card code entity corresponding to the surface coating color of the aluminum single panel, the millimeter value entity corresponding to the thickness of the panel, and the two-dimensional contour point set description entity corresponding to the irregular cut contour. The identified color card code entity, millimeter value entity, and two-dimensional contour point set description entity are encapsulated according to a preset process parameter data structure to form the standardized process parameter query request.

3. The aluminum single-panel production control method based on a large model according to claim 2, characterized in that, Based on the standardized process parameter query request, the entire process knowledge graph is traversed to retrieve the historically successful production batches with the highest correlation to each requirement in the standardized process parameter query request, including: The full-process knowledge graph includes raw material attribute data, equipment operating parameter data, historical process formula data, and finished product quality inspection data. The relationships between different categories of data are structured and stored in the form of nodes and edges in the full-process knowledge graph. Using the color card code entity in the standardized process parameter query request as the starting query node, a graph traversal is performed in the coating material attribute subgraph of the full process knowledge graph to locate all historical production batch nodes that have used color card code coatings. Among all the identified historical production batch nodes, the nodes with mismatched plate thickness parameters are filtered out using the millimeter value entity as a constraint. For the remaining historical production batch nodes, the geometric similarity between their associated contour description data and the two-dimensional contour point set description entity is further calculated, and a set of historical production batch nodes with a geometric similarity higher than a preset threshold is selected. Calculate the comprehensive score of the historical finished product quality inspection data corresponding to each node in the historical production batch node set; The historical production batch node with the highest comprehensive score is selected as the historical successful production batch with the highest correlation. All upstream process data nodes ending at the historical production batch node are extracted from the full process knowledge graph to form the complete process parameter chain.

4. The aluminum single-panel production control method based on a large model according to claim 3, characterized in that, The aforementioned invocation of the attention-based parameter optimization model, combined with real-time equipment status data from the current production line, dynamically fine-tunes and adaptively corrects each parameter in the extracted complete process parameter chain, including: The real-time equipment status data of the current production line is obtained from the production monitoring system. The real-time equipment status data includes the tool wear coefficient of the CNC punch press, the atomization pressure value of the spray gun of the spraying robot, and the current oven temperature distribution data of the baking oven. The original blanking process parameters, original forming process parameters, original welding process parameters, original grinding process parameters, and original spraying process parameters in the complete process parameter chain are input together with the real-time equipment status data into the parameter optimization model based on the attention mechanism. The attention-based parameter optimization model first calculates the correlation weight between each parameter in the real-time equipment status data and each original process parameter in the complete process parameter chain; Based on the calculated correlation weights, the original process parameters in the complete process parameter chain are weighted and corrected, wherein the original process parameters that are highly correlated with equipment status anomalies will receive a greater degree of correction. After the weighted correction, the parameter optimization model based on the attention mechanism outputs adjusted adaptive blanking process parameters, adaptive molding process parameters, adaptive welding process parameters, adaptive grinding process parameters, and adaptive spraying process parameters, which together constitute the optimal production control instruction set.

5. The aluminum single-panel production control method based on a large model according to claim 4, characterized in that, It also includes the steps of executing the optimal production control instruction set and collecting production feedback data: The adaptive blanking process parameters in the optimal production control instruction set are sent to the CNC cutting machine controller in the blanking workshop to drive the CNC cutting machine to blank aluminum plates according to the specified contour and size. The adaptive forming process parameters in the optimal production control instruction set are sent to the CNC bending machine controller in the forming workshop to drive the CNC bending machine to form the sheet metal according to the specified angle and bending sequence. While the material cutting and forming processes are underway, multi-dimensional feedback data is collected from the actual production process through sensor arrays deployed at key nodes of the production line. The multi-dimensional feedback data includes the smoothness measurement value of the aluminum plate cutting edge, the actual angle deviation value of the bending angle, and the timestamp record of the inter-process transmission. The collected multi-dimensional feedback data is compared in real time with the corresponding expected parameters in the optimal production control instruction set to generate a process-level production status deviation report.

6. The aluminum single-panel production control method based on a large model according to claim 5, characterized in that, It also includes the step of iteratively updating the entire process knowledge graph based on production feedback data: After a complete production batch is completed, the multi-dimensional feedback data collected from subsequent processes such as welding, grinding, and spraying are summarized to form a complete production execution trajectory data for a complete production batch. The final quality inspection report of a complete production batch is associated with the production execution trajectory data, and the production batch is marked as a successful batch or a failed batch. Using a graph structure update algorithm, the labeled production execution trajectory data is added as a new production instance node to the full-process knowledge graph; Establish association edges between the new production instance node and the existing raw material attribute nodes, equipment parameter nodes, and process parameter nodes in the full-process knowledge graph. The weights of the association edges are initialized based on the quality of the production batch results. The weights of all associated edges in the full-process knowledge graph are periodically recalculated and updated. The weight updates are based on the statistical information of recently added successful and failed production instances.

7. The aluminum single-panel production control method based on a large model according to claim 6, characterized in that, It also includes using the production instruction parsing engine to handle production exception events: When the online inspection device on the production line detects a product quality abnormality, it automatically generates a natural language alarm message containing a description of the abnormality. The natural language alarm information is input into the production instruction parsing engine; The production instruction parsing engine parses the natural language alarm information to identify the anomaly type, the process in which the anomaly occurred, and the severity level of the anomaly. Based on the identified anomaly type and the process in which the anomaly occurred, the production instruction parsing engine performs a reverse tracing query in the full process knowledge graph to find production records of similar anomalies in the past and the corrective measures taken. Based on the corrective action records retrieved and combined with the current real-time equipment status data, a targeted anomaly handling suggestion instruction is generated. The anomaly handling suggestion instruction includes equipment parameter adjustment, process step suspension, or material replacement scheme.

8. The aluminum single-panel production control method based on a large model according to claim 7, characterized in that, The calculation of the geometric similarity between the associated contour description data and the two-dimensional contour point set describing entity includes: The contour description data associated with the historical production batch nodes is extracted from the knowledge graph and parsed into a set of historical contour points. The entity described by the two-dimensional contour point set is parsed into the target contour point set; The historical contour point set and the target contour point set are subjected to coordinate normalization processing so that the historical contour point set and the target contour point set are in the same size and coordinate system; The iterative nearest point algorithm is applied to calculate the minimum root mean square error required to align the historical contour point set to the target contour point set. The calculated minimum root mean square error is inversely calculated and normalized to the range of zero to one. The resulting value is the geometric similarity.

9. The aluminum single-panel production control method based on a large model according to claim 8, characterized in that, The periodic recalculation and update of the weights of all associated edges in the full-process knowledge graph includes: Set a fixed weight update cycle, and at the end of each weight update cycle, count all newly added successful production instances and failed production instances within the cycle; For each associated edge in the full-process knowledge graph, find all historical production instance nodes that pass through the associated edge, and filter out the new instances that belong to the current update cycle; Statistically analyze the number of successfully produced instances and the number of failed produced instances among the newly selected instances; Based on the number of successful production instances and the number of failed production instances, the current confidence weight of the associated edge is recalculated according to a preset formula. The calculated current confidence weight is used to replace the original weight value of the associated edge, thereby completing the dynamic update of the knowledge association strength in the full-process knowledge graph.

10. A large-scale aluminum single-panel production control system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the aluminum single-panel production control method based on a large model as described in any one of claims 1 to 9.

Images