Mechanical processing work order checking and intelligent completion method based on multi-agent cooperation

Abstract

The application discloses a kind of based on mechanical processing work order verification and intelligent completion method of multi-agent cooperation, comprising: obtaining the multi-source heterogeneous data of mechanical processing field, process semantic embedding model is trained using knowledge distillation technology, extract hidden space representation and construct processing cognitive atlas;Establish a multi-agent collaborative framework, multi-dimensional analysis is carried out on the original work order data, and a multi-layer semantic hypergraph is generated;Based on the process cognitive atlas and the multi-layer semantic hypergraph, the process causal diagram of the work order to be checked is constructed, the processing result distribution is predicted by intervention calculation, and the risk contribution degree is quantified, the root cause elements are located to generate the work order verification report;With the verification report and physical constraint as input, the diffusion model is used for conditional generation and simulation of different decision preferences, and the work completion decision reasoning path set is obtained;Generate candidate completion work order, optimize abnormal scheme through digital twin rehearsal, finally filter and generate the mechanical processing work order that passes verification and completion, improve the efficiency and accuracy of data import.

Description

Technical Field

[0001] This invention relates to the field of machining technology, and in particular to a machining work order verification and intelligent completion method based on multi-agent collaboration. Background Technology

[0002] In the field of machining production management, process documents and work order data are the core basis for guiding production and manufacturing. Their standardization and data quality directly affect the production efficiency and digital management level of enterprises. With the rapid development of intelligent manufacturing technology, machining enterprises have placed higher demands on the completeness, rationality, and executability of work order data. They not only need to accurately record part information, process flows, equipment resources, and process parameters, but also need to ensure the logical consistency and physical feasibility between various process elements. However, the current work order data processing methods commonly used in the industry are still mainly based on manual entry and rule verification. Faced with multi-source heterogeneous data formats, complex and ever-changing processing scenarios, and the need to migrate massive amounts of historical data, these methods suffer from prominent problems such as low efficiency, high error rates, and poor compatibility, making it difficult to meet the requirements of modern manufacturing for data quality and processing timeliness.

[0003] In existing technologies, some management systems attempt to automate data processing through template import or interface integration. However, due to factors such as inconsistent data formats and significant differences in interface standards across enterprise systems, customized development is costly and maintenance is difficult. More importantly, traditional methods only focus on data format conversion and field mapping, lacking an understanding and verification capability of the process logic itself. They cannot identify implicit process conflicts or potential processing risks in work orders, resulting in repeated manual verification and correction after data import. Furthermore, for missing process information, existing technologies often simply fill in default values ​​or leave blanks, failing to intelligently complete the information based on specific processing scenarios and physical constraints, making it difficult to guarantee the actual executability of the generated work orders. Therefore, how to achieve intelligent parsing of multi-format work order data, in-depth verification of process logic, and physical executability completion of missing information has become a pressing technical challenge to be solved in the digital transformation of the machining industry. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a method for mechanical processing work order verification and intelligent completion based on multi-agent collaboration. Its main purpose is to reduce manual intervention, improve the efficiency and accuracy of data import, and lower the application threshold of digital management systems.

[0005] To achieve the above objectives, the first aspect of this invention provides a method for machining work order verification and intelligent completion based on multi-agent collaboration, comprising:

[0006] Acquire static resource data and dynamic process data in the field of machining, use knowledge distillation technology to transfer the general semantic understanding ability of the large language model to the vertical domain, train a process semantic embedding model, extract the latent space representation of process elements based on the process semantic embedding model, and construct a processing cognitive map.

[0007] A multi-agent collaborative framework is established to obtain raw work order data and import it into the multi-agent collaborative framework to perform multi-dimensional identification and parsing of raw work order data of different categories, generating a multi-layer semantic hypergraph containing explicit fields and implicit associations.

[0008] Construct a process cause-effect graph for the work orders to be inspected, introduce a structural cause-effect model to calculate the distribution of processing results that may be caused by the current work order configuration, and analyze the root causes of potential problems by calculating the contribution of each process element to the processing risk, and generate a work order verification report that includes risk prediction and cause location.

[0009] Based on the work order verification report, the known conditions of the current work order and the physical constraints in the processing cognitive map are used as inputs. The diffusion model is combined with historical successful process instances to generate conditions, and different decision preferences are simulated to generate a set of work order completion decision reasoning paths.

[0010] Based on the work order completion decision reasoning path set, several candidate completion work orders are generated. All candidate completion work orders are subjected to process pre-simulation. The differences between the pre-simulation and the historical similar processing conditions are analyzed and adjusted and screened to generate the final machining work order that has passed the verification and completion.

[0011] In this solution, the acquisition of static resource data and dynamic process data in the field of machining, the transfer of the general semantic understanding capability of a large language model to a vertical domain using knowledge distillation technology, the training of a process semantic embedding model, the extraction of latent space representations of process elements based on the process semantic embedding model, and the construction of a processing cognitive graph specifically include:

[0012] Acquire static resource data and dynamic process data in the field of machining. The static resource data includes machine tool parameters, tool specifications and material properties. The dynamic process data includes spindle load curves, cutting force fluctuation data, vibration spectrum characteristics, equipment operation and maintenance records, unstructured process notes and abnormal handling records during the machining process.

[0013] The collected dynamic process data is aligned with the corresponding static resource data according to the processing timestamp to construct an original training dataset containing the correspondence between equipment, tools, materials, process parameters and actual processing effects.

[0014] Unstructured process notes and anomaly handling records are input into a large language model. The model is guided by prompts to identify and extract process knowledge triples contained in the text. The extracted structured process knowledge is then fused with the original training dataset to form a cross-modal training sample containing text descriptions and physical signals.

[0015] A teacher-student training framework is constructed using knowledge distillation technology. A large language model is used as the teacher model, and a lightweight neural network is used as the student model. The semantic feature distribution output by the teacher model is used as the supervision signal to train the student model to learn to map the process requirements described in the text to a high-dimensional semantic vector space, and finally obtains the process semantic embedding model.

[0016] The cross-modal training samples are imported into the process semantic embedding model to obtain the semantic feature vector of the training samples, and concatenated with the corresponding numerical features of the training samples to form the sample feature vector. The variational autoencoder is used to reduce the dimension and reconstruct the sample feature vector to generate the latent space representation vector of each sample.

[0017] Using latent space representation vectors as node features and process similarity between samples as edge weights, an initial process instance graph is constructed. Structural equation modeling is used to model the variable relationships in the graph. Each process element is treated as an endogenous variable, and the path coefficients between variables are estimated through regression analysis. Causal direction constraints are applied based on domain knowledge to construct a processing cognitive graph.

[0018] In this solution, the establishment of a multi-agent collaborative framework, the acquisition of raw work order data, and the import of this data into the framework for multi-dimensional identification and parsing of different categories of raw work order data, generating a multi-layer semantic hypergraph containing explicit fields and implicit associations, specifically includes:

[0019] A multi-agent collaborative framework is established, which includes a comprehensive parsing agent, a verification agent, and a completion agent. The comprehensive parsing agent includes a drawing parsing sub-agent, a document parsing sub-agent, and an image parsing sub-agent.

[0020] Each agent preloads a process semantic embedding model and a processing cognitive graph. At the same time, a shared memory area is established through a blackboard model to store intermediate parsing results. Agents use a message passing protocol to synchronize task states, forming a distributed collaborative parsing environment.

[0021] In a distributed collaborative parsing environment, data types are automatically identified and assigned to corresponding parsing sub-agents based on file extensions and file header features, enabling multi-dimensional identification and parsing of different categories of raw work order data;

[0022] The drawing parsing sub-agent parses the received drawing file through a conditional generative adversarial network. It retrieves the typical clamping scheme codes of similar parts from the processing cognitive map as conditional vectors. Under the constraints of the conditional vectors, the generator outputs a heat map with marked clamping positioning points and finishing allowances. The discriminator evaluates and selects the inference results with the highest confidence and outputs the semantic parsing results of the drawing.

[0023] The document parsing sub-agent converts tables in the document into key-value pairs and segments paragraph text into independent semantic units. It retrieves similar historical process decision chains in the processing cognitive graph as contextual cues through the process semantic embedding model, generates a complete process decision chain restored from abbreviated instructions, and uses entity relationships in the processing cognitive graph for consistency verification. Finally, it outputs the text semantic parsing result containing the complete decision chain.

[0024] The image parsing sub-agent enhances low-quality scanned documents or on-site photos, generates a corrected image from the frontal view through implicit 3D reconstruction, inputs the corrected image into a multi-granularity feature alignment network to extract multi-granularity features in parallel, performs similarity calculation and Bayesian correction with the process term vectors in the processing cognitive map, and outputs the image semantic parsing results.

[0025] Each parsing sub-agent converts the output parsing results into an intermediate representation format with entities as nodes and associations as hyperedges and publishes it to the shared memory area through the blackboard model. The text attributes of each node are converted into semantic feature vectors through the process semantic embedding model. All nodes and hyperedges of the same original work order are integrated to construct a multi-layer semantic hypergraph.

[0026] In this solution, the process cause-effect graph of the work order to be inspected is constructed. A structural cause-effect model is introduced to calculate the distribution of processing results that may be caused by the current work order configuration. By calculating the contribution of each process element to the processing risk, the root cause factors that may cause potential problems are analyzed, and a work order verification report containing risk prediction and cause location is generated. Specifically, it includes:

[0027] A multi-layer semantic hypergraph is obtained and input into the verification agent. Entity nodes and relation edges related to the work order to be inspected are extracted from the processing cognitive graph. Using the multi-layer semantic hypergraph as the query condition, a directed acyclic graph skeleton is generated by retrieving subgraph structures with causal relationships in the processing cognitive graph through a graph traversal algorithm. The skeleton is then connected with the retrieved entity nodes according to the causal direction to form a process causal graph.

[0028] Based on the process causal graph and the introduction of a structural causal model for intervention calculation, the actual process parameter values ​​configured in the work order to be inspected are assigned as intervention variables to the corresponding nodes in the causal graph. After blocking the confusion of all non-causal paths through the backdoor path adjustment method, the intervention effect is transmitted layer by layer along the causal path to form a predicted distribution of the processing results that may be caused by the current work order configuration.

[0029] The processing risk is defined as the distribution of the predicted processing results that exceeds the threshold of the process specification. Counterfactual intervention is carried out for each process element. The value of the target process element is replaced with the typical value of similar successful process instances in the processing cognitive map, and the intervention calculation is recalculated. The change in risk value before and after the counterfactual intervention is used as the risk marginal contribution score of the target process element.

[0030] Based on the calculated risk marginal contribution scores of each process element, the Shapley value method is used to calculate the weighted average of the risk marginal contribution of all possible combinations of process elements, so as to obtain the independent contribution score of each process element to the processing risk. After sorting, process elements with cumulative contribution scores exceeding the preset threshold are selected as potential root cause elements.

[0031] The identified potential root causes are used as query conditions to retrieve information from the processing cognitive graph. The abnormal pattern descriptions and common solutions associated with each potential root cause are obtained using the process knowledge base stored in the graph. These are then integrated with the processing result prediction distribution and risk contribution ranking to form a work order verification report.

[0032] In this solution, based on the work order verification report, and using the known conditions of the current work order and the physical constraints in the processing cognitive map as input, a diffusion model combined with historical successful process examples is used to generate conditions, and different decision preferences are simulated to generate a set of work order completion decision reasoning paths. Specifically, this includes:

[0033] The completion agent reads the work order verification report from the shared memory area and extracts the physical constraints related to the current work order from the processing cognitive graph, including equipment performance limit parameters, recommended cutting range of tools, material machinability range and process sequence constraints. The extracted physical constraints are encoded as hard boundary conditions.

[0034] Retrieve a set of historical successful process instances that are similar to the current work order context from the processing cognitive graph, use the process parameters of each historical successful process instance as the target learning distribution, and use the known process parameters of the current work order as the condition vector input to the diffusion model for iterative generation;

[0035] A decision preference simulation mechanism based on prospect theory is introduced, which takes processing cost, processing efficiency, processing quality and risk aversion as four core decision dimensions and presets different weight coefficient combinations for each dimension to form a preference vector representing different decision styles.

[0036] In the iterative generation process of the diffusion model, preference vectors representing different decision styles are incorporated into the loss function constraints, so that the generated candidate parameter combinations are shifted toward the direction that conforms to the corresponding decision preferences, resulting in multiple sets of candidate completion schemes corresponding to different decision styles.

[0037] The completion agent records the complete generation trajectory of each candidate completion scheme, including the direction and magnitude of parameter vector changes in each iteration and the basis for adjusting each process element. It also associates the similar historical instance identifiers, physical constraints, and preference vector parameters referenced during the generation process to generate a work order completion decision path set containing complete reasoning paths.

[0038] In this solution, the process of generating several candidate completed work orders based on the work order completion decision reasoning path set, performing process simulation on all candidate completed work orders, analyzing the differences between the simulation and historical similar processing conditions, making adjustments and screenings, and generating the final machining work order that has passed verification and completion, specifically includes:

[0039] The completion agent reads the work order completion decision reasoning path set from the shared memory area, and extracts the equipment 3D model, tool geometric parameters, workpiece blank model and material constitutive parameters from the processing cognitive map. After integration, it constructs an independent process pre-simulation environment for each candidate solution through digital twin technology.

[0040] In the constructed process simulation environment, the full-process machining operation simulation is performed for each group of candidate completion schemes. Within each simulation time step, the stress field distribution, temperature field evolution and machine tool dynamic response in the tool-workpiece contact area are solved. The workpiece shape error, residual stress distribution and tool wear after the process are recorded to generate expected process execution indicators.

[0041] Simultaneously, the sensor time-series data generated by simulation is dynamically compared with the actual data of similar historical processes in the processing cognitive map. By calculating the dynamic time warping distance within the same processing stage, the process execution difference between the current completion scheme and historical successful instances is analyzed to obtain the process execution difference metric.

[0042] When the process execution difference metric exceeds the preset difference threshold or the expected process execution index exceeds the process specification range, the corresponding candidate completion scheme is marked as a scheme to be optimized, and the simulation period and process parameters with anomalies are extracted. The process parameters of the current candidate completion scheme are used as the initial state for scheme optimization.

[0043] A deep deterministic strategy gradient algorithm is used to explore and fine-tune the process parameters in a constructed process simulation environment. The physical constraints in the processing cognitive map are used as the action space boundary. After each exploration, the simulation is rerun and a new set of expected process execution indicators and difference measures are calculated. The optimized candidate completion scheme is output through continuous iterative optimization until it meets the stopping criteria.

[0044] All optimized candidate solutions and candidate solutions that meet the initial requirements are comprehensively ranked. The solution with the lowest process execution difference metric and all expected process execution indicators within the specification range is selected as the preferred completion solution for the corresponding decision style. Finally, the final machining work order that passes the verification and completion is generated and pushed out.

[0045] A second aspect of the present invention provides a computer-readable storage medium comprising a method program for verifying and intelligently completing machining work orders based on multi-agent collaboration. When the method program is executed by a processor, it implements the steps of the method program for verifying and intelligently completing machining work orders based on multi-agent collaboration as described in any of the preceding claims. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0047] Figure 1 The first method flowchart of a machining work order verification and intelligent completion method based on multi-agent collaboration provided in an embodiment of the present invention;

[0048] Figure 2 The flowchart of the second method of a machining work order verification and intelligent completion method based on multi-agent collaboration is provided in an embodiment of the present invention.

[0049] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0050] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0052] Figure 1The first method flowchart of a machining work order verification and intelligent completion method based on multi-agent collaboration provided in an embodiment of the present invention;

[0053] like Figure 1 As shown, the present invention provides a first method flowchart for a machining work order verification and intelligent completion method based on multi-agent collaboration, including:

[0054] S102, acquire static resource data and dynamic process data in the field of machining, use knowledge distillation technology to transfer the general semantic understanding ability of the large language model to the vertical domain, train the process semantic embedding model, extract the latent space representation of process elements based on the process semantic embedding model, and construct a processing cognitive map.

[0055] S104, Establish a multi-agent collaborative framework, acquire original work order data and import it into the multi-agent collaborative framework to perform multi-dimensional identification and parsing of different categories of original work order data, and generate a multi-layer semantic hypergraph containing explicit fields and implicit associations.

[0056] S106, construct the process cause-effect diagram of the work order to be inspected, introduce the structural cause-effect model to calculate the distribution of processing results that may be caused by the current work order configuration, and analyze the root cause factors that may cause potential problems by calculating the contribution of each process element to the processing risk, and generate a work order verification report that includes risk prediction and cause location.

[0057] S108, Based on the work order verification report, using the known conditions of the current work order and the physical constraints in the processing cognitive map as input, the diffusion model is used in combination with historical successful process instances to generate conditions, and different decision preferences are simulated to generate a set of work order completion decision reasoning paths;

[0058] S110: Generate several candidate completion work orders based on the work order completion decision reasoning path set; perform process pre-simulation on all candidate completion work orders; analyze the differences between the pre-simulation and historical similar processing conditions; make adjustments and screenings; and generate the final machining work order that has passed verification and completion.

[0059] Furthermore, in a preferred embodiment of the present invention, the steps of acquiring static resource data and dynamic process data in the field of machining, utilizing knowledge distillation technology to transfer the general semantic understanding capability of a large language model to a vertical domain, training a process semantic embedding model, extracting latent space representations of process elements based on the process semantic embedding model, and constructing a processing cognitive graph specifically include:

[0060] Acquire static resource data and dynamic process data in the field of machining. The static resource data includes machine tool parameters, tool specifications and material properties. The dynamic process data includes spindle load curves, cutting force fluctuation data, vibration spectrum characteristics, equipment operation and maintenance records, unstructured process notes and abnormal handling records during the machining process.

[0061] The collected dynamic process data is aligned with the corresponding static resource data according to the processing timestamp to construct an original training dataset containing the correspondence between equipment, tools, materials, process parameters and actual processing effects.

[0062] Unstructured process notes and anomaly handling records are input into a large language model. The model is guided by prompts to identify and extract process knowledge triples contained in the text. The extracted structured process knowledge is then fused with the original training dataset to form a cross-modal training sample containing text descriptions and physical signals.

[0063] A teacher-student training framework is constructed using knowledge distillation technology. A large language model is used as the teacher model, and a lightweight neural network is used as the student model. The semantic feature distribution output by the teacher model is used as the supervision signal to train the student model to learn to map the process requirements described in the text to a high-dimensional semantic vector space, and finally obtains the process semantic embedding model.

[0064] The cross-modal training samples are imported into the process semantic embedding model to obtain the semantic feature vector of the training samples, and concatenated with the corresponding numerical features of the training samples to form the sample feature vector. The variational autoencoder is used to reduce the dimension and reconstruct the sample feature vector to generate the latent space representation vector of each sample.

[0065] Using latent space representation vectors as node features and process similarity between samples as edge weights, an initial process instance graph is constructed. Structural equation modeling is used to model the variable relationships in the graph. Each process element is treated as an endogenous variable, and the path coefficients between variables are estimated through regression analysis. Causal direction constraints are applied based on domain knowledge to construct a processing cognitive graph.

[0066] It is important to note that the first step is to collect comprehensive and diverse basic data. The collected static resource data primarily originates from the enterprise's internal equipment management system, tool management system, and bill of materials. This includes performance parameters such as machine tool model, machining accuracy range, and spindle power; specifications such as the geometric dimensions, coating materials, and recommended cutting parameters of various tools; and attributes such as the workpiece material grade, hardness range, and heat treatment status. Dynamic process data is acquired in real-time through sensor interfaces deployed on the machine tool's CNC system. Examples include spindle load variation curves recorded during machining, fluctuation data of three-dimensional cutting forces, and vibration spectrum characteristics reflecting machining stability. It also includes tool wear cycles and machining anomaly alarm logs extracted from equipment maintenance records. Furthermore, unstructured process notes and on-site anomaly handling records accumulated by experienced process engineers over long-term practice are also crucial components of the dynamic data, containing a wealth of implicit process knowledge that is difficult to quantify.

[0067] After data acquisition, dynamic process data is aligned with corresponding static resource data according to precise processing timestamps. For example, the spindle load data at a certain moment is associated with the tool specifications used, the workpiece material being processed, and the machine tool model, thereby constructing an original training dataset containing a complete correspondence between equipment, tools, materials, process parameters, and actual processing results. Since unstructured process notes and anomaly handling records contain rich causal relationships and expert experience, they need to be transformed into computable structured knowledge. In practice, this text is input into a large language model, and a carefully designed prompting engineering guides the model to identify and extract process knowledge triples from the text. For example, from the description "using a YG8 tool when machining 304 stainless steel easily leads to built-up edge," the association rules between material, tool, and processing phenomena are extracted. The extracted structured process knowledge is then fused with the aforementioned original training dataset to ultimately form a cross-modal training sample containing text descriptions and physical signals.

[0068] To achieve a lightweight and efficient model for practical deployment, a knowledge distillation technique is employed to construct a teacher-student training framework. A large language model serves as the teacher model, while a lightweight neural network designed for industrial scenarios acts as the student model. The semantic feature distribution output by the teacher model when processing process texts is used as a supervision signal to train the student model to learn how to map the process requirements described in the text to a high-dimensional semantic vector space. This approach results in a process semantic embedding model that retains the semantic understanding capabilities of the teacher model while offering the lightweight advantage of rapid deployment at the edge. Cross-modal training samples are input into the trained process semantic embedding model to obtain the semantic feature vector of each sample. This vector is then concatenated with numerical features such as equipment parameters and tool specifications to form a complete sample feature vector. A variational autoencoder is used to reduce and reconstruct these sample feature vectors. The encoder compresses the high-dimensional features into a low-dimensional latent space, and the decoder restores the original features. After training, the encoder output is extracted as the latent space representation vector for each sample. Finally, the latent space representation vectors are used as node features, and the process similarity between samples is used as edge weights to construct an initial process instance graph. Structural equation modeling is used to model the relationships between variables in the graph, treating each process element as an endogenous variable. Path coefficients between variables are estimated through regression analysis, and causal direction constraints are imposed based on domain knowledge, such as specifying that tool type affects cutting force but is not affected by cutting force itself. Through iterative optimization, a machining cognitive graph that simultaneously represents the logical relationships and physical constraints between process elements is finally constructed.

[0069] Furthermore, in a preferred embodiment of the present invention, the step of establishing a multi-agent collaborative framework, acquiring original work order data, and importing it into the multi-agent collaborative framework for multi-dimensional identification and parsing of different categories of original work order data to generate a multi-layer semantic hypergraph containing explicit fields and implicit associations specifically includes:

[0070] A multi-agent collaborative framework is established, which includes a comprehensive parsing agent, a verification agent, and a completion agent. The comprehensive parsing agent includes a drawing parsing sub-agent, a document parsing sub-agent, and an image parsing sub-agent.

[0071] Each agent preloads a process semantic embedding model and a processing cognitive graph. At the same time, a shared memory area is established through a blackboard model to store intermediate parsing results. Agents use a message passing protocol to synchronize task states, forming a distributed collaborative parsing environment.

[0072] In a distributed collaborative parsing environment, data types are automatically identified and assigned to corresponding parsing sub-agents based on file extensions and file header features, enabling multi-dimensional identification and parsing of different categories of raw work order data;

[0073] The drawing parsing sub-agent parses the received drawing file through a conditional generative adversarial network. It retrieves the typical clamping scheme codes of similar parts from the processing cognitive map as conditional vectors. Under the constraints of the conditional vectors, the generator outputs a heat map with marked clamping positioning points and finishing allowances. The discriminator evaluates and selects the inference results with the highest confidence and outputs the semantic parsing results of the drawing.

[0074] The document parsing sub-agent converts tables in the document into key-value pairs and segments paragraph text into independent semantic units. It retrieves similar historical process decision chains in the processing cognitive graph as contextual cues through the process semantic embedding model, generates a complete process decision chain restored from abbreviated instructions, and uses entity relationships in the processing cognitive graph for consistency verification. Finally, it outputs the text semantic parsing result containing the complete decision chain.

[0075] The image parsing sub-agent enhances low-quality scanned documents or on-site photos, generates a corrected image from the frontal view through implicit 3D reconstruction, inputs the corrected image into a multi-granularity feature alignment network to extract multi-granularity features in parallel, performs similarity calculation and Bayesian correction with the process term vectors in the processing cognitive map, and outputs the image semantic parsing results.

[0076] Each parsing sub-agent converts the output parsing results into an intermediate representation format with entities as nodes and associations as hyperedges and publishes it to the shared memory area through the blackboard model. The text attributes of each node are converted into semantic feature vectors through the process semantic embedding model. All nodes and hyperedges of the same original work order are integrated to construct a multi-layer semantic hypergraph.

[0077] It is important to note that a distributed multi-agent collaborative framework needs to be established first. This framework includes a comprehensive parsing agent, a verification agent, and a completion agent. The comprehensive parsing agent is further subdivided into drawing parsing sub-agents, document parsing sub-agents, and image parsing sub-agents. Each agent is pre-loaded with a pre-trained process semantic embedding model and a constructed processing cognitive map, enabling it to understand knowledge in the mechanical processing domain. Agents share a memory area through a blackboard model to store intermediate parsing results and state information. A lightweight message passing protocol is used for task state synchronization and coordination, thus forming a distributed collaborative parsing environment. When processing raw work order data, the data type is automatically identified based on the file extension and header features, and different file types are assigned to the corresponding parsing sub-agents for parallel processing. For example, CAD drawings or image-format drawing files are sent to the drawing parsing sub-agent; text files such as Word documents, Excel spreadsheets, and editable PDFs are sent to the document parsing sub-agent; and image files such as scanned documents and on-site photographs are sent to the image parsing sub-agent.

[0078] The drawing parsing sub-agent uses a conditional generative adversarial network (GAN) to parse drawing files. This agent retrieves typical clamping schemes for similar parts from the machining cognitive graph and encodes them as conditional vectors input to a generator. Under the constraints of these conditional vectors, the generator outputs a heatmap labeled with clamping points and finishing allowances. A discriminator evaluates the generated results and selects the inference with the highest confidence, ultimately outputting a structured semantic parsing result for the drawing. The document parsing sub-agent, when processing text files, first converts tables in the document into structured key-value pairs and segments paragraph text into independent semantic units. For each semantic unit, a process semantic embedding model retrieves similar historical process decision chains from the machining cognitive graph as contextual cues, guiding the large language model to gradually reconstruct the complete process decision chain from abbreviated instructions, including the process name, process sequence, tool selection criteria, and parameter determination logic. The reconstructed decision chain also needs to be validated for consistency using entity relationships in the machining cognitive graph, such as verifying whether the selected tool is suitable for the current workpiece material. Finally, a text semantic parsing result containing the complete decision chain is output. The image parsing agent is specifically designed to handle low-quality scanned documents or on-site photographs. This agent performs implicit 3D reconstruction on the input image and generates a corrected image from a frontal viewpoint using voxel rendering, effectively eliminating distortions caused by shooting angles and uneven lighting. The corrected image is then input into a multi-granularity feature alignment network. This network extracts multi-granularity features in parallel using a character-level convolutional neural network, a word-level bidirectional long short-term memory network, and a paragraph-level self-attention mechanism. These features are then compared with process term vectors in the processing cognitive graph using similarity calculations and Bayesian corrections, ultimately outputting high-confidence image semantic parsing results. For example, for a photograph of an on-site process card taken at an angle with reflections, this agent can correct the image and accurately identify handwritten process descriptions and tool information.

[0079] After each parsing sub-agent completes its processing, it uniformly converts its parsing results into an intermediate representation format with entities as nodes and relationships as hyperedges, and publishes it to the shared memory region through the blackboard model. Further, it retrieves all intermediate representations of the same original work order from the shared memory region, calls the process semantic embedding model to convert the text attributes of each node into semantic feature vectors, and finally integrates all nodes and hyperedges to construct a multi-layer semantic hypergraph. This hypergraph fully represents all explicit field information contained in the original work order data and the implicit relationships between them, providing a rich data foundation for subsequent work order verification and intelligent completion.

[0080] Furthermore, in a preferred embodiment of the present invention, the construction of the process cause-effect graph of the work order to be inspected, the introduction of a structural cause-effect model to calculate the possible distribution of processing results due to the current work order configuration, and the analysis of the root causes of potential problems by calculating the contribution of each process element to the processing risk, generating a work order verification report containing risk prediction and cause location, specifically includes:

[0081] A multi-layer semantic hypergraph is obtained and input into the verification agent. Entity nodes and relation edges related to the work order to be inspected are extracted from the processing cognitive graph. Using the multi-layer semantic hypergraph as the query condition, a directed acyclic graph skeleton is generated by retrieving subgraph structures with causal relationships in the processing cognitive graph through a graph traversal algorithm. The skeleton is then connected with the retrieved entity nodes according to the causal direction to form a process causal graph.

[0082] Based on the process causal graph and the introduction of a structural causal model for intervention calculation, the actual process parameter values ​​configured in the work order to be inspected are assigned as intervention variables to the corresponding nodes in the causal graph. After blocking the confusion of all non-causal paths through the backdoor path adjustment method, the intervention effect is transmitted layer by layer along the causal path to form a predicted distribution of the processing results that may be caused by the current work order configuration.

[0083] The processing risk is defined as the distribution of the predicted processing results that exceeds the threshold of the process specification. Counterfactual intervention is carried out for each process element. The value of the target process element is replaced with the typical value of similar successful process instances in the processing cognitive map, and the intervention calculation is recalculated. The change in risk value before and after the counterfactual intervention is used as the risk marginal contribution score of the target process element.

[0084] Based on the calculated risk marginal contribution scores of each process element, the Shapley value method is used to calculate the weighted average of the risk marginal contribution of all possible combinations of process elements, so as to obtain the independent contribution score of each process element to the processing risk. After sorting, process elements with cumulative contribution scores exceeding the preset threshold are selected as potential root cause elements.

[0085] The identified potential root causes are used as query conditions to retrieve information from the processing cognitive graph. The abnormal pattern descriptions and common solutions associated with each potential root cause are obtained using the process knowledge base stored in the graph. These are then integrated with the processing result prediction distribution and risk contribution ranking to form a work order verification report.

[0086] It should be noted that the verification agent retrieves the previously constructed multi-layer semantic hypergraph from the shared memory region of the blackboard model. This hypergraph fully contains the explicit field information of the current work order to be inspected, as well as the implicit relationships between these fields. The verification agent uses entity nodes in the multi-layer semantic hypergraph as the starting point for queries, extracting all entity nodes and relational edges related to the work order from the processing cognitive graph. Specifically, each node in the hypergraph is used as a query condition. A graph traversal algorithm is used to retrieve subgraph structures containing these nodes and exhibiting causal relationships within the processing cognitive graph. Simultaneously, combined with the directed acyclic graph skeleton pre-learned by the structural equation model in the processing cognitive graph, the retrieved entity nodes are connected according to causal directions, ultimately forming a process causal graph that clearly represents the causal dependencies between various process elements of the work order to be inspected. For example, for a work order containing 45# steel and an external turning process, the process causal graph will show causal chains such as material hardness affecting cutting force, cutting force affecting tool wear, and tool wear affecting surface roughness. Based on the obtained process causal graph, the verification agent introduces a structural causal model for intervention calculations. The actual process parameter values ​​configured in the work order to be inspected are assigned as intervention variables to the corresponding nodes in the causal graph. For example, the actual set cutting speed of 800 rpm is assigned to the speed node. To accurately calculate the intervention effect, all non-causal paths that may cause confusion need to be identified and blocked using a backdoor path adjustment method, ensuring that only causal paths from the intervention variable to the outcome variable are retained. The intervention effect is then propagated layer by layer along these causal paths using path coefficients, ultimately calculating the expected value and confidence interval of the key processing outcome variable under a given work order configuration, forming a predicted distribution of the processing outcome that the current work order configuration may lead to.

[0087] Based on the predicted distribution of the processing results, the verification agent defines the portion exceeding the process specification threshold as processing risk. For example, if the process specification requires that the surface roughness not exceed Ra 2.0 micrometers, then the probability portion exceeding this threshold in the predicted distribution constitutes risk. To quantify the specific contribution of each process element to the risk, counterfactual intervention is performed on each process element individually. Specifically, while keeping the values ​​of other elements unchanged, the value of the target process element is replaced with a typical value from similar successful process instances in the processing cognitive map. For example, the feed rate of the current work order is adjusted from 0.25 mm / revolution to 0.18 mm / revolution, which is common in successful instances. Then, the structural causal model is rerun for intervention calculation. The change in risk value before and after the counterfactual intervention is compared. This change is the marginal contribution score of the element to the risk, reflecting the magnitude of the element's influence on risk formation under the current configuration. Considering that there may be interactions between process elements, the marginal contribution score obtained by simply replacing them one by one cannot eliminate the influence of the element order. Therefore, the Shapley value method is needed to perform a weighted average of the marginal contributions of all possible combinations of process elements. This method, derived from cooperative game theory, can fairly distribute the total risk to each element. By calculating the marginal contribution of each element under all subset combinations and taking the average, the independent contribution score of each process element to the processing risk is finally obtained. The process elements are sorted from high to low contribution, and the set of elements with a cumulative contribution exceeding a preset threshold, such as 80%, is selected as potential root cause elements. Finally, the identified potential root cause elements are used as query conditions to retrieve data from the processing cognitive graph. The abnormal pattern description and common solutions associated with each potential root cause element are obtained using the process knowledge base stored in the graph. All the above information is integrated with the processing result prediction distribution and risk contribution ranking to finally form a complete work order verification report.

[0088] Furthermore, in a preferred embodiment of the present invention, the step of generating conditions based on the work order verification report, using the known conditions of the current work order and the physical constraints in the processing cognitive graph as input, employing a diffusion model combined with historical successful process instances to generate conditions, and simulating different decision preferences to generate a set of work order completion decision reasoning paths, specifically includes:

[0089] The completion agent reads the work order verification report from the shared memory area and extracts the physical constraints related to the current work order from the processing cognitive graph, including equipment performance limit parameters, recommended cutting range of tools, material machinability range and process sequence constraints. The extracted physical constraints are encoded as hard boundary conditions.

[0090] Retrieve a set of historical successful process instances that are similar to the current work order context from the processing cognitive graph, use the process parameters of each historical successful process instance as the target learning distribution, and use the known process parameters of the current work order as the condition vector input to the diffusion model for iterative generation;

[0091] A decision preference simulation mechanism based on prospect theory is introduced, which takes processing cost, processing efficiency, processing quality and risk aversion as four core decision dimensions and presets different weight coefficient combinations for each dimension to form a preference vector representing different decision styles.

[0092] In the iterative generation process of the diffusion model, preference vectors representing different decision styles are incorporated into the loss function constraints, so that the generated candidate parameter combinations are shifted toward the direction that conforms to the corresponding decision preferences, resulting in multiple sets of candidate completion schemes corresponding to different decision styles.

[0093] The completion agent records the complete generation trajectory of each candidate completion scheme, including the direction and magnitude of parameter vector changes in each iteration and the basis for adjusting each process element. It also associates the similar historical instance identifiers, physical constraints, and preference vector parameters referenced during the generation process to generate a work order completion decision path set containing complete reasoning paths.

[0094] It should be noted that the completion agent first reads the work order verification report from the shared memory area of ​​the blackboard model. This report details the missing fields, risk factors, and contribution ranking of each factor in the current work order. Simultaneously, the completion agent extracts physical constraints related to the current work order from the machining cognitive graph. These constraints include extreme parameters of equipment performance, such as the maximum spindle speed and power; cutting ranges recommended by tool manufacturers, such as cutting speed and feed rate ranges; material machinability ranges, such as hardness tolerance ranges and heat treatment requirements; and implicit process sequence constraints in the process specifications, such as the roughing-to-finishing machining sequence. These physical constraints are encoded as hard boundary conditions in the diffusion model generation process, ensuring that all subsequently generated completion solutions remain within the physically realizable range. Based on the above input conditions, the completion agent retrieves a set of historical successful process instances similar to the current work order context from the machining cognitive graph. The retrieval process uses the known process parameters of the current work order as the query vector, calculates the semantic similarity with historical instances in the graph through a process semantic embedding model, and selects the top few instances with the highest similarity. The complete process parameter vectors of these historical successes are used as the target distribution learned during the training phase of the diffusion model. The task of the diffusion model is to generate a combination of process parameters that is as close as possible to this target distribution, given the known conditions of the current work order.

[0095] Considering the varying preferences for process objectives across different production scenarios, the completion agent introduces a decision preference simulation mechanism based on prospect theory. Processing cost, processing efficiency, processing quality, and risk aversion are used as four core decision dimensions, with different weighting coefficient combinations preset for each dimension, forming preference vectors representing different decision styles. Aggressive preferences assign higher weights to processing efficiency while tolerating some risk; robust preferences assign higher weights to risk aversion while accepting slightly higher costs; and balanced preferences take a middle ground among the dimensions. These preference vectors guide the model to generate candidate solutions that conform to different decision styles during subsequent generation. During the iterative generation process of the diffusion model, the preference vectors representing different decision styles are incorporated as constraints into the loss function of each denoising step. The model iterates from pure noise, simultaneously constrained by the known conditions of the current work order, physical boundary conditions, and the guidance of the preference vectors at each step, causing the generated candidate parameter combinations to shift towards the direction that aligns with the corresponding decision preferences. The completion agent fully records the generation process of each set of candidate completion solutions, including the direction and magnitude of parameter vector changes in each iteration from initial noise to final parameters, as well as the basis for adjusting each process element in each denoising step. This information is linked with similar historical instance identifiers referenced during the generation process, the physical constraints used, and preference vector parameters to construct a work order completion decision path set containing complete reasoning paths. This path set not only includes the final completion parameter values ​​but also records the reasoning basis and reference sources for each decision step, providing completely transparent and traceable information for subsequent manual review.

[0096] Figure 2 The flowchart of the second method of a machining work order verification and intelligent completion method based on multi-agent collaboration is provided in an embodiment of the present invention.

[0097] like Figure 2 As shown, the present invention provides a second method flowchart for a machining work order verification and intelligent completion method based on multi-agent collaboration, including:

[0098] S202, the completion agent reads the work order completion decision reasoning path set from the shared memory area, and extracts the equipment 3D model, tool geometric parameters, workpiece blank model and material constitutive parameters from the processing cognitive map. After integration, it constructs an independent process pre-simulation environment for each candidate solution through digital twin technology.

[0099] S204, in the constructed process simulation environment, performs full-process machining simulation for each group of candidate completion schemes, solves the stress field distribution, temperature field evolution and machine tool dynamic response in the tool-workpiece contact area within each simulation time step, records the workpiece morphology error, residual stress distribution and tool wear after the process is completed, and generates expected process execution indicators.

[0100] S206, At the same time, the sensor time series data generated by simulation is dynamically compared with the actual data of similar historical processes in the processing cognitive map. By calculating the dynamic time warping distance in the same processing stage, the process execution difference between the current completion scheme and historical successful instances is analyzed to obtain the process execution difference metric.

[0101] S208, when the process execution difference metric exceeds the preset difference threshold or the expected process execution index exceeds the process specification range, the corresponding candidate completion scheme is marked as a scheme to be optimized, and the simulation period and process parameters with anomalies are extracted, and the process parameters of the current candidate completion scheme are used as the initial state for scheme optimization.

[0102] S210 employs a deep deterministic strategy gradient algorithm to explore and fine-tune process parameters in a constructed process pre-simulation environment. It uses physical constraints in the processing cognitive map as the action space boundary. After each exploration, it reruns the simulation and calculates a new set of expected process execution indicators and difference metrics. Through continuous iterative optimization, it outputs the optimized candidate completion scheme until it meets the stopping criteria.

[0103] S212, comprehensively sort all optimized candidate solutions and candidate solutions that meet the requirements from the beginning, select the solution with the lowest process execution difference metric value and all expected process execution indicators within the specification range as the preferred completion solution corresponding to the decision style, and finally generate the final machining work order that has passed the verification and completion and push it.

[0104] It should be noted that the completion agent first reads the work order completion decision reasoning path set from the shared memory area of ​​the blackboard model. This path set contains multiple sets of candidate completion schemes corresponding to different decision styles, as well as the complete generation trajectory of each scheme. Simultaneously, the completion agent extracts the equipment 3D model, tool geometry parameters, workpiece blank model, and material constitutive parameters related to the current work order from the machining cognitive graph. For example, it extracts the 3D model of the machining center, the geometry of the end mill, the shape of the workpiece blank, and the Johnson-Cook constitutive model parameters of the material. This information is then integrated and input into the digital twin system. The system matches a multibody dynamics model based on the equipment model, calls the cutting force empirical model based on the tool parameters, and applies the thermo-coupling field constitutive equations to the workpiece material, thereby constructing an independent process pre-simulation environment for each candidate scheme.

[0105] In the constructed process simulation environment, a full-process machining simulation was performed on each candidate completion scheme. Within each simulation time step, the stress field distribution, temperature field evolution, and dynamic response of the machine tool spindle and feed system in the tool-workpiece contact area were simultaneously solved. The workpiece morphology error, residual stress distribution, and tool wear were recorded after each operation. These indicators were summarized as the expected process execution indicators for the scheme, such as predicting workpiece dimensional deviation, surface residual stress, and tool wear after a certain operation. During the simulation, real-time generated sensor time-series data, such as spindle power curves, cutting force fluctuation signals, and vibration spectrum characteristics, were dynamically compared with historical actual machining data of similar processes stored in the machining cognitive map. By calculating the dynamic time warping distance between the two time-series sequences within the same machining stage, a process execution difference metric between the current completion scheme and historical successful instances was obtained. This value quantifies the degree of deviation between simulation expectations and actual experience.

[0106] When the process execution difference metric exceeds a preset difference threshold or any indicator in the expected process execution indicator set exceeds the process specification range, the monitoring agent marks the corresponding candidate completion scheme as a scheme to be optimized, extracts the key simulation period and specific process parameters that caused the anomaly, and starts the optimization process with the process parameters of the current candidate completion scheme as the initial state. A deep deterministic strategy gradient algorithm is used to conduct multiple rounds of fine-tuning exploration of the process parameters in the process pre-simulation environment, using the physical constraints in the processing cognitive graph as the action space boundary. After each exploration, the simulation is rerun and a new set of expected process execution indicators and difference metrics are calculated. Optimization is iterated until the difference metric drops below the preset threshold and all expected indicators return to the specification range, or until the maximum number of iterations is reached, at which point the optimized candidate completion scheme is output. Finally, all optimized candidate schemes and candidate schemes that meet the initial requirements are comprehensively ranked, and the scheme with the lowest process execution difference metric and all expected process execution indicators within the specification range is selected as the preferred completion scheme for the corresponding decision style. The preferred schemes for different decision styles are displayed side-by-side to the user, thereby achieving the beneficial effect of assisting production management.

[0107] Furthermore, the method for machining work order verification and intelligent completion based on multi-agent collaboration provided by this invention also includes the following steps:

[0108] The final machining work order and its complete simulation trajectory, process execution expected index set and optimization iteration record generated in the digital twin pre-simulation stage are packaged and encapsulated into a standardized process case. The latent space representation vector of the case is extracted through the process semantic embedding model and stored in the case library of the processing cognitive graph. At the same time, the association edge between the case and the existing entity nodes in the graph is established.

[0109] When a new work order is received that differs from the historical processing scenario, the completion agent retrieves the set of historical process cases with the highest semantic similarity to the current work order from the processing cognitive graph, and selects candidate reference cases by calculating the Euclidean distance between the latent space representation vectors.

[0110] The meta-learning-based fast adaptation algorithm selects candidate reference cases as source tasks, takes the known conditions of the current new part order as the target task, performs a small number of gradient updates on the pre-trained diffusion model, so that the model parameters can quickly adapt to the process distribution characteristics of the new scenario and generate initial candidate completion schemes.

[0111] The generated initial candidate completion scheme is input into the digital twin environment for rapid pre-simulation, and the pre-simulation results are compared with the actual processing effect of the candidate reference cases to identify the process deviation characteristics caused by scene differences. The process deviation characteristics are then backpropagated to the embedding layer of the diffusion model to correct the scene offset.

[0112] The modified diffusion model is applied to the complete completion process of the current new part order to generate an order completion solution adapted to the new scenario. The complete process of this cross-scenario migration, including source case identification, gradient update amount, deviation characteristics and correction records, is stored in the processing cognitive graph as a new knowledge path, realizing the migration, reuse and accumulation of process knowledge between different production scenarios.

[0113] It should be noted that after verifying and completing a single work order, the generated final machining work order, along with the complete simulation trajectory, expected process execution index set, and optimization iteration records generated during the digital twin pre-simulation phase, are uniformly encapsulated to form a standardized process case. The latent space representation vector of this case is extracted using a process semantic embedding model, and this vector is stored as a new node in the case library of the machining cognitive graph. Simultaneously, association edges are established between this node and existing entity nodes such as equipment, tools, and materials in the graph. For example, the blade case is associated with the five-axis machine tool node, ceramic tool node, and high-temperature alloy material node in the graph, integrating the case into the overall knowledge network. When a new work order is received that differs from historical machining scenarios—for example, the same blade machining but with the material changed from high-temperature alloy to titanium alloy, or a different equipment model—the completion agent needs to draw on historical experience but must adapt to the new scenario. In this case, the agent uses the known conditions of the new work order as the query vector and retrieves the set of historical process cases with the highest semantic similarity from the case library of the machining cognitive graph using the process semantic embedding model. Similarity is calculated by comparing the Euclidean distance between latent space representation vectors; the smaller the distance, the closer the process characteristics. For example, the closest examples to the current titanium alloy blade work order might be previous high-temperature alloy blade examples, or another blade example of a different model but with a similar structure. These selected examples will serve as a reference basis for subsequent transfer learning.

[0114] A meta-learning-based fast adaptation algorithm is introduced to achieve efficient knowledge transfer. The core idea of ​​meta-learning is to enable the model to learn quickly. Specifically, this involves using selected candidate reference cases as the source task and the known conditions of the new work order as the target task, performing a small number of gradient updates on a pre-trained diffusion model. This process does not retrain the model; instead, through a few gradient descent steps, the model parameters are quickly adjusted from the process distribution characteristics of the source task to adapt to the process distribution characteristics of the new scenario. For example, a pre-trained model that is good at generating cutting parameters for high-temperature alloys can quickly master the cutting characteristics of titanium alloys after a small number of gradient updates, generating initial candidate completion schemes that meet the requirements of titanium alloy processing. This mechanism avoids the need for retraining with large amounts of historical data every time a new scenario is encountered. To verify the transfer effect and correct potential biases, the generated initial candidate completion schemes are input into a digital twin environment for rapid pre-simulation, and the pre-simulation results are compared item by item with the actual processing effects of the candidate reference cases. By identifying process deviations caused by scenario differences—for example, the significant difference between the cutting force fluctuation pattern in the titanium alloy scheme simulation and the high-temperature alloy reference case, which is due to different material properties—the identified deviation features are transmitted to the embedding layer of the diffusion model through a backpropagation mechanism to correct the model's scenario offset, making the model's understanding of the current scenario more accurate. Finally, the corrected diffusion model is applied to the complete completion process of the current new work order, generating a work order completion scheme that is fully adapted to the new scenario. Simultaneously, the entire process of this cross-scenario migration is recorded, including the source case identifiers used, the specific values ​​of gradient updates, the identified deviation features, and the final correction records. This information is stored as a new knowledge path in the processing cognitive graph. This path not only enriches the case connections in the graph but also provides directly referable migration experience for future similar scenarios, realizing the efficient transfer, reuse, and continuous accumulation and evolution of process knowledge across different production scenarios.

[0115] In another aspect, the present invention provides a computer-readable storage medium including a method program for machining work order verification and intelligent completion based on multi-agent collaboration. When the method program is executed by a processor, it implements the steps of the method program for machining work order verification and intelligent completion based on multi-agent collaboration as described in any of the preceding claims.

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for machining work order verification and intelligent completion based on multi-agent collaboration, characterized in that, include: Acquire static resource data and dynamic process data in the field of machining, use knowledge distillation technology to transfer the general semantic understanding ability of the large language model to the vertical domain, train a process semantic embedding model, extract the latent space representation of process elements based on the process semantic embedding model, and construct a processing cognitive map. A multi-agent collaborative framework is established to obtain raw work order data and import it into the multi-agent collaborative framework to perform multi-dimensional identification and parsing of raw work order data of different categories, generating a multi-layer semantic hypergraph containing explicit fields and implicit associations. Construct a process cause-effect graph for the work orders to be inspected, introduce a structural cause-effect model to calculate the distribution of processing results that may be caused by the current work order configuration, and analyze the root causes of potential problems by calculating the contribution of each process element to the processing risk, and generate a work order verification report that includes risk prediction and cause location. Based on the work order verification report, the known conditions of the current work order and the physical constraints in the processing cognitive map are used as inputs. The diffusion model is combined with historical successful process instances to generate conditions, and different decision preferences are simulated to generate a set of work order completion decision reasoning paths. Based on the work order completion decision reasoning path set, several candidate completion work orders are generated. All candidate completion work orders are subjected to process pre-simulation. The differences between the pre-simulation and the historical similar processing conditions are analyzed and adjusted and screened to generate the final machining work order that has passed the verification and completion.

2. The method for machining work order verification and intelligent completion based on multi-agent collaboration as described in claim 1, characterized in that, The process involves acquiring static resource data and dynamic process data in the machining field, utilizing knowledge distillation technology to transfer the general semantic understanding capabilities of a large language model to a vertical domain, training a process semantic embedding model, extracting latent space representations of process elements based on the process semantic embedding model, and constructing a processing cognitive graph. Specifically, this includes: Acquire static resource data and dynamic process data in the field of machining. The static resource data includes machine tool parameters, tool specifications and material properties. The dynamic process data includes spindle load curves, cutting force fluctuation data, vibration spectrum characteristics, equipment operation and maintenance records, unstructured process notes and abnormal handling records during the machining process. The collected dynamic process data is aligned with the corresponding static resource data according to the processing timestamp to construct an original training dataset containing the correspondence between equipment, tools, materials, process parameters and actual processing effects. Unstructured process notes and anomaly handling records are input into a large language model. The model is guided by prompts to identify and extract process knowledge triples contained in the text. The extracted structured process knowledge is then fused with the original training dataset to form a cross-modal training sample containing text descriptions and physical signals. A teacher-student training framework is constructed using knowledge distillation technology. A large language model is used as the teacher model, and a lightweight neural network is used as the student model. The semantic feature distribution output by the teacher model is used as the supervision signal to train the student model to learn to map the process requirements described in the text to a high-dimensional semantic vector space, and finally obtains the process semantic embedding model. The cross-modal training samples are imported into the process semantic embedding model to obtain the semantic feature vector of the training samples, and concatenated with the corresponding numerical features of the training samples to form the sample feature vector. The variational autoencoder is used to reduce the dimension and reconstruct the sample feature vector to generate the latent space representation vector of each sample. Using latent space representation vectors as node features and process similarity between samples as edge weights, an initial process instance graph is constructed. Structural equation modeling is used to model the variable relationships in the graph. Each process element is treated as an endogenous variable, and the path coefficients between variables are estimated through regression analysis. Causal direction constraints are applied based on domain knowledge to construct a processing cognitive graph.

3. The method for machining work order verification and intelligent completion based on multi-agent collaboration according to claim 1, characterized in that, The establishment of a multi-agent collaborative framework involves acquiring raw work order data and importing it into the framework for multi-dimensional identification and parsing of different categories of raw work order data, generating a multi-layer semantic hypergraph containing explicit fields and implicit associations. Specifically, this includes: A multi-agent collaborative framework is established, which includes a comprehensive parsing agent, a verification agent, and a completion agent. The comprehensive parsing agent includes a drawing parsing sub-agent, a document parsing sub-agent, and an image parsing sub-agent. Each agent preloads a process semantic embedding model and a processing cognitive graph. At the same time, a shared memory area is established through a blackboard model to store intermediate parsing results. Agents use a message passing protocol to synchronize task states, forming a distributed collaborative parsing environment. In a distributed collaborative parsing environment, data types are automatically identified and assigned to corresponding parsing sub-agents based on file extensions and file header features, enabling multi-dimensional identification and parsing of different categories of raw work order data; The drawing parsing sub-agent parses the received drawing file through a conditional generative adversarial network. It retrieves the typical clamping scheme codes of similar parts from the processing cognitive map as conditional vectors. Under the constraints of the conditional vectors, the generator outputs a heat map with marked clamping positioning points and finishing allowances. The discriminator evaluates and selects the inference results with the highest confidence and outputs the semantic parsing results of the drawing. The document parsing sub-agent converts tables in the document into key-value pairs and segments paragraph text into independent semantic units. It retrieves similar historical process decision chains in the processing cognitive graph as contextual cues through the process semantic embedding model, generates a complete process decision chain restored from abbreviated instructions, and uses entity relationships in the processing cognitive graph for consistency verification. Finally, it outputs the text semantic parsing result containing the complete decision chain. The image parsing sub-agent enhances low-quality scanned documents or on-site photos, generates a corrected image from the frontal view through implicit 3D reconstruction, inputs the corrected image into a multi-granularity feature alignment network to extract multi-granularity features in parallel, performs similarity calculation and Bayesian correction with the process term vectors in the processing cognitive map, and outputs the image semantic parsing results. Each parsing sub-agent converts the output parsing results into an intermediate representation format with entities as nodes and associations as hyperedges and publishes it to the shared memory area through the blackboard model. The text attributes of each node are converted into semantic feature vectors through the process semantic embedding model. All nodes and hyperedges of the same original work order are integrated to construct a multi-layer semantic hypergraph.

4. The method for machining work order verification and intelligent completion based on multi-agent collaboration according to claim 1, characterized in that, The process cause-effect diagram of the work order to be inspected is constructed by introducing a structural cause-effect model to calculate the possible distribution of processing results due to the current work order configuration. Furthermore, by calculating the contribution of each process element to processing risk, the root causes of potential problems are analyzed, and a work order verification report containing risk prediction and cause localization is generated. Specifically, this includes: A multi-layer semantic hypergraph is obtained and input into the verification agent. Entity nodes and relation edges related to the work order to be inspected are extracted from the processing cognitive graph. Using the multi-layer semantic hypergraph as the query condition, a directed acyclic graph skeleton is generated by retrieving subgraph structures with causal relationships in the processing cognitive graph through a graph traversal algorithm. The skeleton is then connected with the retrieved entity nodes according to the causal direction to form a process causal graph. Based on the process causal graph and the introduction of a structural causal model for intervention calculation, the actual process parameter values ​​configured in the work order to be inspected are assigned as intervention variables to the corresponding nodes in the causal graph. After blocking the confusion of all non-causal paths through the backdoor path adjustment method, the intervention effect is transmitted layer by layer along the causal path to form a predicted distribution of the processing results that may be caused by the current work order configuration. The processing risk is defined as the distribution of the predicted processing results that exceeds the threshold of the process specification. Counterfactual intervention is carried out for each process element. The value of the target process element is replaced with the typical value of similar successful process instances in the processing cognitive map, and the intervention calculation is recalculated. The change in risk value before and after the counterfactual intervention is used as the risk marginal contribution score of the target process element. Based on the calculated risk marginal contribution scores of each process element, the Shapley value method is used to calculate the weighted average of the risk marginal contribution of all possible combinations of process elements, so as to obtain the independent contribution score of each process element to the processing risk. After sorting, process elements with cumulative contribution scores exceeding the preset threshold are selected as potential root cause elements. The identified potential root causes are used as query conditions to retrieve information from the processing cognitive graph. The abnormal pattern descriptions and common solutions associated with each potential root cause are obtained using the process knowledge base stored in the graph. These are then integrated with the processing result prediction distribution and risk contribution ranking to form a work order verification report.

5. The method for machining work order verification and intelligent completion based on multi-agent collaboration according to claim 1, characterized in that, Based on the work order verification report, and using the known conditions of the current work order and the physical constraints in the processing cognitive map as input, a diffusion model combined with historical successful process instances is used to generate conditions, and different decision preferences are simulated to generate a set of work order completion decision reasoning paths, specifically including: The completion agent reads the work order verification report from the shared memory area and extracts the physical constraints related to the current work order from the processing cognitive graph, including equipment performance limit parameters, recommended cutting range of tools, material machinability range and process sequence constraints. The extracted physical constraints are encoded as hard boundary conditions. Retrieve a set of historical successful process instances that are similar to the current work order context from the processing cognitive graph, use the process parameters of each historical successful process instance as the target learning distribution, and use the known process parameters of the current work order as the condition vector input to the diffusion model for iterative generation; A decision preference simulation mechanism based on prospect theory is introduced, which takes processing cost, processing efficiency, processing quality and risk aversion as four core decision dimensions and presets different weight coefficient combinations for each dimension to form a preference vector representing different decision styles. In the iterative generation process of the diffusion model, preference vectors representing different decision styles are incorporated into the loss function constraints, so that the generated candidate parameter combinations are shifted toward the direction that conforms to the corresponding decision preferences, resulting in multiple sets of candidate completion schemes corresponding to different decision styles. The completion agent records the complete generation trajectory of each candidate completion scheme, including the direction and magnitude of parameter vector changes in each iteration and the basis for adjusting each process element. It also associates the similar historical instance identifiers, physical constraints, and preference vector parameters referenced during the generation process to generate a work order completion decision path set containing complete reasoning paths.

6. The method for machining work order verification and intelligent completion based on multi-agent collaboration according to claim 1, characterized in that, The process of generating several candidate completed work orders based on the work order completion decision reasoning path set, performing process simulation on all candidate completed work orders, analyzing the differences between the simulation and historical similar processing conditions, making adjustments and screenings, and generating the final machining work order that passes verification and completion, specifically includes: The completion agent reads the work order completion decision reasoning path set from the shared memory area, and extracts the equipment 3D model, tool geometric parameters, workpiece blank model and material constitutive parameters from the processing cognitive map. After integration, it constructs an independent process pre-simulation environment for each candidate solution through digital twin technology. In the constructed process simulation environment, the full-process machining operation simulation is performed for each group of candidate completion schemes. Within each simulation time step, the stress field distribution, temperature field evolution and machine tool dynamic response in the tool-workpiece contact area are solved. The workpiece shape error, residual stress distribution and tool wear after the process are recorded to generate expected process execution indicators. Simultaneously, the sensor time-series data generated by simulation is dynamically compared with the actual data of similar historical processes in the processing cognitive map. By calculating the dynamic time warping distance within the same processing stage, the process execution difference between the current completion scheme and historical successful instances is analyzed to obtain the process execution difference metric. When the process execution difference metric exceeds the preset difference threshold or the expected process execution index exceeds the process specification range, the corresponding candidate completion scheme is marked as a scheme to be optimized, and the simulation period and process parameters with anomalies are extracted. The process parameters of the current candidate completion scheme are used as the initial state for scheme optimization. A deep deterministic strategy gradient algorithm is used to explore and fine-tune the process parameters in a constructed process simulation environment. The physical constraints in the processing cognitive map are used as the action space boundary. After each exploration, the simulation is rerun and a new set of expected process execution indicators and difference measures are calculated. The optimized candidate completion scheme is output through continuous iterative optimization until it meets the stopping criteria. All optimized candidate solutions and candidate solutions that meet the initial requirements are comprehensively ranked. The solution with the lowest process execution difference metric and all expected process execution indicators within the specification range is selected as the preferred completion solution for the corresponding decision style. Finally, the final machining work order that passes the verification and completion is generated and pushed out.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a method program for verifying and intelligently completing machining work orders based on multi-agent collaboration. When the method program is executed by a processor, it implements the steps of the method program for verifying and intelligently completing machining work orders based on multi-agent collaboration as described in any one of claims 1 to 6.

Images