Engineering drawing process splitting method and device, product and medium

Abstract

The invention discloses a process splitting method and device for an engineering drawing, a product and a medium, and the process splitting method comprises the steps: receiving the engineering drawing, calling a multi-modal large model to carry out process splitting on the engineering drawing, and generating original process execution data; analyzing to obtain a query statement; retrieving a plurality of sub-knowledge bases in the hierarchical knowledge base until response information is retrieved; optimization is carried out, and optimized process execution data are generated; traversing and retrieving other sub-knowledge bases to obtain comparison information, and judging whether the response information conflicts with the comparison information or not; and if the judgment result is yes, outputting the conflict part of the response information and the comparison information along with the optimized process execution data. By adopting the scheme, the illusion problem of the multi-modal large model can be effectively inhibited, so that the finally output process execution data not only has higher accuracy and credibility, but also can adapt to the differentiated requirements of different factories and customers.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and process design technology, and in particular to a method, device, product, and medium for breaking down engineering drawings into processes. Background Technology

[0002] In the industrial manufacturing sector, breaking down product engineering drawings, such as CAD drawings, into specific processing steps (hereinafter referred to as "order breakdown") is a core task. Traditional manual order breakdown methods are inefficient and not conducive to standardized management. With the development of artificial intelligence, utilizing general multimodal large models in enterprise information systems to understand engineering drawings and automatically generate specific processing steps to complete order breakdown is an important research direction for those skilled in the art.

[0003] In existing technologies, inputting engineering drawings into a general multimodal large model and directly outputting breakdown solutions has significant technical drawbacks: the knowledge held by the general large model mainly comes from publicly available general data and lacks in-depth professional knowledge specific to the manufacturing field. Therefore, when dealing with drawings involving complex process logic, special tolerance requirements, or multi-equipment coordination, the model is prone to "illusions," that is, generating missing elements and unusable process steps, resulting in the reliability of the planning results failing to meet the requirements of industrial production.

[0004] Those skilled in the art have attempted to introduce Retrieval-Augmented Generation (RAG) technology to reduce the impact of "illusions" on the output results. However, when processing complex engineering drawings, this approach results in fragmented search results, making it difficult to output complete and reasonable breakdown results. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide at least one method for decomposing engineering drawings into work processes. By generating query statements and retrieving multiple sub-knowledge bases in a hierarchical knowledge base according to priority, the original work process execution data is optimized. This effectively suppresses the influence of the illusion generated by the general multimodal large model on the work process decomposition results, resulting in higher completeness and practicality of the final output work process execution data.

[0006] In a first aspect, the present invention provides a method for decomposing engineering drawings into work processes, the method comprising: Receive engineering drawings of the product to be processed, and call a pre-trained multimodal large model to decompose the engineering drawings into processes, generating original process execution data, which includes multiple process steps; The multimodal large model is invoked to analyze the original process execution data to obtain a query statement, which is generated by the multimodal large model for at least one of the multiple process steps. Based on the query statement, multiple sub-knowledge bases in the hierarchical knowledge base are retrieved sequentially in descending order of priority until a response is obtained. The multiple sub-knowledge bases have different priorities. The multimodal large model is invoked, and the original process execution data is optimized based on the retrieved response information to generate optimized process execution data; The remaining sub-knowledge bases are traversed and retrieved to obtain comparison information, and it is determined whether the response information conflicts with the comparison information. The remaining sub-knowledge bases refer to all sub-knowledge bases other than the response information. If the judgment result is yes, then the conflict between the response information and the comparison information will be output along with the optimized process execution data.

[0007] Optionally, the plurality of sub-knowledge bases include: The customer preference sub-knowledge base records the preference information of customers who have commissioned the products to be processed. The factory information sub-knowledge base records information about the factory currently producing the product to be processed. The domain sub-knowledge base records general information about the product category to which the product to be processed belongs; The customer preference sub-knowledge base has a higher priority than the factory information sub-knowledge base, and the factory information sub-knowledge base has a higher priority than the domain sub-knowledge base.

[0008] Optionally, the customer's preference information is stored in the customer preference sub-knowledge base indexed by the identifier of the customer or order.

[0009] Optionally, retrieving response information from the hierarchical knowledge base according to the priority of multiple sub-knowledge bases includes: The multiple sub-knowledge bases are queried sequentially from high to low priority until the response information is found. Obtain the response information.

[0010] Optional, also includes: If the response information conflicts with the comparison information, the multimodal large model is invoked to obtain conflict suggestion information, which includes a description of the conflict and suggestions for alternative solutions. A notification message is sent, which includes the conflict suggestion information.

[0011] Optionally, after sending the notification message, the method further includes: Receive the user's decision result in response to the prompt message; If the user's decision result includes an applicable target sub-knowledge base, then the original process execution data is re-optimized based on the comparison information corresponding to the target sub-knowledge base, wherein the target sub-knowledge base is selected from the remaining sub-knowledge bases.

[0012] Optionally, the process steps include multiple elements, and the step of calling the multimodal large model to analyze the original process execution data to obtain a query statement includes: Determine whether each process step is complete based on the aforementioned elements; If the judgment result is negative, then semantic parsing is performed on the process step to extract the missing elements; Based on the missing elements, the query statement is obtained by calling the multimodal large model.

[0013] Optional, also includes: Obtain the update information of the hierarchical knowledge base, which is extracted based on the optimized process execution data; The updated information is reviewed by calling the multimodal large model; The updated information that has passed the review will be stored in the corresponding sub-knowledge base according to the predefined hierarchical rules.

[0014] Optional, also includes: Repeat the following steps until the preset iteration stopping condition is met: The multimodal large model is invoked to analyze the process execution data after the previous optimization, and a query statement is obtained. The query statement is generated by the multimodal large model for at least one process step among the multiple process steps. Based on the query statement, multiple sub-knowledge bases in the hierarchical knowledge base are retrieved sequentially in descending order of priority until the response information is obtained. The multimodal large model is invoked, and the optimized process execution data is further optimized based on the retrieved response information.

[0015] Optionally, the preset iteration stopping condition includes: the number of repeated executions reaches a preset number of iterations.

[0016] Optionally, the preset iteration stopping condition includes: after calling the multimodal large model to analyze the process execution data after the previous optimization, the query statement is not generated.

[0017] Optional, also includes: Compare the optimized process execution data with the original process execution data to identify the changed process steps; An identifier is generated for the changed process steps, which is used to indicate the process steps, corresponding query statements, and response information in the original process execution data.

[0018] In a second aspect, the present invention provides a process breakdown device for engineering drawings, comprising: The drawing receiving and initial disassembly unit is used to receive the engineering drawings of the product to be processed, and call the pre-trained multimodal large model to disassemble the engineering drawings into processes, generating original process execution data, which includes multiple process steps; The query statement generation unit is used to call the multimodal large model to analyze the original process execution data and obtain a query statement, wherein the query statement is generated by the multimodal large model for at least one process step among the multiple process steps; The hierarchical retrieval unit is used to retrieve multiple sub-knowledge bases in the hierarchical knowledge base in descending order of priority based on the query statement until the response information is obtained, wherein the multiple sub-knowledge bases have different priorities; The process optimization unit is used to call the multimodal large model, optimize the original process execution data based on the retrieved response information, and generate optimized process execution data; The conflict handling unit is used to traverse and retrieve the remaining sub-knowledge bases to obtain comparison information, and to determine whether the response information conflicts with the comparison information. The result output unit is used to output the conflict between the response information and the comparison information along with the optimized process execution data if the judgment result is yes.

[0019] Thirdly, the present invention also provides a computer-readable storage medium, which is a non-volatile storage medium or a non-transient storage medium, on which a computer program is stored, wherein the computer program is executed by a computer to perform the process breakdown method of any of the above-mentioned engineering drawings.

[0020] Fourthly, the present invention also provides a computer program product, including a computer program / instruction, wherein when the computer program / instruction is run by a computer, the process breakdown method of any of the above-mentioned engineering drawings is executed.

[0021] Compared with the prior art, the present invention has the following beneficial effects: This invention, through generating query statements and retrieving multiple sub-knowledge bases from a hierarchical knowledge base according to priority, leverages the reasoning and generation capabilities of a large model while simultaneously verifying and correcting its output using structured, prioritized knowledge. This effectively suppresses the influence of illusions on the results, improves the reliability of order splitting results, and enables personalized adaptation of process splitting. After optimizing the original process execution data, the remaining sub-knowledge bases are traversed to obtain further comparison information, and it is determined whether the response information conflicts with the comparison information. Thus, when obtaining optimized process execution data, potential conflicts are also highlighted, avoiding the blind adoption of high-priority sub-knowledge bases while ignoring reasonable constraints in low-priority sub-knowledge bases.

[0022] Once a conflict is detected, the conflict point is output along with the optimization results, rather than concealing the conflict or allowing the model to make its own decision. This approach leverages the personalized advantages of high-priority sub-knowledge bases within the hierarchical knowledge base without masking conflicts between them and low-priority sub-knowledge bases. This solution directly addresses the unreliability issues caused by "illusions" and knowledge conflicts in existing solutions. The output not only includes complete and executable process steps but also conflict warnings, significantly improving the executability of the process decomposition solution in real-world production environments.

[0023] Furthermore, when the response information conflicts with the comparison information, a conflict suggestion information containing a conflict description and alternative suggestions is generated by calling a multimodal large model. This lowers the threshold for users to understand the conflict, provides users with a clear direction for resolution, and ensures that the conflict can reach users in a timely manner, thereby improving the efficiency of human-machine collaboration and ensuring the executability of the process breakdown results.

[0024] Furthermore, by defining process steps as structured objects containing multiple elements, the multimodal large model determines whether the process steps are complete based on the elements, extracts the missing elements for incomplete steps, and then generates query statements based on the missing elements. This enables the guided model to accurately locate knowledge questions. The generated query statements have clear directionality and searchability, solving the shortcomings of general large models that cannot recognize or accurately ask questions when knowledge is insufficient, and suppressing the generation of "illusions". Attached Figure Description

[0025] Figure 1 This is a flowchart of a process breakdown method for engineering drawings according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a process breakdown device for engineering drawings according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] As mentioned in the background section, existing enterprise information systems (hereinafter referred to as "the system") directly use general multimodal large models to break down orders. Although multimodal large models have a certain ability to understand engineering drawings, they lack specialized domain knowledge when directly used to break down process steps, which can easily lead to "illusions" and make the order breakdown results extremely unreliable.

[0028] One method to reduce the impact of "illusions" on output results is to introduce Retrieval-Augmented Generation (RAG) technology. This approach first segments documents such as process manuals and historical engineering cases into text blocks and stores them as vectors. During process planning, engineering drawings are embedded as vectors, and then fragments with semantically similar vectors are retrieved from the text block library. These retrieval results are used as context to assist the large model in generating process plans. However, because the knowledge in the text block library is stored in a flat manner, the retrieval results are fragmented when processing complex engineering drawings, making it difficult to output complete and reasonable process planning solutions.

[0029] While traditional retrieval enhancement generation techniques were introduced to supplement knowledge, they failed to effectively address the reliability issue of order splitting results. The root cause lies in two technical bottlenecks: First, regarding the triggering and execution of knowledge retrieval, traditional RAG solutions employ a global, one-time retrieval strategy, converting the user-input engineering drawings into embedded vectors and directly performing similarity matching. The inherent flaw in this approach is that the constraints of geometric features and tolerance requirements embodied in engineering drawings constitute structured technical information, creating a semantic gap between this information and the textual descriptions in the process knowledge base.

[0030] Secondly, in terms of the construction and organization of the knowledge base, existing solutions typically adopt a flattened text block storage method, which cuts knowledge sources such as process manuals and historical cases into fixed-length text fragments. This method disrupts the inherent logical structure of knowledge, resulting in the loss of hierarchical relationships and correlations between process types, processing parameters, and equipment requirements. When the model needs to handle problems requiring logical reasoning, such as multi-process coordination, process route selection, or satisfaction of personalized constraints, fragmented knowledge blocks are insufficient to support the model's structured and systematic thinking, leading to a lack of completeness in the output process solutions.

[0031] To address this technical problem, this invention provides a method for constructing a structured knowledge base with hierarchical structure and priorities, and performing hierarchical retrieval according to priority during order splitting. First, a multimodal large model is invoked to obtain the original process execution data. Then, the multimodal large model autonomously generates query statements and retrieves response information from the hierarchical knowledge base according to the priority order of customer preferences, factory capabilities, and domain knowledge. Finally, the original process execution data is optimized based on the retrieval results.

[0032] In this embodiment of the invention, by actively questioning the model and retrieving knowledge in a hierarchical manner according to priority, it is possible to utilize the reasoning and generation capabilities of the large model while using structured, priority-based proprietary knowledge to verify and correct its output, thereby effectively suppressing illusions and achieving personalized adaptation of process breakdown.

[0033] To make the above-mentioned objectives, features and beneficial effects of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] The process breakdown method for engineering drawings provided in this embodiment of the invention can be applied to enterprise information systems, such as Manufacturing Execution Systems (MES). Specifically, it can be applied to the order breakdown intelligent agent in the computer-aided design and production preparation stage of the industrial manufacturing field. The enterprise information system can be a cloud-based manufacturing collaboration platform. The backend of this system can run on a cloud server for processing and storing data; the frontend can be based on a webpage or app for inputting engineering drawings and viewing the processing results of process execution data.

[0035] like Figure 1 As shown, Figure 1 This is a flowchart of a process breakdown method for engineering drawings according to the first embodiment of the present invention.

[0036] Specifically, in this embodiment, the method for breaking down engineering drawings into work processes includes the following steps: Step S101 receives the engineering drawings of the product to be processed and calls the pre-trained multimodal large model to decompose the engineering drawings into processes, generating original process execution data, which includes multiple process steps.

[0037] The system receives engineering drawings of the products to be processed uploaded by users, such as DWG or PDF files exported from CAD software. The system integrates or can remotely access a pre-trained multimodal large model (hereinafter referred to as the "model"), such as GPT-4V or Claude 3. This model has been pre-trained in a general domain and possesses certain image recognition and text generation capabilities. The system provides the engineering drawings as input to the model and instructs it to perform process breakdown. The model recognizes information such as part shape, dimensioning, surface roughness symbols, and tolerance requirements in the drawings, and combines this with its built-in general knowledge to output a preliminary process breakdown plan, i.e., the original process execution data. This data can describe multiple process steps in a step-by-step manner, for example: "Step 1: Blanking—Laser cutting; Step 2: Bending—Forming at a 90-degree angle on a bending machine; Step 3: Drilling—Drilling a Φ10mm hole in the center of the workpiece; Step 4: Surface treatment—Powder coating." These process steps are arranged in the order of processing, forming the basis for subsequent optimization processing.

[0038] Step S102 calls the multimodal large model to analyze the original process execution data and obtain a query statement. The query statement is generated by the multimodal large model for at least one process step among multiple process steps.

[0039] After obtaining the original process execution data, the system does not directly output it as the final result. Instead, it calls a multimodal large model to analyze it again. For example, it can analyze the integrity of the original process execution data. This model can be the same model as the one used in the previous process or another isomorphic model to generate query statements. Each process step includes multiple elements. Integrity analysis can analyze each element in the process step to confirm whether any elements are missing or whether the information of some elements is accurate.

[0040] Specifically, the system takes the original process execution data as input and prompts the model: "Please check which steps in the above process have incomplete elements, require additional knowledge for verification or refinement? Please generate one or more queries for these steps." During the analysis, the model evaluates the completeness of the elements in each process step.

[0041] For example, for the step "bending—forming at a 90-degree angle on a bending machine," this step lacks information about the processing equipment. Therefore, the query would be: "What equipment is used to perform this bend to achieve a 90-degree angle?". Or, more specifically, the query could also include: "What is the maximum bending length and tonnage of the bending machine used?"

[0042] For example, consider the query "drilling - drilling Φ10mm hole," but lacking the characteristic elements for drilling. In this case, the characteristic element could be that the customer requires a threaded pilot hole. The generated query could be: "Are there any detailed machining requirements for the Φ10mm hole (such as tapping or reaming)?"

[0043] The generated query statement can be one or more, each statement executing the process steps in the original process data, aiming to obtain supplementary information from the hierarchical knowledge base.

[0044] Step S103 involves retrieving multiple sub-knowledge bases in the hierarchical knowledge base in descending order of priority based on the query statement until the response information is obtained. The multiple sub-knowledge bases have different priorities.

[0045] In one specific embodiment, after generating one or more query statements, a hierarchical retrieval is performed on each query statement. The retrieval strategy is as follows: First, the query statement is searched in the highest-priority sub-knowledge base, which may employ keyword matching, semantic vector similarity retrieval, or a combination of both. If a matching response is found, further retrieval stops, and this response is taken as the result of this query. If no valid information is found in the highest-priority sub-knowledge base, the search proceeds to the next highest-priority sub-knowledge base, and so on, until a response is found in a sub-knowledge base.

[0046] Furthermore, the multiple sub-knowledge bases specifically include: a customer preference sub-knowledge base, a factory information sub-knowledge base, and a domain sub-knowledge base. The customer preference sub-knowledge base records the preferences of customers who commission products for processing, such as a customer requiring all threaded holes to use forming taps. The factory information sub-knowledge base records information about the factory currently producing the products, such as the factory's equipment list (e.g., bending machine tonnage, laser cutting machine power, personnel skill levels, material list, common mold specifications, etc.). The domain sub-knowledge base records general information about the product category to which the products to be processed belong, such as standard sheet metal processing procedures, recommended cutting parameters for common materials, and national standard tolerance levels. The priority of the three sub-knowledge bases is set as follows: the customer preference sub-knowledge base has a higher priority than the factory information sub-knowledge base, and the factory information sub-knowledge base has a higher priority than the domain sub-knowledge base. This priority order reflects the principle of prioritizing customer needs, followed by factory conditions, with general knowledge serving as a fallback for engineering decision-making.

[0047] Specifically, after generating one or more query statements, a hierarchical search is performed on each query statement. The search strategy is as follows: First, the query statement is searched in the highest priority sub-knowledge base (customer preference sub-knowledge base). If a matching response is found, for example, for the query statement "Are there any feature requirements for Φ10mm holes?", the customer preference sub-knowledge base matches "Customer A requires Φ10mm holes to be machined with M10×1.5 threads, with a depth of 15mm", then further searching stops, and this response is taken as the result of this query. If no valid information is found in the customer preference sub-knowledge base, the search then moves to the next highest priority sub-knowledge base (factory information sub-knowledge base), and so on, until a response is found in the domain sub-knowledge base. This method of priority-based hierarchical searching, stopping when a match is found, ensures that the most personalized and specific knowledge is prioritized, while improving search efficiency.

[0048] Step S104 calls the multimodal large model to optimize the original process execution data based on the retrieved response information, and generates optimized process execution data.

[0049] After retrieving the response information, the system provides it along with the original process execution data to the multimodal large model and instructs the model: "Please optimize the original process steps based on the following supplementary knowledge." The model uses the response information as a constraint or guiding condition to modify, supplement, or reorder the corresponding process steps in the original process execution data.

[0050] In some implementations, the original process execution data is "drilling—drilling a Φ10mm hole," while the response information requires "machining an M10×1.5 thread, depth 15mm." Based on this, the model optimizes the original steps into two sub-steps: "Step 3a: Drill an Φ8.5mm pilot hole (depth 17mm, leaving tapping allowance); Step 3b: Tap an M10×1.5 thread (depth 15mm)." Simultaneously, the model may also adjust the order of subsequent steps. For example, when the threaded hole is located in an area inaccessible to the tool after bending and far from the bending deformation zone, tapping should be arranged before bending.

[0051] In other implementations, if the response information comes from a factory information sub-knowledge base, for example, if the customer does not have more detailed requirements for bending, and the response information retrieved from the factory information sub-knowledge base is "the maximum bending length of the factory bending machine is 2m", the model may break down the bending steps in the original process that exceed this length into multiple segments.

[0052] In some implementations, the response information comes from a domain sub-knowledge base. For example, if the model states that "the springback angle of 304 stainless steel bending is usually 2 degrees," then springback compensation will be added to the process parameters.

[0053] After optimizing the response information of all query statements, the system generates optimized process execution data. This data is more complete, executable, and better reflects the actual conditions of the customer and the factory compared to the original data.

[0054] Step S105 involves traversing and retrieving the remaining sub-knowledge bases to obtain comparison information, and determining whether the response information conflicts with the comparison information.

[0055] In this embodiment, after retrieving response information from a sub-knowledge base of a certain priority, such as the customer preference sub-knowledge base, the system does not ignore other sub-knowledge bases with lower priority. To detect conflicts between different levels of knowledge, the system continues to traverse and search other sub-knowledge bases that have not yet been searched, i.e., those with lower priority than the sub-knowledge base where the response information has been found. The purpose is to obtain comparison information; after comparing the response information with the comparison information, it determines whether there are any conflicts in the optimized process execution data.

[0056] Specifically, if the response information comes from the customer preference sub-knowledge base, the system will continue to search the factory information sub-knowledge base and the domain sub-knowledge base in turn to obtain the comparison information; if the response information comes from the factory information sub-knowledge base (second highest priority), the system will continue to search the domain sub-knowledge base to obtain the comparison information; if the response information comes from the domain sub-knowledge base (lowest priority), there is no sub-knowledge base with a lower priority, so there is no need to continue traversing.

[0057] In this context, a conflict refers to a contradiction or inconsistency between the response information and the comparison information. For example, the response information obtained from the customer preference sub-knowledge base is "requires machining M10 threads on a Φ10mm hole." The system then searches the factory information sub-knowledge base and obtains the comparison information "This factory's maximum tapping capability is M8; there are no M10 taps or corresponding fixtures." It then searches the domain sub-knowledge base and obtains the comparison information "The standard pilot hole for an M10 thread is Φ8.5mm; a spiral tap is recommended." The system compares the response information (customer requires M10 threads) with two sets of comparison information: one (the factory lacks M10 tapping capability) and another (the standard pilot hole for an M10 thread is Φ8.5mm; a spiral tap is recommended) to determine if a conflict exists. The result is: the response information and one set of comparison information, indicating the factory lacks M10 tapping capability, constitute a conflict.

[0058] Specifically, conflict determination can be based on predefined rules, such as "the process parameters requested by the customer are not within the capabilities of the factory equipment", or by semantic comparison using a multimodal large model.

[0059] If the judgment result in step S106 is yes, then the conflict between the response information and the comparison information will be output along with the optimized process execution data. When a conflict is determined, the system will not arbitrarily modify customer requirements or ignore factory restrictions, nor will it cover up the contradiction.

[0060] Specifically, the system will output any conflicts between the response information and the comparison information, such as the discrepancy between "customer requires M10 threads" and "factory lacks M10 tapping capability," along with the optimized process execution data to the user (process engineer or production manager). The output format can be a report containing process execution data and conflict information. For example, the optimized process execution data may still retain the "tapping M10×1.5 threads" step, but will output the conflict information as a note: "Warning: The customer preference sub-knowledge base requires M10 threads, but the factory information sub-knowledge base shows no M10 tapping capability. Please make a manual decision."

[0061] This output method suppresses the generation of hallucinations and improves the reliability of the results.

[0062] If the judgment result is negative, meaning that the response information does not conflict with any of the comparison information, then only the optimized process execution data will be output, without the need to add conflict information.

[0063] In some embodiments, steps S102 to S104 described above can be performed iteratively multiple times. That is, the optimized process execution data is used as input again, as will be explained in detail below.

[0064] In some non-limiting embodiments, customer preference information recorded in the customer preference sub-knowledge base is stored in the customer preference sub-knowledge base indexed by the identifier of the customer or order.

[0065] Specifically, when the system needs to retrieve response information from the customer preference sub-knowledge base, it first needs to determine the unique identifier of the customer or order corresponding to the product to be processed. This identifier can be obtained together with the engineering drawings, for example, when the user fills in "Customer Name: Company A" or "Order Number: PO-2026001A" when uploading CAD drawings.

[0066] In practical implementation, within the internal storage structure of the customer preference sub-knowledge base, each customer preference piece of information is associated with one or more identifier fields. For example, the preference information "All threaded holes require the use of extrusion taps" is tagged as belonging to "Customer ID: A"; the preference information "Surface color requires RAL" is tagged as belonging to "Customer ID: A" and further specified to "Order Number: PO-2024001A". When a customer has multiple orders and there are differentiated preferences between different orders, the system supports a two-level index: first, coarse screening by customer identifier, and then fine screening by order identifier.

[0067] During actual retrieval, the system performs the following operations: First, it matches the current order's customer identifier (e.g., "A") against the customer preference sub-knowledge base to obtain all general preference information for that customer. If the current order also provides an order identifier (e.g., "PO-2024001A"), it further searches the customer preference sub-knowledge base for specific preference information bound to that order identifier. This specific preference information has higher priority than general preference information for the same customer. If an entry for the same preference information exists in both a customer-level index and an order-level index, the preference information corresponding to the order-level index overrides the preference information corresponding to the customer-level index. The retrieved results form part of the response information and are used for subsequent optimization.

[0068] In some implementations, the customer preference sub-knowledge base supports dynamic updates. When the system outputs conflicting information, and the user communicates with the customer and reaches a new agreement, the modified preference information can be re-stored using the original identifier as an index, or the original record can be overwritten. For example, if a customer originally requested "M10 thread" but after negotiation changes it to "M8 thread," the system updates this preference information using "Customer ID: A, Order Number: PO-2024001A" as an index, allowing other drawings in the same batch of orders or future orders from the same customer to directly use the updated knowledge.

[0069] Furthermore, response information is retrieved from the hierarchical knowledge base according to the priority of multiple sub-knowledge bases. Specifically, this includes: querying multiple sub-knowledge bases in descending order of priority until response information is found; and then obtaining the response information.

[0070] It should be noted that "query" refers to executing the same query statement sequentially in sub-knowledge bases with different priorities, rather than batch processing multiple query statements simultaneously while querying the same sub-knowledge base. For multiple generated query statements, the system can execute the above operations independently and in parallel for each query statement to improve overall processing efficiency.

[0071] In this embodiment of the invention, the method further includes calling a multimodal large model to obtain conflict suggestion information, which includes a description of the conflict and suggestions for alternative solutions. That is, when the system determines that there is a conflict between the response information and the comparison information, it does not output the conflict directly, but instead further calls the multimodal large model to process the conflict.

[0072] Specifically, the system provides relevant information related to the conflict to the multimodal large model. This relevant information includes, but is not limited to: the query statement that triggered the conflict, the response information and its source sub-knowledge base, the comparison information and its source sub-knowledge base, and the context of the current process step. Upon receiving this information, the model performs two tasks: First, it generates a description of the conflict. This description is a natural language summary of the nature of the conflict, which is clearer and easier to understand than a simple comparison of raw data. For example, the model outputs: "The customer preference sub-knowledge base requires machining M10×1.5 threads on a Φ10mm hole, but the factory information sub-knowledge base shows that the current factory does not have the capability to tap M10 (the maximum tapping capability is M8)." Second, it generates suggestions for alternative solutions. Based on general domain knowledge and reasoning capabilities, the model proposes one or more feasible alternative solutions. These alternative solutions could include negotiating with the customer to persuade them to modify their requirements, adjusting factory configuration, changing the process route, or outsourcing the process.

[0073] In some specific implementations, the model can output suggestions such as: "Alternative 1: Negotiate with the customer to change the M10 thread to an M8 thread; Alternative 2: Purchase an M10 tap and corresponding fixture, and update the factory information sub-knowledge base; Alternative 3: Outsource the threading process to a cooperating factory with M10 tapping capabilities."

[0074] When generating alternative solutions, the model considers factors such as engineering feasibility and cost efficiency, but it does not need to provide the optimal solution. It only needs to provide reasonable options for subsequent human decision-making reference.

[0075] In this embodiment of the invention, after obtaining conflict suggestion information, the system encapsulates it into a notification message and sends it to the relevant user. The sending method can be a pop-up notification in the system interface, an email sent to the user's mailbox, or a push notification to an instant messaging tool.

[0076] In some implementations, after receiving a prompt message containing conflict suggestions, users review the conflict description and suggested alternatives, and make a final decision based on their expertise and the actual business situation. Users input their decision via the system interface, or by replying to emails, messages, etc. The system then receives the user's decision.

[0077] In one specific implementation, if the user's decision includes an applicable target sub-knowledge base, then subsequent re-optimization operations are performed. The applicable target sub-knowledge base refers to the user explicitly specifying one of the conflicting parties—either the sub-knowledge base containing the response information or the sub-knowledge base containing the comparison information—as the final knowledge source, i.e., the target sub-knowledge base. This target sub-knowledge base is selected from the remaining sub-knowledge bases.

[0078] It should be noted that the remaining sub-knowledge bases refer to the sub-knowledge bases used for traversal retrieval to obtain comparison information, and do not include the sub-knowledge bases used to obtain response information when querying by priority.

[0079] For example, if the response information comes from the customer preference sub-knowledge base (highest priority), the remaining sub-knowledge bases include the factory information sub-knowledge base and the domain sub-knowledge base. Users can choose to apply the factory information sub-knowledge base (i.e., abandon customer preferences and apply based on the actual capabilities of the factory) or the domain sub-knowledge base (i.e., apply based on general standards). The user's decision result needs to clearly indicate which sub-knowledge base was selected.

[0080] In other implementations, the system parses the user's decision result to determine whether it contains an indication of the applicable target sub-knowledge base and which specific sub-knowledge base was specified. If the determination result is yes, the system obtains the comparison information corresponding to the target sub-knowledge base.

[0081] Furthermore, the system invokes a multimodal large model and optimizes the original process execution data based on the comparison information.

[0082] Specifically, the user makes a decision to select the applicable knowledge from the factory information sub-knowledge base. At this time, the applicable target sub-knowledge base is the factory information sub-knowledge base. The system then obtains the comparison information "no M10 tapping capability" from the factory information sub-knowledge base and calls the multimodal large model to re-optimize the original process execution data. The optimized result changes the original "tapping M10 threads" step to "tapping M8 threads".

[0083] It should be noted that if the user's decision does not include the applicable target sub-knowledge base, such as if the user chooses to manually modify the process or chooses to ignore the conflict, this step will not be executed. The system can wait for other instructions from the user or directly adopt the result of the user's manual modification.

[0084] In this embodiment, each process step is regarded as a structured operation instruction, which includes at least several of the following elements: operation type (such as cutting, bending, drilling), processing object (such as sheet metal, workpiece features), process parameters (such as size, angle, depth), equipment requirements (such as the required machine tool type), quality requirements (such as tolerance grade), etc.

[0085] In a non-limiting example, a complete process step can be described as "cutting a 3mm thick stainless steel plate into a 200mm × 150mm rectangle using a laser cutting machine", where the operation type is "cutting", the processing object is "stainless steel plate", the process parameters include "thickness 3mm" and "size 200mm × 150mm", and the equipment requirement is "laser cutting machine".

[0086] Based on this, the multimodal large model is invoked to analyze the original process execution data and obtain query statements. Specifically, this includes: determining whether each process step is complete based on the elements; if the determination result is negative, semantic parsing is performed on the process step to extract the missing elements; and based on the missing elements, the multimodal large model is invoked to obtain query statements.

[0087] For example, when the model determines whether a process step is complete based on its built-in knowledge, the basis for the judgment can be a pre-defined list of elements. For instance, for a "drilling" process, the necessary elements include "hole diameter," "hole depth," and "hole position"; for a "bending" process, the necessary elements include "bending angle," "bending length," and "material thickness." If the model determines that a process step lacks any of the above necessary elements, it determines that the process step is incomplete; if all necessary elements are present and their values ​​are reasonable, it determines that the process step is complete.

[0088] Conversely, one step in the original process execution data is described as "drilling." This step lacks elements such as hole diameter, hole depth, and hole position, and is therefore deemed incomplete by the model. Similarly, a step described as "bending 90 degrees" lacks bending length and material thickness, and is also deemed incomplete.

[0089] Furthermore, when the multimodal large model determines that a certain process step is incomplete, the system further calls the model to perform semantic parsing on that process step.

[0090] Specifically, through semantic parsing, the model analyzes the existing text description of the step, identifies the existing elements and their values, and compares them with the list of elements that should be included in the complete process step, thereby determining which element(s) are missing.

[0091] The system uses the extracted missing elements as a basis to generate a query statement by calling a multimodal large model. For each missing element, the model combines the context of the process step (such as part type, material, adjacent processes, etc.) to generate a query statement in natural language. This query statement aims to retrieve the specific knowledge content of the missing elements from the hierarchical knowledge base.

[0092] For example, for the missing "hole diameter" element in the "drilling" step, the model generates the query: "What is the standard hole diameter to be drilled on this part?"; for the missing "hole depth" element, it generates the query: "Is this hole a through hole? If not, what is the required hole depth?"; for the missing "bending length" element, the model can combine the dimensions in the engineering drawings to generate the query: "What is the unfolded length of this bend?"; for the missing "material thickness" element, it generates the query: "What is the thickness of the sheet metal used in the current work order?". Multiple queries can be generated independently or combined into a single query containing multiple questions.

[0093] In some implementations, the system extracts updated information from optimized process execution data and reviews this updated information, which may be done by calling a multimodal large model. The purpose of the review is to confirm the accuracy and applicability of the information. Accuracy review refers to verifying whether the information contradicts known facts or engineering logic. Applicability review refers to determining whether the information has universal value and whether it is suitable for inclusion in the corresponding sub-knowledge base for reuse in subsequent tasks. Updated information that passes the review is stored in the corresponding sub-knowledge base according to predefined hierarchical rules: information involving special customer requirements is stored in the customer preference sub-knowledge base, information involving factory capabilities is stored in the factory information sub-knowledge base, and information involving general processes is stored in the domain sub-knowledge base.

[0094] This embodiment also includes an iterative optimization step for the process execution data. Specifically, the following steps are repeated until a preset iteration stop condition is met: The first step involves using a multimodal large-scale model to analyze the previously optimized process execution data, generating a query statement. This query statement is generated by the multimodal large-scale model for at least one of the multiple process steps. This step is similar to the initial execution, but the analysis object is changed from the original process execution data to the previously optimized process execution data. The second step involves using the query statement to sequentially search multiple sub-knowledge bases in the hierarchical knowledge base in descending order of priority until a response is obtained. The third step involves using the multimodal large-scale model to further optimize the optimized process execution data based on the retrieved response information.

[0095] After each iteration, the system determines whether the preset iteration stop condition is met; if it is met, the iteration stops and the current optimized process execution data is used as the final output; if it is not met, the next iteration begins.

[0096] In some implementations, the preset iteration stopping condition includes: the number of repeated executions reaches a preset number of iterations. For example, the system can be pre-configured to a maximum of 3 iterations; when the number of iterations reaches 3, the iteration stops regardless of the optimization effect. This condition serves as a fallback mechanism to prevent system resources from being exhausted due to the model getting stuck in an infinite loop.

[0097] In other implementations, the preset iteration stopping condition includes: after calling the multimodal large model to analyze the previously optimized process execution data, no query statement is generated. This condition means that the multimodal large model determines that the current optimized process execution data is complete enough and no further knowledge query is needed, at which point iteration stops. Compared to forcibly stopping after reaching a preset number of iterations, this condition better reflects that the process splitting results have met the output criteria and is a stopping condition for better order splitting results.

[0098] In some other implementations, after generating optimized process execution data, traceability identifiers are also included to further prevent "illusions" from interfering with the order splitting results.

[0099] Specifically, the system compares the optimized process execution data with the original process execution data to identify the changed process steps. The comparison can be done step-by-step, for example, by performing a content difference analysis on each process step in the original process execution data and the corresponding process step in the optimized process execution data. Changes include, but are not limited to: adding process steps, deleting process steps, modifying elements within process steps, and adjusting the order of process steps.

[0100] For any identified changed process steps, the system generates an identifier. This identifier indicates the process step in the original process execution data, the corresponding query statement, and the response information. In other words, through this identifier, users can trace why a process step was modified, how it was modified, and what knowledge was used as a basis.

[0101] For example, the original process execution data included a step "Drilling - Drilling a Φ10mm hole," which was later modified to "Drilling - Drilling an Φ8.5mm pilot hole" after optimization. The system generates an identifier for this modification, which is associated with the following information: the original step description ("Drilling - Drilling a Φ10mm hole"), the query that triggered the modification ("Does this hole have threaded or other characteristic limitations?"), and the response information retrieved from the hierarchical knowledge base ("Customer requires M10×1.5 thread, so an Φ8.5mm pilot hole needs to be drilled first").

[0102] This identifier can be presented in several forms. It could be a clickable reference marker or hover tooltip next to the changed process step when outputting optimized process execution data; clicking or hovering the tooltip displays the original step, query statement, and response information associated with the identifier. Alternatively, the system can generate a separate change log, listing each changed process step and its corresponding original step, query statement, and response information in tabular form.

[0103] Based on the above, the scheme of the first embodiment generates query statements and retrieves multiple sub-knowledge bases in the hierarchical knowledge base according to priority. This allows for the utilization of the large model's reasoning and generation capabilities while simultaneously using structured, prioritized knowledge to verify and correct its output. This effectively suppresses the influence of illusions on the results, improves the reliability of the order splitting results, and enables personalized adaptation of process splitting. After optimizing the original process execution data, the remaining sub-knowledge bases are traversed to obtain further comparison information, and it is determined whether the response information conflicts with the comparison information. Thus, when the optimized process execution data is obtained, potential conflicts are also indicated, avoiding the blind adoption of high-priority sub-knowledge bases while ignoring reasonable constraints in low-priority sub-knowledge bases.

[0104] Once a conflict is detected, the conflict point is output along with the optimization results, rather than concealing the conflict or allowing the model to make its own decision. This approach leverages the personalized advantages of high-priority sub-knowledge bases within the hierarchical knowledge base without masking conflicts between them and low-priority sub-knowledge bases. This solution directly addresses the unreliability issues caused by "illusions" and knowledge conflicts in existing solutions. The output not only includes complete and executable process steps but also conflict warnings, significantly improving the executability of the process decomposition solution in real-world production environments.

[0105] Figure 2 This is a schematic diagram of a process decomposition method device 2 for engineering drawings based on a general large language model, according to a second embodiment of the present invention. Those skilled in the art will understand that the process decomposition method device for engineering drawings based on a general large language model in this embodiment can be used to implement the method technical solutions in the above embodiments.

[0106] Specifically, refer to Figure 2 The process decomposition method device 2 for engineering drawings based on a general large language model in this embodiment may include: The drawing receiving and initial disassembly unit 21 is used to receive the engineering drawings of the product to be processed, and call the pre-trained multimodal large model to disassemble the engineering drawings into processes, generating original process execution data, which includes multiple process steps. The query statement generation unit 22 is used to call the multimodal large model to analyze the original process execution data and obtain the query statement. The query statement is generated by the multimodal large model for at least one process step among multiple process steps. The hierarchical retrieval unit 23 is used to retrieve multiple sub-knowledge bases in the hierarchical knowledge base in descending order of priority based on the query statement until the response information is obtained. The multiple sub-knowledge bases have different priorities. The process optimization unit 24 is used to call the multimodal large model, optimize the original process execution data based on the retrieved response information, and generate optimized process execution data. The conflict handling unit 25 is used to traverse and retrieve the remaining sub-knowledge bases to obtain comparison information and determine whether the response information conflicts with the comparison information. The result output unit 26 is used to output the conflict between the response information and the comparison information along with the optimized process execution data if the judgment result is yes.

[0107] In specific implementation, the modules / units included in the various devices and products described in the above embodiments can be software modules / units, hardware modules / units, or a combination of both.

[0108] For example, for various devices and products applied to or integrated into a chip, each module / unit can be implemented using hardware methods such as circuits, or at least some modules / units can be implemented using software programs that run on a processor integrated within the chip, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits; for various devices and products applied to or integrated into a chip module, each module / unit can be implemented using hardware methods such as circuits, and different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components of the chip module, or at least some modules / units can be implemented using hardware methods such as circuits. The components can be implemented using software programs that run on the processor integrated within the chip module. The remaining (if any) modules / units can be implemented using hardware methods such as circuits. For various devices and products applied to or integrated into the terminal, each of its components / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or in different components within the terminal. Alternatively, at least some modules / units can be implemented using software programs that run on the processor integrated within the terminal, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits.

[0109] Furthermore, embodiments of the present invention also disclose a computer-readable storage medium, which is a non-volatile or non-transient storage medium, on which a computer program is stored, and the computer program is executed by a computer. Figure 1 The method and technical solution in the illustrated embodiment. Preferably, the storage medium may include ROM, RAM, disk, or optical disk, etc.

[0110] Furthermore, embodiments of the present invention also disclose a computer program product, including a computer program / instruction, wherein when the computer program / instruction is executed by a computer, the process breakdown method of any of the above-mentioned engineering drawings is executed.

[0111] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for breaking down engineering drawing processes, characterized in that, include: Receive engineering drawings of the product to be processed, and call a pre-trained multimodal large model to decompose the engineering drawings into processes, generating original process execution data, which includes multiple process steps; The multimodal large model is invoked to analyze the original process execution data to obtain a query statement, which is generated by the multimodal large model for at least one of the multiple process steps. Based on the query statement, multiple sub-knowledge bases in the hierarchical knowledge base are retrieved sequentially in descending order of priority until a response is obtained. The multiple sub-knowledge bases have different priorities. The multimodal large model is invoked, and the original process execution data is optimized based on the retrieved response information to generate optimized process execution data; The remaining sub-knowledge bases are traversed and retrieved to obtain comparison information, and it is determined whether the response information conflicts with the comparison information. The remaining sub-knowledge bases refer to all sub-knowledge bases other than the response information. If the judgment result is yes, then the conflict between the response information and the comparison information will be output along with the optimized process execution data.

2. The method for breaking down engineering drawings into work processes as described in claim 1, characterized in that, The multiple sub-knowledge bases include: The customer preference sub-knowledge base records the preference information of customers who have commissioned the products to be processed. The factory information sub-knowledge base records information about the factory currently producing the product to be processed. The domain sub-knowledge base records general information about the product category to which the product to be processed belongs; The customer preference sub-knowledge base has a higher priority than the factory information sub-knowledge base, and the factory information sub-knowledge base has a higher priority than the domain sub-knowledge base.

3. The method for breaking down engineering drawings into work processes as described in claim 2, characterized in that, The customer's preference information is stored in the customer preference sub-knowledge base, indexed by the identifier of the customer or order.

4. The method for breaking down engineering drawings into work processes as described in claim 2, characterized in that, The step of retrieving response information from a hierarchical knowledge base according to the priority of multiple sub-knowledge bases includes: The multiple sub-knowledge bases are queried sequentially from high to low priority until the response information is found. Obtain the response information.

5. The method for breaking down engineering drawings into work processes as described in claim 4, characterized in that, Also includes: If the response information conflicts with the comparison information, the multimodal large model is invoked to obtain conflict suggestion information, which includes a description of the conflict and suggestions for alternative solutions. A notification message is sent, which includes the conflict suggestion information.

6. The method for breaking down engineering drawings into work processes as described in claim 5, characterized in that, Following the sending of the notification message, the following is also included: Receive the user's decision result in response to the prompt message; If the user's decision result includes an applicable target sub-knowledge base, then the original process execution data is re-optimized based on the comparison information corresponding to the target sub-knowledge base, wherein the target sub-knowledge base is selected from the remaining sub-knowledge bases.

7. The method for breaking down engineering drawings into work processes as described in claim 1, characterized in that, The process steps include multiple elements. The process of calling the multimodal large model to analyze the original process execution data and obtaining a query statement includes: Determine whether each process step is complete based on the aforementioned elements; If the judgment result is negative, then semantic parsing is performed on the process step to extract the missing elements; Based on the missing elements, the query statement is obtained by calling the multimodal large model.

8. The method for breaking down engineering drawings into work processes as described in claim 1, characterized in that, Also includes: Obtain the update information of the hierarchical knowledge base, which is extracted based on the optimized process execution data; The updated information is reviewed by calling the multimodal large model; The updated information that has passed the review will be stored in the corresponding sub-knowledge base according to the predefined hierarchical rules.

9. The method for breaking down engineering drawings into work processes as described in claim 1, characterized in that, Also includes: Repeat the following steps until the preset iteration stopping condition is met: The multimodal large model is invoked to analyze the previously optimized process execution data to obtain a query statement, which is generated by the multimodal large model for at least one of the multiple process steps. Based on the query statement, multiple sub-knowledge bases in the hierarchical knowledge base are retrieved sequentially in descending order of priority until the response information is obtained. The multimodal large model is invoked, and the optimized process execution data is further optimized based on the retrieved response information.

10. The method for breaking down engineering drawings into processes as described in claim 9, characterized in that, The preset iteration stopping condition includes: the number of repeated executions reaches the preset number of iterations.

11. The method for breaking down engineering drawings into processes as described in claim 9, characterized in that, The preset iteration stopping condition includes: after calling the multimodal large model to analyze the process execution data after the previous optimization, no query statement is generated.

12. The method for breaking down engineering drawings into processes as described in claim 1, characterized in that, Also includes: Compare the optimized process execution data with the original process execution data to identify the changed process steps; An identifier is generated for the changed process steps, which is used to indicate the process steps, corresponding query statements, and response information in the original process execution data.

13. A process breakdown device for engineering drawings, characterized in that, include: The drawing receiving and initial disassembly unit is used to receive the engineering drawings of the product to be processed, and call the pre-trained multimodal large model to disassemble the engineering drawings into processes, generating original process execution data, which includes multiple process steps; The query statement generation unit is used to call the multimodal large model to analyze the original process execution data and obtain a query statement, wherein the query statement is generated by the multimodal large model for at least one process step among the multiple process steps; The hierarchical retrieval unit is used to retrieve multiple sub-knowledge bases in the hierarchical knowledge base in descending order of priority based on the query statement until the response information is obtained, wherein the multiple sub-knowledge bases have different priorities. The process optimization unit is used to call the multimodal large model, optimize the original process execution data based on the retrieved response information, and generate optimized process execution data; The conflict handling unit is used to traverse and retrieve the remaining sub-knowledge bases to obtain comparison information, and to determine whether the response information conflicts with the comparison information. The result output unit is used to output the conflict between the response information and the comparison information along with the optimized process execution data if the judgment result is yes.

14. A computer-readable storage medium, wherein the computer-readable storage medium is a non-volatile storage medium or a non-transient storage medium, and a computer program is stored thereon, characterized in that, The computer program is executed by a computer to perform the process breakdown method of the engineering drawings according to any one of claims 1 to 12.

15. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the computer, the process breakdown method of the engineering drawings according to any one of claims 1 to 12 is executed.

Images