A flexible production line reconfiguration method that integrates GraphRAG and improved Whisper

Abstract

This application relates to the field of artificial intelligence and automation systems technology, specifically to a flexible production line reconfiguration method integrating GraphRAG and improved Whisper. New workpiece / process requirements serve as input. If the requirements are verbal, the Whisper-S speech recognition model is used to convert the speech description into a text description. The text description serves as input to GraphRAG, which acts as a task decomposer for ControlLLM, breaking down the new workpiece / process task into basic and well-defined sub-tasks. A tool graph containing all modular hardware is manually constructed. ToG, acting as the task planner for ControlLLM, receives the sub-tasks output by GraphRAG and uses a depth-first search algorithm to search for resource nodes and module nodes in the tool graph, obtaining logically related module groups as the solution. This solution constitutes the reconfigured modular flexible production line. The ControlLLM execution engine calls the twin system API through MCP, and the solution generated by ToG is validated on the production line through biomimetic experiments.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and automation systems technology, specifically relating to a flexible production line reconfiguration method that integrates GraphRAG and improved Whisper. Background Technology

[0002] In the field of industrial automation, flexible production line reconfiguration is a key technology for achieving rapid response to market demands and adapting to diverse production needs. However, existing technologies still face numerous limitations in speech recognition, task decomposition, and production line reconfiguration. First, traditional speech recognition models perform poorly in noisy industrial environments, struggling to accurately recognize technical terms, resulting in insufficient accuracy and reliability of speech input. Second, task decomposition methods based on traditional rules or simple machine learning models often struggle to efficiently handle complex new workpiece / process requirements, especially in dynamically changing industrial scenarios, making global optimization and flexible adjustments difficult. Furthermore, existing production line reconfiguration methods typically rely on manual experience or fixed module combinations, lacking intelligent and automated support, leading to low reconfiguration efficiency and failing to meet the actual needs of SMEs for frequent line changes.

[0003] To address the aforementioned issues, while some recent studies have attempted to improve the efficiency of task decomposition and production line reconfiguration using large language models and graph techniques, existing methods still suffer from limitations such as singular knowledge representation, limited reasoning capabilities, and insufficient global optimization. Furthermore, current methods lack efficient collaborative mechanisms across the entire process of speech recognition, task decomposition, and production line reconfiguration, making it difficult to achieve end-to-end intelligent support from speech input to production line reconfiguration. Therefore, there is an urgent need for an efficient and flexible solution that integrates speech recognition, task decomposition, and production line reconfiguration technologies to meet the pressing demand for flexible production line reconfiguration in industrial scenarios. Summary of the Invention

[0004] This application provides a flexible production line reconfiguration method that integrates GraphRAG and improved Whisper. This method can intelligently identify new workpiece / process requirements, automatically generate optimal module combination schemes, control logic and material flow paths, and use digital twins for virtual debugging and verification. It significantly shortens the changeover time of non-standard automated production lines, solves the pain point of "frequent line changes" for small and medium-sized enterprises, and provides efficient and flexible technical support for flexible production line reconfiguration.

[0005] This application provides a flexible production line reconfiguration method that integrates GraphRAG and improved Whisper, the method comprising:

[0006] New workpiece / process requirements are taken as input. If the requirements are verbal, the Whisper-S speech recognition model is used to convert the speech description into a text description.

[0007] The text description serves as input to GraphRAG, which acts as a task decomposer for ControlLLM, breaking down the new job / process task into basic and well-defined subtasks.

[0008] A tool graph containing all modular hardware is manually constructed. ToG acts as the task planner of the ControlLLM, receives the subtasks output by the GraphRAG, and uses a depth-first search algorithm to search the resource nodes and module nodes of the tool graph to obtain a logically related group of modules as a solution. The solution is the reconstructed modular flexible production line.

[0009] The execution engine of ControlLLM calls the twin system API through MCP, and performs production line verification through biomimetic experiments based on the solution generated by ToG.

[0010] In one implementation of this application, the construction process of the Whisper-S speech recognition model includes:

[0011] The Whisper-S model is constructed by replacing the CrossAttention in the 4-layer Transformer Decoder Block of the original Whisper-Tiny model with Overview of codec Cross-attention (OOC-CA).

[0012] Based on a private production environment labeled dataset and the open-source ST-CMDS dataset, the Whisper-S model was optimized using the LoRA (Low-Rank Adaptation) fine-tuning technique.

[0013] In one implementation of this application, the calculation process of OOC-CA includes:

[0014] The encoder features are processed by the Self Attention layer to extract hierarchical adaptive encoder features; the hierarchical adaptive encoder features are then used with the query embedding matrix and the key embedding matrix to calculate the query matrix and the key matrix, respectively; the query matrix and the key matrix are multiplied to obtain the weighted weight matrix; the decoder features are then used with the value embedding matrix to calculate the value matrix; the weighted weight matrix and the value matrix are then multiplied to obtain the output features of the OOC-CA.

[0015] In one implementation of this application, the collection and preprocessing process of the private production environment labeled dataset is as follows:

[0016] For data acquisition, a high-sensitivity directional microphone with a sampling rate of 48kHz is used; the microphone is placed 30-50 cm away from the sound source to capture clear speech signals; the signal-to-noise ratio reaches 25dB to reduce background noise interference; the acquisition strategy emphasizes data diversity, covering speech signals from different devices in operation, ensuring that the model adapts to various industrial scenarios; and audio quality is monitored in real time to avoid the mixing of abnormal data.

[0017] In preprocessing, traditional filters are used to remove background noise; long-term speech signals are segmented into smaller frames or segments for easier subsequent processing; the volume and sampling rate of the speech signal are adjusted to ensure consistency; Mel-frequency cepstral coefficients (MFCCs) are extracted as input to the speech recognition model; and simulated noise and pitch shifting are used for data augmentation to further improve the robustness of the model.

[0018] In one implementation of this application, GraphRAG acts as a task decomposer for ControlLLM, breaking down the new job / process task into basic and well-defined subtasks, specifically including:

[0019] The task decomposition process includes two parts: knowledge graph construction and global search optimization.

[0020] The knowledge graph construction process involves: dividing the input corpus into multiple text units, each containing detailed information about the new workpiece / process requirements and solutions; extracting entities, relationships, and key statements from the text units using LLM to construct an initial knowledge graph; performing hierarchical clustering on the initial knowledge graph using the Leiden algorithm to assign entities to different communities for in-depth analysis; and generating a summary for each community to provide an overview and contextual information of the dataset.

[0021] The global search optimization process is as follows: the new workpiece / process requirement is used as the initial input; the community is designated as the hierarchical level; based on the LLM, semantic retrieval technology is used to retrieve and summarize the community summary to generate a community report; the community report is shuffled and divided into multiple batches; each batch of community reports is divided into text blocks of a predefined size; based on the LLM, a rating intermediate response is generated; the intermediate responses are ranked and filtered, and the most important points are selected to generate an aggregated intermediate response; the aggregated intermediate response is used as context input to the LLM to generate the clear subtask.

[0022] In one implementation of this application, the input corpus is a historically accumulated text of the new workpiece / process requirements and solutions.

[0023] In one implementation of this application, entities in the text unit are extracted using LLM, and the extracted entities specifically include:

[0024] The name of the new workpiece / process, the name / designation of the modular hardware (module), the description of the execution action of the module, the scope of the module, the input parameters of the module, the input materials / middleware of the module, and the output workpiece / middleware of the module.

[0025] In one implementation of this application, the LLM is a Qwen3-7B model, which, based on Prompt, completes the tasks of extracting entities, relations, and key claims from the text units, summarizing the community summary to generate a community report, generating intermediate rating responses, and generating the clarity subtask.

[0026] In one implementation of this application, the clear subtask is generated, and the clear subtask specifically includes: a description of the execution action of the subtask, the scope of the module required to execute the subtask, the input parameters of the module and the material / middleware, and the output workpiece / middleware of the module.

[0027] In one implementation of this application, the solution specifically includes:

[0028] The module combination scheme, control logic, and material flow path are defined as follows: The module combination scheme includes the module combinations used, the logical relationship of the module execution order, the input parameters of the modules, the input materials / middleware of the modules, and the output workpieces / middleware of the modules; the control logic corresponds to the logical relationship of the module execution order in the module combination scheme; and the material flow path corresponds to the logical relationship of the input materials of the modules in the module combination scheme, and these logical relationships are consistent with the logical relationship of the module execution order.

[0029] Compared with the prior art, the significant advantages of this application are as follows:

[0030] Through the above technical solutions, a Whisper-S industrial speech recognition model is constructed to improve the accuracy of speech recognition and the ability to recognize technical terms in industrial noise environments. Simultaneously, a task decomposer is built based on GraphRAG, and through knowledge graph construction and global search optimization, efficient decomposition of new workpiece / process requirements is achieved. Finally, a flexible production line reconfiguration framework based on knowledge graphs and tool-enhanced large models is proposed, combining Whisper-S and GraphRAG technologies to realize a complete process from voice input to production line reconfiguration. This method can intelligently identify new workpiece / process requirements, automatically generate optimal module combination schemes, control logic, and material flow paths, and utilize digital twins for virtual debugging and verification, significantly shortening the changeover time of non-standard automated production lines. Attached Figure Description

[0031] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0032] Figure 1 This is a flowchart illustrating a flexible production line reconfiguration method that integrates GraphRAG and improved Whisper in an embodiment of this application.

[0033] Figure 2 This is a schematic diagram of the Whisper-S decoder network constructed using OOC-CA in this embodiment of the application;

[0034] Figure 3 This is a schematic diagram of a task decomposition process using GraphRAG in an embodiment of this application. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] Traditional speech recognition models perform poorly in noisy industrial environments, struggling to accurately recognize technical terms, resulting in insufficient accuracy and reliability of speech input. Secondly, task decomposition methods based on traditional rules or simple machine learning models often fail to efficiently handle complex new workpiece / process requirements, especially in dynamically changing industrial scenarios, making global optimization and flexible adjustments difficult. Furthermore, existing production line reconfiguration methods typically rely on manual experience or fixed module combinations, lacking intelligent and automated support, leading to low reconfiguration efficiency and failing to meet the actual needs of SMEs with frequent production line changes.

[0037] Based on this, the embodiments of this application provide a flexible production line reconfiguration method that integrates GraphRAG and improved Whisper to solve the problems of insufficient speech recognition accuracy, low task decomposition efficiency, and insufficient production line reconfiguration flexibility in industrial scenarios. It improves speech recognition accuracy and professional terminology recognition rate, enhances task decomposition efficiency and global optimization capabilities, effectively shortens the changeover time of non-standard automated production lines, and provides efficient and flexible technical support for flexible production line reconfiguration.

[0038] The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0039] This application provides a flexible production line reconfiguration method that integrates GraphRAG and improved Whisper, such as Figure 1 As shown, the method may include steps S101-S104:

[0040] S101, new workpiece / process requirements are taken as input. If the requirements are verbal, the Whisper-S speech recognition model is used to convert the speech description into a text description.

[0041] In this application embodiment, the challenges faced by the Whisper-Tiny speech recognition model include:

[0042] Industrial environmental noise interference, namely the noise generated by the operation of workshop equipment, can affect the accuracy of speech recognition;

[0043] Insufficient recognition of technical terms: Industrial scenarios involve a large number of technical terms, such as "modular hardware" and "material flow," which traditional speech recognition models may not be able to accurately recognize.

[0044] In this embodiment of the application, to address the challenges of the Whisper-Tiny speech recognition model, a Whisper-S speech recognition model is constructed. The construction process of the Whisper-S speech recognition model includes:

[0045] The CrossAttention mechanism in the original Whisper-Tiny model's 4-layer Transformer Decoder Block is replaced with Overview of Codec Cross-Attention (OOC-CA) to construct the Whisper-S model. This improvement more efficiently captures global contextual information in the speech signal and enhances the model's robustness to noisy environments. The network structures of the Transformer Decoder in the OOC-CA and improved Whisper-S models are as follows: Figure 2 As shown;

[0046] Based on a private production environment labeled dataset and the open-source ST-CMDS dataset, LoRA (Low-Rank Adaptation) fine-tuning technique is used to optimize the model for specialized terminology in industrial scenarios. LoRA reduces fine-tuning costs by introducing low-rank decomposition into the model parameters, while improving the model's adaptability to specific scenarios.

[0047] Furthermore, the hardware used for model fine-tuning includes: an NVIDIA H200 141GB GPU, an Intel Xeon E5-2680 v4 CPU, and 256GB DDR4 memory.

[0048] Furthermore, the model fine-tuning strategy includes: a batch size of 32, a learning rate of 1e-4, and 50 training epochs.

[0049] In this embodiment of the application, the calculation process of OOC-CA includes:

[0050] The Whisper-S model's 4-layer Transformer Decoder Block should use encoder features adapted to this decoder module. These encoder features are used in the OOC-CA to obtain the query and key matrices, and a pre-normalization operation is performed to calculate the weighted weight matrix based on these matrices. Therefore, the encoder features first pass through a Self-Attention layer to extract hierarchical adaptive encoder features; then, the hierarchical adaptive encoder features are used with the query embedding matrix and key embedding matrix to calculate the query and key matrices, respectively; finally, matrix multiplication is performed on the query and key matrices to obtain the weighted weight matrix. The decoder features are then used with the value embedding matrix to calculate the value matrix. The weighted weight matrix is ​​then multiplied with the value matrix to weight the value matrix, resulting in the OOC-CA output features.

[0051] In this embodiment, the acquisition and preprocessing of industrial scene speech data is a key step in ensuring the accurate operation of the speech recognition model in complex industrial environments. The acquisition and preprocessing process of the private production environment labeled dataset is as follows:

[0052] In terms of data acquisition, the selection of data acquisition equipment is crucial. High-sensitivity directional microphones with a sampling rate of 48kHz are typically used, placed 30-50 cm from the sound source to capture clear speech signals. During acquisition, strict environmental requirements are maintained to ensure a signal-to-noise ratio of 25dB, minimizing background noise interference. Simultaneously, the acquisition strategy emphasizes data diversity, covering speech signals from different devices operating, ensuring the model adapts to various industrial scenarios. Furthermore, real-time monitoring of audio quality prevents the intrusion of abnormal data.

[0053] In preprocessing, the acquired speech data undergoes a series of steps to improve the accuracy and robustness of speech recognition. First, noise reduction is performed using deep learning-based algorithms or traditional filters to effectively remove background noise. Next, long-duration speech signals are segmented into smaller frames or segments for easier subsequent processing. Standardization adjusts the volume and sampling rate of the speech signal to ensure consistency. In the feature extraction stage, key features such as Mel-frequency cepstral coefficients (MFCC) and linear prediction coefficients (LPC) are extracted to provide input for the speech recognition model. Finally, data augmentation techniques, such as adding simulated noise and pitch shifting, further enhance the model's robustness.

[0054] S102, the text description serves as the input to GraphRAG, which acts as a task decomposer for ControlLLM, breaking down new job / process tasks into basic and well-defined subtasks.

[0055] In this embodiment, GraphRAG, acting as a task decomposer for ControlLLM, breaks down new workpiece / process tasks into basic and well-defined subtasks. The task decomposition process is as follows: Figure 3 As shown, the method includes:

[0056] For knowledge graph construction, GraphRAG builds a knowledge graph through text unit segmentation, entity and relation extraction, hierarchical clustering, and community summary generation during the indexing process. This graph reflects the detailed relationships between historical requirements and solutions, providing strong semantic support for task decomposition. Specifically, the input corpus is first divided into multiple text units, each containing detailed information about new workpiece / process requirements and solutions. Then, entities, relations, and key statements are extracted from the text using LLM to construct an initial knowledge graph. Next, the Leiden algorithm is used to perform hierarchical clustering on the initial knowledge graph, assigning entities to different communities for in-depth analysis. Finally, a summary is generated for each community, providing an overview and contextual information of the dataset.

[0057] Global search optimization: During the query process, GraphRAG employs a global search workflow, achieving efficient reasoning for global problems through community report batching, intermediate response rating, and aggregation. This global optimization method considers the optimality of module combination, control logic, and material flow paths from a holistic perspective. Specifically, new workpiece / process requirements serve as initial input; the specified hierarchy level is designated as "community," and based on LLM, semantic retrieval technology is used to retrieve and summarize community summaries, generating community reports. These reports are shuffled and divided into multiple batches; each batch is divided into predefined-sized text blocks, and LLM-rated intermediate responses are generated; these intermediate responses are ranked and filtered, selecting the most important points; the aggregated intermediate responses are used as contextual input to LLM to generate clear subtasks.

[0058] In this embodiment of the application, the input corpus consists of historically accumulated texts of new workpiece / process requirements and solutions.

[0059] In this embodiment of the application, entities are extracted from text units using LLM. The extracted entities specifically include:

[0060] Name of the new workpiece / process, name / designation of the modular hardware (module), description of the module's execution actions, scope of the module, input parameters of the module, input materials / middleware of the module, and output workpiece / middleware of the module.

[0061] In this embodiment, the LLM is the Qwen3-7B model, which, based on Prompt, completes the tasks of extracting entities, relations, and key claims from text units, summarizing community summaries to generate community reports, generating rating intermediate responses, and generating clear sub-tasks.

[0062] In this embodiment of the application, a clarification subtask is generated, which specifically includes:

[0063] The subtask's execution action description, the scope of the module required to execute the subtask, the module's input parameters and materials / middleware, and the module's output workpieces / middleware.

[0064] S103, a tool graph containing all modular hardware is manually constructed. ToG, as the task planner of ControlLLM, receives subtasks output by GraphRAG and uses a depth-first search algorithm to search for resource nodes and module nodes in the tool graph to obtain a logically related group of modules as the solution. The solution is the reconstructed modular flexible production line.

[0065] In this application embodiment, the solution specifically includes:

[0066] The module combination scheme, control logic, and material flow path are defined as follows: The module combination scheme includes the combination of modules used, the logical relationship of module execution order, the input parameters of the modules, the input materials / middleware of the modules, and the output workpieces / middleware of the modules; the control logic corresponds to the logical relationship of module execution order in the module combination scheme; and the material flow path corresponds to the logical relationship of module input materials in the module combination scheme, and the logical relationship is consistent with the logical relationship of module execution order.

[0067] S104, the execution engine of ControlLLM calls the twin system API through MCP, and performs production line verification based on the solution generated by ToG through biomimetic experiments.

[0068] In this embodiment, ControlLLM's execution engine calls the twin system API through MCP to perform biomimetic experiments based on the solution generated by ToG, thereby achieving production line verification, specifically including:

[0069] Solution Generation: ToG (Tool Enhancement Big Model) generates module combination schemes, control logic, and material flow paths based on the tasks decomposed by GraphRAG;

[0070] Virtual debugging: The execution engine inputs these solutions into the digital twin system to simulate the operation of the actual production line and verify the feasibility of the module combination solution;

[0071] Bionic Experiment: By calling the twin system API through the MCP protocol, the execution engine conducts bionic experiments in a virtual environment to ensure the effectiveness and stability of the module combination scheme in actual production;

[0072] Results Feedback: The results of the biomimetic experiment were fed back to ControlLLM for further optimization of module combination schemes and control logic.

[0073] Through the above steps, this application embodiment realizes a complete process from voice input to production line reconstruction, ensuring the feasibility and efficiency of the solution and providing strong technical support for flexible production line reconstruction.

Claims

1. A flexible line reconfiguration method fusing GraphRAG and improved Whisper, characterized in that, The method comprises: New workpiece / process requirements are input, and if the requirements are spoken, a Whisper-S speech recognition model is used to convert the speech description into a text description; The text description is input into a GraphRAG, which is a task decomposer of a ControlLLM, and the GraphRAG decomposes the tasks of the new workpiece / process into basic and clearly defined subtasks; A tool graph ToG containing all modular hardware is artificially constructed, the ToG is a task planner of the ControlLLM, receives the subtasks output by the GraphRAG, searches resource nodes and module nodes of the ToG using a depth-first search algorithm, and obtains a module group with logical relationships as a solution, which is a reconstructed modular flexible production line; An execution engine of the ControlLLM calls a twin system API through an MCP, and performs production line verification through a bionic experiment according to the solution generated by the ToG; The construction process of the Whisper-S speech recognition model comprises: Cross Attention in the 4-layer Transformer Decoder Block of the original Whisper-Tiny model is replaced by an overview cross-attention OOC-CA to construct the Whisper-S model; Based on a private production environment annotation dataset and an open source ST-CMDS dataset, a LoRA fine-tuning technology is used to optimize the Whisper-S model.

2. The flexible line reconfiguration method fusing GraphRAG and improved Whisper according to claim 1, wherein, The calculation process of the OOC-CA comprises: Encoder features are extracted into hierarchical adaptive encoder features through a Self Attention layer; the hierarchical adaptive encoder features are calculated with a query embedding matrix and a key embedding matrix to obtain a query matrix and a key matrix, respectively; the query matrix and the key matrix are subjected to matrix multiplication to obtain a weighted weight matrix; a value matrix is obtained by calculating a decoder feature and a value embedding matrix; the weighted weight matrix and the value matrix are subjected to matrix multiplication to obtain output features of the OOC-CA.

3. The flexible line reconfiguration method fusing GraphRAG and improved Whisper of claim 1, wherein, The collection and preprocessing process of the private production environment annotation dataset comprises: In the collection, a high-sensitivity directional microphone is used, the sampling rate is 48 kHz, the microphone is placed 30-50 cm away from the sound source to capture clear voice signals, and the signal-to-noise ratio reaches 25 dB to reduce the interference of background noise; The collection strategy focuses on data diversity, covers voice signals during operation of different devices, and ensures that the model is adapted to various industrial scenarios; real-time monitoring of audio quality avoids the mixing of abnormal data; In the preprocessing, a traditional filter is used to remove background noise; long-time voice signals are segmented into small frames or segments for subsequent processing; the volume and sampling rate of the voice signals are adjusted to ensure consistency; mel-frequency cepstral coefficients MFCCs are extracted as inputs of the voice recognition model; analog noise and variable speed and pitch are used for data enhancement to further improve the robustness of the model.

4. The flexible line reconfiguration method fusing GraphRAG and improved Whisper of claim 1, wherein, The GraphRAG serves as a task decomposer of the ControlLLM, which decomposes the tasks of the new workpiece / process into basic and clearly defined subtasks, specifically including: The task decomposition process includes two parts: knowledge graph construction and global search optimization: The knowledge graph construction process: divide the input corpus into multiple text units, each containing detailed information about the new workpiece / process requirements and solutions; extract entities, relationships, and key statements from the text units using LLM to construct an initial knowledge graph; use the Leiden algorithm to perform hierarchical clustering on the initial knowledge graph, assign entities to different communities for in-depth analysis; generate summaries for each community to provide an overview and contextual information of the dataset; The global search optimization process: the new workpiece / process requirements serve as the initial input; specify the hierarchical level as the community, use semantic retrieval techniques to retrieve and summarize community summaries based on the LLM, generate community reports, which are shuffled and divided into multiple batches; each batch of community reports is divided into text blocks of a predefined size, and rating intermediate responses are generated based on the LLM; rank and filter the intermediate responses to select the most important points to generate aggregated intermediate responses; use the aggregated intermediate responses as contextual input to the LLM to generate the clear subtasks.

5. The flexible line reconfiguration method of fusing GraphRAG and improved Whisper as claimed in claim 4, wherein, The input corpus is the historical accumulation of new workpiece / process requirement and solution text.

6. The flexible line reconfiguration method of fusing GraphRAG and improved Whisper as claimed in claim 4, wherein, Extract entities from the text units using LLM, including: The name of the new workpiece / process, the name / designation of the modular hardware, the execution action description of the module, the scope of the module, the input parameters of the module, the input materials / intermediates of the module, and the output workpiece / intermediates of the module.

7. The flexible line reconfiguration method of fusing GraphRAG and improved Whisper as claimed in claim 4, wherein, The LLM is the Qwen3-7B model, which completes the tasks of extracting entities, relationships, and key statements from the text units, generating community reports from community summaries, generating rating intermediate responses, and generating clear subtasks based on prompts.

8. The flexible line reconfiguration method of fusing GraphRAG and improved Whisper according to claim 6, wherein, Generate the clear subtasks, which specifically include: the execution action description of the subtask, the scope of the module required to execute the subtask, the input parameters of the module, the input materials / intermediates of the module, and the output workpiece / intermediates of the module.

9. The flexible line reconfiguration method fusing GraphRAG and improved Whisper of claim 1, wherein, The solution specifically includes: Module combination scheme, control logic, and material flow path: the module combination scheme includes the combination of modules used, the logical relationship between module execution, input parameters of the module, input materials / intermediates of the module, and output workpiece / intermediates of the module; the control logic corresponds to the logical relationship between module execution in the module combination scheme; the material flow path corresponds to the logical relationship of the input materials of the module in the module combination scheme, which is consistent with the logical relationship between module execution.

Images