A multi-agent collaborative industrial document automatic generation method and system

Abstract

The present invention discloses a method and system for automated industrial document generation using multi-agent collaboration, comprising: extracting sampled data from each accounting unit based on the agent's task execution instructions; breaking down collaborative tasks into subtasks and clustering the subtasks based on the sampled data generated by each agent; generating draft document fragments for each subtask based on the data characteristics of the subtasks in the clustering results; and injecting knowledge into the draft fragments using evolutionary memory to generate the final industrial document. The final document is unified, professional, and traceable in terms of structure, terminology, and semantics, making it suitable for automatic archiving and report generation scenarios in large-scale industrial collaborative environments, and possessing good versatility and engineering deployment value.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic document generation, and in particular to a multi-agent collaborative automatic industrial document generation method and system. Background Art

[0002] With the rapid development of industrial intelligence and process automation, the frequency and dimensionality of industrial field data collection are increasing. Massive distributed intelligent agents are collaborating in scenarios such as equipment control, condition monitoring, and energy optimization. Traditional industrial documents (such as operation records, operation reports, and debugging logs) written manually or using static templates can no longer meet the needs of rapid archiving and information reconstruction in high-frequency, multi-source, and multi-task collaborative environments.

[0003] Current mainstream industrial document generation methods generally rely on pre-set structural templates, manually scripted rules, or static language models. These methods lack the ability to understand dynamic task contexts and struggle to process the heterogeneous data streams generated by concurrent multi-agent tasks. Furthermore, the control logic of different agents during task execution has complex temporal and causal relationships, making it difficult for traditional document structures based on time segments or static process descriptions to accurately reproduce actual control behaviors and data evolution. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the existing industrial document generation method under multi-agent collaborative tasks has problems such as difficulty in automatic generation, inconsistent structure, and inconsistent semantics.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a multi-agent collaborative industrial document automatic generation method, comprising:

[0007] According to the task execution instructions of the intelligent agent, sample data is extracted from each accounting unit respectively;

[0008] Decompose the collaborative task into subtasks and perform clustering on the subtasks based on the sampled data generated by each agent;

[0009] Based on the data features of the subtasks in the clustering results, a draft fragment of the document is generated for each subtask;

[0010] Using evolutionary memory, knowledge is injected into the draft fragment to generate a final industrial document;

[0011] The accounting unit includes an intelligent agent that completes the same collaborative task during the execution phase of the collaborative task.

[0012] As a preferred embodiment of the method for automatically generating industrial documents through multi-agent collaboration of the present invention, the task execution instructions include operational control commands issued to each agent through the control platform to drive the agent to perform specific functional behaviors in the collaborative task;

[0013] The sampling data includes collecting the running data according to the sampling frequency of each data during the execution of the task by the intelligent agent, and obtaining a data set with a time stamp about each task execution instruction.

[0014] As a preferred solution of the multi-agent collaborative industrial document automatic generation method of the present invention, the subtasks include task segments formed by decomposing the collaborative tasks according to preset functional partitions, process partitions, and data source partitions;

[0015] Assume that the number of functional partitions is n1, the number of process partitions is n2, and the number of data source partitions is n3; then the number of subtasks is: n1+n2+n3;

[0016] When clustering the sampled data, the three types of subtasks are clustered separately to obtain three types of clustering results.

[0017] As a preferred solution of the multi-agent collaborative industrial document automatic generation method of the present invention, wherein: the clustering process of the data source partitions is to cluster the data according to the labels of the data sources to obtain the clustering results;

[0018] The clustering process of the functional partitions and process partitions is specifically as follows:

[0019] Step 1: For the control instruction set of the current collaborative task, according to the mapping relationship between the control instruction and the subtask, obtain the possible attribution of each control instruction in the subtask, and build a candidate set based on the attribution; the kth control instruction The candidate set is ;in, represents the vth subtask in the t partition;

[0020] Step 2: Assuming that each sampled data can only be copied to one subtask, all copying schemes are obtained according to the mapping relationship between the control instructions and the subtasks.

[0021] Step 3: Randomly combine the copy schemes of each sampled data to obtain all combination schemes of the copy schemes of each sampled data; each combination scheme is a clustering result of all the sampled data;

[0022] Step 4: For each combination scheme, calculate the combined fitness of the combination scheme and the fitness of each subtask in the combination scheme; on the premise that the fitness of each subtask meets the corresponding preset constraint value, output the combination scheme with the maximum combined fitness as the clustering result;

[0023] For sampled data labeled with control instruction h, feature encoding is performed using LSTM to obtain feature vector h1. BERT is used to perform text encoding on control instruction h to obtain feature vector h2. Feature vectors h1 and h2 are concatenated to obtain input features for operation instruction h. The feature inputs of each operation instruction in the subtask are integrated to obtain an input feature set. The input feature set and the intrinsic vector of the subtask are input into a pre-trained lightweight multimodal semantic matching network. The output is the matching probability between the subtask and the data within the subtask, which serves as the fitness of the subtask.

[0024] The intrinsic vector includes,a pre-constructed vector representation for each subtask according to,preset partitioning;

[0025] The combined fitness is calculated as the sum of the fitness of each subtask.

[0026] As a preferred solution of the multi-agent collaborative industrial document automatic generation method described in the present invention, the data features include the timestamp and corresponding control instructions corresponding to each data.

[0027] As a preferred solution of the multi-agent collaborative industrial document automatic generation method of the present invention, wherein: the draft fragment includes, in the dimension of control instructions, sampled data, and constructing a data chain connected in series in each control instruction dimension according to timestamps;

[0028] According to the timestamps of the start and end time of each control instruction, each data chain is aligned with the time scale on the time axis.

[0029] As a preferred solution of the multi-agent collaborative industrial document automatic generation method of the present invention, the evolutionary memory includes analyzing the correlation relationship between any two control instructions based on the prior probabilities of the two control instructions under different parameters in different tasks in the historical records;

[0030] The considered parameters include three dimensions: a first dimension, a second dimension, and an intermediate dimension; the first dimension and the second dimension are respectively: the length of the time interval between the two previous control instructions divided by the task amount corresponding to the control instruction;

[0031] If the time intervals of the two control instructions overlap, the intermediate dimension is empty; if the time intervals of the two control instructions do not overlap, the intermediate dimension L is expressed as:

[0032] ;

[0033] in, represents the time interval between control instruction k and control instruction W; Y represents the number of control instructions between control instruction k and control instruction W; y represents the control instruction index between control instruction k and control instruction W; Indicates the task amount corresponding to the y-th control instruction;

[0034] For any two control instructions in the subtask, the reference parameters are calculated and the probability of occurrence of the association relationship is generated according to the reference parameters; and the generated probability is used as the attention weight between the two control instructions;

[0035] The draft fragments are used to perform similarity matching in historical industrial document data, and historical documents with a matching degree exceeding a preset threshold are input as support samples into a meta-learning model; the meta-learning model generates fast adaptation parameters for the current subtask based on the semantic features of the draft fragments and the attention relationship between the control instructions; the adaptation parameters are input into an artificial intelligence document generation model together with the attention weights between the draft fragments and the control instructions, and knowledge is injected during the generation process to output the final industrial document;

[0036] After knowledge injection into the three clustering results respectively, the final industrial documents for the three task splitting results are obtained.

[0037] An industrial document automatic generation system using multi-agent collaboration using any of the methods described in the present invention, wherein: a collection unit extracts sampled data from each accounting unit according to the task execution instructions of the agent;

[0038] A clustering unit that breaks down the collaborative task into subtasks and performs clustering on the subtasks based on the sampled data generated by each agent;

[0039] The fragment generation unit generates draft fragments of documents for each subtask based on the data features in the clustering process;

[0040] The analysis unit uses evolutionary memory to inject knowledge into the draft fragments to generate a final industrial document.

[0041] A computer device comprises: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, the steps of any one of the methods of the present invention are implemented.

[0042] A computer-readable storage medium stores a computer program, wherein: when the computer program is executed by a processor, the steps of any one of the methods of the present invention are implemented.

[0043] Beneficial effects of the present invention: The multi-agent collaborative industrial document automatic generation method provided by the present invention realizes the structured expression of the task process by breaking down the task into subtasks and combining control instructions with sampled data to build a data chain. The introduction of the evolutionary memory mechanism can model the logical association between control instructions based on historical task data, and provide knowledge injection in the generation stage, effectively improving the professionalism and integrity of the document. The historical documents are quickly adapted through the meta-learning method, so that the system has the ability to migrate across tasks. The attention mechanism is used to model the relationship between control instructions, which enhances the logical coherence between paragraphs. The final document is unified, professional and traceable in terms of structure, terminology and semantics. It is suitable for automatic archiving and report generation scenarios in large-scale industrial collaborative environments, and has good versatility and engineering deployment value. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is an overall flow chart of a multi-agent collaborative industrial document automatic generation method provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0047] Example 1, reference Figure 1 , as one embodiment of the present invention, provides a multi-agent collaborative industrial document automatic generation method, comprising:

[0048] S1: According to the task execution instructions of the intelligent agent, sample data is extracted from each accounting unit.

[0049] It should be noted that the accounting unit includes the intelligent agents that complete the same collaborative task during the execution phase of the collaborative task. The task execution instructions include operational control commands issued to each intelligent agent through the control platform to drive the intelligent agent to perform specific functional behaviors in the collaborative task. The sampling data includes the collection of operating data according to the sampling frequency of each data during the execution of the task by the intelligent agent, and the acquisition of a data set with a timestamp about each task execution instruction. The "collaborative task" refers to a complex control task that is initiated by the task scheduling system and completed by multiple intelligent agents in the same time period. The task is usually composed of a set of operational control instructions with an associated relationship. Different instructions are assigned to different intelligent agents for execution. There is a timing dependency or functional coordination relationship between the intelligent agents during the execution process. Collaborative tasks have the characteristics of multi-agent parallel participation, unified task objectives, and strong process complementarity. Its completion depends on multiple execution units to jointly respond, collect and feedback data in the control process within a specific time window. As the basic analysis unit for document generation, collaborative tasks can achieve complete expression of complex control behaviors and unified modeling of multi-source task data, supporting subsequent processing processes such as task decomposition, behavior chain construction, and document structure mapping.

[0050] Furthermore, by treating all agents in the execution phase of the same collaborative task as a single accounting unit, the time window of the sampled data and the task-related context can be unified, facilitating subsequent alignment and cluster analysis in the control dimension. At the same time, the sampled data is organized at the granularity of task execution instructions and carries precise timestamps, which can truly reflect the occurrence, continuation, and feedback process of each control behavior. This provides basic support for the subsequent construction of behavioral chains under the control instruction dimension, the derivation of control logic relationships, and semantic expression fragments. This ensures the document generation system's ability to restore control behaviors and the traceability of the task execution process.

[0051] S2: Decompose the collaborative task into subtasks, and perform clustering on the subtasks based on the sampled data generated by each agent.

[0052] To address the heterogeneous expression requirements of collaborative tasks from different perspectives, a multi-dimensional decomposition and parallel clustering mechanism is introduced to ensure that the generated industrial documents are more closely aligned with the actual task execution structure. The subtasks are generated by decomposing the collaborative tasks according to pre-defined functional, process, and data source partitions (i.e., partitions for different agents). Assuming the number of functional partitions is n1, the number of process partitions is n2, and the number of data source partitions is n3; the number of subtasks is then: n1 + n2 + n3.

[0053] When clustering the sampled data, the three subtask categories are clustered separately, resulting in three clustering results. By independently clustering these three subtask categories, we can obtain three different dimensional views of the task. Furthermore, by injecting knowledge and generating processing into each clustering result, we can output industrial documents covering the three perspectives of "functional logic," "process path," and "role behavior." This enhances the multifaceted, decomposable, and targeted nature of the document content, providing a better structural foundation for archiving, auditing, and analyzing complex tasks.

[0054] The clustering process of the data source partition is to cluster the data according to the data source label to obtain the clustering result. The clustering process of the function partition and process partition is as follows:

[0055] Step 1: For the control instruction set of the current collaborative task, according to the mapping relationship between the control instruction and the subtask, obtain the possible attribution of each control instruction in the subtask, and build a candidate set based on the attribution; the kth control instruction The candidate set is ;in, Represents the vth subtask (possibly one) in the t partition. In this embodiment, the mapping relationship between the control instructions and the subtasks is a preset relationship, that is, in the system modeling stage, based on the functional logic, execution process and agent responsibilities of the collaborative task, each type of control instruction is clearly corresponded to the subtask to which it belongs. This mapping relationship is defined by the engineer in the task configuration template, for example, the "startup instruction" is attributed to the "equipment initialization function partition", the "parameter collection instruction" is attributed to the "data monitoring process partition", and the "equipment maintenance operation performed by agent A" is attributed to the "data source partition corresponding to the agent". This preset mapping relationship can be stored as a table, a configuration file or an embedded task parameter structure as the basis for sampling data attribution judgment and cluster candidate set construction, thereby ensuring the accuracy, consistency and interpretability of task division.

[0056] Step 2: Assuming that each sampled data can only be copied to one subtask, all copying schemes are obtained according to the mapping relationship between the control instructions and the subtasks.

[0057] Step 3: Randomly combine the copying schemes for each sampled data item to obtain all possible combinations of the copying schemes for each sampled data item. Each combination represents a clustering result for all sampled data items. If each sampled data item has multiple possible assignments, all possible clustering results for the "data → subtask" relationship are enumerated through combinations. The one with the best fitness is then selected as the final clustering result.

[0058] Step 4: For each combination scheme, calculate the combined fitness of the combination scheme and the fitness of each subtask in the combination scheme; on the premise that the fitness of each subtask meets the corresponding preset constraint value, output the combination scheme with the maximum combined fitness as the clustering result.

[0059] For the sampled data marked with the control instruction h, LSTM is used to perform feature encoding to obtain the feature vector h1; the control instruction h is text-encoded using BERT to obtain the feature vector h2; the feature vector h1 and the feature vector h2 are concatenated to obtain the input features of the operation instruction h, and the feature input of each operation instruction in the subtask is integrated to obtain the input feature set; the input feature set and the intrinsic vector of the subtask are input into a pre-trained lightweight multimodal semantic matching network; the matching probability between the subtask and the data within the subtask is output as the fitness of the subtask.

[0060] By aligning and fusing sampled data and control instructions in feature space, the inherent discreteness between the two input sources in terms of modality and structural dimensions can be effectively overcome. Specifically, the sampled data, after being encoded by an LSTM network, retains the changing trends and dynamic response characteristics of the time series, while the control instructions are extracted through a BERT network to extract high-dimensional semantic information. These two are then concatenated into a unified representation, enabling the model to capture the combined features of "execution content" and "execution status." This approach not only improves the accuracy of semantic matching but also reduces the model's sensitivity to input format, thereby enhancing generalization. Furthermore, the integrated control instruction input features and the subtask intrinsic vectors are jointly input into a lightweight feedforward neural network. This facilitates efficient and accurate matching while maintaining manageable computational overhead, enabling a bidirectional correlation evaluation between task structure and data content. Compared to traditional rule-based clustering and adaptation methods, this method achieves better subtask partitioning through pre-training, offering greater semantic adaptability, expressive interpretability, and engineering deployment value.

[0061] In this embodiment, the lightweight multimodal semantic matching network belongs to: Feedforward Neural Network (FNN) structure. The front end uses a multimodal encoder to extract features, and the back end uses a fully connected layer to complete fitness scoring. The overall structure is a lightweight multi-input binary classification network based on BERT embedding + numerical feature splicing.

[0062] The intrinsic vector includes a pre-constructed vector representation of each subtask according to a preset partition. The combined fitness is calculated as the sum of the fitness of each subtask.

[0063] When decomposing complex tasks, sampled data can be difficult to directly attribute to a unique subtask, and control behaviors can suffer from semantic overlap and blurred boundaries. By constructing candidate attribution sets and evaluating them together, we can avoid the risk of misjudgment caused by static rule-based divisions. The introduction of a semantic matching model for fitness calculation improves clustering comprehension, enabling more precise division of task behaviors along functional or process dimensions. Finally, by combining optimal solutions for screening, we ensure that each subtask type has greater structural consistency and semantic cohesion after clustering, thereby establishing a clear and interpretable task view for subsequent document generation.

[0064] S3: Based on the data features of the subtasks in the clustering results, a draft fragment of the document is generated for each subtask.

[0065] The data features include the timestamp and control instruction corresponding to each data item. The draft fragment includes the sampled data in the control instruction dimension, and constructs a data chain in series based on the timestamp for each control instruction dimension. Each data chain is aligned with the time scale on the time axis based on the timestamp of the start and end time of each control instruction.

[0066] It should be noted that by classifying the sampled data according to their associated control instructions and connecting them in series in timestamp order to form an instruction-level data chain, the data evolution process of each control behavior in the task can be fully restored. At the same time, by aligning the data chains on a unified time axis based on the start and end time information of the control instructions, structural faults caused by asynchronous execution of intelligent agents or inconsistent data sampling frequencies can be avoided, thereby ensuring that document fragments have good continuity and comparability at the temporal level. This mechanism provides a clear behavioral basis and timing support for subsequent semantic expression, paragraph generation, and structural fusion, ensuring that the generated document draft fragments have engineering characteristics such as traceability, reconstructability, and clear semantics, thereby improving the logical integrity of the document content and the quality of behavioral visualization.

[0067] S4: Using evolutionary memory, knowledge is injected into the draft fragment to generate the final industrial document.

[0068] The evolutionary memory includes analyzing the correlation between two control instructions based on the prior probabilities of any two control instructions under different consideration parameters in different tasks recorded in the history.

[0069] The considered parameters include three dimensions: a first dimension, a second dimension and an intermediate dimension; the first dimension and the second dimension are respectively: the length of the time interval between the two previous and next control instructions divided by the task amount corresponding to the control instructions.

[0070] Furthermore, if the time intervals of the two control instructions overlap, the intermediate dimension is empty; if the time intervals of the two control instructions do not overlap, the intermediate dimension L is expressed as:

[0071] ;

[0072] in, represents the time interval between control instruction k and control instruction W; Y represents the number of control instructions between control instruction k and control instruction W; y represents the control instruction index between control instruction k and control instruction W; Indicates the amount of tasks corresponding to the yth control instruction.

[0073] For any two control instructions in the subtask, a consideration parameter is calculated, and the probability of occurrence of the association relationship is generated according to the consideration parameter; and the generated probability is used as the attention weight between the two control instructions.

[0074] By introducing a time scaling factor based on task load and an intermediate dimension that perceives inter-period density, a fine-grained modeling of behavioral dependencies between control instructions is achieved, transcending the limitations of traditional evaluation methods based on single temporal distance or instruction sequence. The first and second dimensions respectively normalize the execution interval length of a control instruction with the amount of task it carries, enabling the model to perceive task intensity and execution rhythm when constructing attention weights. The intermediate dimension introduces a "gap density" quantification mechanism, assuming that instructions do not overlap. By comprehensively considering the time span and the control density between them, the model improves the accuracy of measuring potential dependencies between instructions. The parameters of this three-dimensional structure can be directly converted into association probability scores, which serve as weight inputs for the attention mechanism. This enables interpretable modeling of the intrinsic coupling of control flows without resorting to deep sequence modeling. This strategy not only improves the contextual consistency of control instruction sequences during document generation but also optimizes the abstraction of control behavior chains by evolving memory, significantly enhancing the structural accuracy and task adaptation robustness of knowledge injection.

[0075] The draft fragment is used to perform similarity matching in historical industrial document data, and the historical documents with a matching degree exceeding a preset threshold are input into the meta-learning model as supporting samples; the meta-learning model generates fast adaptation parameters for the current subtask based on the semantic features of the draft fragment and the attention relationship between the control instructions; the adaptation parameters and the attention weights between the draft fragment and the control instructions are input into the artificial intelligence document generation model, knowledge injection is performed during the generation process, and the final industrial document is output.

[0076] After knowledge injection into the three clustering results respectively, the final industrial documents for the three task splitting results are obtained.

[0077] By analyzing the behavioral characteristics and association probabilities between control instructions in historical tasks, a multi-dimensional association modeling mechanism is constructed. This can quantify the temporal closeness and operational logic relevance of different control behaviors during task evolution, and then form an explanatory attention structure to supplement the behavioral chain logic of the current segment. The introduction of the meta-learning module aims to quickly adapt the style and structural distribution of historical documents to the current subtask, and to extract fine-tuning strategies from experience transfer. By inputting the draft fragments, instruction attention weights, and adaptation parameters into the document generation model, deep injection is achieved in many aspects such as semantic alignment, term reuse, and unified expression style, ensuring that the final document has semantic integrity, task restoration, and style consistency. In addition, the clustering results of the three types of subtasks, namely function, process, and source, are injected separately, and industrial documents with differentiated structures can be generated from multiple perspectives, realizing panoramic expression and structured archiving of behavioral descriptions, further enhancing the actual value of documents in analysis, auditing, and backtracking scenarios.

[0078] Example 2: This embodiment also provides an agent-coordinated industrial document automatic generation system, which includes:

[0079] The acquisition unit extracts sampled data from each accounting unit according to the task execution instructions of the intelligent agent.

[0080] The clustering unit decomposes the collaborative task into subtasks and performs clustering on the subtasks based on the sampled data generated by each agent.

[0081] The fragment generation unit generates draft fragments of the document for each subtask based on the data features in the clustering process.

[0082] The analysis unit uses evolutionary memory to inject knowledge into the draft fragments to generate a final industrial document.

[0083] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0084] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0085] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0086] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the aforementioned embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or combination of the following technologies known in the art may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0087] Example 3 is an embodiment of the present invention, which provides a multi-agent collaborative industrial document automatic generation method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0088] This experiment takes the "intelligent pump station assembly line" in the intelligent manufacturing workshop as the object and constructs a multi-agent collaborative automatic archiving system. The system includes five main types of intelligent agents: assembly robot arm (A), inspection camera (B), mobile transport robot (C), process data monitoring unit (D) and intelligent alarm unit (E).

[0089] The goal was to verify the effectiveness of the proposed automated document archiving solution in multi-perspective task tracing, anomaly identification, and process optimization. The experiment leveraged an MES (Manufacturing Execution System) to issue collaborative tasks, covering the entire process of real-time data collection, encoding, clustering, and document generation. The performance comparison between the proposed solution and existing standalone logging systems was thorough.

[0090] The experimental steps are as follows:

[0091] Experiment initialization and task allocation:

[0092] The production scheduling system uniformly assigns the "pump body intelligent assembly" task to five types of intelligent agents. Each agent automatically receives independent instructions. For example, robotic arm A performs grasping and tightening of the pump casing, inspection camera B performs visual inspection, robot C handles component turnover, monitoring unit D records process parameters such as temperature and pressure in real time, and intelligent alarm unit E detects and reports assembly anomalies. The task breakdown is as follows: functional division (tightening / inspection / transportation / monitoring / alarm), process division (part collection → assembly → inspection → monitoring → anomaly response), and data source division (five intelligent agents A through E).

[0093] Sampling data collection:

[0094] All intelligent agents synchronously collect data such as task status, process parameters, abnormal signals, etc. at a frequency of once every 5 seconds and add precise timestamps.

[0095] A Robotic Arm: gripping torque (Nm), assembly angle (°), action status.

[0096] B Inspection camera: number of defects, defect type, and inspection results.

[0097] C transport robot: moving distance (m), transport time (s), path ID.

[0098] D Monitoring unit: pump body temperature (°C), pressure (kPa).

[0099] E Intelligent alarm: abnormality type, occurrence time, and duration.

[0100] Taking the assembly of a batch of pump bodies as an example, the sampling period covers all process nodes. The following is an example of some raw collected data:

[0101] At the 15th minute, robot arm A detected an assembly angle of 85° and a torque of 18 Nm. Robot arm B detected two defects of the "scratches + deformation" type. Robot C moved 8 meters in 22 seconds, along path P1. Robot arm D monitored a temperature of 56.2°C and a pressure of 220 kPa. The intelligent alarm E reported an "assembly abnormality" once, which lasted for 5 seconds.

[0102] Feature fusion and multi-dimensional clustering: All sampled data is first encoded using an LSTM to encode temporal behavior, and the control instruction text is encoded using a BERT. The two types of features are then concatenated and fed into a lightweight feedforward neural network, which outputs the probability of each data item being assigned to each subtask. Clustering and archiving are performed based on the function, process, and source partitions, with the most adaptable solution being selected. For example, the third set of data for robot arm A (angle 84°, torque 17 Nm, status "tightening") is clustered into the "assembly-process partition-robot arm" after feature fusion.

[0103] Document Behavior Chaining and Draft Generation: Clustering results are categorized by task partition, and the data for the five categories A through E are linked together in a timeline-aligned manner, generating a complete behavior chain for each subtask. For example, the "Detection" subtask in the process partition aggregates all detection data from camera B at the 5th, 10th, 15th, and 20th minutes, organizing it into document segments in chronological order.

[0104] Evolutionary Memory and Knowledge Infusion: The system analyzes the time, workload distribution, and density of detection data and anomaly records from previous batches, calculates the attention weights between new task control instructions, selects the historical documents that best match the current segment, performs knowledge transfer and expression optimization, and outputs multi-perspective archived documents. For example, the third batch of assembled alarm segments is highly similar to the first batch of historical anomaly documents, and their archive format is automatically reused to improve document consistency.

[0105] Experimental Evaluation and Performance Collection: Using five batches of assembly tasks as samples, we compared the proposed solution with an existing single-file archiving system. Key performance metrics were collected: archiving completeness, anomaly identification accuracy, tracing time, behavior chain completeness, multi-perspective coverage, knowledge injection score, and archiving latency. See Table 1 for details.

[0106] Table 1 Data record table

[0107]

[0108] It can be seen from the table data that the solution of the present invention has achieved comprehensive optimization of the archiving process in multi-batch tasks of the intelligent pump station assembly line, and is significantly better than the traditional single-body archiving system in many key indicators.

[0109] First, the archiving completeness rate (85.4% on average) is about 3.6% higher than that of traditional systems, mainly due to the structural innovation of multi-agent division of labor and collaboration and multi-perspective task clustering, which reduces process omissions and information gaps.

[0110] The accuracy of anomaly identification also increased from 77.4% to 82.5%. By integrating multiple data sources, such as abnormal signals and detected defects, and incorporating historical evolution memory and meta-learning adaptive mechanisms, the system can promptly and accurately identify and archive various assembly anomalies.

[0111] The average tracing time has been shortened to 122 seconds, which is more than 30 seconds faster than the traditional system (152 seconds). This shows that through the temporal reconstruction of the behavior chain and the time alignment of the control instructions, anomaly location and tracing are faster, greatly improving production efficiency and safety.

[0112] The completeness of the behavior chain has been significantly improved, with an average of 75.5% for the invention solution and only 70.2% for the traditional system. This reflects that the innovative subtask clustering and timeline alignment algorithms can completely reconstruct complex process chains, facilitating process analysis and process optimization.

[0113] The multi-perspective coverage is fixed at 3, while the traditional one is only 2. This shows that the present invention not only archives in the function and process dimensions, but also can output from multiple perspectives according to the role of the intelligent agent, adapting to more complex auditing and traceability needs.

[0114] The knowledge injection score reached 0.82, compared to a traditional score of 0.57. Meta-learning and historical document migration ensured that document style and content structure were highly consistent with actual process knowledge, effectively avoiding semantic discontinuity and archival fragmentation.

[0115] The archiving delay is shortened to 15.5 seconds, compared with the traditional 23.8 seconds, indicating that the system of the present invention has higher data archiving real-time and concurrency capabilities and is suitable for large-scale industrial scenarios.

[0116] Overall, the method of the present invention comprehensively solves the pain points of existing single-body archiving systems, such as data dispersion, process faults, slow anomaly recognition and difficulty in document knowledge migration, through multi-agent collaboration, data multimodal feature fusion, multi-dimensional sub-task clustering and evolutionary knowledge injection.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multi-agent collaborative industrial document automatic generation method, characterized in that: include: According to the task execution instructions of the intelligent agent, sample data is extracted from each accounting unit respectively; Decompose the collaborative task into subtasks and perform clustering on the subtasks based on the sampled data generated by each agent; Based on the data features of the subtasks in the clustering results, a draft fragment of the document is generated for each subtask; Using evolutionary memory, knowledge is injected into the draft fragment to generate a final industrial document; The subtasks include task segments formed by decomposing the collaborative tasks according to preset functional partitions, process partitions, and data source partitions; Assume that the number of functional partitions is n1, the number of process partitions is n2, and the number of data source partitions is n3; then the number of subtasks is: n1+n2+n3; When clustering the sampled data, the three types of subtasks are clustered separately to obtain three types of clustering results; The draft fragment includes, the sampling data in the control instruction dimension, and constructing a data chain connected in series in each control instruction dimension according to the timestamp; According to the timestamps of the start and end times of each control instruction, each data chain is aligned with the time scale on the time axis; The accounting unit includes, in the execution phase of the collaborative task, agents that complete the same collaborative task; The evolutionary memory includes analyzing the correlation between two control instructions based on the prior probability of any two control instructions under different consideration parameters in different tasks in historical records; The considered parameters include three dimensions: a first dimension, a second dimension, and an intermediate dimension; the first dimension and the second dimension are respectively: the length of the time interval between the two previous control instructions divided by the task amount corresponding to the control instruction; If the time intervals of the two control instructions overlap, the intermediate dimension is empty; if the time intervals of the two control instructions do not overlap, the intermediate dimension L is expressed as: ； in, represents the time interval between control instruction k and control instruction W; Y represents the number of control instructions between control instruction k and control instruction W; y represents the control instruction index between control instruction k and control instruction W; Indicates the task amount corresponding to the y-th control instruction; For any two control instructions in the subtask, a consideration parameter is calculated, and a probability of occurrence of an association relationship is generated based on the consideration parameter; and the generated probability is used as the attention weight between the two control instructions; The draft fragments are used to perform similarity matching in historical industrial document data, and historical documents with a matching degree exceeding a preset threshold are input as support samples into a meta-learning model; the meta-learning model generates fast adaptation parameters for the current subtask based on the semantic features of the draft fragments and the attention relationship between the control instructions; the adaptation parameters are input into an artificial intelligence document generation model together with the attention weights between the draft fragments and the control instructions, and knowledge is injected during the generation process to output the final industrial document; After knowledge injection into the three clustering results respectively, the final industrial documents for the three task splitting results are obtained.

2. The multi-agent collaborative industrial document automatic generation method according to claim 1, characterized in that: The task execution instructions include operational control commands issued to each agent through the control platform to drive the agent to perform specific functional behaviors in the collaborative task; The sampling data includes collecting the running data according to the sampling frequency of each data during the execution of the task by the intelligent agent, and obtaining a data set with a time stamp about each task execution instruction.

3. The multi-agent collaborative industrial document automatic generation method according to claim 2, characterized in that: The clustering process of the data source partition is performed according to the label of the data source to obtain the clustering result; The clustering process of the functional partitions and process partitions is specifically as follows: Step 1: For the control instruction set of the current collaborative task, according to the mapping relationship between the control instruction and the subtask, obtain the possible attribution of each control instruction in the subtask, and build a candidate set based on the attribution; the kth control instruction The candidate set is ;in, represents the vth subtask in the t partition; Step 2: Assuming that each sampled data can only be copied to one subtask, all copying schemes are obtained according to the mapping relationship between the control instructions and the subtasks. Step 3: Randomly combine the copy schemes of each sampled data to obtain all combination schemes of the copy schemes of each sampled data; each combination scheme is a clustering result of all the sampled data; Step 4: For each combination scheme, calculate the combined fitness of the combination scheme and the fitness of each subtask in the combination scheme; on the premise that the fitness of each subtask meets the corresponding preset constraint value, output the combination scheme with the maximum combined fitness as the clustering result; For sampled data labeled with control instruction h, feature encoding is performed using LSTM to obtain feature vector h1. BERT is used to perform text encoding on control instruction h to obtain feature vector h2. Feature vectors h1 and h2 are concatenated to obtain input features for operation instruction h. The feature inputs of each operation instruction in the subtask are integrated to obtain an input feature set. The input feature set and the intrinsic vector of the subtask are input into a pre-trained lightweight multimodal semantic matching network. The output is the matching probability between the subtask and the data within the subtask, which serves as the fitness of the subtask. The intrinsic vector includes,a pre-constructed vector representation for each subtask according to,preset partitioning; The combined fitness is calculated as the sum of the fitness of each subtask.

4. The multi-agent collaborative industrial document automatic generation method according to claim 3, characterized in that: The data features include a timestamp and a corresponding control instruction corresponding to each data.

5. A multi-agent collaborative industrial document automatic generation system using the method according to any one of claims 1 to 4, characterized in that: The acquisition unit extracts sampled data from each accounting unit according to the task execution instructions of the intelligent agent; A clustering unit that breaks down the collaborative task into subtasks and performs clustering on the subtasks based on the sampled data generated by each agent; The fragment generation unit generates draft fragments of documents for each subtask based on the data features in the clustering process; The analysis unit uses evolutionary memory to inject knowledge into the draft fragments to generate a final industrial document.

6. A computer device comprising: memory and processor; The memory stores a computer program, wherein the processor implements the steps of the method according to any one of claims 1 to 4 when executing the computer program.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images