Multi-agent collaborative industrial document automatic generation method and system

Abstract

The invention discloses a multi-agent collaborative industrial document automatic generation method and system, and the method comprises the steps: carrying out the sampling data extraction of each accounting unit according to a task execution instruction of an agent; the collaborative task is disassembled into subtasks, and clustering is carried out on the subtasks according to sampling data generated by each agent; on the basis of the data features of the subtasks in the clustering result, generating draft fragments of the document for each subtask; and performing knowledge injection on the draft segment by using evolutionary memory to generate a final industrial document. And the final document has uniformity, specialty and traceability in the aspects of structure, terms and semantics, is suitable for an automatic filing and report generation scene in a large-scale industrial collaborative environment, and has good universality and engineering deployment value.

Description

Technical Field

[0001] The present invention relates to the technical field of document automatic generation, and particularly to a method and system for automatic generation of industrial documents with multi-agent collaboration. Background Art

[0002] With the rapid development of industrial intelligence and process automation, the acquisition frequency and dimension of industrial field data are continuously increasing, and a large number of distributed agents collaborate in scenarios such as equipment control, status monitoring, and energy consumption optimization. Traditional industrial documents (such as operation records, operation reports, debugging logs, etc.) written manually or with static templates can no longer meet the requirements of rapid archiving and information reconstruction in a high-frequency, multi-source, and multi-task collaboration environment.

[0003] Currently, mainstream industrial document generation methods generally rely on preset structural templates, manual scripting rules, or static language models, lacking the ability to understand dynamic task contexts and being difficult to process heterogeneous data streams generated in multi-agent concurrent tasks. In addition, the control logics of different agents during task execution have complex timing and causal relationships, resulting in difficulty for traditional document structures based on time segments or static process descriptions to accurately restore actual control behaviors and data evolution processes. Summary of the Invention

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

[0005] Therefore, the technical problems solved by the present invention are: existing methods for generating industrial documents under multi-agent collaborative tasks have problems such as difficult automatic generation, inconsistent structures, and non-uniform semantics.

[0006] To solve the above technical problems, the present invention provides the following technical solution: A method for automatic generation of industrial documents with multi-agent collaboration, including: Extracting sampled data for each accounting unit respectively according to the task execution instructions of the agents; Decomposing the collaborative task into subtasks, and clustering on the subtasks according to the sampled data generated by each agent; Generating draft fragments of documents for each subtask based on the data characteristics of the subtasks in the clustering results; Using evolutionary memory to inject knowledge into the draft fragments to generate the final industrial document; The accounting unit includes agents that complete the same collaborative task during the execution stage of the collaborative task.

[0007] As a preferred solution of the industrial document automatic generation method with multi-agent collaboration according to the present invention, wherein: the task execution instruction includes an operational control command sent to each agent by the control platform, driving the agent to perform specific functional behaviors in the collaborative task; The sampled data includes, during the process of the agent executing the task, collecting the operation data according to the sampling frequency of each type of data, obtaining a data set with time stamps for each task execution instruction.

[0008] As a preferred solution of the industrial document automatic generation method with multi-agent collaboration according to the present invention, wherein: the subtask includes a task segment formed by disassembling the collaborative task according to preset functional partitions, process partitions, and data source partitions respectively; Suppose 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, cluster the three types of subtasks respectively to obtain three clustering results.

[0009] As a preferred solution of the industrial document automatic generation method with multi-agent collaboration according to the present invention, wherein: for the clustering process of the data source partition, cluster according to the labels of the data sources respectively to obtain the clustering result; The specific clustering process for the functional partition and the process partition is 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 construct a candidate set according to the attribution situation; the kth control instruction The candidate set of is ; wherein, represents the vth subtask in the t partition; Step 2: Suppose each sampled data can only be copied to one subtask, and obtain all the copy schemes for the sampled data according to the mapping relationship between the control instruction and the subtask; Step 3: Randomly combine the copy schemes of each sampled data to obtain all the combination schemes of each sampled data copy scheme; wherein, each combination scheme is a clustering result of all the sampled data; Step 4: For each combination scheme, calculate the combination fitness of the combination scheme and the fitness of each subtask in the combination scheme; on the premise that the fitness of each subtask satisfies the corresponding preset constraint value, output the combination scheme with the maximum combination fitness as the clustering result; For the sampled data marked with the control instruction h, after feature encoding using LSTM, the feature vector h1 is obtained; using BERT to perform text encoding on the control instruction h, the feature vector h2 is obtained; after concatenating the feature vector h1 and the feature vector h2, the input feature regarding the operation instruction h is obtained. After integrating the feature inputs of each operation instruction in the subtask, the input feature set is obtained; inputting the input feature set and the inherent vector of the subtask into the pre-trained lightweight multi-modal semantic matching network; outputting the matching probability between the subtask and the data within the subtask as the fitness of the subtask. The inherent vector includes the vector representation pre-constructed by each subtask according to the preset partition. The combined fitness is calculated as the sum of the fitnesses of each subtask.

[0010] As a preferred scheme of the multi-agent collaborative industrial document automatic generation method of the present invention, wherein: the data features include the timestamp corresponding to each data and the corresponding control instruction.

[0011] As a preferred scheme of the multi-agent collaborative industrial document automatic generation method of the present invention, wherein: the draft fragment includes constructing a data chain in series on each control instruction dimension according to the timestamp in the dimension of the control instruction for the sampled data. 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.

[0012] As a preferred scheme of the multi-agent collaborative industrial document automatic generation method of the present invention, wherein: the evolutionary memory includes analyzing the correlation relationship between two control instructions according to the prior probabilities of any two control instructions under different consideration parameters in different tasks recorded in the history. The consideration parameters include three dimensions: the first dimension, the second dimension, and the middle dimension; the first dimension and the second dimension are respectively: the length of the time interval between the previous and subsequent control instructions divided by the task volume corresponding to the control instruction. If the time intervals of two control instructions overlap, the middle dimension is empty; if the time intervals of two control instructions do not overlap, the middle dimension L is expressed as: ; wherein, represents the time interval between the control instruction k and the control instruction W; Y represents the number of control instructions between the control instruction k and the control instruction W; y represents the control instruction index between the control instruction k and the control instruction W. represents the task volume corresponding to the y-th control instruction. Between any two control instructions in the subtasks, calculate the consideration parameters, and generate the occurrence probability of the association relationship according to the reference parameters; and use the generated probability as the attention weight between the two control instructions. Use the draft fragment to perform similarity matching in the historical industrial document data, and use the historical documents with a matching degree exceeding the preset threshold as support samples and input them into the meta-learning model; the meta-learning model generates fast adaptation parameters for the current subtask according to the semantic features of the draft fragment and the attention relationship of the control instructions; input the adaptation parameters, the attention weights between the draft fragment and the control instructions into the artificial intelligence document generation model, and perform knowledge injection during the generation process to output the final industrial document. After performing knowledge injection on the three clustering results respectively, obtain the final industrial documents for the three task splitting results.

[0013] An industrial document automatic generation system for multi-agent collaboration using any method described in the present invention, wherein: a collection unit extracts sampled data for each accounting unit according to the task execution instructions of the agent. A clustering unit disassembles the collaborative task into subtasks and performs clustering on the subtasks according to the sampled data generated by each agent. A fragment generation unit generates a draft fragment of the document for each subtask based on the data features during the clustering process. An analysis unit uses evolutionary memory to perform knowledge injection on the draft fragment to generate the final industrial document.

[0014] A computer device includes: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, the steps of any method described in the present invention are implemented.

[0015] A computer-readable storage medium stores a computer program, wherein: when the computer program is executed by a processor, the steps of any method described in the present invention are implemented.

[0016] Advantages 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 decomposing the task into subtasks and constructing a data chain by combining control instructions and sampled data. By introducing an evolutionary memory mechanism, the logical relationship between control instructions can be modeled based on historical task data, and knowledge injection can be provided during the generation stage, effectively improving the professionalism and integrity of the document. The system is made to have cross-task migration ability by quickly adapting to historical documents through a meta-learning method. The attention mechanism is used to model the relationship between control instructions, enhancing the logical coherence between paragraphs. The final document has unity, professionalism, and traceability at the structural, terminological, and semantic levels, is applicable to the scenarios of automatic archiving and report generation in a large-scale industrial collaboration environment, and has good generality and engineering deployment value. Description of the Drawings

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other accompanying drawings without creative efforts based on these drawings.

[0018] Figure 1 It is the overall flowchart of a multi-agent collaborative industrial document automatic generation method provided for the first embodiment of the present invention. Detailed Embodiments

[0019] To make the above objects, features, and advantages of the present invention more obvious and understandable, the detailed embodiments of the present invention will be described below with reference to the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0020] Embodiment 1, referring to Figure 1 , which is an embodiment of the present invention, provides a multi-agent collaborative industrial document automatic generation method, including: S1: Extract sampled data for each accounting unit respectively according to the task execution instructions of the agent.

[0021] It should be noted that the accounting unit includes agents that complete the same collaborative task during the execution phase of the collaborative task. The task execution instructions include operational control commands issued by the control platform to each agent, driving the agent to perform specific functional behaviors in the collaborative task. The sampled data includes data collected from the running data according to the sampling frequency of each type of data during the agent's task execution, resulting in a set of data with timestamps for each task execution instruction. The "collaborative task" refers to a complex control task initiated by the task scheduling system and completed by multiple agents collaborating within the same time period. This task is usually composed of a set of operational control instructions with associated relationships, where different instructions are assigned to different agents for execution, and there are temporal dependencies or functional coordination relationships among the agents during execution. Collaborative tasks have characteristics such as multi-agent parallel participation, unified task goals, and strong process complementarity. Their completion depends on multiple execution units jointly responding, collecting, and feedbacking data during a specific time window. As the basic analysis unit for document generation, collaborative tasks can achieve a complete expression of complex control behaviors and a unified modeling of multi-source task data, supporting subsequent task decomposition, behavior chain construction, and document structure mapping and other processing flows.

[0022] Furthermore, by treating all agents in the execution phase of the same collaborative task as an accounting unit, it is possible to unify the time window of the sampled data and the task association context, facilitating subsequent alignment and clustering analysis in the control dimension. At the same time, the sampled data is organized at the granularity of task execution instructions and has accurate timestamps, which can truly reflect the occurrence, duration, and feedback process of each control behavior, providing basic support for subsequent construction of behavior chains, derivation of control logic relationships, and semantic expression segments under the control instruction dimension. This ensures the restoration ability of the document generation system for control behaviors and the traceability of the task execution process.

[0023] S2: Decompose the collaborative task into subtasks and perform clustering on the subtasks based on the sampled data generated by each agent.

[0024] To solve the problem that collaborative tasks have heterogeneous expression requirements from different perspectives, by introducing a multi-dimensional decomposition and parallel clustering mechanism, the generation results of industrial documents are made more consistent with the actual task execution structure. The subtasks include task segments formed after decomposing the collaborative task according to preset functional partitions, process partitions, and data source partitions (i.e., partitions for different agents). Let the number of functional partitions be n1, the number of process partitions be n2, and the number of data source partitions be n3; then the number of subtasks is: n1 + n2 + n3.

[0025] When clustering the sampled data, the three types of subtasks are clustered separately to obtain three clustering results. By independently clustering these three types of subtasks, we can obtain three different dimensions of task expression views. Then, we perform knowledge injection and generation processing on each type of clustering result, and can output industrial documents covering three perspectives of "functional logic", "process path" and "role behavior", enhance the versatility, decomposability and pertinence of the document content, and provide a better structural foundation for archiving, auditing and analysis of complex tasks.

[0026] The clustering process of the data source partition is to cluster the data according to the label of the data source to obtain the clustering result. The clustering process of the function partition and process partition is 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 according to 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, each type of control instruction is clearly corresponded to the subtask to which it belongs based on the functional logic, execution process and agent responsibilities of the collaborative task. This mapping relationship is defined by engineers 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 a basis for sampling data attribution judgment and clustering candidate set construction, thereby ensuring the accuracy, consistency and interpretability of task division.

[0027] 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.

[0028] Step 3: Randomly combine the copy schemes of each sampled data to obtain all the combination schemes of each sampled data copy scheme; each combination scheme is a clustering result of all sampled data. In the case that each sampled data has multiple possible attributes, all possible clustering results of "data → subtask" are listed by combination enumeration. The one with the best fitness will be selected as the final clustering result.

[0029] 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.

[0030] For the sampled data marked with the control instruction h, after feature encoding using LSTM, the feature vector h1 is obtained; the control instruction h is text-encoded using BERT to obtain the feature vector h2; after concatenating the feature vector h1 and the feature vector h2, the input feature regarding the operation instruction h is obtained. After integrating the feature inputs of each operation instruction in the subtask, the input feature set is obtained; the input feature set and the inherent vector of the subtask are input into the pre-trained lightweight multi-modal semantic matching network; the matching probability between the subtask and the data within the subtask is output as the fitness of the subtask.

[0031] By aligning and fusing the sampled data and the control instruction in the feature space, the natural discreteness problem between the two types of input sources in terms of modal type and structural dimension can be effectively overcome. Specifically, the sampled data retains the change trend and dynamic response characteristics in the time series after being encoded by the LSTM network, while the control instruction extracts high-dimensional semantic information through the BERT network. After concatenating the two, a unified representation is formed, enabling the model to capture the composite features of "execution content" and "execution status" simultaneously. This method not only improves the accuracy of semantic matching but also reduces the sensitivity of the model to the input format, thus having stronger generalization ability. Further, using the integrated control instruction input feature and the subtask inherent vector to jointly input the lightweight feedforward neural network helps to complete efficient and accurate matching judgment while keeping the computational cost controllable, and realizes the two-way correlation evaluation of the task structure and data content. Compared with the traditional rule-based clustering adaptation method, this method can obtain a better subtask division effect through pre-training, and has stronger semantic adaptability, expression interpretability and engineering deployment value.

[0032] In this embodiment, the lightweight multi-modal semantic matching network belongs to the structure of a feedforward neural network (FNN). The front end extracts features through a multi-modal encoder, and the back end uses a fully connected layer to complete the fitness scoring. Overall, it is a lightweight multi-input binary classification network based on BERT embedding + numerical feature concatenation.

[0033] The inherent vector includes the vector representation pre-constructed by each subtask according to the preset partition. The combined fitness is calculated as the sum of the fitness of each subtask.

[0034] During the process of decomposing complex tasks, it is difficult to directly attribute sampled data to a single subtask, and there are problems of semantic overlap and blurred boundaries in control behaviors. By constructing a candidate attribution set and combining evaluations, the risk of misjudgment caused by static rule division can be avoided; introducing a semantic matching model for fitness calculation is to improve the understanding ability of clustering, so that task behaviors can be more accurately divided in terms of function or process dimensions; finally, through the screening of the optimal combined solution, it is ensured that each type of subtask has stronger structural consistency and semantic aggregation after clustering, thus establishing a clear and interpretable task view for subsequent document generation.

[0035] S3: Based on the data characteristics of subtasks in the clustering results, generate draft fragments of documents for each subtask.

[0036] The data characteristics include the timestamp corresponding to each data and the corresponding control instruction. The draft fragments include constructing a data chain in series on each control instruction dimension according to the timestamp of the sampled data in the dimension of the control instruction. 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.

[0037] It should be noted that by classifying the sampled data according to the associated control instructions and concatenating them in timestamp order to form an instruction-level data chain, the data evolution process of each control behavior in the task can be completely restored. At the same time, by unifying the time axis alignment of each data chain in combination with the start and end time information of the control instructions, the structural break caused by the asynchronous execution of the agent or inconsistent data sampling frequency can be avoided, so that the document fragments have good continuity and comparability at the time level. This mechanism provides a clear behavioral basis and timing support for subsequent semantic expression, paragraph generation and structural integration, ensuring that the generated document draft fragments have engineering characteristics such as traceability, reusability and clear semantics, and improving the logical integrity and behavioral visualization quality of the document content.

[0038] S4: Use evolutionary memory to inject knowledge into the draft fragments to generate the final industrial document.

[0039] The evolutionary memory includes analyzing the correlation relationship between any two control instructions under different consideration parameters in different tasks recorded in the history.

[0040] The consideration parameters include three dimensions: the first dimension, the second dimension and the intermediate dimension; the first dimension and the second dimension are respectively: the length of the time interval between the previous and subsequent control instructions divided by the task volume corresponding to the control instruction.

[0041] Further, if there is an overlap in the time intervals of two control instructions, the intermediate dimension is empty; if there is no overlap in the time intervals of two control instructions, the intermediate dimension L is expressed as: ; where 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; represents the task volume corresponding to the y-th control instruction.

[0042] For any two control instructions in the subtasks, calculate the consideration parameters, and generate the occurrence probability of the association relationship according to the consideration parameters; and use the generated probability as the attention weight between the two control instructions.

[0043] By introducing a time scale factor normalized by task load and a cross-period density-aware intermediate dimension, a fine-grained modeling of the behavioral correlation between control instructions is achieved, breaking through the limitations of traditional single-time-distance or instruction-order-based evaluations. The first dimension and the second dimension respectively normalize the execution interval length of the control instruction and the task volume it carries, enabling the model to have the ability to perceive task intensity and execution rhythm when constructing attention weights. The intermediate dimension introduces a "gap density" quantization mechanism on the premise that the instructions do not overlap. By comprehensively considering the time span and the control density during this period, the measurement accuracy of the potential dependence relationship between instructions is enhanced. The consideration parameters composed of this three-dimensional structure can be directly converted into association probability scores and used as the weight input of the attention mechanism, so as to achieve an interpretable modeling of the internal coupling of the control process without introducing deep sequence modeling. This strategy not only improves the context consistency of the control instruction sequence in the document generation process, but also optimizes the abstract ability of the evolutionary memory for the control behavior chain, significantly enhancing the structural accuracy of knowledge injection and the task adaptation robustness.

[0044] Use the draft fragment to perform similarity matching in the historical industrial document data, and use the historical documents with a matching degree exceeding the preset threshold as support samples and input them into the meta-learning model; the meta-learning model generates fast adaptation parameters for the current subtask according to the semantic features of the draft fragment and the attention relationship of the control instructions; input the adaptation parameters, the draft fragment, and the attention weight between the control instructions into the artificial intelligence document generation model, perform knowledge injection during the generation process, and output the final industrial document.

[0045] After performing knowledge injection on the three clustering results respectively, the final industrial documents for the three task splitting results are obtained.

[0046] By analyzing the behavioral characteristics and correlation probabilities between control instructions in historical tasks, a multi-dimensional correlation modeling mechanism is constructed, which can quantify the temporal tightness and operational logic correlation of different control behaviors during task evolution, and then form an interpretable attention structure to strengthen the behavioral chain logic of the current segment. Introducing a meta-learning module aims to quickly adapt to the historical document style and structure distribution for the current subtask, and extract fine-tuning strategies from experience transfer. By jointly inputting the draft segment, instruction attention weights, and adaptation parameters into the document generation model, deep injection in multiple aspects such as semantic alignment, term reuse, and unified expression style is realized, ensuring that the final document has semantic integrity, task restoration degree, and style consistency. In addition, by separately injecting the clustering results of three types of subtasks, namely function, process, and source, industrial documents with differentiated structures can be generated from multiple perspectives, realizing panoramic expression and structured archiving of behavior descriptions, and further enhancing the practical value of the document in analysis, auditing, and retrospective scenarios.

[0047] Embodiment 2. This embodiment also provides an intelligent agent collaborative industrial document automated generation system, which includes: A collection unit, according to the task execution instructions of the intelligent agent, extracts sampling data for each accounting unit respectively.

[0048] A clustering unit, disassembles the collaborative task into subtasks, and performs clustering on the subtasks according to the sampling data generated by each intelligent agent.

[0049] A fragment generation unit, based on the data characteristics during the clustering process, generates draft segments of the document for each subtask.

[0050] An analysis unit, using evolutionary memory, injects knowledge into the draft segment to generate the final industrial document.

[0051] If the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this 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 to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0052] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a definitional sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatuses, or devices. For the 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 combination with an instruction execution system, apparatus, or device.

[0053] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then storing it in a computer memory.

[0054] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0055] Example 3, which is an embodiment of the present invention, provides a method for automatically generating industrial documents with multi-agent collaboration. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0056] 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 agents: an assembly robotic arm (A), an inspection camera (B), a mobile transportation robot (C), a process data monitoring unit (D), and an intelligent alarm unit (E).

[0057] The goal is to verify the actual effects of the described automated document archiving solution in terms of multi-perspective traceability of tasks, anomaly recognition, and process optimization. The experiment relies on the MES (Manufacturing Execution System) to issue collaborative tasks, and comprehensively compares the performance differences between the existing stand-alone logging system and the solution of the present invention in the whole process of real-time data collection, encoding, clustering, and document generation.

[0058] The experimental steps are as follows: Experiment initialization and task assignment: The production scheduling system uniformly issues the "intelligent assembly of pump bodies" task to five types of intelligent agents. Each intelligent agent automatically receives independent instructions. For example, robotic arm A performs the grasping and tightening of the pump housing, inspection camera B conducts appearance inspection, robot C is responsible for part turnover, monitoring unit D records process parameters such as temperature / pressure in real time, and intelligent alarm unit E captures and reports assembly anomalies. The task decomposition scheme is as follows: functional partition (tightening / inspection / transportation / monitoring / alarm), process partition (pick-up → assembly → inspection → monitoring → anomaly response), and data source partition (five intelligent agents A~E).

[0059] Sampling data collection: All intelligent agents synchronously collect data such as task status, process parameters, and anomaly signals at a frequency of once every 5 seconds, and add accurate timestamps.

[0060] Robotic arm A: grasping torque (Nm), assembly angle (°), action status.

[0061] Inspection camera B: number of defects, type of defects, inspection results.

[0062] Transport robot C: moving distance (m), transportation time (s), path ID.

[0063] Monitoring unit D: pump body temperature (°C), pressure (kPa).

[0064] Intelligent alarm E: type of anomaly, occurrence time, duration.

[0065] Taking the assembly of a batch of pump bodies as an example, the sampling period covers all process nodes. The following are some examples of the original collected data: At the 15th minute, robotic arm A collected an assembly angle of 85° and a torque of 18 Nm; inspection camera B detected 2 defects, with the type being "scratch + deformation"; robot C moved 8 m, took 22 s, and the path was P1; monitoring unit D monitored a temperature of 56.2 °C and a pressure of 220 kPa; intelligent alarm E reported "assembly anomaly" once, lasting for 5 s.

[0066] Feature Fusion and Multi-dimensional Clustering: All sampled data are first encoded for temporal behavior by LSTM, and the control instruction text is encoded by BERT. After the two types of features are concatenated, they are fed into a lightweight feed-forward neural network to output the belonging probabilities of each data to each subtask. Clustering and archiving are performed separately for the three partitions of function, process, and source, and the solution with the highest fitness is selected. For example, the third group of data of Robot A (angle 84°, torque 17 Nm, status "tightening") is clustered into "Assembly - Process Partition - Robot Arm" after feature fusion.

[0067] Document Behavior Chain and Draft Generation: The clustering results are classified by task partition, and the five types of data A - E are concatenated in a timeline alignment manner, and a complete behavior chain is generated for each subtask. For example, the "detection" subtask in the process partition aggregates all the detection data of Camera B at the 5th, 10th, 15th, and 20th minutes, and organizes them into a document fragment in chronological order.

[0068] Evolutionary Memory and Knowledge Injection: The system analyzes the time, task volume distribution, and density of detection data and anomaly records in previous batches, calculates the attention weights between new task control instructions, selects the historical document most matching the current fragment for knowledge transfer and expression optimization, and outputs a multi-perspective archived document. For example, the warning fragment of the 3rd batch of assembly is highly similar to the historical anomaly document of the 1st batch, and its archiving format is automatically reused to improve document consistency.

[0069] Experimental Evaluation and Performance Collection: Taking 5 batches of assembly tasks as samples, comparing the proposed solution of the present invention with the existing single-entity archiving system, the following main performance indicators are collected: archiving completeness rate, anomaly recognition accuracy, traceability duration, behavior chain integrity, multi-perspective coverage, knowledge injection score, and archiving delay. Specifically as shown in Table 1.

[0070] Table 1 Data Record Table

[0071] It can be seen from the tabular data that the solution of the present invention has achieved a comprehensive optimization of the archiving process in multiple batches of tasks on the intelligent pumping station assembly line, and is significantly superior to the traditional single-entity archiving system in multiple key indicators.

[0072] First of all, the archiving completeness rate (average 85.4%) has increased by about 3.6% compared with the traditional system, mainly due to the structural innovation of multi-agent division of labor and multi-perspective task clustering, which reduces process omissions and information breaks.

[0073] The anomaly recognition accuracy has also increased from 77.4% to 82.5%. By unifying and correlating multi-source data such as anomaly signals and detection defects and introducing the historical evolution memory and meta-learning adaptive mechanism, the system can identify and archive various assembly anomalies in a timely and accurate manner.

[0074] The average tracing duration is 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 control instructions, anomaly localization and tracing are more rapid, greatly improving production efficiency and safety guarantee.

[0075] The integrity of the behavior chain is significantly improved. The average of the invention solution is 75.5%, while the traditional system is only 70.2%. This reflects that the innovative sub-task clustering and time-axis alignment algorithms can completely reconstruct complex process links, facilitating process analysis and process optimization.

[0076] The multi-perspective coverage is fixed at 3, while the traditional one is only 2. This shows that the present invention can not only archive in terms of function and process dimensions, but also output from multiple perspectives according to the roles of agents, adapting to more complex auditing and tracing requirements.

[0077] The knowledge injection score is as high as 0.82, while the traditional one is only 0.57. Meta-learning and historical document migration make the document style, content structure highly consistent with the actual process knowledge, effectively avoiding semantic breaks and fragmented archiving.

[0078] The archiving delay is shortened to 15.5 seconds, while the traditional one is 23.8 seconds. This indicates that the system of the present invention has higher data archiving real-time performance and concurrency ability, and is applicable to large-scale industrial scenarios.

[0079] Generally speaking, through multi-agent collaboration, multi-modal feature fusion of data, multi-dimensional sub-task clustering and evolutionary knowledge injection, the method of the present invention comprehensively solves the pain points of the existing single-entity archiving system, such as scattered data, process faults, slow anomaly recognition and difficult migration of document knowledge.

[0080] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An industrial document automated generation method for multi-agent collaboration, characterized in that Including: Extract sampling data for each accounting unit respectively according to the task execution instructions of the agent; Decompose the collaborative task into subtasks, and perform clustering on the subtasks according to the sampling data generated by each agent; Generate draft fragments of documents for each subtask based on the data characteristics of the subtasks in the clustering results; Use evolutionary memory to inject knowledge into the draft fragments to generate the final industrial document; The accounting unit includes agents that complete the same collaborative task during the execution stage of the collaborative task.

2. The multi-agent collaborative industrial document automatic generation method according to claim 1, wherein: The task execution instructions include operational control commands sent to each agent by the control platform to drive the agent to perform specific functional behaviors in the collaborative task; The sampling data includes, during the process of the agent executing the task, collecting the running data according to the sampling frequency of each type of data to obtain a data set with time stamps for each task execution instruction.

3. The multi-agent collaborative industrial document automated generation method according to claim 2, characterized in that: The subtasks include task fragments formed by decomposing the collaborative task according to preset functional partitions, process partitions, and data source partitions respectively; Let , the number of functional partitions be n1, the number of process partitions be n2, and the number of data source partitions be n3; then the number of the subtasks is: n1 + n2 + n3; When clustering the sampling data, cluster the three types of subtasks respectively to obtain three clustering results.

4. The multi-agent collaborative industrial document automatic generation method according to claim 3, wherein: For the clustering process of the data source partition, cluster according to the labels of the data sources respectively to obtain the clustering result; The specific clustering process for the functional partition and the process partition is as follows: Step 1: For the set of control instructions for the current collaborative task, based on the mapping relationship between control instructions and subtasks, obtain the possible attribution of each control instruction in the subtasks, and construct a candidate set according to the attribution situation; the candidate set of the k-th control instruction is ; where represents the v-th subtask in the t partition; Step 2: Assume that each sampling data can only be copied to one subtask, and obtain all copy schemes according to the mapping relationship between the control instructions and the subtasks of the sampling data; Step 3: Randomly combine the copy schemes of each sampling data to obtain all combination schemes of each sampling data copy scheme; among them, each combination scheme is a clustering result of all sampling data; Step 4: For each combination scheme, calculate the combination 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 combination fitness as the clustering result; For the sampling data marked with the control instruction h, after feature encoding using LSTM, obtain the feature vector h1; use BERT to perform text encoding on the control instruction h to obtain the feature vector h2; after splicing the feature vector h1 and the feature vector h2, obtain the input feature regarding the operation instruction h, and after integrating the feature inputs of each operation instruction in the subtask, obtain the input feature set; input the input feature set and the inherent vector of the subtask into the pre-trained lightweight multi-modal semantic matching network; output the matching probability between the subtask and the data within the subtask as the fitness of the subtask; The inherent vector includes the vector representation pre-constructed by each subtask according to the preset partition; The combination fitness is calculated as the sum of the fitness of each subtask.

5. The method for automatically generating industrial documents with multi-agent collaboration according to claim 4, characterized in that: The data characteristics include the time stamp corresponding to each data and the corresponding control instruction.

6. The multi-agent collaborative industrial document automatic generation method according to claim 5, wherein: The draft segment includes constructing a concatenated data chain for each dimension of the control instruction based on the time stamp in the dimension of the sampled data. Align each data chain with the time scale on the time axis according to the time stamps of the start and end times of each control instruction.

7. The industrial document automated generation method with multi-agent collaboration according to claim 6, characterized in that: The evolutionary memory includes analyzing the correlation relationship between any two control instructions under different consideration parameters in different tasks recorded in the historical records. The consideration parameters include three dimensions: the first dimension, the second dimension, and the intermediate dimension; the first dimension and the second dimension are respectively: the length of the time interval between the previous and the next control instructions divided by the task volume corresponding to the control instruction. If the time intervals of two control instructions overlap, the intermediate dimension is empty; if the time intervals of two control instructions do not overlap, the intermediate dimension L is expressed as: ； Among them, 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; represents the task volume corresponding to the y-th control instruction; For any two control instructions in the subtask, calculate the consideration parameters, generate the occurrence probability of the correlation relationship according to the consideration parameters; and use the generated probability as the attention weight between the two control instructions. Use the draft segment to perform similarity matching in the historical industrial document data, and use the historical documents with a matching degree exceeding the preset threshold as support samples to be input into the meta-learning model; the meta-learning model generates fast adaptation parameters for the current subtask according to the semantic features of the draft segment and the attention relationship of the control instructions; input the adaptation parameters, the draft segment, and the attention weight between the control instructions into the artificial intelligence document generation model, and perform knowledge injection during the generation process to output the final industrial document. After performing knowledge injection on the three clustering results respectively, obtain the final industrial documents for the three task splitting results.

8. An industrial document automated generation system for multi-agent collaboration using the method according to any one of claims 1-7, characterized in that: The acquisition unit extracts sampled data for each accounting unit respectively according to the task execution instructions of the intelligent agent. The clustering unit disassembles the collaborative task into subtasks and performs clustering on the subtasks according to the sampled data generated by each intelligent agent. The segment generation unit generates draft segments of documents for each subtask based on the data features during the clustering process. The analysis unit uses the evolutionary memory to perform knowledge injection on the draft segment to generate the final industrial document.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of the method according to any one of claims 1-7 are implemented.

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

Images