Vertical domain large model multi-agent collaborative reasoning system based on dynamic feedback

Abstract

The vertical field large model multi-agent collaborative reasoning system based on dynamic feedback comprises the following steps: dividing a process sub-sequence according to industrial field process information and setting monitoring points; constructing an industrial whole-process model; constructing a vertical field large model based on an industrial knowledge graph, a fault library, an optimization library and a process library, which contains three expert branches of equipment diagnosis, process adaptation and operation optimization; splitting a user task into sub-tasks and assigning the sub-tasks to corresponding agents to construct a dynamic collaborative network and call the expert branches to perform equipment state monitoring and reasoning; training a resource prediction model through historical data to dynamically adjust the cloud and industrial field computing power and resources; checking the initial reasoning result for facts, compliance and feasibility, generating a global execution scheme through conflict resolution and multi-dimensional fusion, and triggering re-reasoning or manual intervention if the quality threshold is not reached, thereby realizing whole-link closed-loop management of the industrial field, improving decision-making efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a multi-agent collaborative reasoning system for large-scale vertical domain models based on dynamic feedback. Background Technology

[0002] With the rapid development of industrial intelligence, artificial intelligence technology is increasingly widely used in the industrial field (such as equipment operation and maintenance, production optimization, quality control, and energy management), becoming a core driving force for promoting intelligent manufacturing upgrades, improving industrial production efficiency, and reducing production losses. However, existing technologies show significant shortcomings in the application of single large-scale models in the industrial field: First, the depth of industrial knowledge is insufficient; a single large-scale model cannot cover the detailed knowledge of equipment parameters, production processes, fault cases, and safety standards in the industrial field. Second, reasoning is prone to errors; equipment operation data in industrial scenarios is complex and subject to many interference factors, and a single large-scale model is not sufficiently adaptable to industrial data, easily leading to problems such as misjudgment of faults and unreasonable process optimization suggestions, and reasoning errors cannot be corrected in real time. Third, it cannot adapt to the real-time requirements of industry; the reasoning speed of a single large-scale model is slow, making it difficult to handle tasks with high real-time requirements such as equipment fault warnings and real-time process adjustments in industrial settings.

[0003] To address the limitations of a single large model, existing technologies have proposed multi-agent collaborative reasoning schemes, which involve multiple agents working together to complete complex industrial reasoning tasks. However, traditional multi-agent systems still have several shortcomings in industrial applications: First, they employ a static division of labor model, where the roles and responsibilities of agents are fixed, making it impossible to dynamically adjust based on the complexity of industrial tasks, changes in equipment operating status, and differences in fault types. This results in poor flexibility and difficulty in adapting to the dynamic needs of industrial scenarios. Second, the reasoning processes of each agent are independent, lacking effective collaborative optimization and synchronization with industrial knowledge. This prevents the system's reasoning capabilities from self-evolving and makes it difficult to adapt to the dynamic needs of industrial equipment updates, process upgrades, and the addition of new fault types. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a large-scale multi-agent collaborative reasoning system for vertical domains based on dynamic feedback, comprising a cloud layer, a perception layer, a simulation layer, a memory layer, a vertical domain base layer, a multi-agent collaborative layer, a resource allocation layer, and a dynamic feedback layer.

[0005] The perception layer is used to set up on-site data monitoring points based on the process information entered into the current industrial site data, and to collect indicator data at each point.

[0006] The simulation layer is used to construct a full-process industrial processing model based on the index data and processing scenario information at each point.

[0007] The memory layer is used to build knowledge graphs, fault libraries, optimization libraries, and process libraries for industrial sites;

[0008] The vertical domain base layer is used to build a large vertical domain model based on the memory layer. The large vertical domain model includes equipment diagnosis expert branches, process adaptation expert branches, and operation and maintenance optimization expert branches.

[0009] The multi-agent collaboration layer is used to construct various types of agents, break down the reasoning task, assign sub-tasks to corresponding agents, construct a dynamic collaboration network based on the agents paired with each sub-task, monitor the device status, and perform reasoning based on the expert branches paired with each type of agent in the dynamic collaboration network, and output the initial reasoning result.

[0010] The resource allocation layer is used to predict resources for the dynamic collaborative network, dynamically expand and shrink resources based on the resource prediction results, and allocate computing resources and industrial site resources for various types of intelligent agents in the dynamic collaborative network.

[0011] The dynamic feedback layer is used to perform fact verification, compliance verification, and feasibility verification on the initial inference results output by each expert branch. Based on the verification results, it generates inference results to be merged or performs expert branch re-inference. It performs structured decomposition and unified coding on the inference results to be merged, and performs conflict resolution analysis, core fact layer fusion, execution action layer fusion, effect layer fusion, and effect evaluation on the structured decomposition and unified coding results to generate a global execution plan.

[0012] Furthermore, the process of setting up on-site data monitoring points and collecting indicator data from each point in the perception layer includes:

[0013] Obtain the process information of the current industrial site data entry, extract the characteristics of the process unit based on the process information, and break down the industrial site data entry process into several process sub-sequences according to the characteristics of the process unit.

[0014] On-site data monitoring points are set up in each process subsequence, and industrial monitoring indicators of each on-site data monitoring point are obtained by data retrieval based on the functional attributes in the process unit characteristics of the corresponding process subsequence.

[0015] The on-site data monitoring points acquire indicator data in real time based on industrial monitoring indicators, mark the monitoring time, and set the monitoring cycle.

[0016] Furthermore, the process of constructing a full-process industrial processing model in the simulation layer includes:

[0017] The process involves acquiring the physical entities of each process subsequence in the current industrial field data entry process, constructing a digital space, creating 3D models of the physical entities of each process subsequence, mapping the 3D models to the digital space, setting up API interfaces on the 3D models of each process subsequence, preprocessing the data format of the indicator data collected from the field data monitoring points of the corresponding process subsequence, generating twin data based on the preprocessed indicator data, and matching the twin data with the 3D models of the corresponding process subsequences in the digital space to generate digital twin models.

[0018] Acquire different processing scenario information for each process subsequence, store the different processing scenario information in digital space, and combine the 3D models of each process subsequence in the digital twin model with the processing scenario information of each process subsequence to generate a full-process industrial processing model.

[0019] Furthermore, the process of building a large-scale vertical domain model from the vertical domain foundation layer includes:

[0020] Based on the memory layer, fault datasets, process datasets, and operation and maintenance datasets are constructed. These datasets are used as training samples. The training samples are preprocessed, data augmented, standardized, and multimodal fused to generate input samples. A large vertical domain model is constructed based on the MoE sparse architecture. The core input features and triggering features of each type of expert branch in the large vertical domain model are set. The input samples are compared with the core input features and triggering features of each type of expert branch to obtain the expert branch corresponding to the input sample. Then, the shared encoder is pre-trained and the expert branches are trained independently until the loss function is stable. The model parameters are saved, and the trained large vertical domain model is output.

[0021] Furthermore, the process of constructing a dynamic cooperative network based on the pairing of agents for each subtask includes:

[0022] The system acquires the inference task input by the user in the current industrial setting. Based on the core objective, industrial scenario constraints, and time benchmark in the inference task, the inference task is broken down into several sub-tasks. The objectives, constraints, priorities, and time requirements of each sub-task are determined. Based on the objectives, constraints, priorities, and time requirements of each sub-task, various types of intelligent agents are selected to obtain the core intelligent agent and auxiliary intelligent agent corresponding to each sub-task. A dynamic collaborative network is constructed based on the core intelligent agent, auxiliary intelligent agent, and task allocation intelligent agent of each sub-task.

[0023] In the dynamic network, each core agent inputs the industrial monitoring indicators of each process subsequence into the expert branch corresponding to its core agent, and outputs the initial inference results based on the expert branch.

[0024] Furthermore, the resource allocation layer dynamically expands and shrinks resources, allocating computing resources and industrial site resources for various types of intelligent agents in the dynamic collaborative network. This process includes:

[0025] A resource allocation history database is constructed to store resource allocation data of the dynamic collaborative network under different processing scenarios within several historical monitoring periods. A resource prediction model is constructed, and the full data in the resource allocation history database is extracted as training data. The resource prediction model is trained using the training data to obtain the completed resource prediction model.

[0026] Input the current processing scenario information and the dynamic collaborative network into the resource prediction model, and obtain the computing power resource demand curve and industrial site resource demand curve of each type of intelligent agent in the dynamic collaborative network in the current monitoring period based on the resource prediction model.

[0027] Monitor resources in the cloud and industrial sites to obtain current computing power resources and industrial site resources. Based on the computing power resource demand curves and industrial site resource demand curves of various types of intelligent agents in the current monitoring period, dynamically expand and shrink resources in the cloud and industrial sites in advance.

[0028] The system obtains the load rate and priority of each type of intelligent agent in the dynamic collaborative network. Based on the priority and load rate of each type of intelligent agent, the system sorts the intelligent agents in ascending order and constructs a resource allocation queue. Starting from the first position of the resource allocation queue, the system allocates computing resources and industrial field resources to each type of intelligent agent in the resource allocation queue in turn, thereby obtaining the computing resources and industrial field resources of each type of intelligent agent.

[0029] Furthermore, the equipment condition monitoring process includes:

[0030] In the dynamic network, each auxiliary agent inputs the time-series numerical values ​​of industrial monitoring indicators for each process subsequence of the current monitoring period into the equipment status monitoring model. The model compares these values ​​with preset threshold values ​​for each process subsequence, obtaining the cumulative time during which each industrial monitoring indicator is outside its corresponding threshold. Based on the cumulative time and the preset threshold values, the model determines the intermediate inference result for each process subsequence. If the cumulative time of any industrial monitoring indicator in a process subsequence exceeds the preset threshold, the intermediate inference result indicates an abnormality in that indicator. Conversely, if the cumulative time of all industrial monitoring indicators in a process subsequence is less than or equal to the preset threshold, the intermediate inference result indicates normal equipment operation.

[0031] Furthermore, the process of performing fact verification, compliance verification, and feasibility verification on the initial inference results output by each expert branch, and generating the inference result to be merged or performing expert branch re-inference based on the verification results includes:

[0032] For the initial inference result output by the equipment diagnosis expert branch, the abnormal industrial monitoring indicators in the initial inference result are extracted and compared with the abnormal industrial monitoring indicators in the intermediate inference result. If the abnormal industrial monitoring indicators in the initial inference result do not include the abnormal industrial monitoring indicators in the intermediate inference result, the initial inference result is marked as unqualified, and a verification deviation report is generated and sent back to the expert branch for re-inference. If the abnormal industrial monitoring indicators in the initial inference result include the abnormal industrial monitoring indicators in the intermediate inference result, the initial inference result is marked as a result to be fused.

[0033] For the initial inference results output by the process adaptation expert branch and the operation and maintenance optimization expert branch, compliance verification and feasibility verification are performed on the initial inference results. If the compliance verification or feasibility verification of the initial inference results fails, the initial inference results are marked as unqualified, and a verification deviation report is generated and sent back to the expert branch for re-inference. If both the compliance verification and feasibility verification of the initial inference results pass, the initial inference results are marked as inference results to be merged.

[0034] Furthermore, the process of generating a global execution plan includes:

[0035] A fusion temporary database is constructed, and the reasoning results to be fused from the three types of expert branches are standardized, decomposed, and uniformly encoded. The reasoning results to be fused are decomposed into standardized core elements and uniformly stored in the fusion temporary database. Conflict resolution analysis is performed on the standardized core elements of the three types of expert branches in the fusion temporary database. Then, using the triggering events in the standardized core elements as the core anchor points, the core factual conclusions of the standardized core elements of the three types of expert branches are correlated and mapped to construct causal relationship links. Based on the causal relationship links, the execution action instructions, timing requirements, constraints, and resource requirements in the standardized core elements of the three types of expert branches are fused to generate a global execution plan.

[0036] The expected effects of the standardized core elements of the three types of expert branches are integrated to obtain multi-dimensional expected goals. The multi-dimensional expected goals are used as evaluation indicators. The indicator weights and fuzzy rule bases of the evaluation indicators are set according to the reasoning task input by the user. Based on the fuzzy rule base, the membership matrix of the multi-dimensional expected goals with respect to the preset quality level is obtained through fuzzy comprehensive evaluation. The quality level of the multi-dimensional expected goals is obtained according to the membership matrix and indicator weights.

[0037] If the quality level of the multi-dimensional expected target is less than the preset quality level threshold, an expected target deviation report is generated and sent back to the expert branch for re-inference. If the quality level of the multi-dimensional expected target is greater than or equal to the preset quality level threshold, the numerical time series of industrial monitoring indicators of each process subsequence in the current monitoring cycle, processing scenario information, and global execution plan are input into the industrial processing full-process model for simulation to obtain the multi-dimensional simulation target and the corresponding quality level. If the quality level corresponding to the multi-dimensional simulation target is less than the quality level threshold, a simulation deviation report is generated based on the multi-dimensional simulation target and sent back to the expert branch for re-inference. If the quality level corresponding to the multi-dimensional simulation target is greater than or equal to the quality level threshold, the global execution plan is input into the agent for execution.

[0038] Furthermore, the process of conducting conflict resolution analysis on the standardized core elements of the three expert branches includes:

[0039] If the execution action instructions in the standardized core elements of an expert branch do not meet the constraints in the standardized core elements of other expert branches, a rigid conflict report for the expert branch is generated and sent back to the expert branch for re-reasoning.

[0040] If the core fact conclusion A in the standardized core elements of an expert branch is inconsistent with the core fact conclusion B in the standardized core elements of other expert branches, then the specific parameter variables corresponding to the core fact conclusions of the expert branch and the other expert branches are extracted respectively.

[0041] The core factual conclusions of the expert branch and other expert branches are respectively imported into the industrial processing full-process model for simulation, and the numerical time series of industrial monitoring indicators of each process subsequence in the previous monitoring cycle and the processing scenario information are imported into the model for simulation, and the data of the conclusion A verification group and the conclusion B verification group are obtained.

[0042] The data from the validation groups of Conclusion A and Conclusion B are compared with the time-series values ​​of industrial monitoring indicators for each process subsequence in the current monitoring cycle. The average goodness of fit of the data from the validation groups of Conclusion A and Conclusion B is obtained. If the average goodness of fit of the validation group data is greater than or equal to a preset goodness of fit threshold, the expert branch corresponding to the validation group data with the smaller average goodness of fit is extracted from the two validation groups of data. A rigid conflict report of the expert branch is generated and sent back to the expert branch for re-reasoning. If the average goodness of fit of the two validation groups of data is less than the preset goodness of fit threshold, an automated validation failure report is generated, triggering human expert intervention as a backup.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] 1. Addressing the common problems of "data silos," "poor real-time performance," and "insufficient reliability" in existing industrial AI systems, this invention achieves comprehensive data collection, real-time processing, and intelligent analysis of industrial field data through mechanisms such as multi-source data fusion (perception layer), digital twin (simulation layer), and multi-agent collaboration (collaboration layer). For example, the perception layer breaks down the data entry process according to process sub-sequences, solving the fragmentation problem of traditional data collection; the digital twin model in the simulation layer, combined with processing scenario information, realizes the visualization and simulation of industrial processes, providing more intuitive support for decision-making; and the multi-agent collaboration layer improves the system's adaptability to complex scenarios through dynamic network adjustments.

[0045] 2. The system's dynamic feedback mechanism and multi-agent collaborative architecture enable real-time monitoring, intelligent reasoning, and adaptive optimization in industrial settings. For example, equipment status monitoring uses auxiliary agents to compare monitoring data with preset thresholds in real time, promptly identifying anomalies; the dynamic feedback layer verifies and integrates the reasoning results to generate a global execution plan, ensuring the feasibility and effectiveness of decisions. These functions drive the transformation of industry from "experience-driven" to "data-driven," improving production efficiency and product quality, and promoting industrial intelligence and digital transformation.

[0046] 3. The MoE architecture of the vertical domain foundation layer allows for flexible adjustment of the core input features and triggering conditions of expert branches according to the needs of industrial scenarios, improving the system's adaptability to vertical domains; the dynamic network construction mechanism of the multi-agent collaboration layer can select agents according to task requirements to achieve optimized resource allocation; and the dynamic expansion and contraction mechanism of the resource allocation layer ensures stable operation of the system under different loads. These features enable the system to adapt to the needs of different industrial scenarios, exhibiting good flexibility and scalability. Attached Figure Description

[0047] Figure 1 This is a flowchart of a multi-agent collaborative reasoning system for a large vertical domain model based on dynamic feedback, as described in an embodiment of this application.

[0048] Figure 2 This is the overall architectural framework of the vertical domain large model in this application embodiment. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] like Figure 1As shown, the vertical domain large model multi-agent collaborative reasoning system based on dynamic feedback includes a cloud, and the cloud communication connection includes a perception layer, simulation layer, memory layer, vertical domain base layer, multi-agent collaboration layer, resource allocation layer and dynamic feedback layer.

[0051] The perception layer is used to set up on-site data monitoring points based on the process information entered into the current industrial site data, and to collect indicator data at each point.

[0052] The simulation layer is used to construct a full-process industrial processing model based on the index data and processing scenario information at each point.

[0053] The memory layer is used to build knowledge graphs, fault libraries, optimization libraries, and process libraries for industrial sites;

[0054] The knowledge graph includes nodes such as fault type (mechanical fault, electrical fault, control system fault), fault characteristics (vibration data threshold, temperature threshold, abnormal current range), root cause of fault, and handling plan, enabling the visualization and related query of fault knowledge.

[0055] Fault Database: Stores case information for each fault, including fault occurrence time, operating parameters, fault characteristics, reasoning path, handling plan, execution result, and feedback correction parameters, for subsequent rapid fault location and handling;

[0056] Optimization library: Stores operation and maintenance optimization strategies (such as tool replacement cycle optimization and equipment inspection cycle optimization) and collaborative reasoning optimization parameters to improve operation and maintenance efficiency;

[0057] Process Library: Stores the processing specifications of process subsequences, processing parameters of different workpieces, equipment operation manuals and other knowledge, which are used for knowledge support for collaborative reasoning and training of operation and maintenance personnel;

[0058] The vertical domain base layer is used to build a large vertical domain model based on the memory layer. The large vertical domain model includes equipment diagnosis expert branches, process adaptation expert branches, and operation and maintenance optimization expert branches.

[0059] The multi-agent collaboration layer is used to construct various types of agents, break down the reasoning task, assign sub-tasks to corresponding agents, construct a dynamic collaboration network based on the agents paired with each sub-task, monitor the device status, and perform reasoning based on the expert branches paired with each type of agent in the dynamic collaboration network, and output the initial reasoning result.

[0060] The resource allocation layer is used to predict resources for the dynamic collaborative network, dynamically expand and shrink resources based on the resource prediction results, and allocate computing resources and industrial site resources for various types of intelligent agents in the dynamic collaborative network.

[0061] The dynamic feedback layer is used to perform fact verification, compliance verification, and feasibility verification on the initial inference results output by each expert branch. Based on the verification results, it generates inference results to be merged or performs expert branch re-inference. It performs structured decomposition and unified coding on the inference results to be merged, and performs conflict resolution analysis, core fact layer fusion, execution action layer fusion, effect layer fusion, and effect evaluation on the structured decomposition and unified coding results to generate a global execution plan.

[0062] It should be further explained that, in the specific implementation process, the sensing layer sets up on-site data monitoring points based on the current industrial site data entry process information, and the process of collecting indicator data at each point includes:

[0063] Obtain the process information of the current industrial site data entry, extract the characteristics of the process unit (the inherent attributes, functional attributes, constraint attributes and data attributes of the smallest process execution unit that cannot be further divided in industrial production) based on the process information, and divide the industrial site data entry process into several process sub-sequences according to the characteristics of the process unit.

[0064] On-site data monitoring points are set up in each process subsequence, and industrial monitoring indicators of each on-site data monitoring point are obtained by data retrieval based on the functional attributes in the process unit characteristics of the corresponding process subsequence.

[0065] The on-site data monitoring points acquire indicator data in real time based on industrial monitoring indicators, mark the monitoring time, and set the monitoring cycle.

[0066] It should be further explained that, in the specific implementation process, the simulation layer constructs a full-process industrial processing model based on the indicator data and processing scenario information at each point, including:

[0067] The process involves acquiring the physical entities of each process subsequence in the current industrial field data entry process, constructing a digital space, creating 3D models of the physical entities of each process subsequence, mapping the 3D models to the digital space, setting up API interfaces on the 3D models of each process subsequence, preprocessing the data format of the indicator data collected from the field data monitoring points of the corresponding process subsequence, generating twin data based on the preprocessed indicator data, and matching the twin data with the 3D models of the corresponding process subsequences in the digital space to generate digital twin models.

[0068] Acquire different processing scenario information for each process subsequence. The processing scenario information includes: process parameter fluctuation values, load rate, ambient temperature 20-25℃, ambient pressure 0.1-0.12MPa, humidity 40%-60%, energy consumption, product model, batch quantity, etc. Store the different processing scenario information in the digital space, and combine the 3D model of each process subsequence in the digital twin model with the processing scenario information of each process subsequence to generate a full-process industrial processing model.

[0069] It should be further explained that, in the specific implementation process, the generation of the industrial processing full-process model is based on the recent stable production line full-process operation data (production index parameters, equipment operation parameters, environmental parameters, product quality data, etc. of each operation data monitoring node). The industrial processing full-process model is subjected to static + dynamic dual calibration to ensure that the deviation of the model's parameter response, quality transfer, equipment linkage and the physical production line is ≤±2%, which meets the accuracy requirements of process testing.

[0070] Static calibration: Match the 3D model parameters of the core equipment in each process subsequence (such as metering pump speed-flow rate, stretching roller speed ratio-stretching ratio, cutting machine cutter spacing-length accuracy) to the physical equipment;

[0071] Dynamic calibration: Simulate the normal production state of the physical production line to verify that the linkage response of the whole process parameters and the transmission law of intermediate quality indicators in the model are consistent with the physical production line. For example, after adjusting the spinning temperature, the adaptation changes of the crystallinity of the nascent filament and the cooling wind speed in the model need to match the physical production line.

[0072] It should be further explained that, in the specific implementation process, the process of building a large-scale vertical domain model in the vertical domain base layer includes:

[0073] Based on the memory layer, a fault dataset, a process dataset, and an operation and maintenance dataset (containing 10,000+ fault cases, 5,000+ process parameters, and 3,000+ operation and maintenance records) are constructed. These datasets are used as training samples. The training samples undergo preprocessing, data augmentation, standardization, and multimodal fusion to obtain input samples. A large-scale vertical domain model is constructed based on the MoE sparse architecture. The equipment diagnosis expert branch is responsible for fault feature extraction, fault type identification, and root cause analysis. The process adaptation expert branch is responsible for determining the impact of faults on processing quality in conjunction with the processing technology. The operation and maintenance optimization expert branch is responsible for generating fault handling solutions and operation and maintenance optimization strategies. Core input features and triggering features for each type of expert branch in the large-scale vertical domain model are defined. The input samples are compared with the core input features and triggering features of each type of expert branch to obtain the corresponding expert branch. Then, a shared encoder pre-training and independent expert branch training are performed until the loss function is stable. The model parameters are saved, and the completed large-scale vertical domain model is output.

[0074] Building a large-scale model for a vertical domain is a complex process involving multiple steps such as model selection, training, validation, and testing. The following is a detailed supplementary explanation of this process:

[0075] Based on the fault dataset (10,000+ fault cases), the process dataset (5,000+ process parameters), and the operation and maintenance dataset (3,000+ operation and maintenance records), a unified industrial structured dataset is constructed, incorporating "process parameter fluctuation values" as a key feature into each dataset, as shown in Table 1 below:

[0076] Table 1

[0077]

[0078] Data augmentation: For fluctuations in process parameters, the "fluctuation range expansion" method is used to generate incremental samples (e.g., for a fault case with an 8% temperature fluctuation, the simulated sample range of 7%-9% is expanded) to solve the problem of insufficient data in small-sample fault scenarios.

[0079] Standardization: Normalize all process parameter fluctuations to the [0,1] interval and unify the input feature scale;

[0080] Multimodal fusion: It integrates process parameter fluctuation values ​​(numerical type) with fault text descriptions (text type) and equipment vibration curves (time series type) into a unified input vector.

[0081] The large-scale vertical domain model adopts a MoE sparse architecture of "shared encoder + expert branch decoder + gated routing network". It dynamically activates the corresponding expert branch based on input features (especially process parameter fluctuations). Its overall architecture framework is as follows: Figure 2 As shown.

[0082] The specific designs of each type of expert branch are shown in Table 2 below:

[0083] Table 2

[0084]

[0085] Gated networks use "whether the fluctuation value of process parameters exceeds the limit" as the trigger feature to dynamically allocate input samples to corresponding expert branches, for example:

[0086] When the fluctuation value of process parameters in the input sample is greater than the steady-state threshold: the equipment diagnosis expert branch is activated first (activation weight 0.7), and the process adaptation branch (0.2) and operation and maintenance optimization branch (0.1) are activated as auxiliary branches.

[0087] When the input sample is a process parameter fluctuation value, process type, or quality standard: the process adaptation expert branch (0.8) is activated first, and the equipment diagnosis branch (0.1) and operation and maintenance optimization branch (0.1) are activated secondarily.

[0088] When the input sample is a process parameter fluctuation value, operation and maintenance record, or equipment status: the operation and maintenance optimization expert branch (0.8) is activated first, and the equipment diagnosis branch (0.1) and process adaptation branch (0.1) are activated as auxiliary branches.

[0089] Then, the shared encoder pre-training and expert branch independent training processes are performed:

[0090] Phase 1: Shared encoder pre-training:

[0091] Training objective: To enable the shared encoder to learn common features of industrial scenarios (especially the semantics of process parameter fluctuations).

[0092] Input sample source: Unlabeled samples from the entire dataset (accounting for 60%);

[0093] Training tasks:

[0094] Mask Language Modeling (MLM): Randomizes process parameter fluctuations in the input mask and allows the model to predict the mask value;

[0095] Fluctuation value classification: Predict whether the fluctuation value of process parameters exceeds the preset steady-state threshold (binary classification task);

[0096] Training parameters: Batch size = 64, learning rate = 1e-4, training epochs = 10, optimizer = AdamW.

[0097] Phase 2: Independent training of expert branches (core: task adaptation to process parameter fluctuation values):

[0098] 1. Equipment Diagnostic Expert Branch Training

[0099] Input sample source: Fault dataset (10000+ cases), divided into training / validation / test sets in an 8:1:1 ratio;

[0100] Training tasks:

[0101] Fault type classification (multiple classifications: such as mechanical faults / electrical faults / control system faults);

[0102] Root cause analysis (generative task: input process parameter fluctuation values ​​+ fault characteristics, output root cause text).

[0103] Key constraint: Add "volatility correlation weight" to the loss function (the higher the correlation, the greater the sample weight);

[0104] Training parameters: BatchSize=32, learning rate=5e-5, Epoch=15, loss function=cross-entropy loss (classification) + cosine similarity loss (generation).

[0105] 2. Process Adaptation Expert Branch Training

[0106] Input sample sources: quality-related samples from the process dataset and the fault dataset;

[0107] Training tasks:

[0108] Volatility-Quality Correlation Prediction (Regression Task: Input volatility value, output quality deviation value);

[0109] Process parameter adjustment suggestions (Generative task: input fluctuation exceeds limit, output parameter adjustment scheme);

[0110] Training parameters: BatchSize=32, learning rate=5e-5, Epoch=12, loss function=MSE loss (regression) + perplexity loss (generation).

[0111] 3. Operations and Maintenance Optimization Expert Branch Training

[0112] Input sample sources: sample handling solutions from the operation and maintenance dataset and the fault dataset;

[0113] Training tasks:

[0114] Maintenance type prediction (multiple categories: inspection / repair / parts replacement);

[0115] Operation and maintenance plan generation (generative task: input fluctuation limit value + device status, output operation and maintenance steps);

[0116] Key constraint: The generated solution must include a "fluctuation value recovery target" (e.g., temperature fluctuation ≤3% after maintenance).

[0117] Training parameters: BatchSize=32, learning rate=5e-5, Epoch=12, loss function=cross-entropy loss (classification) + BLEU score loss (generation).

[0118] Phase 3: Gated Network and Branch Fusion Training

[0119] Training objective: To optimize the branch activation strategy of the gating network so that fluctuations in process parameters accurately drive the corresponding branches;

[0120] Training data: Mixed dataset (faults + processes + operations and maintenance, in a 1:1:1 ratio).

[0121] Training strategy:

[0122] By fixing the expert branch parameters and training only the gating network, the gating network can accurately allocate activation weights based on the fluctuation values ​​of the input process parameters.

[0123] Unfreeze all parameters and perform joint fine-tuning (reducing the learning rate to 1e-5) to optimize inter-branch collaboration;

[0124] Loss function: Combines the task loss of each branch with the gated routing loss (penalizing erroneous activations, such as samples with excessive fluctuations not activating the device diagnosis branch).

[0125] It should be further explained that, in the specific implementation process, the core value of the vertical domain large model (based on the MoE sparse architecture, containing three major expert branches) is to provide professional industrial reasoning capabilities without directly connecting to industrial field tasks. The core value of the various intelligent agents (equipment diagnostic intelligent agents, process reasoning intelligent agents, etc.) built is to transform the reasoning results of the large model into implementable industrial operations, connect to sub-tasks, coordinate interactions, and execute specific actions, serving as a "bridge" between the large model and the industrial field.

[0126] The preset agent types and core capabilities are as follows:

[0127] Data processing intelligent agent: industrial multi-source data acquisition, preprocessing (denoising, normalization, outlier correction), data format conversion, data transmission and archiving;

[0128] The intelligent device for equipment diagnosis monitors equipment status, extracts fault features, locates faults, analyzes root causes, and provides early warnings of faults.

[0129] Operation and maintenance execution intelligent agent: generating fault handling solutions, formulating operation and maintenance plans, executing operation and maintenance processes, and verifying operation and maintenance results;

[0130] Process reasoning agent: monitoring process parameters, analyzing parameter fluctuations, optimizing process parameters, and adjusting process schemes;

[0131] Intelligent quality inspection agent: real-time product quality detection, quality anomaly analysis, and correlation analysis between quality and process parameters;

[0132] Energy consumption optimization intelligent agent: real-time energy consumption monitoring, energy consumption analysis, energy consumption allocation optimization, and energy-saving scheme generation;

[0133] Task allocation agent: subtask scheduling, agent collaboration and coordination, conflict resolution, and collaboration effect monitoring;

[0134] Knowledge retrieval agent: industrial knowledge base retrieval, knowledge matching, knowledge updating, and knowledge push (providing knowledge support for other agents).

[0135] The specific correspondences between intelligent agents and the three expert branches of large models are as follows:

[0136] Equipment Diagnosis Expert Branch → Equipment Diagnosis Agent: The expert branch is responsible for fault feature extraction, fault type judgment, and root cause analysis (core reasoning). The equipment diagnosis agent takes over the reasoning results and performs field tasks such as equipment status monitoring, fault location feedback, and anomaly warning. At the same time, it feeds back field data (such as equipment operating parameters and actual fault conditions) to the branch to help the branch optimize the reasoning accuracy.

[0137] Process Adaptation Expert Branch → Process Reasoning Agent + Quality Inspection Agent: The expert branch is responsible for judging the impact of faults on processing quality and optimizing process parameters (core reasoning) based on the processing technology. The process reasoning agent takes over the parameter optimization results and performs process parameter adjustments and fluctuation monitoring. The quality inspection agent takes over the quality impact analysis results and performs quality inspection and anomaly feedback. Together, they implement the reasoning of the process adaptation branch and simultaneously transmit back quality data and process parameter fluctuation data to support branch iteration.

[0138] Operation and Maintenance Optimization Expert Branch → Operation and Maintenance Execution Agent + Energy Consumption Optimization Agent: The expert branch is responsible for generating fault handling plans and operation and maintenance optimization strategies (core reasoning). The operation and maintenance execution agent takes over the handling plans and operation and maintenance plans, and executes fault handling, inspection and maintenance. The energy consumption optimization agent takes over the energy consumption optimization reasoning results, and executes energy consumption monitoring and energy-saving adjustments. The two simultaneously provide the expert branch with operation and maintenance effects and energy consumption data to assist in strategy optimization.

[0139] It should be further explained that, in the specific implementation process, the reasoning task is broken down, subtasks are assigned to corresponding agents, and a dynamic cooperative network is constructed based on the paired agents for each subtask. This process includes:

[0140] The inference task (such as equipment lifecycle maintenance and global production process optimization) input by the user in the current industrial setting is obtained. Based on the core objectives of the inference task (the ultimate demands of the user input inference task, such as the core objective of equipment lifecycle maintenance being "reducing equipment failure rate by ≥30%, extending equipment lifespan by ≥15%, and reducing maintenance costs by ≥20%"; and the core objective of global production process optimization being "improving product qualification rate by ≥5%, reducing energy consumption by ≥10%, and optimizing production cycle by ≥8%"), industrial scenario constraints (safety constraints, such as equipment maintenance must not affect production safety and process adjustments must not exceed safety thresholds), compliance constraints, such as environmental indicators and quality standards, equipment constraints, such as equipment runtime sequence and access restrictions, and resource constraints, such as computing power and the number of industrial tool interfaces) and time benchmarks (total execution cycle, such as a 1-year equipment lifecycle maintenance cycle and a 3-month global production process optimization cycle), the inference task is broken down into several sub-tasks. The objectives, constraints, priorities, and time requirements of each sub-task are determined, as shown in Table 3 below.

[0141] Table 3

[0142]

[0143] Based on the objectives, constraints, priorities, and time requirements of each subtask, different types of intelligent agents are selected to obtain the core and auxiliary intelligent agents corresponding to each subtask, specifically:

[0144] Step s1: Extract the core requirements of the subtask (e.g., the core requirements of the subtask "Real-time Device Status Monitoring" are "Data Acquisition, Anomaly Identification, and Real-time Feedback").

[0145] Step s2: Compare the capability boundaries of the intelligent agents and select the core intelligent agents (e.g., "Real-time Device Status Monitoring" matches "Device Diagnosis Intelligent Agent" as the core intelligent agent).

[0146] Step s3: Determine whether an auxiliary intelligent agent is needed (e.g., "Real-time Device Status Monitoring" requires the collection of multi-source data, and a "Data Processing Intelligent Agent" is matched as an auxiliary intelligent agent to be responsible for data preprocessing and transmission).

[0147] Step s4: Verify the rationality of the matching (check whether the agent's capabilities meet the sub-task constraints, such as data acquisition delay and accuracy requirements). If not, adjust the agent matching scheme (such as replacing the agent with a high-computing-power agent or adding an auxiliary agent).

[0148] A dynamic collaborative network is constructed based on the core intelligent agents and auxiliary intelligent agents of each sub-task, as well as the task allocation intelligent agent;

[0149] In the dynamic network, each core agent inputs the industrial monitoring indicators of each process subsequence into the expert branch corresponding to each core agent. Based on the initial inference results output by the expert branch, the initial inference results output by the equipment diagnosis expert branch include: fault feature extraction results, fault type determination, root cause localization conclusion, fault impact range prediction, and fault level assessment.

[0150] The initial inference results output by the process adaptation expert branch include: assessment of the impact of faults / operating condition changes on processing quality, optimization / adjustment schemes for process parameters, process adaptation suggestions, and prediction of quality deviations;

[0151] The initial inference results output by the Operation and Maintenance Optimization Expert branch include: fault handling solutions, operation and maintenance / maintenance plan optimization, equipment lifespan adaptation strategies, and energy consumption optimization solutions;

[0152] The dynamic collaborative network adopts a "layered collaboration + dynamic networking" architecture, consisting of three layers:

[0153] Scheduling layer: led by task allocation agents, responsible for sub-task scheduling, agent coordination, conflict resolution, and monitoring of coordination effects, serving as the core hub of the collaborative network;

[0154] Execution layer: Composed of core and auxiliary intelligent agents for each subtask, responsible for the specific execution of the subtask, data feedback, and result output;

[0155] Support layer: Composed of knowledge retrieval agents and data processing agents, it provides data and knowledge support to the execution layer agents to ensure the smooth execution of subtasks.

[0156] In the collaborative network, the agents interact in real time through a pre-defined interaction protocol (compliant with industrial communication standards). The interaction logic is divided into three categories: "data interaction, command interaction, and result interaction," as detailed below:

[0157] Data interaction: Auxiliary agents (such as data processing agents and knowledge retrieval agents) push the required data and knowledge to the core agent, and the core agent feeds back to the auxiliary agents to adjust the data and knowledge requirements; for example, the equipment diagnosis agent sends a "device vibration data acquisition request" to the data processing agent, the data processing agent preprocesses the data and pushes the vibration data, and if the data quality is not up to standard, the equipment diagnosis agent feeds back a "data re-acquisition request".

[0158] Command interaction: The task allocation agent sends subtask execution instructions and scheduling instructions to each execution agent, and the execution agents report the execution status (such as "in execution", "completed", "abnormally interrupted") to the task allocation agent; for example, the task allocation agent sends a "fault handling instruction" to the operation and maintenance execution agent, and the operation and maintenance execution agent reports the handling progress and results.

[0159] Results interaction: The core intelligent agent pushes the execution results of subtasks to related intelligent agents to support the execution of related subtasks; for example, the device diagnosis intelligent agent pushes the "root cause result of the fault" to the operation and maintenance execution intelligent agent to support the operation and maintenance execution intelligent agent in generating a handling plan; the operation and maintenance execution intelligent agent pushes the "fault handling effect" to the device diagnosis intelligent agent to support the device diagnosis intelligent agent in verifying the accuracy of the root cause analysis.

[0160] Interaction constraints: Interaction delay ≤ 1s (for subtasks with high real-time requirements, such as fault location, interaction delay ≤ 500ms); interaction data adopts a standardized format to ensure that each agent can recognize it; the entire interaction process is recorded for collaborative effect analysis and fault tracing.

[0161] It should be further explained that, in the specific implementation process, the resource allocation layer performs resource prediction on the dynamic collaborative network, and dynamically expands and shrinks resources according to the resource prediction results. The process of allocating computing resources and industrial site resources for various types of intelligent agents in the dynamic collaborative network includes:

[0162] A resource allocation history database is constructed to store resource allocation data of dynamic collaborative networks under different processing scenarios within several historical monitoring periods. A resource prediction model is constructed using the LightGBM regression model. The full data in the resource allocation history database is extracted as training data. The resource prediction model is trained using the training data to obtain the completed resource prediction model.

[0163] Input the current processing scenario information and the dynamic collaborative network into the resource prediction model, and obtain the computing power resource demand curve and industrial site resource demand curve of each type of intelligent agent in the dynamic collaborative network in the current monitoring period based on the resource prediction model.

[0164] In the dynamic collaborative network, task allocation agents monitor resources in the cloud and industrial sites, acquiring current computing resources (including edge CPU / GPU load, cloud computing node occupancy, computing response latency), token resources (remaining industrial knowledge base call tokens, token consumption rate), storage resources (industrial data storage capacity, storage read / write speed, archived data occupancy) and industrial site resources (including industrial equipment computing power (equipment's built-in processor load, data processing capability), and industrial tool availability status (equipment control interfaces, quality inspection interfaces, online / offline status of processing tools, occupancy status)). Based on the computing resource demand curves of each type of agent in the current monitoring period and the industrial site resource demand curves (the resource demand curves are plotted with time within the current monitoring period on the horizontal axis and the resource demand values ​​of each type of agent on the vertical axis), dynamic expansion and contraction of resources in the cloud and industrial sites are performed in advance. For example:

[0165] Expanding computing resources: If it is predicted that the GPU load demand of the process inference agent will reach 80% during a certain period of the demand curve (such as 10:00-10:30), while the current GPU load is only 60%, the resource scheduling module will automatically send an expansion request to the cloud to add GPU computing nodes, increase the GPU load to 85% in advance, and reserve 5% redundancy to avoid insufficient resources during peak periods.

[0166] Expanding computing power for industrial equipment: If it is predicted that the computing power demand of the equipment diagnostic agent will reach 70% at a certain time, while the current computing power load of the equipment is 55%, the equipment operating parameters can be adjusted in advance to release some of the computing power of non-core tasks and allocate it to the equipment diagnostic agent first.

[0167] Industrial tool interface expansion: If it is predicted that the call frequency of the quality inspection interface will reach 28 times / minute (close to the upper limit of 30 times / minute) during a certain period of the demand curve, check the idle quality inspection interfaces in advance, start the backup interfaces, and ensure that the number of available interfaces meets the demand; at the same time, adjust the interface call priority to give priority to the interface usage needs of high-priority intelligent agents.

[0168] Obtain the load rate (load rate = actual resource usage / rated capacity × 100%) and priority of each type of agent in the dynamic collaborative network. Sort the agents in ascending order based on their priority and load rate, and construct a resource allocation queue (the higher the priority and the greater the load rate, the higher the ranking of the agent in the resource allocation queue). Starting from the first position in the resource allocation queue, allocate computing resources to each type of agent in the queue sequentially. (For agents requiring complex inference (such as equipment diagnostic agents and process inference agents), prioritize allocating GPU computing nodes to improve inference efficiency; for agents requiring only basic data processing (such as data processing agents), allocate CPU computing nodes to avoid wasting high computing resources; for subtasks with high real-time requirements (such as fault location), allocate computing resources to the edge to reduce inference latency.) Global optimization subtasks (such as operation and maintenance plan optimization, allocating cloud computing power, and using high computing power to complete complex calculations) and industrial site resource allocation (allocating corresponding industrial tool interfaces and data acquisition resources according to the task requirements of intelligent agents. For example, when a quality inspection intelligent agent performs a real-time inspection task, the quality inspection equipment interface is allocated first to ensure real-time acquisition of inspection data; when an operation and maintenance execution intelligent agent performs fault handling, it is given priority to obtain the right to use the equipment control interface to ensure that the handling action is implemented in a timely manner. For intelligent agents that frequently call the industrial knowledge base (such as operation and maintenance execution intelligent agents and knowledge call intelligent agents), sufficient token resources are allocated to avoid token exhaustion affecting task execution; for intelligent agents that need to store a large amount of data (such as data processing intelligent agents), sufficient storage resources are allocated to ensure smooth data archiving and reading and writing), and computing power resources and industrial site resources of various types of intelligent agents are acquired.

[0169] It should be further explained that, in the specific implementation process, the equipment status monitoring process includes:

[0170] In the dynamic network, each auxiliary agent inputs the time-series numerical values ​​of industrial monitoring indicators for each process subsequence of the current monitoring period into the equipment condition monitoring model. The equipment condition monitoring model compares the time-series numerical values ​​of industrial monitoring indicators for each process subsequence with the preset industrial monitoring indicator thresholds for each process subsequence, obtains the cumulative time during which the industrial monitoring indicators for each process subsequence are not within the corresponding preset industrial monitoring indicator thresholds, and determines the intermediate inference result of each process subsequence based on the cumulative time of the industrial monitoring indicators for each process subsequence and the preset time thresholds. If the cumulative time of a certain industrial monitoring indicator in a process subsequence is greater than the preset time threshold, the intermediate inference result of the process subsequence is that the industrial monitoring indicator is abnormal. If the cumulative time of all industrial monitoring indicators in a process subsequence is less than or equal to the preset time threshold, the intermediate inference result of the process subsequence is that the equipment is normal.

[0171] It should be further explained that, in the specific implementation process, the initial inference results output by each expert branch are subject to fact verification, compliance verification, and feasibility verification. The process of generating the inference result to be merged or performing expert branch re-inference based on the verification results includes:

[0172] For the initial inference result output by the equipment diagnosis expert branch, the abnormal industrial monitoring indicators in the initial inference result are extracted and compared with the abnormal industrial monitoring indicators in the intermediate inference result. If the abnormal industrial monitoring indicators in the initial inference result do not include the abnormal industrial monitoring indicators in the intermediate inference result, the initial inference result is marked as unqualified, and a verification deviation report is generated based on the comparison result and sent back to the expert branch for re-inference. If the abnormal industrial monitoring indicators in the initial inference result include the abnormal industrial monitoring indicators in the intermediate inference result, the initial inference result is marked as a result to be fused.

[0173] For the initial inference results output by the process adaptation expert branch and the operation and maintenance optimization expert branch, compliance verification and feasibility verification are performed on the initial inference results, including:

[0174] Pre-set compliance verification benchmark library (stored in the memory layer): national / industry mandatory standards, enterprise safety production system, process SOP, quality control system, equipment manufacturer rated parameters and maintenance specifications, environmental emission limits, special equipment safety regulations, etc.; Feasibility verification benchmark library (dynamically updated, from the perception layer + dynamic feedback layer): real-time equipment status data, production scheduling data, material / spare parts / tooling inventory data, personnel qualification and shift data, historical execution effect data, on-site resource availability status data, etc.

[0175] For the process parameter adjustment schemes (adjustment values, timing, and step size of parameters such as temperature, pressure, speed, and feed rate) in the initial inference results output by the process adaptation expert branch, and the equipment fault handling schemes (shutdown / non-shutdown handling steps, component replacement suggestions, and parameter calibration schemes) and equipment energy consumption optimization schemes (operating power adjustment, start-up and shutdown timing optimization, and energy distribution schemes) in the initial inference results output by the operation and maintenance optimization expert branch:

[0176] First, perform the highest priority safety compliance verification (e.g., whether the fault handling / maintenance plan complies with the safety production standards (e.g., prohibition of live maintenance, live operation, and unprotected high-altitude work), whether the shutdown maintenance complies with the safe shutdown and power-off tagging procedures, etc.). If the safety compliance verification fails, it is directly rejected, a "Non-compliance Reason Report" is generated, and it is sent back to the corresponding expert branch to trigger re-reasoning.

[0177] If safety and compliance are passed, the remaining compliance verification dimensions are executed (e.g., whether the operation and maintenance plan and component replacement comply with the equipment manufacturer's maintenance specifications and model matching requirements; whether the maintenance cycle adjustment is shorter than the manufacturer's mandatory maintenance cycle). If non-compliance is found, optimization suggestions are generated and sent back to the expert branch for re-reasoning.

[0178] After the full compliance verification is passed, a feasibility verification is performed (such as whether the process parameter adjustment is within the current actual capacity of the equipment (not the rated parameters, and the aging, wear and failure status of the equipment need to be considered); whether the equipment adjustment system can achieve the adjustment rate and control accuracy required by the plan). If it is not feasible, a "Reasons for Infeasibility and Optimization Directions" is generated and sent back to the expert branch for re-reasoning.

[0179] After both verifications pass, a "Verification Pass Report" is generated. The initial inference result is marked as the inference result to be fused, and the final inference result is sent to the corresponding intelligent agent (process inference intelligent agent / operation and maintenance execution intelligent agent) for execution. At the same time, the entire verification process data is stored in the knowledge accumulation layer.

[0180] It should be further explained that, in the specific implementation process, the process of structurally decomposing and uniformly encoding the results of the fusion reasoning, and then performing conflict resolution analysis, core fact layer fusion, execution action layer fusion, effect layer fusion, and effect evaluation on the structured decomposition and uniform encoding results to generate a global execution plan includes:

[0181] A temporary fusion database is constructed, and the inference results to be fused from the three expert branches are standardized, decomposed, and uniformly coded. The inference results to be fused are decomposed into standardized core elements and uniformly stored in the temporary fusion database. The standardized core elements are shown in Table 4 below:

[0182] Table 4

[0183]

[0184] Conflict resolution analysis was conducted on the standardized core elements of the three expert branches. Using the triggering events within these standardized core elements as the core anchor points, a correlation mapping was performed on the core factual conclusions of the standardized core elements across the three expert branches. This constructed a causal link of "equipment diagnosis root cause → process adaptation impact → operation and maintenance optimization and handling," specifically:

[0185] Using the fault / equipment status determination of the equipment diagnosis expert branch as the core benchmark, anchoring the underlying facts: if the equipment diagnosis branch determines "spindle bearing wear, radial runout exceeds tolerance by 0.02mm", then all conclusions of the process adaptation and operation and maintenance optimization branches must revolve around this core fact.

[0186] The process adaptation expert branch supplements the impact of this fact on the entire production process: such as "excessive spindle radial runout will cause the workpiece cylindricity to exceed the tolerance by 0.03mm, and the corresponding batch product defect rate is expected to increase by 12%", thus perfecting the impact chain of the fact.

[0187] The Operations and Maintenance Optimization Experts branch supplemented the closed loop of the implementation of this fact: such as "the machine is shut down to replace the bearing, the estimated downtime is 2 hours, and the corresponding scheduling window is 22:00 on the same day to 0:00 the next day", thus completing the logical closed loop of fact-impact-action.

[0188] Subsequently, based on the causal relationship chain, the execution action instructions, timing requirements, constraints, and resource requirements in the standardized core elements of the three types of expert branches are integrated to generate a global execution plan, including:

[0189] Action Classification and Deduplication: The action instructions of the three branches are unified into four categories, and duplicate and redundant actions are eliminated.

[0190] Preparatory actions (such as spare parts retrieval, personnel scheduling, and scheduling window confirmation).

[0191] Real-time control actions (such as temporary adjustments to process parameters, and equipment de-load operation).

[0192] Core handling actions (such as shutdown for maintenance, component replacement, and process parameter calibration)

[0193] Post-verification actions (such as equipment accuracy calibration, trial production, and quality inspection)

[0194] Timing alignment: Following the logic of industrial production, actions are sequentially ordered, clearly defining the executing entity, time node, and triggering conditions for each action, forming a time-sequential action execution list. Example: Timing fusion of actions related to bearing wear events:

[0195] Constraint fusion: Integrate all constraints from the three branches, and take the most stringent rigid boundary as the constraint line for the execution action, such as safety thresholds, compliance requirements, equipment rated parameters, and scheduling window limits, to ensure that all actions comply with the constraint requirements.

[0196] Resource demand integration: Integrate the resource requirements of all actions and submit them to the industrial resource scheduling module in a unified manner to complete the allocation and reservation of resources such as computing power, interfaces, spare parts, and personnel in advance, so as to avoid resource gaps during execution.

[0197] The expected effects of the standardized core elements of the three expert branches are integrated to obtain multi-dimensional expected goals, including: equipment dimension: failure elimination rate, equipment life extension ratio; quality dimension: product qualification rate improvement ratio, quality deviation control effect; efficiency dimension: production downtime, production cycle time impact; cost dimension: operation and maintenance cost, material loss, capacity loss; energy consumption dimension: energy consumption optimization ratio, energy consumption control. These multi-dimensional expected goals are used as evaluation indicators. Based on the user-input reasoning task, indicator weights and a fuzzy rule base are set. Based on the fuzzy rule base, fuzzy comprehensive evaluation is used to obtain the membership matrix of the multi-dimensional expected goals for a preset quality level. The quality level of the multi-dimensional expected goals is then obtained based on the membership matrix and indicator weights.

[0198] The fuzzy rule base is defined as follows:

[0199] Let the quality levels be (Excellent, Good, Average, Poor), and the set of evaluation indicators be: equipment, quality, efficiency, cost, energy consumption, with sub-indicators (used as a premise for the rules):

[0200] Equipment: Fault elimination rate, equipment life extension rate;

[0201] Quality: Product qualification rate improvement rate, quality deviation control effectiveness;

[0202] Efficiency: Impact of production downtime and production cycle time;

[0203] Costs: Operation and maintenance costs, material losses, and production capacity losses;

[0204] Energy consumption: energy consumption optimization ratio, energy consumption control;

[0205] The fuzzy quantization values ​​and industrial thresholds for each indicator are set as shown in Table 5 below:

[0206] Table 5

[0207]

[0208] The process of obtaining the quality level of multi-dimensional expected targets based on the membership matrix and indicator weights includes:

[0209] The evaluation index weights and membership matrix of the evaluation index are fused by formula to obtain the fuzzy comprehensive evaluation matrix of the evaluation index. The membership degree of the multi-dimensional expected target to different quality levels is obtained according to the fuzzy comprehensive evaluation matrix. The quality level with the highest membership degree corresponding to the multi-dimensional expected target is selected and the quality level with the highest membership degree corresponding to the multi-dimensional expected target is taken as the quality level of the multi-dimensional expected target.

[0210] The formula is as follows:

[0211] ;

[0212] in, The fuzzy comprehensive evaluation matrix of the evaluation indicators. To evaluate the indicator weights, For the membership matrix, "This indicates that the elements at corresponding positions in the weight matrix and membership matrix of the evaluation index are multiplied together." The weighting parameter is used to balance the weight matrix and membership matrix in the fuzzy comprehensive evaluation matrix used to control the evaluation index.

[0213] If the quality level of the multi-dimensional expected target is less than the preset quality level threshold, an expected target deviation report is generated and sent back to the expert branch for re-inference. If the quality level of the multi-dimensional expected target is greater than or equal to the preset quality level threshold, the numerical time series of industrial monitoring indicators of each process subsequence in the current monitoring cycle, processing scenario information, and global execution plan are input into the industrial processing full-process model for simulation to obtain the multi-dimensional simulation target and the corresponding quality level. If the quality level corresponding to the multi-dimensional simulation target is less than the quality level threshold, a simulation deviation report is generated based on the multi-dimensional simulation target and sent back to the expert branch for re-inference. If the quality level corresponding to the multi-dimensional simulation target is greater than or equal to the quality level threshold, the global execution plan is input into the agent for execution.

[0214] It should be further explained that, in the specific implementation process, the conflict resolution analysis of the standardized core elements of the three expert branches includes:

[0215] If the execution action instructions in the standardized core elements of an expert branch do not meet the constraints in the standardized core elements of other expert branches, a rigid conflict report for the expert branch is generated and sent back to the expert branch for re-reasoning.

[0216] If the core fact conclusion A in the standardized core elements of an expert branch is inconsistent with the core fact conclusion B in the standardized core elements of other expert branches, then the specific parameter variables corresponding to the core fact conclusions of the expert branch and the other expert branches are extracted respectively. For example, if the core fact conclusion A for the expert branch is spindle bearing wear, and the core fact conclusion B for the other expert branches is a conflict of feed rate parameter error, then:

[0217] The specific parameter variables corresponding to core fact conclusion A are: bearing failure characteristic frequencies (inner ring / outer ring / rolling element characteristic peaks) in the vibration spectrum, radial runout time-domain variation trend, spindle temperature rise slope, current fluctuation characteristics, and historical similar fault characteristic matching degree.

[0218] The specific parameter variables corresponding to core fact conclusion B are: the deviation between feed rate and process SOP, the temporal correspondence between feed rate adjustment and quality deviation, the quality change pattern under different feed rates, and the matching degree of characteristics of similar historical process problems.

[0219] The core factual conclusions of the expert branch and other expert branches, their corresponding exclusive parameter variables, the numerical time series of industrial monitoring indicators for each process subsequence in the previous monitoring cycle, and the processing scenario information are imported into the industrial processing full-process model for simulation. Data from the conclusion A verification group and conclusion B verification group are obtained, specifically:

[0220] Import the time sequence of industrial monitoring indicators and processing scenario information of each process subsequence in the previous monitoring cycle into the industrial processing full-process model to restore the initial state of equipment, process parameters and production conditions before the anomaly occurred, and ensure that the initial environment of the digital space is 100% aligned with the physical site.

[0221] The simulation results using the controlled variable method are shown in Table 6 below:

[0222] Table 6

[0223]

[0224] The data from validation groups A and B are compared with the time-series values ​​of industrial monitoring indicators for each process subsequence in the current monitoring period to obtain the average goodness of fit between the data from validation groups A and B. ;in, For the goodness of fit, The simulation output value corresponding to the data in either the validation group of conclusion A or the validation group of conclusion B. The actual collected values ​​are the time series of industrial monitoring indicators for the current monitoring period. If the average fit of the validation group data is greater than or equal to the preset fit threshold (0.9), the expert branch corresponding to the validation group data with the smaller average fit is extracted from the two validation group data, a rigid conflict report of the expert branch is generated, and the expert branch is sent back for re-inference.

[0225] If the average fit of the two validation groups is less than the preset fit threshold, and a unique conclusion cannot be drawn through automated validation, an automated validation failure report will be generated, triggering human expert intervention as a backup.

[0226] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A multi-agent collaborative reasoning system for large-scale vertical domain models based on dynamic feedback, characterized in that: Including the cloud, the cloud communication connection includes a perception layer, simulation layer, memory layer, vertical domain base layer, multi-agent collaboration layer, resource allocation layer and dynamic feedback layer; The perception layer is used to set up on-site data monitoring points based on the process information entered into the current industrial site data, and to collect indicator data at each point. The simulation layer is used to construct a full-process industrial processing model based on the index data and processing scenario information at each point. The memory layer is used to build knowledge graphs, fault libraries, optimization libraries, and process libraries for industrial sites; The vertical domain base layer is used to build a large vertical domain model based on the memory layer. The large vertical domain model includes equipment diagnosis expert branches, process adaptation expert branches, and operation and maintenance optimization expert branches. The multi-agent collaboration layer is used to construct various types of agents, break down the reasoning task, assign sub-tasks to corresponding agents, construct a dynamic collaboration network based on the agents paired with each sub-task, monitor the device status, and perform reasoning based on the expert branches paired with each type of agent in the dynamic collaboration network, and output the initial reasoning result. The resource allocation layer is used to predict resources for the dynamic collaborative network, dynamically expand and shrink resources based on the resource prediction results, and allocate computing resources and industrial site resources for various types of intelligent agents in the dynamic collaborative network. The dynamic feedback layer is used to perform fact verification, compliance verification, and feasibility verification on the initial inference results output by each expert branch. Based on the verification results, it generates inference results to be merged or performs expert branch re-inference. It performs structured decomposition and unified coding on the inference results to be merged, and performs conflict resolution analysis, core fact layer fusion, execution action layer fusion, effect layer fusion, and effect evaluation on the structured decomposition and unified coding results to generate a global execution plan.

2. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 1, characterized in that, The process of setting up on-site data monitoring points and collecting indicator data from each point in the perception layer includes: Obtain the process information of the current industrial site data entry, extract the characteristics of the process unit based on the process information, and break down the industrial site data entry process into several process sub-sequences according to the characteristics of the process unit. On-site data monitoring points are set up in each process subsequence, and industrial monitoring indicators of each on-site data monitoring point are obtained by data retrieval based on the functional attributes in the process unit characteristics of the corresponding process subsequence. The on-site data monitoring points acquire indicator data in real time based on industrial monitoring indicators, mark the monitoring time, and set the monitoring cycle.

3. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 2, characterized in that, The process of constructing a full-process industrial processing model in the simulation layer includes: The process involves acquiring the physical entities of each process subsequence in the current industrial field data entry process, constructing a digital space, creating 3D models of the physical entities of each process subsequence, mapping the 3D models to the digital space, setting up API interfaces on the 3D models of each process subsequence, preprocessing the data format of the indicator data collected from the field data monitoring points of the corresponding process subsequence, generating twin data based on the preprocessed indicator data, and matching the twin data with the 3D models of the corresponding process subsequences in the digital space to generate digital twin models. Acquire different processing scenario information for each process subsequence, store the different processing scenario information in digital space, and combine the 3D models of each process subsequence in the digital twin model with the processing scenario information of each process subsequence to generate a full-process industrial processing model.

4. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 3, characterized in that, The process of building a large-scale vertical domain model from the vertical domain base layer includes: Based on the memory layer, fault datasets, process datasets, and operation and maintenance datasets are constructed. These datasets are used as training samples. The training samples are preprocessed, data augmented, standardized, and multimodal fused to generate input samples. A large vertical domain model is constructed based on the MoE sparse architecture. The core input features and triggering features of each type of expert branch in the large vertical domain model are set. The input samples are compared with the core input features and triggering features of each type of expert branch to obtain the expert branch corresponding to the input sample. Then, the shared encoder is pre-trained and the expert branches are trained independently until the loss function is stable. The model parameters are saved, and the trained large vertical domain model is output.

5. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 4, characterized in that, The process of constructing a dynamic cooperative network based on the pairing of agents for each subtask includes: The system acquires the inference task input by the user in the current industrial setting. Based on the core objective, industrial scenario constraints, and time benchmark in the inference task, the inference task is broken down into several sub-tasks. The objectives, constraints, priorities, and time requirements of each sub-task are determined. Based on the objectives, constraints, priorities, and time requirements of each sub-task, various types of intelligent agents are selected to obtain the core intelligent agent and auxiliary intelligent agent corresponding to each sub-task. A dynamic collaborative network is constructed based on the core intelligent agent, auxiliary intelligent agent, and task allocation intelligent agent of each sub-task. In the dynamic network, each core agent inputs the industrial monitoring indicators of each process subsequence into the expert branch corresponding to its core agent, and outputs the initial inference results based on the expert branch.

6. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 5, characterized in that, The resource allocation layer dynamically expands and shrinks resources, allocating computing resources and industrial site resources for various types of intelligent agents in the dynamic collaborative network. This process includes: A resource allocation history database is constructed to store resource allocation data of the dynamic collaborative network under different processing scenarios within several historical monitoring periods. A resource prediction model is constructed, and the full data in the resource allocation history database is extracted as training data. The resource prediction model is trained using the training data to obtain the completed resource prediction model. Input the current processing scenario information and the dynamic collaborative network into the resource prediction model, and obtain the computing power resource demand curve and industrial site resource demand curve of each type of intelligent agent in the dynamic collaborative network in the current monitoring period based on the resource prediction model. Monitor resources in the cloud and industrial sites to obtain current computing power resources and industrial site resources. Based on the computing power resource demand curves and industrial site resource demand curves of various types of intelligent agents in the current monitoring period, dynamically expand and shrink resources in the cloud and industrial sites in advance. The system obtains the load rate and priority of each type of intelligent agent in the dynamic collaborative network. Based on the priority and load rate of each type of intelligent agent, the system sorts the intelligent agents in ascending order and constructs a resource allocation queue. Starting from the first position of the resource allocation queue, the system allocates computing resources and industrial field resources to each type of intelligent agent in the resource allocation queue in turn, thereby obtaining the computing resources and industrial field resources of each type of intelligent agent.

7. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 6, characterized in that, The equipment condition monitoring process includes: In the dynamic network, each auxiliary agent inputs the time-series numerical values ​​of industrial monitoring indicators for each process subsequence of the current monitoring period into the equipment status monitoring model. The model compares these values ​​with preset threshold values ​​for each process subsequence, obtaining the cumulative time during which each industrial monitoring indicator is outside its corresponding threshold. Based on the cumulative time and the preset threshold values, the model determines the intermediate inference result for each process subsequence. If the cumulative time of any industrial monitoring indicator in a process subsequence exceeds the preset threshold, the intermediate inference result indicates an abnormality in that indicator. Conversely, if the cumulative time of all industrial monitoring indicators in a process subsequence is less than or equal to the preset threshold, the intermediate inference result indicates normal equipment operation.

8. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 7, characterized in that, The process of performing fact verification, compliance verification, and feasibility verification on the initial inference results output by each expert branch, and generating the inference result to be merged or performing expert branch re-inference based on the verification results, includes: For the initial inference result output by the equipment diagnosis expert branch, the abnormal industrial monitoring indicators in the initial inference result are extracted and compared with the abnormal industrial monitoring indicators in the intermediate inference result. If the abnormal industrial monitoring indicators in the initial inference result do not include the abnormal industrial monitoring indicators in the intermediate inference result, the initial inference result is marked as unqualified, and a verification deviation report is generated and sent back to the expert branch for re-inference. If the abnormal industrial monitoring indicators in the initial inference result include the abnormal industrial monitoring indicators in the intermediate inference result, the initial inference result is marked as a result to be fused. For the initial inference results output by the process adaptation expert branch and the operation and maintenance optimization expert branch, compliance verification and feasibility verification are performed on the initial inference results. If the compliance verification or feasibility verification of the initial inference results fails, the initial inference results are marked as unqualified, and a verification deviation report is generated and sent back to the expert branch for re-inference. If both the compliance verification and feasibility verification of the initial inference results pass, the initial inference results are marked as inference results to be merged.

9. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 8, characterized in that, The process of generating a global execution plan includes: A fusion temporary database is constructed, and the reasoning results to be fused from the three types of expert branches are standardized, decomposed, and uniformly encoded. The reasoning results to be fused are decomposed into standardized core elements and uniformly stored in the fusion temporary database. Conflict resolution analysis is performed on the standardized core elements of the three types of expert branches in the fusion temporary database. Then, using the triggering events in the standardized core elements as the core anchor points, the core factual conclusions of the standardized core elements of the three types of expert branches are correlated and mapped to construct causal relationship links. Based on the causal relationship links, the execution action instructions, timing requirements, constraints, and resource requirements in the standardized core elements of the three types of expert branches are fused to generate a global execution plan. The expected effects of the standardized core elements of the three types of expert branches are integrated to obtain multi-dimensional expected goals. The multi-dimensional expected goals are used as evaluation indicators. The indicator weights and fuzzy rule bases of the evaluation indicators are set according to the reasoning task input by the user. Based on the fuzzy rule base, the membership matrix of the multi-dimensional expected goals with respect to the preset quality level is obtained through fuzzy comprehensive evaluation. The quality level of the multi-dimensional expected goals is obtained according to the membership matrix and indicator weights. If the quality level of the multi-dimensional expected target is less than the preset quality level threshold, an expected target deviation report is generated and sent back to the expert branch for re-inference. If the quality level of the multi-dimensional expected target is greater than or equal to the preset quality level threshold, the numerical time series of industrial monitoring indicators of each process subsequence in the current monitoring cycle, processing scenario information, and global execution plan are input into the industrial processing full-process model for simulation to obtain the multi-dimensional simulation target and the corresponding quality level. If the quality level corresponding to the multi-dimensional simulation target is less than the quality level threshold, a simulation deviation report is generated based on the multi-dimensional simulation target and sent back to the expert branch for re-inference. If the quality level corresponding to the multi-dimensional simulation target is greater than or equal to the quality level threshold, the global execution plan is input into the agent for execution.

10. The vertical domain large-scale model multi-agent collaborative reasoning system based on dynamic feedback according to claim 9, characterized in that, The process of conducting conflict resolution analysis on the standardized core elements of the three types of expert branches includes: If the execution action instructions in the standardized core elements of an expert branch do not meet the constraints in the standardized core elements of other expert branches, a rigid conflict report for the expert branch is generated and sent back to the expert branch for re-reasoning. If the core fact conclusion A in the standardized core elements of an expert branch is inconsistent with the core fact conclusion B in the standardized core elements of other expert branches, then the specific parameter variables corresponding to the core fact conclusions of the expert branch and the other expert branches are extracted respectively. The core factual conclusions of the expert branch and other expert branches are respectively imported into the industrial processing full-process model for simulation, and the numerical time series of industrial monitoring indicators of each process subsequence in the previous monitoring cycle and the processing scenario information are imported into the model for simulation, and the data of the conclusion A verification group and the conclusion B verification group are obtained. The data from the validation groups of Conclusion A and Conclusion B are compared with the time-series values ​​of industrial monitoring indicators for each process subsequence in the current monitoring cycle. The average goodness of fit of the data from the validation groups of Conclusion A and Conclusion B is obtained. If the average goodness of fit of the validation group data is greater than or equal to a preset goodness of fit threshold, the expert branch corresponding to the validation group data with the smaller average goodness of fit is extracted from the two validation groups of data. A rigid conflict report of the expert branch is generated and sent back to the expert branch for re-reasoning. If the average goodness of fit of the two validation groups of data is less than the preset goodness of fit threshold, an automated validation failure report is generated, triggering human expert intervention as a backup.

Images