Industrial large model real-time decision assistance system

Abstract

This invention discloses a real-time decision support system for large-scale industrial models, comprising a data module, a workflow module, an intelligent agent module, and a security module. The data module is responsible for collecting multi-source heterogeneous industrial data and reconstructing contextual knowledge. The workflow module is used to compile and convert natural language commands from human-machine interaction into deterministic industrial control links and control high concurrency speed. The intelligent agent module is responsible for building a heterogeneous digital employee cluster and performing conflict arbitration and consistency locking for cross-departmental multi-objective decision-making. The security module is used to build a security red line defense mechanism for the system, implementing access security control and full operation traceability. Through the collaborative operation of these modules, this invention establishes a shared data foundation to eliminate the illusion of large models, eliminates computational randomness to ensure that decision-making reasoning logic and security thresholds remain constant, and ultimately achieves high-performance, second-level response streaming security output and an intelligent decision support closed loop for human-machine co-driving.

Description

Technical Field

[0001] This invention relates to a real-time decision support system for industrial large-scale models, belonging to the technical field of industrial large-scale models. Background Technology

[0002] With the deepening advancement of industry and intelligent manufacturing, large language models, with their powerful natural language understanding and generalized reasoning capabilities, are gradually penetrating deeper into vertical fields such as industrial control and production support decision-making, from general-domain dialogue and interaction. In complex flexible manufacturing environments, unexpected events such as sudden equipment failures, emergency order insertions, and supply chain fluctuations often require the scheduling of multi-source data and comprehensive cross-departmental assessments. Traditional industrial automation systems are mostly based on rigid preset rules or fixed scripts, making it difficult to cope with multivariable and highly dynamic on-site anomalies and emergencies. Therefore, introducing large industrial models for real-time decision support has become an inevitable trend in the evolution of smart factories.

[0003] However, existing industrial large-scale model-assisted decision-making systems have shortcomings in practical engineering applications and complex real-world production conditions: they lack physical mechanism constraints and cross-protocol alignment, making them highly susceptible to "black box illusions." When applying large models, existing decision-making systems typically simply input industrial sensor values ​​as prompts. Due to the significant protocol barriers between multi-source heterogeneous industrial data (such as numerical data like temperature, torque, and yield) and the text-based natural language that large models excel at processing, and the lack of cognitive constraints on the objective common sense of the physical world and the intricate mechanisms of industry, the models are prone to generating "illusions" that severely deviate from actual on-site conditions during inference, leading to misjudgments that violate objective physical laws or process formulations. This poses significant safety and quality risks in demanding industrial production. Furthermore, they lack adaptive arbitration capabilities for cross-departmental, multi-objective conflicts. Real-world industrial assisted decision-making often has far-reaching consequences; for example, a production line shutdown not only involves equipment maintenance but also directly impacts existing production schedules, material inventory turnover, and overall financial costs. Existing technologies typically employ a single, centralized, large model to handle all problems, failing to consider the professional depth of different industrial positions; or each business system operates independently, lacking a collaborative hub capable of driving multiple roles of "digital employees" to conduct multiple rounds of intent alignment and performing global optimal arbitration and logical consistency locking when multi-objective deadlock conflicts occur, resulting in the final output decision instructions often being incomplete. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a real-time decision support system for industrial large models that eliminates illusions based on deep fusion of multi-source data, supports conflict arbitration of heterogeneous intelligent agents, and has high-concurrency streaming speed control and full-link compliance audit traceability capabilities.

[0005] To address the aforementioned technical problems, this invention proposes a real-time decision support system for large industrial models, comprising a data module, a workflow module, an intelligent agent module, and a security module. The data module is used to acquire multi-source heterogeneous industrial data and reconstruct contextual knowledge, including an acquisition unit and a fusion unit; The acquisition unit is used to capture the dynamic production feature flow of the entire plant area in real time and concurrently via the bus; The fusion unit integrates the retrieval enhancement generation architecture and the model context protocol to align the retrieved expert knowledge with real-time production dynamic data and reconstruct and generate a feature semantic enhancement dataset that includes complete industrial operating conditions. The workflow module is used to compile and convert natural language commands for human-machine interaction into deterministic links for industrial control and to control the speed of high concurrency. It includes a generation unit and a flow control unit. The generation unit has a built-in semantic parsing operator, which is used to automatically identify business intentions and control variables, and automatically reconstruct and generate industrial automation workflows; The flow control unit is configured with a token distributor and a physical computing thread pool dynamic adjuster to perform token flow control distribution and dynamic load balancing among multiple hardware components during high-frequency processing periods. The intelligent agent module is used to construct a heterogeneous digital employee cluster and perform conflict arbitration and consistency locking for cross-departmental multi-objective decision-making, including a collaboration unit and a scheduling unit; The collaborative unit is used to manage and run multiple heterogeneous, dedicated distributed intelligent agents for specific industrial positions in parallel. The scheduling unit acts as a shared communication bus to drive each agent to perform intent alignment and data interaction, performs conflict arbitration on the unidirectional decision results fed back by each agent, merges them into a globally optimal collaborative decision-making auxiliary instruction, and has a built-in logical consistency verification engine to perform reduction constraints on the output parameters through adversarial verification. The security module is used to build the system's security red line defense mechanism, including an access control unit and an auditing unit; The permission unit has a built-in three-level dynamic control matrix of roles, spaces, and functions, which is used to implement fine-grained data space isolation and access security control; The audit unit is equipped with a hardware-level full-operation audit engine, which is used to fully record the underlying behavior of model inference and bind data link fingerprint logs to the final decision conclusion.

[0006] The above includes the following steps: S1. Capture spoken natural language speech stream through the workflow module, and parse and extract the composite intent containing the core business intent and control target variables.

[0007] S2. The data module sends a full-speed sampling command to the perception node based on the composite intent to hard switch to high-frequency sampling state, and simultaneously collects cross-domain multimodal production feature data streams, and combines retrieval enhancement generation with expert knowledge fusion to encapsulate it into a feature semantic enhancement dataset. S3. Activate multiple cross-departmental heterogeneous dedicated distributed intelligent agents through the intelligent agent module, carry out multiple rounds of intent alignment and data interaction based on the shared memory base, and perform arbitration through the conflict judgment matrix operator. Combined with logical consistency verification, lock and generate the optimal collaborative decision-making auxiliary instruction. S4. The workflow module's flow control unit uses a streaming word distribution mechanism to perform hardware-level high-concurrency speed control, and seamlessly splits tasks to cascaded heterogeneous devices for distributed load balancing when encountering peak conditions. S5. By comparing the security module with the three-level control matrix, unauthorized interception and verification are performed, full-scale operation behavior is captured and data link fingerprint is bound for compliance hard audit, and finally the streaming security output of the digital decision-making solution is completed.

[0008] In step S3 above, when the agent module handles content that is outside its assigned role and goes beyond its boundaries, it specifically executes the following secondary adaptive fault-tolerant sub-step: S31. The dedicated intelligent agent calculates the semantic similarity dot product between the instruction text and the ontology vocabulary through the local domain classifier, and outputs the local domain confidence score. S32. If the confidence level is lower than the preset threshold or the reverse constraint feature word is hit, the local inference chain of the large model is forcibly broken, and the instruction is repackaged into an out-of-bounds data packet and marked as an abnormal suspension by the abnormal feature operator. S33. By reversing the out-of-bounds data packet back through the shared communication bus inside the intelligent agent module, the global intent routing engine in the scheduling unit is immediately awakened. S34. The global intent routing engine performs multi-dimensional correlation comparison of out-of-bounds instructions in the global intent mapping matrix to re-identify the true target's domain, rewrites the routing header label, clears historical confidence, accurately directs and transfers the instructions, and sends them to the input queue of the target-adaptive agent.

[0009] The generation unit of the above workflow module specifically includes: The system captures spoken natural language speech streams at high frequencies, uses acoustic beamforming algorithms for pre-stage denoising, and filters out noise using bandpass hardware filters and adaptive spectral subtraction algorithms. The clean audio stream is then input into an automated speech recognition acoustic model fine-tuned with a built-in industrial corpus, decoded, and converted into a standard text character stream. A sequence labeling algorithm is run to perform named entity recognition, extracting spatial and asset entity slots, as well as state and constraint feature slots online. A syntactic tree dependency analysis algorithm is used to parse the grammatical dominance relationships between words, segmenting long sentences into clause structures with independent control intentions. By comparing with a preset industrial action mapping matrix, the system identifies the core verbs and object mappings of each clause as target intentions, constructs topological dependency chains based on sequential logic, and unifies heterogeneous multi-target intentions into a unified composite intention.

[0010] The data acquisition unit of the above data module specifically includes: When the generation unit identifies and reconstructs a composite intent or an alarm signal triggered by an abnormal operating condition by the field equipment, the acquisition unit, as the master control node, sends a full-speed sampling command to each sensing node, production line control unit and business system interface, and controls each heterogeneous data channel to hard switch from low-frequency monitoring mode to full-speed high-frequency sampling state; and simultaneously captures field physical sensing feature data, production line status and micro-physical attitude data, management and operation and external scene context data. The local microcontroller unit forcibly marks the concurrently received data streams with a globally unified timestamp distributed by the system crystal oscillator, calculates the time difference in a rolling manner, and encapsulates signals with time differences less than or equal to a preset synchronization time deviation threshold through cross-modal multi-dimensional matrix alignment, generating an aligned dataset that is streamed to the fusion unit.

[0011] The fusion unit of the aforementioned data module specifically includes: The parsing operator extracts the physical identifier and instantaneous anomaly feature vector from the aligned dataset as a high-dimensional query key. Nearest neighbor similarity matching is calculated in a local private high-dimensional vector database, retrieving the expert rule text with the highest matching degree and historical best-case scenario data. Cross-protocol format conversion is performed using the model context protocol. Through a preset industrial semantic mapping template, the numerical dynamic production data returned from each channel is automatically mapped into working condition statements to eliminate heterogeneous source protocol barriers. The generated working condition statements are incrementally concatenated with the retrieved expert knowledge text. Semantic embedding overlay technology is used to perform deep fusion at the text semantic level. A system-preset global task prompt shell is added to the front-end and back-end to package and reconstruct a feature semantic enhancement dataset, which is then streamed into the shared memory of the intelligent agent module's collaborative unit.

[0012] The aforementioned collaborative unit includes parallel scheduling agents, inventory agents, financial agents, and operations and maintenance agents. The operations and maintenance agent, after receiving the feature semantic enhancement dataset, employs a closed-loop control mechanism for the re-identification and transfer of out-of-bounds content, specifically including: Each dedicated agent calls the local domain classifier, performs semantic similarity dot product calculation between the received instruction text and the local domain ontology vocabulary, outputs the local domain confidence score, and verifies whether the instruction contains explicit inverse constraint feature words. If the local domain confidence is lower than the preset safety threshold or the reverse constraint feature word is hit, the local inference chain of the current agent is forcibly broken, the instruction is marked as an out-of-bounds exception and suspended, and the exception feature operator is used to repackage and combine it into a standard out-of-bounds data packet. By reversing the out-of-bounds data packets through the shared communication bus, the global intent routing engine inside the scheduling unit is activated. The global intent routing engine performs multi-dimensional correlation comparison on the original instruction in the global intent mapping matrix to re-identify the real target's domain, rewrites the routing header label and clears the local confidence verification history, accurately directs and moves the instruction to the target-adaptive agent's pending input queue.

[0013] The scheduling unit of the aforementioned intelligent agent module specifically includes: A dynamically updated shared blackboard data area is established within a private secure computing domain. The bus listener frequently polls the one-way decision drafts generated by each agent and writes them into this area. By using the conflict determination matrix operator for rolling monitoring, when a multi-target deadlock conflict event is triggered, the output link of the decision scheme is immediately cut off. The reverse feedback mechanism is called to repackage the blocking text and the correction opinion as strong constraint feedback variables into high-priority prompt word increments, which are then thrown back to the source agent to drive secondary inference to modify the control parameters. When the parameter conflict mutation rate is continuously lower than the convergence threshold, the optimal collaborative decision auxiliary instruction that takes into account the overall situation is selected through multi-index weighted logic. The internal logical consistency verification engine extracts the control parameters in the instruction and performs spatial dimension reduction constraints on the safety threshold red line matrix pre-stored in the register. If the boundary is exceeded, it reverses and forcibly offsets and rewrites the lock to the safety red line critical value. The current inference logic chain is compared with historical similar audit logs across time and space. If the morphological deviation exceeds the preset mutual trust threshold, the model is determined to be hallucinating and the logic is reset. The pre-stored baseline workflow is forcibly invoked to perform correction and replacement.

[0014] The aforementioned workflow module specifically includes: The internal token unit extracts the expected response latency from each concurrent request and pushes the requests into different hardware execution queues according to the urgency of the latency, thus reconstructing a streaming buffer with latency gradients. Real-time monitoring of hardware core temperature, memory usage, and computation time levels; dynamic rewriting of the underlying computing kernel's thread allocation registers; expansion of the parallel computing thread pool for high-priority tasks. A streaming word rate control and distribution mechanism is adopted during the text generation lifecycle to dynamically adjust the token injection rate of the word bucket algorithm to control the character throughput per second. When demand exceeds the safety limit of single-machine computing power, the internal expansion unit polls the remaining available computing power and bandwidth of other integrated hardware devices in the factory, seamlessly splits and packages the analysis tasks that exceed the capacity into independent data streams, and dynamically distributes them to other cascaded idle hardware for physical elastic distribution via industrial Ethernet.

[0015] The permission and audit units within the aforementioned security module specifically include: The access control operator of the permission unit captures the globally unique identity of the current request initiator and compares it with the three-level dynamic control matrix of the system, which includes role permissions, data space permissions and function permissions. If any level fails, the highest priority blocking and suspension decision is initiated. The audit unit's hardware-level full-scale operation audit engine records the execution of underlying behaviors throughout the entire lifecycle, including data access, model interface calls, and agent parameter modification. The reverse link tracing algorithm is activated to correlate the control conclusions derived by each agent with the original data source it calls using a causal graph. A unique data link fingerprint log is calculated and bound and written into an immutable private long-term storage library. Finally, the network security gateway is driven to provide streaming security output to the execution mechanism.

[0016] This invention has positive effects: (1) This invention integrates the data module to enhance the generation architecture and model context protocol, and performs cross-protocol alignment and deep fusion of multi-source heterogeneous expert knowledge and cross-domain multimodal real-time production dynamic data to reconstruct and generate a feature semantic enhancement dataset containing complete industrial site conditions, thus establishing a deterministic shared data foundation to eliminate the black box illusion of large models. At the same time, in conjunction with the logical consistency verification engine of the intelligent agent module, the space dimension threshold reduction constraint and time dimension adversarial consistency verification are performed, effectively eliminating the computational randomness of large language model inference and ensuring the constant consistency of auxiliary decision control command output in the time and operator dimensions.

[0017] (2) This invention achieves accurate decoupling of heterogeneous multi-target intents and deep recognition of multi-condition composite intents in complex long sentences of natural language by using the industrial instruction semantic parsing operator and syntactic tree dependency analysis algorithm built into the generation unit of the workflow module. It can automatically reconstruct and generate a directed acyclic workflow including node dependency topology graph online, reducing manual repetitive work. (3) This invention relies on the token unit and extension unit of the flow control unit to introduce streaming word rate control distribution, hardware-level dynamic adjustment of the computing thread pool, and multi-machine horizontal flexible expansion and dynamic load balancing mechanism of the task flow. It effectively prevents hardware deadlock under high concurrency, ensures that the overall large model inference computing throughput rate of the whole plant is kept stable in the high performance range, and ensures low latency control response of the system when facing high frequency peak batch analysis tasks. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a schematic diagram of the module and unit. Detailed Implementation

[0020] (Example 1) This invention provides a real-time decision support system for industrial large-scale models. The system is built on a privately deployed integrated industrial large-scale model, serving as an edge central computing platform that houses the factory's intelligent brain and digital employee cluster. It establishes bidirectional data cascading with a cloud-based large-scale model server through a network security gateway, and utilizes the factory's internal industrial bus to electrically connect or achieve bidirectional data cascading with various sensing nodes, manufacturing execution system terminals, enterprise resource planning terminals, and early warning execution mechanisms in the production workshop.

[0021] The industrial large-scale model real-time decision support system includes a data module, a workflow module, an intelligent agent module, and a security module. The intelligent agent module, as the core decision-making and collaborative control hub, performs cross-domain data synchronization with the data module, workflow module, and security module, respectively.

[0022] The specific details of each of the above modules are as follows: The data module is located at the input bus gateway of the edge computing all-in-one machine, and is responsible for acquiring multi-source heterogeneous industrial data and reconstructing contextual knowledge. The data module includes an acquisition unit and a fusion unit; The acquisition unit is used to poll and concurrently capture the dynamic production feature flow of the entire plant in real time through industrial Ethernet interface or bus channel, including on-site physical sensing feature data of workshop assembly line sensors, operation data and micro-physical attitude data of manufacturing execution system (MES) terminals, and global path planning and external environment context data of enterprise resource planning system (ERP) terminals.

[0023] The fusion unit integrates the Retrieval Enhanced Generation (RAG) architecture and the Model Context Protocol (MCP) to retrieve pre-stored lean expert guidance rules, historical fault maintenance logs, and process recipe texts from the system's private database. Through vector similarity retrieval, it associates the aforementioned expert knowledge with the real-time production dynamic data acquired by the acquisition unit at the millisecond level, reconstructing and generating a semantic feature-enhanced dataset that includes complete industrial operating conditions.

[0024] The workflow module is integrated into the core central processing unit of the edge computing power appliance. It is used to compile and translate natural language commands from human-machine interaction into deterministic industrial control links and to control high-concurrency speed. The workflow module includes a generation unit and a flow control unit. The generation unit has a built-in zero-threshold application workflow building engine. It is used to receive spoken natural language assistance requests from operators, and automatically identify business intentions and control variables using semantic parsing operators of large language models. It automatically builds and generates industrial automation workflows, AI assistant parameters and business intelligence (BI) dashboard configurations for specific business scenarios online, reducing repetitive manual work.

[0025] The flow control unit is equipped with a token distributor and a physical computing thread pool dynamic adjuster. It is used to perform flow control during high-frequency peak processing periods such as industrial shift changes or system inspections. Through dynamic load balancing scheduling among multiple all-in-one hardware units, it ensures that the overall inference computing throughput rate of the system is stably maintained in the high-performance range when processing large batches of analysis tasks, achieving second-level control response.

[0026] The intelligent agent module, privately hosted and running within the private secure computing domain of the edge computing appliance, is used to build heterogeneous digital employee clusters and perform conflict arbitration and consistency locking for cross-departmental multi-objective decision-making. The intelligent agent module includes a collaboration unit and a scheduling unit; The collaboration unit is used to manage and run multiple heterogeneous, dedicated, distributed intelligent agents for specific industrial positions in parallel. These include a production scheduling intelligent agent that performs flexible production planning calculations, an inventory intelligent agent that performs precise material throughput alignment, a financial intelligent agent that performs plant-wide operating cost actuarial calculations, and an operation and maintenance intelligent agent that performs on-site fault diagnosis and standardized repair guidance.

[0027] The scheduling unit, as a shared communication bus for the distributed intelligent agent cluster, drives the various intelligent agents within the collaborative unit to perform multi-round intent alignment and data interaction. It performs multi-index alignment and conflict arbitration on the unidirectional decision results fed back by each intelligent agent, and merges them into a globally optimal decision-making auxiliary instruction. At the same time, the scheduling unit has a built-in logical consistency verification engine, which uses an adversarial verification mechanism to apply multiple reduction constraints to the inference output of the large model, ensuring that the output decision inference logic remains consistent with the safety threshold for the same industrial production problem under different time periods and different operator-triggered scenarios.

[0028] The security module is positioned at the boundary defense point of data flow to build the system's security red-line defense mechanism. The security module includes an access control unit and an auditing unit. The permission unit has a built-in three-level dynamic control matrix of role permissions, data space permissions, and function permissions, which is used to implement fine-grained data space isolation and access security control for users, workshop equipment, and intelligent agents accessing the system.

[0029] The audit unit is equipped with a hardware-level full-operation audit engine, which is used to perform tamper-proof full audit and traceability of all data access behaviors, large model API call frequency and Agent parameter modification records that occur during the system life cycle, and to establish a complete data link log for all key decision conclusions derived from large model inference, ensuring that key conclusions can be traced back to the lowest level of original sensor data source.

[0030] In the overall operation, the workflow of each module of this system is as follows: First, the workflow module captures real-time industrial site demands and orchestrates the topology flow. The linkage data module deeply integrates real-time operating conditions with expert knowledge across protocols, establishing a shared data foundation to eliminate the illusion of large models. Next, the agent module schedules multiple heterogeneous distributed agents to conduct cross-departmental intent alignment and conflict arbitration, outputting globally optimal collaborative instructions. Subsequently, the security module relies on a three-level permission control matrix and full audit to intercept unauthorized behavior and establish traceable logs. Finally, the flow control unit within the workflow performs high-performance streaming speed control and load balancing, ensuring safe streaming output of decisions with second-level response during peak periods.

[0031] The specific workflow of the unit within the module is as follows: Operators input spoken natural language control commands through the interactive terminal, activating the generation unit within the workflow module. The generation unit utilizes built-in semantic parsing operators to identify and extract the core business intent, control target variables, and associated workshop entities from the commands online. Upon identifying a composite intent, the generation unit sends a high-priority bus scheduling request to the acquisition unit within the data module. The acquisition unit responds to the request, acting as the master control node, and issues full-speed sampling commands in parallel to all sensing nodes in the workshop via the industrial Ethernet interface. This controls each data channel to dynamically switch from low-frequency, low-power monitoring mode to full-speed, high-frequency sampling mode, capturing cross-domain, multimodal, heterogeneous feature streams, including on-site physical sensor characteristic data, equipment status and yield data, and operation management and delivery time data.

[0032] The acquisition unit marks all concurrently received heterogeneous data streams with a globally unified timestamp. All signals with time differences less than or equal to a preset synchronization time deviation threshold are encapsulated using a multi-dimensional matrix alignment. The resulting aligned dataset is then streamed to the fusion unit within the data module. Upon receiving the dataset, the fusion unit extracts device identifiers and anomaly features. It performs rapid nearest-neighbor vector similarity comparisons in a private high-dimensional vector database, retrieving the highest-matching set of expert rule texts and historical best-case scenario data. The fusion unit utilizes a model context protocol to align textual expert knowledge with numerical dynamic production data across protocols. It then uses semantic embedding overlay technology to uniformly encapsulate and reconstruct a feature-enhanced dataset containing a complete industrial context. Subsequently, the fusion unit streams this dataset in real-time into the shared memory bus of the collaborating units within the intelligent agent module, establishing a shared data foundation for the distributed intelligent agent cluster.

[0033] Simultaneously with data establishment, the generation unit within the workflow module, based on the identified business intent, retrieves the corresponding atomic operation nodes from the standard control template library. Through online automatic reconstruction using a graph relationship network, it generates a directed acyclic workflow including a node-dependent topology graph. This workflow, along with corresponding background parameters, is synchronously streamed to the scheduling unit within the agent module. Based on the received workflow, the scheduling unit adaptively wakes up and coordinates multiple cross-departmental heterogeneous roles within the collaboration unit, such as the production scheduling agent, inventory agent, financial agent, and operations and maintenance agent. Each dedicated agent concurrently extracts the contextual semantic features required for its respective role from the shared data base for local large-scale model inference and performs multiple rounds of intent alignment within the shared data space. The scheduling unit collects unidirectional decision-making schemes from each agent, runs a conflict arbitration algorithm to comprehensively evaluate positive and negative feedback on various indicators, and performs multi-objective convergence to select the optimal collaborative decision-making auxiliary instruction that balances plant-wide efficiency and financial costs. Simultaneously, the scheduling unit's internal logical consistency verification engine performs adversarial verification to eliminate computational randomness in the large-scale model, ensuring that the decision-making inference logic and safety thresholds remain constant across time and operator dimensions.

[0034] The scheduling unit sends the optimal collaborative decision-making assistance instruction derived from arbitration to the flow control unit within the workflow module. The token unit within the flow control unit initiates a dynamic distribution mechanism for streaming inference threads, dynamically rewriting the priority registers of the underlying computing threads based on the expected response latency of different concurrent requests. When peak batch analysis tasks, such as shift changes, cause the single-machine load to exceed the threshold, the expansion unit within the flow control unit issues load balancing instructions, horizontally distributing the tasks to other cascaded integrated hardware, ensuring that the overall inference computing throughput rate of the entire plant remains stable at a high-performance range of no less than 500 tokens per second, guaranteeing decision-making response within seconds.

[0035] Before the decision results are finally output to the terminal screen or edge execution mechanism, the permission unit within the security module first performs an unauthorized interception verification against the receiver by comparing the three-level dynamic control matrix composed of role permissions, data space permissions, and functional permissions, preventing the leakage of confidential processes or financial formulas across spaces. After verification, the audit unit within the security module automatically captures modification records and model interface call behaviors in the entire multi-agent collaborative debate chain, binding the generated conclusion with complete data links and inference logs, and performing a full compliance hard audit. After confirming that there is no risk of sensitive information leakage, the audit unit removes the interception status and persistently stores the logs in the database. Finally, the master control scheduling command securely outputs the digitally intelligent decision solution, which has been traced with trustworthiness, to the field terminal, completing the full-unit adaptive control closed loop of the entire decision support system.

[0036] The aforementioned generation unit specifically includes: The underlying implementation control logic and technical implementation process of the generation unit within the workflow module for parsing and extracting voice signals and accurately identifying and determining complex intentions are as follows: The generation unit captures the spoken natural language speech stream input by the operator in a high-frequency streaming manner through a local wireless voice interaction terminal. Before performing speech-to-text conversion, the generation unit first performs low-level noise reduction processing: Acoustic beamforming and pre-amplifier denoising: The interactive terminal uses a multi-microphone array, and the generation unit calls the built-in acoustic beamforming algorithm to target and lock the operator's speaking position and suppress physical environmental noise in non-target directions; at the same time, the original audio signal is filtered out in real time by a bandpass hardware filter and an adaptive spectral subtraction algorithm to remove mechanical vibration noise and random white noise at fixed frequencies.

[0037] Industrial acoustic feature stream decoding: The denoised, clean audio stream is fed in real-time into the automated speech recognition acoustic model built into the generation unit. This acoustic model is pre-embedded with a specialized vocabulary that has been deeply fine-tuned from a corpus of industrial vertical domains, including equipment identification codes, standard process terms, and workshop job abbreviations. The acoustic model decodes the temporal and frequency domain features of the audio with high precision and converts them into a standard text character stream.

[0038] When operators issue complex long sentences involving multiple overlapping business operations, the generation unit, relying on built-in industrial instruction semantic parsing operators and graph network dependency analysis matrices, achieves deep recognition and accurate decoupling of complex intentions across multiple operating conditions through the following steps: Industrial named entity and status slot extraction: The generation unit first runs a sequence labeling algorithm to perform named entity recognition on the streaming text character stream, and retrieves and extracts space and asset entity slots (such as automatically marking the location entity as "A line" and the equipment entity as "mixed flow equipment") and status and constraint feature slots (such as marking the abnormal event status as "stuck and stopped" and the time feature variable as "3 hours").

[0039] Semantic Dependency Parsing and Semantic Sentence Segmentation: The generation unit invokes a syntactic tree dependency parsing algorithm to analyze the grammatical dominance relationships between words in a long sentence. By capturing logical conjunctions and core verbs in the text, the algorithm semantically segments the complex sentence into two or more clause structures with independent control intentions, breaking the black box of a single long sentence.

[0040] Composite Intent Matrix Reconstruction and Action Mapping: For the first clause "Please provide maintenance instructions", the unit compares with the preset industrial action mapping matrix to identify its core verb as "instructions" and object as "maintenance", thus accurately mapping it to the first target intent: equipment fault diagnosis and operation and maintenance control decision support; For the second clause "readjust today's production schedule", the generating unit identifies its core verb as "readjust" and its object as "production schedule", thus mapping it to the second target intent: flexible dynamic scheduling of multiple production lines.

[0041] Dependency topology chain construction: The generation unit associates the identified first and second target intentions, and automatically determines the causal triggering dependency relationship between the two intentions based on the sequential logic of the syntax tree. Finally, the generation unit unifies and defines this group of interrelated heterogeneous multi-target intentions with spatiotemporal dependencies as a composite intention. The aforementioned data acquisition unit specifically includes: The technical implementation process of the specific acquisition control mechanism of the acquisition unit within the data module, the data content acquired in parallel, and the data flow and transmission target is as follows: Linked wake-up and hard switching of sampling modes: The acquisition unit maintains real-time cascading with the generation unit of the workflow module through the industrial backbone bus. When the generation unit identifies and reconstructs the operator's multi-command complex intent across business operations through semantic parsing operators, or when the field automation equipment triggers an abnormal operating condition alarm signal, the system performs central scheduling and activates the data synchronization acquisition process. As the core master control node for data sampling, the acquisition unit, upon receiving the trigger activation signal, sends high-priority bus commands for full-speed sampling in parallel to each sensing node on the bus, the production line PLC control unit, and the database interface of the upper-level business system. After responding to the bus commands, each heterogeneous data channel uniformly hard switches from a low-frequency, low-power suspended monitoring mode to a full-speed, high-frequency sampling state.

[0042] Parallel acquisition content: Under full-speed, high-frequency sampling, the acquisition units capture three types of cross-domain, multimodal, heterogeneous feature data streams in parallel: On-site physical sensing characteristic data includes real-time working pressure on the surface of workshop equipment, lateral torque impact of steering column and robotic arm base, continuous temperature field data of key bearings and heating modules of equipment, multi-axis acceleration and angular velocity impact values ​​of drive shaft, real-time light pressure inside or outside the workshop, and environmental temperature and humidity characteristic values.

[0043] Production line status and microscopic physical attitude data: including the instantaneous utilization rate of each production line's current process, the relative deviation speed of the lane keeping line or material trajectory line, the real-time yield rate of the production line, the spatial motion contour of the robotic arm's hands, and the instantaneous three-axis acceleration variation rate of the vehicle body or equipment retrieved by the high-precision inertial navigation unit. Management and operation contextual data: including real-time process formula attributes of products, delivery deadline of current orders, raw material inventory balance data, congestion level index of the current driving or production segment, and geometric topological risk characteristics of the path ahead.

[0044] Data Processing and Flow: The local microcontroller unit built into the acquisition unit forcibly marks all concurrently received data streams with a globally unified timestamp distributed by the system crystal oscillator. The acquisition unit continuously calculates and compares the timestamp differences of each signal, and performs cross-modal multi-dimensional matrix alignment and encapsulation on all multimodal heterogeneous signals with time differences less than or equal to a preset synchronization time deviation threshold. This transforms the signals into a unified structured aligned dataset, which is then transmitted in real time and streaming to the fusion unit inside the data module via the high-speed data bus inside the edge computing power all-in-one machine.

[0045] The aforementioned fusion unit specifically includes: After receiving the dataset streamed from the acquisition unit, the fusion unit within the data module performs deep fusion of the data and expert knowledge through the following steps, ultimately encapsulating it into a feature semantic enhancement dataset: Knowledge Base Retrieval and Feature Extraction: The fusion unit utilizes a built-in retrieval enhancement and generation technology architecture. The fusion unit's local parsing operator performs high-frequency feature extraction on the received aligned dataset, extracting unique physical identifiers for the current equipment, instantaneous anomaly feature vectors from sensors, and current order process status labels. Subsequently, the fusion unit uses the physical identifier and anomaly feature vectors as high-dimensional query keys to perform rapid nearest-neighbor similarity matching calculations in a locally deployed vector database. The vector database pre-stores historical fault maintenance logs, standard production process recipes, and lean expert guidance guidelines. Through similarity comparison, the fusion unit retrieves and displays the set of expert control rule texts and historical best-case scenario data that best match the current workshop conditions and equipment fault characteristics within milliseconds.

[0046] Cross-protocol data textualization: The fusion unit utilizes a built-in model context protocol to perform cross-protocol data stream alignment and format conversion. The fusion unit correlates the structured, numerical, dynamic production data acquired by the acquisition unit with retrieved textual expert knowledge across protocols. Specifically, the fusion unit automatically translates the physical sensor values ​​(such as instantaneous bearing temperature, macroscopic torque of tubing, and real-time yield percentage of the production line) transmitted from various data channels into textual statements conforming to human natural language logic, thereby eliminating protocol barriers between heterogeneous data sources.

[0047] Semantic concatenation and multimodal dataset encapsulation: The fusion unit uses the translated working condition statements as the basic context of the current production state, and incrementally concatenates them with the retrieved expert process recipe text and fault diagnosis tree logs. Through semantic embedding overlay technology, the fusion unit performs deep fusion of the three core elements—the large model foundation, expert knowledge, and industrial dynamic data—at the text semantic level, thereby establishing a deterministic boundary to eliminate the black-box illusion of the large model. Finally, the fusion unit automatically adds system-preset global task prompt words to the front and back ends of the concatenated text, packages it to generate a feature-enhanced semantic dataset containing complete industrial site working conditions, and streams it to the collaborating units within the intelligent agent module.

[0048] The aforementioned collaborative units specifically include: After receiving the feature semantic enhancement dataset, the collaborative units within the intelligent agent module employ the following private customized training mechanism, judgment and recognition logic, and closed-loop control process for handling out-of-bounds content: Private Customized Training Mechanism for Dedicated Intelligent Agents: The scheduling, inventory, finance, and operations intelligent agents within the collaborative unit are all based on an open-source general-purpose large language model foundation that has been privately fine-tuned and reconstructed with tool binding. Each dedicated intelligent agent is trained and generated using the following customized data fine-tuning scheme during factory release or system deployment: Training of the production scheduling agent: A dedicated training set is built based on the historical production scheduling logs of the manufacturing execution system, the workshop assembly line scheduling instruction set, and the dynamic control rules of multiple processes; through supervised fine-tuning of the process, the model base deeply grasps the reasoning thinking chain of multi-line flexible diversion and capacity optimization alignment under sudden constraints such as abnormal assembly line shutdowns and order changes.

[0049] Training of the inventory agent: The agent is trained using the factory material requirements planning log, the raw material coding dictionary of the enterprise resource planning system, and the safety stock level control rules; the focus is on strengthening its ability to break down multi-level bills of materials and the accuracy of its assessment of safety fluctuations in the upstream and downstream of the supply chain.

[0050] Training of the financial intelligence agent: Supervised fine-tuning based on factory manufacturing cost allocation rules, logistics transfer cost matrix, and workshop asset depreciation model; enabling it to have the professional intent to dynamically calculate multi-dimensional financial losses and generate marginal contribution analysis reports for different production scheduling plans.

[0051] Training of the operation and maintenance intelligent agent: Deeply bind to the standard official maintenance manuals, lean expert guidance rules, and physical sensor fault code registry of various production equipment in the factory; enable it to perform semantic alignment of multi-channel hardware data and focus on the streaming guided reasoning of fault root cause localization and standard maintenance workflow.

[0052] The decision-making and recognition logic of each specialized agent: After the feature semantic enhancement dataset is streamed into the shared data space of the agent module, each specialized agent performs decision-making and recognition locally in parallel using its built-in domain ontology vocabulary and trigger feature extraction operators. The production scheduling agent identifies and determines the production status by searching the shared data space in real time for production status features such as equipment downtime and delivery delays. Once it detects that the real-time yield rate or utilization rate of the production line deviates from the normal benchmark value, it adaptively determines that a production scheduling conflict event has occurred and automatically extracts the time window of the conflict's impact and the affected processes, and initiates dynamic production scheduling reasoning.

[0053] Inventory agent identification: Taking the production scheduling adjustment draft output by the production scheduling agent as input, the product models and production quantities involved in the draft are extracted and cross-compared with the underlying bill of materials dictionary to automatically calculate the instantaneous demand for raw materials; then, it is compared with the current material inventory balance data to determine the size. If the demand is greater than the safety stock level, a material shortage blocking signal is identified and output.

[0054] Financial intelligence agent identification: Extract the turnover distance caused by logistics adjustments and the penalty threshold caused by delayed delivery in the plan. Through a multi-indicator weighted algorithm, identify and determine online whether the overall financial cost of the plan is within an economically reasonable bounded range.

[0055] The intelligent operation and maintenance agent determines and identifies the following: It directly polls the field physical sensor feature data in the feature dataset to extract the physical vibration frequency, instantaneous working pressure and temperature field data of each device and key bearings; when the sensor value exceeds the preset safety red line threshold, it automatically identifies and matches the fault tree and outputs accurate equipment diagnosis conclusions.

[0056] Re-identification and transfer closed-loop control mechanism for out-of-bounds content: If a dedicated intelligent agent receives heterogeneous and atypical demand content that is outside its job scope in its dedicated input channel or task queue, the system performs adaptive re-identification and safe transfer through the following four control stages: Phase 1: Local Domain Confidence Detection: Each dedicated agent first invokes its local domain classifier to calculate the semantic similarity dot product between the received instruction text and the local domain ontology vocabulary, outputting a local domain confidence score between zero and one. Simultaneously, the agent also verifies whether the instruction includes explicit inverse constraint feature words.

[0057] Phase Two: Out-of-bounds Anomaly Suspension and Encapsulation: If the calculated local domain confidence score is lower than the preset safe passage threshold, or if a reverse constraint feature word is forcibly hit, the current agent determines that the request content is out of bounds. The system immediately forcibly suspends the current agent's local large model inference chain for this instruction, and marks the instruction status as "out-of-bounds anomaly suspension." The system then uses an anomaly feature operator to re-encapsulate and combine the original instruction, the agent identifier that reported the error, and the timestamp into a standard out-of-bounds data packet.

[0058] Phase 3: Multi-Agent Bus Back-Throw and Scheduling Center Reception: The current agent throws the out-of-bounds data packet back via the shared communication bus within the agent module. The scheduling unit within the agent module listens for and captures the out-of-bounds data packet in real time via the bus, and the global intent routing engine within the scheduling unit is immediately activated.

[0059] Phase Four: Global Re-identification and Precise Directed Transfer: The scheduling unit's global intent routing engine performs cross-service multi-instruction secondary feature parsing on the original instructions in the out-of-bounds data packets. The routing engine re-identifies the true target domain (e.g., financial control domain) of the instruction by performing multi-dimensional correlation comparisons within the global intent mapping matrix. After identification, the scheduling unit rewrites the routing header label of the data packet, clears its local confidence verification history, precisely directs it, and pushes it into the pending input queue of the target adaptive agent (e.g., the financial agent), achieving adaptive fault tolerance and closed-loop flow for non-self-service applications.

[0060] The aforementioned scheduling unit specifically includes: The scheduling unit within the intelligent agent module serves as the collaborative brain and communication bus of the distributed intelligent agent cluster. Its specific scheduling control mechanism, cross-departmental multi-objective conflict arbitration logic, and logical consistency locking technical implementation process are as follows: Distributed intelligent agent bus orchestration and dynamic workflow scheduling mechanism: The scheduling unit establishes a dynamically updated "shared blackboard data area" within the private secure computing domain of the edge computing power appliance. The master control state machine inside the scheduling unit maintains the execution status of each node in the workflow in real time (e.g., waiting, activating, suspended, completed). The scheduling unit parses the node dependencies in the directed acyclic graph. When it determines that a certain stage has the conditions for parallel computing, the scheduling unit simultaneously issues wake-up bus commands to multiple corresponding intelligent agents through the internal communication bus, and seamlessly streams the current feature semantic enhancement dataset to the local input queues of both, triggering distributed local inference between them. The bus listener of the scheduling unit frequently polls the output ports of each intelligent agent. Once an intelligent agent completes inference, the scheduling unit adds the intelligent agent identifier and global timestamp to its generated one-way decision draft and streams it into the shared blackboard data area.

[0061] The adaptive arbitration control logic for cross-departmental multi-objective conflicts: The scheduling unit has a built-in conflict determination matrix operator. When an agent writes a decision plan in the blackboard area, the scheduling unit automatically passes that plan as an independent variable into the rule constraint boundaries of other agents. If other constraint boundary agents write blocking signals or correction opinions in the blackboard area, the scheduling unit's conflict determination matrix operator determines that a multi-objective deadlock conflict event has been triggered through textual logic comparison.

[0062] After determining a conflict, the scheduling unit immediately disconnects the output link from the current decision scheme to the subsequent flow control module. The scheduling unit invokes the reverse feedback mechanism to automatically repackage the blocking texts and correction suggestions thrown out by other intelligent agents as strongly constrained feedback variables; The scheduling unit uses the encapsulated feedback packet as a high-priority prompt word increment and throws it back into the input channel of the source agent, forcibly driving it to perform secondary large model local inference based on the new constraint boundary; In the secondary inference, the source agent modifies its control parameters, flexibly diverts production factors, and writes them back into the blackboard area.

[0063] The scheduling unit continuously monitors the aforementioned multi-round interaction process. When the conflict variation rate of the interaction parameters of each agent in the blackboard area is lower than the preset convergence threshold for two consecutive cycles, or when the system's preset maximum evolution round is reached, the scheduling unit uses multi-index weighted evaluation logic to perform scalarized benefit projection on the currently converged solution, and selects a balance point that takes into account both global efficiency and the lowest marginal cost as the final optimal collaborative decision-making auxiliary instruction, thereby completing intelligent arbitration.

[0064] A logical consistency locking mechanism based on generation reduction constraints: The scheduling unit has a built-in logical consistency verification engine that triggers constraints on the execution conditions of the optimal collaborative decision-making auxiliary instructions. Spatial Dimension Roles and Threshold Reduction Constraints: The consistency verification engine pre-stores a safety threshold red line matrix for different production conditions in its local register. Before the decision instructions generated by the agent cluster are output, the verification engine forcibly imposes hard textual constraints on the control parameters in the instructions. If the values ​​generated by the large model inference exceed the safety boundaries of the expert rules, the engine performs a reverse forced hedging, locking and rewriting them as the safety red line critical value.

[0065] Adversarial consistency verification in the time dimension: When the same type of industrial production problem in the same workshop is triggered again at different time periods or by different operators, the consistency verification engine will perform cross-temporal adversarial comparison between the currently generated inference logic chain and the historical similar audit logs stored in the security module audit unit library.

[0066] Logical closed-loop reset and safe release: If the verification engine calculates that the deviation between the current generated decision logic direction and the historical standard handling link exceeds the preset mutual trust threshold, the system determines that the model is hallucinating, immediately starts the logical reset, and forcibly calls the pre-stored expert deterministic benchmark workflow to correct and replace the large model text, thereby ensuring that the control instructions output by the auxiliary decision system have absolute constant consistency in time and operator dimensions.

[0067] The aforementioned flow control unit specifically includes: The specific workflow of the flow control unit within the workflow module, the high-concurrency rate control mechanism, and the technical implementation process of cluster horizontal scaling are as follows: High-frequency peak request streaming token rate control and dynamic thread allocation: After receiving structured control command data, the flow control unit initiates the underlying real-time peak shaving and rate control process: Dynamic priority queue reconstruction: The token unit inside the flow control unit first extracts the expected response latency from each concurrent request. Based on the urgency of the latency, the token unit dynamically pushes concurrent large model inference requests into different hardware execution queues, reconstructing a streaming buffer with latency gradients.

[0068] Hardware-level dynamic adjustment of the computing thread pool: The token unit monitors the hardware core temperature, memory usage, and computation time level of the CPU and GPU of the edge computing all-in-one machine in real time. Based on the priority weights of different queues, the flow control unit stream-rewrites the thread allocation registers of the underlying computing kernel, dynamically expanding the parallel computing thread pool for high-priority tasks and compressing the computing bandwidth of low-latency sensitive tasks to prevent hardware deadlock caused by high concurrency.

[0069] Streaming Token Generation Rate Control: During the lifecycle of local inference text generation in the large language model, the token unit adopts a streaming token rate control distribution mechanism. By dynamically adjusting the token injection rate in the token bucket algorithm, the character throughput of the large model outputting to specific interactive terminals per second is controlled. While ensuring second-level response, it adaptively addresses the instantaneous impact of conflict-triggered traffic on the all-in-one machine's high-speed cache.

[0070] Multi-machine horizontal flexible expansion and dynamic load balancing mechanism for hardware computing power: When the concurrent large-scale analysis demand across the entire plant exceeds the computing power safety limit of a single integrated machine, the expansion unit inside the flow control unit is activated, initiating cross-hardware distributed load balancing scheduling. Multi-machine cascaded network status polling: The expansion unit polls the remaining available computing power bandwidth and task queuing latency of other local all-in-one hardware devices deployed in the factory through the local private network at high frequency, and maintains a global computing power topology matrix in real time.

[0071] Seamless horizontal task splitting and elastic task distribution: Once the load of a single main control unit exceeds the preset overload threshold, the expansion unit immediately initiates distributed task distribution control. Without interrupting the current computing link, the expansion unit seamlessly splits and packages concurrent batch analysis and peak tasks exceeding the single machine's capacity into independent bus data streams, which are then dynamically distributed to other cascaded idle all-in-one hardware via gigabit industrial Ethernet, achieving zero-threshold flexible expansion of hardware computing power at the physical layer.

[0072] Multi-machine alignment of heterogeneous model scheduling: During flexible expansion of computing across multiple devices, the expansion unit maintains hard synchronization with the agent module scheduling unit, coordinating heterogeneous model instances deployed on different all-in-one machines to ensure that the intermediate text results of multi-agent computing after splitting can be back-converged with a unified data structure.

[0073] High-throughput performance assurance and data flow transmission: Through dynamic thread distribution of the token unit and horizontal scaling across multiple machines by the expansion unit, the flow control unit ensures that the overall industrial large-scale model inference computing throughput of the entire plant is stably maintained at a high-performance range of no less than 500 tokens per second ($\ge500\text{tokens / s}$) when the system encounters peak task processing, strictly limiting the overall system decision response latency under high concurrency to within seconds. After completing processing, the flow control unit streams the decision instructions to the permission unit and audit unit within the security module.

[0074] The specific contents of the aforementioned authorization units and audit units include: The three-tiered dynamic control mechanism for the role, space, and function of the permission unit: The permission unit is positioned at the boundary defense end of data flow and decision output, serving as the highest-priority interception and access control barrier of the system. When the flow control unit sends collaborative decision-making assistance instructions, the permission unit performs hard-core level over-authority interception verification through the following steps: Multi-dimensional feature credential extraction: The access control operator of the permission unit first captures the globally unique identity of the current request initiator (operator, terminal workshop equipment, or digital employee intelligent agent initiating cross-domain call), and reads the security feature credential carried in the identity online.

[0075] Cross-validation of the three-level permission matrix: The permission unit will seamlessly integrate the extracted security feature credentials into the system's built-in "role-space-function" three-level dynamic control matrix for three-dimensional cross-alignment comparison. First-level role permission verification: Verify whether the initiator has the core master control permission to access the current process big model, based on their job level or equipment security level. Second-order data space permission verification: Performs precise isolation verification of sensitive industrial asset spaces involved in large model decisions or requests to ensure that the initiator's access boundaries do not exceed the scope of permissions across spaces; Third-level function permission verification: Further verify whether the initiator has the permission to execute the current specific function (e.g., ordinary operators only have the function of viewing BI dashboards, while high-risk control functions such as circuit breaking, rewriting, and agent skill modification are only open to the system administrator role).

[0076] Adaptive interception and secure access control: If any of the above-mentioned level-one verifications fails, the permission unit immediately initiates the highest-priority blocking, forcibly suspends the current decision output, and throws an unauthorized alarm to the terminal to prevent the risk of leakage of core sensitive data; if all three levels of permission verifications pass, the permission unit releases the blocking state and streams the collaborative decision-making assistance instruction to the audit unit inside the security module.

[0077] The audit unit's full lifecycle full traceability and data link tracking mechanism: After receiving the compliance decision instruction from the authorized unit, the audit unit initiates a highly reliable and trustworthy AI audit and full traceability process within the final time window before the decision support results are finally output to the workshop execution mechanism. High-frequency, full-data capture of full operational domain features: The audit unit's built-in hardware-level full-data operation audit engine records all underlying behaviors that occur throughout the decision-making lifecycle. Specific capture objects include, but are not limited to: data access behavior of each sensor within the data module, interface call frequency and prompt text of the local and cloud-based large model, and Agent parameter and skill modification records that occur during collaborative debate within the agent module.

[0078] Data Link Fingerprint Binding: To ensure industrial-grade interpretability for every core conclusion generated by the large model, the audit unit initiates a reverse link tracing algorithm. The algorithm automatically correlates the final control conclusions derived by each agent with the most original data source accessed in the shared blackboard data area, creating a causal graph and calculating and binding a unique data link fingerprint log for each complete inference logic chain. This fingerprint allows factory management to conduct traceable causal investigations at any time.

[0079] Long-term log retention and streaming secure release closed loop: After confirming that the full audit is error-free and there is no risk of sensitive information leakage, the audit unit writes the complete audit log, including operation traces, model call library (trace) trajectories, and data link fingerprints, into the system's built-in, tamper-proof, private long-term log retention library. After writing, the audit unit completely releases its control lock on the bus, and the master control scheduling instructions seamlessly drive the all-in-one machine's network security gateway. The final intelligent decision-making solution, after trusted arbitration and compliance audit, is then streamed securely output to the industrial workshop operation screen, AGV scheduling center, or edge actuator with a second-level response, thus ultimately completing the adaptive control flow closed loop of the entire decision support system across all units.

[0080] (Example 2) The core mixing equipment on Line A of the flexible manufacturing workshop suddenly stopped due to overheating and jamming of the main bearing. The on-site operator input the following request into the all-in-one machine via voice terminal: "The mixing equipment on Line A has suddenly overheated and stopped. It is estimated that it will take 3 hours to repair. Please provide repair instructions and readjust today's production schedule." The workflow module's generation unit performed beamforming denoising, ASR vocabulary decoding, and called the syntactic tree dependency analysis operator to accurately identify and parse the complex long sentence into a multi-instruction composite intent data packet that includes two interdependent instructions: "Operation and Maintenance Control" and "Production Scheduling".

[0081] Sampling hard switching and establishment of a shared data base: The generation unit instantly wakes up the acquisition unit of the data module. The acquisition unit sends a full-speed sampling command to the plant backbone bus, controlling each channel to switch from suspended monitoring mode to full-speed high-frequency sampling state, and simultaneously captures the current and bearing continuous temperature field data before the mixed-flow equipment shutdown, the list of incomplete orders in the MES system, and the weight of today's delivery options and material inventory balance in the ERP system. Immediately afterwards, the fusion unit uses the Model Context Protocol (MCP) to retrieve the standard maintenance guide and expert knowledge text corresponding to this model of equipment from the private high-dimensional vector database, automatically translates the numerical bearing temperature data into human declarative sentences, incrementally concatenates and encapsulates it with the retrieved maintenance text, and streams it into the shared memory of the internal collaborative unit of the intelligent agent module, establishing a unified shared blackboard data base.

[0082] Agent Debate and Conflict Arbitration: Within the collaborative unit, parallel "digital employee" roles are simultaneously activated, operating based on a local inference engine (deployed with Qwen / DeepSeek model instances) without requiring manual coding. The scheduling agent first calculates and outputs a draft scheduling adjustment based on the new context: Option 1 diverts orders from line A to line B, which has redundant capacity; Option 2 diverts them to line C. At this point, the scheduling unit inputs this draft as a new independent variable to other agents via a shared blackboard bus. After verification, the inventory agent sends a blocking signal indicating insufficient inventory of specific raw materials on line B, and the finance agent also reports that the excessively long turnaround path to line B results in excessive logistics costs. The scheduling unit uses a multi-round feedback adjustment mechanism to push the gap and excess variables back to the scheduling agent for iterative re-evaluation. In the secondary inference, the scheduling agent modifies the computing power weights, ultimately converging to the optimal solution: 70% of urgent orders are diverted to line C, and the remaining 30% are suspended pending line A's recovery. The scheduling unit then executes conflict arbitration through multi-indicator quantified benefit projection. Meanwhile, the consistency verification engine inside the scheduling unit performs security reduction and countermeasure constraints on the output parameters to eliminate model random illusions.

[0083] Peak flow control and security audit output closed loop: The scheduling unit transfers the determined decision plan to the flow control unit. During peak concurrency periods coinciding with shift changes (shift handover) across the entire plant, the flow control unit's token unit reconstructs the dynamic priority queue, adjusts the computing kernel thread pool, and employs streaming token rate control distribution. Simultaneously, the extension unit polls the global computing power topology matrix, horizontally distributing excess batch analysis tasks to other cascaded integrated hardware devices, ensuring that the plant's inference throughput rate remains consistently within the high-performance range.

[0084] Before the final decision is output, the security module's permission unit verifies and intercepts unauthorized access against the three-level permission matrix to ensure that the current shift's operators' access boundaries are compliant. After successful verification, the audit unit captures the full operational characteristics of the entire decision-making interaction cycle and uses a reverse link tracing algorithm to bind the final production scheduling and maintenance conclusions with the original bearing overheating sensor data and ERP bill of materials data, generating a unique data link fingerprint log. This achieves causal retrospective, reliable, full-volume audit traceability and storage. Finally, the system completely unlocks and streams the precise decision (streaming the mixed-flow equipment main bearing overheating troubleshooting steps to the on-site operator's tablet, while simultaneously pushing the reconstructed daily production schedule dashboard to the workshop's central control screen) to the terminal with a second-level response time, perfectly realizing a closed-loop intelligent decision-making assistance system for human-machine collaboration.

[0085] Obviously, the above embodiments are merely examples to clearly illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all embodiments here. However, these obvious variations or modifications derived from the spirit of the present invention are still within the protection scope of the present invention.

Claims

1. A real-time decision support system for large industrial models, characterized in that, It includes a data module, a workflow module, an intelligent agent module, and a security module; The data module is used to acquire multi-source heterogeneous industrial data and reconstruct contextual knowledge, including an acquisition unit and a fusion unit; The acquisition unit is used to capture the dynamic production feature flow of the entire plant area in real time and concurrently via the bus; The fusion unit integrates the retrieval enhancement generation architecture and the model context protocol to align the retrieved expert knowledge with real-time production dynamic data and reconstruct and generate a feature semantic enhancement dataset that includes complete industrial operating conditions. The workflow module is used to compile and convert natural language commands for human-machine interaction into deterministic links for industrial control and to control the speed of high concurrency. It includes a generation unit and a flow control unit. The generation unit has a built-in semantic parsing operator, which is used to automatically identify business intentions and control variables, and automatically reconstruct and generate industrial automation workflows; The flow control unit is configured with a token distributor and a physical computing thread pool dynamic adjuster to perform token flow control distribution and dynamic load balancing among multiple hardware components during high-frequency processing periods. The intelligent agent module is used to construct a heterogeneous digital employee cluster and perform conflict arbitration and consistency locking for cross-departmental multi-objective decision-making, including a collaboration unit and a scheduling unit; The collaborative unit is used to manage and run multiple heterogeneous, dedicated distributed intelligent agents for specific industrial positions in parallel. The scheduling unit acts as a shared communication bus to drive each agent to perform intent alignment and data interaction, performs conflict arbitration on the one-way decision results fed back by each agent, merges them into a globally optimal collaborative decision-making auxiliary instruction, and has a built-in logical consistency verification engine to perform reduction constraints on the output parameters through adversarial verification. The security module is used to build the system's security red line defense mechanism, including an access control unit and an auditing unit; The permission unit has a built-in three-level dynamic control matrix of roles, spaces, and functions, which is used to implement fine-grained data space isolation and access security control; The audit unit is equipped with a hardware-level full-operation audit engine, which is used to fully record the underlying behavior of model inference and bind data link fingerprint logs to the final decision conclusion.

2. A real-time decision support method for an industrial large-scale model of the system as described in claim 1, characterized in that, Includes the following steps: S1. Capture spoken natural language speech stream through the workflow module, and parse and extract the composite intent containing the core business intent and control target variables; S2. The data module sends a full-speed sampling command to the perception node based on the composite intent to hard switch to high-frequency sampling state, and simultaneously collects cross-domain multimodal production feature data streams, and combines retrieval enhancement generation with expert knowledge fusion to encapsulate it into a feature semantic enhancement dataset. S3. Activate multiple cross-departmental heterogeneous dedicated distributed intelligent agents through the intelligent agent module, carry out multiple rounds of intent alignment and data interaction based on the shared memory base, and perform arbitration through the conflict judgment matrix operator, combined with logical consistency verification to lock and generate the optimal collaborative decision-making auxiliary instruction. S4. The workflow module's flow control unit uses a streaming word distribution mechanism to perform hardware-level high-concurrency speed control, and seamlessly splits tasks to cascaded heterogeneous devices for distributed load balancing when encountering peak conditions. S5. By comparing the security module with the three-level control matrix, unauthorized interception and verification are performed, full-scale operation behavior is captured and data link fingerprint is bound for compliance hard audit, and finally the streaming security output of the digital decision-making solution is completed.

3. The real-time decision support method for large industrial models according to claim 2, characterized in that, In step S3, when the agent module processes content that is outside its assigned role and goes beyond its boundaries, it specifically performs the following secondary adaptive fault-tolerant sub-step: S31. The dedicated intelligent agent calculates the semantic similarity dot product between the instruction text and the ontology vocabulary through the local domain classifier, and outputs the local domain confidence score. S32. If the confidence level is lower than the preset threshold or the reverse constraint feature word is hit, the local inference chain of the large model is forcibly broken, and the instruction is repackaged into an out-of-bounds data packet and marked as an abnormal suspension by the abnormal feature operator. S33. By reversing the out-of-bounds data packet back through the shared communication bus inside the intelligent agent module, the global intent routing engine in the scheduling unit is immediately awakened. S34. The global intent routing engine performs multi-dimensional correlation comparison of out-of-bounds instructions in the global intent mapping matrix to re-identify the true target's domain, rewrites the routing header label, clears historical confidence, accurately directs and transfers the instructions, and sends them to the input queue of the target-adaptive agent.

4. The real-time decision support system for large industrial models according to claim 1, characterized in that, The workflow module's generation unit specifically includes: The system captures spoken natural language speech streams at high frequencies, uses acoustic beamforming algorithms for pre-stage denoising, and filters out noise using bandpass hardware filters and adaptive spectral subtraction algorithms. The clean audio stream is then input into an automated speech recognition acoustic model fine-tuned with a built-in industrial corpus, decoded, and converted into a standard text character stream. A sequence labeling algorithm is run to perform named entity recognition, extracting spatial and asset entity slots, as well as state and constraint feature slots online. A syntactic tree dependency analysis algorithm is used to parse the grammatical dominance relationships between words, segmenting long sentences into clause structures with independent control intentions. By comparing with a preset industrial action mapping matrix, the system identifies the core verbs and object mappings of each clause as target intentions, constructs topological dependency chains based on sequential logic, and unifies heterogeneous multi-target intentions into a unified composite intention.

5. The industrial large-scale model real-time decision support system according to claim 4, characterized in that, The data acquisition unit of the data module specifically includes: When the generation unit identifies and reconstructs a composite intent or an alarm signal triggered by an abnormal operating condition by the field equipment, the acquisition unit, as the master control node, sends a full-speed sampling command to each sensing node, production line control unit and business system interface, and controls each heterogeneous data channel to hard switch from low-frequency monitoring mode to full-speed high-frequency sampling state; and simultaneously captures field physical sensing feature data, production line status and micro-physical attitude data, management and operation and external scene context data. The local microcontroller unit forcibly marks the concurrently received data streams with a globally unified timestamp distributed by the system crystal oscillator, calculates the time difference in a rolling manner, and encapsulates signals with time differences less than or equal to a preset synchronization time deviation threshold through cross-modal multi-dimensional matrix alignment, generating an aligned dataset that is streamed to the fusion unit.

6. The real-time decision support system for large industrial models according to claim 5, characterized in that, The data module's fusion unit specifically includes: The parsing operator extracts the physical identifier and instantaneous anomaly feature vector from the aligned dataset as a high-dimensional query key. Nearest neighbor similarity matching is calculated in a local private high-dimensional vector database, retrieving the expert rule text with the highest matching degree and historical best-case scenario data. Cross-protocol format conversion is performed using the model context protocol. Through a preset industrial semantic mapping template, the numerical dynamic production data returned from each channel is automatically mapped into working condition statements to eliminate heterogeneous source protocol barriers. The generated working condition statements are incrementally concatenated with the retrieved expert knowledge text. Semantic embedding overlay technology is used to perform deep fusion at the text semantic level. A system-preset global task prompt shell is added to the front-end and back-end to package and reconstruct a feature semantic enhancement dataset, which is then streamed into the shared memory of the intelligent agent module's collaborative unit.

7. The real-time decision support system for large industrial models according to claim 6, characterized in that, The collaborative unit includes parallel scheduling agents, inventory agents, financial agents, and operations and maintenance agents. The operations and maintenance agent, after receiving the feature semantic enhancement dataset, employs a closed-loop control mechanism for the re-identification and transfer of out-of-bounds content, specifically including: Each dedicated agent calls the local domain classifier, performs semantic similarity dot product calculation between the received instruction text and the local domain ontology vocabulary, outputs the local domain confidence score, and verifies whether the instruction contains explicit inverse constraint feature words. If the local domain confidence is lower than the preset safety threshold or the reverse constraint feature word is hit, the local inference chain of the current agent is forcibly broken, the instruction is marked as an out-of-bounds exception and suspended, and the exception feature operator is used to repackage and combine it into a standard out-of-bounds data packet. By reversing out-of-bounds data packets through the shared communication bus, the global intent routing engine inside the scheduling unit is activated. The global intent routing engine performs multi-dimensional correlation comparison on the original instruction in the global intent mapping matrix to re-identify the real target's domain, rewrites the routing header label and clears the local confidence verification history, accurately directs and transfers the instruction to the target-adaptive agent's pending input queue.

8. The industrial large-scale model real-time decision support system according to claim 7, characterized in that, The scheduling unit of the intelligent agent module specifically includes: A dynamically updated shared blackboard data area is established within a private secure computing domain. The bus listener frequently polls the one-way decision drafts generated by each agent and writes them into this area. By using the conflict determination matrix operator for rolling monitoring, when a multi-target deadlock conflict event is triggered, the output link of the decision scheme is immediately cut off. The reverse feedback mechanism is called to repackage the blocking text and the correction opinion as strong constraint feedback variables into high-priority prompt word increments, which are then thrown back to the source agent to drive secondary inference to modify the control parameters. When the parameter conflict mutation rate is continuously lower than the convergence threshold, the optimal collaborative decision auxiliary instruction that takes into account the overall situation is selected through multi-index weighted logic. The internal logical consistency verification engine extracts the control parameters in the instruction and performs spatial dimension reduction constraints on the safety threshold red line matrix pre-stored in the register. If the boundary is exceeded, it reverses and forcibly offsets and rewrites the lock to the safety red line critical value. The current inference logic chain is compared with historical similar audit logs across time and space. If the morphological deviation exceeds the preset mutual trust threshold, the model is determined to be hallucinating and the logic is reset. The pre-stored baseline workflow is forcibly invoked to perform correction and replacement.

9. The real-time decision support system for large industrial models according to claim 1, characterized in that, The workflow module specifically includes: The internal token unit extracts the expected response latency from each concurrent request and pushes the requests into different hardware execution queues according to the urgency of the latency, thus reconstructing a streaming buffer with latency gradients. Real-time monitoring of hardware core temperature, memory usage, and computation time levels; dynamic rewriting of the underlying computing kernel's thread allocation registers; expansion of the parallel computing thread pool for high-priority tasks. A streaming word rate control and distribution mechanism is adopted during the text generation lifecycle to dynamically adjust the token injection rate of the word bucket algorithm to control the character throughput per second. When demand exceeds the safety limit of single-machine computing power, the internal expansion unit polls the remaining available computing power and bandwidth of other integrated hardware devices in the factory, seamlessly splits and packages the analysis tasks that exceed the capacity into independent data streams, and dynamically distributes them to other cascaded idle hardware for physical elastic distribution via industrial Ethernet.

10. The real-time decision support system for large industrial models according to claim 1, characterized in that, The permission unit and audit unit within the security module specifically include: The access control operator of the permission unit captures the globally unique identity of the current request initiator and compares it with the three-level dynamic control matrix of the system, which includes role permissions, data space permissions and function permissions. If any level fails, the highest priority blocking and suspension decision is initiated. The audit unit's hardware-level full-scale operation audit engine records the execution of underlying behaviors throughout the entire lifecycle, including data access, model interface calls, and agent parameter modification. The reverse link tracing algorithm is activated to correlate the control conclusions derived by each agent with the original data source it calls using a causal graph. A unique data link fingerprint log is calculated and bound and written into an immutable private long-term storage library. Finally, the network security gateway is driven to provide streaming security output to the execution mechanism.

Images