Industrial large model adaptation method and device based on cross-mode linkage and comparative learning

Abstract

The invention provides an industrial large model adaptation method and device based on cross-mode linkage and comparative learning, and relates to the field of data processing. According to the method, historical operation data and production line working conditions in the industrial field are obtained; according to the historical operation data and the working condition of the production line, a digital twinning environment of the industrial large model is constructed, and the digital twinning environment comprises a first agent, a second agent and a third agent; injecting a mixed fault signal into the digital twin environment through the first agent; performing fault type identification on the mixed fault signal through a second intelligent agent to obtain an identification result; performing comparative learning on the recognition result through a third agent to obtain a chess playing result; and dynamically adjusting the type and intensity of the digital twin environment based on the chess playing result, generating an adaptation strategy, and adjusting the industrial large model according to the adaptation strategy. By implementing the technical scheme provided by the invention, the adaptation accuracy of the industrial large model can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and in particular to a method and device for industrial large-scale model adaptation based on cross-model joint construction and comparative learning. Background Art

[0002] With the deepening of intelligent manufacturing, intelligent monitoring and diagnosis methods based on large-scale pre-trained models have received widespread attention in the fields of fault warning, predictive maintenance, quality inspection, etc.

[0003] However, existing methods primarily rely on single or limited modal data collected offline, along with fixed model architectures and training processes. These methods struggle to cope with the distribution drift caused by complex and changing production environments and equipment evolution. Because traditional industrial large-scale model adaptation is often based on centralized, batch-labeled data, the model's ability to adapt to new operating conditions after launch is severely limited. With the accumulation of factors such as equipment aging, process adjustments, and sensor replacement, the gap between offline training models and real-time production data continues to widen, resulting in lower adaptation accuracy for industrial large-scale models.

[0004] Therefore, there is an urgent need for industrial large-scale model adaptation methods and devices based on cross-model joint construction and comparative learning. Summary of the Invention

[0005] This application provides an industrial large model adaptation method and device based on cross-model joint construction and comparative learning, which is convenient for improving the adaptation accuracy of industrial large models.

[0006] In the first aspect of the present application, an industrial large model adaptation method based on cross-model joint construction and comparative learning is provided, the method comprising: obtaining historical operating data and production line conditions in the industrial field; constructing a digital twin environment of the industrial large model based on the historical operating data and the production line conditions, the digital twin environment comprising a first intelligent agent, a second intelligent agent and a third intelligent agent; injecting a mixed fault signal into the digital twin environment through the first intelligent agent; identifying the fault type of the mixed fault signal through the second intelligent agent to obtain an identification result; performing comparative learning on the identification result through the third intelligent agent to obtain a game result; based on the game result, dynamically adjusting the type and strength of the digital twin environment, generating an adaptation strategy, and adjusting the industrial large model according to the adaptation strategy.

[0007] By employing the above technical solution, a virtual agent system consisting of a first agent, a second agent, and a third agent is constructed. This system can simulate various mixed fault signals in a simulation environment and continuously optimize model performance based on a real-time feedback mechanism. In particular, the first agent's active injection of mixed faults enhances the "adversarial" and "diverse" nature of the training data, thereby improving the robustness and fault differentiation capabilities of large models in complex real-world scenarios. This method integrates a comparative learning strategy with evolutionary game theory, transforming the model tuning process into a continuous game and adaptation process among multiple agents. This achieves a closed-loop, self-driven process from data generation, model training, to performance evaluation. The third agent, using a comparative learning mechanism, deeply explores the differences in the model's performance between similar faults, effectively exposing model weaknesses and driving adjustments to fault types and injection intensity within the digital twin environment, thereby fostering more targeted training strategies. Ultimately, through the continuously evolved adaptation strategy, targeted fine-tuning of the large industrial model is achieved, thereby improving its adaptation accuracy.

[0008] Optionally, the obtaining of historical operating data and production line operating conditions in the industrial field specifically includes: system docking with PLC, DCS and MES through OPCUA, MQTT and RESTful API, and real-time acquisition of temperature, pressure, vibration, acoustics, energy consumption, work orders and operation logs of industrial equipment in the industrial field to obtain first modal data; obtaining the sensor position, model, sampling frequency and maintenance record of the production line where the industrial equipment is located to obtain second modal data; denoising, interpolating and aligning the first modal data and the second modal data to obtain the historical operating data and the production line operating conditions.

[0009] By adopting this technical solution, seamless integration with key systems such as PLCs, DCSs, and MESs is achieved through standard communication protocols such as OPC UA, MQTT, and RESTful APIs. This not only ensures high-frequency, low-latency acquisition of key operating parameters such as temperature, pressure, vibration, acoustics, and energy consumption at the industrial site, but also fully integrates unstructured information such as work orders and operation logs to form primary modal data covering equipment operating status and human-machine interaction behaviors. Further contextual information, such as the spatial configuration of equipment on the production line, sensor model and sampling parameters, and historical maintenance records, is collected to construct secondary modal data, enhancing the data's structure and semantic integrity.

[0010] Optionally, constructing a digital twin environment of the industrial large model based on the historical operating data and the production line operating conditions specifically includes: using a three-dimensional simulation engine to replicate the workshop layout, machine geometry and sensor installation position based on the historical operating data and the production line operating conditions; loading historical sensor curves and operating condition parameters into the initial digital twin model, and driving the initial digital twin model to operate through kinematic and thermodynamic models to obtain the digital twin environment of the industrial large model.

[0011] By adopting the above technical solutions and building a highly restored industrial digital twin environment, a highly realistic, dynamically controllable virtual test platform is provided for the intelligent agent training and adaptive evolution of large industrial models, which significantly improves the efficiency of model development and the adaptability of actual deployment.

[0012] Optionally, injecting a mixed fault signal into the digital twin environment through the first intelligent agent specifically includes: determining a multimodal simulation feature based on the historical operating data and the production line operating conditions; randomly extracting at least two simulation features from the multimodal simulation feature through the first intelligent agent, and synthesizing a mixed fault signal according to a time series; and adding the mixed fault signal to the digital twin environment through the first intelligent agent according to the time of occurrence, intensity of occurrence, and duration of the fault.

[0013] By adopting the above technical solution, first, based on historical operating data and operating parameters, multimodal simulation features are systematically extracted, including but not limited to vibration spectra, temperature anomaly curves, acoustic mutations, energy consumption transitions, etc. These features cover multi-dimensional and multi-level equipment status performances and have strong industrial scenario orientation.

[0014] Optionally, the second intelligent agent is used to identify the fault type of the mixed fault signal to obtain an identification result, which specifically includes: using the second intelligent agent, using pre-trained visual, vibration and acoustic coding units to map the mixed fault signal into an initial feature vector; using an adaptive attention mechanism to dynamically weight the initial feature vector to form a fused feature vector; performing classification inference on the fused feature vector, and outputting a fault category label and confidence to obtain the identification result.

[0015] By adopting the above technical solution, the second intelligent agent is used to intelligently identify mixed fault signals, and a fault identification framework with multimodal perception and dynamic feature adaptation capabilities is constructed, which significantly improves the diagnostic accuracy and discrimination robustness of large industrial models under complex working conditions.

[0016] Optionally, the third intelligent agent is used to perform comparative learning on the recognition results to obtain a game result, which specifically includes: controlling the third intelligent agent to extract target samples that do not meet the preset fault category in the recognition results from the same model evolution node as positive samples; controlling the third intelligent agent to extract target samples that meet the preset fault category in the recognition results from the same model evolution node as negative samples; determining the first fusion feature vector corresponding to the positive sample and the second fusion feature vector corresponding to the negative sample; and generating the game result according to the confidence of the positive sample and the negative sample at the same model evolution node.

[0017] By adopting the above technical solution, the third agent first actively identifies samples that do not match the preset fault category from the same model evolution node based on the current recognition results as "positive samples", and samples that match the preset category but have low confidence or blurred boundaries as "negative samples", forming comparative sample pairs. This sample selection method is not derived from external annotations, but is combined with adaptive judgment under the model evolution state, reflecting strong scenario relevance and training targeting. Subsequently, by extracting the fused feature vectors of positive and negative samples and comparing and analyzing them in combination with their confidence in the evolution node, it is possible to quantify the model's current recognition bias, modal imbalance, or attention loss in fault type differentiation, thereby generating game results as a basis for optimization.

[0018] Optionally, based on the game result, the type and strength of the digital twin environment are dynamically adjusted, an adaptation strategy is generated, and the industrial large model is adjusted according to the adaptation strategy, specifically including: obtaining abnormal modes and abnormal parameters according to the game result; inputting the abnormal modes and abnormal parameters into the online fine-tuning framework to perform lightweight weight updates and attention distribution retraining to obtain the adaptation strategy; based on the environment configuration, fine-tuning log and performance indicators in the adaptation strategy, the industrial large model is adapted and adjusted.

[0019] By employing the aforementioned technical solution, this method first uses game results to identify abnormal modes (such as acoustic interference or vibration misjudgment) and abnormal parameters (such as frequency distortion and amplitude drift) involved in the current model's recognition deviations. This accurately locates weak links in the model's perception and understanding, thereby enabling "problem-driven" fine-tuning input. This abnormal information is then fed into the online fine-tuning framework, which rapidly corrects the model through lightweight weight updates and attention distribution retraining. This avoids the computational resource consumption and convergence bottlenecks of traditional retraining while preserving the structural advantages and stability of the original model. Finally, based on comprehensive information generated by the adaptation strategy, including the environmental configuration (such as the twin environment perturbation factor), fine-tuning logs (such as the number of iterations and convergence rate), and performance metrics (such as accuracy and response latency), targeted adaptation adjustments are made to the large industrial model.

[0020] In the second aspect of the present application, an industrial large model adaptation device based on cross-model joint construction and comparative learning is provided, the device including an acquisition module and a processing module, wherein the acquisition module is used to acquire historical operating data and production line conditions in the industrial field; the processing module is used to construct a digital twin environment of the industrial large model based on the historical operating data and the production line conditions, and the digital twin environment includes a first intelligent agent, a second intelligent agent and a third intelligent agent; the processing module is also used to inject mixed fault signals into the digital twin environment through the first intelligent agent; the processing module is also used to identify the fault type of the mixed fault signal through the second intelligent agent to obtain an identification result; the processing module is also used to perform comparative learning on the identification result through the third intelligent agent to obtain a game result; the processing module is also used to dynamically adjust the type and strength of the digital twin environment based on the game result, generate an adaptation strategy, and adjust the industrial large model according to the adaptation strategy.

[0021] In a third aspect of the present application, an electronic device is provided, which includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs the method described above.

[0022] In a fourth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores instructions. When the instructions are executed, the method described above is executed.

[0023] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: By constructing a virtual agent system consisting of a first agent, a second agent, and a third agent, it is possible to simulate various mixed fault signals in a simulation environment and continuously optimize model performance based on a real-time feedback mechanism. In particular, the first agent's active injection of mixed faults makes the training data more "adversarial" and "diverse," thereby improving the robustness and fault differentiation capabilities of large models in complex real-world scenarios. This method integrates a comparative learning strategy with evolutionary game theory, transforming the model tuning process into a continuous game and adaptation process among multiple agents. This achieves a closed-loop, self-driven process from data generation, model training, to performance evaluation. The third agent, using a comparative learning mechanism, deeply explores the differences in the model's performance between similar faults, effectively exposing model weaknesses and driving adjustments to fault types and injection intensity within the digital twin environment, thereby fostering more targeted training strategies. Ultimately, through the continuously evolved adaptation strategy, targeted fine-tuning of the large industrial model is achieved, thereby improving the adaptation accuracy of the large industrial model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flowchart of an industrial large-scale model adaptation method based on cross-model joint construction and comparative learning provided in an embodiment of the present application; Figure 2 Another flowchart of the industrial large-scale model adaptation method based on cross-model joint construction and comparative learning provided in an embodiment of the present application; Figure 3 A schematic diagram of a module of an industrial large-scale model adaptation device based on cross-model joint construction and comparative learning provided in an embodiment of the present application; Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0025] Explanation of the reference numerals: 31, acquisition module; 32, processing module; 41, processor; 42, communication bus; 43, user interface; 44, network interface; 45, memory. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0027] In the description of the embodiments of this application, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a concrete manner.

[0028] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0029] With the continuous deepening of intelligent manufacturing, intelligent monitoring and diagnosis technologies based on large-scale pre-trained models are playing an increasingly important role in key links such as fault warning, predictive maintenance, and quality control.

[0030] Despite this, current mainstream methods often rely on single or small amounts of modal data collected offline, using fixed model architectures and static training processes, making it difficult to cope with the distribution drift problems caused by dynamic changes in working conditions and continuous evolution of equipment in real production environments. Especially after model deployment, traditional adaptation methods are highly dependent on centralized batch-labeled data, resulting in weak adaptability of the model when faced with new working conditions, new equipment, or sudden conditions. Over time, the accumulation of factors such as equipment aging, process parameter adjustments, and sensor replacements makes the deviation between the original training data and real-time production data more obvious, ultimately resulting in a significant decrease in the adaptation accuracy of large industrial models in actual scenarios, making it difficult to meet the actual needs of intelligent manufacturing for high reliability and high adaptability.

[0031] In order to solve the above technical problems, this application provides an industrial large model adaptation method based on cross-model joint construction and contrastive learning, referring to Figure 1 , Figure 1 This is a flow chart of an industrial large-scale model adaptation method based on cross-model joint construction and comparative learning provided in an embodiment of the present application. The method is applied to a server and includes steps S110 to S160, which are as follows: S110. Obtain historical operating data and production line conditions in the industrial field.

[0032] Specifically, the server will send data requests to programmable logic controllers (PLCs), distributed control systems (DCSs), manufacturing execution systems (MESs), etc. through standard communication protocols (such as OPC UA, MQTT, RESTful API, etc.) on a regular or real-time basis to obtain temperature, pressure, vibration, speed, energy consumption, operating status, and operation logs such as each maintenance, material change, and alarm, thereby fully collecting the "operation history" of the equipment and the "working condition information" of the production line.

[0033] Taking an automobile engine assembly line as an example, the backend server first connects to the assembly line's PLC via OPC UA to obtain real-time information on key physical quantities such as spindle torque and assembly pressure. Simultaneously, it connects to the MES system via a RESTful API to download work order numbers, operator records, and inspection results for each batch of engines. The collected data is first stored in a time-series database and centrally managed, along with sensor location, model, and calibration records provided by field engineers. This allows the server to instantly review and replay operational trajectories and operating condition changes over the past several months or even years, providing a solid data foundation for subsequent fault analysis, digital twin modeling, and intelligent diagnosis.

[0034] In one possible implementation, historical operating data and production line conditions in the industrial field are obtained, specifically including: system docking with PLC, DCS, and MES through OPC UA, MQTT, and RESTful API to obtain real-time temperature, pressure, vibration, acoustics, energy consumption, work orders, and operation logs of industrial equipment in the industrial field to obtain first modal data; obtaining the sensor location, model, sampling frequency, and maintenance records of the production line where the industrial equipment is located to obtain second modal data; and denoising, interpolating, and aligning the first and second modal data to obtain historical operating data and production line conditions.

[0035] Specifically, the server seamlessly connects with the programmable logic controllers (PLCs), distributed control systems (DCSs), and manufacturing execution systems (MESs) within the shop floor via three mainstream industrial communication protocols: OPC UA, MQTT, and RESTful APIs. This enables the system to pull real-time data on temperature, pressure, vibration, acoustic signals, and energy consumption, as well as work orders and operation logs generated for each maintenance or alarm event. This generates so-called "first modality data," which comprehensively reflects the dynamic parameters and operational history of the equipment during operation.

[0036] Next, in order to make these sensor data more usable, the system will also collect additional information about the location, brand model, sampling frequency, and maintenance records of the sensors installed at each node on the production line. This is called "second modality data" and provides contextual information about the equipment and sensors. Subsequently, the first and second modality data are denoised, interpolated, and time-aligned to eliminate occasional noise interference, fill in missing data points, and ensure that records from different sources correspond to each other in the time dimension. For example, on a packaging production line, after obtaining the vibration curve and acoustic signal of the packaging machine in real time through MQTT, combined with the daily calibration log of the sensor, all data are uniformly mapped to the same time axis, which can obtain clean and complete historical operation records and production line conditions, laying a solid foundation for subsequent digital twin simulation and intelligent fault diagnosis.

[0037] S120. Build a digital twin environment of the industrial large model based on historical operating data and production line conditions. The digital twin environment includes a first intelligent agent, a second intelligent agent, and a third intelligent agent.

[0038] Specifically, the backend server uses collected historical operating data (such as temperature curves, vibration waveforms, and energy consumption curves) and production line operating information (such as sensor layout, equipment configuration, and maintenance records) to recreate a highly realistic "digital twin" factory scene in virtual space. This scene includes not only three-dimensional models of the equipment and physical motion patterns, but also simulated sensor signal streams. To enable the model to evolve and optimize itself, the system pre-installs three different intelligent agents in the twin environment: the first agent is responsible for actively injecting complex fault signals into the environment, the second agent is responsible for multimodal fault identification of the injected signals, and the third agent conducts comparative learning based on the identification results, thus forming an adversarial, self-driven optimization closed loop.

[0039] Taking an automobile engine assembly line as an example, the server first imports a year's worth of torque sensor, temperature sensor, and acoustic sensor data into a 3D assembly workshop model and dynamically drives the model based on actual operating conditions. Subsequently, on this virtual assembly line, the first agent simulates compound fault scenarios such as "insufficient lubrication + part misalignment." The second agent, using its built-in visual, vibration, and acoustic analysis modules, identifies the types of these fault signals and outputs diagnostic results. The third agent performs a self-supervised comparison of the recognition results with pre-set fault labels, continuously evaluating the recognition accuracy and providing feedback to the system. This facilitates online adjustment of the twin environment's fault injection strategy and model parameters, ultimately making the large industrial model more robust and adaptable on the real production line.

[0040] In one possible implementation, a digital twin environment of a large industrial model is constructed based on historical operating data and production line conditions, specifically including: using a three-dimensional simulation engine to replicate the workshop layout, machine geometry, and sensor installation positions based on historical operating data and production line conditions; loading historical sensor curves and operating condition parameters into an initial digital twin model, and driving the initial digital twin model to run through kinematic and thermodynamic models to obtain a digital twin environment of the large industrial model.

[0041] Specifically, the system first uses a 3D simulation engine (such as Unity or Unreal) to accurately recreate the overall layout of the workshop, including the plant structure, equipment spatial location, and conveyor line routing. Simultaneously, based on CAD or on-site survey data, it replicates the geometry of the mechanical equipment and mounts virtual sensors that correspond to their real counterparts in the corresponding locations. Next, previously collected sensor timing curves for temperature, vibration, acoustics, and other parameters, along with production line operating parameters (such as speed setting, load level, and process temperature), are loaded into this "initial model." Driven by physical simulation of equipment kinematics (such as shaft rotation and robot arm displacement) and thermodynamics (such as heat conduction and heat dissipation), the model not only visually reproduces equipment operation but also outputs dynamic curves that match historical data in real time on the virtual sensor side. This creates a highly realistic and controllable digital twin environment for large industrial models.

[0042] Taking an electronics assembly line as an example, the team first constructed the three assembly lines, two inspection stations, and surrounding auxiliary facilities of the assembly workshop in a 3D engine. 3D models of key equipment, such as the robotic arm and screw machine, were placed in place according to their actual dimensions. The team then imported the screw machine's torque fluctuation curve, vibration sensor data, and assembly temperature records from the past year into the corresponding virtual sensor nodes. Finally, the kinematics module simulated the robot arm's trajectory control, and the thermodynamics module reproduced the machine's temperature changes under high load. This enabled the entire digital twin environment to not only achieve nearly identical operating conditions to the actual production line but also flexibly incorporate new operating conditions or failure scenarios into subsequent agent training, providing a reliable, low-risk virtual testbed for the adaptive optimization and online evaluation of large industrial models.

[0043] S130. Injecting a mixed fault signal into the digital twin environment through the first intelligent agent.

[0044] Specifically, within the digital twin's virtual factory environment, the server drives the first agent (also known as a "challenger" or "fault injector") to proactively create scenarios combining multiple fault types and "inject" these synthesized fault signals into virtual sensors or simulation models to simulate complex faults that might occur on a real production line. During this process, the first agent randomly or strategically selects two or more fault modes based on multimodal fault characteristics extracted from historical operating data and production line conditions (such as sudden vibration surges, temperature rises, and abnormal acoustic signals). These are combined according to the order, intensity, and duration of the faults. These composite signals are then sent to the corresponding sensor channels of the digital twin environment through the underlying interface, allowing subsequent agents to experience the fault evolution process in a safe and controllable virtual scenario, which is "neither a single mode nor more complex than pure synthesis."

[0045] For example, in a digital twin model of a high-pressure pumping station, the first agent might simultaneously inject high-frequency vibration peaks caused by bearing wear and temperature rise signals caused by reduced lubricant viscosity at the 120th second of a simulated run. The former is reflected in the vibration waveform via a virtual vibration sensor, while the latter is presented in the temperature curve via a virtual temperature probe. Subsequently, at the 150th second, the agent superimposes motor current fluctuations to further increase the complexity of fault scenarios. This entire process generates multi-cascading fault sequences with high frequency and repeatability, independent of actual equipment. This allows fault recognition models trained in the digital twin environment to continuously hone their resolution and emergency response strategies across a virtually unlimited and highly realistic set of custom fault combinations.

[0046] In one possible implementation, a mixed fault signal is injected into the digital twin environment through the first intelligent agent, specifically including: determining multimodal simulation features based on historical operating data and production line conditions; randomly extracting at least two simulation features from the multimodal simulation features through the first intelligent agent, and synthesizing a mixed fault signal according to a time series; and adding the mixed fault signal to the digital twin environment through the first intelligent agent according to the time of occurrence, intensity of occurrence, and duration of the fault.

[0047] Specifically, the server first extracts a variety of "simulated features" (such as high-frequency vibration patterns caused by bearing wear, temperature ramp-up curves caused by insufficient lubrication, and acoustic pulses generated by valve leakage) from existing multi-source data such as temperature, vibration, acoustics, and energy consumption, as well as operating condition information such as work orders and maintenance records. Subsequently, the first intelligent agent does not simply repeat a single pattern. Instead, it selects at least two from this pool of simulated features through "random combination" or "strategic selection" and splices and superimposes them in a predetermined or random time sequence to generate a "hybrid fault signal" with dual complexity in both temporal and modal dimensions. Finally, the intelligent agent is also responsible for precisely controlling the start time, intensity level, and duration of each fault sub-signal, and injecting it into the corresponding virtual sensor channel in the twin environment through the underlying interface to ensure that subsequent identification and game-playing stages can be trained and evaluated in dynamic fault scenarios close to reality.

[0048] For example, in a digital twin environment for a chemical pumping station, the server first extracts high-frequency vibration bands corresponding to "minor bearing wear" and small leakage acoustic signals corresponding to "seal aging" from historical operating data as simulated features. The first agent might trigger a "bearing wear" vibration spike at the 200th second, add a "seal leakage" acoustic pulse at the 220th second, and then superimpose a "motor current fluctuation" simulated feature at the 250th second to construct a three-stage hybrid fault process. The duration and intensity of each sub-signal are set by the agent based on previous game feedback or random strategies. For example, the vibration peak is set to a moderate intensity lasting 5 seconds, the acoustic pulse is set to a short, high intensity of 1 second, and the current fluctuation is set to be slight but lasting 10 seconds. This repeated combination and injection not only realistically reproduces complex scenarios in which multiple faults occur simultaneously or sequentially, but also trains the model's detection and adaptability capabilities from multiple angles and time sequences in a safe and controllable virtual scenario.

[0049] S140. Using the second intelligent agent, identify the fault type of the mixed fault signal to obtain an identification result.

[0050] Specifically, the second agent simultaneously perceives multiple signal streams from the virtual environment, including visual, vibration, acoustic, and temperature signals. Combining various pre-trained feature extraction units, it transforms the raw fault waveform into a more semantically informative internal representation. Subsequently, through an adaptive fusion strategy, it automatically weighs the importance of different modalities in the current scenario and dynamically adjusts the weights of each signal, allowing the most diagnostically valuable features to play a leading role in the decision-making process. Ultimately, the agent outputs a clear fault category label and provides a corresponding confidence level or priority indicator, facilitating targeted processing and optimization of the identification results in subsequent processes.

[0051] For example, consider the drive motor on a packaging production line. When the first agent is fed high-frequency vibrations caused by bearing wear and then a temperature rise due to insufficient lubrication, the second agent simultaneously captures the vibration signals through a virtual vibration sensor and acquires temperature fluctuation data using a built-in thermal imaging simulator. It identifies typical bearing fault patterns in the vibration signals and compares and integrates them with abnormal temperature trends, automatically determining whether bearing wear is the primary fault or insufficient lubrication is a secondary factor. Ultimately, the system reports "Major Fault: Bearing Wear (90% Confidence), Minor Fault: Insufficient Lubrication (65% Confidence)" to provide precise guidance for maintenance decisions.

[0052] In one possible implementation, a second intelligent agent is used to identify the fault type of a mixed fault signal to obtain an identification result, specifically including: using the second intelligent agent, using pre-trained visual, vibration, and acoustic coding units, mapping the mixed fault signal into an initial feature vector; using an adaptive attention mechanism to dynamically weight the initial feature vector to form a fused feature vector; performing classification inference on the fused feature vector, and outputting a fault category label and confidence to obtain an identification result.

[0053] Specifically, the second agent first receives a variety of fault signals injected into the virtual environment and maps these raw signals into a series of "initial feature vectors" through its pre-trained visual, vibration, and acoustic encoding units. Specifically, whether it is the thermal imaging captured by the camera, the high-frequency vibration waveform recorded by the vibration sensor, or the acoustic noise collected by the microphone, they are all converted into vector representations with semantic information by the corresponding encoding modules. These vectors retain the key diagnostic features of their respective channels, but are incompatible due to their heterogeneous sources. To this end, the second agent further introduces an adaptive attention mechanism to dynamically weight all initial feature vectors, assigning different weights to each dimension based on its "diagnostic value" in the current fault scenario, thereby generating a fused feature vector that not only integrates multimodal information but also highlights the most critical signals, ensuring more accurate subsequent judgments.

[0054] Based on this, the second agent inputs the fused feature vector into its classification reasoning module. By matching and comparing the fault category patterns learned during training, it ultimately outputs a specific fault type label and its corresponding confidence score. For example, in a chemical pumping station, when the system captures the superimposed "abnormal pump shaft vibration" and "seal leakage acoustic pulse" signals, the encoding unit extracts the vibration spectrum characteristics and acoustic pulse characteristics, respectively. The attention mechanism may give higher weight to the vibration characteristics, as historical data shows that bearing anomalies are more critical to the overall fault diagnosis. Classification reasoning then combines these two pieces of information to arrive at a diagnosis of "Major fault: bearing fatigue (92% confidence), Minor fault: seal leakage (68% confidence)," providing a clear and reliable basis for subsequent maintenance decisions.

[0055] S150. Using a third intelligent agent, comparative learning is performed on the recognition results to obtain a game result.

[0056] Specifically, the server activates a third agent (also known as a "game evaluator") in the digital twin environment to conduct self-supervised comparative learning on the fault identification results output by the second agent to generate a "game result"—that is, to quantify the second agent's ability to distinguish different fault types and its weaknesses at the current evolution node. Specifically, at the same time point or the same fault evolution node, the third agent extracts samples with large recognition deviations and inconsistent with the preset results as "positive samples," and samples with relatively accurate recognition and high confidence as "negative samples." By comparing the internal feature representations of these two types of samples, the third agent can discover the second agent's blind spots in fault confusion boundaries or feature overlap areas, and score each comparison result, thereby forming a comprehensive "game score" that covers multiple dimensions of information such as accuracy differences and modal weight imbalances.

[0057] For example, in a digital twin scenario involving wind turbine blade inspection, a second agent misclassified blade microcracks as blade wear, while accurately identifying blade bending. In this case, a third agent would use the misclassified examples (cracks identified as wear) as positive examples and the bending examples as negative examples, extracting and comparing their fused feature vectors. After internal comparative learning, the third agent would discover that the second agent's confidence distributions at the boundary between the crack and wear feature spaces are too close, resulting in a low match score and marking this boundary region as requiring reinforcement. Meanwhile, a high score would be assigned for the easily identifiable bending scenario, indicating that the current model has sufficient discriminatory power for this fault type. Ultimately, these match scores and accompanying sample feature feedback will guide subsequent environmental perturbations and model fine-tuning strategies, driving the second agent's discriminative capabilities in challenging scenarios.

[0058] In one possible implementation, a third intelligent agent is used to perform comparative learning on the recognition results to obtain a game result, specifically including: controlling the third intelligent agent to extract target samples that do not meet the preset fault category in the recognition results from the same model evolution node as positive samples; controlling the third intelligent agent to extract target samples that meet the preset fault category in the recognition results from the same model evolution node as negative samples; determining the first fusion feature vector corresponding to the positive sample and the second fusion feature vector corresponding to the negative sample; and generating a game result based on the confidence of the positive sample and the negative sample at the same model evolution node.

[0059] Specifically, the third agent (the evaluator) first selects two key types of samples from the same model evolution node: samples whose recognition results are inconsistent with the preset fault labels, exhibiting confusion or misjudgment, are considered "positive samples," while samples whose recognition results fully match the preset labels and have a high confidence level are considered "negative samples." The third agent then extracts the fused feature vectors of these two types of samples, which are compact representations of the modal information fused through adaptive attention. The feature vectors of the positive samples are referred to as the first fused feature vector, and the feature vectors of the negative samples are referred to as the second fused feature vector. The third agent then compares the confidence levels of these two vectors at the same evolution node. If the confidence level of the first fused feature vector is significantly lower than that of the second fused feature vector, it indicates that the model has not yet clearly distinguished the fault type. Otherwise, it indicates that the fault type has been well understood by the model. Based on this targeted, scenario-consistent sample comparison, the third agent ultimately generates the "game result," a quantitative evaluation report that accurately identifies the model's current strengths and weaknesses across different fault types.

[0060] For example, suppose in a digital twin environment of a printing press, the second agent incorrectly identifies a "nozzle clog" as "cartridge idle" but accurately identifies a "paper deviation" scenario. In this case, the third agent uses several nozzle clog samples that were incorrectly identified as "cartridge idle" as positive examples and correctly identifies "paper deviation" as negative examples. It extracts fused feature vectors from the two sets of samples and compares their corresponding recognition confidences at the current evolution node. If the average confidence for the nozzle clog samples is only 50%, while the average confidence for the paper deviation samples is 90%, the third agent generates a game result indicating that the boundary between "nozzle clog" and "cartridge idle" is too vague and needs to be strengthened. It also assigns a high score to "paper deviation," indicating that the model has sufficient discrimination for this category. This game result directly guides the focus of subsequent fault injection and the direction of online fine-tuning, making the model more sensitive and accurate in detecting subtle differences between nozzle clogs and cartridge idle in real production.

[0061] S160. Based on the game results, dynamically adjust the type and strength of the digital twin environment, generate an adaptation strategy, and adjust the industrial large model according to the adaptation strategy.

[0062] Specifically, the server uses the "game results" generated by the third agent as feedback to automatically optimize the configuration of fault injection in the virtual twin environment and formulate corresponding model adjustment plans accordingly to achieve continuous and accurate adaptation. On the one hand, the system will analyze the "weak links" marked in the game results, such as insufficient modal signal weight of a certain fault type, low recognition confidence, or overly vague boundaries, and map this information into "abnormal modes" and "abnormal parameters" that need to be enhanced. Then, based on this abnormal information, the environmental perturbation scheduler will make targeted adjustments to the fault type (such as adding a certain fault sub-mode), fault intensity (such as increasing the signal amplitude or extending the duration), and injection timing in the next round of simulation, forming a set of dynamically evolving "adaptation strategies" to continuously increase the training intensity of the model in the most error-prone areas.

[0063] After obtaining the adaptation strategy, the system will input the abnormal modes and abnormal parameters into the online fine-tuning framework to perform lightweight updates on the key components of the model (such as the attention distribution of the fusion layer or the weights of the feature extraction unit). With the help of this closed-loop process of "problem strategy fine-tuning", the model can quickly focus on and repair its own deficiencies exposed in the game without the need for training from scratch or using a large number of labeled samples. For example, if the game results indicate that the model performs poorly in distinguishing "seal leakage" from "bearing looseness", the system will increase the signal strength of these two faults at the same time in the next twin injection, and give priority to enhancing the attention weights of acoustic or vibration features in the fine-tuning link. Through continuous feedback and iteration, the industrial large model can eventually achieve efficient resolution and robust diagnosis of the fault pair in both simulation and real working conditions.

[0064] In one possible implementation, refer to Figure 2 , Figure 2 Another flow chart of the industrial large model adaptation method based on cross-model joint construction and contrastive learning provided in an embodiment of the present application includes steps S210 to S230, and the above steps are as follows: S210, according to the game results, obtain abnormal modes and abnormal parameters; S220, input the abnormal modes and abnormal parameters into the online fine-tuning framework to perform lightweight weight updates and attention distribution retraining to obtain an adaptation strategy; S230, based on the environment configuration, fine-tuning log and performance indicators in the adaptation strategy, adapt and adjust the industrial large model.

[0065] Specifically, for example, if the game results indicate that the model frequently misjudges the fault type based on acoustic signals, then "acoustics" becomes an abnormal mode; or if the vibration amplitude threshold for a specific frequency band is found to be too low to trigger a sensitive response, then the amplitude threshold for that frequency band becomes an abnormal parameter. Once this information is captured, it is fed into the online fine-tuning framework. Without affecting the overall architecture, the system uses lightweight weight fine-tuning and attention distribution retraining to allow the model to automatically adjust its focus on abnormal modes and sensitivity to key parameters, generating a highly targeted "adaptation strategy."

[0066] After obtaining the adaptation strategy, the system will make final adjustments to the industrial large-scale model using the environment configuration, fine-tuning log, and performance metrics recorded in the strategy. The environment configuration section specifies which fault types, signal strengths, or disturbance factors should be prioritized during the next simulation or deployment. The fine-tuning log details the weight update range, number of learning steps, and changes in attention distribution. Performance metrics measure key performance indicators such as accuracy, recall, or response latency before and after the model adjustment. For example, if the strategy indicates that the acoustic noise amplitude needs to be increased by 20% and the attention weight of this channel needs to be strengthened in the next round of twin simulation, the system will adjust the fine-tuning script accordingly, increasing the weights of the relevant layers of the acoustic encoder and redistributing the attention weights. Finally, the system will verify the results in real or virtual environments, ensuring that the fault identification accuracy has increased from the original 75% to over 90%. This closed-loop process of "problem-strategy-fine-tuning-verification-iteration" enables the model to continuously optimize itself, maintaining high adaptability and reliability in complex working conditions.

[0067] This application also provides an industrial large model adaptation device based on cross-model joint construction and comparative learning, referring to Figure 3 , Figure 3 A schematic diagram of a module for an industrial large-scale model adaptation device based on cross-model joint construction and comparative learning provided in an embodiment of the present application. The device is a server, comprising an acquisition module 31 and a processing module 32. The acquisition module 31 acquires historical operating data and production line conditions in the industrial field; the processing module 32 constructs a digital twin environment of the industrial large-scale model based on the historical operating data and production line conditions. The digital twin environment includes a first agent, a second agent, and a third agent; the processing module 32 injects mixed fault signals into the digital twin environment through the first agent; the processing module 32 identifies the fault type of the mixed fault signals through the second agent to obtain an identification result; the processing module 32 performs comparative learning on the identification result through the third agent to obtain a game result; and based on the game result, the processing module 32 dynamically adjusts the type and strength of the digital twin environment, generates an adaptation strategy, and adjusts the industrial large-scale model according to the adaptation strategy.

[0068] In one possible implementation, the acquisition module 31 acquires historical operating data and production line conditions in the industrial field, specifically including: the processing module 32 conducts system docking with PLC, DCS and MES through OPC UA, MQTT and RESTful API, and acquires the temperature, pressure, vibration, acoustics, energy consumption, work orders and operation logs of industrial equipment in the industrial field in real time to obtain first modal data; the acquisition module 31 acquires the sensor location, model, sampling frequency and maintenance records of the production line where the industrial equipment is located to obtain second modal data; the processing module 32 denoises, interpolates and aligns the first modal data and the second modal data to obtain historical operating data and production line conditions.

[0069] In one possible implementation, the processing module 32 constructs a digital twin environment of the industrial large model based on historical operating data and production line operating conditions, specifically including: the processing module 32 uses a three-dimensional simulation engine to replicate the workshop layout, machine geometry, and sensor installation position based on historical operating data and production line operating conditions; the processing module 32 loads the historical sensor curves and operating condition parameters into the initial digital twin model, and drives the initial digital twin model to run through kinematic and thermodynamic models to obtain the digital twin environment of the industrial large model.

[0070] In one possible implementation, the processing module 32 injects a mixed fault signal into the digital twin environment through the first intelligent agent, specifically including: the processing module 32 determines the multimodal simulation features based on historical operating data and production line conditions; the processing module 32 randomly extracts at least two simulation features from the multimodal simulation features through the first intelligent agent, and synthesizes the mixed fault signal according to the time series; the processing module 32 adds the mixed fault signal to the digital twin environment through the first intelligent agent according to the time of occurrence, intensity of occurrence and duration of the fault.

[0071] In one possible implementation, the processing module 32 uses a second intelligent agent to identify the fault type of the mixed fault signal and obtain an identification result, which specifically includes: the processing module 32 uses a second intelligent agent to map the mixed fault signal into an initial feature vector using pre-trained visual, vibration, and acoustic coding units; the processing module 32 uses an adaptive attention mechanism to dynamically weight the initial feature vector to form a fused feature vector; the processing module 32 performs classification inference on the fused feature vector, and outputs a fault category label and confidence to obtain an identification result.

[0072] In one possible implementation, the processing module 32 performs comparative learning on the recognition results through the third intelligent agent to obtain the game result, which specifically includes: the processing module 32 controls the third intelligent agent to extract target samples that do not meet the preset fault category in the recognition results from the same model evolution node as positive samples; the processing module 32 controls the third intelligent agent to extract target samples that meet the preset fault category in the recognition results from the same model evolution node as negative samples; the processing module 32 determines the first fusion feature vector corresponding to the positive sample and the second fusion feature vector corresponding to the negative sample; the processing module 32 generates the game result according to the confidence of the positive sample and the negative sample at the same model evolution node.

[0073] In one possible implementation, the processing module 32 dynamically adjusts the type and strength of the digital twin environment based on the game results, generates an adaptation strategy, and adjusts the industrial large model according to the adaptation strategy, specifically including: the processing module 32 obtains abnormal modes and abnormal parameters based on the game results; the processing module 32 inputs the abnormal modes and abnormal parameters into the online fine-tuning framework to perform lightweight weight updates and attention distribution retraining to obtain the adaptation strategy; the processing module 32 adapts and adjusts the industrial large model based on the environment configuration, fine-tuning logs and performance indicators in the adaptation strategy.

[0074] It should be noted that the above embodiments provide devices that implement their functions using only the division of the above functional modules as examples. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0075] This application also provides an electronic device, referring to Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include: at least one processor 41, at least one network interface 44, a user interface 43, a memory 45, and at least one communication bus 42.

[0076] The communication bus 42 is used to realize the connection and communication between these components.

[0077] The user interface 43 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 43 may also include a standard wired interface and a wireless interface.

[0078] The network interface 44 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0079] The processor 41 may include one or more processing cores. Using various interfaces and circuits, the processor 41 connects to various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 45, as well as accesses data stored in the memory 45, to perform various server functions and process data. Optionally, the processor 41 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field programmable gate array (FPGA), or a programmable logic array (PLA). The processor 41 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display screen; and the modem handles wireless communications. It is understood that the modem may also be implemented as a separate chip, rather than integrated into the processor 41.

[0080] Among them, the memory 45 may include a random access memory (RAM) or a read-only memory (Read Only Memory). Optionally, the memory 45 includes a non-transitory computer readable storage medium. The memory 45 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 45 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 45 may also be optionally at least one storage device located away from the aforementioned processor 41. As Figure 4 As shown, the memory 45 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an application program of an industrial large model adaptation method based on cross-model joint construction and comparative learning.

[0081] exist Figure 4In the electronic device shown, the user interface 43 is mainly used to provide an input interface for the user and obtain data input by the user; and the processor 41 can be used to call the application program of the industrial large model adaptation method based on cross-model joint construction and comparative learning stored in the memory 45. When executed by one or more processors, the electronic device executes one or more methods in the above embodiments.

[0082] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application.

[0083] The present application also provides a computer-readable storage medium storing instructions, which, when executed by one or more processors, enable an electronic device to execute one or more of the methods described in the above embodiments.

[0084] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0085] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0086] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0087] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory includes various media that can store program code, such as USB flash drives, mobile hard drives, magnetic disks, or optical disks.

[0089] The above is only an exemplary embodiment of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure. This application is intended to cover any variation, use or adaptive change of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art that are not recorded in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. An industrial large-scale model adaptation method based on cross-model joint construction and contrastive learning, characterized by: The method comprises: Obtain historical operating data and production line conditions in the industrial field; Constructing a digital twin environment of an industrial model based on the historical operating data and the production line operating conditions, wherein the digital twin environment includes a first agent, a second agent, and a third agent; injecting a mixed fault signal into the digital twin environment through the first agent; Using the second intelligent agent, performing fault type identification on the mixed fault signal to obtain an identification result; By means of the third agent, comparative learning is performed on the recognition results to obtain a game result; Based on the game results, the type and strength of the digital twin environment are dynamically adjusted, an adaptation strategy is generated, and the industrial large model is adjusted according to the adaptation strategy.

2. The industrial large model adaptation method based on cross-model joint construction and contrastive learning according to claim 1 is characterized in that: The acquisition of historical operating data and production line conditions in the industrial field specifically includes: Through OPC UA, MQTT, and RESTful API, the system is connected to PLC, DCS, and MES to obtain real-time temperature, pressure, vibration, acoustics, energy consumption, work orders, and operation logs of industrial equipment in the industrial field to obtain first modal data; Obtaining the sensor location, model, sampling frequency, and maintenance record of the production line where the industrial equipment is located to obtain second modal data; Denoising, interpolation, and alignment are performed on the first modal data and the second modal data to obtain the historical operation data and the production line operating conditions.

3. The industrial large model adaptation method based on cross-model joint construction and contrastive learning according to claim 1 is characterized in that: The digital twin environment of the industrial model is constructed based on the historical operation data and the production line working conditions, specifically including: Based on the historical operating data and the production line operating conditions, a 3D simulation engine is used to replicate the workshop layout, machine geometry, and sensor installation locations; The historical sensor curves and operating parameters are loaded into the initial digital twin model, and the initial digital twin model is driven to run through the kinematic and thermodynamic models to obtain the digital twin environment of the industrial large model.

4. The industrial large model adaptation method based on cross-model joint construction and contrastive learning according to claim 1 is characterized in that: The injecting of a mixed fault signal into the digital twin environment through the first intelligent agent specifically includes: determining a multimodal simulation feature based on the historical operating data and the production line operating conditions; randomly extracting at least two simulation features from the multimodal simulation features through the first agent, and synthesizing a mixed fault signal according to a time series; Through the first intelligent agent, the mixed fault signal is added to the digital twin environment according to the fault occurrence time, occurrence intensity and duration.

5. The industrial large model adaptation method based on cross-model joint construction and contrastive learning according to claim 1 is characterized in that: The performing fault type identification on the mixed fault signal by the second intelligent agent to obtain an identification result specifically includes: Mapping the mixed fault signal into an initial feature vector by the second agent using pre-trained visual, vibration, and acoustic encoding units; Adopting an adaptive attention mechanism to dynamically weight the initial feature vector to form a fused feature vector; Classification reasoning is performed on the fused feature vector, and a fault category label and confidence level are output to obtain the recognition result.

6. The industrial large model adaptation method based on cross-model joint construction and contrastive learning according to claim 1 is characterized in that: The third agent performs comparative learning on the recognition results to obtain a game result, specifically including: Controlling the third intelligent agent to extract target samples that do not meet the preset fault category in the identification results from the same model evolution node as positive samples; Controlling the third intelligent agent to extract target samples that meet a preset fault category from the recognition results from the same model evolution node as negative samples; Determine a first fused feature vector corresponding to the positive pair sample and a second fused feature vector corresponding to the negative pair sample; The game result is generated according to the confidence levels of the positive and negative samples at the same model evolution node.

7. The industrial large model adaptation method based on cross-model joint construction and contrastive learning according to claim 1 is characterized in that: Based on the game results, the type and strength of the digital twin environment are dynamically adjusted, an adaptation strategy is generated, and the industrial model is adjusted according to the adaptation strategy, specifically including: Obtaining abnormal modes and abnormal parameters according to the game result; Inputting the abnormal modality and the abnormal parameters into an online fine-tuning framework to perform lightweight weight update and attention distribution retraining to obtain the adaptation strategy; Based on the environment configuration, fine-tuning logs and performance indicators in the adaptation strategy, the industrial large model is adapted and adjusted.

8. Industrial large model adaptation device based on cross-model joint construction and contrastive learning, characterized by: The device comprises an acquisition module (31) and a processing module (32), wherein: The acquisition module (31) is used to acquire historical operating data and production line operating conditions in the industrial field; The processing module (32) is used to construct a digital twin environment of the industrial model based on the historical operation data and the production line working conditions, wherein the digital twin environment includes a first intelligent agent, a second intelligent agent, and a third intelligent agent; The processing module (32) is further configured to inject a mixed fault signal into the digital twin environment through the first intelligent agent; The processing module (32) is further configured to perform fault type identification on the mixed fault signal through the second intelligent agent to obtain an identification result; The processing module (32) is further configured to perform comparative learning on the recognition results through the third intelligent agent to obtain a game result; The processing module (32) is further configured to dynamically adjust the type and strength of the digital twin environment based on the game result, generate an adaptation strategy, and adjust the industrial large model according to the adaptation strategy.

9. An electronic device, characterized in that: The electronic device comprises a processor (41), a memory (45), a user interface (43) and a network interface (44), wherein the memory (45) is used to store instructions, the user interface (43) and the network interface (44) are both used to communicate with other devices, and the processor (41) is used to execute the instructions stored in the memory (45) so that the electronic device executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 7 is performed.

Images