A strategy self-evolution industrial vertical domain large model self-optimization learning method and system

Abstract

The application discloses a strategy self-evolution industrial vertical domain large model self-optimization learning method and system, relates to the technical field of industrial artificial intelligence, and comprises the following steps: acquiring a current working condition description, inputting the working condition description into an industrial vertical domain large model, and generating an execution strategy matched with the current working condition; executing an industrial task based on the execution strategy, generating an execution result, acquiring a feedback result corresponding to the execution result, recording the execution strategy and the feedback result, and forming a strategy and result mapping table; according to the strategy template, the historical performance of each strategy template is counted, the strategy template is classified according to the historical performance, and the evolution of the execution strategy is driven according to the classification result and the feedback result. The method improves the adaptability of the strategy and the working condition, the collaborative efficiency of the strategy optimization and the model parameter optimization, the updating accuracy of the model under the delayed feedback, the credibility of the training data, and the safety and reliability of the online self-optimization in the industrial scene.

Description

Technical Field

[0001] This invention relates to the field of industrial artificial intelligence technology, specifically to a self-optimizing learning method and system for large-scale industrial vertical models with self-evolutionary strategies. Background Technology

[0002] As the application scale of large-scale models in the industrial field continues to expand, industrial vertical-domain large-scale models often use supervised fine-tuning or reinforcement learning to complete domain adaptation. In this process, the task execution strategy (including which sensors to select, which features to extract, and the order of inference) is statically preset by engineers once, and cannot be automatically adjusted with changes in real-time operating conditions. When operating conditions change, the model still uses the old strategy, which makes it difficult to guarantee the output quality. The strategy itself is also independent of the model parameters and cannot be improved in synergy with the model's capabilities.

[0003] In the above situation, even if the model parameters can be continuously updated through fine-tuning, the execution strategy remains in the initial state set manually and cannot learn from the feedback. At the same time, the feedback signals from the industrial field are often delayed for hours or even days from model output to collection. The feedback data is used with equal weight when it is mismatched with the current model state that has undergone multiple rounds of updates. This will cause outdated feedback to conflict with the effective patterns recently learned by the model, interfering with the current model update process.

[0004] Furthermore, feedback signals from industrial sites may be distorted due to sensor failures, communication interruptions, violations of process rules, or human annotation biases. If these signals are used directly for model training without verification, the model will learn incorrect correlation patterns. At the same time, if there is a lack of continuous quality monitoring and degradation protection mechanisms after the model is updated online, it will be difficult to detect and recover from capability degradation caused by the new version in a timely manner, which seriously affects the safety and reliability of online self-optimization in industrial scenarios. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is that existing industrial vertical domain large model task execution methods have the following problems: strategy depends on manual static preset and cannot evolve dynamically with working conditions; strategy optimization and model parameter update are disconnected and have low coordination efficiency; and how to utilize the delayed, sparse and potentially distorted feedback signals in industrial sites to enable the model to self-optimize online.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a self-optimizing learning method for a large industrial vertical domain model with self-evolving strategies, comprising: acquiring a description of the current operating condition; inputting the description of the operating condition into a large industrial vertical domain model to generate an execution strategy matching the current operating condition; executing industrial tasks based on the execution strategy to generate execution results; acquiring feedback results corresponding to the execution results; recording the execution strategies and feedback results in correspondence to form a strategy-result mapping table; statistically analyzing the historical performance of each strategy template according to the strategy templates; classifying the strategy templates according to their historical performance; and driving the evolution of the execution strategy based on the classification results and feedback results.

[0008] As a preferred embodiment of the self-evolutionary industrial vertical domain large model self-optimization learning method described in this invention, the execution strategy is parsed and mapped to a set of machine-executable parameters; the data acquisition range is mapped to the calling parameters of the data acquisition interface; the feature extraction method is mapped to the selection of signal processing functions; the inference step arrangement is mapped to the task execution order; and the judgment threshold is mapped to the numerical comparison condition, so that the execution strategy is directly transformed from a text description into an executable parameter sequence.

[0009] As a preferred embodiment of the self-evolutionary learning method for large industrial vertical models in the strategy self-evolution of the present invention, the method comprises: forming a strategy-result mapping table, classifying and storing the execution strategies according to the process sub-domain, and statistically analyzing the historical performance of each strategy template according to the process sub-domain, so that the historical performance statistics of each strategy template correspond to the working condition characteristics of the process sub-domain.

[0010] As a preferred embodiment of the self-evolutionary industrial vertical domain large-scale model self-optimization learning method of the present invention, the following steps are taken: classifying strategy templates according to historical performance includes: counting the number of times each strategy template received correct feedback and the total number of times it was used; calculating the accuracy of each strategy template; marking strategy templates with an accuracy reaching a first threshold as high-quality strategies; marking strategy templates with an accuracy less than or equal to a second threshold as strategies to be optimized; marking strategy templates with an accuracy between the first and second thresholds as candidate strategies for continuing to accumulate feedback; and classifying each strategy template into different evolution stages based on the statistical results.

[0011] As a preferred embodiment of the self-optimization learning method for industrial vertical domain large-scale models with self-evolutionary strategies described in this invention, the execution of strategy evolution includes: statistically analyzing the error rate according to the process sub-domain; when the error rate exceeds a trigger threshold, using the strategy to be optimized and the corresponding error results in the process sub-domain as training data to fine-tune the industrial vertical domain large-scale model; when the number of times the strategy template receives correct feedback reaches a set number, the strategy template is marked as a stable strategy, and the stable strategy is used as a directly reused strategy for subsequent executions of similar tasks.

[0012] As a preferred embodiment of the self-optimization learning method for large industrial vertical models with self-evolutionary strategies described in this invention, the fine-tuning of the large industrial vertical model includes: constructing strategy training samples and result training samples; the strategy training samples take the current working condition description and the execution strategy used in the previous step as input, and the corrected execution strategy as the target output; the result training samples take the current working condition description and the corrected execution strategy as input, and the correct execution result as the target output; the strategy training samples and result training samples are jointly trained in the same batch, and the weighted sum of the loss of the strategy training samples and the loss of the result training samples is used as the total loss.

[0013] As a preferred embodiment of the self-evolutionary industrial vertical domain large-scale model self-optimization learning method described in this invention, the following is defined: In joint training, the timeliness weight of each training sample is determined based on the delay time of the feedback result; training samples corresponding to feedback results with delay times within the adoption threshold are fully involved in training; training samples corresponding to feedback results with delay times exceeding the discard threshold are not involved in training; training samples corresponding to feedback results with delay times between the adoption and discard thresholds participate in training with a decay weight, which decreases as the delay time increases.

[0014] As a preferred embodiment of the self-evolutionary industrial vertical domain large-scale model self-optimization learning method described in this invention, the following steps are performed: before the feedback results participate in training, physical constraint verification, process rule verification, and similar working condition consistency verification are performed on the feedback results; physical constraint verification includes verifying whether the sensor data involved in the feedback results are within the physical range, whether the data change rate is within the physical reasonable range, and whether the key sensor combination is self-consistent; process rule verification includes verifying whether the model output is within the operating range specified by the process standard, whether the feedback results contradict the existing rules in the process knowledge base, and whether the quality inspection judgment results are consistent with the process quantity trend; similar working condition consistency verification includes: retrieving historical feedback records under similar working conditions from the historical feedback database, counting the number of historical feedbacks consistent with the current feedback annotation conclusion, and calculating the consistency ratio. When the consistency ratio is lower than the minimum consistency threshold, the current feedback is marked as questionable feedback; when the feedback conclusion is a continuous numerical type, the percentile ranking of the current feedback conclusion in the set of historical feedback results under similar working conditions is calculated. When the percentile ranking is in an extreme position range, the current feedback is marked as questionable feedback; only the feedback results obtained through physical constraint verification, process rule verification, and similar working condition consistency verification participate in training.

[0015] As a preferred embodiment of the self-evolutionary industrial vertical domain large model self-optimization learning method described in this invention, the following steps are taken: after the industrial vertical domain large model is updated, the quality indicators of the model are continuously monitored; when any quality indicator deteriorates beyond the alarm threshold and continues to reach the confirmation time, the industrial vertical domain large model is rolled back to the most recent historical version with the best quality; when the updated version continues to run stably and reaches the target time, the version is solidified as the new baseline, and subsequent incremental updates are based on the new baseline.

[0016] Another objective of this invention is to provide a self-optimizing learning system for large-scale industrial vertical models with self-evolving strategies. This system can statistically analyze the historical performance of each strategy template according to the strategy template, classify the strategy templates based on the historical performance, and drive the evolution of execution strategies based on the classification results and feedback results. This solves the problem that current methods for executing large-scale industrial vertical model tasks contain a disconnect between strategy generation and execution, and cannot adaptively adjust based on historical performance.

[0017] As a preferred embodiment of the self-evolving industrial vertical domain large-scale model self-optimization learning system described in this invention, it includes: a strategy generation module, a strategy evolution module, a model fine-tuning module, and a verification management module; the strategy generation module is used to obtain working condition descriptions and generate execution strategies in structured text form, the execution strategies including data collection range, feature extraction methods, inference step arrangement, and judgment thresholds; the strategy evolution module is used to receive execution strategies and corresponding feedback results, form a strategy-result mapping table, statistically analyze the historical performance of each strategy template and classify the strategy templates, and update the execution strategies according to the classification results; the model fine-tuning module is used to construct training samples based on the corrected execution strategies and correct execution results, and update the parameters of the industrial vertical domain large-scale model; the verification management module is used to perform validity verification on the feedback results, monitor the quality indicators of the updated industrial vertical domain large-scale model, and roll back to a historical stable version when the quality deteriorates.

[0018] The beneficial effects of this invention are: The self-evolving industrial vertical-domain large-scale model self-optimization learning method provided by this invention transforms task execution strategies from manual static configuration to dynamic objects that are automatically generated by the model and continuously evolve based on feedback through a strategy self-generation and evolution mechanism. Each time an industrial task is executed, the model automatically generates an execution strategy oriented towards the current working condition and drives strategy self-optimization and solidification based on feedback tags, solving the problems of strategy-working-condition disconnect and strategy optimization-model parameter optimization in existing methods. Through a two-stage joint fine-tuning method of strategy-execution, using the execution strategy as an intermediate layer context, it expands the optimization of only the conclusion generation effect to the joint optimization of strategy + result. This not only optimizes the strategy output capability but also the capability to improve the strategy based on the previous step. Simultaneously, it solves the problem of outdated feedback interference caused by industrial feedback delays through a feedback timeliness weighting mechanism. Through a three-layer progressive feedback validity verification mechanism, physical constraint checks, process rule checks, and similar working-condition consistency checks are performed before feedback is used for model updates, filtering out unreliable feedback caused by sensor failures, non-model factor interference, and human subjective bias. Meanwhile, through quality monitoring and automatic rollback mechanisms, it is ensured that when online updates cause capability degradation, it can automatically revert to a historical stable version, thus guaranteeing the security and reliability of online self-optimization in industrial scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is an overall flowchart of the self-optimization learning method for large industrial vertical models with self-evolutionary strategies in this invention.

[0021] Figure 2 This is a schematic diagram illustrating the execution strategy self-generation and evolution mechanism of the self-optimization learning method for large industrial vertical models in this invention.

[0022] Figure 3 The LoRA training loss trend diagram is used for fault diagnosis of excessive vibration of generator in thermal power unit using the self-optimization learning method of industrial vertical domain large model for strategy self-evolution in this invention.

[0023] Figure labeling: 100, Large-scale industrial vertical model; 200, Execution strategy; 201, Execution result; 300, Feedback result; 400, Strategy template; 401, Excellent strategy; 402, Strategy to be optimized. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0027] Example 1

[0028] Reference Figure 1 This is the first embodiment of the present invention, which provides a self-optimization learning method for large industrial vertical models with self-evolutionary strategies.

[0029] Specifically, the current operating condition description is obtained, including equipment status data, process parameter data, and historical trend data. This description is input into the industrial vertical domain large model 100. The industrial vertical domain large model 100 performs a forward inference process to generate an execution strategy 200 that matches the current operating condition. The execution strategy 200 is output in structured text format, including which sensor data should be collected for this task, which features should be extracted, the order of inference steps, and key judgment thresholds. After generating the execution strategy 200, the industrial task is executed based on it. The execution strategy 200 is parsed and mapped to a set of machine-executable parameters, specifically: sensor selection is mapped to data acquisition interface call parameters, feature extraction is mapped to signal processing function selection, inference steps are mapped to the execution order, and judgment thresholds are mapped to numerical comparison conditions. The industrial vertical domain large model 100 concatenates the execution strategy 200 with the original operating condition information and uses an agent to generate an execution result 201 based on this strategy. The execution result 201 is a diagnostic conclusion or parameter recommendation.

[0030] Obtain the feedback result 300 corresponding to the execution result 201. The feedback result 300 includes information on whether it is correct or incorrect. Record the execution strategy 200 and the corresponding feedback result 300 to form a strategy-result mapping table. Classify and store the execution strategies 200 according to the process subdomain. Statistically analyze the historical performance of each strategy template 400 according to the process subdomain, so that the historical performance statistics of each strategy template 400 correspond to the operating conditions of the process subdomain. Statistically analyze the historical performance of each strategy template 400 according to the strategy template 400, specifically: count the number of times each strategy template 400 obtained correct feedback and the total number of times it was used, and calculate the accuracy rate of each strategy template 400. Classify the strategy templates 400 according to the accuracy rate: strategy templates 400 with an accuracy rate reaching the first threshold are marked as high-quality strategies 401, strategy templates 400 with an accuracy rate less than or equal to the second threshold are marked as strategies to be optimized 402, and strategy templates 400 with an accuracy rate between the first and second thresholds are marked as candidate strategies for further accumulating feedback and will not participate in the evolution for the time being. The typical range for the first threshold is 0.8 to 0.95, and the typical range for the second threshold is 0.3 to 0.6, with the second threshold being less than the first threshold. Based on the classification results and feedback results 300, the execution strategy 200 evolves. According to the error rate statistics for each process subdomain, when the error rate exceeds the trigger threshold, the optimization strategy 402 corresponding to the process subdomain and its corresponding error results are used as training data to fine-tune the industrial vertical domain large model 100. This allows the industrial vertical domain large model 100 to avoid known inefficient patterns when generating strategies in the next iteration. The typical range for the trigger threshold is 0.1 to 0.3. When the strategy template 400 receives correct feedback a set number of times, it is marked as a stable strategy. Subsequent similar tasks can directly reuse this stable strategy without needing to regenerate it each time. The typical range for the set number of times is 5 to 20.

[0031] When fine-tuning the industrial vertical domain large model 100, strategy training samples and result training samples are constructed and jointly trained in the same batch. The total loss is the weighted sum of the losses of the strategy training samples and the result training samples. During joint training, the timeliness weight of each training sample is determined based on the latency of the feedback result 300. The total loss is the sum of the timeliness weight of each training sample multiplied by the weighted sum of the losses of the strategy training samples and the result training samples. Parameter updates only apply to the LoRA adapter; the weights of the base model remain unchanged. Before the feedback result 300 participates in training, it undergoes a three-layer progressive verification: physical constraint verification, process rule verification, and similar working condition consistency verification. Only the feedback result 300 that has passed the physical constraint verification, process rule verification, and similar working condition consistency verification participates in training. After the industrial vertical domain large model 100 is updated, the model's quality indicators are continuously monitored. When any quality indicator deteriorates beyond the alarm threshold and continues to reach the confirmation time, the industrial vertical domain large model 100 is automatically rolled back to the most recent historical version with the best quality. Once the updated version has been running stably for a certain period of time, it will be fixed as the new baseline, and all subsequent incremental adjustments will be based on the new baseline.

[0032] Example 2

[0033] Reference Figure 1 and Figure 2 This is the second embodiment of the present invention. This embodiment provides a method for mapping structured execution strategy 200 to machine executable parameters, which solves the problem that the abstract strategy text generated by large models cannot be directly used in industrial control systems, resulting in a semantic gap in strategy execution.

[0034] Specifically, the structured execution strategy 200 generated by the industrial vertical domain large model 100 includes fields such as data acquisition range, feature extraction method, inference step arrangement and judgment threshold. After parsing, the execution strategy 200 is mapped to a set of parameters that can be executed by the machine: the data acquisition range is mapped to the calling parameters of the data acquisition interface, the feature extraction method is mapped to the selection of signal processing functions, the inference step arrangement is mapped to the task execution order, and the judgment threshold is mapped to the numerical comparison conditions.

[0035] Taking the execution strategy 200 under a target load of 500MW as an example, this strategy, after parsing, is mapped as follows: calling the SIS historical database and operation ticket interface as data acquisition parameters; selecting temperature rise trend and load slope signal processing functions as feature extraction methods; executing the inference steps sequentially in the order of "whether the temperature rise is too fast → whether there is a load adjustment → whether the load adjustment speed is too fast"; and comparing the numerical output of each step with "bearing housing temperature rise > 10℃, load increase rate > 5MW / minute". Through the above mapping, the strategy text generated by the industrial vertical domain large model 100 is directly converted into an executable parameter sequence, and there is no semantic gap between the description and execution of the strategy. After the execution strategy 200 is implemented, its performance needs to be continuously recorded and evaluated to determine whether the strategy is suitable for the current operating conditions and to decide whether to adjust it, thus entering the strategy evaluation and evolution stage.

[0036] Furthermore, after each output of the large industrial domain model 100, the execution strategy 200 and the feedback result 300 (correct / incorrect) of that output are recorded together to form a "strategy-result" mapping table. This table is then stored and statistically analyzed according to the process sub-domains. The process sub-domains are divided based on factors such as equipment type and process steps. For example, separate strategy-result mapping tables are established for different fields such as thermal power, wind power, and chemical industry. This ensures that the historical performance statistics of each strategy template 400 correspond to the operating conditions of the process sub-domains, avoiding evaluation distortion caused by cross-domain strategy performance interference.

[0037] The historical performance of each strategy template 400 is statistically analyzed and represented as follows: , in, For accuracy, For strategy The number of times you receive the correct feedback. For strategy Total number of times it was used.

[0038] Strategies are categorized based on accuracy. Strategies with an accuracy greater than or equal to the high-quality strategy threshold are labeled as high-quality strategy 401; those with an accuracy less than or equal to the strategy threshold for optimization are labeled as strategy for optimization 402; and strategies with accuracy between the high-quality and strategy thresholds continue to accumulate feedback samples but do not participate in evolution. Both the high-quality and strategy for optimization thresholds are empirical values. The typical range for the high-quality strategy threshold is 0.8–0.95; a higher value indicates greater stability for the selected high-quality strategy 401. The typical range for the strategy for optimization threshold is 0.3–0.6; a higher value indicates more aggressive self-optimization in the industrial vertical large-scale model 100. In practical engineering, these values ​​are usually adjusted according to the needs of strategy quality and evolution pace.

[0039] like Figure 2As shown, the current working condition description and the execution strategy 200 used in the previous step are obtained and input into the industrial vertical domain large model 100 for inference to generate the execution strategy 200 corresponding to the current working condition, and its accuracy is verified. Next, it is determined whether the accuracy of the execution strategy 200 is greater than the high-quality strategy threshold, i.e., the first threshold: if the result is yes, the strategy is marked as a high-quality strategy 401; if the result is no, it is further determined whether the accuracy of the execution strategy 200 is less than the optimization strategy threshold, i.e., the second threshold: if the result is yes, the strategy is marked as optimization strategy 402, triggering the strategy evolution process, and the strategy and its corresponding error feedback are used as training data to fine-tune the industrial vertical domain large model 100; if the result is no, the accuracy of the strategy is between the first threshold and the second threshold, marked as a candidate strategy, and feedback samples continue to be accumulated, but it does not participate in the evolution for the time being. After the industrial vertical domain large model 100 corresponding to optimization strategy 402 is fine-tuned, the optimized execution strategy 200 is output, and its accuracy is re-verified in subsequent working conditions, forming a closed-loop iteration. Through the above process, the strategy template 400 is classified into different evolution stages (high-quality solidification stage, optimization stage, and candidate accumulation stage) based on statistical results, thereby realizing the adaptive classification and on-demand evolution of the execution strategy 200.

[0040] In the thermal power plant vibration diagnosis dataset, for operational influence strategy template 400, the accuracy rate was statistically obtained as follows: The strategy marked as high-quality (401) and targeting bearing failure strategy template (400) have a statistical accuracy rate of [percentage missing]. Templates marked as strategy 402 to be optimized are analyzed by process sub-domain, and the error rate is expressed as follows: , in, For the process subfield Error rate, For the process subfield Number of times error feedback is received For the process subfield The total number of outputs.

[0041] When the error rate exceeds the trigger threshold, strategy evolution is triggered. The system fine-tunes the industrial vertical domain large model 100 with the historical optimization strategy 402 and corresponding error results of the process subdomain. This allows the industrial vertical domain large model 100 to avoid statistically analyzed inefficient patterns when generating strategies in the next iteration. The trigger threshold is an empirical value, typically ranging from 0.1 to 0.3. Higher values ​​indicate stricter requirements for strategy stability; however, it should not be set too high, as the large model cannot guarantee complete accuracy, and the operating conditions of different data may vary significantly. Overly strict settings can easily lead to the industrial vertical domain large model 100 continuously triggering optimization. For example, in the thermal power vibration diagnosis dataset, the overall error rate of the process subdomain is statistically obtained... This triggers fine-tuning of the historical optimization strategy 402 and its erroneous results in the subdomain, enabling the subsequent execution strategy 200 generated by the industrial vertical domain large model 100 to avoid known inefficient patterns. After the strategy evolution is triggered, the specific improvement methods for the industrial vertical domain large model 100 involve the joint training of strategy samples and result samples.

[0042] Example 3

[0043] Reference Figure 1 and Figure 3 This is the third embodiment of the present invention. This embodiment provides a method for joint fine-tuning of strategy and result in two stages, which solves the problem that the existing fine-tuning method only optimizes the ability to generate the final conclusion, and cannot improve the model's ability to generate and correct the execution strategy 200, resulting in the repeated occurrence of similar strategy errors.

[0044] Specifically, when strategy template 400 receives correct feedback a preset number of times, the strategy is marked as a stable strategy. Subsequent similar tasks can directly reuse this strategy without needing to regenerate it each time, thus reducing inference overhead. The preset number of times is an empirical value, typically ranging from 5 to 20. A higher value indicates a more stringent assessment of strategy stability; a value of 5 is commonly used to strike a balance between stability and timely reuse. For example, in the case of thermal power plant vibration diagnosis, the number of correct responses from this high-quality strategy 401 on historical data... If the conditions for solidification are met, it can be used as a direct reuse strategy for subsequent execution of similar tasks.

[0045] For strategies that have not met the solidification conditions, when feedback indicates that the industrial vertical domain large model 100 outputs an error, two training samples are constructed: the strategy training sample takes the current operating condition description and the execution strategy 200 used in the previous step as input, and the corrected execution strategy 200 as the target output. The corrected strategy can be generated by manual review or automatically generated by the process knowledge base according to the error type; the result training sample takes the current operating condition description and the corrected execution strategy 200 as input, and the correct diagnostic conclusion or parameter recommendation as the target output. The two samples are jointly trained in the same fine-tuning batch, and the total loss is the weighted sum of the two, expressed as: , in, For the total loss, The cross-entropy loss is calculated on the training samples of the strategy. The result is the cross-entropy loss on the training samples. To train weights for the policy, when At that time, the industrial vertical domain large model 100 focuses on learning a better strategy generation method; when At that time, the Industrial Vertical Large Model 100 focuses on learning the correct way to generate results under a given strategy.

[0046] Weight The value should be set according to actual needs; the lower the value, the more emphasis is placed on learning outcomes, and the higher the value, the more emphasis is placed on learning strategies. In thermal power plant vibration diagnosis experiments, the value should be set accordingly. This allows the large-scale industrial vertical model 100 to achieve a significant improvement in strategy generation capabilities while retaining its result generation capabilities. For example... Figure 3 The loss curves shown on the thermal power plant vibration diagnostic data when performing LoRA incremental fine-tuning indicate that, with the increase of training steps, the two types of losses gradually converge and stabilize at a low level after about 100 steps. This shows that the two-stage joint fine-tuning achieves stable optimization on both the strategy generation and result generation sides. The joint fine-tuning corrects the strategy generation method while correcting the results, enabling the industrial vertical domain large model 100 to expand from only having the ability to correct results to having the ability to generate both strategies and results, thus avoiding the recurrence of the same type of error.

[0047] Example 4

[0048] Reference Figure 1 This is the fourth embodiment of the present invention. This embodiment provides a time-weighted feedback filtering and three-layer progressive verification method to solve the problem that delayed arrival of feedback signals in industrial fields causes outdated data to interfere with the current model state, and that unreliable feedback such as sensor failure, violation of process rules, and manual annotation bias is directly used for training, thereby reducing the reliability of the model.

[0049] Specifically, due to the significant arrival delay of feedback signals from the industrial field, if all feedback is used with equal weight, outdated feedback will interfere with the current state of the large industrial vertical model 100. Therefore, it is necessary to perform time-attenuation processing on the delayed feedback. Simultaneously, the reliability of the feedback signals also needs to be verified before entering training. Let the feedback delay time be... (The time interval from the output of the large industrial vertical model 100 to receiving feedback), the time-delay weight is expressed as: , in, As a time-decrease weight, To provide feedback on the delay time, Assuming an adoption threshold, feedback delayed within this threshold is entirely used for training. This is the decay coefficient, controlling the exponential decay rate, typically ranging from 0.005 / hour to 0.05 / hour. The threshold for discarding is set; feedback with a delay exceeding this value is discarded directly, typically ranging from 48 to 168 hours.

[0050] The parameters are determined by the amount of data in the actual application; the larger the data volume, the smaller the feedback interval. The smaller the value, The larger the value, Adjustments were then made to match the data pace. For example, in the thermal power plant vibration diagnostic dataset, the median feedback delay was 62 hours, which is within the adoption threshold. Hours and Discard Threshold Between hours, the timeliness weight is approximately 0.705, meaning that this part of the feedback only participates in training with about 70% of the weight. The greater the delay, the smaller the weight, and the impact of outdated feedback on the current state of the large industrial vertical model 100 is significantly weakened.

[0051] Applying a timeliness weight to the loss term of each training sample, the weighted total loss of the fine-tuning batch is expressed as: , in, To fine-tune the total loss of the batch, For batch size, For the first Feedback delay time for each sample For the first The cross-entropy loss of the training samples of the strategy. For the first The result is the cross-entropy loss on the training samples.

[0052] Parameter updates only apply to the LoRA adapter; the weights of the base model remain unchanged, as shown below: , in, This is the parameter increment matrix for the LoRA adapter. The learning rate, typically taking the value of... ~ If the value is too large, the large industrial vertical model 100 may fail to converge; if it is too small, the convergence will be too slow.

[0053] Before feedback result 300 is used in training, it must undergo physical constraint verification, process rule verification, and similar operating condition consistency verification in sequence. Physical constraint verification checks whether the sensor data involved in the feedback is within the physical range, whether the data change rate is within a physically reasonable range, and whether the key sensor combination is self-consistent. Feedback that fails this check is marked as sensor abnormal and will not be used for updating the industrial vertical domain large model 100. Process rule verification checks whether the output of the industrial vertical domain large model 100 is within the operating range specified by the process standard, whether feedback result 300 contradicts existing rules in the process knowledge base, and whether the quality inspection judgment results are consistent with the process quantity trend. Feedback that fails this check is marked as abnormal and awaits manual review and confirmation. Similar operating condition consistency verification retrieves historical feedback records under similar operating conditions from the historical feedback database based on the operating condition characteristics (equipment type, process parameter range, fault mode category, etc.) corresponding to the current feedback. Typical values ​​range from 10 to 50. The number of conclusions in historical feedback regarding similar working conditions that are consistent with the conclusions marked in the current feedback is [number missing]. The consistency ratio is: , in, For the consistency ratio, For the number of consistent values, This represents the number of searches.

[0054] like (Minimum consistency threshold, typically 0.2~0.3) indicates that the correct result of this feedback annotation is inconsistent with the conclusions of most similar working conditions in history, and there is a high probability that it is an incorrect annotation. It is marked as questionable and will not be automatically used for updating the large industrial vertical model 100, but will enter the manual review queue. For example, in the thermal power vibration diagnosis dataset, using a process parameter range (70% load, vibration exceeding limits, 1st harmonic dominant) with an annotation result of "foundation loosening" as the search condition, the following results were retrieved from the historical feedback database: Of the 100 similar working condition records, only 2 were labeled as having loose basic conditions, which is consistent with the current feedback labeling conclusion. The consistency ratio was approximately 0.067, which is lower than the minimum consistency threshold of 0.2. Therefore, the feedback was deemed questionable and was removed.

[0055] If the feedback conclusion is a continuous numerical value (such as prediction deviation or temperature value), then percentile deviation is used for judgment: calculate the percentile ranking of the current feedback conclusion in the set of 300 historical feedback results under similar working conditions. If the percentile ranking is below 5% or above 95%, that is, the current feedback result is at an extreme position in the historical distribution, it is judged as inconsistent and marked as questionable. The minimum consistency threshold and percentile boundaries (5%, 95%) should be determined based on the differences in the actual working condition distribution of the data. When the differences in working condition distribution are large, a more lenient threshold should be adopted to avoid excessive rejection.

[0056] Only after feedback from the aforementioned three layers of verification does the system proceed to the two-stage joint fine-tuning process described above. After fine-tuning, the system continuously monitors the quality indicators of the industrial vertical domain large model 100, including output adoption rate, error rate, and key indicator anomaly rate. When any indicator deteriorates beyond the alarm threshold and is continuously confirmed for a certain period, the system automatically rolls back to the most recent historical version with the best quality. The confirmation time typically ranges from 2 to 6 hours. Once a version has been running stably for a certain period and meets the target time, that version is solidified as the new baseline. All subsequent incremental fine-tuning is based on the new baseline, with the target time typically ranging from 12 to 24 hours. Both the confirmation time and the target time are empirical values; larger values ​​indicate a more conservative and stringent judgment, resulting in a more robust response but also a longer lag.

[0057] Example 5

[0058] As a fifth embodiment of the present invention, a self-optimization learning method for large industrial vertical models with self-evolutionary strategies is provided. To verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0059] First, the base model adopts the Qwen3.6-27B large model, incrementally trained based on industrial datasets. After being merged with the original weights, it is used for root cause analysis and fault handling suggestions for generator vibration anomalies. On-site data systems that can be accessed include: the plant-level monitoring information system (SIS) (real-time data such as unit load, shaft vibration, bearing vibration, and bearing housing temperature), the inspection system (abnormal sound records, vibration spectrum analysis reports, coupling oscillation inspection records, and other inspection data), the ERP defect management system (defect data such as defect work orders), the operation ticket management system (load adjustment, start-up and shutdown records, and other data), the work ticket management system (maintenance and repair operation records, and other data), and the periodic work management system (reports on dynamic balancing tests and alignment checks, and other data).

[0060] The training dataset used contains 393 entries, consisting of input data corresponding to various faults that triggered mechanism threshold alarms and root causes of faults identified by experts. These include 221 entries related to operational impact (R1), 57 entries related to coupling alignment issues (R2), 23 entries related to bearing failures (R3), 31 entries related to dynamic imbalance (R4), and 61 entries related to foundation loosening (R5). The parameter settings are as follows: high-quality strategy threshold is 0.8, strategy to be optimized threshold is 0.3, number of solidification iterations is 20, self-optimization trigger threshold is 0.2, strategy training weight is 0.6, adoption threshold is 48 hours, discard threshold is 168 hours, decay coefficient is 0.025 / hour, and learning rate is... The batch size is set to 8, and the minimum consistency threshold is set to 0.2.

[0061] Setting a baseline control: All 393 data points were validated on a base model that had not been optimized and trained using this method. The validation results showed that only 179 judgments were correct, while the remaining 214 judgments included 106 judgments with ambiguous conclusions that could not pinpoint the root cause and 108 judgments that were incorrect. The overall error rate was approximately 0.54, which is far higher than the self-optimization trigger threshold of 0.2, thus triggering the self-optimization of the industrial vertical domain large model 100.

[0062] During the self-optimization process, initial execution strategies 200 were first generated for two typical cases: R1 (operational impact) and R3 (bearing failure), and their performance was verified. For case R1, the operating condition at that moment was a target load of 500MW (83% rated load), with the generator bearing housing temperature 11℃ higher than the normal operating temperature. The key influencing factor was the unit increasing load in response to grid dispatch requirements. After receiving the operating condition description, the base model automatically generated the initial strategy. The data collection scope included bearing housing temperature data, generator load adjustments over the past 24 hours, and start-stop operation records. Feature extraction methods included temperature rise curves and load adjustment curves. The reasoning steps were: "Is the temperature rise too fast? → Is there a load adjustment? → Is the load adjustment speed too fast? → True → R1; False → R3." The judgment thresholds were a bearing housing temperature rise greater than 10℃ and a load increase rate greater than 5MW / minute. Based on this initial strategy, the base model output the diagnosis "Load increase too fast, it is recommended to avoid sudden load changes," consistent with the expert diagnosis. The strategy achieved an accuracy rate of approximately 0.86 (190 / 221) on historical data samples, which is higher than the high-quality strategy threshold of 0.8. It is marked as a high-quality strategy 401, and the number of correct judgments (190) is greater than the number of times it has been solidified (20), so it can be solidified and used.

[0063] For case R3, the operating conditions at that moment were a target load of 450MW (75% rated load), generator shaft vibration of 0.13mm, bearing vibration of 0.07mm, and vibration increase of 0.015mm within 10 minutes. The initial strategy generated by the base model relied solely on vibration values ​​and the vibration spectrum for judgment, without referencing historical activity information such as operation tickets and work permits to investigate the causes. It could not distinguish between various possible root causes such as poor coupling alignment, bearing wear, dynamic imbalance, and loose foundation. The output diagnosis, "Generator vibration exceeds the standard; it is recommended to stop the machine and check the bearings and couplings, and perform dynamic balancing tests if necessary," only gave a general inspection conclusion without pointing out specific root causes, giving troubleshooting priorities, or referencing any operation tickets or maintenance records to assist in the judgment. The accuracy rate of this strategy on historical data samples was approximately 0.17 (4 / 23), which is lower than the threshold of 0.3 for the strategy to be optimized, and it is marked as strategy 402 to be optimized.

[0064] When constructing the training dataset, the strategy training samples were constructed by using the strategy to be optimized (402) as the previous strategy, the superior strategy (401), and the expert-designed judgment chain as the current strategy. The result training samples were constructed using the execution of the superior strategy (401) and the manually reviewed correct results. Due to the long time span of the dataset, some data collection times exceeded the adoption threshold of 48 hours. The median feedback delay, sorted in descending order of collection time, was 62 hours, corresponding to a timeliness weight of approximately 0.705 for the samples. This means that this part of the feedback only participated in the training with about 70% weight.

[0065] The input data in the dataset undergoes validity checks: For vibration sensors with a range of 0-50 mm / s, data exceeding this range, such as readings of 120 mm / s, are removed through physical constraint checks; for 600MW units, data requiring load increases exceeding the rated range, such as 95% load, are removed through process rule checks; cross-similarity matching is performed on all samples to remove data with similar operating conditions but contradictory conclusions. For example, in the R5 annotation case mentioned above, only 2 out of 30 similar operating condition records were annotated as R5, with a consistency ratio of approximately 0.067, which is lower than the minimum consistency threshold of 0.2, so this data is removed.

[0066] After mixing the two samples, LoRA incremental fine-tuning was performed on the large industrial vertical domain model 100, with a learning rate of [value missing]. The batch size was set to 8. After fine-tuning, the quality of the industrial vertical domain large model 100 was evaluated on 154 data points in the validation set. The strategy generated by the fine-tuned industrial vertical domain large model 100 includes feature extraction covering shaft vibration / bearing vibration peak value and rise rate, vibration spectrum (first and third harmonic component proportions), bearing housing temperature rise rate, coupling swing amplitude, generator load adjustment in the past 24 hours, start-stop operation records, abnormal generator noise records, coupling swing records, recent generator maintenance and bearing maintenance operation records, generator dynamic balance test reports, coupling alignment detection records, etc. The inference steps are arranged in the order of S1 checking operation impact → S2 checking coupling alignment → S3 checking bearing wear → S4 checking dynamic balance → S5 checking foundation looseness → comprehensive judgment, and specific trigger conditions and judgment thresholds are configured for each step. The new strategy establishes a clear mapping with the data source, the data referenced in each investigation step has a clear path, and physical inspection methods are given when data is missing or uncertain.

[0067] On the validation set of 154 data points, the overall error rate of the fine-tuned industrial vertical domain large model 100 was 19 / 154, approximately 0.12, which is significantly lower than the self-optimization trigger threshold of 0.2, reaching the standard for stable operation. This version was then solidified as the new baseline, and all subsequent incremental fine-tuning was based on the new baseline.

[0068] Compared to the base model without this method (overall error rate of approximately 0.54), the error rate of the industrial vertical domain large model 100 on the validation set decreased to approximately 0.12 after adopting this method, demonstrating a significant performance improvement. Simultaneously, the consistency, interpretability, and troubleshooting ability of the industrial vertical domain large model 100 output with expert diagnoses were significantly improved. This invention transforms the task execution strategy 200 from a manually static configuration into a dynamic object automatically generated by the industrial vertical domain large model 100 and continuously evolving based on feedback through the self-generation and feedback-driven evolution of the execution strategy 200. Through joint fine-tuning in the strategy-execution two-stage process, the strategy generation method is corrected while simultaneously correcting the results. A time-effect decay mechanism addresses the problem of outdated feedback interference caused by industrial feedback delays. A three-layer progressive feedback validity verification and quality monitoring automatic rollback mechanism ensures the safety and reliability of online self-optimization in industrial scenarios.

[0069] Example 6

[0070] As a sixth embodiment of the present invention, an industrial vertical domain large-scale model self-optimization learning system with self-evolutionary strategy is provided, including a strategy generation module, a strategy evolution module, a model fine-tuning module, and a verification management module.

[0071] The strategy generation module is used to obtain the working condition description and generate an execution strategy 200 in the form of structured text. The execution strategy 200 includes the data collection range, feature extraction method, inference step arrangement and judgment threshold.

[0072] The strategy evolution module is used to receive the execution strategy 200 and the corresponding feedback result 300, form a strategy-result mapping table, count the historical performance of each strategy template 400 and classify the strategy template 400, and update the execution strategy 200 according to the classification result.

[0073] The model fine-tuning module is used to construct training samples based on the corrected execution strategy 200 and the correct execution result 201, and to update the parameters of the industrial vertical domain large model 100.

[0074] The verification management module is used to perform validity verification on the feedback result 300 and monitor the quality indicators of the updated industrial vertical domain large model 100, and roll back to the historical stable version when the quality deteriorates.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A self-optimization learning method for large-scale industrial vertical models with self-evolutionary strategies, characterized in that, include: Obtain the current working condition description, input the working condition description into the industrial vertical domain large model (100), and generate an execution strategy (200) that matches the current working condition. The industrial task is executed based on the execution strategy (200), the execution result (201) is generated, the feedback result (300) corresponding to the execution result (201) is obtained, and the execution strategy (200) and the feedback result (300) are recorded to form a strategy-result mapping table. The historical performance of the statistical strategy template (400) is analyzed, and the strategy template (400) is classified according to the historical performance. The evolution of the execution strategy (200) is driven by the classification results and the feedback results (300).

2. The self-optimization learning method for large-scale industrial vertical models with strategy self-evolution as described in claim 1, characterized in that: The execution strategy (200) is parsed and mapped to a set of machine-executable parameters; The data acquisition range is mapped to the call parameters of the data acquisition interface, the feature extraction method is mapped to the selection of the signal processing function, the inference step arrangement is mapped to the task execution order, and the judgment threshold is mapped to the numerical comparison condition, so that the execution strategy (200) is directly transformed from a text description into an executable parameter sequence.

3. The self-optimization learning method for large-scale industrial vertical models with strategy self-evolution as described in claim 1 or 2, characterized in that: The formation strategy and result mapping table classifies and stores the execution strategies (200) according to the process sub-domains, and statistically analyzes the historical performance of each strategy template (400) according to the process sub-domains, so that the historical performance statistics of each strategy template (400) correspond to the working conditions of the process sub-domains.

4. The self-optimization learning method for large-scale industrial vertical models with self-evolutionary strategies as described in claim 3, characterized in that: The strategy template (400) is categorized based on its historical performance, including: Count the number of times each strategy template (400) received correct feedback and the total number of times it was used, and calculate the accuracy rate of each strategy template (400); The strategy template (400) whose accuracy reaches the first threshold is marked as a high-quality strategy (401). The policy template (400) with an accuracy rate less than or equal to the second threshold is marked as the policy to be optimized (402). The policy template (400) with an accuracy between the first threshold and the second threshold is marked as a candidate policy for continuing to accumulate feedback; Each strategy template (400) is classified into different evolutionary stages based on statistical results.

5. The self-optimization learning method for large-scale industrial vertical models with self-evolutionary strategies as described in claim 4, characterized in that: The evolution of the execution strategy (200) includes, According to the statistical error rate of the process sub-domain, when the error rate exceeds the trigger threshold, the optimization strategy (402) and the corresponding error result of the process sub-domain are used as training data to fine-tune the industrial vertical domain large model (100). When the number of times the strategy template (400) receives correct feedback reaches a set number, the strategy template (400) is marked as a stable strategy, and the stable strategy is used as a direct reuse strategy for subsequent execution of similar tasks.

6. The self-optimization learning method for large-scale industrial vertical models with self-evolutionary strategies as described in claim 5, characterized in that: Fine-tuning the large industrial vertical model (100) includes, Construct strategy training samples and result training samples. The strategy training samples take the current working condition description and the execution strategy (200) used in the previous step as input and the modified execution strategy (200) as the target output. The training samples take the current working condition description and the modified execution strategy (200) as input, and the correct execution result (201) as the target output; The policy training samples and the result training samples are jointly trained in the same batch, and the total loss is the weighted sum of the loss of the policy training samples and the loss of the result training samples.

7. The self-optimization learning method for large-scale industrial vertical models with self-evolutionary strategies as described in claim 6, characterized in that: In the joint training, the time weight of each training sample is determined based on the delay time of the feedback result (300); Feedback results (300) with a delay time within the adoption threshold are given full participation in training; The training samples corresponding to feedback results (300) whose delay time exceeds the discard threshold will not be used for training; The training samples corresponding to the feedback results (300) whose delay time is between the adoption threshold and the discard threshold participate in the training according to the decay weight, which decreases as the delay time increases.

8. The self-optimization learning method for large-scale industrial vertical models with strategy self-evolution as described in claim 6 or 7, characterized in that: It also includes performing physical constraint verification, process rule verification and similar working condition consistency verification on the feedback result (300) before the feedback result (300) is used for training; The physical constraint verification includes verifying whether the sensor data involved in the feedback result (300) is within the physical range, whether the data change rate is within the physical reasonable range, and whether the key sensor combination is self-consistent. The process rule verification includes verifying whether the output of the verification model is within the operating range specified by the process standard, whether the feedback result (300) contradicts the existing rules in the process knowledge base, and whether the quality inspection judgment result is consistent with the process quantity trend. The similar working condition consistency verification includes: retrieving historical feedback records under similar working conditions from the historical feedback database, counting the number of historical feedbacks that are consistent with the current feedback labeling conclusion and calculating the consistency ratio; when the consistency ratio is lower than the minimum consistency threshold, the current feedback is marked as questionable feedback. When the feedback conclusion is a continuous numerical type, calculate the percentile ranking of the current feedback conclusion in the set of historical feedback results (300) under similar working conditions. When the percentile ranking is in an extreme position range, mark the current feedback as questionable feedback. Only the feedback results (300) of the physical constraint verification, process rule verification and similar working condition consistency verification are used for training.

9. The self-optimization learning method for large-scale industrial vertical models with self-evolutionary strategies as described in claim 8, characterized in that: It also includes continuously monitoring the quality indicators of the industrial vertical domain large model (100) after the update; When any of the quality indicators deteriorates beyond the alarm threshold and continues to reach the confirmation time, the industrial vertical large model (100) will be rolled back to the most recent historical version with the best quality. When the updated version runs stably for a certain period of time, the version will be fixed as the new baseline, and subsequent incremental updates will be based on the new baseline.

10. A self-optimizing learning system for a large-scale industrial vertical model with self-evolutionary strategy, employing the self-optimizing learning method for a large-scale industrial vertical model with self-evolutionary strategy as described in any one of claims 1 to 9, characterized in that: It includes a strategy generation module, a strategy evolution module, a model fine-tuning module, and a verification management module; The strategy generation module is used to obtain the working condition description and generate an execution strategy (200) in the form of structured text. The execution strategy (200) includes the data collection range, feature extraction method, inference step arrangement and judgment threshold. The strategy evolution module is used to receive the execution strategy (200) and the corresponding feedback result (300), form a strategy-result mapping table, count the historical performance of each strategy template (400) and classify the strategy template (400), and update the execution strategy (200) according to the classification result. The model fine-tuning module is used to construct training samples based on the corrected execution strategy (200) and the correct execution result (201), and update the parameters of the industrial vertical domain large model (100); The verification management module is used to perform validity verification on the feedback result (300) and monitor the updated quality indicators of the industrial vertical domain large model (100), and roll back to the historical stable version when the quality deteriorates.

Images