State-driven multi-modal model autonomous evolution method for power grid AIOS

Abstract

The application discloses a state-driven multi-modal model autonomous evolution method for power grid AIOS. The application organizes seven kinds of information, including tasks, data, models, target basis, constraints and feedback, by the current model update state, so that the AIOS model autonomous evolution no longer depends on the fixed training process. The low-value samples which can be directly solved by language templates, historical frequency or regional background are eliminated through target basis effectiveness screening, so that the training resources are concentrated on high-value samples which must rely on target basis data to correctly judge. The model output explicitly references the operation field, time window segment, image area, log record or variable combination in the target basis data through reinforcement training and process record feedback, so as to improve the explainability of diagnosis and treatment suggestions. By including the release constraints and manual confirmation requirements into the current model update state, the safety risk caused by the direct online of the model autonomous evolution is reduced.

Description

Technical Field

[0001] This invention relates to the technical field of power grid artificial intelligence operating systems and intelligent analysis of power grid operation, specifically to an autonomous evolution method for state-driven multimodal models for power grid AIOS. Background Technology

[0002] With the advancement of new power system construction, the large-scale integration of distributed new energy sources, the popularization of online equipment monitoring technology, and the development of the power grid artificial intelligence operating system (AIOS), massive amounts of multi-source heterogeneous data are continuously generated at power production sites, covering various types such as operation measurements, synchronous phasors, equipment images, dispatch logs, fault work orders, historical cases, and manual handling feedback.

[0003] The aforementioned data includes continuous operating information such as voltage, current, load, renewable energy output, temperature, humidity, and electricity price, as well as business tags such as overload, transformer faults, stability margin, line defects, and handling results. Together, these data form the data foundation for model tasks in AIOS, such as anomaly identification, fault diagnosis, load forecasting, defect location, and handling suggestion generation.

[0004] Currently, multimodal models in power grid AIOS are typically deployed as relatively static inference components. Specifically, after pre-training, fine-tuning, or knowledge injection at the central system, model parameters are fixed and distributed to the business system, and are no longer updated based on changes in on-site data. This deployment method is adequate for routine recognition and question-answering tasks, but it has significant shortcomings when facing continuous changes in the actual operating environment:

[0005] When the load curve of a certain area shifts, the proportion of new energy output increases significantly, seasonal temperature and humidity fluctuate, equipment aging leads to changes in state distribution, sensor quality declines, or new defect samples appear for the first time, the model still relies on the historical statistical patterns, language templates, and regional background priors fixed in the old parameters for reasoning, and cannot transform the newly added field data into the model's judgment ability in a timely manner.

[0006] Furthermore, in the actual operation of the power grid AIOS, different model update tasks have significantly different requirements for data sources, data types, training actions, and deployment constraints. For cases with slight data distribution drift, supplementing with a small number of high-value samples and performing lightweight incremental training can meet the requirements. However, for application scenarios involving high-risk equipment fault diagnosis, cross-regional model deployment, or strict safety constraints, it is necessary to further combine multiple stages such as target data validity screening, data reinforcement training, training course reconstruction, model validation, canary release, and manual review. If a fixed preset training process is uniformly adopted, it will lead to a series of problems such as wasted training resources, an excessively high proportion of invalid samples, delayed model update response, lack of interpretable basis for model output, and uncontrollable release risks. Summary of the Invention

[0007] The technical problem this invention aims to solve is that existing power grid AIOS multimodal models struggle to dynamically determine subsequent update steps based on task objectives, equipment objects, model capabilities, available data, and release constraints. Furthermore, they lack the ability to perform process-level feedback evaluation and rule adjustments during the model update process. The goal is to provide a state-driven autonomous evolution method for power grid AIOS models, thereby resolving the aforementioned technical problems.

[0008] This invention is achieved through the following technical solution:

[0009] In a first aspect, the present invention provides an autonomous evolution method for a state-driven multimodal model of a power grid AIOS, comprising:

[0010] Receive model update tasks from the power grid artificial intelligence operating system (AIOS) and collect available data related to the model update tasks;

[0011] The model update task is analyzed to determine the task description information;

[0012] The available data is validated to obtain the target data; wherein, the target data is the data on which the model makes its judgments.

[0013] Based on the task description information and target data, the current model update status is generated;

[0014] Based on the current model update state, determine the next processing unit from the preset processing unit set;

[0015] Execute the next processing unit and obtain the processing result. Update the current model update status based on the processing result and continue to determine the subsequent processing units until the preset release conditions, preset stop conditions, or preset processing step limit are met.

[0016] Record each processing unit and its processing results to form a model update process record;

[0017] Model update results are generated based on the model update process record;

[0018] Feedback and evaluation are conducted based on the model update results and model update process records, and the selection rules for processing units in subsequent model update tasks are adjusted based on the feedback and evaluation results.

[0019] Furthermore, the available data includes one or more of the following: power grid operation measurement data, equipment status data, image inspection data, dispatch log data, historical fault case data, model inference records, and manual feedback data;

[0020] The power grid operation measurement data includes one or more of the following: voltage, current, active power consumption, reactive power, power factor, photovoltaic output, wind power output, grid power supply, voltage fluctuation rate, temperature, humidity, electricity price, and predicted load.

[0021] The equipment status data includes one or more of the following: overload status, transformer fault status, stable status, stability margin, response time, power injection, and coupling coefficient;

[0022] The image inspection data includes one or more of the following: transmission line images, insulator images, infrared images, defect area markings, and equipment appearance images.

[0023] Furthermore, the process of parsing the model update task to determine the task description information includes:

[0024] Extract the task objective, device object, model version, and business scenario from the model update task; wherein, the task objective is used to characterize the type of capability to be improved in this model update, the device object is used to characterize the type of power grid equipment or regional node involved in this model update, the model version is used to record the parameter summary and current capability indicators of the model to be updated, and the business scenario is used to characterize the application conditions in which this model update is performed;

[0025] The extracted task objective, device object, model version, and business scenario are organized into the task description information.

[0026] Furthermore, the step of validating the available data to obtain the target data includes:

[0027] The factual output is obtained based on the complete multimodal input, the task description reference output is obtained based on the task description baseline input, and the basis occlusion reference output is obtained based on the occlusion candidate basis input; wherein, the task description baseline input includes: task objective, problem description, device object, business scenario and output format constraints, but does not include the field data in the available data;

[0028] Based on the output differences between the fact output, the task description reference output, and the basis occlusion reference output, candidate basis whose output differences satisfy preset conditions are determined as the target basis data;

[0029] The target data includes one or more of the following: running fields, time window segments, image regions, log records, case records, human feedback records, and combinations of variables.

[0030] Furthermore, the step of generating the current model update state based on task description information and target basis data includes:

[0031] Based on the task description information and target basis data, the current model update status is generated; wherein, the current model update status includes: task status field, data status field, model status field, basis status field, uncertainty field, constraint status field and feedback status field;

[0032] The task status field is used to record the task objective, device object, and business scenario.

[0033] The data status field is used to record the sample source, modality type, data quality, regional sample pool, and proportion of newly added samples;

[0034] The model status field is used to record the model version, model parameter summary, capability evaluation results, and capability items to be optimized; the basis status field is used to record the target basis data.

[0035] The uncertainty field is used to record diagnostic confidence, reward variance, basis missing items, and items requiring manual review.

[0036] The constraint status field is used to record security constraints, release constraints, computing resource constraints, data compliance constraints, and manual confirmation requirements;

[0037] The feedback status field is used to record manual adoption status, operation and maintenance handling results, model rollback results, and subsequent business performance.

[0038] Furthermore, the preset processing unit set includes:

[0039] The system comprises a data governance unit, a validity-based screening unit, a validity-based reinforcement unit, a curriculum restructuring unit, an incremental training unit, a model validation unit, a release control unit, and a feedback write-back unit.

[0040] The data governance unit is used to align and control the quality of multi-source power grid data.

[0041] The validity screening unit is used to screen high-value samples that require model judgment based on target data.

[0042] The basis reinforcement unit is used to construct training samples including input samples, target output labels and target basis labels, and to jointly train the model through target output loss, basis reference loss and basis consistency reward, so that the model outputs reference target basis data.

[0043] The course reconstruction unit is used to organize training courses according to task type, region of origin, device type, modality type, sample difficulty, and target basis type.

[0044] The incremental training unit is used to update adapter parameters, fusion layer parameters, output head parameters, or retrieve enhancement module parameters when the base encoder is frozen or the backbone parameters are shared, and to maintain old knowledge through historical sample playback, distillation constraints, or parameter regularization.

[0045] The model verification unit is used to verify the capabilities and security of the updated model.

[0046] The release control unit is used to determine whether the model is released, grayscale, rolled back, or manually reviewed.

[0047] The feedback write-back unit is used to write business feedback into the AIOS model update context.

[0048] Furthermore, based on the current model update state, the next processing unit is determined from the preset processing unit set, including:

[0049] Based on the current model update status, the fitness score of each processing unit is calculated; wherein, the fitness score is determined by weighting the necessity of the step, the update benefit, the execution cost, the constraint risk and the historical feedback correction item.

[0050] The processing unit with the highest fit score is determined as the next processing unit;

[0051] The historical feedback correction item is obtained by statistically analyzing the output quality improvement of each processing unit in the same task, the hit rate improvement, the verification pass status, the average execution cost, and the rollback record, and is used to improve or reduce the adaptability score of the corresponding processing unit in subsequent similar tasks.

[0052] Secondly, the present invention provides a state-driven multimodal model auto-evolution device for power grid AIOS, comprising:

[0053] Receiving module: Receives model update tasks from the power grid artificial intelligence operating system (AIOS) and collects available data related to the model update tasks;

[0054] Parsing module: Parses the model update task to determine the task description information;

[0055] Validation module: Performs validity verification on the available data to obtain target basis data; wherein, the target basis data is the data on which the model makes its judgments;

[0056] First generation module: Generates the current model update status based on task description information and target data;

[0057] Determine module: Based on the current model update state, determine the next processing unit from the preset processing unit set;

[0058] Update module: Execute the next processing unit and obtain the processing result, update the current model update status according to the processing result, and continue to determine the subsequent processing units until the preset release conditions, preset stop conditions, or preset processing step limit are met;

[0059] Recording module: Records each processing unit and its processing results, forming a model update process record;

[0060] The second generation module generates model update results based on the model update process records.

[0061] Feedback and Evaluation Module: Provides feedback and evaluation based on the model update results and model update process records, and adjusts the processing unit selection rules in subsequent model update tasks based on the feedback and evaluation results.

[0062] Thirdly, the present invention provides an electronic device, comprising: a memory, and one or more processors communicatively connected to the memory; the memory stores instructions executable by the one or more processors, the instructions being executed by the one or more processors to cause the one or more processors to implement the method described above.

[0063] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0064] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0065] This invention unifies the organization of seven categories of information—task, data, model, target basis, constraints, and feedback—by utilizing the current model update status, freeing the AIOS model's autonomous evolution from dependence on a fixed training process. By filtering and eliminating low-value samples that can be directly solved using language templates, historical frequencies, or regional background through the validity of target basis data, training resources are concentrated on high-value samples that require target basis data for accurate judgment. Through reinforcement training based on data and feedback from process records, the model output explicitly references runtime fields, time windows, image regions, log records, or variable combinations from the target basis data, improving the interpretability of diagnostic and treatment recommendations. By incorporating release constraints and manual confirmation requirements into the current model update status, the safety risks associated with directly deploying the model's autonomous evolution are reduced, ensuring that model releases meet the safety and compliance requirements of power production. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0067] Figure 1 This is one of the flowcharts for a state-driven multimodal model autoevolution method for power grid AIOS provided in the embodiments of this specification;

[0068] Figure 2 This is the second flowchart of a state-driven multimodal model autoevolution method for power grid AIOS provided in the embodiments of this specification;

[0069] Figure 3 This is the third flowchart of a state-driven multimodal model autoevolution method for power grid AIOS provided in the embodiments of this specification;

[0070] Figure 4 This is a block diagram of an electronic device provided in the embodiments of this specification. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0072] like Figures 1-3 As shown, this embodiment provides an autonomous evolution method for a state-driven multimodal model of a power grid AIOS, including:

[0073] Step S11: Receive the model update task from the power grid artificial intelligence operating system (AIOS) and collect the available data related to the model update task.

[0074] In this embodiment, the model update task is triggered in AIOS and is a task that updates or adapts the capabilities of the deployed multimodal model. The multimodal model is a machine learning model, specifically a deep learning-based multimodal neural network model. This model includes a basic encoder, a fusion layer, and an output head, used to fuse power grid operation measurement data, equipment status data, and image inspection data and output corresponding judgment results.

[0075] Specifically, the trigger source may include:

[0076] AIOS detected a continuous decline in the confidence level of the model output in a certain area and triggered an update request.

[0077] Maintenance personnel initiate a model update request in the work order system;

[0078] Automatically triggered when new equipment or new fault type samples reach a preset cumulative threshold;

[0079] This is triggered when the data distribution difference between the target region and the source region exceeds a preset offset threshold in a cross-regional deployment scenario.

[0080] In this embodiment, the model update task is transmitted to AIOS in the form of a structured task message, which includes fields such as task type identifier, trigger source, region identifier, and device identifier. AIOS determines the device object, region range, and time window corresponding to this model update task based on the above fields, and collects the related multi-source multimodal data.

[0081] Specifically, AIOS pulls the following seven types of data from the data platform, scheduling system, inspection system, work order system, and manual feedback interface: power grid operation measurement data, equipment status data, image inspection data, scheduling log data, historical fault case data, model inference records, and manual feedback data.

[0082] More specifically, the power grid operation measurement data may include: voltage, current, active power consumption, reactive power, power factor, photovoltaic output, wind power output, grid power supply, voltage fluctuation rate, temperature, humidity, electricity price, and predicted load.

[0083] The equipment status data may include: overload status, transformer fault status, stable status, stability margin, response time, power injection, and coupling coefficient.

[0084] The image inspection data may include: images of transmission lines, images of insulators, infrared images, defect area markings, and images of equipment appearance.

[0085] The scheduling log data may include: scheduling operation records, alarm records, and event sequence records.

[0086] The historical fault case data may include: fault type, fault cause, handling measures, and handling results.

[0087] The model inference record may include: the historical inference inputs, inference outputs, and confidence levels of the model to be updated on the target device and target region.

[0088] Manual feedback data can include: annotations of the adoption of model inference results by operations and maintenance personnel, correction suggestions, and false alarm / missed alarm markers.

[0089] In this embodiment, the aggregated available data can be initially collected according to device identifier and timestamp to form an original data set associated with the model update task.

[0090] Step S12: Analyze the model update task and determine the task description information.

[0091] In this embodiment, AIOS can extract task description information from the structured messages of the model update task.

[0092] Specifically, the task description information may include four fields: task objective, device object, model version, and business scenario.

[0093] More specifically, the task objective characterizes the type of capability to be improved in this model update, which may include improved anomaly identification accuracy, improved fault diagnosis accuracy, reduced load prediction error, improved defect location capability, or improved data interpretation capability. The value of the task objective is mapped from the task type identifier in the model update task message.

[0094] The device object is used to characterize the type of power grid equipment or regional node involved in this model update, specifically it can be a transformer, transmission line, switch cabinet, distribution substation or regional node.

[0095] The model version is used to record the parameter summary and current capability metrics of the model to be updated; the parameter summary is represented by model parameter hash values ​​or parameter statistics. Current capability metrics may include one or more of the following metrics: task accuracy on target devices and target areas, fault identification rate, load prediction error, defect location mAP, and hit rate. The model version value can be obtained from the AIOS model registry.

[0096] The business scenario is used to characterize the application conditions under which this model update takes place. Specifically, it can be an online monitoring scenario, a fault handling scenario, a scheduling assistance scenario, an inspection and identification scenario, or a cross-regional adaptation scenario.

[0097] Step S13: Verify the validity of the available data to obtain the target basis data; wherein, the target basis data is the data on which the model makes its judgment.

[0098] In this embodiment, the target basis data refers to data that substantially contributes to the model's ability to identify anomalies, diagnose faults, predict loads, or identify defects. For any candidate basis, if the change in the model's output after masking the candidate basis meets a preset condition, and the model can obtain the expected judgment result when the candidate basis is retained, then the candidate basis is determined as the target basis data.

[0099] Specifically, AIOS can verify the validity of evidence through a three-input comparison mechanism.

[0100] In one possible and specific implementation, firstly, a complete multimodal input is constructed, which includes all available data collected in step S11, such as operational measurements, device status, images, logs, cases, inference records, and human feedback data corresponding to the current sample. The complete multimodal input is then fed into the model to be updated to obtain the factual output.

[0101] Next, a task description baseline input is constructed, which includes: task objectives, problem description, equipment objects, business scenarios, and output format constraints. It excludes on-site data such as operational measurements, equipment status values, image content, log text, historical case details, and human feedback conclusions. The task description baseline input is then fed into the model to be updated to obtain a task description reference output.

[0102] Next, the input for masking candidate criteria is constructed, and the candidate criteria data in the complete multimodal input are masked item by item (the granularity of candidate criteria includes a single running field, continuous time window, image patch or detection box, variable combination, log fragment and case fragment). The masked input is then fed into the model to be updated to obtain the criterion masking reference output.

[0103] Next, AIOS determines the role of the occluded candidate evidence in the model's judgment based on the output differences between the factual output, the task description reference output, and the occlusion reference output.

[0104] More specifically, for classification tasks, output differences are determined by the decrease in confidence level of the target category or the change in classification results.

[0105] For regression or load forecasting tasks, output discrepancies are determined by either the absolute error increment or the mean square error increment.

[0106] For detection or localization tasks, output differences are determined by changes in mAP, IoU, or localization box offset.

[0107] When the difference between the fact output and the reference output after masking a candidate basis meets a preset threshold condition, the candidate basis is determined as the target basis data.

[0108] The specific form of the target data may include: operational fields (such as voltage fluctuation rate, power factor), time window segments (such as continuous measurement sequences within a preset time before the fault), image regions (such as conductor strand breakage areas, insulator damage areas, infrared anomaly areas), log records (such as alarm records, scheduling operation records), case records (such as fault causes and handling measures in historical fault cases), manual feedback records (such as correction opinions and false alarm markers from maintenance personnel), and variable combinations (such as the combined relationship between response time, power injection, coupling coefficient, and stability margin).

[0109] Step S14: Generate the current model update status based on the task description information and target data.

[0110] In this embodiment, the current model update state is used to characterize the real-time processing conditions of the current model update task and serves as the basis for determining the next processing action. The current model update state includes a structured state object and a numerical state vector generated based on the structured state object.

[0111] In one possible and specific implementation, AIOS generates structured state objects in the following manner:

[0112] Extract the task objective, device object, and business scenario from the task description information and populate the task status field.

[0113] From the available data collected in step S11, determine the sample source, modality type, data quality, regional sample pool, and proportion of new samples, and populate the data status field. Data quality is calculated by a weighted sum of field completeness rate, cross-modal time or device alignment rate, valid value proportion, and data timeliness score.

[0114] Determine the model parameter summary, capability evaluation results, and capabilities to be optimized from the model version in the task description information, and populate the model status field.

[0115] Fill the target data obtained in step S13 into the basis status field, and record the field identifier, time window index, image region coordinates, log record identifier, case record identifier and variable combination identifier corresponding to the target data.

[0116] The uncertainty field is populated with diagnostic confidence, reward variance, missing basis, and items requiring manual review. Diagnostic confidence can be extracted from the inference output of the model to be updated on the target samples. Reward variance can be statistically obtained from the periodic records of the model to be updated on the probe sample set. Missing basis can be determined by fields or regions in the model output that are missing references, as identified during basis validity verification. Items requiring manual review can be selected from samples with basis consistency scores below a preset threshold.

[0117] The constraint status field is populated with security constraints, release constraints, computing resource constraints, data compliance constraints, and manual confirmation requirements. Security constraints are mapped from the power grid safety operation procedures. Release constraints are read from the AIOS model release management strategy. Computing resource constraints are determined by obtaining currently available computing and storage resources from the AIOS resource scheduler. Data compliance constraints are read from the data governance strategy. Manual confirmation requirements are determined from the online approval rules corresponding to the business scenario.

[0118] The feedback status field is populated with information on manual adoption, operational handling results, model rollback results, and subsequent business performance. Specifically, the manual adoption status, operational handling results, model rollback results, and subsequent business performance are obtained by querying records related to the current device and region from the AIOS feedback record database.

[0119] After the data is populated, AIOS organizes the above seven fields into a structured state object.

[0120] In one possible and specific implementation, AIOS obtains the numerical state vector from the structured state object mapping in the following way:

[0121] The data quality score is mapped by weighting and summing the field completeness rate, cross-modal time or device alignment rate, valid value ratio, and data timeliness score in the data status field.

[0122] The verification results of the target criteria data in the status field are mapped to the target criteria validity score. This score is obtained by weighted summation of the output evaluation change after masking candidate criteria and the output evaluation change when only the task text is retained.

[0123] Map the capabilities to be optimized and the capability evaluation results in the model status field to the degree to which the model needs improvement.

[0124] Map the diagnostic confidence, reward variance, and pending review items in the uncertainty field to uncertainty scores.

[0125] Map the security constraints, compliance constraints, and release constraints in the constraint status field to the online risk score.

[0126] The manual adoption status, operation and maintenance handling results, and subsequent business performance in the feedback status field are mapped to the feedback credibility score F.

[0127] The numerical state vector S = [data quality score, target basis validity score, model improvement level, uncertainty score, launch risk score, feedback credibility score] is used for the adaptation score calculation and automatic scheduling of subsequent processing units.

[0128] Step S15: Based on the current model update state, determine the next processing unit from the preset processing unit set.

[0129] In this embodiment, the preset processing unit set includes multiple callable processing units pre-configured in AIOS.

[0130] Specifically, the set of preset processing units may include a data governance unit, a validity-based filtering unit, a reinforcement unit, a course reconstruction unit, an incremental training unit, a model verification unit, a release control unit, and a feedback write-back unit.

[0131] Each type of processing unit encapsulates its own execution logic for data processing, model training, or validation and deployment. AIOS can call any of these units by passing the current model update status and available data.

[0132] AIOS can calculate the fitness score of each processing unit based on the current model update state.

[0133] Specifically, for the numerical state vector, AIOS first matches each component in the numerical state vector with the triggering rules of each processing unit to calculate the necessity of the steps. Then, it estimates the update benefit of each processing unit based on the model improvement level and data quality score in the numerical state vector. Next, it assesses the constraint risk of each processing unit based on the ceiling risk score and constraint status in the numerical state vector. Finally, it retrieves historical feedback correction items for each processing unit from the AIOS feedback record library for similar historical tasks.

[0134] AIOS adds step necessity, update benefits, and historical feedback correction items as supplementary items, and subtracts execution costs and constraint risks by taking their inverse values. The weighted sum is then used to obtain the fit score for each processing unit. The fit scores of each processing unit are compared, and the processing unit with the highest fit score is selected as the next processing unit.

[0135] Step S16: Execute the next processing unit and obtain the processing result. Update the current model update status according to the processing result, and continue to determine the subsequent processing units until the preset release conditions, preset stop conditions, or preset processing step limit are met.

[0136] In this embodiment, the current model update state and available data can be passed as input parameters to the processing unit. After executing its encapsulated internal logic, the processing unit outputs the processing result.

[0137] Specifically, the data governance unit outputs cleaned and aligned sample data, the validity screening unit outputs a list of retained high-value samples and removed low-value samples, the reinforcement unit outputs the update amount based on reward results and model parameters, the course reconstruction unit outputs the reconstructed training course configuration, the incremental training unit outputs the incremental parameter update amount, the model validation unit outputs a validation report, the release control unit outputs the gray-scale release results or manual review comments, and the feedback write-back unit outputs context update confirmation.

[0138] AIOS updates the current model update status based on the processing results.

[0139] In one possible and specific implementation, changes in data quality, sample composition, model parameters, state, uncertainty, or feedback involved in the processing results are written into the corresponding fields of the structured state object, and the updated numerical state vector is obtained by remapping from the updated structured state object. The updated current model update state serves as the input for selecting the next processing step.

[0140] In this embodiment, AIOS continues to determine the subsequent processing unit based on the updated current model update state, and repeats the above-mentioned loop process of calling, updating and reselecting.

[0141] The loop terminates when any of the following conditions are met:

[0142] The preset release conditions must be met, that is, the model capability evaluation results must reach the preset capability threshold, the security constraint verification must be passed, the data compliance constraint verification must be passed, and there must be no items pending manual review.

[0143] The preset stopping conditions are met, that is, after multiple consecutive iterations, the model's performance indicators no longer improve and the reward variance has stabilized or all high-value samples in the currently available data have been utilized.

[0144] Reaching the preset processing step limit: The number of loops reaches the preset maximum processing step threshold.

[0145] The three termination conditions mentioned above are checked by AIOS at the beginning of each loop. If any condition is met, the loop stops and proceeds to step S17.

[0146] Step S17: Record each processing unit and its processing results to form a model update process record.

[0147] In this embodiment, AIOS generates a corresponding process record entry after each processing unit completes execution. The process record entry includes a processing unit identifier, execution sequence number, call timestamp, input state snapshot, output result summary, and execution time, used to record the call order of each processing unit, execution results, and the changes in the current model update state. The process record entry can be stored in key-value pair format.

[0148] After the loop terminates, AIOS arranges all process record entries according to the processing unit number to form a complete model update process record for this model update.

[0149] Step S18: Generate model update results based on the model update process record.

[0150] In one possible and specific implementation, AIOS summarizes and extracts the following information from the model update process records to generate model update results.

[0151] Sample pool update results: Extract statistical data on sample addition, sample removal, and sample label update from the output summary of each processing unit in the process record, and summarize them to describe the changes in the regional sample pool after this model update.

[0152] Model parameter increment: Extract the model parameter increment from the process record entry corresponding to the incremental training unit, and record the parameter summary comparison before and after the update.

[0153] Model capability evaluation results: Extract capability indicators such as task accuracy, fault identification rate, load prediction error, defect location mAP, and basis hit rate from the process record entries corresponding to the model verification unit in the verification report.

[0154] Based on the interpretation results: the reference information of the target basis data and the change trajectory of the basis consistency score are extracted from the output result summary of each processing unit in the process record, and summarized as an explanation of the improvement of the model basis interpretation capability of this model update.

[0155] Model release recommendation: Extract the release recommendation type from the process record entry corresponding to the release control unit. The release recommendation type is one of the following: direct release, canary release, continue training, manual review, or model rollback.

[0156] Step S19: Provide feedback and evaluation based on the model update results and model update process records, and adjust the processing unit selection rules in subsequent model update tasks based on the feedback and evaluation results.

[0157] In this embodiment, AIOS provides feedback evaluation based on model update results and model update process records. The feedback evaluation comprehensively considers model output quality, consistency of objective criteria, model update benefits, task response costs, constraint satisfaction, and execution failures. Specifically, model output quality is determined based on changes in task accuracy or fault identification rate of the model on the target device and target region before and after the update.

[0158] The consistency score is calculated based on the consistency evaluation value, which is determined by comparing the target references in the updated model output with the manually confirmed references.

[0159] The updated benefit evaluation value is calculated based on the sample utilization efficiency and the improvement in model capability in this update.

[0160] The response cost evaluation value is calculated based on the total number of training steps, total training latency, cost of calling each processing unit, and release waiting time from step S15 to step S18 of this model update, after normalization.

[0161] The constraint satisfaction evaluation value is determined based on whether the current model update meets security constraints, data compliance constraints, model release constraints, and manual review requirements. The maximum value is taken when all constraints are met; if any constraint is violated, the value is deducted according to the severity of the violation.

[0162] The failure penalty evaluation value is assigned when the model update triggers a model rollback, a security event occurs after release, or the operation and maintenance handling results are abnormal.

[0163] In one possible and specific implementation, AIOS can determine historical feedback correction items for each processing unit based on feedback evaluation results. These historical feedback correction items are statistically derived from improvements in output quality, basis hit rate, validation pass rate, average execution cost, and rollback records for each processing unit within the same type of task. A processing unit's historical feedback correction item is positive when its average output quality and basis hit rate improvements are positive, its validation pass rate is high, and its rollback rate is low within the same type of historical task. Conversely, a processing unit's historical feedback correction item is negative when its average execution cost is high, its validation pass rate is low, or its rollback rate is high within the same type of historical task.

[0164] AIOS writes the updated historical feedback corrections into the corresponding processing unit and task type record entries in the feedback record library.

[0165] In one possible and specific implementation, AIOS uses historical feedback correction items to assess the suitability of subsequent similar tasks. When a historical feedback correction item is positive, the suitability score of the corresponding processing unit is increased; when a historical feedback correction item is negative, the suitability score of the corresponding processing unit is decreased, thereby adjusting the priority of that processing unit in being identified as the next processing unit.

[0166] Through the aforementioned feedback and adjustment mechanism, AIOS continuously optimizes the selection strategy of processing units during the execution of multiple model update tasks.

[0167] In some implementations, the available data includes one or more of the following: power grid operation measurement data, equipment status data, image inspection data, dispatch log data, historical fault case data, model inference records, and manual feedback data;

[0168] The power grid operation measurement data includes one or more of the following: voltage, current, active power consumption, reactive power, power factor, photovoltaic output, wind power output, grid power supply, voltage fluctuation rate, temperature, humidity, electricity price, and predicted load.

[0169] The equipment status data includes one or more of the following: overload status, transformer fault status, stable status, stability margin, response time, power injection, and coupling coefficient;

[0170] The image inspection data includes one or more of the following: transmission line images, insulator images, infrared images, defect area markings, and equipment appearance images.

[0171] In some implementations, parsing the model update task to determine task description information includes:

[0172] Step S121: Extract the task objective, device object, model version, and business scenario from the model update task; wherein, the task objective is used to characterize the type of capability to be improved in this model update, the device object is used to characterize the type of power grid equipment or regional node involved in this model update, the model version is used to record the parameter summary and current capability indicators of the model to be updated, and the business scenario is used to characterize the application conditions under which this model update is performed.

[0173] In this embodiment, the model update task can be passed to AIOS in the form of a structured task message. In addition to the task type identifier, trigger source, associated region identifier, and device identifier fields, the task message may also include the target model identifier field and the deployment environment identifier field.

[0174] In one possible and specific implementation, when extracting task objectives, AIOS reads the task type identifier field, queries a preset mapping table between task type identifiers and task objectives, and converts the task type identifier into the corresponding task objective. Specifically, the anomaly detection model updates the task mapping to improve anomaly detection accuracy; the fault diagnosis model updates the task mapping to improve fault diagnosis accuracy; the load forecasting model updates the task mapping to reduce load forecasting error; the defect detection model updates the task mapping to improve defect location capability; and the cross-regional model adaptation task mapping is based on improved explanatory power.

[0175] When extracting a device object, AIOS reads the associated device identifier field and the associated region identifier field. If the associated device identifier field is not empty, AIOS queries the AIOS model service platform for the device type corresponding to that identifier. The device type is one of transformer, transmission line, switchgear, or distribution substation. AIOS identifies this device type as the device object. If the associated device identifier field is empty but the associated region identifier field is not empty, AIOS identifies the value of the associated region identifier field as the device object. In this case, the value of the device object is the region node identifier of the target region.

[0176] When extracting the model version, AIOS uses the target model identifier field as the query key to retrieve the registration information of the model to be updated from the AIOS model service platform. It then combines the model parameter summary and current capability indicators from the registration information to form the model version. The model parameter summary is either the hash value or parameter statistics of the model parameters, and the parameter statistics are one of the following: parameter mean, parameter variance, or parameter norm.

[0177] When extracting business scenarios, AIOS reads the trigger source field and deployment environment identifier field, and converts them into business scenarios according to a preset mapping rule table. When the trigger source is an online monitoring alarm, it is mapped to an online monitoring scenario; when the trigger source is a fault work order, it is mapped to a fault handling scenario; when the trigger source is a scheduling request, it is mapped to a scheduling assistance scenario; when the trigger source is an inspection task, it is mapped to an inspection identification scenario; and when the deployment environment identifier is inconsistent with the original deployment area of ​​the target model, it is mapped to a cross-region adaptation scenario.

[0178] Step S122: Organize the extracted task objective, device object, model version, and business scenario into the task description information.

[0179] In this embodiment, AIOS writes the task objective, device object, model version, and business scenario extracted in step S121 into the task description information according to a preset field mapping relationship. The task description information can be stored in key-value pair format and used as input for the validity verification in step S13 and for generating the current model update status in step S14.

[0180] In some implementations, the step of validating the available data to obtain the target data includes:

[0181] Step S131: Obtain factual output based on complete multimodal input, obtain task description reference output based on task description baseline input, and obtain basis occlusion reference output based on occlusion candidate basis input; wherein, the task description baseline input includes: task objective, problem description, device object, business scenario and output format constraints, and does not include field data in the available data.

[0182] In this embodiment, AIOS can generate first-level single-granularity candidate criteria from the available data collected in step S11.

[0183] Specifically, individual operational fields are extracted one by one; continuous time windows slide across the measurement time series with preset steps; image blocks or detection boxes are segmented or extracted from inspection images according to a grid; variable combinations are selected from combinations of values ​​of two or more variables from equipment status data; log segments are extracted from scheduling logs according to event boundaries; case segments are extracted from historical fault cases according to the handling stage; and manual feedback segments are extracted from manual feedback data according to feedback items. Each candidate criterion is assigned a unique identifier.

[0184] In this embodiment, AIOS can construct a set of three inputs for each candidate criterion: a complete multimodal input, a task description baseline input, and a masking input. The masking input is obtained by setting the portion of the complete multimodal input corresponding to the candidate criterion to zero, the mean, or a mask label.

[0185] AIOS sequentially feeds the three inputs into the model to be updated, obtaining the factual output, the task description reference output, and the occlusion reference output, respectively.

[0186] Step S132: Based on the output difference between the fact output, the task description reference output, and the basis occlusion reference output, determine the candidate basis that satisfies the preset conditions as the target basis data;

[0187] The target data includes one or more of the following: running fields, time window segments, image regions, log records, case records, human feedback records, and combinations of variables.

[0188] In this embodiment, AIOS can calculate the output difference value of each candidate criterion and compare it with a preset threshold, which can be determined by AIOS based on task type, device object and historical verification statistics.

[0189] When the output difference value of a candidate criterion is lower than a preset threshold, it means that the model can still maintain the output by relying on language prior, historical frequency or regional background after the candidate criterion is masked, and the candidate criterion is excluded.

[0190] When the output difference value is not less than the preset threshold, it indicates that the candidate basis has a significant contribution to the current model judgment, and the candidate basis is determined as the target basis data.

[0191] In this embodiment, AIOS can combine the top-ranked candidate criteria based on output difference values ​​in two ways: same-type combination and cross-modal combination.

[0192] Specifically, the same-type combination merges multiple candidate criteria under the same modality type into one combined candidate criteria, and the cross-modal combination merges candidate criteria under different modality types into one combined candidate criteria. The combined candidate criteria are verified again according to the three-input construction method of step S131 and the output difference calculation method of this step. Combined candidate criteria whose output difference values ​​meet the preset threshold are also determined as target criteria data.

[0193] In this embodiment, AIOS classifies the target data according to data granularity and data source. The classification results include running fields, time window segments, image regions, log records, case records, human feedback records, and variable combinations.

[0194] In some implementations, generating the current model update state based on task description information and target basis data includes:

[0195] Step S141: Generate the current model update status based on the task description information and the target basis data; wherein, the current model update status includes: task status field, data status field, model status field, basis status field, uncertainty field, constraint status field and feedback status field;

[0196] The task status field is used to record the task objective, device object, and business scenario.

[0197] The data status field is used to record the sample source, modality type, data quality, regional sample pool, and proportion of newly added samples;

[0198] The model status field is used to record the model version, model parameter summary, capability evaluation results, and capability items to be optimized; the basis status field is used to record the target basis data.

[0199] The uncertainty field is used to record diagnostic confidence, reward variance, basis missing items, and items requiring manual review.

[0200] The constraint status field is used to record security constraints, release constraints, computing resource constraints, data compliance constraints, and manual confirmation requirements;

[0201] The feedback status field is used to record manual adoption status, operation and maintenance handling results, model rollback results, and subsequent business performance.

[0202] In this embodiment, AIOS takes the task description information output in step S12 and the target basis data output in step S13 as input, updates the task state, data state, model state, basis state, uncertainty state, constraint state and feedback state, and generates a structured state object and corresponding numerical state vector accordingly.

[0203] In some embodiments, the preset processing unit set includes:

[0204] The system comprises a data governance unit, a validity-based screening unit, a validity-based reinforcement unit, a curriculum restructuring unit, an incremental training unit, a model validation unit, a release control unit, and a feedback write-back unit.

[0205] The data governance unit is used to align and control the quality of multi-source power grid data.

[0206] In this embodiment, the internal execution process of the data governance unit includes three sub-steps: outlier detection and processing, time window alignment, and modality alignment.

[0207] Specifically, in the outlier detection and processing sub-step, AIOS performs distribution statistics on each field of the power grid operation measurement data, calculates the mean and standard deviation of each field, marks values ​​that deviate from the mean by more than three times the standard deviation as outliers, and replaces them with the median of that field in adjacent time windows.

[0208] For enumerated fields in device status data, check whether the values ​​are within the preset set of valid values. Values ​​not in the set are marked as invalid and removed from the sample.

[0209] In the time window alignment sub-step, AIOS uses the device identifier as the primary key and the sampling timestamp as the sorting key to perform time window alignment on the time series data of the same device from different data sources.

[0210] For data sources with inconsistent sampling frequencies, the data source with the highest sampling frequency is used as the benchmark, and other data sources are subjected to linear interpolation or nearest neighbor filling to align the data points of each data source to a unified sampling time on the time axis.

[0211] In the modal alignment sub-step, AIOS associates image samples and measurement samples acquired by the same device within the same time period into cross-modal sample pairs based on the correspondence between image acquisition timestamps and measurement data timestamps. Furthermore, based on event timestamps and device identifiers, it associates two types of unstructured text data—scheduling logs and historical fault cases—with the corresponding time periods' measurement and image data. After data governance is completed, AIOS updates the data quality score in the data status field.

[0212] The validity screening unit is used to screen high-value samples that require model judgment based on target data.

[0213] In this embodiment, the validity screening unit performs batch screening of samples based on the three-input comparison verification.

[0214] AIOS calculates the output difference value for each sample. When the output difference value reaches a preset threshold, the corresponding sample is identified as a high-value sample and retained in the subsequent training process; when the output difference value does not reach the preset threshold, the corresponding sample is identified as a low-value sample and removed from the subsequent training samples.

[0215] After the screening is completed, AIOS updates the proportion of new samples and the composition of the regional sample pool in the data status field according to the sample retention results.

[0216] The basis reinforcement unit is used to construct training samples including input samples, target output labels and target basis labels, and to jointly train the model through target output loss, basis reference loss and basis consistency reward, so that the model outputs reference target basis data.

[0217] In this embodiment, the internal execution process of the reinforcement unit may specifically include two sub-steps: training sample construction and joint training.

[0218] More specifically, in the training sample construction sub-step, AIOS extracts target basis labels from the target basis data determined in step S13, and the target basis labels are represented in the form of structured annotations.

[0219] The `Run Field` class records the field identifier and the field value range; the `Time Window Segment` class records the start and end indexes of the time window; the `Image Region` class records the coordinates of the top left and bottom right corners of the region; the `Log Record` class records the log record identifier; the `Case Record` class records the case record identifier; the `Human Feedback Record` class records the feedback item identifier; and the `Variable Combination` class records the combination of field identifiers for each variable.

[0220] In the joint training sub-step, AIOS uses a weighted sum of the target output loss, citation loss, and citation consistency reward as the total loss to update the model parameters using gradients. Specifically, the target output loss is calculated from the cross-entropy or mean squared error between the model's predicted output on the input samples and the target output label; the citation loss is calculated from the difference between the citation identifier in the model output and the target citation label; and the citation consistency reward is converted from the consistency score between the citation in the model output and the manually verified citation.

[0221] During training, AIOS periodically records the accuracy rate and consistency score. When the accuracy rate, consistency score, or manually confirmed matching degree of the criteria reach the preset threshold, and the task output quality is not lower than the preset lower limit, the criteria enhancement is considered successful.

[0222] The course reconstruction unit is used to organize training courses according to task type, regional source, equipment type, modality type, sample difficulty, and target basis type.

[0223] In this embodiment, the internal execution process of the course reconstruction unit includes two sub-steps: probe sample evaluation and course reconstruction.

[0224] Specifically, in the probe sample evaluation sub-step, AIOS extracts probe sample sets from the available data in stratified categories based on task type, region of origin, device type, modality type, and sample difficulty. Within each update cycle, AIOS uses the model to be updated to infer the probe sample set and records the output quality, consistency score, and reward variance for each layer of probe samples. The reward variance is the variance of the reward signal output by the model across multiple inferences on the probe samples.

[0225] In the course restructuring sub-step, AIOS triggers course restructuring when the reward variance is lower than the preset change threshold for a continuous preset number of update cycles, and the proportion of high-value samples to all samples is lower than the preset sample proportion.

[0226] More specifically, the reorganized training course is organized hierarchically according to seven attributes: task type, regional source, equipment type, modality type, sample difficulty, target basis type, and sample newness / oldness. It is also sampled from newly added high-value samples, historical high-value samples, boundary samples, negative samples, and manual feedback samples according to a preset sampling ratio.

[0227] Historical high-value samples are stored in a replay queue, which maintains the most recent preset number of historical high-value samples in a first-in, first-out manner. The recombined training course maintains a positive-to-negative sample ratio that does not exceed a preset balance threshold.

[0228] The incremental training unit is used to update adapter parameters, fusion layer parameters, output header parameters, or retrieve enhancement module parameters when the base encoder is frozen or the backbone parameters are shared, and to maintain old knowledge through historical sample playback, distillation constraints, or parameter regularization.

[0229] In this embodiment, the internal execution process of the incremental training unit may specifically include two sub-steps: parameter freezing and selective updating, and old knowledge retention.

[0230] Specifically, in the parameter freezing and selective update sub-steps, AIOS freezes the basic encoder parameters or shared backbone parameters of the model to be updated. The frozen parameters do not participate in gradient updates during training. AIOS selectively updates at least one of the adapter parameters, fusion layer parameters, output head parameters, and retrieval enhancement module parameters. The training batch consists of newly added high-value samples, historical replay samples, negative samples, and cross-regional adaptation samples mixed in a preset ratio. AIOS performs forward and backward propagation calculations on the unfrozen parameters batch by batch.

[0231] In the old knowledge retention sub-step, AIOS enables at least one of the following three mechanisms:

[0232] The historical sample replay mechanism mixes high-value historical samples from the replay queue into the current training batch at a preset replay ratio.

[0233] The distillation constraint mechanism uses the output of the model on the training samples before the update as the soft label, and adds the distillation loss between the current model output and the soft label as an auxiliary loss term to the total loss.

[0234] The parameter regularization mechanism applies L2 regularization constraints to the changes in non-frozen parameters, limiting the magnitude of parameter updates.

[0235] The model verification unit is used to verify the capabilities and security of the updated model.

[0236] In this embodiment, the internal execution process of the model verification unit includes two sub-steps: capability verification and security verification.

[0237] Specifically, in the capability verification sub-step, AIOS uses a reserved verification sample set to perform offline evaluation of the updated model. The verification sample set does not overlap with the training sample set in terms of device and region.

[0238] AIOS calculates the updated model's performance on the validation sample set based on five performance indicators: task accuracy, fault identification rate, load prediction error, defect location mAP, and basis hit rate, and compares them item by item with the corresponding indicators of the model before the update.

[0239] In the security verification sub-step, AIOS uses a security rule verification set to verify the updated model. The security rule verification set includes preset power grid security rule templates and data compliance verification rules. Each security rule template defines a prohibited output pattern. If the model output matches any prohibited output pattern or contains non-compliant data references, the security verification fails.

[0240] The release control unit is used to decide whether to release the model, perform a grayscale test, rollback, or manual review.

[0241] In this embodiment, the internal execution flow of the release control unit may specifically include:

[0242] AIOS determines the release action based on the verification report output by the model verification unit and the constraint status field in the current model update status, according to preset decision rules. Specifically, if all capability indicators meet the preset capability thresholds, security verification passes, and there are no violations in the constraint status, direct release is decided. If capability indicators meet the thresholds but low-risk items exist in the security verification, a canary release is decided, and the updated model is first deployed to a preset proportion of non-critical regional nodes for trial operation. If some capability indicators do not meet the thresholds but security verification passes, training continues. If capability indicators do not meet the thresholds or security verification fails and cannot be resolved by continuing training, manual review is decided. If a security event occurs during canary operation or the false alarm rate exceeds the preset threshold, model rollback is decided, and the released model is reverted to the previous version.

[0243] The feedback write-back unit is used to write the business feedback generated by this model update task into the AIOS model update context.

[0244] In this embodiment, after the feedback write-back unit is invoked, AIOS collects the results of manual confirmation, operation and maintenance handling, model false alarm and false alarm records, rollback information and subsequent business performance, and stores them in a preset feedback record format.

[0245] AIOS writes feedback records to the AIOS feedback record library. The feedback record library is indexed by device identifier and region identifier, and updates the manual adoption status, operation and maintenance handling results, and subsequent business performance records in the feedback status field.

[0246] In some implementations, the next processing unit is determined from a preset set of processing units based on the current model update state, including:

[0247] Step S151: Based on the current model update state, calculate the fitness score of each processing unit; wherein the fitness score is determined by weighting the necessity of the step, the update benefit, the execution cost, the constraint risk and the historical feedback correction item.

[0248] In this embodiment, AIOS pre-sets trigger rules for each type of processing unit.

[0249] Specifically, the trigger rule for the data governance unit is that the data quality score is lower than a preset quality threshold. The trigger rule for the validity screening unit is that the proportion of new samples exceeds a preset proportion threshold and the variance of the target validity score distribution exceeds a preset variance threshold.

[0250] The triggering rules for the reinforcement unit are as follows: the missing reference count of the target data exceeds the preset missing threshold; the triggering rules for the course reconstruction unit are: the reward variance is lower than the preset change threshold within a consecutive preset number of periods and the proportion of high-value samples is lower than the preset proportion threshold; the triggering rules for the incremental training unit are: the course reconstruction unit has output the reconstructed training course configuration and the model improvement level exceeds the preset improvement threshold; and the triggering rules for the model validation unit and the release control unit are: the incremental training unit has completed parameter updates.

[0251] AIOS matches the current model update state with each triggering rule. The higher the degree of matching, the greater the value of the step necessity.

[0252] In this implementation, AIOS uses the historical statistical average of the updated revenue of each processing unit in the same historical task as a benchmark, and makes corrections based on the deviation between the current model's level of improvement and the historical average. When the deviation is positive, the revenue estimate is increased, and when the deviation is negative, the revenue estimate is decreased.

[0253] In this embodiment, AIOS reads the execution cost of each processing unit from a preset execution cost table. The execution cost table records the normalized mean of three cost components: computation, training and data processing costs, manual review or release waiting costs, and external call or operation management costs. The execution cost is obtained by weighting and summing the three cost components according to preset weights.

[0254] In this implementation, AIOS pre-defines a risk profile for each type of processing unit. The risk profile represents the types of security risks, data compliance risks, and business constraint risks that may be triggered during the execution of that unit, along with their baseline probabilities. AIOS multiplies the online risk score by the baseline probabilities of the corresponding risks in the risk profile and sums them up to obtain the constraint risks.

[0255] In this embodiment, AIOS uses the current task type and device object as query conditions to obtain historical feedback records of each processing unit in the same type of task from the feedback record library, and calculates the output quality improvement rate, the hit rate improvement rate, the verification pass rate, the average execution cost, and the rollback rate accordingly.

[0256] Among them, the output quality improvement rate is determined based on the difference in task accuracy before and after the model update; the basis hit rate improvement rate is determined based on the difference in basis hit rate before and after the model update; the validation pass rate is the ratio of the number of validation passes to the total number of executions of the processing unit; the average execution cost is the average of the normalized values ​​of resource consumption for each execution; and the rollback rate is the ratio of the number of rollbacks after release to the total number of releases.

[0257] AIOS performs a weighted calculation on the above metrics. Output quality improvement rate, basis hit rate improvement rate, and validation pass rate are positive evaluation metrics, while average execution cost and rollback rate are negative evaluation metrics. The weighted calculation results are normalized to the range [-1, 1] and used as historical feedback correction terms for the corresponding processing units.

[0258] In this embodiment, AIOS can multiply the necessity of the step, the benefit of the update, and the historical feedback correction item by the corresponding preset weight coefficients, and multiply the execution cost and constraint risk by the corresponding preset weight coefficients after taking their inverse values. The sum of these items yields the adaptation score of each processing unit.

[0259] Step S152: The processing unit with the highest fit score is determined as the next processing unit. The historical feedback correction item is obtained by statistically analyzing the output quality improvement, hit rate improvement, verification pass rate, average execution cost, and rollback records of each processing unit in the same type of task, and is used to increase or decrease the fit score of the corresponding processing unit in subsequent similar tasks.

[0260] In this embodiment, AIOS sorts the adaptation scores of each processing unit from largest to smallest, and determines the processing unit with the largest adaptation score as the next processing unit.

[0261] In one implementation, data quality is... Calculate according to the following formula:

[0262] ;

[0263] in, Indicates the field completeness rate. Indicates cross-modal time or device alignment rate. Indicates the proportion of effective values. This represents the timeliness score of the data, with each item ranging from 0 to 1. and This is a data quality weight parameter, which can be determined by business rules, historical task statistics, or validation set search.

[0264] In one implementation, the objective is categorized based on effectiveness. Calculate according to the following formula:

[0265] ;

[0266] ;

[0267] in, This represents the output evaluation value of the target result under complete input. This represents the output evaluation value after masking candidate criteria data. This indicates the output evaluation value when only the task text is retained; the output evaluation value is the normalized transformation of the target category probability, classification confidence, or cross-entropy in classification tasks, the inverse normalized value of the prediction error in regression or load prediction tasks, the normalized evaluation value of mAP, IoU, hit rate, or localization box offset in detection or localization tasks, and the normalized evaluation value based on hit rate, consistency score, or manual confirmation based on matching degree in interpretation and generation tasks. and This is the validity weight parameter. When... Not less than the preset threshold When this is done, the corresponding field, time window, image region, or record will be selected as the target data.

[0268] In one implementation, the current model update state is quantized into a state vector as follows:

[0269] ;

[0270] in, For data quality score, Based on the effectiveness of the objective, The model needs improvement. For uncertainty, To implement risk scoring, The feedback credibility score is used.

[0271] In one implementation, the current model update state includes both a structured state object and a numerical state vector mapped from the structured state object. The structured state object is... It means that, among them, For task status fields, For data status fields, For model status fields, Based on the status field, For uncertain fields, For constraint status fields, This is for feedback status fields. The structured status object stores the original fields for rule judgment, process recording, manual review, and traceable auditing.

[0272] Numerical state vector correspondence formula ,in, It is obtained by mapping the field completeness rate, modal alignment rate, valid value ratio, and timeliness in the data status field. It is obtained by mapping the target based on validity in the status field. It is obtained by mapping the capability items to be optimized in the model state field to the capability evaluation results. It is obtained by mapping diagnostic confidence, reward variance, and items pending review from the uncertainty field. It is obtained by mapping the security, compliance, and release constraints in the constraint status field. It is obtained by mapping the manual adoption status, operation and maintenance handling results, and subsequent business performance in the feedback status field. The structured status object is used for condition triggering and audit traceability, while the numerical status vector is used for adaptability scoring and automatic scheduling.

[0273] In one implementation, fit scoring First, calculate based on the base score and historical feedback correction items:

[0274] ;

[0275] in, Indicates the necessity of the step. Indicates updated earnings. Indicates execution cost, Indicates constraint risk, Indicates historical feedback correction items; , , and Normalize to 0 to 1 Normalize to -1 to 1; , , , and Select weight parameters for the steps. System selection. The largest processing step is used as the next processing step.

[0276] In one implementation, , , , and Determine in the following manner:

[0277] ;

[0278] in, The degree to which the triggering rule is satisfied for processing unit j. The degree of abnormality of the relevant status fields, There is a gap in capabilities that need to be optimized.

[0279] ;

[0280] in, To process the historical gains of unit j in similar tasks, The coverage value of the current sample for the target device, region, and modality. This provides room for improvement in target capabilities.

[0281] ;

[0282] in, Due to computation, training, and data processing costs, Costs associated with manual review or waiting for publication. Costs associated with external calls or operational management.

[0283] ;

[0284] in, Risks related to power safety and model release. For data compliance risks, Risks are determined by business constraints or manually confirmed.

[0285] ;

[0286] in, To improve the output quality of similar historical missions. Based on the improvement of hit rate, To verify the pass status, To average execution cost, To publish the rollback record; when When positive, improve the adaptability score of the corresponding processing unit in subsequent similar tasks. When the score is negative, the adaptation score of the corresponding processing unit is reduced.

[0287] In one implementation, the t-th step process is represented as:

[0288] ;

[0289] in, This indicates the current model update state at step t. This represents the processing unit selected in step t. This indicates the processing function corresponding to the current processing unit. This represents the processing result, and U represents the state update function. The task terminates when the current model update state meets the release condition, the stop condition, or the maximum number of processing steps is reached.

[0290] In one implementation, the model updates the overall score. Calculate according to the following formula:

[0291] ;

[0292] in, For task accuracy, For fault identification rate, Based on hit rate, To ensure a high pass rate, To normalize training costs, To publish the rollback rate; , , , , and To update the performance evaluation weight parameters.

[0293] According to an embodiment of the present invention, an electronic device is provided; please refer to... Figure 4 The electronic device in this embodiment may include one or more of the following components: a processor, a network interface, memory, non-volatile memory, and one or more application programs, wherein the one or more application programs may be stored in non-volatile memory and configured to be executed by one or more processors, and the one or more programs are configured to perform the methods as described in the foregoing method embodiments.

[0294] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A state-driven multimodal model auto-evolution method for power grid AIOS, characterized in that, include: Receive model update tasks from the power grid artificial intelligence operating system (AIOS) and collect available data related to the model update tasks; The model update task is analyzed to determine the task description information; The available data is validated to obtain the target data; wherein, the target data is the data on which the model makes its judgments. Based on the task description information and target data, the current model update status is generated; Based on the current model update state, determine the next processing unit from the preset processing unit set; Execute the next processing unit and obtain the processing result. Update the current model update status based on the processing result and continue to determine the subsequent processing units until the preset release conditions, preset stop conditions, or preset processing step limit are met. Record each processing unit and its processing results to form a model update process record; Model update results are generated based on the model update process record; Feedback and evaluation are conducted based on the model update results and model update process records, and the selection rules for processing units in subsequent model update tasks are adjusted based on the feedback and evaluation results.

2. The method according to claim 1, characterized in that, The available data includes one or more of the following: power grid operation measurement data, equipment status data, image inspection data, dispatch log data, historical fault case data, model inference records, and manual feedback data; The power grid operation measurement data includes one or more of the following: voltage, current, active power consumption, reactive power, power factor, photovoltaic output, wind power output, grid power supply, voltage fluctuation rate, temperature, humidity, electricity price, and predicted load. The equipment status data includes one or more of the following: overload status, transformer fault status, stable status, stability margin, response time, power injection, and coupling coefficient; The image inspection data includes one or more of the following: transmission line images, insulator images, infrared images, defect area markings, and equipment appearance images.

3. The method according to claim 1, characterized in that, The process of parsing the model update task to determine the task description information includes: Extract the task objective, device object, model version, and business scenario from the model update task; wherein, the task objective is used to characterize the type of capability to be improved in this model update, the device object is used to characterize the type of power grid equipment or regional node involved in this model update, the model version is used to record the parameter summary and current capability indicators of the model to be updated, and the business scenario is used to characterize the application conditions in which this model update is performed; The extracted task objective, device object, model version, and business scenario are organized into the task description information.

4. The method according to claim 1, characterized in that, The process of validating the available data to obtain the target data includes: The factual output is obtained based on the complete multimodal input, the task description reference output is obtained based on the task description baseline input, and the basis occlusion reference output is obtained based on the occlusion candidate basis input; wherein, the task description baseline input includes: task objective, problem description, device object, business scenario and output format constraints, but does not include the field data in the available data; Based on the output differences between the fact output, the task description reference output, and the basis occlusion reference output, candidate basis whose output differences satisfy preset conditions are determined as the target basis data; The target data includes one or more of the following: running fields, time window segments, image regions, log records, case records, human feedback records, and combinations of variables.

5. The method according to claim 4, characterized in that, The process of generating the current model update state based on task description information and target data includes: Based on the task description information and target basis data, the current model update status is generated; wherein, the current model update status includes: task status field, data status field, model status field, basis status field, uncertainty field, constraint status field and feedback status field; The task status field is used to record the task objective, device object, and business scenario. The data status field is used to record the sample source, modality type, data quality, regional sample pool, and proportion of newly added samples; The model status field is used to record the model version, model parameter summary, capability evaluation results, and capability items to be optimized; the basis status field is used to record the target basis data. The uncertainty field is used to record diagnostic confidence, reward variance, basis missing items, and items requiring manual review. The constraint status field is used to record security constraints, release constraints, computing resource constraints, data compliance constraints, and manual confirmation requirements; The feedback status field is used to record manual adoption status, operation and maintenance handling results, model rollback results, and subsequent business performance.

6. The method according to claim 1, characterized in that, The preset processing unit set includes: The system comprises a data governance unit, a validity-based screening unit, a validity-based reinforcement unit, a curriculum restructuring unit, an incremental training unit, a model validation unit, a release control unit, and a feedback write-back unit. The data governance unit is used to align and control the quality of multi-source power grid data. The validity screening unit is used to screen high-value samples that require model judgment based on target data. The basis reinforcement unit is used to construct training samples including input samples, target output labels and target basis labels, and to jointly train the model through target output loss, basis reference loss and basis consistency reward, so that the model outputs reference target basis data. The course reconstruction unit is used to organize training courses according to task type, region of origin, device type, modality type, sample difficulty, and target basis type. The incremental training unit is used to update adapter parameters, fusion layer parameters, output head parameters, or retrieve enhancement module parameters when the base encoder is frozen or the backbone parameters are shared, and to maintain old knowledge through historical sample playback, distillation constraints, or parameter regularization. The model verification unit is used to verify the capabilities and security of the updated model. The release control unit is used to determine whether the model is released, grayscale, rolled back, or manually reviewed. The feedback write-back unit is used to write business feedback into the AIOS model update context.

7. The method according to claim 6, characterized in that, Based on the current model update state, the next processing unit is determined from the preset processing unit set, including: Based on the current model update status, the fitness score of each processing unit is calculated; wherein, the fitness score is determined by weighting the necessity of the step, the update benefit, the execution cost, the constraint risk and the historical feedback correction item. The processing unit with the highest fit score is determined as the next processing unit; The historical feedback correction item is obtained by statistically analyzing the output quality improvement of each processing unit in the same task, the hit rate improvement, the verification pass status, the average execution cost, and the rollback record, and is used to improve or reduce the adaptability score of the corresponding processing unit in subsequent similar tasks.

8. A state-driven multimodal model auto-evolution device for power grid AIOS, characterized in that, include: Receiving module: Receives model update tasks from the power grid artificial intelligence operating system (AIOS) and collects available data related to the model update tasks; Parsing module: Parses the model update task to determine the task description information; Validation module: Performs validity verification on the available data to obtain target basis data; wherein, the target basis data is the data on which the model makes its judgments; First generation module: Generates the current model update status based on task description information and target data; Determine module: Based on the current model update state, determine the next processing unit from the preset processing unit set; Update module: Execute the next processing unit and obtain the processing result, update the current model update status according to the processing result, and continue to determine the subsequent processing units until the preset release conditions, preset stop conditions, or preset processing step limit are met; Recording module: Records each processing unit and its processing results, forming a model update process record; The second generation module generates model update results based on the model update process records. Feedback and Evaluation Module: Provides feedback and evaluation based on the model update results and model update process records, and adjusts the processing unit selection rules in subsequent model update tasks based on the feedback and evaluation results.

9. An electronic device, characterized in that, include: A memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors to cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images