Inverter fault diagnosis method and device, electronic equipment and readable storage medium

Abstract

The application discloses an inverter fault diagnosis method and device, electronic equipment and a readable storage medium, and relates to the technical field of new energy power stations. The method comprises the following steps: in response to a fault diagnosis triggering instruction, at least one target diagnosis task for a target inverter is determined; a plurality of candidate diagnosis conclusion labels corresponding to the target diagnosis task are acquired, and runtime sequence data of the target inverter in a time window associated with the target diagnosis task is acquired; task information of the target diagnosis task, each candidate diagnosis conclusion label and the runtime sequence data are input into a pre-trained target time sequence large model, so that the target time sequence large model is used to calculate a matching loss value of each candidate diagnosis conclusion label and the runtime sequence data under the target diagnosis task; and a diagnosis result of the target diagnosis task is determined based on the matching loss value corresponding to each candidate diagnosis conclusion label. The application can improve the accuracy of inverter fault diagnosis and reduce the maintenance cost of the inverter fault diagnosis system.

Description

Technical Field

[0001] This application relates to the field of new energy power plant technology, and in particular to an inverter fault diagnosis method, device, electronic equipment and readable storage medium. Background Technology

[0002] Inverters are the core equipment of new energy power plants, and they continuously generate a large amount of time-series data such as power, voltage, current, and temperature during operation. Various inverter faults typically manifest on the operating curve as trend changes, local fluctuations, abnormal variable linkages, or deviations in control state. Therefore, diagnosing inverter faults usually requires a comprehensive assessment of the combined characteristics of multiple measurement points over a period of time.

[0003] Currently, inverter fault diagnosis mainly relies on manual experience, fixed threshold rules, and mechanistic models designed for single faults. While these methods perform well in scenarios with clear judgment boundaries, they also have significant limitations: different faults require separate rule design and feature extraction, resulting in low algorithm reusability; fixed rules are prone to false negatives or missed faults, and as the number of fault types increases, the rule branches continue to expand, significantly increasing maintenance costs. Summary of the Invention

[0004] The main objective of this application is to provide an inverter fault diagnosis method, device, electronic equipment, and readable storage medium, which aims to improve the accuracy of inverter fault diagnosis and reduce the maintenance cost of the inverter fault diagnosis system.

[0005] This application provides a method for diagnosing inverter faults, the method comprising: In response to a fault diagnosis trigger command, determine at least one target diagnostic task for the target inverter; Obtain multiple candidate diagnostic conclusion tags corresponding to the target diagnostic task, as well as the runtime sequence data of the target inverter within the time window associated with the target diagnostic task; The task information of the target diagnostic task, each of the candidate diagnostic conclusion labels, and the runtime sequence data are input into a pre-trained target time series large model, so that the target time series large model can calculate the matching loss value between each of the candidate diagnostic conclusion labels and the runtime sequence data under the target diagnostic task. The diagnostic result of the target diagnostic task is determined based on the matching loss value corresponding to each of the candidate diagnostic conclusion labels.

[0006] In one embodiment, the step of determining the diagnostic result of the target diagnostic task based on the matching loss value corresponding to each of the candidate diagnostic conclusion labels includes: The matching loss value corresponding to each candidate diagnostic conclusion label is normalized to obtain the confidence level of each candidate diagnostic conclusion label; The difference between the highest confidence level and the second highest confidence level among the confidence levels of each candidate diagnostic conclusion label is detected to be greater than a preset difference threshold, and the highest confidence level is greater than a first preset confidence threshold. If so, the diagnostic result of the target diagnostic task is determined to be the candidate diagnostic conclusion label corresponding to the highest confidence level; If not, the diagnostic result of the target diagnostic task is determined to be pending manual review.

[0007] In one embodiment, before the step of determining the diagnostic result of the target diagnostic task as the candidate diagnostic conclusion label corresponding to the highest confidence level, the method further includes: The runtime sequence data is input into a pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; wherein, the mechanism model performs fault diagnosis based on preset rules; If the diagnostic report output by the mechanism model can match a diagnostic conclusion, and the confidence level of the matched target diagnostic conclusion is greater than the second preset confidence threshold, then verify whether the target diagnostic conclusion is consistent with the candidate diagnostic conclusion label corresponding to the highest confidence level. If they match, the diagnostic result of the target diagnostic task is determined to be the candidate diagnostic conclusion label corresponding to the highest confidence level.

[0008] In one embodiment, before the step of inputting the task information of the target diagnostic task, each of the candidate diagnostic conclusion labels, and the runtime sequence data into the pre-trained target temporal large model, the method further includes: The runtime sequence data is input into a pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; wherein, the mechanism model performs fault diagnosis based on preset rules; If the diagnostic report output by the mechanism model can match a diagnostic conclusion, and the confidence level of the matched target diagnostic conclusion is greater than the second preset confidence threshold, then the target diagnostic conclusion is taken as the diagnostic result of the target diagnostic task. If the diagnostic report output by the mechanistic model cannot match a diagnostic conclusion, or if the confidence level of the matched target diagnostic conclusion is less than or equal to the second preset confidence threshold, the step of inputting the task information of the target diagnostic task, the labels of each candidate diagnostic conclusion, and the runtime sequence data into the pre-trained target time series large model is executed.

[0009] In one embodiment, the construction steps of the target time series large model include: Obtain a multi-task training sample set. Each training sample in the multi-task training sample set includes sample task information, a candidate diagnostic conclusion label sample set, runtime sequence sample data, and answer text. The answer text is composed of a preset fixed answer prefix and a real diagnostic conclusion label. The pre-trained temporal model is fine-tuned based on the multi-task training sample set, and the training loss value of the real diagnostic conclusion label in the answer text of each training sample is calculated by using a loss mask during the fine-tuning process. With the goal of minimizing the training loss value of each training sample, the model parameters of the pre-trained temporal large model are iteratively updated to generate the target temporal large model.

[0010] In one embodiment, the step of fine-tuning the pre-trained temporal large model based on the multi-task training sample set includes: The multi-task training sample set is subjected to balanced sampling according to the preset task-level sampling weight and category-level sampling weight to obtain the balanced multi-task training sample set. Based on the balanced sampling of the multi-task training sample set, the pre-trained temporal large model is fine-tuned. The task-level sampling weight is used to balance the sample quantity distribution among different diagnostic tasks, and the category-level sampling weight is used to balance the sample quantity distribution among different candidate diagnostic conclusion labels within the same diagnostic task.

[0011] In one embodiment, after the step of generating the target time series large model, the method further includes: Having obtained the training and validation sample sets for the new diagnostic task, the target time series model is incrementally fine-tuned based on the training sample set for the new diagnostic task. Based on the validation sample set of the newly added diagnostic task and the validation sample set associated with the target time series large model before incremental fine-tuning, the effectiveness of the target time series large model after incremental fine-tuning is verified. If the validity verification is successful, the incrementally fine-tuned target time series model is run online.

[0012] In one embodiment, after the step of validating the effectiveness of the incrementally fine-tuned target time-series large model based on the validation sample set associated with the newly added diagnostic task and the validation sample set associated with the target time-series large model before incremental fine-tuning, the method further includes: If the validity verification fails, the incrementally fine-tuned target time series model is run in offline auxiliary mode. The diagnostic report output by the incrementally fine-tuned target time series model is sent to the manual reviewer as a review reference.

[0013] In one embodiment, the step of obtaining multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task includes: Using the task information of the target diagnostic task as an index, multiple candidate diagnostic conclusion tags corresponding to the target diagnostic task are retrieved from the preset multi-task diagnostic configuration table.

[0014] Furthermore, to achieve the above objectives, this application also provides an inverter fault diagnosis device, the inverter fault diagnosis device comprising: The task determination module is used to determine at least one target diagnostic task for the target inverter in response to a fault diagnosis trigger command. The data acquisition module is used to acquire multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task, as well as the runtime sequence data of the target inverter within the time window associated with the target diagnostic task; The fault diagnosis module is used to input the task information of the target diagnosis task, each of the candidate diagnosis conclusion labels, and the runtime sequence data into a pre-trained target time series large model, so as to calculate the matching loss value between each of the candidate diagnosis conclusion labels and the runtime sequence data under the target diagnosis task through the target time series large model; and to determine the diagnosis result of the target diagnosis task based on the matching loss value corresponding to each of the candidate diagnosis conclusion labels.

[0015] In addition, to achieve the above objectives, this application also provides an electronic device, the electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the inverter fault diagnosis method as described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the inverter fault diagnosis method as described above.

[0017] This application provides an inverter fault diagnosis method, comprising: responding to a fault diagnosis trigger command, determining at least one target diagnosis task for a target inverter; acquiring multiple candidate diagnosis conclusion labels corresponding to the target diagnosis task, and runtime sequence data of the target inverter within the time window associated with the target diagnosis task; inputting the task information of the target diagnosis task, each candidate diagnosis conclusion label, and the runtime sequence data into a pre-trained target time series large model, so as to calculate the matching loss value between each candidate diagnosis conclusion label and the runtime sequence data under the target diagnosis task through the target time series large model; and determining the diagnosis result of the target diagnosis task based on the matching loss value corresponding to each candidate diagnosis conclusion label.

[0018] Therefore, the technical solution provided in this application pre-trains a target time-series large model capable of calculating the matching loss value between each candidate diagnostic conclusion label and the inverter's runtime timing data under a given diagnostic task. Thus, after determining at least one target diagnostic task for the target inverter, multiple candidate diagnostic conclusion labels corresponding to the predefined target diagnostic task are obtained, along with the runtime timing data of the target inverter within the relevant time window required to achieve the target diagnostic task. The task information of the target diagnostic task, each candidate diagnostic conclusion label, and the runtime timing data are then input into the target time-series large model. This model directly calculates the matching loss value between each candidate diagnostic conclusion label and the runtime timing data under the target diagnostic task, i.e., assessing the degree of matching between the operating mode presented by the runtime timing data and each candidate diagnostic conclusion under the target diagnostic task. Based on this, the diagnostic result of the target diagnostic task can be accurately determined.

[0019] In summary, the technical solution provided in this application, when diagnosing inverter faults, uses a unified candidate diagnostic conclusion label system, which eliminates the need to redesign the algorithm link when adding new fault types, thereby reducing the maintenance and expansion costs of the multi-fault diagnosis system. Moreover, the diagnostic results are always determined based on the matching loss value within the preset candidate label range, avoiding the uncertainty of generative model output, and the accuracy of fault diagnosis is also high. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the inverter fault diagnosis method provided in the first embodiment of this application. Figure 2 This is a schematic diagram of the inverter fault diagnosis device provided in the embodiments of this application; Figure 3 This is a schematic diagram of the hardware operating environment involved in the embodiments of this application.

[0023] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0025] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0026] The execution subject of the inverter fault diagnosis method of this application can be an electronic device with data processing, network communication and program operation functions, such as the control equipment of a new energy power station or an inverter containing the control equipment, etc. This embodiment does not specifically limit it.

[0027] The following description uses an electronic device as the execution subject to illustrate the various embodiments.

[0028] This application presents a first embodiment of an inverter fault diagnosis method; please refer to [link / reference]. Figure 1 The inverter fault diagnosis method may include steps S10 to S40: Step S10: In response to the fault diagnosis trigger command, determine at least one target diagnostic task for the target inverter; A fault diagnosis trigger command is a signal or message used to initiate a fault diagnosis process. It can be automatically generated by the inverter monitoring system when abnormal operating parameters are detected, or it can be issued periodically by the central control center according to a preset inspection plan, or it can be manually issued by maintenance personnel through a human-machine interface. This embodiment does not specifically limit this. The fault diagnosis trigger command carries at least the identification information of the inverter that needs to be diagnosed, and the task information of the diagnostic task to be executed. The target inverter is the inverter that currently needs fault diagnosis. The target diagnostic task refers to the diagnostic task that needs to be executed for the target inverter. One fault diagnosis trigger command can correspond to one or more target diagnostic tasks; this embodiment does not specifically limit this.

[0029] When multiple diagnostic tasks exist, they can be summarized according to task priority, fault severity, confidence level, and / or mechanism verification results. If multiple tasks output contradictory conclusions, they can be prioritized for manual review. If multiple tasks point to different causes of abnormality in the same inverter, combined diagnostic results can be formed for subsequent work order dispatch and cause localization.

[0030] Step S20: Obtain multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task, as well as the runtime sequence data of the target inverter within the time window associated with the target diagnostic task; The multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task are a set of possible diagnostic results predefined for the target diagnostic task. For example, for a heat dissipation anomaly diagnostic task, the multiple candidate diagnostic conclusion labels may include normal heat dissipation, fan failure, air duct blockage, and the influence of high ambient temperature.

[0031] The runtime sequence data of the target inverter within the time window associated with the target diagnostic task refers to the runtime sequence data of the target inverter within the relevant time window required to achieve the target diagnostic task. The length and end position of the time window associated with the target diagnostic task are related to the characteristics of the target diagnostic task. For example, for an inefficient diagnostic task that requires observing trend changes, its time window may be the past hour; for an instruction anomaly diagnostic task that requires observing instantaneous anomalies, its time window may only be 5 minutes before and after. The runtime sequence data is the sequence data of various operating parameters generated by the target inverter within the time window. When acquiring runtime sequence data, specific measurement points are not limited. Instead, several relevant operating curves, such as power curves, voltage curves, current curves, and temperature curves, are selected according to the needs of the target diagnostic task and arranged in a uniform timestamp alignment.

[0032] In one feasible implementation, the step of obtaining multiple candidate diagnostic conclusion tags corresponding to the target diagnostic task may include: using the task information of the target diagnostic task as an index, searching in a preset multi-task diagnostic configuration table to obtain multiple candidate diagnostic conclusion tags corresponding to the target diagnostic task.

[0033] A multi-task diagnostic configuration table is a pre-built, structured data table or configuration file used to centrally manage the configuration information of all inverter diagnostic tasks. Each record in this table corresponds to a diagnostic task and contains at least two fields: task information and a list of candidate diagnostic conclusion tags. Through this multi-task diagnostic configuration table, the system eliminates the need to maintain separate diagnostic logic and tag sets for each type of fault in the code, thus achieving unified registration and management of diagnostic tasks.

[0034] When setting the candidate diagnostic conclusion label set for a diagnostic task, for tasks with a hierarchical diagnostic structure, a graded labeling approach can be adopted, dividing the candidate diagnostic conclusion label set into primary labels and secondary labels. Primary labels are used for coarse-grained preliminary judgments, such as "operating normally" and "operating abnormally." Secondary labels are used to further determine the specific cause of the abnormality after a preliminary judgment. For example, when the primary judgment result is "operating abnormally," it can be further subdivided into inverter heat dissipation abnormality, control command abnormality, AC voltage abnormality, or other sub-causal types. For samples where there is no clear basis for judgment, labels requiring manual confirmation or undeterminable can also be set in the candidate diagnostic conclusion label set to avoid the model being forced to output a specific diagnostic conclusion when the evidence is insufficient.

[0035] Task information is a collection of all metadata used to define and describe a diagnostic task. It may include, but is not limited to, task name, diagnostic object, applicable inverter model range, required operating curve type, priority, review strategy, and output fields. Among them, the diagnostic object specifies the inverter device or group of devices targeted by the diagnostic task; the applicable inverter model range limits the inverter manufacturers and models to which the diagnostic task is applicable, avoiding the incorrect application of diagnostic logic to incompatible inverter devices; the required operating curve type defines which time-series data points need to be collected when performing the diagnostic task; the priority determines the processing order when multiple tasks are performed in parallel; the review strategy defines under what conditions the diagnostic conclusions need to be pushed to human review; and the output fields specify the format in which the diagnostic conclusions are written into the diagnostic report or alarm message.

[0036] This organizational method, indexed by task information and constrained by the candidate tag list, enables unified registration and management of multiple types of fault diagnosis tasks. Therefore, when adding a new fault diagnosis type, only a new configuration record needs to be added to the multi-task diagnosis configuration table, populated with the corresponding task information and candidate tag list, without requiring any modification to the code logic in the diagnosis process.

[0037] In addition, for diagnostic tasks that are supported by existing mechanism algorithms, the results of the mechanism algorithms can be marked as a preliminary screening condition in the review strategy of their task information. For new fault tasks that have not yet been fully verified, the running mode can be set to offline auxiliary mode first. After the offline evaluation using historical data and the diagnostic effect reach the preset stable conditions, the running mode can be switched to online mode to formally connect to the real-time diagnostic process.

[0038] In this embodiment, a multi-task diagnostic configuration table is set up to uniformly register and centrally manage the candidate diagnostic conclusion tags corresponding to different inverter fault diagnostic tasks. Therefore, after responding to a fault diagnostic trigger command and determining the target diagnostic task, the task information of that target diagnostic task can be used as an index to quickly and accurately obtain multiple candidate diagnostic conclusion tags corresponding to that target diagnostic task. This organization method, using task information as an index and the candidate tag list as a constraint, achieves unified registration and management of multiple types of fault diagnostic tasks. When adding a new inverter fault diagnostic type, only a new configuration record needs to be added to the multi-task diagnostic configuration table, populated with the corresponding task information and candidate tag list, without modifying any code logic in the diagnostic process. This effectively reduces the maintenance cost and expansion complexity of the multi-fault diagnostic system.

[0039] Step S30: Input the task information of the target diagnosis task, the candidate diagnosis conclusion labels and the runtime sequence data into the pre-trained target time series large model, so as to calculate the matching loss value between each candidate diagnosis conclusion label and the runtime sequence data under the target diagnosis task through the target time series large model. The target time-series large model refers to a pre-trained and fine-tuned deep learning model capable of processing multivariate time-series data and understanding and executing the specific instruction to "calculate a matching loss value based on the input data and candidate labels." The matching loss value measures the degree of agreement between the target inverter's runtime time-series data and the candidate diagnostic conclusion labels under the current diagnostic task. A lower matching loss value indicates a higher degree of agreement between the runtime time-series data and the state described by the corresponding candidate diagnostic conclusion label.

[0040] In one feasible implementation, before inputting the task information of the target diagnostic task, the labels of each candidate diagnostic conclusion, and the runtime sequence data into the pre-trained target time series large model, the obtained runtime sequence data can be cleaned and its quality verified. Specifically, the time series measurement point data and necessary metadata fields required to complete the target diagnostic task are retained in the runtime sequence data, and the remaining original fields are removed to reduce the data dimension of the model input. Furthermore, the integrity of the retained time series data is checked. When the data missing rate exceeds the preset ratio, the effective time coverage is insufficient, or there are numerical anomalies, it can be marked as undiagnosable and the subsequent process can be terminated to avoid low-quality samples interfering with the model's judgment.

[0041] Among them, necessary metadata refers to basic information that is unrelated to diagnostic tasks but is useful for data location and traceability, such as the inverter ID to which the data belongs, timestamp, site information, etc.

[0042] Step S40: Determine the diagnostic result of the target diagnostic task based on the matching loss value corresponding to each candidate diagnostic conclusion label.

[0043] The diagnostic result of the target diagnostic task refers to the final output conclusion about this target diagnostic task. It can be one of multiple candidate diagnostic conclusion labels, or it can be a preset state that indicates that the system cannot make a judgment automatically, such as "awaiting manual review".

[0044] In one feasible implementation, step S40 may include steps S41 to S44: Step S41: Normalize the matching loss value corresponding to each candidate diagnostic conclusion label to obtain the confidence level of each candidate diagnostic conclusion label. Confidence score is a quantitative indicator that represents the degree of matching between runtime sequence data and corresponding candidate diagnostic conclusion labels, obtained after transforming the matching loss value. The higher the confidence score of a candidate diagnostic conclusion label, the higher the degree of matching between the target time series model and the current runtime sequence data, and the more likely the candidate diagnostic conclusion label is to be a correct diagnosis.

[0045] Normalization refers to the process of converting the matching loss value corresponding to each candidate diagnostic conclusion label into a probability distribution through a preset mathematical transformation, so that the sum of the confidence scores after transformation is 1, so as to facilitate subsequent comparison and analysis.

[0046] Step S42: Detect whether the difference between the highest confidence level and the second highest confidence level among the confidence levels of each candidate diagnostic conclusion label is greater than a preset difference threshold, and whether the highest confidence level is greater than a first preset confidence threshold. The highest confidence level refers to the maximum confidence level among all candidate diagnostic conclusion labels. The second highest confidence level refers to the confidence level of the candidate diagnostic conclusion label that is second only to the highest confidence level. The preset difference threshold is a pre-set confidence difference threshold used to determine whether the model's tendency towards the optimal candidate diagnostic conclusion label is sufficiently clear. It can be a default value or flexibly set by the user according to the actual situation; this embodiment does not impose specific limitations on it. The first preset confidence threshold is a pre-set confidence threshold used to determine whether the model's acceptance of the optimal candidate diagnostic conclusion label reaches the standard for automatic output. It can be a default value or flexibly set by the user according to the actual situation; this embodiment does not impose specific limitations on it.

[0047] Step S43, if yes, then determine the diagnostic result of the target diagnostic task as the candidate diagnostic conclusion label corresponding to the highest confidence level; Step S44: If not, then determine that the diagnosis result of the target diagnosis task is pending manual review.

[0048] "Pending manual review" is a preset diagnostic result status, indicating that the system cannot automatically determine a diagnostic conclusion based on the current evidence and requires manual judgment by maintenance personnel.

[0049] In this implementation, the matching loss value is normalized to a confidence level, and the final diagnostic result output method is determined based on the difference between the highest and second-highest confidence levels, and whether the highest confidence level itself meets a preset threshold. When both threshold conditions are met, it indicates that the target time-series large model has a clear and unique tendency towards a certain candidate diagnostic conclusion label, and the candidate diagnostic conclusion label corresponding to the highest confidence level can be directly used as the diagnostic result of the target diagnostic task. When either threshold condition is not met, it indicates that the target time-series large model has hesitation or uncertainty in judging the current runtime time-series data, and the diagnostic result of the target diagnostic task can be determined to be subject to manual review by operations and maintenance personnel. Thus, this decision-making mechanism based on dual threshold verification can effectively reduce the risk of misjudgment caused by the model forcibly giving deterministic conclusions on boundary samples and difficult samples while ensuring the efficiency of automated diagnosis, thereby improving the accuracy and reliability of the diagnostic results.

[0050] This embodiment does not specifically limit the implementation of step S40. For example, in other feasible implementations, the candidate diagnostic conclusion label corresponding to the minimum matching loss value among the candidate diagnostic conclusion labels can be directly used as the diagnostic result of the target diagnostic task.

[0051] Based on the above, the technical solution provided in this embodiment pre-trains a target time-series large model capable of calculating the matching loss value between each candidate diagnostic conclusion label and the inverter's runtime timing data under a given diagnostic task. Therefore, after determining at least one target diagnostic task for the target inverter, multiple candidate diagnostic conclusion labels corresponding to the predefined target diagnostic task are obtained, along with the runtime timing data of the target inverter within the relevant time window required to achieve the target diagnostic task. Then, the task information of the target diagnostic task, each candidate diagnostic conclusion label, and the runtime timing data are input into the target time-series large model. This allows the target time-series large model to directly calculate the matching loss value between each candidate diagnostic conclusion label and the runtime timing data under the target diagnostic task, thus evaluating the degree of matching between the operating mode presented by the runtime timing data and each candidate diagnostic conclusion under the target diagnostic task. Based on this, the diagnostic result of the target diagnostic task can be accurately determined.

[0052] Therefore, the technical solution provided in this embodiment, when implementing inverter fault diagnosis, uses a unified candidate diagnostic conclusion label system, so that adding new fault types does not require redesigning the algorithm link, thereby reducing the maintenance and expansion costs of the multi-fault diagnosis system. Moreover, the diagnostic results are always determined based on the matching loss value within the preset candidate label range, avoiding the uncertainty of generative model output, and the accuracy of fault diagnosis is also high.

[0053] Based on the first embodiment described above, a second embodiment of the inverter fault diagnosis method of this application is proposed. In the second embodiment, before step S43, the inverter fault diagnosis method may further include steps S401 to S403: Step S401: Input runtime sequence data into the pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; wherein, the mechanism model performs fault diagnosis based on preset rules; A mechanistic model is a deterministic model that diagnoses faults based on pre-defined logical rules, physical laws, or threshold conditions, rather than a neural network model trained on data. For example, a typical mechanistic rule is: when the internal temperature is greater than 85°C and the fan speed is less than 800 rpm, a fan fault is diagnosed. Mechanistic models are characterized by clear judgment chains and low computational overhead.

[0054] When the mechanistic model is a rule engine, after inputting runtime sequence data into the rule engine, the rule engine can execute the diagnostic rules associated with the target diagnostic task one by one. When all the preconditions of a rule are met, the rule engine outputs the corresponding diagnostic conclusion and its preset confidence level.

[0055] When the mechanism model is a physical model based on physical formulas, the runtime sequence data is input into the physical model to substitute the relevant parameters in the runtime sequence data into the physical equations, thereby calculating the theoretical expected value; based on the numerical range of the difference between the theoretical expected value and the actual measured value, the physical model outputs the corresponding diagnostic conclusion and the confidence level calculated based on the difference.

[0056] Step S402: If the diagnostic report output by the mechanism model can match the diagnostic conclusion and the confidence of the matched target diagnostic conclusion is greater than the second preset confidence threshold, verify whether the target diagnostic conclusion is consistent with the candidate diagnostic conclusion label corresponding to the highest confidence. The second pre-set reliability threshold is a pre-defined confidence threshold used to evaluate the reliability of the diagnostic conclusions output by the mechanistic model. It can be a default value or can be flexibly set by the user according to the actual situation. This embodiment does not impose specific limitations on it. The target diagnostic conclusion refers to the specific diagnostic conclusion matched by the mechanistic model when the output conditions are met. The target diagnostic conclusion is one of the candidate diagnostic conclusion labels corresponding to the target diagnostic task.

[0057] Step S403: If they match, then determine the diagnostic result of the target diagnostic task as the candidate diagnostic conclusion label corresponding to the highest confidence level.

[0058] In this embodiment, a mechanistic model is introduced on top of the time-series large-scale model to construct a dual-model cross-validation diagnostic decision-making mechanism. Under this mechanism, the system automatically outputs the final diagnostic result of the target diagnostic task only when the mechanistic model can provide a diagnostic conclusion that meets the reliability requirements, and this diagnostic conclusion is completely consistent with the candidate diagnostic conclusion label pointed to by the highest confidence level of the target time-series large-scale model. This decision-making method, which mutually verifies the data-driven model and the rule-driven model, effectively combines the ability of the time-series large-scale model to identify complex multivariate curve patterns with the advantages of the mechanistic model's transparent and highly interpretable judgment logic. This reduces the risk of misjudgment caused by a single model's training data bias or insufficient boundary sample recognition ability, thereby further improving the accuracy and reliability of inverter fault diagnosis.

[0059] Furthermore, in one feasible implementation, if the target diagnostic conclusion is found to be inconsistent with the candidate diagnostic conclusion label corresponding to the highest confidence level, the diagnostic result of the target diagnostic task can be determined to be pending manual review and handed over to the operation and maintenance personnel for manual judgment.

[0060] Based on the first and / or second embodiments described above, a third embodiment of the inverter fault diagnosis method of this application is proposed. In the third embodiment, before step S30, the inverter fault diagnosis method may further include steps S301 to S303: Step S301: Input runtime sequence data into the pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; wherein, the mechanism model performs fault diagnosis based on preset rules; Step S302: If the diagnostic report output by the mechanism model can match the diagnostic conclusion, and the confidence of the matched target diagnostic conclusion is greater than the second preset confidence threshold, the target diagnostic conclusion is taken as the diagnostic result of the target diagnostic task. Step S303: If the diagnostic report output by the mechanism model cannot match a diagnostic conclusion, or if the confidence level of the matched target diagnostic conclusion is less than or equal to the second preset confidence threshold, the step of inputting the task information of the target diagnostic task, the labels of each candidate diagnostic conclusion, and the runtime sequence data into the pre-trained target time series large model is executed.

[0061] In this embodiment, a mechanistic model is introduced as a pre-screening step before calling the target time-series large model to construct a hierarchical diagnostic mechanism that rapidly screens the mechanism model and deeply supplements it with the time-series large model. Under this mechanism, for diagnostic tasks where the mechanistic model can match the diagnostic conclusion and the reliability meets the threshold requirements, the diagnostic conclusion of the mechanistic model can be directly output as the diagnostic result of the relevant diagnostic task, thereby fully utilizing the advantages of the mechanistic model's low computational overhead and fast response speed, and reducing unnecessary model inference resource consumption. For difficult samples where the mechanistic model cannot match or the reliability is insufficient, the target time-series large model is then used for supplementary diagnosis to leverage the time-series large model's strong ability to recognize complex multivariate curve patterns and its wide coverage. This hierarchical diagnostic mechanism can effectively improve the efficiency and accuracy of inverter fault diagnosis.

[0062] Based on the first, second, and / or third embodiments described above, a fourth embodiment of the inverter fault diagnosis method of this application is proposed. In the fourth embodiment, after step S40, the inverter fault diagnosis method may further include: The key information generated during the execution of the target diagnosis task will be recorded as diagnosis records, and the runtime sequence data corresponding to the diagnosis records that meet the preset return conditions and their reviewed diagnosis conclusions will be returned to the training sample library. The diagnostic record includes at least one of the following: task information of the target diagnostic task, version identifier of the target time series large model used, confidence level corresponding to each candidate diagnostic conclusion label, final diagnostic result, diagnostic conclusion output by the mechanism model, and final conclusion after manual review.

[0063] The preset backfeedback conditions include at least one of the following: low-confidence samples whose final diagnostic results require manual review; misjudged samples whose manual review conclusions are inconsistent with the original output conclusions of the target time-series large model or mechanistic model; diagnostic samples belonging to newly added fault types; and diagnostic samples that are disputed during the manual review process. For samples that meet the above backfeedback conditions, their runtime sequence data, task information, and correctly diagnosed conclusion labels after manual confirmation can be backfeeded to the multi-task training sample library of the target time-series large model.

[0064] Diagnostic records are data entries that are structured and stored after a complete fault diagnosis process. These records are used for post-diagnosis tracking of results, statistical analysis of model performance, and screening of training samples for retraining. Target retraining samples are those that meet preset retraining conditions and will be reintegrated into the training sample library for subsequent incremental fine-tuning of the model. Target retraining samples are typically those that supplement the current model's diagnostic capabilities, such as samples where the model's judgment is uncertain, samples where there is a discrepancy between model and human judgment, and samples belonging to new fault types.

[0065] This embodiment introduces a diagnostic record saving and sample feedback mechanism after the diagnostic results are output, transforming each diagnostic process into a potential data source for continuous model optimization. For samples where the target time-series large-scale model's judgment is uncertain, deviates from the mechanistic model or human judgment, or belongs to newly added fault types, they can be automatically fed back to the training sample library, forming a targeted supplement to the diagnostic shortcomings of the target time-series large-scale model. Subsequently, through incremental fine-tuning and canary release, the target time-series large-scale model can gradually digest the new knowledge and boundary patterns carried by the fed-back samples while ensuring that existing diagnostic capabilities do not decline, ultimately forming a continuously evolving capability that can cover multiple types of inverter faults and support unified multi-task diagnosis.

[0066] Based on the first, second, third, and / or fourth embodiments described above, a fifth embodiment of the inverter fault diagnosis method of this application is proposed. In the fifth embodiment, the construction step of the target time-series large model may include steps S01 to S03: Step S01: Obtain a multi-task training sample set. Each training sample in the multi-task training sample set includes sample task information, a candidate diagnostic conclusion label sample set, runtime sequence sample data, and answer text. The answer text is composed of a preset fixed answer prefix and a real diagnostic conclusion label. A multi-task training sample set is a collection of samples used to train a target time-series model, containing training samples from multiple different diagnostic tasks. Unlike traditional methods that train a separate model for each fault, the multi-task training sample set mixes samples of multiple fault types together, enabling the same model to learn the decision boundaries of different diagnostic tasks in a unified training process.

[0067] Sample task information refers to the task information of the diagnostic task to which the training sample belongs; the candidate diagnostic conclusion label sample set refers to the set of candidate diagnostic conclusion labels corresponding to the diagnostic task to which the training sample belongs; runtime sequence sample data refers to the motion time sequence data extracted from the inverter's historical operating data and corresponding to the diagnostic task to which the training sample belongs; the answer text refers to the standard output text in the training sample that serves as a supervision signal, which is composed of a preset fixed answer prefix and the true diagnostic conclusion label. For example, if the fixed answer prefix is ​​"Diagnosis result:" and the true diagnostic conclusion label is "Fan failure", then the answer text is "Diagnosis result: Fan failure". The fixed answer prefix provides a unified text anchor point during training and inference, enabling the model to stably locate the output position that needs attention. The true diagnostic conclusion label refers to the correct diagnostic conclusion corresponding to the training sample, which has been manually labeled or confirmed through other reliable methods. It can come from historical fault work orders, manual review records, or diagnostic cases confirmed by the mechanism model.

[0068] Step S02: Fine-tune the pre-trained temporal large model based on the multi-task training sample set, and calculate the training loss value of the real diagnostic conclusion label in the answer text of each training sample through the loss mask during the fine-tuning process. The pre-trained time series large model is a deep learning model that has been pre-trained on general time series data, possessing the ability to recognize and represent patterns in multivariate time series sequences. This pre-trained time series large model serves as a base model; instead of being trained from scratch for a single fault type, it undergoes domain-specific fine-tuning on a pre-constructed multi-task training sample set, gradually learning common anomaly patterns in inverter operating curves and differentiated judgment boundaries under different diagnostic tasks.

[0069] A loss mask is a selective masking mechanism applied when calculating training loss values. It specifies which positions in the input sequence are included in the loss function calculation and which are ignored. In this embodiment, the loss mask is configured to calculate the loss value only for the position of the true diagnostic conclusion label located after the fixed answer prefix in the answer text, while not calculating the loss value for the input part before the fixed answer prefix. For true diagnostic conclusion labels composed of multiple tokens, the average or weighted average of the loss values ​​of each token within the label segment can be calculated as the training loss value for that training sample to avoid training bias caused by the different label lengths of different diagnostic tasks.

[0070] Through the aforementioned loss masking mechanism, the optimization objective of the model during fine-tuning shifts from generating a complete diagnostic explanation to accurately outputting the correct diagnostic conclusion label from the candidate label set for the current diagnostic task. This ensures that the behavior of the fine-tuned model remains consistent with the method used in online inference, where the diagnostic conclusion is determined by matching the loss value within the candidate label set. In other words, the model focuses on identifying the best match from the pre-defined candidate diagnostic conclusion label set, rather than freely generating a diagnostic description.

[0071] The training loss value refers to the prediction error of a pre-trained time series large-scale model on the true diagnostic conclusion labels in the training samples. Specifically, it is the difference between the predicted probability distribution of each candidate diagnostic conclusion label output by the pre-trained time series large-scale model and the ideal probability distribution targeting the true diagnostic conclusion labels. The smaller the training loss value, the more accurate the model's diagnostic judgment on the corresponding training samples.

[0072] When calculating the training loss value of the true diagnostic conclusion label in the answer text of each training sample using a loss mask, the position indices of the starting and ending words of the true diagnostic conclusion label in the answer text can be recorded when constructing the input sequence of the training samples. After the model's forward propagation is complete, a loss mask vector of the same length as the input sequence is generated, with only the mask value at the position corresponding to the true diagnostic conclusion label word set to 1, and the rest set to 0. The loss mask is then multiplied element-wise by the original loss value at each position, and the average is calculated to obtain the final training loss value. Alternatively, the loss function interface of the model training framework can be directly used, setting the loss weights of the fixed answer prefix and preceding word positions to zero during the loss value calculation stage, retaining only the loss weight at the position of the true diagnostic conclusion label as a valid value. This embodiment does not specifically limit the implementation method of step S02.

[0073] To accommodate the expansion needs of multiple diagnostic tasks, during fine-tuning, task identifiers corresponding to the sample task information can be embedded in the input layer of the pre-trained temporal model. This allows the model to perceive the diagnostic task to which the current training sample belongs, thereby learning the judgment patterns corresponding to different diagnostic tasks. Furthermore, independent low-rank adaptation parameters can be maintained for different diagnostic tasks or similar families of diagnostic tasks. These low-rank adaptation parameters refer to trainable low-rank matrices inserted into the key attention layer or feedforward layer of the pre-trained temporal model. During fine-tuning, the main parameters of the pre-trained temporal model are frozen, and only the inserted low-rank adaptation parameters are updated. This significantly reduces the computational resource consumption during training. Simultaneously, when adding a new fault diagnosis task, expansion can be completed simply by supplementing the training samples for that new fault diagnosis task and fine-tuning the corresponding low-rank adaptation parameters, without needing to retrain and deploy the entire temporal model.

[0074] For multiple diagnostic tasks with strong correlation, the same set of low-rank adaptation parameters can be shared to improve the knowledge transfer capability between different diagnostic tasks. When loading the model in the online inference service, the corresponding task identifier and / or low-rank adaptation parameters can be dynamically selected or loaded based on the task information of the current target diagnostic task, enabling the same base model to serve multiple different diagnostic tasks.

[0075] In one feasible implementation, the above fine-tuning process can be divided into three stages: The first stage is the general diagnostic adaptation stage, where training samples containing various diagnostic task expression formats are used to perform initial fine-tuning on the pre-trained time series model, enabling the model to become familiar with the task information expression methods of diagnostic tasks, the input format of runtime sequence curves, and the answer text format formed by concatenating fixed answer prefixes and diagnostic conclusion labels. The second stage is the task-specific reinforcement stage, where corresponding specialized training samples are used to perform reinforcement fine-tuning for each diagnostic task, enabling the model to gradually learn the unique judgment boundaries and fault mode characteristics of each diagnostic task. The third stage is the cross-task hybrid training and difficult example playback stage, where training samples from multiple diagnostic tasks are mixed for training, and the sampling probability of difficult examples is increased during the training process to reduce the impact of new diagnostic tasks on the performance of existing diagnostic tasks and ensure the overall diagnostic effect of the model in multi-task scenarios.

[0076] In one feasible implementation, step S02 may include steps S021-S022: Step S021: Perform balanced sampling on the multi-task training sample set according to the preset task-level sampling weights and category-level sampling weights to obtain the balanced multi-task training sample set. Task-level sampling weights refer to the weight coefficients assigned to each diagnostic task in the multi-task training sample set, used to control the probability of the training sample corresponding to that diagnostic task being selected in each sampling. Category-level sampling weights refer to the weight coefficients assigned to each candidate diagnostic conclusion label within the same diagnostic task, used to control the probability of the training sample corresponding to that candidate diagnostic conclusion label being selected in each sampling. Balanced sampling refers to the process of selectively sampling the multi-task training sample set according to preset task-level and category-level sampling weights, so that the sample distribution among different diagnostic tasks and among candidate diagnostic conclusion labels within the same diagnostic task tends to be balanced after sampling. Balanced sampling does not change the content of the original training sample set, but only changes the frequency with which each sample is fed into the model for parameter updates during training.

[0077] When performing balanced sampling on the multi-task training sample set according to preset task-level sampling weights and category-level sampling weights, before the start of each training iteration, the comprehensive sampling probability of each training sample can be calculated based on the task-level sampling weight of the diagnostic task to which it belongs and the category-level sampling weight of the candidate diagnostic conclusion label to which it belongs. Then, training samples are randomly selected from the multi-task training sample set according to the comprehensive sampling probability to form the training data of the current batch, so that diagnostic tasks and fault categories with smaller sample sizes have more opportunities to appear in the training batch. Alternatively, after each training cycle, the diagnostic performance index of each diagnostic task on the validation set can be statistically analyzed. For diagnostic tasks with diagnostic performance indices significantly lower than the average level, their task-level sampling weights are automatically increased; for candidate diagnostic conclusion labels with low diagnostic performance indices within the same diagnostic task, their category-level sampling weights are automatically increased, increasing the attention to weak tasks and weak categories in the next training cycle, forming an adaptive balanced sampling strategy. This embodiment does not specifically limit the implementation of step S021.

[0078] Step S022: Fine-tune the pre-trained temporal large model based on the multi-task training sample set after equalization sampling; Among them, task-level sampling weights are used to balance the distribution of sample numbers among different diagnostic tasks, while category-level sampling weights are used to balance the distribution of sample numbers among different candidate diagnostic conclusion labels within the same diagnostic task.

[0079] This implementation introduces a balanced sampling mechanism before model fine-tuning, adjusting the sample distribution between diagnostic tasks and among candidate diagnostic conclusion labels within each diagnostic task according to preset task-level and category-level sampling weights. This effectively avoids the problem in multi-task mixed training scenarios where diagnostic tasks with large sample sizes or high-frequency fault categories dominate model parameter updates, leading to the model ignoring diagnostic tasks with small sample sizes or low-frequency fault categories. This ensures that the model has equal learning opportunities for each diagnostic task and each fault category, thereby improving the overall diagnostic balance and generalization ability of the time-series model in multi-task, multi-category scenarios.

[0080] In addition, when managing multi-task training sample sets, training samples with inconsistent conclusions between the mechanistic model and the time series model, low confidence, or significant disputes during manual review can be included in the difficult sample pool for focused use in subsequent training.

[0081] Step S03: With the goal of minimizing the training loss value of each training sample, iteratively update the model parameters of the pre-trained temporal large model to generate the target temporal large model.

[0082] This embodiment achieves joint learning of multiple inverter fault diagnosis tasks within the same model by constructing a unified multi-task training sample set and fine-tuning the pre-trained time-series model using a loss mask mechanism. During training, the loss mask only calculates the training loss value for the true diagnostic conclusion label portion of the answer text, focusing the model's parameter optimization on improving the ability to identify correct conclusions from the candidate label set, rather than the ability to generate the entire text sequence. This training method is consistent with the approach used in online inference, where diagnostic conclusions are determined by matching loss values ​​within the candidate label set, avoiding a misalignment between the training and inference objectives. This allows the generated target time-series model to stably and accurately output diagnostic conclusions under various diagnostic tasks. Furthermore, since multiple diagnostic tasks share the same base model for training, common patterns between different tasks can be shared and learned by the model, while the differentiated judgment boundaries of each diagnostic task can be mastered by the model through guidance from sample task information, effectively improving the overall generalization ability and diagnostic coverage of the model.

[0083] Based on the first, second, third, fourth, and / or fifth embodiments described above, a sixth embodiment of the inverter fault diagnosis method of this application is proposed. In the sixth embodiment, after step S03, the inverter fault diagnosis method may further include steps S04 to S06: Step S04: After obtaining the training sample set and validation sample set for the new diagnostic task, perform incremental fine-tuning on the target time series model based on the training sample set for the new diagnostic task. The training sample set for the newly added diagnostic task is a set of samples constructed for incremental fine-tuning of the target time-series large model, specifically for new fault types or new diagnostic needs. The format of each training sample in this training sample set is consistent with that of the multi-task training sample set in step S01, and also includes sample task information, a set of candidate diagnostic conclusion label samples, runtime sequence sample data, and answer text.

[0084] The validation sample set for the new diagnostic task is a set of samples that complements the training sample set for the new diagnostic task and is used to evaluate the diagnostic performance of the incrementally fine-tuned target time-series model on the new diagnostic task. The sample format of this validation sample set is consistent with that of the training sample set for the new diagnostic task, but the samples in the validation sample set do not overlap with those in the training sample set to ensure the objectivity of the validation results.

[0085] Incremental fine-tuning of the target time-series model involves making small updates to the model parameters using the training sample set of the newly added diagnostic task, based on the already trained target time-series model. Incremental fine-tuning does not retrain the model from scratch, but rather gradually enables it to acquire diagnostic capabilities for the new diagnostic task while preserving its existing diagnostic capabilities.

[0086] In one feasible implementation, training samples for new diagnostic tasks can be continuously collected. When the accumulated number of samples reaches a preset minimum training quantity threshold, an incremental fine-tuning process is automatically triggered. At this time, the current model parameters can be loaded from the target time-series large model that is already running online as initial parameters. Using the training sample set of the new diagnostic tasks, the model parameters are updated for a limited number of rounds using the loss mask fine-tuning method consistent with step S02, so that the model can learn the decision boundaries of the new diagnostic tasks while retaining its existing diagnostic capabilities.

[0087] In another feasible implementation, a fixed update cycle can be set. At the end of each cycle, the training samples of all newly added diagnostic tasks collected during that cycle are summarized, and the target time-series model is incrementally fine-tuned. This implementation is suitable for operation and maintenance scenarios where the growth of newly added diagnostic task samples is stable and the real-time requirements are not high.

[0088] Step S05: Based on the validation sample set of the newly added diagnostic task and the validation sample set associated with the target time series large model before incremental fine-tuning, verify the effectiveness of the target time series large model after incremental fine-tuning. The validation sample set associated with the target time series model before incremental fine-tuning is the validation sample set consisting of samples from existing diagnostic tasks that were saved and associated with this version of the model when the target time series model was initially trained. This validation sample set is used to detect whether incremental fine-tuning has caused a degradation in the original diagnostic capabilities of the model.

[0089] Validation refers to the process of quantitatively evaluating the target time series model after incremental fine-tuning. The evaluation content may include: whether the model's diagnostic performance on new diagnostic tasks has reached a usable level, and whether the model's diagnostic performance on existing diagnostic tasks has maintained or improved the level before fine-tuning.

[0090] In a first feasible implementation, step S05 may include: comparing the diagnostic performance indicators of the target time-series large model before and after incremental fine-tuning based on the validation sample set associated with the target time-series large model before incremental fine-tuning, thereby confirming whether the model's diagnostic performance for existing diagnostic tasks has significantly decreased; evaluating whether the diagnostic performance of the target time-series large model for the new diagnostic tasks after incremental fine-tuning reaches a preset usable standard based on the validation sample set of the new diagnostic tasks; if both validations pass, the validity validation is deemed successful; otherwise, the validity validation fails. In a second feasible implementation, step S05 may further include: based on the first feasible implementation of step S05, additionally introducing the validation sample set of the new diagnostic tasks to validate the target time-series large model before incremental fine-tuning, comparing the performance of the target time-series large model for the new diagnostic tasks before and after incremental fine-tuning, to quantify the performance improvement brought about by incremental fine-tuning, as one of the comprehensive reference bases for validity validation. This embodiment does not specifically limit the implementation of step S05.

[0091] Step S06: If the validity verification passes, run the incrementally fine-tuned target time series model online.

[0092] If the validity verification is successful, the online target time-series model can be directly replaced with the incrementally fine-tuned target time-series model, and all subsequent diagnostic tasks will use the new version of the model for inference. Alternatively, the incrementally fine-tuned target time-series model can be deployed to at least a few sites or some diagnostic tasks for trial operation, and diagnostic performance indicators can be continuously monitored. During the trial operation, the remaining sites or tasks will still use the target time-series model before incremental fine-tuning. After the trial operation confirms stability, all sites and diagnostic tasks can be gradually switched to the incrementally fine-tuned target time-series model.

[0093] This embodiment achieves continuous expansion and smooth upgrades of the diagnostic model by introducing incremental fine-tuning and validity verification mechanisms after the target time-series large-scale model is generated and put into operation. When new diagnostic task requirements arise, the target time-series large-scale model already in operation is incrementally fine-tuned using the training sample set of the new diagnostic task. This allows the target time-series large-scale model to gradually master the diagnostic capabilities for new fault types while retaining its existing diagnostic capabilities. After the incremental fine-tuning is completed, the model's learning effect on the new task and its retention effect on the existing task are tested using dual validation sets. Only after both validations pass is the incrementally fine-tuned target time-series large-scale model put into online operation. This allows the diagnostic coverage of the inverter fault diagnosis system to gradually expand with actual needs. The access of new fault types no longer requires retraining and deploying a complete model, reducing the time cost and computational resource consumption of system expansion. At the same time, validity verification ensures the stability of the existing diagnostic capabilities of the target time-series large-scale model during continuous evolution.

[0094] Based on the sixth embodiment described above, a seventh embodiment of the inverter fault diagnosis method of this application is proposed. In the seventh embodiment, after step S05, the inverter fault diagnosis method may further include steps S07 to S08: Step S07: If the validity verification fails, run the incrementally fine-tuned target time series model in offline assisted mode; Step S08: The diagnostic report output by the incrementally fine-tuned target time series large model is sent to the manual review terminal as a review reference.

[0095] The diagnostic report output by the incrementally fine-tuned target time-series large model includes at least the final diagnostic conclusion, and may also include the matching loss value or confidence level corresponding to each candidate diagnostic conclusion label. This embodiment does not specifically limit this. The manual review terminal refers to the terminal or system interface that receives review reference information and allows maintenance personnel to make manual judgments. The manual review terminal can be the maintenance workstation of the central control center, the maintenance application on a mobile terminal, or a dedicated review management system, etc. This embodiment does not specifically limit this.

[0096] Offline auxiliary mode is a model operation state distinct from online operation mode. In offline auxiliary mode, the target time-series model after incremental fine-tuning is not connected to the real-time diagnostic process. Its output diagnostic conclusions are not automatically sent to the operation and maintenance system or alarm chain as final diagnostic results, but are only provided as reference information for manual review. Offline auxiliary mode is suitable for the transitional stage where the model's effectiveness verification fails after an update, but the model has already acquired a certain diagnostic capability for new tasks.

[0097] If the validity verification fails, the incrementally fine-tuned target time-series model can be deployed to the background environment, receiving the same runtime time-series data in parallel with the online target time-series model. The incrementally fine-tuned target time-series model can perform diagnostic inference and output diagnostic reports normally in the background, but these reports are not directly pushed to the alarm system or maintenance ticket system; instead, they are sent to the manual review end for reference and comparison during manual review. Alternatively, the incrementally fine-tuned target time-series model can be deployed as an offline service, and manually triggered by the reviewer to perform diagnostic analysis on specific samples only when the reviewer needs assistance.

[0098] This embodiment runs the incrementally fine-tuned target time-series model in offline auxiliary mode when validity verification fails, and sends its output diagnostic report as review reference information to the human reviewer. Therefore, when the incrementally fine-tuned target time-series model cannot be directly deployed due to degradation of existing task performance or failure to meet new task performance standards, the model is not directly discarded but transformed into an offline auxiliary tool, whose diagnostic conclusions can still provide valuable reference information for human review. This mechanism ensures that the stability and reliability of the online diagnostic link are not impacted by insufficiently validated models, while fully utilizing the diagnostic capability improvements brought about by incremental fine-tuning, thus effectively balancing the security and evolution efficiency of the diagnostic system.

[0099] This application also provides an inverter fault diagnosis device. Please refer to... Figure 2 The inverter fault diagnosis device may include: Task determination module 10 is used to determine at least one target diagnostic task for the target inverter in response to a fault diagnosis trigger command; The data acquisition module 20 is used to acquire multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task, as well as the runtime sequence data of the target inverter within the time window associated with the target diagnostic task. The fault diagnosis module 30 is used to input the task information of the target diagnosis task, the labels of each candidate diagnosis conclusion, and the runtime sequence data into the pre-trained target time series large model, so as to calculate the matching loss value between each candidate diagnosis conclusion label and the runtime sequence data under the target diagnosis task through the target time series large model; and determine the diagnosis result of the target diagnosis task based on the matching loss value corresponding to each candidate diagnosis conclusion label.

[0100] In one embodiment, the fault diagnosis module 30 is further configured to: The matching loss value corresponding to each candidate diagnostic conclusion label is normalized to obtain the confidence level of each candidate diagnostic conclusion label; Check whether the difference between the highest confidence level and the second highest confidence level among the confidence levels of each candidate diagnostic conclusion label is greater than a preset difference threshold, and whether the highest confidence level is greater than a first preset confidence threshold. If so, then the diagnostic result of the target diagnostic task is determined as the candidate diagnostic conclusion label corresponding to the highest confidence level; If not, the diagnostic result of the target diagnostic task is determined to be pending manual review.

[0101] In one embodiment, the fault diagnosis module 30 is further configured to: Runtime sequence data is input into a pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; the mechanism model performs fault diagnosis based on preset rules. If the diagnostic report output by the mechanism model can match the diagnostic conclusion, and the confidence of the matched target diagnostic conclusion is greater than the second preset confidence threshold, verify whether the target diagnostic conclusion is consistent with the candidate diagnostic conclusion label corresponding to the highest confidence. If they match, the diagnostic result of the target diagnostic task is determined as the candidate diagnostic conclusion label corresponding to the highest confidence level.

[0102] In one embodiment, the fault diagnosis module 30 is further configured to: Runtime sequence data is input into a pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; the mechanism model performs fault diagnosis based on preset rules. If the diagnostic report output by the mechanism model can match the diagnostic conclusion, and the confidence of the matched target diagnostic conclusion is greater than the second preset confidence threshold, the target diagnostic conclusion will be used as the diagnostic result of the target diagnostic task. If the diagnostic report output by the mechanistic model cannot match a diagnostic conclusion, or if the confidence level of the matched target diagnostic conclusion is less than or equal to the second preset confidence threshold, the step of inputting the task information of the target diagnostic task, the labels of each candidate diagnostic conclusion, and the runtime sequence data into the pre-trained target time series large model is executed.

[0103] In one embodiment, the inverter fault diagnosis device further includes a model building module 40, used for: Obtain a multi-task training sample set. Each training sample in the multi-task training sample set includes sample task information, a sample set of candidate diagnostic conclusion labels, runtime sequence sample data, and answer text. The answer text is composed of a preset fixed answer prefix and the real diagnostic conclusion label. The pre-trained temporal model is fine-tuned based on a multi-task training sample set. During the fine-tuning process, the training loss value of the true diagnostic conclusion label in the answer text of each training sample is calculated by using a loss mask. With the goal of minimizing the training loss of each training sample, the model parameters of the pre-trained temporal large model are iteratively updated to generate the target temporal large model.

[0104] In one embodiment, the model building module 40 is further configured to: The multi-task training sample set is balanced by performing equal sampling according to the preset task-level sampling weights and category-level sampling weights to obtain the balanced multi-task training sample set. Fine-tuning of the pre-trained temporal large model is performed based on the multi-task training sample set after equalization sampling. Among them, task-level sampling weights are used to balance the distribution of sample numbers among different diagnostic tasks, while category-level sampling weights are used to balance the distribution of sample numbers among different candidate diagnostic conclusion labels within the same diagnostic task.

[0105] In one embodiment, the model building module 40 is further configured to: Having obtained the training and validation sample sets for the new diagnostic task, the target time series model is incrementally fine-tuned based on the training sample set for the new diagnostic task. The effectiveness of the target time series model after incremental fine-tuning is verified based on the validation sample set of the newly added diagnostic task and the validation sample set associated with the target time series model before incremental fine-tuning. If the validity verification is successful, the target time series model with incremental fine-tuning is run online.

[0106] In one embodiment, the model building module 40 is further configured to: If the validity verification fails, the incrementally fine-tuned target time series model is run in offline assisted mode. The diagnostic report output by the incrementally fine-tuned target time series model is sent to the manual reviewer as a review reference.

[0107] In one embodiment, the data acquisition module 20 is further configured to: Using the task information of the target diagnostic task as an index, multiple candidate diagnostic conclusion tags corresponding to the target diagnostic task are retrieved from the preset multi-task diagnostic configuration table.

[0108] The inverter fault diagnosis device provided in this application can improve the accuracy of inverter fault diagnosis and reduce the maintenance cost of the inverter fault diagnosis system. Compared with the prior art, the beneficial effects of the inverter fault diagnosis device provided in this application are the same as those of the inverter fault diagnosis method provided in the above embodiments, and other technical features in the inverter fault diagnosis device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0109] This application also provides an electronic device, which may include: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the inverter fault diagnosis method in the above embodiments.

[0110] The following is for reference. Figure 3 It shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of this application. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0111] like Figure 3As shown, the electronic device may include a processing unit 101 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory 102 or a program loaded from storage device 103 into random access memory 104. Random access memory 104 also stores various programs and data required for the operation of the electronic device. The processing unit 101, read-only memory 102, and random access memory 104 are interconnected via bus 105. Input / output interface 106 is also connected to bus 105. Typically, the following systems can be connected to input / output interface 106: input devices 107 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 108 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 103 including, for example, magnetic tape, hard disks, etc.; and communication devices 109. Communication device 109 allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although the diagrams show electronic devices with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0112] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 103, or installed from read-only memory 102. When the computer program is executed by processing device 101, it performs the functions defined in the methods of the embodiments of this application.

[0113] The electronic device provided in this application, employing the inverter fault diagnosis method described in the above embodiments, can improve the accuracy of inverter fault diagnosis and reduce the maintenance cost of the inverter fault diagnosis system. Compared with the prior art, the beneficial effects of the electronic device provided in this application are the same as those of the inverter fault diagnosis method provided in the above embodiments, and other technical features in the electronic device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0114] It should be understood that various parts of the embodiments of this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0115] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the above claims.

[0116] This application also provides a computer-readable storage medium storing a computer program that can run on a processor. The computer program is used to execute the inverter fault diagnosis method in the above embodiments.

[0117] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0118] The aforementioned computer-readable storage medium may be included in an electronic device or may exist independently without being assembled into an electronic device.

[0119] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by an electronic device, the electronic device causes the electronic device to: respond to a fault diagnosis trigger command, determine at least one target diagnostic task for the target inverter; acquire multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task, and runtime timing data of the target inverter within the time window associated with the target diagnostic task; input the task information of the target diagnostic task, each candidate diagnostic conclusion label, and the runtime timing data into a pre-trained target time-series large model, so as to calculate the matching loss value between each candidate diagnostic conclusion label and the runtime timing data under the target diagnostic task through the target time-series large model; and determine the diagnostic result of the target diagnostic task based on the matching loss value corresponding to each candidate diagnostic conclusion label.

[0120] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0121] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0122] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0123] The computer-readable storage medium provided in this application embodiment stores computer-readable program instructions for executing the above-described inverter fault diagnosis method, which can improve the accuracy of inverter fault diagnosis and reduce the maintenance cost of the inverter fault diagnosis system. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application embodiment are the same as the beneficial effects of the inverter fault diagnosis method provided in the above embodiments, and will not be repeated here.

[0124] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the inverter fault diagnosis method described above.

[0125] The computer program product provided in this application can improve the accuracy of inverter fault diagnosis and reduce the maintenance cost of the inverter fault diagnosis system. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the inverter fault diagnosis method provided in the above embodiments, and will not be repeated here.

[0126] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for diagnosing inverter faults, characterized in that, The method includes: In response to a fault diagnosis trigger command, determine at least one target diagnostic task for the target inverter; Obtain multiple candidate diagnostic conclusion tags corresponding to the target diagnostic task, as well as the runtime sequence data of the target inverter within the time window associated with the target diagnostic task; The task information of the target diagnostic task, each of the candidate diagnostic conclusion labels, and the runtime sequence data are input into a pre-trained target time series large model, so that the target time series large model can calculate the matching loss value between each of the candidate diagnostic conclusion labels and the runtime sequence data under the target diagnostic task. The diagnostic result of the target diagnostic task is determined based on the matching loss value corresponding to each of the candidate diagnostic conclusion labels.

2. The method as described in claim 1, characterized in that, The step of determining the diagnostic result of the target diagnostic task based on the matching loss value corresponding to each of the candidate diagnostic conclusion labels includes: The matching loss value corresponding to each candidate diagnostic conclusion label is normalized to obtain the confidence level of each candidate diagnostic conclusion label; The difference between the highest confidence level and the second highest confidence level among the confidence levels of each candidate diagnostic conclusion label is detected to be greater than a preset difference threshold, and the highest confidence level is greater than a first preset confidence threshold. If so, the diagnostic result of the target diagnostic task is determined to be the candidate diagnostic conclusion label corresponding to the highest confidence level; If not, the diagnostic result of the target diagnostic task is determined to be pending manual review.

3. The method as described in claim 2, characterized in that, Before the step of determining the diagnostic result of the target diagnostic task as the candidate diagnostic conclusion label corresponding to the highest confidence level, the method further includes: The runtime sequence data is input into a pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; wherein, the mechanism model performs fault diagnosis based on preset rules; If the diagnostic report output by the mechanism model can match a diagnostic conclusion, and the confidence level of the matched target diagnostic conclusion is greater than the second preset confidence threshold, then verify whether the target diagnostic conclusion is consistent with the candidate diagnostic conclusion label corresponding to the highest confidence level. If they match, the diagnostic result of the target diagnostic task is determined to be the candidate diagnostic conclusion label corresponding to the highest confidence level.

4. The method as described in claim 1, characterized in that, Before the step of inputting the task information of the target diagnostic task, the candidate diagnostic conclusion labels, and the runtime sequence data into the pre-trained target time-series large model, the method further includes: The runtime sequence data is input into a pre-trained mechanism model to obtain a diagnostic report output by the mechanism model; wherein, the mechanism model performs fault diagnosis based on preset rules; If the diagnostic report output by the mechanism model can match a diagnostic conclusion, and the confidence level of the matched target diagnostic conclusion is greater than the second preset confidence threshold, then the target diagnostic conclusion is taken as the diagnostic result of the target diagnostic task. If the diagnostic report output by the mechanistic model cannot match a diagnostic conclusion, or if the confidence level of the matched target diagnostic conclusion is less than or equal to the second preset confidence threshold, the step of inputting the task information of the target diagnostic task, the labels of each candidate diagnostic conclusion, and the runtime sequence data into the pre-trained target time series large model is executed.

5. The method according to any one of claims 1 to 4, characterized in that, The construction steps of the target time series large model include: Obtain a multi-task training sample set. Each training sample in the multi-task training sample set includes sample task information, a candidate diagnostic conclusion label sample set, runtime sequence sample data, and answer text. The answer text is composed of a preset fixed answer prefix and a real diagnostic conclusion label. The pre-trained temporal model is fine-tuned based on the multi-task training sample set, and the training loss value of the real diagnostic conclusion label in the answer text of each training sample is calculated by using a loss mask during the fine-tuning process. With the goal of minimizing the training loss value of each training sample, the model parameters of the pre-trained temporal large model are iteratively updated to generate the target temporal large model.

6. The method as described in claim 5, characterized in that, The step of fine-tuning the pre-trained temporal large model based on the multi-task training sample set includes: The multi-task training sample set is subjected to balanced sampling according to the preset task-level sampling weight and category-level sampling weight to obtain the balanced multi-task training sample set. Based on the balanced sampling of the multi-task training sample set, the pre-trained temporal large model is fine-tuned. The task-level sampling weight is used to balance the sample quantity distribution among different diagnostic tasks, and the category-level sampling weight is used to balance the sample quantity distribution among different candidate diagnostic conclusion labels within the same diagnostic task.

7. The method as described in claim 5, characterized in that, After the step of generating the target time series large model, the method further includes: Having obtained the training and validation sample sets for the new diagnostic task, the target time series model is incrementally fine-tuned based on the training sample set for the new diagnostic task. Based on the validation sample set of the newly added diagnostic task and the validation sample set associated with the target time series large model before incremental fine-tuning, the effectiveness of the target time series large model after incremental fine-tuning is verified. If the validity verification is successful, the incrementally fine-tuned target time series model is run online.

8. The method as described in claim 7, characterized in that, After the step of validating the effectiveness of the incrementally fine-tuned target time-series large model based on the validation sample set associated with the newly added diagnostic task and the validation sample set associated with the target time-series large model before incremental fine-tuning, the method further includes: If the validity verification fails, the incrementally fine-tuned target time series model is run in offline auxiliary mode. The diagnostic report output by the incrementally fine-tuned target time series model is sent to the manual reviewer as a review reference.

9. The method according to any one of claims 1 to 4, characterized in that, The step of obtaining multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task includes: Using the task information of the target diagnostic task as an index, multiple candidate diagnostic conclusion tags corresponding to the target diagnostic task are retrieved from the preset multi-task diagnostic configuration table.

10. An inverter fault diagnosis device, characterized in that, The inverter fault diagnosis device includes: The task determination module is used to determine at least one target diagnostic task for the target inverter in response to a fault diagnosis trigger command. The data acquisition module is used to acquire multiple candidate diagnostic conclusion labels corresponding to the target diagnostic task, as well as the runtime sequence data of the target inverter within the time window associated with the target diagnostic task; The fault diagnosis module is used to input the task information of the target diagnosis task, each of the candidate diagnosis conclusion labels, and the runtime sequence data into a pre-trained target time series large model, so as to calculate the matching loss value between each of the candidate diagnosis conclusion labels and the runtime sequence data under the target diagnosis task through the target time series large model; and to determine the diagnosis result of the target diagnosis task based on the matching loss value corresponding to each of the candidate diagnosis conclusion labels.

11. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the inverter fault diagnosis method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the inverter fault diagnosis method as described in any one of claims 1 to 9.

Images