Fusion of large model and multi-source heterogeneous wind turbine fault diagnosis tracing method and system

Abstract

The present application belongs to the technical field of intelligent operation and maintenance, and relates to a method and system for fusing large models and multi-source heterogeneous wind turbine fault diagnosis and tracing. Real-time operation parameters of each monitoring subsystem of the wind turbine are obtained through a data access gateway, effective working condition segments are extracted and input into multiple heterogeneous fault detection models in parallel, preliminary warning signals output are comprehensively judged to determine fault signs and initial types, after confirmation of existence, the initial fault types and related parameters are structured into a warning query request for cross-system causal correlation analysis, cross-validation results are generated and matched with the initial fault type to retrieve the root cause tracing path, fault root components and complete fault propagation chains are output, and a fault diagnosis report is generated to automatically create a maintenance work order. The present application significantly improves the accuracy of cross-system fault correlation identification, utilizes the directed propagation characteristics of the knowledge graph to realize rapid positioning of the fault root and transparent display of the propagation path, and effectively shortens the response cycle from early warning to the issuance of maintenance tasks.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology, and relates to a method and system for tracing the faults of wind turbines that integrate large models and multi-source heterogeneous wind turbines. Background Technology

[0002] Wind turbine generators typically operate in complex and variable natural environments. Their critical components, such as gearboxes, generators, and main shaft bearings, are prone to progressive or sudden failures under long-term alternating loads, including mechanical wear, fatigue cracks, and insulation aging. Most existing wind turbine condition monitoring systems employ an independent data acquisition and analysis model for each subsystem. Vibration monitoring, temperature monitoring, electrical parameter monitoring, and oil level monitoring subsystems each issue alarm signals based on their own preset fixed criteria. However, because the monitored physical quantities, sampling frequencies, and alarm threshold settings of each subsystem differ, this fragmented monitoring approach easily generates numerous isolated alarms or false alarms. Furthermore, it fails to identify the true origin of the fault from the interconnected changes in multiple physical field parameters. This makes it difficult for maintenance personnel to quickly locate the root cause of the fault when faced with massive amounts of alarm information, often requiring them to check each subsystem one by one, consuming significant manpower and time.

[0003] To address the aforementioned problem of multi-system alarm fusion, some existing technologies attempt to aggregate the outputs of various monitoring subsystems onto a unified data platform, employing voting mechanisms or weighted scoring methods to comprehensively evaluate multiple alarm signals. While these methods formally achieve the convergence of multi-source information, their evaluation rules remain based on pre-set static weights or fixed voting thresholds, failing to fully consider the causal timing relationships and parameter change directions of different fault types propagating across subsystems. More critically, the early warning logic of existing systems relies entirely on offline configurations of historical fault case libraries or expert rule bases. When novel fault combinations or cross-system abnormal linkage patterns emerge that are not explicitly defined in the rule base, the system lacks the ability for autonomous reasoning and cross-domain knowledge association, making it difficult to provide physically consistent explanations of fault propagation chains, thus resulting in insufficient credibility of fault root cause determination results.

[0004] In recent years, large language models have demonstrated powerful capabilities in general knowledge question answering and semantic understanding. However, their application in industrial equipment fault diagnosis still faces challenges such as poor domain knowledge adaptability, insufficient interpretability of output results, and difficulty in integration with existing enterprise information systems. Furthermore, existing knowledge graph technologies in fault diagnosis are largely limited to static fault phenomenon-fault cause mapping queries, failing to form a closed-loop linkage with real-time operating data, dynamic reasoning output from large language models, and maintenance management processes in enterprise resource planning systems. This results in low automation across the entire chain from fault warning to maintenance work order generation, making it difficult to directly translate fault tracing results into executable maintenance tasks. Therefore, there is an urgent need for a fault diagnosis method that can organically integrate multi-model fusion warning, cross-system reasoning of large language models, structured tracing using knowledge graphs, and interaction with enterprise work order systems. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art by providing a method for fault diagnosis and tracing of wind turbine units that integrates large-scale models and multi-source heterogeneous wind turbines, including: S1. Obtain the real-time operating parameters of each monitoring subsystem of the wind turbine through the data access gateway, and extract effective operating condition segments from the real-time operating parameters according to the preset operating condition filtering conditions. S2, input the effective operating condition segment into multiple heterogeneous fault detection models respectively, obtain the preliminary warning signals output by each model, comprehensively evaluate each preliminary warning signal, and determine whether the wind turbine currently has fault symptoms and the initial fault type corresponding to the fault symptoms; S3, when the fault symptoms are confirmed, the initial fault type and the relevant operating parameters in the effective operating condition segment are constructed into an early warning query request and sent to the large language model service interface. The large language model performs cross-system comparison analysis on the early warning query request based on its internal knowledge and generates a cross-validation result containing cross-system correlation information. S4, match the cross-validation results with the initial fault type. When the match is successful, use the cross-system association information in the cross-validation results as an index to retrieve the root cause tracing path corresponding to the fault symptoms from the fault knowledge graph, and output the fault root cause component and fault propagation chain indicated by the root cause tracing path. S5. Based on the fault root cause component and fault propagation chain, generate a fault diagnosis report for the wind turbine and push the fault diagnosis report to the work order management module of the enterprise resource planning system, whereby the work order management module creates a maintenance work order.

[0006] Further, S1 includes: The data access gateway establishes communication links with the vibration monitoring subsystem, temperature monitoring subsystem, electrical parameter monitoring subsystem, and oil monitoring subsystem of the wind turbine, respectively. Through each of the communication links, it synchronously acquires the real-time operating parameters output by each monitoring subsystem at different data acquisition frequencies. The real-time operating parameters include the timestamp of each monitoring subsystem. For each monitoring subsystem, the real-time operating parameters are time-aligned according to their timestamps to form a multi-dimensional operating parameter sequence under a unified time reference. A preset set of operating condition filtering conditions is read. The set of operating condition filtering conditions contains multiple discrimination conditions for determining whether the wind turbine is currently in a steady-state operating condition. The parameter value corresponding to each time point in the multi-dimensional operating parameter sequence is compared with the discrimination conditions one by one. A continuous time interval that simultaneously satisfies all discrimination conditions is selected. The multi-dimensional operating parameter sequence within the continuous time interval is extracted as an effective operating condition segment.

[0007] Further, S2 includes: The effective operating condition segment is simultaneously distributed to the detection model based on signal waveform analysis, the detection model based on statistical threshold discrimination, and the detection model based on physical mechanism. Each detection model processes the effective operating condition segment, and each detection model independently outputs a preliminary warning signal. The preliminary warning signal includes the fault existence marker determined by the model and the fault type marker determined by the model. Collect all preliminary warning signals output by the detection models, count the number of detection models marked as having a fault, and determine that the wind turbine currently has fault symptoms when the number exceeds a preset majority consensus threshold. When a fault symptom is detected, the fault type labels output by each detection model are voted on and statistically analyzed. The fault type label with the most votes is determined as the initial fault type corresponding to the fault symptom. At the same time, the characteristic fragments of operating condition parameters on which each detection model outputs the initial fault type are recorded as supporting evidence for the initial fault type.

[0008] Further, S3 includes: After confirming the existence of the fault symptoms, an operating parameter sequence related to the initial fault type is extracted from the effective operating condition segment. The operating parameter sequence includes the operating parameters of the monitoring subsystem involved in the initial fault type and the operating parameters of other monitoring subsystems that have energy transfer association with the monitoring subsystem. The name of the initial fault type and the operating parameter sequence are combined into a structured query text. The structured query text includes a fault type description field, a parameter value field, and a time evolution description field. The structured query text is sent as an early warning query request to the deployed large language model service interface, instructing the large language model to perform cross-system causal correlation analysis on the early warning query request. In response to the instruction, the large language model extracts historical fault cases that match the fault type description field and the parameter value field from its internal parameterized knowledge, and parses the parameter linkage change relationship between multiple subsystems before and after the fault from the historical fault cases. The parameter linkage change relationship is organized into a cross-system causal link description, wherein each cross-system causal link description includes upstream parameter change nodes and downstream parameter change nodes and the causal transmission direction between them. The large language model also retrieves the handling measures and verification conclusions for the fault type corresponding to the early warning query request recorded in the historical fault cases based on the cross-system causal link description, merges the handling measures and verification conclusions with the cross-system causal link description, and generates cross-validation results expressed in natural language. The cross-validation results include at least the cross-system causal link description, the associated historical fault case number, and a summary of the historical handling conclusions.

[0009] Further, S4 includes: Extract the set of subsystem names and the set of parameter change directions involved in the cross-system causal link description from the cross-validation results. Match the set of subsystem names and the set of parameter change directions with the initial fault type. When all subsystems included in the set of subsystem names are associated with the initial fault type in the preset fault propagation topology table, and the parameter change direction in the set of parameter change directions is consistent with the change direction recorded in the preset fault propagation topology table, it is confirmed that the cross-validation results match the initial fault type successfully. After a successful match, the cross-system causal link description in the cross-validation results is used as the retrieval index to traverse the nodes and edges in the fault knowledge graph. Each node in the fault knowledge graph corresponds to a wind turbine component or a monitoring parameter type. Each edge corresponds to the fault propagation relationship between two nodes. The direction of the edge indicates the direction of fault propagation. Each edge also includes the operating condition label required for the fault propagation relationship to be established. The starting node is found in the fault knowledge graph to match the upstream parameter change node in the retrieval index. Starting from the starting node, the intermediate nodes that match the downstream parameter change node in the retrieval index are traversed sequentially along the edge direction until the terminating node that matches the fault root cause type indicated by the historical handling conclusion summary in the cross-validation results is reached. All nodes and edges traversed during the traversal are arranged in order to form the root cause tracing path, wherein the starting node and the intermediate nodes correspond to the transmission components in the fault propagation chain, and the terminating node corresponds to the fault root cause component.

[0010] Furthermore, in step S4, after matching the cross-validation result with the initial fault type, if the match fails, the following steps are also included: The difference information between the cross-validation results and the initial fault type is extracted into a difference feature vector, which includes fault type name inconsistency items and cross-system causal link direction inconsistency items. The difference feature vector is fed back to the large language model service interface, instructing the large language model to readjust its internal knowledge retrieval scope based on the difference feature vector, and generate updated cross-validation results based on the adjusted retrieval scope; The updated cross-validation result is matched again with the initial fault type. If the match is successful, the process is transferred to follow the procedure for successful matching. If the match is unsuccessful, a manual intervention identifier is sent to the operation and maintenance terminal. The initial fault type, the cross-validation result, and the updated cross-validation result are packaged into a data packet to be reviewed and sent to the operation and maintenance terminal.

[0011] Furthermore, the construction method of the fault knowledge graph includes: Multiple fault cases are extracted in advance from the design documents, historical fault records and maintenance manuals of the wind turbine. Each fault case includes a description of the fault phenomenon, the identification of the root cause component, the sequence of intermediate components through which the fault propagates, and the direction of parameter transmission between each intermediate component. For each fault case, the monitoring parameter type in the fault phenomenon description is taken as the first node, the fault root cause component identifier is taken as the second node, and each component in the intermediate component sequence through which the fault propagation path passes is taken as an intermediate node. According to the parameter transmission direction, a first directed edge between the first node and the first intermediate node, a second directed edge between adjacent intermediate nodes, and a third directed edge between the last intermediate node and the second node are established in sequence. Each directed edge takes the parameter transmission direction as the direction of the edge, and the operating condition mark required for the establishment of the fault propagation relationship is added to each directed edge. All nodes and directed edges established from all fault cases are merged into the same graph structure to form the fault knowledge graph. The normal value range and alarm threshold range of the monitoring parameters corresponding to each node are recorded in the fault knowledge graph, and the response delay time range between the parameter change direction of the upstream node and the parameter change direction of the downstream node is recorded on each directed edge.

[0012] Further, S5 includes: The component number, component name, and physical location of the fault-causing component, as well as the component number and component name of each component in the fault propagation chain, are arranged into a fault propagation sequence table according to the time sequence of fault propagation. The fault propagation sequence table is associated with the operating parameter change curves within the corresponding time interval of the effective operating condition segment to generate a fault diagnosis report containing parameter change curve annotations. The fault diagnosis report, in accordance with the order of the fault propagation sequence table, sequentially annotates the parameter change trend description and parameter over-limit time of each component during the fault propagation process. The fault diagnosis report is pushed to the work order management module of the enterprise resource planning system via the application programming interface. After receiving the fault diagnosis report, the work order management module parses the fault root cause component number in the fault diagnosis report, queries the corresponding spare parts inventory information and maintenance tool information from the spare parts inventory database of the enterprise resource planning system according to the fault root cause component number, and merges the spare parts inventory information and maintenance tool information with the fault diagnosis report. The work order management module generates a maintenance work order that includes at least the fault root cause component information, the fault propagation sequence table, the spare parts list, and the maintenance tool list, and assigns the maintenance work order to the corresponding maintenance team terminal.

[0013] Furthermore, after comprehensively evaluating each of the preliminary warning signals to determine that the wind turbine currently exhibits signs of a fault, and before constructing a warning query request from the initial fault type and the relevant operating parameters in the effective operating condition segment, the process further includes: Based on the timestamp of the effective operating condition segment, retrieve the historical operating parameter sequence within a preset duration range before and after the corresponding time interval of the effective operating condition segment from the historical cache of the data access gateway; The historical operating parameter sequence is compared with the effective operating condition segment. The trend change slope of each monitoring parameter in the historical operating parameter sequence is calculated. Based on the trend change slope, it is determined whether the fault symptom corresponding to the effective operating condition segment is a sudden fault symptom or a gradual fault symptom. When a sudden fault symptom is identified, a sudden fault marker is added to the early warning query request, and the relevant operating parameters in the effective operating condition segment are arranged in chronological order as a mutation parameter sequence. The mutation parameter sequence is then constructed as part of the relevant operating parameters into the early warning query request. When a gradual fault symptom is identified, a gradual fault marker is added to the early warning query request, and the relevant operating parameters in the effective operating condition segment are merged into a continuously changing parameter sequence according to the time order retrieved from the historical cache and the current time order. The continuously changing parameter sequence is then constructed into the early warning query request as part of the relevant operating parameters.

[0014] According to a second aspect of the present invention, the present invention claims protection for a fault diagnosis and tracing system for integrated large-scale models and multi-source heterogeneous wind turbines, comprising: One or more processors; A memory storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the aforementioned method for tracing and diagnosing faults in a large-scale integrated model and multi-source heterogeneous wind turbines.

[0015] This invention belongs to the field of intelligent operation and maintenance technology, and relates to a method and system for fault diagnosis and tracing of wind turbine units that integrates large-scale models and multi-source heterogeneous systems. It acquires real-time operating parameters of each monitoring subsystem of the wind turbine unit through a data access gateway, extracts effective operating condition segments, and inputs them in parallel into multiple heterogeneous fault detection models. The output preliminary warning signals are comprehensively evaluated to determine fault symptoms and their initial types. After confirmation, the initial fault type and related parameters are constructed into a warning query request for cross-system causal correlation analysis. Cross-validation results are generated, matched with the initial fault type, and the root cause tracing path is retrieved. The root cause component and the complete fault propagation chain are output, and a fault diagnosis report is generated, automatically creating a maintenance work order. This invention significantly improves the accuracy of cross-system fault correlation identification, utilizes the directed propagation characteristics of knowledge graphs to achieve rapid fault root cause location and transparent display of propagation paths, effectively shortening the response cycle from warning to maintenance task issuance. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the workflow of a fault diagnosis and tracing method for a multi-source heterogeneous wind turbine that integrates a large model and multiple sources, as claimed in an embodiment of the present invention. Figure 2 The second workflow diagram is shown for a fault diagnosis and tracing method for a fusion large model and multi-source heterogeneous wind turbine generator claimed in an embodiment of the present invention. Figure 3 The third flowchart is a method for tracing the faults of a large-scale integrated model and a multi-source heterogeneous wind turbine, as claimed in the embodiments of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application.

[0018] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a method for fault diagnosis and tracing of wind turbine units that integrates large-scale models and multi-source heterogeneous wind turbines, comprising: S1. Obtain the real-time operating parameters of each monitoring unit of the wind turbine through the data access gateway, and extract effective operating condition segments from the real-time operating parameters according to the preset operating condition filtering conditions. S2, input the effective operating condition segment into multiple heterogeneous fault detection models respectively, obtain the preliminary warning signals output by each model, comprehensively evaluate each preliminary warning signal, and determine whether the wind turbine currently has fault symptoms and the initial fault type corresponding to the fault symptoms; S3, when the fault symptoms are determined to exist, the initial fault type and the relevant operating parameters in the effective operating condition segment are constructed into an early warning query request and sent to the large language model service interface. The large language model performs cross-comparison analysis on the early warning query request based on its internal knowledge and generates cross-validation results containing cross-correlation information. S4, match the cross-validation results with the initial fault type. When the match is successful, use the cross-association information in the cross-validation results as an index to retrieve the root cause tracing path corresponding to the fault symptoms from the fault knowledge graph, and output the fault root cause component and fault propagation chain indicated by the root cause tracing path. S5. Based on the fault root cause component and fault propagation chain, generate a fault diagnosis report for the wind turbine and push the fault diagnosis report to the work order management module of the enterprise resource planning system, whereby the work order management module creates a maintenance work order.

[0019] In this embodiment, a 3.0 MW doubly-fed asynchronous wind turbine generator set in a wind farm is used as the implementation object. This generator set is equipped with vibration monitoring, temperature monitoring, electrical parameter monitoring, and oil level monitoring. During generator operation, a data access gateway, acting as a front-end data aggregation device, establishes communication connections with the data acquisition terminals of the four monitoring units via an industrial Ethernet network. The vibration monitoring terminal is installed at the generator's main shaft bearing housing and the planetary position of the gearbox, outputting vibration acceleration waveforms at a sampling rate of 5120 points per second. The temperature monitoring unit outputs temperature readings every 200 milliseconds using platinum resistance thermometers distributed in the generator windings, gearbox oil sump, and main bearing. The electrical parameter monitoring unit outputs the effective values ​​and phase angles of three-phase voltage and current every 20 milliseconds using voltage and current transformers. The oil level monitoring unit outputs the concentration of metal particles and moisture content in the oil every 10 minutes using an online particle counter and moisture sensor in the gearbox lubrication oil circuit.

[0020] After receiving the real-time operating parameters from each component, the data access gateway aligns data from different sampling frequencies on the time axis based on the unified timestamp contained in each data packet and a GPS-synchronized clock. For example, it aligns every 512 sampling points of the vibration waveform with the time points of the temperature readings, forming a unified time series in milliseconds. Subsequently, the data access gateway reads a pre-stored set of operating condition screening conditions from its internal memory. This set includes: the unit's output power is stable between 85% and 95% of the rated power, the wind speed change rate does not exceed 0.5 meters per second, and the generator speed fluctuation does not exceed 2% of the rated speed. The data access gateway iterates through the power value, wind speed change rate, and speed value at each time point in the aligned time series, comparing them one by one with the above conditions. When all conditions are met at all time points within a continuous period, this continuous time interval is marked as a steady-state operating condition. For example, if all parameters meet the conditions within a 330-second interval from 14:23:15 to 14:28:45, the data access gateway extracts all multi-dimensional operating parameter sequences within this interval, names them as valid operating condition segment A, and temporarily stores them in the gateway's circular buffer. At the same time, the start and end timestamps of this segment are recorded in the log file.

[0021] Furthermore, referring to Figure 2 S1 includes: The data access gateway establishes communication links with the vibration monitoring unit, temperature monitoring unit, electrical parameter monitoring unit, and oil monitoring unit of the wind turbine, respectively. Through each of the communication links, it synchronously acquires the real-time operating parameters output by each monitoring unit at different data acquisition frequencies. The real-time operating parameters include the timestamp of each monitoring unit. For each monitoring sub-database, the real-time operating parameters are time-aligned according to their timestamps to form a multi-dimensional operating parameter sequence under a unified time reference. Read the preset set of operating condition filtering conditions, which contains multiple discrimination conditions for determining whether the wind turbine is currently in a steady-state operating condition; The parameter value corresponding to each time point in the multi-dimensional operating parameter sequence is compared with the discrimination conditions one by one, and the continuous time interval that simultaneously satisfies all the discrimination conditions is selected. The multi-dimensional operating parameter sequence in the continuous time interval is extracted as an effective operating condition segment.

[0022] In this embodiment, the data access gateway is configured with four independent communication ports, each corresponding to one of the four monitoring sub-monitors. Upon startup, the data access gateway first sends a synchronization handshake message to the data acquisition terminal of the vibration monitoring sub-monitor. After a successful handshake, a real-time data stream channel based on the UDP protocol is established, and vibration data continuously flows into the gateway's receiving buffer at a frequency of 5120 Hz. Simultaneously, the gateway sends a read command to the RS-485 bus of the temperature monitoring sub-monitor, and temperature data is returned in response at a frequency of 5 Hz. The gateway sends a periodic polling request to the Modbus / TCP server of the electrical parameter monitoring sub-monitor, acquiring a set of three-phase electrical parameters every 50 milliseconds. The gateway sends a query command to the serial interface of the oil monitoring sub-monitor, acquiring oil analysis results every 10 minutes.

[0023] Because the data arrival rates of each data packet differ, a timestamp alignment module is installed inside the gateway. This module uses the GPS second pulse as a reference and normalizes the microsecond-level timestamp attached to each data packet. The gateway uses every 20 milliseconds as a basic time slot. Within each time slot, the root mean square value of the 512 sampling points closest to the center of that time slot in the vibration data, the temperature reading at that moment, the electrical parameter values, and the most recent oil data are combined into a multi-dimensional record if no new oil data is available in that time slot. For example, in the time slot from 14:23:15.000 to 14:23:15.020, the root mean square value of vibration is 1.2 m / s², the generator winding temperature is 78.5 degrees Celsius, the gearbox oil sump temperature is 62.3 degrees Celsius, the A-phase current is 450 amps, the voltage is 690 volts, and the oil particle concentration is 12 particles per milliliter.

[0024] Subsequently, the set of operating condition screening criteria is stored in a structured file on the gateway's solid-state drive. This file contains three discrimination rules: Rule 1, the active power of the unit must be between 2.55 MW and 2.85 MW; Rule 2, the absolute value of the change in wind speed sensor measurement within 5 consecutive seconds does not exceed 2.5 m / s; Rule 3, the change in generator speed within 5 consecutive seconds does not exceed 1.5% of the rated speed. The gateway applies these three rules to each multi-dimensional record in each time slot. When a record has a power value of 2.70 MW, a wind speed of 11.2 m / s with a maximum wind speed difference of 1.8 m / s in the previous 5 seconds, and a speed of 1780 rpm with a maximum difference of 12 rpm in the previous 5 seconds, the record meets the rules. The gateway continuously scans and finds all time slots within the 330 seconds from 14:23:15 to 14:28:45 that meet the rules. Therefore, all multi-dimensional records within this time period are copied to the valid operating condition segment storage area and the segment is numbered EF001.

[0025] Furthermore, referring to Figure 3 S2 includes: The effective operating condition segment is simultaneously distributed to the detection model based on signal waveform analysis, the detection model based on statistical threshold discrimination, and the detection model based on physical mechanism. Each detection model processes the effective operating condition segment, and each detection model independently outputs a preliminary warning signal. The preliminary warning signal includes the fault existence marker determined by the model and the fault type marker determined by the model. Collect all preliminary warning signals output by the detection models, count the number of detection models marked as having a fault, and determine that the wind turbine currently has fault symptoms when the number exceeds a preset majority consensus threshold. When a fault symptom is detected, the fault type labels output by each detection model are voted on and statistically analyzed. The fault type label with the most votes is determined as the initial fault type corresponding to the fault symptom. At the same time, the characteristic fragments of operating condition parameters on which each detection model outputs the initial fault type are recorded as supporting evidence for the initial fault type.

[0026] In this embodiment, the control simultaneously distributes the effective operating condition segment EF001 to three detection model processes, and these three models run in three independent computing containers.

[0027] The first detection model is a waveform analysis model based on vibration envelope demodulation. This model performs bandpass filtering and envelope demodulation on the vibration acceleration signal to extract the gear meshing frequency and the amplitude of its sideband. When the model processes the vibration data in EF001, it finds that the amplitude of the second harmonic sideband of the planetary meshing frequency of the gearbox is about 40% higher than the normal operating baseline. The fault mode matching table inside the model associates this feature with the gear tooth surface wear fault type. Therefore, the model outputs a preliminary warning signal, in which the fault presence is marked as a fault exists, the fault type is marked as gear tooth surface wear, and the specific frequency range of the abnormal increase is attached.

[0028] The second detection model is a statistical threshold-based detection model. This model sets dynamic thresholds for generator winding temperature, gearbox oil sump temperature, and main shaft bearing temperature, which are automatically adjusted according to output power and cooling oil temperature. During the time period corresponding to EF001, the generator winding temperature reached a maximum of 92.3 degrees Celsius, while the dynamic threshold at this power was 90.0 degrees Celsius, and the temperature exceeded the limit for approximately 120 seconds; simultaneously, the gearbox oil sump temperature reached 68.5 degrees Celsius, while the dynamic threshold was 67.0 degrees Celsius. Based on the duration and magnitude of the temperature exceedance, the model determined that there were signs of a fault, marking the fault type as overheating fault and the presence of an output fault as a fault.

[0029] The third detection model is a physical mechanism-based model. This model incorporates the thermodynamic and dynamic coupling equations of the gearbox. After inputting the speed, torque, oil temperature, and vibration value, it calculates the gear meshing stiffness reduction rate by calculating the relationship between the thermal expansion coefficient and the bearing clearance. When the parameters of EF001 are input, the stiffness reduction rate calculated by the model reaches 12%, exceeding its internally set alarm line of 8%. Moreover, this reduction rate is positively correlated with the increase of high-frequency vibration components. Therefore, the model determines the fault type as abnormal gear pair meshing and outputs a fault existence marker as a fault exists.

[0030] The outputs of the three models were collected and the number of faults marked as present was 3, exceeding the preset majority consensus threshold of 2. Therefore, it was determined that the wind turbine unit did indeed have fault symptoms. Next, a voting statistics were performed on the fault type labels output by each model: the two models output gear tooth surface wear and gear pair meshing abnormality, respectively. These two labels actually belong to the same major category of gear faults in physical terms. There is a fault type synonym mapping table inside, which maps gear pair meshing abnormality to the major category of gear tooth surface wear. Therefore, gear tooth surface wear received 2 votes and overheating fault received 1 vote. Gear tooth surface wear received the most votes, so this fault type was determined as the initial fault type. At the same time, the frequency range provided by the waveform analysis model and the stiffness reduction rate provided by the mechanism model were stored in memory as supporting evidence.

[0031] Further, S3 includes: After confirming the presence of the fault symptoms, an operating parameter sequence related to the initial fault type is extracted from the effective operating condition segment. The operating parameter sequence includes the operating parameters of the monitoring sub-monitor involved in the initial fault type and the operating parameters of other monitoring sub-monitors that have energy transfer association with the monitoring sub-monitor. The name of the initial fault type and the operating parameter sequence are combined into a structured query text. The structured query text includes a fault type description field, a parameter value field, and a time evolution description field. The structured query text is sent as an early warning query request to the deployed large language model service interface, instructing the large language model to perform cross-causal association analysis on the early warning query request. In response to the instruction, the large language model extracts historical fault cases that match the fault type description field and the parameter value field from its internal parameterized knowledge, and parses the parameter linkage change relationship between multiple sub-faults before and after the fault occurs from the historical fault cases. The parameter linkage change relationship is organized into a cross-causal link description, wherein each cross-causal link description includes the upstream parameter change node and the downstream parameter change node and the causal transmission direction between them. The large language model also retrieves the handling measures and verification conclusions for the fault type corresponding to the early warning query request recorded in the historical fault cases based on the cross-causal link description, merges the handling measures and verification conclusions with the cross-causal link description, and generates cross-validation results expressed in natural language. The cross-validation results include at least the cross-causal link description, the associated historical fault case number, and a summary of the historical handling conclusions.

[0032] In this embodiment, after determining the initial fault type as gear tooth surface wear, the control extracts a sequence of operating parameters related to this fault type from the effective operating condition segment EF001. Specifically, based on the internally stored fault propagation topology table, it is found that the gear tooth surface wear fault type directly affects the gearbox vibration amplitude and the concentration of metal particles in the oil. The gearbox vibration is transmitted to the generator end cover through the bearing, causing generator vibration. At the same time, the increase in gearbox temperature affects the cooling efficiency and thus changes the oil temperature. Therefore, the following sequence of operating parameters is extracted: root mean square value of vibration acceleration of gearbox input shaft, root mean square value of vibration acceleration of gearbox output shaft, root mean square value of vibration acceleration of generator drive end, gearbox oil sump temperature, cooling water outlet temperature, oil ferrographic particle count, and generator winding temperature. These seven parameters are arranged in chronological order, with each parameter accompanied by a description of its changing trend. For example, the vibration of the gearbox input shaft gradually increases from 1.0 m / s² to 1.8 m / s² and remains at a high level.

[0033] The initial fault type name, gear tooth wear, is combined with the sequence of the seven parameters mentioned above to form a structured query text. This text is described in JSON format and includes a fault type field, parameter name and value range field, and time evolution description field. This structured query text is sent via HTTPS protocol to the large language model service interface deployed in the isolated area. The backend of this interface is a domain-fine-tuned large language model with 13 billion parameters. The request also includes an instruction: Based on the known knowledge of wind turbine gearbox faults, please analyze whether the provided fault characteristic parameters are causally related to other sub-factors such as electrical and cooling, and list relevant historical cases and handling conclusions.

[0034] Upon receiving the request, the large language model retrieves historical fault cases from its internal parametric knowledge that match the gear tooth wear and the provided vibration, temperature, and oil parameters. The model outputs a natural language description, indicating that case C-2023-045 in the historical case library recorded similar vibration and temperature characteristics. In this case, gear tooth wear increased meshing impact, which in turn caused gearbox vibration to be transmitted to the generator end through the foundation, resulting in uneven generator air gap and an increase in the negative sequence current component in the electrical parameters. At the same time, metal particles generated by wear enter the lubricating oil, leading to an increase in oil particle count and an increase in oil temperature due to increased frictional heat, which in turn increases the cooling load and the cooling water outlet temperature. The model organizes these relationships into one cross-causal link: Gear tooth wear (mechanical) → Increased gearbox vibration (mechanical) → Increased generator end vibration (mechanical / electrical interface) → Change in generator air gap (electrical) → Increased negative sequence current (electrical); and another link: Gear tooth wear → Increased frictional heat (thermodynamic) → Increased oil temperature (lubrication) → Increased cooling load (cooling). The model also includes the corrective action from case C-2023-045, which involves replacing the planetary gears in the gearbox and readjusting the meshing clearance. The verification conclusion is that the vibration returned to normal levels after the replacement, and the oil particle count decreased to the acceptable range after 24 hours of operation. The large language model combines all of the above to generate a complete cross-validation result, which is returned to the control system in natural language text.

[0035] Further, S4 includes: Extract the set of sub-names and the set of parameter change directions involved in the cross-causal link description from the cross-validation results. Match the set of sub-names and the set of parameter change directions with the initial fault type. When all the sub-names included in the set of sub-names are associated with the initial fault type in the preset fault propagation topology table, and the parameter change direction in the set of parameter change directions is consistent with the change direction recorded in the preset fault propagation topology table, it is confirmed that the cross-validation results match the initial fault type successfully. After a successful match, the cross-causal link description in the cross-validation results is used as the retrieval index to traverse the nodes and edges in the fault knowledge graph. Each node in the fault knowledge graph corresponds to a wind turbine component or a monitoring parameter type. Each edge corresponds to the fault propagation relationship between two nodes. The direction of the edge indicates the direction of fault propagation. Each edge also includes the operating condition label required for the fault propagation relationship to be established. The starting node is found in the fault knowledge graph to match the upstream parameter change node in the retrieval index. Starting from the starting node, the intermediate nodes that match the downstream parameter change node in the retrieval index are traversed sequentially along the edge direction until the terminating node that matches the fault root cause type indicated by the historical handling conclusion summary in the cross-validation results is reached. All nodes and edges traversed during the traversal are arranged in order to form the root cause tracing path, wherein the starting node and the intermediate nodes correspond to the transmission components in the fault propagation chain, and the terminating node corresponds to the fault root cause component.

[0036] In this embodiment, after receiving the cross-validation result text returned by the large language model, the natural language parsing module extracts a set of sub-names from the cross-causal link description, including mechanical gearbox, mechanical generator terminal, electrical negative sequence current, lubrication, and cooling. Simultaneously, it extracts a set of parameter change directions, such as increased vibration, increased temperature, increased particle count, and increased negative sequence current. The set of sub-names is then compared with a preset fault propagation topology table associated with the initial fault type, gear tooth wear. This topology table predefines downstream sub-types that gear tooth wear may trigger, including gearbox, generator, lubricating oil, and cooling, and specifies that the parameter change directions in these sub-types should be increased vibration, increased oil temperature, increased particle count, and increased cooling water temperature. Since all sub-types included in the cross-validation result appear in the topology table, and the parameter change directions are completely consistent, the cross-validation result is confirmed to be a successful match with the initial fault type.

[0037] After a successful match, the cross-causal link description in the cross-validation results is used as the retrieval index to access the fault knowledge graph stored in the graph database. The nodes of this knowledge graph include component nodes such as gearbox input shaft bearing, planetary gears, sun gear, gearbox output shaft, generator drive end bearing, generator winding, lubricating oil filter, and cooler, as well as monitoring parameter nodes such as vibration amplitude, temperature value, particle concentration, and current harmonics. Edges represent fault propagation relationships, and each edge has a direction. For example, an edge pointing from the planetary gear to the gearbox input shaft bearing indicates that a planetary gear failure will cause abnormal vibration of the input shaft bearing. This edge is also marked with operating conditions indicating that the load is greater than 80% and the speed is within the rated range.

[0038] The cross-validation results identified the initial parameter node, "Increased Vibration of the Gearbox Input Shaft." The corresponding node, "Vibration Amplitude of the Gearbox Input Shaft," was located in the knowledge graph. This node is connected to the planetary gear node via an edge pointing from planetary gear to vibration amplitude. Therefore, the planetary gear was chosen as the starting node. The process then traversed along this edge, starting from the planetary gear, passing through the gearbox output shaft node, then the generator drive bearing node, then the generator winding node, and finally the cooler node. Simultaneously, the historical handling conclusion summary in the cross-validation results mentioned replacing the gearbox planetary gears; therefore, the final termination node should point to the planetary gear component. The node sequence traversed was: planetary gear → gearbox output shaft → generator drive bearing → generator winding → cooler. The corresponding edge direction indicates that the fault originated from the planetary gear, propagated through the output shaft to the generator, causing the winding temperature to rise, which in turn affected the cooler load. The planetary gear was identified as the root cause of the fault, and the entire node sequence was output as the fault propagation chain.

[0039] Furthermore, in step S4, after matching the cross-validation result with the initial fault type, if the match fails, the following steps are also included: The difference information between the cross-validation results and the initial fault type is extracted into a difference feature vector, which includes fault type name inconsistency items and cross-causal link direction inconsistency items. The difference feature vector is fed back to the large language model service interface, instructing the large language model to readjust its internal knowledge retrieval scope based on the difference feature vector, and generate updated cross-validation results based on the adjusted retrieval scope; The updated cross-validation result is matched again with the initial fault type. If the match is successful, the process is transferred to follow the procedure for successful matching. If the match is unsuccessful, a manual intervention identifier is sent to the operation and maintenance terminal. The initial fault type, the cross-validation result, and the updated cross-validation result are packaged into a data packet to be reviewed and sent to the operation and maintenance terminal.

[0040] In this embodiment, it is assumed that in another scenario, the initial fault type is determined to be generator bearing overheating. However, the cross-causal link mentioned in the cross-validation results returned by the large language model points to excessive harmonics in the electrical system leading to increased copper losses, which in turn causes the generator to overheat. Furthermore, the model does not mention the unique vibration characteristics of the bearing. The control compares the fault type name "electrical harmonic overheating" in the cross-validation results with the initial type "generator bearing overheating" and finds that they are inconsistent. At the same time, the cross-link description in the cross-validation results is from electrical to thermodynamic, while the link direction corresponding to generator bearing overheating in the preset fault propagation topology table should be from mechanical friction to temperature rise, which is inconsistent.

[0041] These two discrepancies are extracted to form a discrepancy feature vector, which contains two fields: a type name inconsistency marker and a link direction inconsistency marker. This discrepancy feature vector is then appended to a new query request and sent again to the large language model service interface, with the instruction: "Please ignore electrical harmonic factors and focus on retrieving historical cases related to bearing friction, poor lubrication, and vibration characteristics." Upon receiving the adjusted request, the large language model re-retrieves another case, C-2022-112, from its internal knowledge base. This case describes how aging grease in a generator bearing leads to increased friction, resulting in a temperature rise, and the appearance of bearing fault characteristic frequencies in the vibration spectrum. Based on this, the model generates updated cross-validation results, where the cross-link becomes: bearing friction (mechanical) → bearing temperature rise (thermodynamic) → lubrication (decreased lubricating oil viscosity).

[0042] The control system will again match the updated cross-validation results with the initial type "generator bearing overheating". This time, the sub-name set includes "bearing" and "lubricating oil", and the parameter changes in the direction of temperature increase and viscosity decrease, which is completely consistent with the topology table, resulting in a successful match. If the second match still fails, for example, if the link returned by the model is still incorrect, an alarm message will be sent to the maintenance terminal requesting manual intervention: "The fault type cannot match the model cross-validation results. Please review manually." At the same time, the initial fault type, the first cross-validation result, the second cross-validation result, and all original operating condition parameters will be packaged into a compressed data package and sent to the shift leader's workstation via email for manual judgment.

[0043] Furthermore, the construction method of the fault knowledge graph includes: Multiple fault cases are extracted in advance from the design documents, historical fault records and maintenance manuals of the wind turbine. Each fault case includes a description of the fault phenomenon, the identification of the root cause component, the sequence of intermediate components through which the fault propagates, and the direction of parameter transmission between each intermediate component. For each fault case, the monitoring parameter type in the fault phenomenon description is taken as the first node, the fault root cause component identifier is taken as the second node, and each component in the intermediate component sequence through which the fault propagation path passes is taken as an intermediate node. According to the parameter transmission direction, a first directed edge between the first node and the first intermediate node, a second directed edge between adjacent intermediate nodes, and a third directed edge between the last intermediate node and the second node are established in sequence. Each directed edge takes the parameter transmission direction as the direction of the edge, and the operating condition mark required for the establishment of the fault propagation relationship is added to each directed edge. All nodes and directed edges established from all fault cases are merged into the same graph structure to form the fault knowledge graph. The normal value range and alarm threshold range of the monitoring parameters corresponding to each node are recorded in the fault knowledge graph, and the response delay time range between the parameter change direction of the upstream node and the parameter change direction of the downstream node is recorded on each directed edge.

[0044] In this embodiment, prior to deployment, the engineering team collected 320 wind turbine fault records from the past five years for the wind farm, as well as design documents and maintenance manuals provided by the equipment manufacturer. Each fault case was organized into a uniform format: case number, fault phenomenon description (e.g., abnormal gearbox vibration accompanied by periodic impacts), root cause component (e.g., fatigue spalling of planetary gear teeth), sequence of intermediate components through which the fault propagates (e.g., planetary gear → sun gear → gearbox output shaft → coupling → generator input shaft), and parameter transmission direction between components (e.g., planetary gear vibration → sun gear vibration → output shaft torque fluctuation → coupling misalignment → generator shaft bending).

[0045] For each case, construct the corresponding graph structure. Taking case C-2023-045 as an example, first, take the monitoring parameter type "vibration acceleration amplitude" from the fault phenomenon description as the first node, named "Vibration Amplitude Gearbox". Then, take the planetary gear, the root cause component, as the second node. Next, take the sun gear, gearbox output shaft, and coupling from the intermediate component sequence as three intermediate nodes. Then, establish directed edges according to the parameter transmission direction: from the vibration amplitude gearbox node to the sun gear node, the edge direction indicating that the change in vibration amplitude will affect the operating state of the sun gear; from the sun gear to the gearbox output shaft; from the gearbox output shaft to the coupling; and finally from the coupling to the planetary gear. Since the ultimate root cause is the planetary gear, but the propagation direction is reversed? According to the description, the actual propagation direction is from the root cause to the phenomenon, but in the construction, the first node is the phenomenon, the second node is the root cause, and the intermediate nodes are from the root cause to the phenomenon according to the propagation direction; therefore, the direction should be from the planetary gear to the coupling, output shaft, sun gear, and vibration amplitude. Based on the parameter transmission direction, the fault should originate from the planetary gear, be transmitted through the sun gear, output shaft, and coupling to the vibration amplitude, so the directed edges should be: planetary gear → sun gear, sun gear → output shaft, output shaft → coupling, coupling → vibration amplitude. Therefore, when establishing nodes and edges, the planetary gear should be the starting point, pointing sequentially to intermediate nodes, and finally to the phenomenon node. During construction, the first node should be the root component, the second node should be the phenomenon parameter, and the intermediate nodes should be the transmission components. Directed edges should be established from the root node to the first intermediate node, then to the next, and finally to the phenomenon parameter node. Operating condition markers should be added to each edge, such as load between 80% and 100% and speed stable at the rated value. Simultaneously, the response delay time range should be recorded; for example, the time delay from planetary gear vibration to sun gear vibration is approximately 0.2 milliseconds.

[0046] After the above processing, all 320 cases were merged into a single graph database. During merging, if multiple cases had the same node name, such as a planetary gear, the node was reused. If the same edge had different operating conditions, multiple operating condition tags were added to the edge. The resulting fault knowledge graph contained approximately 800 nodes and 2000 directed edges. Each node, in addition to its name, stored the normal value range and alarm threshold interval of the monitoring parameter. Each edge stored the corresponding response delay time range, such as 0.1-0.5 milliseconds. This knowledge graph was deployed in the form of a Neo4j graph database for subsequent retrieval.

[0047] Further, S5 includes: The component number, component name, and physical location of the fault-causing component, as well as the component number and component name of each component in the fault propagation chain, are arranged into a fault propagation sequence table according to the time sequence of fault propagation. The fault propagation sequence table is associated with the operating parameter change curves within the corresponding time interval of the effective operating condition segment to generate a fault diagnosis report containing parameter change curve annotations. The fault diagnosis report, in accordance with the order of the fault propagation sequence table, sequentially annotates the parameter change trend description and parameter over-limit time of each component during the fault propagation process. The fault diagnosis report is pushed to the work order management module of the enterprise resource planning system via the application programming interface. After receiving the fault diagnosis report, the work order management module parses the fault root cause component number in the fault diagnosis report, queries the corresponding spare parts inventory information and maintenance tool information from the spare parts inventory database of the enterprise resource planning system according to the fault root cause component number, and merges the spare parts inventory information and maintenance tool information with the fault diagnosis report. The work order management module generates a maintenance work order that includes at least the fault root cause component information, the fault propagation sequence table, the spare parts list, and the maintenance tool list, and assigns the maintenance work order to the corresponding maintenance team terminal.

[0048] In this embodiment, the identified root cause of the fault is the planetary gear, and the fault propagation chain is planetary gear → gearbox output shaft → generator drive end bearing → generator winding → cooler. Each component in the above chain is arranged into a fault propagation sequence table according to the propagation order. Each row in the table includes the component number obtained from the equipment ledger, the component name, and the physical location of the component. For example, the planetary gear is located inside the gearbox, in the second-stage planetary gear set, with the location number GB-02. The corresponding time interval 14:23:15 to 14:28:45 is also extracted from the effective operating condition segment EF001, and the monitoring parameter change curve of each component within this interval is associated with the chain sequence.

[0049] Specifically, a multi-page PDF document is generated. The first page is an overview, including the wind turbine number and the time of fault discovery. The second page is a fault propagation sequence table, listing the parameter change trends of the planetary gear, output shaft, bearing, winding, and cooler in chronological order. For example, in the planetary gear stage, the vibration amplitude increases from 1.0 to 1.8, and the impact pulse is obvious; in the output shaft stage, the torque fluctuation frequency is consistent with the impact; in the bearing stage, the generator drive end temperature rises to 75 degrees Celsius; in the winding stage, the winding temperature rises to 92 degrees Celsius; in the cooler stage, the cooling water outlet temperature rises from 38 to 45 degrees Celsius. Starting from the third page, parameter change curves are displayed, with each curve labeled with the time when the parameter exceeded the limit, such as 14:25:30, when the vibration amplitude first exceeded the alarm line.

[0050] After the report is generated, the PDF file and structured data, including the fault root cause component number GB-02-PLANET, are pushed to the work order management module of the Enterprise Resource Planning (ERP) system via a RESTful API. Upon receiving the data, the work order management module parses the fault root cause component number as GB-02-PLANET, then queries the ERP spare parts inventory database. It finds two planetary gear spare parts in warehouse number two, with the corresponding dedicated puller tool stored in the tool library's third cabinet. The work order management module merges this information with the fault diagnosis report and automatically generates a maintenance work order. The work order includes: work order number WO-2026-07-12-001, faulty equipment name 3.0 MW doubly-fed generator unit, fault root cause component planetary gear, fault propagation sequence table copied from the report, required spare part planetary gear (inventory code SP-203), and required tools gearbox-specific puller and torque wrench. The work order management module assigns the work order to the second shift team responsible for mechanical maintenance that day according to the duty roster. The work order details are displayed on the team's terminal, and a text message reminder is sent to the team leader.

[0051] Furthermore, after comprehensively evaluating each of the preliminary warning signals to determine that the wind turbine currently exhibits signs of a fault, and before constructing a warning query request from the initial fault type and the relevant operating parameters in the effective operating condition segment, the process further includes: Based on the timestamp of the effective operating condition segment, retrieve the historical operating parameter sequence within a preset duration range before and after the corresponding time interval of the effective operating condition segment from the historical cache of the data access gateway; The historical operating parameter sequence is compared with the effective operating condition segment. The trend change slope of each monitoring parameter in the historical operating parameter sequence is calculated. Based on the trend change slope, it is determined whether the fault symptom corresponding to the effective operating condition segment is a sudden fault symptom or a gradual fault symptom. When a sudden fault symptom is identified, a sudden fault marker is added to the early warning query request, and the relevant operating parameters in the effective operating condition segment are arranged in chronological order as a mutation parameter sequence. The mutation parameter sequence is then constructed as part of the relevant operating parameters into the early warning query request. When a gradual fault symptom is identified, a gradual fault marker is added to the early warning query request, and the relevant operating parameters in the effective operating condition segment are merged into a continuously changing parameter sequence according to the time order retrieved from the historical cache and the current time order. The continuously changing parameter sequence is then constructed into the early warning query request as part of the relevant operating parameters.

[0052] In this embodiment, after comprehensively evaluating and determining the existence of fault symptoms but before constructing an early warning query request, the operation involves judging whether the fault symptoms are sudden or gradual and adjusting the content of the query request accordingly. After determining the existence of fault symptoms and obtaining the initial fault type, the system does not immediately construct an early warning query request. Instead, based on the timestamp of the valid operating condition segment, it first sends a retrieval instruction to the historical cache of the data access gateway, requesting the acquisition of the historical operating parameter sequence for the 10 minutes before the time interval 14:23:15 to 14:28:45 (i.e., 14:13:15 to 14:23:15) and the following 2 minutes (14:28:45 to 14:30:45) corresponding to the valid operating condition segment.

[0053] The retrieved historical parameters are combined with the corresponding parameters in the effective operating condition segment to form a continuous time series. For vibration parameters, the average slope of the parameter in the 10 minutes prior to the start time of the effective operating condition segment (14:23:15) is calculated, which is the change in the root mean square value of vibration per second. For example, in the historical data, the vibration value slowly rises from 0.9 to 1.0, with a positive but small slope; while in the effective operating condition segment, the vibration value jumps sharply from 1.0 to 1.8 within 330 seconds, with a significantly larger slope. The slope in the effective operating condition segment is compared with the historical trend slope. If the slope in the effective segment exceeds five times the historical slope, it is judged as a sign of sudden failure; if the slopes are similar and continue to rise, it is judged as a sign of gradual failure; in this example, because the vibration value jumped sharply in a short period of time, it is judged as a sign of sudden failure.

[0054] If a fault is determined to be sudden, a sudden fault flag field is added to the early warning query request. Vibration parameters in the valid operating condition segment are arranged chronologically, one point per millisecond, to form a sudden change parameter sequence. This entire sequence is incorporated into the early warning query request as part of the relevant operating parameters, allowing the large language model to focus on the transient characteristics at the moment of the sudden change. If the fault is determined to be a gradual fault symptom, such as a temperature parameter slowly rising over the past hour and continuing to rise steadily within the valid operating condition segment, a gradual fault flag is added to the early warning query request. All time series retrieved from the historical buffer from 14:13:15 to 14:30:45 are merged into a continuously changing parameter sequence, also incorporated into the relevant operating parameters, enabling the large language model to analyze long-term trends. Through these differentiated processing steps, subsequent cross-validation analysis by the large language model can more effectively focus on the nature and characteristics of the fault, thereby improving the accuracy of the cross-validation results.

[0055] According to a second embodiment of the present invention, the present invention claims protection for a fault diagnosis and tracing system for integrated large-scale models and multi-source heterogeneous wind turbines, comprising: One or more processors; A memory storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the aforementioned method for tracing and diagnosing faults in a large-scale integrated model and multi-source heterogeneous wind turbines.

[0056] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A method for fault diagnosis and tracing of wind turbine units that integrates large-scale models and multi-source heterogeneous wind turbines, characterized in that, include: S1. Obtain the real-time operating parameters of each monitoring subsystem of the wind turbine through the data access gateway, and extract effective operating condition segments from the real-time operating parameters according to the preset operating condition filtering conditions. S2, input the effective operating condition segment into multiple heterogeneous fault detection models respectively, obtain the preliminary warning signals output by each model, comprehensively evaluate each preliminary warning signal, and determine whether the wind turbine currently has fault symptoms and the initial fault type corresponding to the fault symptoms; S3, when the fault symptoms are confirmed, the initial fault type and the relevant operating parameters in the effective operating condition segment are constructed into an early warning query request and sent to the large language model service interface. The large language model performs cross-system comparison analysis on the early warning query request based on its internal knowledge and generates a cross-validation result containing cross-system correlation information. S4, match the cross-validation results with the initial fault type. When the match is successful, use the cross-system association information in the cross-validation results as an index to retrieve the root cause tracing path corresponding to the fault symptoms from the fault knowledge graph, and output the fault root cause component and fault propagation chain indicated by the root cause tracing path. S5. Based on the fault root cause component and fault propagation chain, generate a fault diagnosis report for the wind turbine and push the fault diagnosis report to the work order management module of the enterprise resource planning system, whereby the work order management module creates a maintenance work order.

2. The method according to claim 1, characterized in that, S1 includes: The data access gateway establishes communication links with the vibration monitoring subsystem, temperature monitoring subsystem, electrical parameter monitoring subsystem, and oil monitoring subsystem of the wind turbine, respectively. Through each of the communication links, it synchronously acquires the real-time operating parameters output by each monitoring subsystem at different data acquisition frequencies. The real-time operating parameters include the timestamp of each monitoring subsystem. For each monitoring subsystem, the real-time operating parameters are time-aligned according to their timestamps to form a multi-dimensional operating parameter sequence under a unified time reference. A preset set of operating condition filtering conditions is read. The set of operating condition filtering conditions contains multiple discrimination conditions for determining whether the wind turbine is currently in a steady-state operating condition. The parameter value corresponding to each time point in the multi-dimensional operating parameter sequence is compared with the discrimination conditions one by one. A continuous time interval that simultaneously satisfies all discrimination conditions is selected. The multi-dimensional operating parameter sequence within the continuous time interval is extracted as an effective operating condition segment.

3. The method according to claim 1, characterized in that, The S2 includes: The effective operating condition segment is simultaneously distributed to the detection model based on signal waveform analysis, the detection model based on statistical threshold discrimination, and the detection model based on physical mechanism. Each detection model processes the effective operating condition segment, and each detection model independently outputs a preliminary warning signal. The preliminary warning signal includes the fault existence marker determined by the model and the fault type marker determined by the model. Collect all preliminary warning signals output by the detection models, count the number of detection models marked as having a fault, and determine that the wind turbine currently has fault symptoms when the number exceeds a preset majority consensus threshold. When a fault symptom is detected, the fault type labels output by each detection model are voted on and statistically analyzed. The fault type label with the most votes is determined as the initial fault type corresponding to the fault symptom. At the same time, the characteristic fragments of operating condition parameters on which each detection model outputs the initial fault type are recorded as supporting evidence for the initial fault type.

4. The method according to claim 1, characterized in that, The S3 includes: After confirming the existence of the fault symptoms, an operating parameter sequence related to the initial fault type is extracted from the effective operating condition segment. The operating parameter sequence includes the operating parameters of the monitoring subsystem involved in the initial fault type and the operating parameters of other monitoring subsystems that have energy transfer association with the monitoring subsystem. The name of the initial fault type and the operating parameter sequence are combined into a structured query text. The structured query text includes a fault type description field, a parameter value field, and a time evolution description field. The structured query text is sent as an early warning query request to the deployed large language model service interface, instructing the large language model to perform cross-system causal correlation analysis on the early warning query request. In response to the instruction, the large language model extracts historical fault cases that match the fault type description field and the parameter value field from its internal parameterized knowledge, and parses the parameter linkage change relationship between multiple subsystems before and after the fault from the historical fault cases. The parameter linkage change relationship is organized into a cross-system causal link description, wherein each cross-system causal link description includes upstream parameter change nodes and downstream parameter change nodes and the causal transmission direction between them. The large language model also retrieves the handling measures and verification conclusions for the fault type corresponding to the early warning query request recorded in the historical fault cases based on the cross-system causal link description, merges the handling measures and verification conclusions with the cross-system causal link description, and generates cross-validation results expressed in natural language. The cross-validation results include at least the cross-system causal link description, the associated historical fault case number, and a summary of the historical handling conclusions.

5. The method according to claim 4, characterized in that, The S4 includes: Extract the set of subsystem names and the set of parameter change directions involved in the cross-system causal link description from the cross-validation results. Match the set of subsystem names and the set of parameter change directions with the initial fault type. When all subsystems included in the set of subsystem names are associated with the initial fault type in the preset fault propagation topology table, and the parameter change direction in the set of parameter change directions is consistent with the change direction recorded in the preset fault propagation topology table, it is confirmed that the cross-validation results match the initial fault type successfully. After a successful match, the cross-system causal link description in the cross-validation results is used as the retrieval index to traverse the nodes and edges in the fault knowledge graph. Each node in the fault knowledge graph corresponds to a wind turbine component or a monitoring parameter type. Each edge corresponds to the fault propagation relationship between two nodes. The direction of the edge indicates the direction of fault propagation. Each edge also includes the operating condition label required for the fault propagation relationship to be established. The starting node is found in the fault knowledge graph to match the upstream parameter change node in the retrieval index. Starting from the starting node, the intermediate nodes that match the downstream parameter change node in the retrieval index are traversed sequentially along the edge direction until the terminating node that matches the fault root cause type indicated by the historical handling conclusion summary in the cross-validation results is reached. All nodes and edges traversed during the traversal are arranged in order to form the root cause tracing path, wherein the starting node and the intermediate nodes correspond to the transmission components in the fault propagation chain, and the terminating node corresponds to the fault root cause component.

6. The method according to claim 1, characterized in that, In step S4, after matching the cross-validation result with the initial fault type, if the match fails, the following further step is taken: The difference information between the cross-validation results and the initial fault type is extracted into a difference feature vector, which includes fault type name inconsistency items and cross-system causal link direction inconsistency items. The difference feature vector is fed back to the large language model service interface, instructing the large language model to readjust its internal knowledge retrieval scope based on the difference feature vector, and generate updated cross-validation results based on the adjusted retrieval scope; The updated cross-validation result is matched again with the initial fault type. If the match is successful, the process is transferred to follow the procedure for successful matching. If the match is unsuccessful, a manual intervention identifier is sent to the operation and maintenance terminal. The initial fault type, the cross-validation result, and the updated cross-validation result are packaged into a data packet to be reviewed and sent to the operation and maintenance terminal.

7. The method according to claim 1, characterized in that, The methods for constructing the fault knowledge graph include: Multiple fault cases are extracted in advance from the design documents, historical fault records and maintenance manuals of the wind turbine. Each fault case includes a description of the fault phenomenon, the identification of the root cause component, the sequence of intermediate components through which the fault propagates, and the direction of parameter transmission between each intermediate component. For each fault case, the monitoring parameter type in the fault phenomenon description is taken as the first node, the fault root cause component identifier is taken as the second node, and each component in the intermediate component sequence through which the fault propagation path passes is taken as an intermediate node. According to the parameter transmission direction, a first directed edge between the first node and the first intermediate node, a second directed edge between adjacent intermediate nodes, and a third directed edge between the last intermediate node and the second node are established in sequence. Each directed edge takes the parameter transmission direction as the direction of the edge, and the operating condition mark required for the establishment of the fault propagation relationship is added to each directed edge. All nodes and directed edges established from all fault cases are merged into the same graph structure to form the fault knowledge graph. The normal value range and alarm threshold range of the monitoring parameters corresponding to each node are recorded in the fault knowledge graph, and the response delay time range between the parameter change direction of the upstream node and the parameter change direction of the downstream node is recorded on each directed edge.

8. The method according to claim 1, characterized in that, The S5 includes: The component number, component name, and physical location of the fault-causing component, as well as the component number and component name of each component in the fault propagation chain, are arranged into a fault propagation sequence table according to the time sequence of fault propagation. The fault propagation sequence table is associated with the operating parameter change curves within the corresponding time interval of the effective operating condition segment to generate a fault diagnosis report containing parameter change curve annotations. The fault diagnosis report, in accordance with the order of the fault propagation sequence table, sequentially annotates the parameter change trend description and parameter over-limit time of each component during the fault propagation process. The fault diagnosis report is pushed to the work order management module of the enterprise resource planning system via the application programming interface. After receiving the fault diagnosis report, the work order management module parses the fault root cause component number in the fault diagnosis report, queries the corresponding spare parts inventory information and maintenance tool information from the spare parts inventory database of the enterprise resource planning system according to the fault root cause component number, and merges the spare parts inventory information and maintenance tool information with the fault diagnosis report. The work order management module generates a maintenance work order that includes at least the fault root cause component information, the fault propagation sequence table, the spare parts list, and the maintenance tool list, and assigns the maintenance work order to the corresponding maintenance team terminal.

9. The method according to claim 1, characterized in that, After comprehensively evaluating each of the preliminary warning signals to determine that the wind turbine currently exhibits signs of a fault, and before constructing the warning query request from the initial fault type and the relevant operating parameters in the effective operating condition segment, the process further includes: Based on the timestamp of the effective operating condition segment, retrieve the historical operating parameter sequence within a preset duration range before and after the corresponding time interval of the effective operating condition segment from the historical cache of the data access gateway; The historical operating parameter sequence is compared with the effective operating condition segment. The trend change slope of each monitoring parameter in the historical operating parameter sequence is calculated. Based on the trend change slope, it is determined whether the fault symptom corresponding to the effective operating condition segment is a sudden fault symptom or a gradual fault symptom. When a sudden fault symptom is identified, a sudden fault marker is added to the early warning query request, and the relevant operating parameters in the effective operating condition segment are arranged in chronological order as a mutation parameter sequence. The mutation parameter sequence is then constructed as part of the relevant operating parameters into the early warning query request. When a gradual fault symptom is identified, a gradual fault marker is added to the early warning query request, and the relevant operating parameters in the effective operating condition segment are merged into a continuously changing parameter sequence according to the time order retrieved from the historical cache and the current time order. The continuously changing parameter sequence is then constructed into the early warning query request as part of the relevant operating parameters.

10. A fault diagnosis and tracing system for wind turbine generators integrating large-scale models and multi-source heterogeneous wind turbines, characterized in that, include: One or more processors; A memory having stored one or more programs, which, when executed by one or more processors, cause the one or more processors to implement a fault diagnosis and tracing method for a fusion large model and multi-source heterogeneous wind turbine as described in any one of claims 1 to 9.

Images