Data-driven multi-dimensional intelligent diagnosis and fault warning system for wind farm

Abstract

The present application relates to the field of wind power generation technology, and discloses a wind farm multi-dimensional intelligent diagnosis and fault early warning system based on data driving.The present application comprises a state perception quantization module, a state assumption generation module, a verification generation implementation module, a confidence degree synthesis calculation module and a diagnosis report generation module; the system generates state signals by identifying abnormalities and using dynamic threshold values to monitor wind turbine operation data in real time; generates fault root assumption based on diagnosis knowledge graph and historical cases; plans and executes multi-dimensional verification chain for each fault root assumption, obtains quantitative evidence from power curve, operation data, log records and other sources; fuses the evidence and calculates the reliability of each fault root assumption through a confidence degree synthesis algorithm, and outputs a high confidence degree diagnosis conclusion; finally, a structured report containing abnormal information, diagnosis results, visualized evidence and operation and maintenance suggestions is automatically generated; the present application improves the operation and maintenance efficiency and reliability of the wind farm.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and in particular to a data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms. Background Technology

[0002] With the transformation of the global energy structure, wind power, as a clean and renewable energy source, has experienced rapid development and widespread application. Wind farms typically consist of a large number of wind turbine units, and the stability and reliability of their operation directly affect power generation efficiency and economic benefits. However, during long-term operation, wind turbine units inevitably experience various faults due to mechanical wear, electrical failures, environmental factors, and other reasons. Traditional wind farm fault diagnosis and early warning methods mainly rely on manual inspections, periodic maintenance, and simple threshold-based alarm systems. These methods have the following significant technical problems:

[0003] When a system issues an anomaly alarm, it can usually only indicate that a certain parameter is abnormal, without being able to further infer the specific fault type and potential cause. This requires maintenance personnel to invest a lot of time and energy in troubleshooting, resulting in low diagnostic efficiency and susceptibility to expert experience, making standardization and automation difficult. Traditional fault diagnosis lacks a mechanism for the fusion and verification of multi-dimensional evidence. A single data source or a single analysis method often cannot provide sufficiently reliable diagnostic evidence, easily leading to misjudgments. In complex fault scenarios, it is necessary to comprehensively consider data from multiple aspects such as mechanical, electrical, and performance aspects, and to perform cross-validation to arrive at an accurate diagnostic conclusion. Existing systems are insufficient in this regard, resulting in low confidence in diagnostic conclusions and affecting the accuracy of decision-making. There is an urgent need for a system that can achieve multi-dimensional data-driven, intelligent diagnosis and fault early warning to improve the operational reliability and maintenance efficiency of wind farms. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing wind farm fault diagnosis and early warning systems in terms of real-time performance, accuracy, automation, and in-depth analysis capabilities. It proposes a data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms to achieve early detection, accurate diagnosis, and efficient early warning of wind turbine faults.

[0005] To achieve the above objectives, this application adopts the following technical solution: a data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms, comprising: a state perception and quantification module, used to monitor in real time the operating data of each wind turbine in the wind farm in terms of power generation performance, mechanical state, and electrical state, identify and quantify anomalies in key performance indicators, and generate structured state signals, wherein the state signals at least include anomaly type, equipment identifier, occurrence time, and quantification deviation value; a state hypothesis generation module, used to receive the state signals and, based on a pre-set diagnostic knowledge graph and historical case library, generate multiple potential fault root cause hypotheses associated with the current abnormal state, constituting an initial diagnostic hypothesis set; and a verification and implementation module, used to dynamically plan and generate a verification path for each fault root cause hypothesis in the initial diagnostic hypothesis set, including at least one verification path. The system comprises a multi-dimensional verification chain for verification tasks, which acquire evidence from power curves, operational data, log records, neighbor comparisons, and historical trends. The verification generation and implementation module further executes tasks in the verification chain in parallel or sequentially, outputting quantitative evidence generated by each task. A confidence synthesis and calculation module receives and integrates multi-dimensional quantitative evidence from different verification tasks, calculates a confidence score for each fault root cause hypothesis using a preset confidence synthesis algorithm, sorts and filters all fault root cause hypotheses based on the confidence scores, and outputs at least one high-confidence diagnostic conclusion. A diagnostic report generation module automatically generates a structured diagnostic report based on the high-confidence diagnostic conclusion and its corresponding multi-dimensional evidence. The report includes at least the core anomaly, diagnostic conclusion, visual charts supporting the evidence, and operational recommendations.

[0006] Preferably, the state-aware quantization module is specifically used to: set dynamic thresholds associated with historical operating conditions for multiple key performance indicators such as power generation, availability, power curve K-value, vibration amplitude, and bearing temperature; continuously collect real-time operating data of the wind turbine through a data interface; compare the real-time operating data with the dynamic thresholds of the corresponding indicators; when any indicator data continuously exceeds the threshold range for a preset time, an anomaly is determined and a state signal is generated; wherein, the quantization deviation value is the percentage deviation between the abnormal indicator data and the median of the corresponding dynamic threshold.

[0007] Preferably, the state hypothesis generation module is specifically used for: maintaining and updating a diagnostic knowledge graph, where nodes represent specific abnormal states or root causes of faults, edges represent causal or statistical correlations between the two, and each edge is assigned a weight value to characterize the correlation strength; when a state signal is received, the module uses the abnormal type in the signal as the query starting point and traverses the graph to find all directly connected root cause nodes of faults; the module sorts the found root causes of faults in descending order according to the weight values ​​and outputs the sorted list as the initial set of diagnostic hypotheses; wherein, the correlation strength weight value is calculated based on the frequency of co-occurrence of the abnormal state and the root cause of faults in the historical case library and the confidence of the causal relationship.

[0008] Preferably, the verification generation implementation module has a built-in structured verification task library. Each record in the library defines a verification task, including the task name, target verification dimension, required data source or analysis model interface, and estimated execution time. For each root cause hypothesis in the initial diagnostic hypothesis set, a set of tasks that can provide relevant evidence is selected from the verification task library based on the physical mechanism and common characteristics of the root cause hypothesis. Based on the logical dependencies between tasks, execution costs, and the urgency of verifying the root cause hypothesis, the execution sequence of these tasks is dynamically planned to form a dedicated verification chain for the root cause hypothesis. By calling the corresponding data analysis service in the system, the tasks in the verification chain are executed sequentially, and the raw results output by the tasks are quantified into standardized evidence strength values.

[0009] Preferably, when planning the verification chain, the verification generation implementation module is also used to: calculate a dynamic priority score for each candidate verification task before execution begins. The calculation of the dynamic priority score takes into account the following factors: the average discrimination of the task against all current root cause hypotheses of failure, the uncertainty of the task's historical execution results, the execution frequency of the task in recent diagnoses, and the information complementarity between the evidence expected to be provided by the task and the evidence of the planned tasks; and prioritize the task with the highest dynamic priority score to be added to the verification chain.

[0010] Preferably, the confidence synthesis calculation module is specifically used for: receiving multi-dimensional quantitative evidence for the same root cause hypothesis of failure; assigning a basic confidence coefficient to each piece of evidence, which is positively correlated with the historical accuracy of the verification task that generated the evidence and the quality rating of the data source; fusing the multi-dimensional evidence assigned with confidence coefficients, and finally outputting a comprehensive confidence score between 0 and 1 to represent the probability that the root cause hypothesis of failure is true.

[0011] Preferably, when fusing multi-dimensional evidence, the specific method for handling evidence conflicts is as follows: when a high degree of conflict is detected in the support of different dimensions of evidence for the same root cause hypothesis, the algorithm will automatically reduce the weight of conflicting evidence in the synthesis and introduce an additional global uncertainty measure. The value of the global uncertainty measure will be reflected in the final comprehensive confidence score range to reflect the reliability level of the diagnostic conclusion.

[0012] Preferably, before outputting the initial diagnostic hypothesis set, the state hypothesis generation module is further configured to: invoke a density statistical algorithm to preliminarily score each fault root cause hypothesis in the initial diagnostic hypothesis set. The algorithm uses the static correlation strength between the state signal and the fault root cause hypothesis, the duration of the signal from its occurrence to the current moment, and the severity of the anomalies exhibited by the signal in multiple technical dimensions as core input parameters to calculate a density score; and reorder the initial diagnostic hypothesis set based on this score to optimize the allocation order of subsequent verification resources.

[0013] The preferred expression for the density statistics algorithm is: ;in Indicates the root cause hypothesis of the failure The tightness score; This represents the set of currently observed state signals; Indicates status signal Root cause hypothesis The static correlation strength is derived from historical statistics and ranges from 0 to 1; Indicates status signal The duration, in hours; Indicates status signal The cross-dimensional weights are calculated by comprehensively considering the degree of anomaly in power generation performance, mechanical condition, and electrical condition, with a value range of 0 to 2. This indicates the validation of root cause hypotheses in historical diagnostics. The average time required, in minutes.

[0014] The preferred expression for calculating the dynamic priority score is: ;in Indicates verification task Dynamic priority scoring; This represents the set of highly dense root cause hypotheses of failure that are currently to be verified; Indicates verification task For the hypothesis The distinguishability, i.e. the success rate of the task in history in effectively identifying or eliminating the root cause hypothesis of the failure, has a value range of 0 to 1. Indicates verification task Hypothesis of the root cause of the failure The uncertainty of the outcome, i.e. the standard deviation or volatility of historical results, has a value range of 0 to 1; Indicates verification task Number of executions in the last 24 hours; Indicates verification task The degree of information overlap between the data and the set of planned or executed tasks, i.e., the historical data. The result is the average similarity to the results of other tasks, with a value range of 0 to 1.

[0015] The technical effects and advantages of this invention are as follows: This invention utilizes a state-aware quantification module for real-time monitoring and dynamic threshold comparison, enabling timely detection of early anomalies and avoiding false alarms and missed alarms caused by traditional fixed thresholds. The state hypothesis generation module, combined with a diagnostic knowledge graph and historical case library, can directly infer multiple potential root cause hypotheses from abnormal states, achieving in-depth diagnosis from "anomaly" to "root cause," reducing reliance on manual investigation. The verification generation and implementation module dynamically plans a multi-dimensional verification chain, acquiring evidence from different data sources and fusing the evidence through a confidence synthesis calculation module to calculate confidence scores for root cause hypotheses, significantly improving the reliability and credibility of diagnostic conclusions. When planning the verification chain, the verification generation and implementation module considers task differentiation, uncertainty, execution frequency, and information complementarity, dynamically calculating priorities to ensure effective utilization of verification resources, avoid unnecessary verification tasks, and improve diagnostic efficiency. Finally, a structured report is automatically generated, containing core anomalies, diagnostic conclusions, visualized evidence, and operational suggestions, providing clear and intuitive decision support for operations personnel and improving operational efficiency. Attached Figure Description

[0016] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0017] Figure 1 This is a schematic diagram of the module topology of the present invention. Detailed Implementation

[0018] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0019] Example 1: Refer to Figure 1 As shown, the present invention provides a technical solution: a data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms, which mainly includes the following modules: a state perception quantification module, a state hypothesis generation module, a verification generation and implementation module, a confidence level synthesis calculation module, and a diagnosis report generation module.

[0020] State perception quantification module: responsible for real-time collection and preliminary processing of wind turbine operation data; continuously monitoring the operation data of each wind turbine in the wind farm in multiple dimensions such as power generation performance (e.g., power generation, power curve K value), mechanical status (e.g., vibration amplitude, bearing temperature) and electrical status (e.g., voltage, current, frequency).

[0021] Specifically, this module sets dynamic thresholds for several key performance indicators, such as power generation, availability, power curve K-value, vibration amplitude, and bearing temperature, each associated with its historical operating conditions. These dynamic thresholds are adaptively adjusted based on historical data statistical analysis, machine learning models, or expert experience to accommodate the normal fluctuation range of wind turbines under different environments and loads. Real-time operating data of the wind turbines is continuously collected via a data interface, which may originate from SCADA systems, sensor networks, PLCs, etc. The real-time operating data is compared with the dynamic thresholds of the corresponding indicators. When any indicator data continuously exceeds the threshold range for a preset duration, for example, exceeding the threshold for 10 consecutive minutes, an anomaly is determined. Once an anomaly is determined, the module generates a structured status signal. The status signal includes at least: the anomaly type, such as "abnormal power curve K-value" or "overheating bearing temperature"; equipment identification, such as "#12 wind turbine"; and the occurrence time and quantification deviation value.

[0022] Among them, the quantitative deviation value is the percentage deviation between the abnormal indicator data and the corresponding dynamic threshold median. For example, if the normal temperature range of the bearing is 60±5℃ and the actual temperature reaches 70℃, then the quantitative deviation value is (70-60) / 60=16.7%. This quantitative deviation value can intuitively reflect the severity of the abnormality.

[0023] State hypothesis generation module: It is responsible for making preliminary inferences about possible root causes of failure based on the perceived abnormal signals; it receives state signals from the state perception quantification module and generates multiple potential root cause hypotheses of failure associated with the current abnormal state based on the pre-set diagnostic knowledge graph and historical case library, forming an initial diagnostic hypothesis set.

[0024] Specifically, this module maintains and updates a diagnostic knowledge graph. Nodes in this graph represent specific abnormal states, such as "abnormal power curve K-value" or root causes of faults, such as "blade icing," "pitch system failure," or "poor gearbox lubrication." Edges represent causal or statistical relationships between the two, and each edge is assigned a weight value to characterize the strength of the relationship. For example, there might be an edge from "abnormal power curve K-value" to "blade icing" with a weight value of 0.8. When a state signal is received, the module uses the abnormal type in the signal as the starting point for the query and traverses the graph to find all directly connected root cause nodes. The found root causes are sorted in descending order according to their weight values, and the sorted list is output as the initial set of diagnostic hypotheses. The weight value of the relationship strength is calculated based on the frequency of co-occurrence of the abnormal state and the root cause of the fault in the historical case library, as well as the confidence level of the causal relationship. For example, if "abnormal power curve K-value" and "blade icing" co-occur 100 times in historical cases, and 80 of them are ultimately diagnosed as blade icing, then the weight value of the relationship strength will be relatively high.

[0025] Before outputting the initial set of diagnostic hypotheses, this module also invokes a density statistical algorithm to preliminarily score each root cause hypothesis in the initial set. This algorithm uses the static correlation strength between the state signal and the root cause hypothesis, the duration of the signal from its occurrence to the current moment, and the severity of the anomalies exhibited by the signal across multiple technical dimensions as core input parameters to calculate a density score. Based on this score, the initial set of diagnostic hypotheses is reordered to optimize the allocation order of subsequent validation resources.

[0026] The expression for the density statistics algorithm is: ;in Indicates the root cause hypothesis of the failure The tightness score; This represents the set of currently observed state signals; Indicates status signal Root cause hypothesis The static correlation strength is derived from historical statistics and ranges from 0 to 1; Indicates status signal The duration, in hours; Indicates status signal The cross-dimensional weight is calculated by comprehensively considering the degree of anomaly in multiple dimensions such as power generation performance, mechanical condition, and electrical condition, with a value range of 0 to 2. This indicates the validation of root cause hypotheses in historical diagnostics. The average time required is in minutes. The algorithm comprehensively considers the correlation between the signal and the root cause hypothesis, the duration of the anomaly, the severity of the anomaly, and the verification cost, so that the root cause hypothesis with a high tightness score can be verified first.

[0027] The verification generation and implementation module is responsible for designing and executing verification tasks for each root cause hypothesis to obtain supporting or refuting evidence. For each root cause hypothesis in the initial diagnostic hypothesis set, it dynamically plans and generates a multi-dimensional verification chain containing at least one verification task. The verification tasks are used to obtain evidence from different dimensions such as power curves, operating data, log records, comparison with neighboring machines, and historical trends.

[0028] This module has a built-in structured verification task library. Each record in the library defines a verification task, including: task name, such as "power curve analysis", "vibration spectrum analysis", "SCADA log retrieval"; target verification dimension, such as "power generation performance" and "mechanical condition"; data source or analysis model interface to be called, such as "historical power data interface", "vibration sensor data interface", "log database interface", "machine learning prediction model"; and estimated execution time.

[0029] For each fault root cause hypothesis in the initial diagnostic hypothesis set, a set of tasks that can provide relevant evidence are selected from the verification task library based on the physical mechanism and common characteristics of the fault root cause hypothesis; for example, for the hypothesis of "blade icing", tasks such as "power curve analysis" and "blade temperature sensor data analysis" may be selected.

[0030] Based on the logical dependencies between tasks, such as the requirement that some analyses must be performed after data cleaning; the execution costs, such as the need for large amounts of computing resources for some tasks; and the urgency of verifying the root cause hypothesis, the execution sequence of these tasks is dynamically planned to form a dedicated verification chain for the root cause hypothesis. For example, tasks with low cost, short execution time, and the ability to quickly eliminate most root cause hypotheses can be prioritized.

[0031] Before execution begins, a dynamic priority score is calculated for each candidate validation task. The calculation of the dynamic priority score takes into account the following factors: the average discrimination of the task against all current root cause hypotheses, i.e., the task's ability to effectively distinguish between different root cause hypotheses; the uncertainty of the task's historical execution results, i.e., the stability of the task's results; the frequency of the task's execution in recent diagnoses; and the complementarity of the evidence expected to be provided by the task with the evidence from the planned tasks. The task with the highest dynamic priority score is selected to join the validation chain.

[0032] The expression for calculating the dynamic priority score is: ;in Indicates verification task Dynamic priority scoring; This represents the set of highly dense root cause hypotheses of failure that are currently to be verified; Indicates verification task Hypothesis of the root cause of the failure The distinguishability, i.e. the success rate of the task in history in effectively identifying or eliminating the root cause hypothesis of the failure, has a value range of 0 to 1. Indicates verification task Hypothesis of the root cause of the failure The uncertainty of the outcome, i.e. the standard deviation or volatility of historical results, has a value range of 0 to 1; Indicates verification task Number of executions in the last 24 hours; Indicates verification task The degree of information overlap between the data and the set of planned or executed tasks, i.e., the historical data. The results are the average similarity with the results of other tasks, with a value range of 0 to 1. By calling the corresponding data analysis service in the system, the tasks in the verification chain are executed in parallel or sequentially, and the original results output by the tasks, such as the deviation of the power curve and the characteristic value of the vibration spectrum, are quantified into standardized evidence strength values, for example, between -1 and 1, where -1 indicates strong rebuttal, 1 indicates strong support, and 0 indicates neutrality.

[0033] The confidence synthesis calculation module is responsible for synthesizing all evidence and making a final decision on the root cause hypothesis of the failure; it receives and integrates multi-dimensional quantitative evidence from different verification tasks.

[0034] Specifically, the module receives multi-dimensional quantitative evidence for the same root cause hypothesis of the failure; it assigns a basic credibility coefficient to each piece of evidence, which is positively correlated with the historical accuracy of the verification task that generated the evidence and the quality rating of the data source; for example, the credibility coefficient of evidence from high-precision sensors is higher than that from manually entered logs.

[0035] The multi-dimensional evidence with assigned confidence coefficients is fused together, and a confidence score is calculated for each fault root cause hypothesis using a pre-defined Dempster-Shafer evidence theory algorithm. Finally, a comprehensive confidence score between 0 and 1 is output to indicate the probability that the fault root cause hypothesis is true.

[0036] When fusing multi-dimensional evidence, the specific handling of evidence conflicts is as follows: when a high degree of conflict is detected in the support of different dimensions of evidence for the same root cause hypothesis, for example, one piece of evidence strongly supports hypothesis A, while another piece of evidence strongly refutes hypothesis A, the algorithm will automatically reduce the weight of conflicting evidence in the synthesis and introduce an additional global uncertainty measure. The value of the global uncertainty measure will be reflected in the final comprehensive confidence score range. For example, instead of outputting a single score, a confidence range [0.6, 0.8] will be output to reflect the reliability level of the diagnostic conclusion. This helps to avoid misjudgments caused by errors in single evidence or data noise.

[0037] Diagnostic report generation module: responsible for presenting diagnostic results to users in a clear and easy-to-understand way; automatically generating structured diagnostic reports based on high-confidence diagnostic conclusions and their corresponding multi-dimensional evidence.

[0038] The report includes at least the core anomaly, such as "the K value of the power curve of #12 wind turbine is consistently low"; diagnostic conclusions, such as "high confidence diagnosis of blade icing"; visual charts supporting the evidence, such as power curve comparison charts, blade temperature trend charts, vibration spectrum charts, etc., to intuitively present the evidence; and operation and maintenance recommendations, such as "it is recommended to immediately perform blade de-icing operations" and "check the pitch system sensors". The operation and maintenance recommendations can be automatically generated based on the diagnostic conclusions and historical maintenance experience, and can be optimized based on user feedback.

[0039] Example 2: Based on Example 1, the state-aware quantization module further optimizes its dynamic threshold setting and anomaly detection logic.

[0040] Dynamic thresholds are based on machine learning anomaly detection models, such as isolated forests, local anomaly factors, or deep learning autoencoders; they learn the complex patterns of wind turbines under normal operating conditions and identify anomalous data points that deviate from these patterns, thereby providing more accurate dynamic thresholds and anomaly determination.

[0041] In addition, the "preset duration" in the anomaly detection is also adaptive; for some key indicators, such as the main bearing temperature, even if it exceeds the threshold for a short time, it may indicate a serious problem, so the duration can be set to short; while for some highly volatile indicators, such as wind speed, a longer duration may be required to determine an anomaly; this adaptive duration can be configured according to the historical characteristics and failure modes of the indicator.

[0042] In addition to percentage deviation, quantification deviation values ​​can also be standardized Z-scores or Mahalanobis distances to better reflect the degree to which data points deviate from the normal distribution, providing more refined input for subsequent generation of root cause hypotheses.

[0043] Example 3: Building upon Example 1, the diagnostic knowledge graph of the state hypothesis generation module adopts a more complex ontological structure, containing richer entity relationships, such as "component-fault," "fault-symptom," "fault-cause," and "cause-solution." The construction and updating of the graph combine natural language processing technology to automatically extract knowledge from historical maintenance records and expert reports.

[0044] The parameters in the density statistical algorithm are dynamically adjusted and learned; for example, static association strength. Bayesian updates can be used to continuously optimize based on new diagnostic cases; cross-dimensional weights. Principal component analysis or factor analysis can be used to automatically identify the combined effects of anomalies in different dimensions; average time consumption It can then be updated in real time based on the execution logs of the actual verification task.

[0045] By reordering the initial set of diagnostic hypotheses through density scores, it is possible to ensure that validation resources are prioritized for the most likely and urgent root cause hypotheses, thereby improving diagnostic efficiency. For example, if a root cause hypothesis has a high density score, it may be validated first even if its association weight in the knowledge graph is not the highest, because it may be an anomaly with a long duration and high severity.

[0046] Example 4: Based on Example 1, the verification generation implementation module introduces a reinforcement learning algorithm when planning the verification chain; the system can regard each diagnosis process as a sequential decision problem, and learn the optimal verification chain planning strategy by continuously trying different verification chain planning strategies and obtaining rewards based on the accuracy and efficiency of the final diagnosis.

[0047] The factors in the formula for calculating dynamic priority scores, such as discrimination... Uncertainty Number of executions Information overlap All of these can be statistically analyzed and updated in real time using the execution results of historical verification tasks; for example, if a task has successfully distinguished two similar root cause hypotheses of failure multiple times in history, its discriminative power can be increased. The uncertainty will increase accordingly; if the results of a task fluctuate frequently, its uncertainty will increase. It will improve; information complementarity The calculation can employ concepts such as information entropy or mutual information; if a task can provide information that is completely different from the planned task, its information complementarity is high, and its priority will be increased accordingly; this helps to avoid duplicate verification and ensure the efficiency of the verification chain.

[0048] The execution of the verification chain can employ parallel computing frameworks, such as Apache Spark, to accelerate data analysis and model inference processes and shorten the diagnostic cycle.

[0049] Example 5: Based on Example 1, the confidence synthesis calculation module adopts a more complex evidence fusion model; for example, the Dempster-Shafer evidence theory can better handle uncertainty and unknown information, use a basic probability assignment function to represent the degree of support of evidence for the root cause hypothesis of the failure, and use combination rules to fuse evidence from different sources.

[0050] In addition to reducing the weight of conflicting evidence and introducing uncertainty metrics, an expert feedback mechanism can also be introduced to handle conflicting evidence. When the system detects highly conflicting evidence, it can send an alert to the operation and maintenance experts to request human intervention for judgment and feed back the experts' judgment results to the system to optimize subsequent evidence fusion algorithms and credibility coefficient allocation.

[0051] The allocation of credibility coefficients can also be dynamic; for example, the credibility of evidence from power curve analysis may be lower at low wind speeds because power fluctuations are greater at low wind speeds; while it is higher when wind speeds are stable; this dynamic credibility helps to improve the accuracy of evidence fusion.

[0052] Example 6: Based on Example 1, the diagnostic report generation module can be further enhanced in terms of intelligence and interactivity.

[0053] The system automatically generates a priority list of recommendations based on factors such as the severity of the diagnostic findings, the operating status of the wind turbine, spare parts inventory, and the maintenance personnel's schedule. For example, in the case of an emergency fault, the system can automatically generate a shutdown order and an emergency repair work order.

[0054] The report supports multilingual output and provides customizable templates to suit the needs of different users.

[0055] The report includes embedded interactive visualizations, allowing users to click on data points within the charts to view the raw data or more detailed analysis results.

[0056] In addition, the diagnostic report generation module can be integrated with the wind farm's asset management system or enterprise resource planning system to automatically create maintenance work orders, spare parts applications, etc., realizing management from diagnosis to execution.

[0057] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A data-driven, multi-dimensional intelligent diagnostic and fault early warning system for wind farms, characterized in that: include: The state awareness quantification module is used to monitor the real-time operating data of each wind turbine in the wind farm in terms of power generation performance, mechanical status, and electrical status, identify and quantify anomalies in key performance indicators, and generate structured state signals. The state signals at least include anomaly type, equipment identifier, occurrence time, and quantification deviation value. The state hypothesis generation module is used to receive the state signals and, based on a pre-set diagnostic knowledge graph and historical case library, generate multiple potential fault root cause hypotheses associated with the current abnormal state, forming an initial diagnostic hypothesis set. The verification generation implementation module is used to dynamically plan and generate a multi-dimensional verification chain containing at least one verification task for each fault root cause hypothesis in the initial diagnostic hypothesis set. The verification task is used to obtain evidence from power curves, operating data, log records, neighbor comparisons, and historical trends. The verification generation implementation module is further used to execute the tasks in the verification chain in parallel or sequentially and output the quantified evidence generated by each task. The confidence synthesis calculation module receives and integrates multi-dimensional quantitative evidence from different verification tasks, calculates a confidence score for each root cause hypothesis using a preset confidence synthesis algorithm, and sorts and filters all root cause hypotheses based on the confidence scores, outputting at least one high-confidence diagnostic conclusion. The diagnostic report generation module automatically generates a structured diagnostic report based on the high-confidence diagnostic conclusion and its corresponding multi-dimensional evidence. The report includes at least the core anomaly, diagnostic conclusion, visualization charts supporting the evidence, and operational recommendations. Before outputting the initial diagnostic hypothesis set, the state hypothesis generation module further uses a density statistical algorithm to preliminarily score each root cause hypothesis in the initial diagnostic hypothesis set. This algorithm uses the static correlation strength between the state signal and the root cause hypothesis, the duration of the signal from its occurrence to the current moment, and the severity of the anomalies exhibited by the signal in multiple technical dimensions as core input parameters to calculate a density score. Based on this score, the initial diagnostic hypothesis set is reordered to optimize the allocation order of subsequent verification resources. The expression for the density statistical algorithm is: ;in Indicates the root cause hypothesis of the failure The tightness score; This represents the set of currently observed state signals; Indicates status signal Root cause hypothesis The static correlation strength is derived from historical statistics and ranges from 0 to 1; Indicates status signal The duration, in hours; Indicates status signal The cross-dimensional weights are calculated by comprehensively considering the degree of anomaly in power generation performance, mechanical condition, and electrical condition, with a value range of 0 to 2. This indicates the validation of root cause hypotheses in historical diagnostics. The average time required, in minutes.

2. The data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms according to claim 1, characterized in that, The state-aware quantization module is specifically used to: set dynamic thresholds associated with historical operating conditions for multiple key performance indicators such as power generation, availability, power curve K-value, vibration amplitude, and bearing temperature; continuously collect real-time operating data of the wind turbine through a data interface; compare the real-time operating data with the dynamic thresholds of the corresponding indicators; when any indicator data continuously exceeds the threshold range for a preset time, an anomaly is determined, and the state signal is generated; wherein, the quantization deviation value is the percentage deviation between the abnormal indicator data and the median of the corresponding dynamic threshold.

3. The data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms according to claim 2, characterized in that, The state hypothesis generation module is specifically used for: maintaining and updating a diagnostic knowledge graph, where nodes represent specific abnormal states or root causes of faults, edges represent causal or statistical correlations between them, and each edge is assigned a weight value to characterize the correlation strength; when a state signal is received, the module uses the abnormal type in the signal as the query starting point and traverses the graph to find all directly connected root cause nodes of faults; the module sorts the found root causes of faults in descending order according to the weight values ​​and outputs the sorted list as the initial diagnostic hypothesis set; wherein, the weight value of the correlation strength is calculated based on the frequency of co-occurrence of the abnormal state and the root cause of faults in the historical case library and the confidence of the causal relationship.

4. The data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms according to claim 3, characterized in that, The verification generation and implementation module has a built-in structured verification task library. Each record in the library defines a verification task, including the task name, target verification dimension, required data source or analysis model interface, and estimated execution time. For each root cause hypothesis in the initial diagnostic hypothesis set, a set of tasks that can provide relevant evidence is selected from the verification task library based on the physical mechanism and common characteristics of the root cause hypothesis. Based on the logical dependencies between tasks, execution costs, and the urgency of verifying the root cause hypothesis, the execution sequence of these tasks is dynamically planned to form a dedicated verification chain for the root cause hypothesis. By calling the corresponding data analysis service within the system, the tasks in the verification chain are executed sequentially, and the raw results output by the tasks are quantified into standardized evidence strength values.

5. The data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms according to claim 4, characterized in that, When planning the verification chain, the verification generation implementation module is also used to: calculate a dynamic priority score for each candidate verification task before execution begins. The calculation of the dynamic priority score takes into account the following factors: the average discrimination of the task against all current root cause hypotheses, the uncertainty of the task's historical execution results, the execution frequency of the task in recent diagnoses, and the information complementarity between the evidence expected to be provided by the task and the evidence of the planned tasks; and prioritize the task with the highest dynamic priority score to be added to the verification chain.

6. The data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms according to claim 5, characterized in that, The confidence synthesis calculation module is specifically used for: receiving multi-dimensional quantitative evidence for the same root cause hypothesis; assigning a basic confidence coefficient to each piece of evidence, which is positively correlated with the historical accuracy of the verification task that generated the evidence and the quality rating of the data source; fusing the multi-dimensional evidence with assigned confidence coefficients, and finally outputting a comprehensive confidence score between 0 and 1 to represent the probability that the root cause hypothesis is true.

7. The data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms according to claim 6, characterized in that, When the multi-dimensional evidence is fused, the specific method for handling evidence conflicts is as follows: when a high degree of conflict is detected in the support of different dimensions of evidence for the same root cause hypothesis, the algorithm will automatically reduce the weight of conflicting evidence in the synthesis and introduce an additional global uncertainty measure. The value of the global uncertainty measure will be reflected in the final comprehensive confidence score range to reflect the reliability level of the diagnostic conclusion.

8. The data-driven multi-dimensional intelligent diagnosis and fault early warning system for wind farms according to claim 7, characterized in that, The expression for calculating the dynamic priority score is: ;in Indicates verification task Dynamic priority scoring; This represents the set of highly dense root cause hypotheses of failure that are currently to be verified; Indicates verification task Hypothesis of the root cause of the failure The distinguishability, i.e. the success rate of the task in history in effectively identifying or eliminating the root cause hypothesis of the failure, has a value range of 0 to 1. Indicates verification task Hypothesis of the root cause of the failure The uncertainty of the outcome, i.e. the standard deviation or volatility of historical results, has a value range of 0 to 1; Indicates verification task Number of executions in the last 24 hours; Indicates verification task The degree of information overlap between the data and the set of planned or executed tasks, i.e., the historical data. The result is the average similarity to the results of other tasks, with a value range of 0 to 1.

Images