State evaluation and fault early warning method and system for wind generating set

Abstract

The invention discloses a wind generating set state evaluation and fault early warning method and system, and relates to the technical field of wind generating set state monitoring, and the method comprises the steps: obtaining the operation monitoring data and control feedback data of a wind generating set; constructing an equivalent load characterization quantity based on the wind regime parameters and the unit operation state parameters; determining a response offset characteristic of the unit component relative to a load response reference based on the equivalent load characterization quantity, and obtaining a component state evaluation result; constructing a variable pitch control residual error and a yaw control residual error based on the control feedback data to obtain a control chain abnormity evaluation result; constructing a multi-source evidence set and carrying out Bayesian network probabilistic reasoning to obtain an abnormal risk confidence coefficient and an abnormal source judgment result; and matching the early warning level and outputting a result. According to the method, distinguishing of wind condition disturbance, control chain abnormity and part degradation abnormity is achieved, the accuracy, reliability and engineering practicability of state evaluation and fault early warning are improved, and the false alarm rate and the missing report rate are reduced.

Description

Technical Field

[0001] This invention relates to the field of wind turbine generator condition monitoring technology, specifically to a method and system for wind turbine generator condition assessment and fault early warning. Background Technology

[0002] With the continuous increase in single-unit capacity of wind turbine generators, the increasing diameter of rotors, and the increasingly complex operating environment of wind farms, wind turbine generator condition monitoring and fault early warning technology has gradually evolved from single-parameter monitoring to comprehensive assessment technology based on multi-source data fusion. Existing related technologies typically rely on the interfaces of generator monitoring systems, condition monitoring systems, and control systems to collect operational data such as wind speed, power, rotational speed, vibration, temperature, current, voltage, and pitch and yaw parameters. These data are then combined with statistical analysis, fault diagnosis models, health index calculations, or predictive maintenance strategies to identify and provide early warnings about the operating status of key generator components. Simultaneously, with the deepening application of data-driven methods, mechanistic modeling methods, and probabilistic reasoning methods in the field of industrial equipment health management, wind turbine generator condition assessment is developing towards load perception, component-level identification, anomaly attribution, and tiered early warning output.

[0003] However, existing wind turbine condition assessment and fault early warning technologies still have significant limitations. First, many existing methods rely directly on absolute values ​​or fixed thresholds of monitored quantities such as vibration, temperature, and power for anomaly judgment, failing to adequately characterize the intermediate process of wind disturbances transmitting to the turbine load. This leads to the coupling of influences between changes in the external wind environment, changes in control actions, and component degradation, making it difficult to accurately identify whether component responses truly deviate from the healthy baseline under the same or similar load conditions. Second, existing technologies do not pay sufficient attention to the execution status of the pitch and yaw control links, often conflating anomalies such as control system execution deviations, response hysteresis, and action mismatches with mechanical component degradation anomalies. This makes it difficult to effectively distinguish between control chain anomalies and component anomalies, thus reducing the accuracy of anomaly source identification. Third, existing technologies in multi-source information fusion mostly remain at the level of simple weighting, threshold superposition, or single-model classification, lacking a unified probabilistic reasoning mechanism that can simultaneously integrate load-side evidence, component-side evidence, and control-side evidence. Consequently, it is difficult to output early warning results with reliable quantification and anomaly source identification, and it is also difficult to achieve hierarchical early warning output adapted to actual operation and maintenance needs. Therefore, existing technologies are still insufficient to simultaneously achieve wind disturbance impact isolation, independent identification of control chain anomalies, accurate assessment of component status, and reliable attribution of anomaly sources. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing wind turbine generator condition assessment and fault early warning methods fail to effectively characterize the process of wind disturbances being transmitted to the unit load, have difficulty distinguishing between control chain anomalies and component anomalies, lack multi-source information fusion capabilities and a reliable attribution mechanism for anomaly sources, and have problems with how to achieve accurate assessment of wind turbine generator condition, anomaly source determination and graded fault early warning output.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for wind turbine generator condition assessment and fault early warning, comprising the following steps: S1: Acquire the operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The control feedback data includes pitch control data, yaw control data and corresponding execution feedback data. S2: Based on the wind condition parameters and the unit operating status parameters, construct an equivalent load characterization quantity that represents the relationship between wind disturbance and unit load transmission; S3: Based on the equivalent load characterization quantity and the unit operating status parameters, determine the response offset characteristics of the unit components relative to the corresponding load response benchmark, and obtain the component status assessment results; S4: Based on the control feedback data, construct the pitch control residual and yaw control residual to obtain the control chain anomaly evaluation results; S5: Construct a multi-source evidence set based on the component status assessment results, the control chain anomaly assessment results, and the equivalent load characterization quantity, and perform probabilistic reasoning on the multi-source evidence set based on a Bayesian network to obtain the anomaly risk confidence level and anomaly source determination results; S6: Based on the abnormal risk confidence level and the abnormal source determination result, match the corresponding early warning level and output the wind turbine generator status assessment result and fault early warning result.

[0007] As a preferred embodiment of the wind turbine generator condition assessment and fault early warning method of the present invention, step S1 specifically includes: The system acquires operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The wind condition parameters include real-time wind speed, wind direction deviation, and turbulence intensity. The generator set operation status parameters include generator speed, output power, and actual values ​​of blade pitch angle. The control feedback data includes pitch control data, yaw control data, and corresponding execution feedback data. The pitch control data includes a target pitch angle command sequence output by the pitch controller. The yaw control data includes a target yaw angle command sequence output by the yaw controller. The execution feedback data includes an actual pitch angle feedback sequence corresponding to the target pitch angle command sequence and an actual nacelle azimuth angle feedback sequence corresponding to the target yaw angle command sequence. The operation monitoring data and control feedback data are processed by time alignment, missing value filling, outlier removal and data format unification, and then collected according to the preset analysis window to form a dataset to be analyzed.

[0008] As a preferred embodiment of the wind turbine generator condition assessment and fault early warning method of the present invention, step S2 specifically includes: Feature extraction is performed on the wind condition parameters to obtain the wind disturbance description results; Based on the wind disturbance description results and the unit operating status parameters, a load mapping model is established to characterize the correspondence between wind parameters, unit operating status parameters and unit load levels. Input the wind condition parameters and unit operating status parameters in the current analysis window into the load mapping model to obtain the load characterization results corresponding to the current analysis window.

[0009] As a preferred embodiment of the wind turbine generator set condition assessment and fault early warning method of the present invention, in step S2, the load characterization results include thrust direction load characterization results and torque direction load characterization results. The thrust direction load characterization results are used to characterize the load level transmitted along the main shaft direction after the wind turbine receives wind, and the torque direction load characterization results are used to characterize the load level transmitted to the generator input side by the wind turbine rotation driving force through the main shaft and transmission chain. The thrust direction load characterization results and the torque direction load characterization results are unified and organized to form the equivalent load characterization quantity.

[0010] As a preferred embodiment of the wind turbine generator condition assessment and fault early warning method of the present invention, in step S3, the load response benchmark is established in the following manner: Using the equivalent load characterization quantity as input and the component response parameters corresponding to the component to be evaluated as output, a regression model describing the mapping relationship between load conditions and component response is established. Input the equivalent load characterization quantity corresponding to the current analysis window into the regression model to obtain the baseline response result of the component to be evaluated under the current load conditions.

[0011] As a preferred embodiment of the wind turbine generator set condition assessment and fault early warning method of the present invention, wherein: in step S3, the response offset feature is a standardized offset; The standardized offset is obtained by comparing the difference between the actual response value of the component to be evaluated in the current analysis window and the baseline response result with the response fluctuation scale under the same load conditions in historical healthy samples. Based on whether the standardized offset exceeds the allowable fluctuation range of the baseline response, the degree of excess, and the duration of occurrence, the current state of the component to be evaluated is classified to form a component state evaluation result.

[0012] As a preferred embodiment of the wind turbine generator condition assessment and fault early warning method of the present invention, step S4 specifically includes: Extract the target pitch angle command sequence of the pitch controller and the actual pitch angle feedback sequence of the pitch drive from the control feedback data, as well as the target yaw angle command sequence of the yaw controller and the actual cabin azimuth angle feedback sequence from the yaw encoder. Based on the target pitch angle command sequence and the actual pitch angle feedback sequence, calculate the pitch instantaneous residual sequence, pitch response delay, and pitch cumulative deviation. Based on the target yaw angle command sequence and the actual cabin azimuth feedback sequence, calculate the yaw instantaneous residual sequence and the yaw start delay; The control chain anomaly assessment results are obtained by comparing the instantaneous residual sequence of pitch, the pitch response delay, the cumulative pitch deviation, the instantaneous residual sequence of yaw, and the yaw start delay with the control residual reference benchmark corresponding to the historical healthy operation phase.

[0013] As a preferred embodiment of the wind turbine generator condition assessment and fault early warning method of the present invention, step S5 specifically includes: Component-side evidence is constructed based on the component status assessment results, control-side evidence is constructed based on the control chain anomaly assessment results, and load-side evidence is constructed based on the equivalent load characterization quantity. A multi-source evidence set is formed by the component-side evidence, the control-side evidence, and the load-side evidence. The multi-source evidence set is input into a pre-established Bayesian network, which includes anomaly source nodes and evidence nodes corresponding to the multi-source evidence set. The anomaly source nodes include wind condition excitation anomalies, control chain execution anomalies, and component body degradation anomalies. Based on the evidence status of each type of evidence in the multi-source evidence set within the current analysis window, calculate the posterior probability of each anomaly source category under the current evidence conditions; The anomaly source category with the highest posterior probability is taken as the anomaly source determination result, and the highest posterior probability value is output as the anomaly risk confidence level.

[0014] As a preferred embodiment of the wind turbine generator condition assessment and fault early warning method of the present invention, step S6 specifically includes: Based on the abnormal risk confidence level obtained in step S5, it is matched with the preset abnormal risk confidence level range to determine the warning level corresponding to the current analysis window. The warning level includes normal state, attention state, warning state and alarm state. When matching the warning level, the abnormal source determination result obtained in step S5 is used for auxiliary adjustment. When the abnormal source determination result is wind excitation abnormality, the warning level is constrained. Based on the anomaly source determination result, the anomaly risk confidence level, and the warning level, the wind turbine generator status assessment result and fault warning result corresponding to the current analysis window are generated.

[0015] Secondly, embodiments of the present invention provide a wind turbine generator condition assessment and fault early warning system, comprising: Data acquisition module: acquires the operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The control feedback data includes pitch control data, yaw control data and corresponding execution feedback data. Load characterization module: Based on the wind condition parameters and the unit operating status parameters, construct an equivalent load characterization quantity that characterizes the relationship between wind disturbance and unit load; Condition assessment module: Based on the equivalent load characterization quantity and the unit operating status parameters, determine the response offset characteristics of the unit components relative to the corresponding load response benchmark, and obtain the component condition assessment results; Control evaluation module: Based on the control feedback data, construct pitch control residuals and yaw control residuals to obtain control chain anomaly evaluation results; Probabilistic reasoning module: Based on the component state assessment results, the control chain anomaly assessment results, and the equivalent load characterization quantity, a multi-source evidence set is constructed, and probabilistic reasoning is performed on the multi-source evidence set based on a Bayesian network to obtain the anomaly risk confidence level and anomaly source determination results; Early warning output module: Based on the confidence level of the abnormal risk and the determination result of the abnormal source, match the corresponding early warning level and output the status assessment result and fault early warning result of the wind turbine generator set.

[0016] The beneficial effects of this invention are as follows: By introducing an equivalent load characterization quantity, this invention makes the relationship between wind disturbance and unit load transmission explicit. On this basis, a component response benchmark is established and response offset features are extracted, thereby improving the pertinence and accuracy of component condition assessment. At the same time, by constructing pitch control residuals and yaw control residuals, the difference between control chain anomalies and component-specific anomalies is realized. Furthermore, by performing Bayesian network probabilistic inference on evidence from the load side, component side, and control side, the confidence level of anomaly risk and the determination result of anomaly source are obtained, thereby enabling the output of graded early warning results with credible quantification and anomaly source orientation, improving the reliability and engineering practicality of wind turbine condition assessment and fault early warning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is an overall flowchart of a wind turbine generator condition assessment and fault early warning method provided in the first embodiment of the present invention; Figure 2 The diagram shows the module connection of a wind turbine generator condition assessment and fault early warning system provided in the third embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0019] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for wind turbine generator condition assessment and fault early warning is provided.

[0020] S1: Acquire the operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The control feedback data includes pitch control data, yaw control data and corresponding execution feedback data.

[0021] In this embodiment, step S1 is to acquire the operation monitoring data and control feedback data of the wind turbine generator set to form the basic dataset required for subsequent status assessment and fault early warning.

[0022] The operational monitoring data includes wind condition parameters and unit operating status parameters. The wind condition parameters characterize the current external wind environment of the wind turbine generator and may include real-time wind speed, wind direction, wind direction deviation, turbulence intensity, and gust variation information. The real-time wind speed can be acquired by a nacelle anemometer or a wind field anemometer; the wind direction and wind direction deviation can be acquired by a wind direction measuring device or a unit orientation detection device; and the turbulence intensity and gust variation information can be calculated based on wind speed fluctuation data within a preset time window. By acquiring these wind condition parameters, environmental input data can be provided for constructing the equivalent load characterization in subsequent steps.

[0023] The unit operating status parameters are used to characterize the operating status of the wind turbine generator set under current operating conditions, and may include generator speed, output power, main shaft vibration information, nacelle vibration information, tower vibration information, gearbox temperature, main shaft bearing temperature, generator temperature, current, voltage, and actual blade pitch angle. It should be noted that the unit operating status parameters can be selected according to different unit configurations, monitoring conditions, and components to be evaluated, and are not limited to including all of the above parameters simultaneously. In some embodiments, the above operating status parameters can be continuously acquired through existing monitoring channels, vibration sensors, temperature sensors, electrical parameter acquisition modules, and the unit monitoring system.

[0024] Furthermore, the control feedback data includes pitch control data, yaw control data, and corresponding execution feedback data. The pitch control data characterizes the control state in the pitch control link and may include pitch action trigger records, pitch adjustment direction records, pitch adjustment amplitude records, pitch adjustment frequency records, and pitch control channel status records. In some embodiments, the pitch control data may specifically represent a target pitch angle command sequence output by the pitch controller. The yaw control data characterizes the control state in the yaw control link and may include yaw action trigger records, yaw adjustment direction records, yaw adjustment amplitude records, yaw adjustment frequency records, and yaw control channel status records. In some embodiments, the yaw control data may specifically represent a target yaw angle command sequence output by the yaw controller. The execution feedback data characterizes the actual response results of the above control actions on the actuator side and may include actual pitch angle feedback, actual nacelle azimuth angle feedback, execution in place status, execution delay status, and action completion status. In this way, the control feedback data can characterize both the control state of the control link and the actual response of the actuator, thus providing a data foundation for constructing pitch control residuals and yaw control residuals in subsequent steps.

[0025] In some implementations, the operational monitoring data and control feedback data can be acquired through existing monitoring and acquisition systems of the wind turbine generator set, such as the generator set monitoring system, condition monitoring system, or controller data interface. The acquired data can be continuously collected at a preset sampling frequency, which can be set according to the monitoring object and analysis requirements, for example, using second-level sampling or higher frequency sampling. For data from different sources, timestamp information can be uniformly added so that subsequent processing can be performed according to the same analysis cycle.

[0026] It should also be noted that after acquiring the above data, preprocessing can be performed. This preprocessing may include time alignment, missing value imputation, outlier removal, and data format standardization. Time alignment ensures that wind parameters, unit operating status parameters, and control feedback data correspond at the same analysis time, avoiding inaccuracies in subsequent load characterization and response analysis due to inconsistent sampling times. Missing value imputation supplements data with short-term interruptions or partial missing data. Outlier removal removes abnormal records that clearly exceed equipment operating boundaries or sensor ranges. Data format standardization organizes data from different sources into a unified field structure for subsequent retrieval.

[0027] Furthermore, the operational monitoring data and control feedback data can be collected according to a preset analysis window to form a dataset to be analyzed. The analysis window can be a fixed-duration window or a rolling update window corresponding to changes in the unit's operating status, thereby improving the adaptability of the dataset to continuous changes in operating conditions. Within each analysis window, wind parameters, unit operating status parameters, and control feedback data are organized accordingly, and time identifiers, parameter category identifiers, and data source identifiers are added to form structured data records. The dataset to be analyzed can be directly used for subsequent steps such as constructing equivalent load representations, evaluating component status, and assessing control chain anomalies.

[0028] In this embodiment, after acquiring and organizing the operation monitoring data and control feedback data in the above manner, the basic input data corresponding to the current analysis cycle can be obtained. This provides wind condition and operation status information for constructing the equivalent load characterization quantity in step S2, provides state observation information for determining the response offset characteristics in step S3, and provides control link related data support for constructing the pitch control residual and yaw control residual in step S4.

[0029] S2: Based on the wind condition parameters and the unit operating status parameters, construct an equivalent load characterization quantity that represents the relationship between wind disturbance and unit load.

[0030] In this embodiment, step S2 is used to construct an equivalent load characterization quantity based on the wind condition parameters and unit operating status parameters obtained in step S1, representing the relationship between wind disturbance and unit load transfer. It should be noted that the equivalent load characterization quantity is not a complete solution of the actual stress state of the unit, but rather a load characterization result constructed for subsequent component status assessment, used to reflect the comprehensive load level on the unit after the current wind disturbance is affected by the aerodynamic action of the wind turbine and the unit's operating conditions.

[0031] Specifically, wind condition parameters and generator operating status parameters within the current analysis window can be extracted from the dataset to be analyzed formed in step S1. The wind condition parameters may include real-time wind speed, wind direction, wind direction deviation, turbulence intensity, and gust variation information; the generator operating status parameters may include generator speed, output power, actual blade pitch angle, and nacelle azimuth angle. Further, feature extraction processing can be performed on the wind condition parameters. This feature extraction processing may include: determining the representative wind speed level of the current window based on the wind speed data within the analysis window; determining the degree of wind speed fluctuation based on continuous wind speed sampling results; determining the degree of yaw deviation based on wind direction data and nacelle azimuth angle; determining the turbulence intensity level based on the dispersion of wind speed changes within the analysis window; and determining the degree of gust disturbance based on short-term wind speed spikes or drops. Through the above processing, a description of the wind disturbance corresponding to the current analysis window can be obtained.

[0032] Furthermore, after obtaining the wind disturbance description results, an aerodynamic load mapping relationship can be constructed by combining the unit's operating state parameters. It should be noted that in this embodiment, the aerodynamic load mapping relationship is preferably implemented using a pre-established load mapping model. This load mapping model is used to characterize the correspondence between wind parameters, unit operating state parameters, and unit load levels. The inputs to the load mapping model may include at least wind speed, wind direction deviation, turbulence intensity, generator speed, actual blade pitch angle, and nacelle azimuth angle. The output of the load mapping model is the load characterization result corresponding to the current analysis window. It should also be noted that the load mapping model can be established based on aerodynamic simulation data generated during the design phase, calibration data generated during the commissioning phase, or historical sample data accumulated during the stable operation phase of the target unit. Specifically, it can use a lookup table mapping method, an empirical calibration mapping method, or a proxy mapping method formed through offline training. This embodiment does not impose a unique limitation on this; as long as an effective mapping relationship between wind disturbance and unit load can be formed, it can be used in this invention.

[0033] In some implementations, the load mapping model can be implemented using a lookup table mapping method. Specifically, a multi-dimensional correspondence table of wind speed, pitch angle, engine speed, and wind direction deviation can be pre-established for the target unit, and the load characterization results under different combinations of operating conditions can be stored in the correspondence table. During actual operation, the wind speed, pitch angle, engine speed, and wind direction deviation in the current analysis window are input into the correspondence table, and the load characterization results under the current operating condition are obtained through matching or interpolation. When using this method, the required input parameters are clear, the mapping process is straightforward, and it is easy to call up in the field system.

[0034] In other implementations, the load mapping model can be implemented using an empirical calibration mapping method. Specifically, multiple sets of sample data from the target unit during its historical stable operation phase can be selected first. Operating records under different wind speeds, pitch angles, speeds, and yaw deviations are grouped and organized. Then, based on the correspondence between power changes and operating condition changes on the unit's load performance in each set of samples, load level mapping rules corresponding to different operating condition intervals are formed. These load level mapping rules can be determined using piecewise linear mapping, interval assignment mapping, or mapping based on the distribution center of historical samples. During actual operation, the corresponding mapping rule is invoked according to the operating condition interval to which the current analysis window belongs, obtaining the load characterization result for the current analysis window. Using this method, there is no need to introduce a complex real-time solution process, making it suitable for engineering deployment.

[0035] Furthermore, in this embodiment, the equivalent load characterization quantity is preferably composed of thrust direction load characterization results and torque direction load characterization results. The thrust direction load characterization result characterizes the load level transmitted along the main shaft direction to the main shaft, nacelle, and tower structure after the wind turbine receives wind; the torque direction load characterization result characterizes the load level transmitted from the wind turbine's rotational driving force to the generator input side via the main shaft and transmission chain. Specifically, the thrust direction load characterization result can be determined based on wind speed, wind direction deviation, actual pitch angle, and nacelle azimuth angle; the torque direction load characterization result can be determined based on wind speed, generator speed, output power, and actual pitch angle. It should also be noted that the thrust direction load characterization result and torque direction load characterization result can be output separately from the load mapping model, or the load level results in the two directions can be obtained separately first, and then a unified equivalent load characterization quantity can be formed according to a preset integration rule.

[0036] In this embodiment, it is preferable to unify and organize the thrust direction load characterization results and torque direction load characterization results as the equivalent load characterization quantity output. Specifically, when the subsequent step S3 establishes load response benchmarks for different components, the direction load characterization results more relevant to the target component can be directly called; when the subsequent step S5 constructs a multi-source evidence set, the thrust direction load characterization results and torque direction load characterization results can be used together as load-side evidence input.

[0037] Furthermore, after obtaining the equivalent load characterization, it can be standardized in scale. This standardization may include: mapping the load characterization results obtained in the current analysis window to the rated operating condition scale range corresponding to the target unit; performing benchmark alignment on the current results based on the load characterization results under the same operating condition label in the historical stable operation samples of the target unit; and calibrating and correcting the load characterization results based on the unit model, rated capacity, and turbine parameters. Through this standardization, the equivalent load characterizations obtained under different analysis windows can have a consistent basis for comparison, so that subsequent step S3 can call the load response benchmark under the corresponding operating condition for comparison.

[0038] It should also be noted that, in specific implementation, an equivalent load characterization quantity can be independently constructed for each analysis window, and then stored after associating the equivalent load characterization quantity with a time identifier, unit identifier, and operating condition label. When using a rolling update window, the wind condition parameters and unit operating status parameters within the current window can be re-extracted after each window update, and the load mapping model can be re-called to calculate the new equivalent load characterization quantity, so that subsequent component status assessments are always based on the load characterization results corresponding to the latest operating condition.

[0039] Furthermore, it should be noted that in constructing the equivalent load characterization quantity in this step, parameters belonging to the component response side, such as nacelle vibration, tower vibration, and main shaft vibration, are not introduced as core construction inputs. This ensures that the equivalent load characterization quantity retains its load-side characterization attributes, avoiding the circular dependency problem in subsequent step S3 where response results are used to participate in load construction and then load results are used to reverse-interpret response anomalies. By using operating condition parameters such as wind conditions, generator speed, output power, actual blade pitch angle, and nacelle azimuth angle as the main inputs, the equivalent load characterization quantity can be placed before the component response in terms of causality, providing an independent load basis for determining subsequent response offset characteristics.

[0040] In this embodiment, the equivalent load characterization quantity constructed in the above manner can reflect the result of wind disturbances being transmitted to the unit load within the current analysis window, and serves as the input basis for determining the response offset characteristics of unit components relative to the corresponding load response benchmark in step S3. Simultaneously, the equivalent load characterization quantity can also participate in anomaly risk probability inference as part of the multi-source evidence set in step S5, thus playing an intermediate load characterization role in the entire technical solution, connecting external wind disturbances, unit operating conditions, and subsequent anomaly attribution analysis.

[0041] S3: Based on the equivalent load characterization quantity and the unit operating status parameters, determine the response offset characteristics of the unit components relative to the corresponding load response benchmark, and obtain the component status assessment results.

[0042] In this embodiment, step S3 is used to determine the response offset characteristics of unit components relative to the corresponding load response benchmark based on the equivalent load characterization quantity constructed in step S2 and the unit operating status parameters obtained in step S1, and thereby obtain the component status assessment result. It should be noted that this step does not directly judge whether the unit components are abnormal based on the absolute magnitude of vibration values, temperature values, or electrical parameters, but rather compares the actual response of the unit components within the current analysis window with the benchmark response under the same or similar load conditions to identify whether the unit components have abnormal offsets that exceed the normal load response pattern.

[0043] Specifically, this step first establishes a load response benchmark. This benchmark characterizes the response level that the target unit should exhibit under healthy conditions and under different load conditions. Further, time periods confirmed to be in a healthy state from the target unit's historical operating data can be selected as historical health samples. These health states can be filtered based on criteria such as no fault records, no abnormal alarms, no abnormal maintenance records, and continuous and stable operating data. Then, the equivalent load characterization quantity constructed in step S2 and the unit operating status parameters obtained in step S1 are extracted from the historical health samples. The samples are then organized according to load conditions to establish a correspondence between "load conditions and component responses."

[0044] In one implementation, the load response benchmark can be established using a regression model. Specifically, the equivalent load characterization quantity can be used as input, and the response parameters corresponding to the component to be evaluated can be used as output to establish a mapping relationship between load and response. The response parameters can be selected according to the type of component to be evaluated. For example, for a gearbox, bearing temperature, lubricating oil temperature, or vibration amplitude can be selected as response parameters; for a spindle, spindle vibration amplitude or bearing temperature can be selected as response parameters; for a generator, generator temperature, current fluctuation parameters, or power fluctuation parameters can be selected as response parameters; for tower and nacelle structures, acceleration parameters or vibration displacement parameters can be selected as response parameters. The regression model can be a linear regression model, a polynomial regression model, or other regression models that can describe the mapping relationship between load and response. Through the regression model, given the current equivalent load characterization quantity, the benchmark predicted value of the corresponding component response under healthy conditions and its allowable fluctuation range can be obtained.

[0045] In another implementation, the load response benchmark can be established using a segmented statistical model. Specifically, the equivalent load characteristics in historical healthy samples can be divided into several load intervals based on their numerical values, and representative values ​​and fluctuation ranges of the corresponding component response parameters can be statistically analyzed within each load interval. The representative values ​​can be determined using the mean, median, or stable interval representative values, and the fluctuation range can be determined based on the distribution of historical healthy samples within that load interval. In practical applications, the component response benchmark under the corresponding interval is called for comparison based on the load interval to which the equivalent load characteristics in the current analysis window belong. This method has a clear implementation path and is suitable for scenarios with high real-time requirements.

[0046] Furthermore, after establishing the load response benchmark, actual response parameters corresponding to the component to be evaluated can be extracted for the current analysis window. Specifically, vibration parameters, temperature parameters, current fluctuation parameters, power fluctuation parameters, or other state parameters related to the component to be evaluated can be extracted from the unit operating status parameters, and these parameters can be used as the actual response results of the component within the current analysis window. Subsequently, the equivalent load characterization quantity corresponding to the current analysis window is input into the aforementioned load response benchmark to obtain the benchmark response results and their allowable fluctuation range under the corresponding load conditions.

[0047] After obtaining the current actual response result and the corresponding baseline response, the response offset feature can be calculated. The response offset feature characterizes the degree of deviation of the current actual response from the healthy baseline response. In this embodiment, the response offset feature can be represented in one or more of the following forms: First, it can be represented by an absolute offset, i.e., the difference between the current actual response value and the baseline response result; second, it can be represented by a relative offset, i.e., the ratio of the difference to the baseline response result; third, it can be represented by a standardized offset, i.e., comparing the difference with the response fluctuation scale under the same load conditions in historical healthy samples to form an offset result with a unified comparison basis. It should be noted that using a standardized offset is beneficial for unified judgment under different components, different parameter dimensions, and different operating conditions.

[0048] It should also be noted that different offset extraction methods can be used for different types of response parameters. For temperature-related response parameters, since their changes are relatively slow, the average offset or temperature rise offset within the analysis window can be used as the response offset feature. For vibration-related response parameters, since their transient fluctuations are more obvious, the peak value offset, RMS value offset, or fluctuation range offset within the analysis window can be used as the response offset feature. For electrical response parameters, the current fluctuation offset, power offset, or change in electrical imbalance can be used as the response offset feature. By selecting the appropriate offset extraction method according to the type of response parameter, the obtained response offset features can better reflect the actual operating characteristics of the component.

[0049] Furthermore, after obtaining the response offset characteristics, a component status assessment result can be generated according to preset judgment rules. Specifically, the current status of the component to be assessed can be classified according to whether the response offset characteristics exceed the allowable fluctuation range, the degree of exceedance, and the duration of occurrence. The component status assessment result can be represented as normal state, state of concern, abnormal state, and severely abnormal state, or as healthy, slightly abnormal, moderately abnormal, and severely abnormal. This embodiment does not impose a unique limitation on this, as long as it can reflect the operating status of the unit components under the current load conditions.

[0050] In some implementations, the above processing can be performed separately for multiple unit components to form a set of component condition assessment results. Specifically, corresponding response parameters can be extracted for the main shaft, gearbox, generator, bearings, nacelle structure, and tower structure, respectively, and corresponding load response benchmarks can be called to calculate the response offset characteristics and generate corresponding component condition assessment results. Furthermore, for different monitoring locations within the same component, corresponding response benchmarks can be established separately. For example, temperature response benchmarks can be established for different bearing locations in the gearbox, or temperature or current response benchmarks can be established for different locations in the generator. This approach can improve the refinement of component condition assessment.

[0051] It should also be noted that this step maintains the causal relationship between load and response during execution. The equivalent load representation constructed in step S2 is used to characterize the load input under the current operating condition, and the unit operating status parameters obtained in step S1 are used to characterize the actual response results of the components. The load response benchmark is established based on historical healthy samples, and the response results within the current analysis window are not included in the benchmark construction. In this way, the analysis logic of load as an antecedent and response as a consequence can be guaranteed, thereby avoiding causal confusion in subsequent anomaly judgment.

[0052] Furthermore, in specific implementation, the response offset features and component status assessment results obtained within each analysis window can be associated and stored with the corresponding time identifier, unit identifier, component identifier, and operating condition label. When a rolling analysis window is used, the actual response parameters within the current window are re-extracted after each window update, the corresponding load response benchmark is re-called, and the response offset features and component status assessment results are recalculated, thereby forming a continuously updated component status assessment sequence. The component status assessment sequence can serve as an important component of the multi-source evidence set in subsequent step S5, used to participate in anomaly risk probability reasoning and anomaly source determination.

[0053] In this embodiment, step S3 determines the response offset characteristics of unit components relative to the corresponding load response benchmark based on the equivalent load characterization quantity and unit operating status parameters, thus forming a component status assessment result. This component status assessment result reflects whether the unit components exhibit abnormal behavior beyond normal response patterns under the current load conditions, and provides component-side evidence to support probabilistic reasoning about the source of the anomaly in subsequent steps.

[0054] S4: Based on the control feedback data, construct the pitch control residual and yaw control residual to obtain the control chain anomaly evaluation results.

[0055] In this embodiment, step S4 is used to construct pitch control residuals and yaw control residuals based on the control feedback data obtained in step S1, and to obtain control chain anomaly assessment results based on the control residuals. It should be noted that the focus of this step is not the response state of the unit's mechanical components, but rather the execution accuracy and response characteristics of the control chain itself. By constructing pitch control residuals and yaw control residuals, independent evidence related to the control execution state can be provided for subsequent anomaly source determination, thereby distinguishing between load fluctuations or response changes caused by control chain anomalies and abnormal performance caused by component degradation.

[0056] Specifically, the data sequence related to the pitch control system is first extracted from the control feedback data obtained in step S1. This data sequence includes at least: the target pitch angle command sequence issued by the pitch controller, and the actual pitch angle response sequence fed back by the pitch drive. Furthermore, auxiliary parameters reflecting the operating state of the pitch actuator can also be extracted, such as the pitch motor current sequence, hydraulic station pressure sequence, or pitch action status indicator. The above data sequences should be aligned with a unified time base to ensure that the target command and actual feedback at the same moment correspond accurately.

[0057] After obtaining the target pitch angle command sequence and the actual pitch angle response sequence, pitch control residuals can be constructed. These residuals characterize the command tracking deviation of the pitch control system during execution. In one embodiment, instantaneous residuals can be used; that is, for each sampling moment, the difference between the target pitch angle and the actual pitch angle at that moment is calculated to obtain an instantaneous pitch control residual sequence. These instantaneous residuals reflect the static tracking deviation at each moment. When the residuals are consistently non-zero or consistently exceed a preset range, it indicates that the pitch control system may have static deviation or zero-point drift problems.

[0058] Furthermore, to reflect the dynamic response characteristics of the pitch control system to changes in commands, a response delay feature can be constructed. Specifically, the time point at which the actual pitch angle response begins to change can be identified near the moment the target pitch angle command changes, and the time difference between the two can be calculated as the pitch response delay. For continuous changes, a sliding window approach can be used to statistically analyze the average or maximum delay between the command change and the actual response change within the window. When the response delay significantly exceeds the historical normal level, it indicates that the pitch actuator may have problems such as slow response, execution lag, incomplete execution, or local jamming.

[0059] It should also be noted that a cumulative deviation feature can be constructed for the pitch control system. This cumulative deviation feature can be obtained by summing or accumulating the instantaneous residuals within the analysis window, reflecting the cumulative tracking error of the pitch control system over a period of time. When the cumulative deviation continues to increase, it indicates that the pitch system may have problems such as trend deviation, inability to effectively track even after frequent corrections, or decreased execution stability.

[0060] Similarly, for the yaw control system, data sequences related to the yaw control system are extracted from the control feedback data. These data sequences include at least: the target yaw angle command sequence issued by the yaw controller, and the actual cabin azimuth angle feedback sequence from the yaw encoder. Furthermore, auxiliary information such as the yaw drive motor current sequence, yaw brake status indicators, and yaw action status records can also be extracted.

[0061] It should be noted that yaw control systems typically exhibit intermittent operation characteristics. This means the aircraft does not continuously fine-tune the yaw, but rather triggers a yaw action once the wind deviation reaches a preset condition, and maintains a relatively stable state after the action ends. Therefore, the construction of yaw control residuals should consider not only the static difference between the target yaw angle and the actual cabin azimuth angle, but also the initiation delay of the yaw action, the tracking situation during the action, and the stability maintenance after the action ends.

[0062] Based on the target yaw angle command sequence and the actual cabin azimuth feedback sequence, yaw control residuals can be constructed. In one implementation, instantaneous residuals are used, that is, the difference between the target yaw angle and the actual cabin azimuth angle at each moment is calculated to obtain the instantaneous yaw control residual sequence. The instantaneous yaw control residuals can reflect the degree of wind deviation of the yaw control system under static conditions. When the residuals are continuously large, it indicates that the yaw system may have problems such as inaccurate wind alignment, response mismatch, or inadequate execution.

[0063] Furthermore, considering that yaw control systems typically employ intermittent actions, yaw response characteristics can be constructed. Specifically, the start and end times of yaw actions can be identified, the time delay from the issuance of the yaw command to the actual start of the yaw action can be statistically analyzed, and the tracking of the cabin azimuth angle to the target angle during the yaw action can be recorded. When the yaw start delay is too long, the tracking deviation during yaw is too large, or the cabin azimuth angle repeatedly oscillates near the target after the action ends, it indicates that the yaw link may have problems such as execution lag, frequent corrections, local instability, or degraded drive performance.

[0064] In some implementations, a cumulative yaw deviation characteristic can also be constructed. This cumulative yaw deviation characteristic can be obtained by analyzing the correlation between wind direction deviation and yaw control residual within an analysis window. For example, when the wind direction continuously changes and the yaw action fails to keep up in time, both the yaw control residual and the wind direction deviation will increase simultaneously. In this case, it can be used as an assessment basis for evaluating sluggish response or insufficient coordination in the yaw control link.

[0065] After obtaining the relevant characteristics of the pitch control residuals and yaw control residuals, control chain anomaly assessment results can be further generated. The control chain anomaly assessment results include at least one or more of the following: statistical values ​​of pitch instantaneous residuals, statistical values ​​of pitch response delays, numerical values ​​of cumulative pitch deviations, statistical values ​​of yaw instantaneous residuals, statistical values ​​of yaw start delays, statistical values ​​of yaw tracking deviations, and a comprehensive assessment indicator reflecting the overall execution status of the control chain.

[0066] It should be noted that, to determine whether the control residuals are in an abnormal state, a control residual reference benchmark can be established based on data from historical healthy operation phases. Specifically, a time period in the target unit's historical operation that has been confirmed as having a normal control link with no related fault records is selected. The distribution characteristics of pitch control residuals and yaw control residuals during this time period are statistically analyzed, including the mean, standard deviation, quantiles, and normal fluctuation range. In practical applications, the control residual characteristics within the current analysis window are compared with the historical reference benchmark. When the residual characteristics exceed a preset threshold or significantly deviate from the historical distribution, an anomaly in the control link is determined.

[0067] Furthermore, the severity of control chain anomalies can be graded. For example, when the instantaneous residual error of pitch occasionally exceeds the limit but the duration is short and the degree of exceedance of the historical benchmark is minor, it can be assessed as a mild anomaly or a state of concern; when the pitch response delay is persistently large, the yaw tracking deviation continues to accumulate, or the control action is frequently corrected, it can be assessed as a moderate anomaly; when the pitch control system exhibits a large and persistent deviation, the yaw link fails to execute properly for a long period, or the control process shows obvious local instability, it can be assessed as a severe anomaly. The results of the control chain anomaly assessment can be expressed in the form of anomaly level, anomaly score, or anomaly label.

[0068] It should also be noted that the control chain anomaly assessment results constructed in this step will serve as an important component of the multi-source evidence set in step S5. In subsequent probabilistic reasoning, the control chain anomaly assessment results, along with the component state assessment results obtained in step S3 and the equivalent load characterization constructed in step S2, will be used as evidence inputs to calculate the probabilities of different anomaly source categories. By introducing the control chain anomaly assessment results, the easily confused control chain anomalies and component degradation anomalies can be effectively distinguished, thereby improving the accuracy of anomaly source determination.

[0069] In some implementations, a corresponding control chain anomaly assessment result can be independently constructed for each analysis window and stored in association with a time stamp, unit identifier, and operating condition label. When a rolling update window is used, the control feedback data within the current window is re-extracted after each window update, the relevant characteristics of the pitch control residual and yaw control residual are recalculated, and the control chain anomaly assessment result is updated, thereby forming a control chain state sequence that is continuously updated over time.

[0070] In this embodiment, step S4 constructs the pitch control residual and yaw control residual based on the control feedback data, and obtains the control chain anomaly assessment result accordingly. This assessment result reflects the execution accuracy and response characteristics of the pitch control system and yaw control system within the current analysis window, providing independent control-side evidence to distinguish between control chain anomalies and component degradation anomalies in subsequent steps.

[0071] S5: Construct a multi-source evidence set based on the component status assessment results, the control chain anomaly assessment results, and the equivalent load characterization quantity, and perform probabilistic reasoning on the multi-source evidence set based on a Bayesian network to obtain the anomaly risk confidence level and anomaly source determination results.

[0072] In this embodiment, step S5 is used to construct a multi-source evidence set based on the component state assessment results obtained in step S3, the control chain anomaly assessment results obtained in step S4, and the equivalent load characterization quantity constructed in step S2. Then, based on a Bayesian network, probabilistic reasoning is performed on the multi-source evidence set to obtain the anomaly risk confidence level and anomaly source determination results. It should be noted that this step is not a simple superposition of the output results of the preceding steps, but rather places multiple types of evidence representing the component response state, control chain state, and external load excitation state under a unified probabilistic reasoning framework. By analyzing the conditional correlation between various types of evidence and different anomaly sources, a comprehensive judgment is made on the credibility of the anomaly occurrence and the anomaly source within the current analysis window.

[0073] Specifically, a multi-source evidence set can be constructed based on the outputs of steps S3, S4, and S2. This multi-source evidence set can include component-side evidence, control-side evidence, and load-side evidence. Component-side evidence reflects the response deviation of unit components under current load conditions and may include the state assessment results, response deviation characteristic values, deviation significance indicators, and component state levels for each component to be evaluated. Control-side evidence reflects the execution consistency and response stability of the pitch and yaw links within the current analysis window and may include pitch control residual characteristics, yaw control residual characteristics, control chain anomaly assessment results, and control chain anomaly levels. Load-side evidence reflects the result of wind disturbances being transmitted to the unit load under the current operating condition and may include the equivalent load characterization quantity constructed in step S2, load interval identifiers, operating condition labels, and load characterization results in the thrust and torque directions. Through this method, analysis results from different sources and with different functions can be organized into structured evidence that can uniformly participate in subsequent probabilistic inference.

[0074] Furthermore, when constructing a multi-source evidence set, various types of evidence can be uniformly organized. This uniform organization may include evidence field standardization, evidence state normalization, and evidence time window correspondence processing. Evidence field standardization is used to organize component-side evidence, control-side evidence, and load-side evidence according to a unified data structure; evidence state normalization is used to organize evidence from different sources into state results or level results suitable for Bayesian network input; and evidence time window correspondence processing is used to ensure that all evidence entering the same probabilistic inference corresponds to the same analysis window, avoiding distortion of inference results due to time misalignment.

[0075] It should be noted that the evidence items in the multi-source evidence set can be selected based on the monitoring capabilities of the target unit, the types of components to be evaluated, and actual deployment requirements. In one embodiment, the component status assessment results of the main shaft, gearbox, and generator can be selected as the main component-side evidence; in another embodiment, the status assessment results of the nacelle structure and tower structure can also be included simultaneously. Similarly, for control-side evidence, the main characteristic results of pitch control residuals and yaw control residuals can be prioritized; for load-side evidence, the unified equivalent load characterization quantity of the current analysis window can be prioritized, and the load characterization results in the thrust direction and torque direction can also be further included. As long as the selected evidence can reflect the external excitation, component response, and control chain status respectively, the reasoning requirements of this invention can be met.

[0076] Furthermore, after obtaining the multi-source evidence set, a Bayesian network model can be established for anomaly source determination. The Bayesian network model may include anomaly source nodes and evidence nodes corresponding to the multi-source evidence set. The anomaly source nodes represent possible anomaly source categories within the current analysis window, and the evidence nodes receive specific evidence items from the multi-source evidence set. The values ​​of various types of evidence within the current analysis window serve as the input states for the evidence nodes. Through the conditional associations between each anomaly source node and each evidence node, probabilistic reasoning regarding anomaly sources is achieved.

[0077] The anomaly source nodes can include at least wind condition excitation anomalies, control chain execution anomalies, and component degradation anomalies. Wind condition excitation anomalies characterize abnormal load fluctuations caused by drastic changes in the external wind environment; control chain execution anomalies characterize execution deviations, response hysteresis, or action mismatches in the pitch control system or yaw control system; and component degradation anomalies characterize abnormal responses of unit components under normal load conditions caused by factors such as wear, fatigue, cracks, or poor lubrication. By setting these anomaly source categories, the Bayesian network can attribute observed multi-source evidence to the most probable anomaly source during the inference process.

[0078] In one implementation, the conditional probability relationships in the Bayesian network can be determined through statistical learning based on historical fault samples. Specifically, operational data with clearly identified anomaly sources from the target unit's historical operation can be collected. The component status assessment results, control chain anomaly assessment results, and equivalent load characterization quantities within the corresponding analysis window before the anomaly occurred are correlated with the anomaly source label. The value distribution of various types of evidence under different anomaly source conditions is statistically analyzed, thereby forming the corresponding conditional probability relationships. In another implementation, the conditional probability relationships can also be determined based on simulation data calibration. Specifically, different anomaly scenarios can be simulated using a wind turbine simulation model, and the characteristics of multi-source evidence under each scenario can be recorded. Conditional probability relationships are then formed through statistical analysis. Alternatively, they can be set based on expert experience. That is, domain technicians conduct qualitative analysis of the possible manifestations of various types of evidence under different anomaly sources based on the wind turbine's operating mechanism, fault propagation path, and control chain execution rules, and determine the corresponding conditional probability relationships accordingly. The above methods can be used individually or in combination.

[0079] Furthermore, after the Bayesian network model is established, the multi-source evidence set corresponding to the current analysis window can be input into the Bayesian network model for probabilistic inference. Specifically, the component state evaluation results, control chain anomaly evaluation results, and equivalent load representation quantities within the current analysis window can be converted into the current states of the corresponding evidence nodes. Then, based on the pre-established conditional probability relationships in the Bayesian network, the posterior probabilities of different anomaly source categories under the current evidence conditions are calculated. The posterior probability is used to characterize the credibility of a certain anomaly source category within the current analysis window.

[0080] After obtaining the posterior probabilities of each anomaly source category, the anomaly source category with the highest posterior probability can be used as the anomaly source determination result within the current analysis window, and this highest posterior probability value can be output as the anomaly risk confidence score. It should be noted that the anomaly risk confidence score characterizes the degree of matching between the current observational evidence and the determined anomaly source. The higher the anomaly risk confidence score, the higher the support of the current evidence for the anomaly source, and the higher the credibility of the anomaly determination result. If the posterior probabilities of multiple anomaly source categories are close, it indicates that the current multi-source evidence has a relatively weak ability to distinguish anomaly sources. In this case, the current anomaly risk confidence score can be combined with the current score for comprehensive processing during subsequent warning level matching.

[0081] Furthermore, in specific implementations, the Bayesian network can support continuous updates of evidence. When a rolling analysis window is used, the multi-source evidence set of the current window can be reconstructed after each window update, and probabilistic inference can be re-executed, thereby forming a continuously updated sequence of anomaly risk confidence and anomaly source determination. In this way, the development process of the unit's abnormal state over time can be tracked, providing dynamic input for the early warning level matching in step S6.

[0082] It should also be noted that the role of the Bayesian network in this step is to perform unified reasoning on the conditional relationships between multi-source evidence, rather than replacing the specific analysis processes of steps S2, S3, and S4. The equivalent load representation output of step S2 is used to reflect the intensity of external excitation; the component status assessment result output of step S3 is used to reflect the response deviation of unit components under the current load conditions; and the control chain anomaly assessment result output of step S4 is used to reflect the execution status of the control chain. By simultaneously introducing the above three types of results into the same reasoning framework, the determination of the anomaly source can be based on the joint support of three types of information: load, component, and control.

[0083] In this embodiment, step S5 constructs a multi-source evidence set based on component status assessment results, control chain anomaly assessment results, and equivalent load characterization quantities. Then, probabilistic reasoning is performed on this multi-source evidence set using a Bayesian network to obtain anomaly risk confidence and anomaly source determination results. The anomaly risk confidence is used to characterize the credibility of anomaly occurrence within the current analysis window, and the anomaly source determination results are used to characterize whether the current anomaly is more likely to originate from wind-induced excitation, control chain execution, or component degradation, thus providing a reasoning basis for matching early warning levels and outputting fault early warning results in step S6.

[0084] S6: Based on the abnormal risk confidence level and the abnormal source determination result, match the corresponding early warning level and output the wind turbine generator status assessment result and fault early warning result.

[0085] In this embodiment, step S6 is used to match the corresponding early warning level based on the abnormal risk confidence level and abnormal source determination result obtained in step S5, and output the wind turbine generator set status assessment result and fault early warning result.

[0086] Specifically, the anomaly risk confidence level and anomaly source determination result output in step S5 can be read first. The anomaly risk confidence level can be the degree of credibility corresponding to the current primary anomaly source obtained through probabilistic inference in step S5, used to characterize the credibility of the anomaly determination result within the current analysis window; the anomaly source determination result is used to characterize which of the following is more likely to originate from: wind condition excitation anomaly, control chain execution anomaly, or component degradation anomaly. Further, the anomaly risk confidence level can be mapped according to a pre-set warning level matching rule to determine the warning level corresponding to the current analysis window.

[0087] In one implementation, the warning levels can be divided into normal state, attention state, warning state, and alarm state. The normal state indicates that no obvious anomalies are found within the current analysis window, or the confidence level of the anomaly risk is within the normal fluctuation range. The attention state indicates that some signs of anomalies have appeared within the current analysis window, but the confidence level of the anomalies is relatively low, making continued observation suitable. The warning state indicates that the confidence level of the anomalies within the current analysis window is high, and there is a clear source of the anomaly, requiring entry into a state of close attention. The alarm state indicates that the confidence level of the anomalies within the current analysis window is high, and the anomaly manifestation is persistent or obvious, requiring the output of a clear fault warning result. It should be noted that the classification of the warning levels is not limited to the above four-level format and can be adjusted to other levels according to the operation and maintenance strategy of the target unit.

[0088] Furthermore, the early warning level matching rules can be set according to the range of abnormal risk confidence levels. Specifically, different abnormal risk confidence levels can be preset, and when the current abnormal risk confidence level falls into a certain range, the corresponding early warning level is matched. For example, when the abnormal risk confidence level is in a low range, it can be matched as a "watch" status; when the abnormal risk confidence level is in a medium range, it can be matched as a "warning" status; and when the abnormal risk confidence level is in a high range, it can be matched as an "alarm" status. The abnormal risk confidence level range can be preset based on the target unit's historical operating data, historical anomaly records, and wind farm operation and maintenance experience, and can also be adjusted according to different turbine models and different site operation characteristics.

[0089] It should also be noted that the anomaly source determination result can be used as an auxiliary reference when matching warning levels. When the anomaly source determination result is an abnormal wind condition, the warning level can be appropriately constrained while maintaining the anomaly risk confidence level, to avoid excessive warnings due to short-term environmental disturbances. When the anomaly source determination result is an abnormal control chain execution or an abnormal component degradation, the corresponding warning level can be matched according to the duration and severity of the anomaly. By incorporating both the anomaly risk confidence level and the anomaly source determination result into the warning level matching process, the warning output can be more consistent with the actual operating characteristics of wind turbine generators.

[0090] Furthermore, after determining the warning level, wind turbine generator status assessment results and fault warning results can be constructed. The status assessment results are used to characterize the overall status of the unit within the current analysis window, and may include at least the time stamp corresponding to the current analysis window, the unit identifier, the anomaly risk confidence level, the anomaly source determination result, the warning level, and a summary of key evidence related to the determination result. The fault warning results are used to characterize whether the current unit has entered an abnormal state requiring close attention, and the corresponding output format can be determined according to the warning level.

[0091] In some implementations, the fault warning result may also include maintenance prompts as optional extended output. These maintenance prompts can be generated based on the warning level and the determination of the anomaly source, such as "It is recommended to strengthen monitoring," "It is recommended to plan a shutdown for maintenance," or "It is recommended to immediately shut down for inspection." It should be noted that these maintenance prompts are intended to assist maintenance personnel in understanding the current warning result and do not constitute a necessary limitation of this step.

[0092] It should also be noted that the fault warning results output in this step are not required to directly trigger unit shutdown or control strategy adjustments. Instead, they are used to provide abnormal status alerts and risk level assessment criteria for the operation and maintenance system, monitoring platform, or manual on-duty personnel. In other words, the focus of this step's output is to achieve hierarchical expression of abnormal results and standardized output of warning information, rather than replacing the control logic of the unit's control system itself.

[0093] In some implementations, the status assessment results and fault warning results generated within the current analysis window can be associated and stored with time stamps, unit identifiers, operating condition tags, and historical warning records. When a rolling analysis window is used, the abnormal risk confidence level and abnormal source determination results can be re-acquired with the formation of each new analysis window, and the corresponding warning level can be re-matched, thereby forming a continuously updated sequence of warning results.

[0094] It should be noted that the component status assessment result formed in step S3, the control chain anomaly assessment result formed in step S4, and the warning level output in this step represent different levels of results. Specifically, the component status assessment result characterizes the response deviation of unit components under corresponding load conditions, the control chain anomaly assessment result characterizes the execution status of the control link, and the warning level characterizes the comprehensive risk level formed based on the anomaly risk confidence level and the anomaly source determination result. By distinguishing between different levels of results, this invention can reflect both specific anomalies on the component and control sides, and also output a unified warning result for the overall machine operating status.

[0095] In this embodiment, step S6 matches the corresponding early warning level based on the anomaly risk confidence level and the anomaly source determination result, and outputs the wind turbine generator status assessment result and fault early warning result. The status assessment result reflects the overall operating status of the unit within the current analysis window, and the fault early warning result characterizes whether the unit has entered an abnormal state requiring close attention and its corresponding risk level, thereby providing the final output results for the operation monitoring and fault early warning of the wind turbine generator.

[0096] Example 2 is the second embodiment of the present invention, which differs from the previous embodiment in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art or the current technical solution, can be embodied in the form of a software product. This current computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0098] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0099] Example 3, referring to Figure 2 As an embodiment of the present invention, a wind turbine generator condition assessment and fault early warning system is provided, which includes a data acquisition module, a load characterization module, a condition assessment module, a control assessment module, a probabilistic reasoning module, and an early warning output module; Data acquisition module: acquires the operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The control feedback data includes pitch control data, yaw control data and corresponding execution feedback data. Load characterization module: Based on the wind condition parameters and the unit operating status parameters, construct an equivalent load characterization quantity that characterizes the relationship between wind disturbance and unit load; Condition assessment module: Based on the equivalent load characterization quantity and the unit operating status parameters, determine the response offset characteristics of the unit components relative to the corresponding load response benchmark, and obtain the component condition assessment results; Control evaluation module: Based on the control feedback data, construct pitch control residuals and yaw control residuals to obtain control chain anomaly evaluation results; Probabilistic reasoning module: Based on the component state assessment results, the control chain anomaly assessment results, and the equivalent load characterization quantity, a multi-source evidence set is constructed, and probabilistic reasoning is performed on the multi-source evidence set based on a Bayesian network to obtain the anomaly risk confidence level and anomaly source determination results; Early warning output module: Based on the confidence level of the abnormal risk and the determination result of the abnormal source, match the corresponding early warning level and output the status assessment result and fault early warning result of the wind turbine generator set.

[0100] Example 4 is an embodiment of the present invention, which provides a method for wind turbine generator condition assessment and fault early warning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / comparative experiments.

[0101] This embodiment selects a GW-2500 wind turbine generator set from a wind farm as the test object. The unit has a rated power of 2500kW, a rotor diameter of 115m, and a hub height of 90m. The test period is set from March 1, 2024 to April 15, 2024, with continuous data collection for 45 days to verify the practical application effect of the present invention in wind disturbance identification, control chain anomaly identification, component status assessment, anomaly source determination, and graded early warning output. The unit is equipped with a nacelle anemometer, wind direction measurement device, vibration sensor, temperature sensor, electrical parameter acquisition module, and control system data interface, which can acquire wind parameters, unit operating status parameters, and control feedback data in real time. The data sampling frequency is set to 1Hz, and the data is collected in 10-minute analysis windows to form a dataset that can be directly used for subsequent load characterization, component status assessment, control chain anomaly identification, and probabilistic reasoning.

[0102] Before the experiment began, basic models and reference benchmarks for each step were established based on historical healthy operation samples. Specifically, the operating data of the unit during the period from November 2023 to February 2024, during which it operated continuously and stably without fault records, abnormal alarms, or abnormal maintenance, was selected as the healthy sample, lasting for a total of 90 days. For step S2, a load mapping model was established based on the simulation data formed during the unit design phase and the stable operating data in the healthy sample. The load mapping model was implemented using a lookup table mapping method. Its inputs included wind speed, wind direction deviation, turbulence intensity, generator speed, actual blade pitch angle, and nacelle azimuth angle. Its outputs included equivalent thrust load and equivalent torque load, used to form the equivalent load characterization quantity corresponding to the current analysis window. For step S3, the gearbox high-speed shaft bearing temperature was selected as the response parameter of the component to be evaluated. A load response benchmark model was established based on the healthy sample. In this embodiment, a Gaussian process regression method was used, with the equivalent load characterization quantity as input and bearing temperature as output, to obtain the benchmark response results and allowable fluctuation range corresponding to different load conditions under the healthy state. For step S4, based on the control residual characteristics of the pitch control system and yaw control system statistically analyzed from health samples, a control residual reference benchmark is formed. This benchmark includes the normal fluctuation range of the pitch instantaneous residual mean, pitch response delay, cumulative pitch deviation, yaw instantaneous residual, and yaw start delay. For step S5, a Bayesian network is established by combining historical fault samples and on-site operation and maintenance experience. The network includes three anomaly source nodes: wind condition excitation anomaly, control chain execution anomaly, and component degradation anomaly. It also includes three evidence nodes: component-side evidence, control-side evidence, and load-side evidence, used for probabilistic inference of anomaly sources within each analysis window.

[0103] During the experiment, the unit was continuously monitored and analyzed according to steps S1 to S6 of this invention. After each 10-minute analysis window, the wind condition parameters, unit operating status parameters, and control feedback data within the current window were extracted from the dataset to be analyzed. For step S2, the average wind speed, average wind direction deviation, average turbulence intensity, average generator speed, average blade pitch angle, and average nacelle azimuth angle within the current window were input into the load mapping model to obtain the equivalent thrust load and equivalent torque load, and these two were then unified to form the equivalent load characterization quantity for the current window. For step S3, this equivalent load characterization quantity was input into the load response benchmark model for the high-speed shaft bearing temperature of the gearbox to obtain the bearing temperature benchmark response result and allowable fluctuation range under the current load conditions. This result was then compared with the bearing temperature actually collected within the current window, and the standardized offset was calculated to form the component condition assessment result. For step S4, the target pitch angle command sequence output by the pitch controller and the actual pitch angle feedback sequence fed back by the pitch driver are extracted from the control feedback data in the current window. The instantaneous pitch residual sequence, pitch response delay, and cumulative pitch deviation are calculated. Simultaneously, the target yaw angle command sequence output by the yaw controller and the actual nacelle azimuth feedback sequence fed back by the yaw encoder are extracted. The instantaneous yaw residual sequence and yaw start delay are calculated. The above control residual characteristics are then compared with the control residual reference benchmark to obtain the control chain anomaly assessment result. For step S5, the equivalent load characterization quantity is used to form load-side evidence, the component state assessment result is used to form component-side evidence, and the control chain anomaly assessment result is used to form control-side evidence. These are then input into a pre-established Bayesian network for probabilistic inference to obtain the posterior probabilities of three types of anomaly sources—wind excitation anomaly, control chain execution anomaly, and component degradation anomaly—under the current evidence conditions. The maximum posterior probability value is used as the anomaly risk confidence level, and the corresponding anomaly source is used as the anomaly source determination result. For step S6, the confidence level of abnormal risks is matched according to the pre-set confidence level range, and the warning level is divided into normal state, attention state, warning state and alarm state. At the same time, auxiliary adjustments are made in combination with the abnormal source determination result. When the abnormal source determination result is wind condition excitation abnormality, the warning level is constrained to avoid excessive warning caused by short-term environmental disturbances. Finally, the status assessment result and fault warning result corresponding to the current analysis window are output.

[0104] To verify the effectiveness of this invention in identifying control chain anomalies, a controlled disturbance test was conducted on the pitch control link of the unit during the experiment. Specifically, from 10:00 on April 5, 2024 to 16:00 on April 8, 2024, a slight hysteresis disturbance was applied to the pitch actuator link, causing a static deviation of approximately 0.3° between the actual pitch angle and the target pitch angle, accompanied by a response delay of more than 0.5 seconds, to simulate the condition of slight jamming and slow response of the pitch actuator. Relevant data were continuously collected during the experiment, and key analysis windows before the fault occurred, in the early stage of the fault, during the fault development stage, during the fault significant stage, and after the fault was cleared were selected as representative results and recorded in the table below. The key data recorded are shown in Table 1: Table 1: Experimental Data Recording Table

[0105] As can be seen from the data in Table 1, this invention can identify control chain execution anomalies earlier and more stably in this continuous fault evolution scenario, and can distinguish them from component degradation anomalies and wind excitation anomalies. First, at 08:00 on April 5, 2024, before the fault occurred, the average instantaneous residual of pitch was only 0.12°, the pitch response delay was only 0.21s, the cumulative pitch deviation was 0.84°·s, the standardized offset between the actual bearing temperature and the baseline value was -0.3, and all types of posterior probabilities were at a low level. Ultimately, no anomalies were output, indicating that there is a good correspondence between the baseline model established by the healthy sample and the actual operating state. Secondly, immediately after the fault occurred at 10:00 AM on April 5, 2024, the mean instantaneous residual of the pitch control increased from 0.12° to 0.28°, the pitch response delay increased from 0.21s to 0.68s, and the cumulative pitch deviation increased to 2.47°·s. Meanwhile, the bearing temperature normalization offset was only 0.5, indicating that no significant abnormal offset had yet occurred on the component side, but a significant execution mismatch had already deviated from the healthy baseline on the control chain side. At this point, the invention increased the posterior probability of control chain execution anomaly to 0.35 and determined the source of the anomaly to be control chain execution anomaly, outputting a state of concern, demonstrating its ability to provide directional early warning information in the early stages of the fault. Thirdly, as the disturbance continued, the posterior probability of control chain execution anomaly rose to 0.79 at 9:00 AM on April 6, 2024, and the warning level was raised to a warning state; by 2:00 PM on April 8, 2024, the posterior probability of control chain execution anomaly further increased to 0.94, outputting an alarm state. Throughout the process, the posterior probability of control chain execution anomalies showed a continuous upward trend, and the warning level was upgraded step by step according to the path of attention, warning, and alarm. This indicates that the anomaly risk confidence level output by the present invention can reflect the degree of fault evolution well and can provide maintenance personnel with a basis for risk judgment over time.

[0106] Furthermore, the component condition assessment results demonstrate that this invention possesses excellent anomaly source differentiation capabilities. Although the bearing temperature normalization offset in the table gradually increased from 0.5 to 2.8, it never exceeded the high offset range used for severe component degradation determination. Moreover, the posterior probability of component degradation anomaly remained between 0.15 and 0.22 throughout the entire experiment, significantly lower than the posterior probability of control chain execution anomaly. This indicates that in the current scenario, component response changes are mainly indirectly affected by control chain execution deviations, rather than directly caused by component degradation itself. Simultaneously, the posterior probability of wind excitation anomaly remained at a low level, indicating that although wind speed and power fluctuated normally during this period, they did not constitute a dominant anomaly source. Therefore, this invention, by introducing three types of evidence—equivalent load characterization, component condition assessment results, and control chain anomaly assessment results—and utilizing Bayesian networks for unified reasoning, can avoid misjudging control chain anomalies as component failures and also prevent misjudging normal operating condition fluctuations as abnormal events. If the traditional fixed threshold method is used, an alarm is only triggered when the instantaneous residual pitch exceeds 0.5° or the response delay exceeds 1.0s. In this experiment, a significant alarm would not be triggered until 14:00 on April 8, 2024. However, the present invention identified the abnormal trend in the control chain as early as 10:00 on April 5, 2024, providing an early warning more than three days earlier. This result shows that the present invention can not only improve the sensitivity of identifying anomalies in the control chain of wind turbine generators, but also improve the interpretability and engineering applicability of the early warning results through anomaly source attribution and confidence quantification. Thus, it demonstrates significant technical advantages over existing technologies in false alarm suppression, early identification, anomaly differentiation, and tiered output.

[0107] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for condition assessment and fault early warning of wind turbine generator sets, characterized in that, include: S1: Acquire the operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The control feedback data includes pitch control data, yaw control data and corresponding execution feedback data. S2: Based on the wind condition parameters and the unit operating status parameters, construct an equivalent load characterization quantity that represents the relationship between wind disturbance and unit load transmission; S3: Based on the equivalent load characterization quantity and the unit operating status parameters, determine the response offset characteristics of the unit components relative to the corresponding load response benchmark, and obtain the component status assessment results; S4: Based on the control feedback data, construct the pitch control residual and yaw control residual to obtain the control chain anomaly evaluation results; S5: Construct a multi-source evidence set based on the component status assessment results, the control chain anomaly assessment results, and the equivalent load characterization quantity, and perform probabilistic reasoning on the multi-source evidence set based on a Bayesian network to obtain the anomaly risk confidence level and anomaly source determination results; S6: Based on the abnormal risk confidence level and the abnormal source determination result, match the corresponding early warning level and output the wind turbine generator status assessment result and fault early warning result.

2. The method for wind turbine generator condition assessment and fault early warning as described in claim 1, characterized in that, Step S1 specifically includes: The system acquires operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The wind condition parameters include real-time wind speed, wind direction deviation, and turbulence intensity. The generator set operation status parameters include generator speed, output power, and actual values ​​of blade pitch angle. The control feedback data includes pitch control data, yaw control data, and corresponding execution feedback data. The pitch control data includes a target pitch angle command sequence output by the pitch controller. The yaw control data includes a target yaw angle command sequence output by the yaw controller. The execution feedback data includes an actual pitch angle feedback sequence corresponding to the target pitch angle command sequence and an actual nacelle azimuth angle feedback sequence corresponding to the target yaw angle command sequence. The operation monitoring data and control feedback data are processed by time alignment, missing value filling, outlier removal and data format unification, and then collected according to the preset analysis window to form a dataset to be analyzed.

3. The wind turbine generator condition assessment and fault early warning method as described in claim 2, characterized in that, Step S2 specifically includes: Feature extraction is performed on the wind condition parameters to obtain the wind disturbance description results; Based on the wind disturbance description results and the unit operating status parameters, a load mapping model is established to characterize the correspondence between wind parameters, unit operating status parameters and unit load levels. Input the wind condition parameters and unit operating status parameters in the current analysis window into the load mapping model to obtain the load characterization results corresponding to the current analysis window.

4. The wind turbine generator condition assessment and fault early warning method as described in claim 3, characterized in that, In step S2, the load characterization results include thrust direction load characterization results and torque direction load characterization results. The thrust direction load characterization results are used to characterize the load level transmitted along the main shaft direction after the wind turbine receives wind, and the torque direction load characterization results are used to characterize the load level transmitted from the wind turbine rotation driving force to the generator input side through the main shaft and transmission chain. The thrust direction load characterization results and the torque direction load characterization results are unified and organized to form the equivalent load characterization quantity.

5. The method for wind turbine generator condition assessment and fault early warning as described in claim 4, characterized in that, In step S3, the load response benchmark is established in the following manner: Using the equivalent load characterization quantity as input and the component response parameters corresponding to the component to be evaluated as output, a regression model describing the mapping relationship between load conditions and component response is established. Input the equivalent load characterization quantity corresponding to the current analysis window into the regression model to obtain the baseline response result of the component to be evaluated under the current load conditions.

6. The method for wind turbine generator condition assessment and fault early warning as described in claim 5, characterized in that, In step S3, the response offset feature is a standardized offset. The standardized offset is obtained by comparing the difference between the actual response value of the component to be evaluated in the current analysis window and the baseline response result with the response fluctuation scale under the same load conditions in historical healthy samples. Based on whether the standardized offset exceeds the allowable fluctuation range of the baseline response, the degree of excess, and the duration of occurrence, the current state of the component to be evaluated is classified to form a component state evaluation result.

7. The wind turbine generator condition assessment and fault early warning method as described in claim 6, characterized in that, Step S4 specifically includes: Extract the target pitch angle command sequence of the pitch controller and the actual pitch angle feedback sequence of the pitch drive from the control feedback data, as well as the target yaw angle command sequence of the yaw controller and the actual cabin azimuth angle feedback sequence from the yaw encoder. Based on the target pitch angle command sequence and the actual pitch angle feedback sequence, calculate the pitch instantaneous residual sequence, pitch response delay, and pitch cumulative deviation. Based on the target yaw angle command sequence and the actual cabin azimuth feedback sequence, calculate the yaw instantaneous residual sequence and the yaw start delay; The control chain anomaly assessment results are obtained by comparing the instantaneous residual sequence of pitch, the pitch response delay, the cumulative pitch deviation, the instantaneous residual sequence of yaw, and the yaw start delay with the control residual reference benchmark corresponding to the historical healthy operation phase.

8. The method for wind turbine generator condition assessment and fault early warning as described in claim 7, characterized in that, Step S5 specifically includes: Component-side evidence is constructed based on the component status assessment results, control-side evidence is constructed based on the control chain anomaly assessment results, and load-side evidence is constructed based on the equivalent load characterization quantity. A multi-source evidence set is formed by the component-side evidence, the control-side evidence, and the load-side evidence. The multi-source evidence set is input into a pre-established Bayesian network, which includes anomaly source nodes and evidence nodes corresponding to the multi-source evidence set. The anomaly source nodes include wind condition excitation anomalies, control chain execution anomalies, and component body degradation anomalies. Based on the evidence status of each type of evidence in the multi-source evidence set within the current analysis window, calculate the posterior probability of each anomaly source category under the current evidence conditions; The anomaly source category with the highest posterior probability is taken as the anomaly source determination result, and the highest posterior probability value is output as the anomaly risk confidence level.

9. The method for wind turbine generator condition assessment and fault early warning as described in claim 8, characterized in that, Step S6 specifically includes: Based on the abnormal risk confidence level obtained in step S5, it is matched with the preset abnormal risk confidence level range to determine the warning level corresponding to the current analysis window. The warning level includes normal state, attention state, warning state and alarm state. When matching the warning level, the abnormal source determination result obtained in step S5 is used for auxiliary adjustment. When the abnormal source determination result is wind excitation abnormality, the warning level is constrained. Based on the anomaly source determination result, the anomaly risk confidence level, and the warning level, the wind turbine generator status assessment result and fault warning result corresponding to the current analysis window are generated.

10. A wind turbine generator condition assessment and fault early warning system, used to implement the wind turbine generator condition assessment and fault early warning method as described in any one of claims 1 to 9, characterized in that, include: Data acquisition module: acquires the operation monitoring data and control feedback data of the wind turbine generator set. The operation monitoring data includes wind condition parameters and generator set operation status parameters. The control feedback data includes pitch control data, yaw control data and corresponding execution feedback data. Load characterization module: Based on the wind condition parameters and the unit operating status parameters, construct an equivalent load characterization quantity that characterizes the relationship between wind disturbance and unit load; Condition assessment module: Based on the equivalent load characterization quantity and the unit operating status parameters, determine the response offset characteristics of the unit components relative to the corresponding load response benchmark, and obtain the component condition assessment results; Control evaluation module: Based on the control feedback data, construct pitch control residuals and yaw control residuals to obtain control chain anomaly evaluation results; Probabilistic reasoning module: Based on the component state assessment results, the control chain anomaly assessment results, and the equivalent load characterization quantity, a multi-source evidence set is constructed, and probabilistic reasoning is performed on the multi-source evidence set based on a Bayesian network to obtain the anomaly risk confidence level and anomaly source determination results; Early warning output module: Based on the confidence level of the abnormal risk and the determination result of the abnormal source, match the corresponding early warning level and output the status assessment result and fault early warning result of the wind turbine generator set.

Images