Wind turbine generator variable pitch power supply health diagnosis method and system based on data monitoring

Abstract

The embodiment of the application provides a wind turbine variable pitch power health diagnosis method and system based on data monitoring, which, in real-time diagnosis, applies corresponding working condition criteria to self-adaptive normalization processing of monitoring data, effectively strips strong interference brought by external variables, and extracts a pure feature vector reflecting only the internal health condition of the variable pitch power. Subsequently, the pure feature is input into a reconstruction model trained only on health data for abnormality evaluation. The model judges whether the current state is abnormal through the ability of reconstructing the input data; for healthy input, the reconstruction error is small, and for abnormal input, a significant reconstruction error is generated. This method can accurately identify the real early degradation signal from the strong interference background, improving the accuracy and reliability of diagnosis.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, specifically to a method and system for health diagnosis of wind turbine pitch power supply based on data monitoring. Background Technology

[0002] Wind power, as a clean and renewable energy source, occupies an increasingly important position in the global energy structure. Wind turbines are the core equipment of wind power generation, and their operational safety and stability directly affect the economic benefits of the entire wind farm. The pitch power supply, as the power source for the pitch system, is typically installed inside the hub in harsh environments. Its ability to provide stable and reliable power under various operating conditions is crucial for the safe shutdown and efficient operation of the entire unit. Given the high failure rate of the pitch power supply and its critical role in the system, developing a method to accurately and promptly diagnose its health status is of great practical significance for ensuring the safety of wind turbines and improving operation and maintenance efficiency.

[0003] However, existing methods for diagnosing the health of wind turbine pitch power supplies face significant challenges in practical applications. Traditional diagnostic methods often rely on setting fixed thresholds for monitoring parameters, triggering alarms when these thresholds are exceeded. The drawback of this approach is its strong lag and inability to effectively adapt to the complex and variable working environment and operating conditions of wind turbines. The performance parameters of pitch power supplies, such as charging and discharging voltage and current, not only change with the decline in their own health but are also severely affected by external environmental factors and turbine operating conditions. For example, in frigid winter conditions, the internal resistance of the power supply increases significantly, causing changes in its charging and discharging response characteristics and producing data appearances similar to internal aging. Under turbulent or gusty conditions, the control system instructs the pitch system to perform frequent, small-amplitude movements, resulting in voltage and current data exhibiting fluctuation patterns drastically different from those under stable wind conditions. These normal data fluctuations caused by external factors, coupled with subtle fault characteristics resulting from early performance degradation of the power supply itself, make traditional diagnostic methods highly prone to misjudgment, leading to a persistently high false alarm rate. This not only consumes a lot of maintenance manpower and resources to troubleshoot non-existent faults, but may also cause maintenance personnel to become complacent about the alarm system, thus missing the opportunity to provide early warning of real faults.

[0004] Therefore, an optimized health diagnosis solution for wind turbine pitch power supplies is desired. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the prior art, and provides a method and system for health diagnosis of wind turbine pitch power supply based on data monitoring.

[0006] In a first aspect, embodiments of the present invention provide a method for health diagnosis of wind turbine pitch power supply based on data monitoring, comprising: The acquired real-time monitoring data of the wind turbines is time-series aligned to obtain aligned time-series data; The aligned time series data is divided into working condition spaces and labeled to obtain data with working condition labels; Extract the normalized parameters corresponding to the data with working condition labels from the normalized parameter mapping table to obtain the time series of normalized feature vectors; The time series of normalized feature vectors are reconstructed and anomaly scored using a pre-trained autoencoder model to obtain the time series of health status indicators. The time series of health status indicators are aggregated to obtain a health index, and a trend chart of the health index of wind turbine pitch power supply is generated based on the health index.

[0007] In the above-mentioned data-based wind turbine pitch power supply health diagnosis method, the acquired wind turbine monitoring data is time-series aligned to obtain aligned time series data. This includes: based on timestamps, resampling and cleaning the wind turbine monitoring data to obtain aligned time series data. The wind turbine monitoring data includes ambient temperature, wind speed, blade motor current, pitch battery voltage, and pitch battery current.

[0008] In the above-mentioned data monitoring-based wind turbine pitch power supply health diagnosis method, the aligned time series data is divided into operating condition spaces and labeled to obtain data with operating condition labels. This includes: matching each operating condition data in the aligned time series data with the operating condition space to divide each operating condition data into the corresponding operating condition space; and labeling each operating condition data with operating condition labels based on the divided operating condition space to obtain data with operating condition labels.

[0009] In the above-mentioned data-based wind turbine pitch power supply health diagnosis method, the step of generating the normalized parameter mapping table includes: extracting data from historical data on the healthy operating cycle of the pitch power supply as a health dataset; grouping the health dataset according to operating condition labels to obtain a set of health data groups; for each health data group in the set of health data groups, calculating the power supply diagnostic characteristics of the same type of operating condition data within the group as normalized parameters, the power supply diagnostic characteristics including mean and standard deviation; and storing the normalized parameters and their corresponding operating condition labels as key-value pairs in the normalized parameter mapping table.

[0010] In the aforementioned data-based wind turbine pitch power supply health diagnosis method, the key feature is that the time series of normalized feature vectors obtained by extracting normalized parameters corresponding to the data with operating condition labels from the normalized parameter mapping table includes: extracting normalized parameters corresponding to the operating condition labels from the normalized parameter mapping table for the operating condition data of each operating condition space in the data with operating condition labels; standardizing the various types of operating condition data based on the normalized parameters to obtain normalized feature vectors; and regularizing the multiple normalized feature vectors according to the time dimension to obtain the time series of normalized feature vectors.

[0011] In the above-mentioned data monitoring-based wind turbine pitch power supply health diagnosis method, the normalization process is performed on various types of operating condition data to obtain a normalized feature vector based on normalization parameters. This includes: subtracting the mean of the operating condition data from the mean of the corresponding operating condition interval, and then dividing the difference by the standard deviation of the corresponding operating condition interval to obtain a normalized value; and combining the normalized values ​​corresponding to various types of operating condition data into a normalized feature vector.

[0012] In the above-mentioned data monitoring-based wind turbine pitch power supply health diagnosis method, the time series of normalized feature vectors is reconstructed and anomaly scored through a pre-trained autoencoder model to obtain the time series of health status indicators. This includes: reconstructing the normalized feature vectors through a pre-trained autoencoder model to obtain the reconstructed vectors; calculating the reconstruction error between the reconstructed vectors and the normalized feature vectors; and comparing the reconstruction error with a preset anomaly threshold to obtain the health status indicators.

[0013] In the above-mentioned data monitoring-based wind turbine pitch power supply health diagnosis method, the calculation of the reconstruction error between the reconstruction vector and the normalized feature vector includes: using the mean square error algorithm to calculate the reconstruction error between the reconstruction vector and the normalized feature vector.

[0014] In the aforementioned data-monitoring-based health diagnosis method for wind turbine pitch power supplies, the time series of health status indicators are aggregated to obtain a health index. This includes: aggregating the time series of health status indicators using the following aggregation formula, where the aggregation formula is:

[0015] in, The health index at the previous moment. The health index is updated at the current moment. The trainingable decay coefficient, This represents the number of times the health status flag is marked as abnormal within a predetermined time window.

[0016] Secondly, embodiments of the present invention provide a wind turbine pitch power supply health diagnosis system based on data monitoring, comprising: The time-series alignment processing module is used to perform time-series alignment processing on the acquired real-time monitoring data of wind turbines to obtain aligned time-series data; The module for acquiring data with working condition labels is used to divide the aligned time series data into working condition spaces and label the data to obtain data with working condition labels; The feature vector time series extraction module is used to extract the normalization parameters corresponding to the data with working condition labels from the normalization parameter mapping table to obtain the time series of normalized feature vectors. The health status indicator time series acquisition module is used to reconstruct and anomaly score the time series of normalized feature vectors through a pre-trained autoencoder model to obtain the health status indicator time series. The power health index trend chart generation module is used to aggregate and calculate the time series of health status indicators to obtain the health index, and generate a wind turbine pitch power health index trend chart based on the health index.

[0017] Compared with existing technologies, this invention proposes a data monitoring-based health diagnosis method for wind turbine pitch power supplies. First, it divides the complex operating state of the wind turbine into multiple independent operating condition spaces based on key external variables such as ambient temperature and wind speed. For each operating condition space, a dedicated health status benchmark model is established using historical health data. During real-time diagnosis, the monitoring data is first assigned to its corresponding operating condition space, and adaptive normalization is performed using the health benchmark of that space. This step effectively isolates normal data fluctuations caused by changes in environment and operating conditions, obtaining a feature vector that only reflects the intrinsic health status of the pitch power supply. Subsequently, this purified feature vector is input into a reconstruction model that has only learned the health pattern. If the model cannot accurately reconstruct this vector, resulting in a large reconstruction error, it indicates that the power supply has deviated from its true normal performance under that specific operating condition. This two-step strategy can accurately identify early degradation features from a strong interference background, thereby improving the accuracy and reliability of diagnosis and reducing the false alarm rate. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1This is a flowchart of a data monitoring-based health diagnosis method for wind turbine pitch power supply according to an embodiment of the present invention; Figure 2 This is a data flow diagram of a wind turbine pitch power supply health diagnosis method based on data monitoring according to an embodiment of the present invention. Figure 3 A flowchart illustrating the generation of a normalized parameter mapping table for a data monitoring-based wind turbine pitch power supply health diagnosis method according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the time series of a wind turbine pitch power supply health diagnosis method based on data monitoring according to an embodiment of the present invention, which extracts normalized parameters corresponding to data with operating condition labels from a normalized parameter mapping table to obtain normalized feature vectors. Figure 5 This is a block diagram of a wind turbine pitch power supply health diagnosis system based on data monitoring, according to an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0022] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0023] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0024] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention; it should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0025] Existing health diagnosis methods for wind turbine pitch power supplies are ill-suited to complex and variable operating environments, and cannot effectively distinguish between power supply performance degradation and normal data fluctuations caused by changes in the external environment and operating conditions, resulting in a high false alarm rate. Therefore, this application proposes a data monitoring-based health diagnosis method for wind turbine pitch power supplies. Specifically, the process first performs time-series alignment on the acquired real-time monitoring data of the wind turbine. Next, the aligned data, especially variables reflecting the external environment and operating conditions, is matched with a preset operating condition space, assigning a precise operating condition label to each data record. Then, normalized parameters corresponding perfectly to the current data's operating condition label are extracted from a normalized parameter mapping table generated based on historical health data. These parameters are then used to standardize the real-time power supply diagnostic features, resulting in a normalized feature vector that eliminates the influence of operating conditions. Subsequently, this purified feature vector is input into an autoencoder model pre-trained only with healthy samples for data reconstruction. By calculating the reconstruction error between the original input vector and the model's reconstructed vector and comparing it with a preset threshold, the system determines whether a real anomaly exists at the current moment and outputs a health status indicator. Finally, the health status indicators on the time series are aggregated to generate a health index that intuitively reflects the long-term performance degradation trend of the pitch power supply, as well as a trend chart of the wind turbine pitch power supply health index.

[0026] Figure 1 This is a flowchart of a data monitoring-based health diagnosis method for wind turbine pitch power supply according to an embodiment of the present invention. Figure 2 This is a data flow diagram illustrating a data monitoring-based wind turbine pitch power supply health diagnosis method according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the wind turbine pitch power supply health diagnosis method and system based on data monitoring according to an embodiment of the present invention includes the following steps: S100, performing time-series alignment processing on the acquired real-time monitoring data of the wind turbine to obtain aligned time-series data; S200, dividing the aligned time-series data into operating condition spaces and labeling the data to obtain data with operating condition labels; S300, extracting the normalized parameters corresponding to the data with operating condition labels from the normalized parameter mapping table to obtain the time series of normalized feature vectors; S400, reconstructing and anomaly scoring the time series of normalized feature vectors through a pre-trained autoencoder model to obtain the time series of health status indicators; S500, performing aggregation operations on the time series of health status indicators to obtain a health index, and generating a wind turbine pitch power supply health index trend chart based on the health index.

[0027] Specifically, in step S100, the acquired real-time monitoring data of the wind turbine is time-series aligned to obtain aligned time-series data. It should be understood that since the monitoring data of the wind turbine originates from multiple sensors and subsystems with different sampling frequencies and communication protocols—for example, minute-level SCADA system data and second-level condition monitoring system data—the original data suffers from time misalignment and asynchrony. Therefore, in the technical solution of this application, the acquired real-time monitoring data of the wind turbine is time-series aligned to integrate independent data streams from diverse sources and with varying frequencies into a unified time-based, complete wide-table dataset. This provides a high-quality, highly reliable data foundation for subsequent operating condition classification, feature extraction, and model analysis, ensuring that the correlation between environmental, operating condition, and power state parameters is real and effective at the same time point, thereby avoiding analytical distortion caused by data misalignment.

[0028] Specifically, in this embodiment, the acquired monitoring data of the wind turbine is time-series aligned to obtain aligned time-series data. This includes: resampling and cleaning the monitoring data of the wind turbine based on timestamps to obtain aligned time-series data. The monitoring data of the wind turbine includes ambient temperature, wind speed, blade motor current, pitch battery voltage, and pitch battery current. More specifically, firstly, the required data streams are collected from multiple data sources, such as the SCADA system and condition monitoring system of the wind turbine. The former contains data such as pitch battery voltage, current, and wind speed with a ten-minute statistical cycle, while the latter contains ambient temperature data sampled at the second level. Next, a unified timestamp grid is constructed as the benchmark for data alignment, with a target time resolution of one minute. Subsequently, resampling operations are performed on data of different frequencies to adapt to the timestamp grid: for ambient temperature data sampled at the second level, it is downsampled to a one-minute resolution by calculating the average value within each minute; for SCADA data with a ten-minute cycle, it is upsampled using a forward padding method, that is, the value of each ten-minute cycle is assigned to ten one-minute time points within that cycle. After resampling, all data streams are merged according to a unified timestamp grid. During this process, data cleaning is performed to check and handle missing values ​​caused by sensor or communication interruptions. Short-term missing data points can be filled using linear interpolation. The final output is a time series data table with all monitored variables precisely aligned on the same time axis.

[0029] Specifically, in step S200, the aligned time-series data is divided into operating condition spaces and labeled to obtain data with operating condition tags. It should be understood that the normal range of performance parameters of wind turbine pitch power supplies, such as voltage and current, is closely related to the external environment and operating conditions. Even a healthy power supply can exhibit drastically different data characteristics under different temperatures or wind speeds. Therefore, in the technical solution of this application, the aligned time-series data is further divided into operating condition spaces and labeled to establish a clear context for each data point, discretizing the continuously changing operating state into a finite number of standardized scenarios for comparison. This provides accurate prerequisites for subsequently eliminating the influence of operating conditions and implementing adaptive diagnosis, enabling the model to evaluate data performance under specific scenarios, rather than making general, sweeping judgments.

[0030] More specifically, in a specific example of this application, the aligned time series data is divided into working condition spaces and labeled to obtain data with working condition labels, including: matching each working condition data in the aligned time series data with the working condition space to divide each working condition data into the corresponding working condition space; and labeling each working condition data with working condition labels based on the working condition space after data division to obtain data with working condition labels.

[0031] More specifically, firstly, the dimensions of the operational condition space and the threshold values ​​for each dimension are defined. Here, ambient temperature and wind speed are selected as the two core dimensions. Ambient temperature is divided into three intervals: below 0 degrees Celsius is the low-temperature zone, 0 to 25 degrees Celsius is the normal-temperature zone, and above 25 degrees Celsius is the high-temperature zone. Similarly, wind speed is also divided into three intervals: below 5 meters per second is the low-wind-speed zone, 5 to 12 meters per second is the medium-wind-speed zone, and above 12 meters per second is the high-wind-speed zone. By combining the above dimensional intervals, a two-dimensional operational condition space containing nine independent operational condition subspaces is constructed, and a unique operational condition label is assigned to each operational condition space. For example, the operational condition label for low temperature and high wind speed is denoted as C13. Subsequently, the aligned time series data table output from the previous process is traversed. For each row of data in the table, the values ​​of ambient temperature and wind speed are extracted and compared with the preset interval thresholds to determine the operational condition subspace to which the row of data belongs. After determining its affiliation, the unique working condition label corresponding to that working condition subspace, such as C13, is written as a new data field into the row of data, completing the labeling of that working condition data. This process is applied to the entire time series dataset, and the final data obtained is data with working condition labels added to the original aligned data.

[0032] Specifically, in step S300, normalized parameters corresponding to the data with operating condition labels are extracted from the normalized parameter mapping table to obtain the time series of normalized feature vectors. It should be understood that, after operating condition segmentation and labeling, the original values ​​of the power supply diagnostic features are still absolute quantities, coupled with the dual influence of the equipment's own health status and external operating conditions. This results in data from different operating conditions lacking direct comparability, making it difficult to establish a unified health assessment benchmark. Therefore, in the technical solution of this application, normalized parameters corresponding to the data with operating condition labels are further extracted from the normalized parameter mapping table to obtain the time series of normalized feature vectors. This allows for the invocation of a specific health status statistical benchmark based on the specific operating condition scenario of each data point, and adaptive standardization processing of the original data. This transforms the multi-dimensional original monitoring data into a new feature space where the value of the data point has been stripped of the differential influence of external operating conditions, representing only the degree of deviation from the health baseline of its respective operating condition. This achieves normalization and direct comparison of data from different operating conditions, laying the foundation for the subsequent construction of a globally unified anomaly detection model.

[0033] Figure 3 This is a flowchart illustrating the generation of a normalized parameter mapping table for a data monitoring-based wind turbine pitch power supply health diagnosis method according to an embodiment of the present invention. Figure 3 As shown, the normalized parameter mapping table generation process includes the following steps: S301, extracting data from the historical data of the pitch power supply's healthy operating cycle as a health dataset; S302, grouping the health dataset according to the operating condition label to obtain a set of health data groups; S303, for each health data group in the set of health data groups, calculating the power supply diagnostic characteristics of the same type of operating condition data within the group as normalized parameters, the power supply diagnostic characteristics including mean and standard deviation; S304, storing the normalized parameters and their corresponding operating condition labels as key-value pairs in the normalized parameter mapping table.

[0034] In steps S301 and S302, data from the healthy operating cycle of the pitch power supply is extracted from historical data as a health dataset. The health dataset is then grouped according to operating condition labels to obtain a set of health data groups. It should be understood that the core of the adaptive normalization strategy is to establish a benchmark that accurately reflects the healthy operating status of the equipment under different operating conditions. This benchmark must originate from real and pure historical data; any unfiltered data containing fault or degradation information will contaminate the accuracy of the benchmark. Therefore, in the technical solution of this application, data from the healthy operating cycle of the pitch power supply is extracted from historical data as a health dataset, and the health dataset is grouped according to operating condition labels to obtain a set of health data groups. This constructs a highly reliable, structured reference database, where each data subset uniquely corresponds to a standard health behavior pattern under a specific operating condition. This provides pure and independent sample input for subsequent calculation of the specific normalization parameters for each operating condition, ensuring that the generated normalization parameter mapping table can truly reflect the health status of the equipment in various scenarios, thereby guaranteeing the reliability and accuracy of the diagnostic model.

[0035] More specifically, in a concrete example of this application, by reviewing the operation and maintenance logs and historical alarm records of the target wind turbine, the first six months after the turbine is connected to the grid are selected as the healthy operating cycle of its pitch power supply. During this cycle, the equipment is in a brand-new state and there are no related fault records. Next, all monitoring data for these six months are completely extracted from the historical database, and the same time-series alignment and data cleaning process as in the aforementioned embodiments is performed to ensure consistency in data format and processing methods. Subsequently, this batch of aligned health cycle data is input into the defined operating condition space partitioning and data labeling process, assigning an operating condition label to each data record. Finally, based on the generated operating condition label column, a data grouping operation is performed on the health dataset, grouping data rows with the same operating condition label together to form independent data subsets, i.e., health data groups. For example, all data with label C13 are grouped into one data group, all data with label C22 are grouped into another data group, and so on, ultimately generating a set of nine independent health data groups, equal to the number of operating condition spaces.

[0036] In step S303, for each health data group in the set of health data groups, the power supply diagnostic characteristics of the same type of operating condition data within the group are calculated as normalization parameters. These power supply diagnostic characteristics include the mean and standard deviation. It should be understood that since the health data group set generated in the previous process is still a raw dataset containing a large number of data points, it cannot be directly used as a parameter for real-time data standardization. It is necessary to extract and solidify the statistical regularities within it to form a concise and quantifiable health benchmark. Therefore, in the technical solution of this application, for each health data group in the set of health data groups, the power supply diagnostic characteristics of the same type of operating condition data within the group are further calculated as normalization parameters. These power supply diagnostic characteristics include the mean and standard deviation. This condenses the distribution characteristics of the health power supply performance parameters under each operating condition scenario into two core indicators representing its central tendency: the mean and the standard deviation of its dispersion. In this way, a complete normalization parameter mapping table can be finally constructed. This table uses the operating condition label as an index and the corresponding mean and standard deviation as content, providing an efficient and accurate parameter query basis for subsequent adaptive normalization steps.

[0037] More specifically, in a concrete example of this application, the calculation process of the normalization parameter is as follows. First, the set of nine health data groups generated in the previous process is traversed. Taking the health data group with the operating condition label C13 as an example, this group contains all health data points collected under low temperature and high wind speed conditions. Within this data group, the data column containing the diagnostic feature of pitch battery voltage is located. The arithmetic mean of the health voltage under this operating condition is calculated by summing all values ​​in this column and dividing by the total number of data points. Next, the difference between each voltage value in this column and the calculated mean is calculated. These differences are squared, summed, divided by the total number of data points, and then the square root is taken to obtain the standard deviation of the health voltage under this operating condition. Using the same calculation method, the mean and standard deviation of other power supply diagnostic features such as pitch battery current within this data group are also calculated. After completing the parameter calculation for all diagnostic features within the C13 operating condition data group, continue to traverse the next healthy data group, such as C11 or C12, and repeat the above calculation process until the mean and standard deviation of all nine healthy data groups have been calculated and recorded.

[0038] In step S304, the normalized parameters and their corresponding operating condition labels are stored as key-value pairs in the normalized parameter mapping table. It should be understood that since the normalized parameters such as the mean and standard deviation calculated in the previous steps for each operating condition are still in a discrete and unstructured form at this point, without effective organization, parameter queries during the real-time diagnostic phase will face problems of low efficiency and matching difficulties. Therefore, in the technical solution of this application, the normalized parameters and their corresponding operating condition labels are further stored as key-value pairs in the normalized parameter mapping table. This systematically and structurally stores the health benchmarks for all operating conditions, creating a query dictionary with operating condition labels as direct indexes and the complete set of normalized parameters for that operating condition as content. This ultimately forms an efficient, ordered, and persistent parameter library. In the real-time diagnostic process, the corresponding normalized parameters can be instantly retrieved through operating condition labels, thereby improving the computational efficiency and engineering practicality of the entire adaptive normalization process.

[0039] More specifically, in a concrete example of this application, the construction process of the normalized parameter mapping table is as follows: First, a key-value pair data structure, such as a hash table or dictionary, is initialized in memory as the carrier for storing the mapping table. Next, all normalized parameters calculated in the preceding steps are integrated. Taking the parameter with operating condition label C13 as an example, its multiple values, including the average voltage, standard deviation of voltage, average current, and standard deviation of current, are combined into a structured object or nested dictionary. Subsequently, using the operating condition label C13 as the key and the aforementioned combined parameter object as the value, the first key-value pair entry is inserted into the hash table. Thereafter, using the same operating mode, the remaining eight operating condition labels, such as C11 and C12, and their corresponding parameter calculation results are processed sequentially, storing them one by one as key-value pairs into the hash table. After all the parameters for all operating conditions have been stored, this fully filled hash table constitutes the final normalized parameter mapping table. This mapping table can be serialized into a file format and stored on disk for loading and calling by the diagnostic program at runtime.

[0040] Figure 4 This is a flowchart illustrating the time series of a wind turbine pitch power supply health diagnosis method based on data monitoring according to an embodiment of the present invention, which extracts normalized parameters corresponding to data with operating condition labels from a normalized parameter mapping table to obtain a normalized feature vector. Figure 4As shown, step S300 includes: S310, extracting the normalization parameters corresponding to the working condition labels from the normalization parameter mapping table for the working condition data of each working condition space in the data with working condition labels; S320, performing normalization processing on various types of working condition data based on the normalization parameters to obtain normalized feature vectors; S330, regularizing multiple normalized feature vectors according to the time dimension to obtain the time series of normalized feature vectors.

[0041] In step S310, for the operating condition data of each operating condition space in the data with operating condition labels, the normalized parameters corresponding to their operating condition labels are extracted from the normalized parameter mapping table. It should be understood that each piece of data with operating condition labels acquired in real time only contains the original monitoring value and classification label, while the health benchmarks necessary for adaptive standardization processing, namely the mean and standard deviation, are stored in a separate normalized parameter mapping table. Therefore, in the technical solution of this application, the normalized parameters corresponding to their operating condition labels are further extracted from the normalized parameter mapping table for the operating condition data of each operating condition space in the data with operating condition labels. This dynamically and accurately matches and loads the health status benchmark parameters of the scenario to which the current data point to be processed belongs. This ensures that subsequent standardization calculations are based on the correct benchmark that fully corresponds to the current operating condition, thus providing the necessary and readily available input parameters for effectively removing the influence of operating conditions and achieving accurate health status assessment.

[0042] More specifically, in a concrete example of this application, firstly, during the initialization phase, the pre-built normalized parameter mapping table file is loaded from local storage and parsed into an in-memory hash table data structure for high-speed querying. Next, when a real-time data stream, after time-alignment and condition labeling, enters this processing stage, the condition label field of that data record is read. For example, the condition label for the current data point is C13. Subsequently, the string C13 is used as the query key to search within the in-memory hash table. This search operation returns the value associated with the key C13, which is a structured object containing the mean and standard deviation of all diagnostic features for that condition. For example, it includes parameters such as a mean of 545 volts and a standard deviation of 10 volts for the pitch battery voltage. Finally, this successfully extracted parameter structured object, along with the original monitoring value of the current data point, is passed to the next step of the normalization process. This extraction operation is repeated for each new real-time data point.

[0043] In step S320, various types of operating condition data are standardized based on normalization parameters to obtain normalized feature vectors. It should be understood that after parameter extraction, the real-time data points to be processed and their corresponding health baseline parameters remain two independent entities; the data itself has not changed, and its absolute value is still strongly influenced by external operating conditions. Therefore, in the technical solution of this application, various types of operating condition data are further standardized based on normalization parameters to obtain normalized feature vectors. This performs the core decoupling operation, that is, through mathematical transformation, converting the original monitoring value into a dimensionless relative value that only reflects the degree of deviation from the corresponding operating condition health baseline. This ultimately generates a pure feature vector unaffected by external environment and operating conditions, allowing subsequent anomaly detection models to make judgments in a unified and stable feature space, thereby improving the accuracy and robustness of diagnosis.

[0044] More specifically, in a specific example of this application, the normalization process for various types of operating condition data is performed based on the normalization parameter to obtain a normalized feature vector, including: subtracting the mean of the operating condition data from the mean of the operating condition interval, and then dividing the difference by the standard deviation of the operating condition interval to obtain a normalized value; and combining the normalized values ​​corresponding to various types of operating condition data into a normalized feature vector.

[0045] In other words, more specifically, the standardization process and feature vector construction are as follows: First, the real-time data points and their corresponding normalization parameters from the previous process are received. For example, the original value of the pitch battery voltage at the current data point is 535 volts, and the mean voltage corresponding to its matching operating condition label C13 is 545 volts with a standard deviation of 10 volts. Next, the pitch battery voltage is standardized by subtracting the mean of the operating condition data from the mean of its corresponding operating condition interval, and then dividing the difference by the standard deviation of the corresponding operating condition interval to obtain the normalized value. This method is also used for all other power supply diagnostic features, such as the pitch battery current, at the current data point, performing the same standardization calculation using their respective mean and standard deviation to obtain their respective normalized values. After the normalized values ​​of all diagnostic features have been calculated, the normalized values ​​corresponding to various types of operating condition data are combined into a normalized feature vector. Each element of this vector corresponds to a standardized diagnostic feature, and this vector will serve as the input for the next process.

[0046] In step S330, multiple normalized feature vectors are regularized according to the time dimension to obtain a time series of normalized feature vectors. It should be understood that since the normalized feature vectors generated in the preceding steps are instantaneous state snapshots at discrete time points, a single snapshot may not fully reflect the dynamic behavior anomalies caused by performance degradation. Many early fault characteristics are reflected in the evolution patterns of the data over time. Therefore, in the technical solution of this application, multiple normalized feature vectors are further regularized according to the time dimension to obtain a time series of normalized feature vectors, thereby combining independent instantaneous state data points into a sequence data structure with a time context. This allows for the construction of an input form suitable for subsequent deep learning models to perform time-series pattern analysis, enabling the model not only to evaluate the state at a single moment but also to learn and judge whether the evolution trajectory of the state over a period of time conforms to a healthy pattern, thereby improving the detection capability of dynamic anomalies and early minor faults.

[0047] More specifically, in a concrete example of this application, a two-dimensional data matrix structure is initialized to store the time series. The rows of this matrix correspond to the time steps, and the columns correspond to the dimensions of the normalized feature vectors. After the previous process generates a normalized feature vector at each time point, such as 14:01, this vector is appended as a new row to the end of this two-dimensional matrix. This process is executed continuously over time; feature vectors generated at 14:02, 14:03, etc., are appended as new rows sequentially, strictly maintaining their chronological order. For example, after processing three consecutive time points, the three independent feature vectors generated are then normalized into three rows of data in this matrix. This two-dimensional matrix containing multiple rows of data is the final time series of normalized feature vectors, which will be input as a whole into the subsequent autoencoder model for processing.

[0048] In step S400, the time series of the normalized feature vector is reconstructed and anomaly scored using a pre-trained autoencoder model to obtain a time series of health status indicators. It should be understood that although the time series of the feature vector after normalization eliminates interference from external operating conditions, it is still high-dimensional and complex data. A mechanism is needed to autonomously learn health patterns and quantitatively evaluate unknown patterns to achieve automated anomaly detection. Therefore, in the technical solution of this application, the time series of the normalized feature vector is further reconstructed and anomaly scored using a pre-trained autoencoder model to obtain a time series of health status indicators. This utilizes the accurate memory and reproduction capability of the autoencoder model, trained only on a health dataset, for normal patterns. By calculating the model's reconstruction error against real-time input data, the degree of difference between the current data and all known health patterns is quantified. This transforms the complex pattern matching problem into a simple numerical comparison problem. By judging whether the reconstruction error exceeds a preset threshold, a clear normal or abnormal health status indicator is generated for each time point, thereby achieving accurate and real-time monitoring and judgment of the pitch power supply's health status.

[0049] More specifically, in a concrete example of this application, the process of reconstructing and scoring a time series of normalized feature vectors using a pre-trained autoencoder model to obtain a health status indicator includes: reconstructing the normalized feature vectors using the pre-trained autoencoder model to obtain a reconstructed vector; calculating the reconstruction error between the reconstructed vector and the normalized feature vector; and comparing the reconstruction error with a preset anomaly threshold to obtain a health status indicator. In particular, calculating the reconstruction error between the reconstructed vector and the normalized feature vector includes: using a mean squared error algorithm to calculate the reconstruction error between the reconstructed vector and the normalized feature vector.

[0050] More specifically, the time series of the normalized feature vector generated in the previous process—for example, a matrix representing data from the past 60 minutes—is input into an autoencoder model already trained on a health dataset. Upon receiving this input, the model is reconstructed, and a reconstructed vector with the same dimensions as the input matrix is ​​output. Next, the reconstruction error between the reconstructed vector and the normalized feature vector is calculated. This calculation uses the mean squared error algorithm, which calculates the square of the difference between each corresponding element in the input matrix and the reconstructed matrix, sums all these squared differences, and averages them to obtain a single scalar value as the final reconstruction error. Finally, the calculated reconstruction error is compared with a preset anomaly threshold to obtain a health status indicator. For example, if the calculated reconstruction error is 0.8 and the preset anomaly threshold is 1.2, the current health status is recorded as normal because the error does not exceed the threshold; conversely, if the reconstruction error is 1.5, the current time is recorded as abnormal. This process is continuously executed as the time window slides, generating a series of health status indicators ordered by time, i.e., a time series of health status indicators.

[0051] Specifically, in step S500, the time series of health status indicators are aggregated to obtain a health index, which is expressed by the following formula:

[0052] in, The health index at the previous moment. The health index is updated at the current moment. The trainingable decay coefficient, This represents the number of times the health status flag is marked as abnormal within a predetermined time window.

[0053] It is understandable that the health status indicator time series generated in the preceding steps is essentially a series of discrete, instantaneous normal or abnormal judgments. While this form can provide real-time alerts, it is relatively volatile and lacks a macroscopic representation of the long-term performance evolution trend of the equipment, which is not conducive to maintenance personnel assessing the cumulative damage and overall health status of components. Therefore, in the technical solution of this application, the time series of health status indicators are further aggregated to obtain a health index, thereby transforming discrete abnormal events into a continuously changing health indicator that can quantitatively represent the long-term cumulative degradation of the pitch power supply. In this way, a smooth, intuitive, and physically meaningful health index curve can be output. This curve not only reflects the current health level of the equipment but also reveals the rate and trend of its performance degradation, providing a key decision-making basis for implementing predictive maintenance strategies and optimizing spare parts management.

[0054] In summary, the data monitoring-based health diagnosis method for wind turbine pitch power supplies according to embodiments of the present invention is explained. First, the complex operating state of the wind turbine is divided into multiple independent operating condition spaces based on key external variables such as ambient temperature and wind speed. For each operating condition space, a dedicated health status benchmark model is established using historical health data. During real-time diagnosis, the monitoring data is first assigned to its corresponding operating condition space, and adaptive normalization is performed using the health benchmark of that space. This step effectively isolates normal data fluctuations caused by changes in environment and operating conditions, obtaining a feature vector that only reflects the inherent health status of the pitch power supply. Subsequently, this purified feature vector is input into a reconstruction model that has only learned the health pattern. If the model cannot accurately reconstruct this vector, resulting in a large reconstruction error, it indicates that the power supply has deviated from its true normal performance under that specific operating condition. This two-step strategy can accurately identify early degradation characteristics from a strong interference background, thereby improving the accuracy and reliability of diagnosis and reducing the false alarm rate.

[0055] Furthermore, a data monitoring-based health diagnostic system for wind turbine pitch power supplies is also provided.

[0056] Figure 5 This is a block diagram of a wind turbine pitch power supply health diagnosis system based on data monitoring, according to an embodiment of the present invention. Figure 5 As shown, the wind turbine pitch power supply health diagnosis system 100 based on data monitoring according to an embodiment of the present invention includes: a time-series alignment processing module 110, used to perform time-series alignment processing on the acquired real-time monitoring data of the wind turbine to obtain aligned time-series data; a data acquisition module 120 with operating condition labels, used to divide the aligned time-series data into operating condition spaces and label the data to obtain data with operating condition labels; a feature vector time-series extraction module 130, used to extract the normalized parameters corresponding to the data with operating condition labels from the normalized parameter mapping table to obtain the time series of normalized feature vectors; a health status indicator time-series acquisition module 140, used to reconstruct and anomaly score the time series of normalized feature vectors through a pre-trained autoencoder model to obtain the time series of health status indicators; and a power supply health index trend chart generation module 150, used to perform aggregation operations on the time series of health status indicators to obtain a health index, and generate a wind turbine pitch power supply health index trend chart based on the health index.

[0057] As described above, the wind turbine pitch power supply health diagnosis system based on data monitoring according to embodiments of the present invention can be implemented in a collaborative working environment between the edge computing unit at the wind turbine site and the back-end server platform of the wind farm control center, such as an industrial control computer deployed inside the wind turbine nacelle or tower base, and communicate with the wind farm's SCADA monitoring system or production operation management platform. In one possible implementation, the wind turbine pitch power supply health diagnosis system based on data monitoring according to embodiments of the present invention can be integrated into the SCADA monitoring system or production operation management platform as a software module or hardware module. For example, on the back-end server side, the system can be a background service program running in its operating system, including functions such as data timing alignment, operating condition space partitioning, adaptive normalization, and health index aggregation, or it can be an independent intelligent diagnosis and health management engine developed for the wind farm; of course, the autoencoder reconstruction and anomaly scoring model used to perform real-time diagnosis in the system can also be embedded in dedicated industrial computing hardware, such as an embedded AI chip or dedicated accelerator card on the edge computing gateway close to the data acquisition source, to accelerate the real-time identification and alarm response process of early minor faults in the pitch power supply.

[0058] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for health diagnosis of wind turbine pitch power supply based on data monitoring, characterized in that, include: The acquired real-time monitoring data of the wind turbines is time-series aligned to obtain aligned time-series data; The aligned time series data is divided into working condition spaces and labeled to obtain data with working condition labels; Extract the normalized parameters corresponding to the data with working condition labels from the normalized parameter mapping table to obtain the time series of normalized feature vectors; The time series of normalized feature vectors are reconstructed and anomaly scored using a pre-trained autoencoder model to obtain the time series of health status indicators. The time series of health status indicators are aggregated to obtain a health index, and a trend chart of the health index of wind turbine pitch power supply is generated based on the health index.

2. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to claim 1, characterized in that, The acquired monitoring data of the wind turbines are time-series aligned to obtain aligned time-series data. This includes: resampling and cleaning the monitoring data of the wind turbines based on timestamps to obtain aligned time-series data. The monitoring data of the wind turbines includes ambient temperature, wind speed, blade motor current, pitch battery voltage, and pitch battery current.

3. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to claim 1, characterized in that, The aligned time series data is divided into working condition spaces and labeled to obtain data with working condition labels, including: The various operating condition data in the aligned time series data are matched with the operating condition space to divide each operating condition data into the corresponding operating condition space. Based on the working condition space after data partitioning, each working condition data is labeled with a working condition tag to obtain data with working condition tags.

4. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to any one of claims 1 to 3, characterized in that, The steps for generating the normalized parameter mapping table include: Extract data from the historical data of the healthy operating cycle of the pitch power supply as a health dataset; The health dataset is grouped according to the working condition label to obtain a set of health data groups; For each health data group in the set of health data groups, calculate the power supply diagnostic characteristics of the same type of operating condition data within the group as a normalization parameter. The power supply diagnostic characteristics include the mean and standard deviation. The normalized parameters and their corresponding operating condition labels are stored as key-value pairs in the normalized parameter mapping table.

5. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to any one of claims 1 to 3, characterized in that, The time series of normalized feature vectors is obtained by extracting the normalized parameters corresponding to the data with working condition labels from the normalized parameter mapping table, including: For the working condition data of each working condition space in the data with working condition labels, extract the normalized parameters corresponding to their working condition labels from the normalized parameter mapping table; Normalized feature vectors are obtained by standardizing various types of working condition data based on normalization parameters. Multiple normalized feature vectors are regularized according to the time dimension to obtain a time series of normalized feature vectors.

6. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to claim 5, characterized in that, Normalized feature vectors are obtained by standardizing various types of working condition data based on normalization parameters, including: The normalized value is obtained by subtracting the mean of the operating condition data from the mean of the operating condition interval, and then dividing the difference by the standard deviation of the operating condition interval. The normalized values ​​corresponding to various types of working condition data are combined into a normalized feature vector.

7. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to any one of claims 1 to 3, characterized in that, The time series of normalized feature vectors are reconstructed and anomaly scored using a pre-trained autoencoder model to obtain time series indicators of health status, including: The normalized feature vectors are reconstructed using a pre-trained autoencoder model to obtain the reconstructed vectors. Calculate the reconstruction error between the reconstructed vector and the normalized eigenvector; The reconstruction error is compared with a preset anomaly threshold to obtain a health status indicator.

8. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to claim 7, characterized in that, The reconstruction error between the reconstructed vector and the normalized eigenvector is calculated, including: using the mean squared error algorithm to calculate the reconstruction error between the reconstructed vector and the normalized eigenvector.

9. The method for health diagnosis of wind turbine pitch power supply based on data monitoring according to any one of claims 1 to 3, characterized in that, The health index is obtained by aggregating the time series of health status indicators, including: aggregating the time series of health status indicators using the following aggregation formula, where the aggregation formula is: in, The health index at the previous moment. The health index is updated at the current moment. The trainingable decay coefficient, This represents the number of times the health status flag is marked as abnormal within a predetermined time window.

10. A health diagnosis system for wind turbine pitch power supply based on data monitoring, characterized in that, include: The time-series alignment processing module is used to perform time-series alignment processing on the acquired real-time monitoring data of wind turbines to obtain aligned time-series data; The working condition labeled data acquisition module is used to divide the aligned time series data into working condition spaces and label the data to obtain working condition labeled data; The feature vector time series extraction module is used to extract the normalized parameters corresponding to the data with working condition labels from the normalized parameter mapping table to obtain the time series of normalized feature vectors. The health status indicator time series acquisition module is used to reconstruct and anomaly score the time series of normalized feature vectors through a pre-trained autoencoder model to obtain the health status indicator time series. The power health index trend chart generation module is used to aggregate and calculate the time series of health status indicators to obtain the health index, and generate a wind turbine pitch power health index trend chart based on the health index.

Images