A wind turbine data preprocessing method and device, a terminal device and a medium

Abstract

The application provides a wind turbine data preprocessing method and device, terminal equipment and medium, comprising: in response to a diagnosis instruction of a wind turbine fault diagnosis model, obtaining target operation data associated with the diagnosis instruction; inputting CMS data into a pre-trained integrated convolutional neural network model for anomaly detection classification, identifying that the CMS data belongs to normal signals or abnormal types; performing corresponding data repair processing on the CMS data according to the identified abnormal type to obtain repaired CMS data; and performing outlier elimination and missing value filling on SCADA data to obtain repaired SCADA data. The application can increase the effective sample size of the downstream fault diagnosis model, thereby improving the accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method, apparatus, terminal equipment, and medium for preprocessing data of wind turbine generators. Background Technology

[0002] Wind turbines generate massive amounts of operational data during operation, including SCADA (Supervisory Control and Data Acquisition) data such as temperature, pressure, power, and speed, as well as CMS (Condition Monitoring System) data reflecting the vibration of various components. This data contains crucial information reflecting the turbine's operating status and is essential for evaluating wind turbine performance, monitoring its health, and diagnosing and predicting faults. By processing and analyzing wind turbine operational data, real-time monitoring, fault warnings, and performance optimization can be achieved, effectively improving the power plant's operational efficiency and reliability while significantly reducing maintenance costs.

[0003] For the large amount of raw data generated during wind turbine operation, systematic and scientific preprocessing to obtain high-quality raw data can significantly improve the accuracy of model analysis. However, current technologies acquire wind turbine operating data related to fault diagnosis, process single data sources and single data anomalies, and then input them into the fault diagnosis model for diagnosis or prediction. This data processing method fails to comprehensively address the complexity and diversity of wind turbine operating data and does not adequately remove anomalies, resulting in low accuracy in fault diagnosis. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, device, terminal equipment and medium for preprocessing wind turbine data, so as to increase the effective sample size of downstream fault diagnosis models and improve the accuracy of fault diagnosis.

[0005] In a first aspect, the present invention provides a method for preprocessing wind turbine data, the method comprising the following steps:

[0006] In response to the diagnostic commands from the wind turbine fault diagnosis model, target operating data associated with the diagnostic commands are obtained from the wind power operation database; the target operating data includes SCADA data reflecting the unit's operating conditions and CMS data reflecting component vibration.

[0007] CMS data is input into a pre-trained ensemble convolutional neural network model for anomaly detection and classification, identifying whether the CMS data belongs to normal signals or anomalies. The ensemble convolutional neural network model adopts a Bagging ensemble strategy and is composed of multiple independent convolutional neural network sub-models. Anomaly types include signal asymmetry, zero-point drift, and data jump anomalies.

[0008] Based on the identified anomaly type, corresponding data repair processing is performed on the CMS data to obtain the repaired CMS data; the data repair processing includes different mathematical transformations or interpolation methods for signal asymmetry, zero-point drift and data jump anomalies respectively;

[0009] Outlier removal and missing value filling are performed on the SCADA data to obtain repaired SCADA data; the repaired CMS data and the repaired SCADA data are used as target input data to input the fault diagnosis model.

[0010] Optionally, the target operational data associated with the diagnostic command can be obtained from the wind power operation database, including:

[0011] The diagnostic instructions of the wind turbine fault diagnosis model are analyzed to determine the target component to be diagnosed and the corresponding fault type.

[0012] A data index is established based on the fault type. Candidate SCADA data associated with the target component is retrieved from the wind power SCADA database, and candidate CMS data associated with the target component is retrieved from the CMS database.

[0013] Extract wind speed and active power values ​​from candidate SCADA data, and identify the time interval in which the wind turbine is generating electricity based on the wind speed and active power values.

[0014] Extract data segments from candidate SCADA data and candidate CMS data that fall within the time interval, and combine them as the target running data.

[0015] Optionally, the CMS data is input into a pre-trained ensemble convolutional neural network model for anomaly detection and classification, identifying whether the CMS data belongs to a normal signal or an anomaly type, including:

[0016] A raw dataset containing various vibration anomaly types was created and labeled.

[0017] The Bootstrap self-sampling method was used to generate multiple distinct sub-training sets from the original dataset;

[0018] Multiple convolutional neural network sub-models with identical structures are trained in parallel using multiple sub-training sets;

[0019] In the prediction phase, the CMS data to be detected is input into multiple convolutional neural network sub-models, and the outputs of all convolutional neural network sub-models are integrated by voting or averaging to determine the final anomaly classification result.

[0020] Optionally, the convolutional layers in the integrated convolutional neural network model extract the physical features of the vibration signals in the CMS data through convolutional kernels. The feature extraction process satisfies the following formula:

[0021]

[0022] in, Indicates the first The first convolutional layer Each output feature map represents the deep features of the extracted vibration signal; Indicates the first The first convolutional layer Vibration data segment corresponding to each receptive field; Indicates the first The first convolutional layer The weight matrix of each convolution kernel corresponds to a specific vibration mode filter; This is the bias value; Represents a nonlinear activation function; This represents the convolution operation.

[0023] Optionally, based on the identified anomaly type, perform corresponding data repair processing on the CMS data, specifically including:

[0024] If the identified anomaly type is signal asymmetry anomaly, then zero-mean processing is performed on the CMS data to eliminate DC component deviation;

[0025] If the identified anomaly type is zero-point drift anomaly, then detrending processing is performed on the CMS data to eliminate baseline drift that changes slowly over time;

[0026] If the identified anomaly type is a data jump anomaly, then the jump point in the CMS data is identified, the jump point is marked as an invalid value, and the vibration value of the point is reconstructed using a cubic spline interpolation algorithm based on the normal data before and after the jump point, so that it is restored to normal data that conforms to physical continuity.

[0027] Optionally, outlier removal and missing value imputation can be performed on the SCADA data, specifically including:

[0028] Identify outliers in SCADA data caused by communication interruptions or sensor malfunctions; outliers include NULL values, NA values, constant values, and consecutive 0 values.

[0029] Interpolate outliers;

[0030] For temperature data in SCADA data, a method of aggregating and aligning the timelines of data collected from all units in the field is adopted, using the operating parameters of units of the same model on the same timeline to fill in the missing temperature parameters.

[0031] Optionally, the method also includes triggering a sensor fault alarm when the proportion of abnormal data exceeds a preset threshold, including:

[0032] After processing the CMS data, the ratio of abnormal data to normal data is calculated.

[0033] If the ratio is greater than the preset threshold, it is determined that the data acquisition sensor has a physical fault, the input of data to the fault diagnosis model is terminated, and a sensor maintenance alarm signal is output.

[0034] If the proportion is less than or equal to the preset threshold, the repaired data will be input into the fault diagnosis model for fault diagnosis of unit components.

[0035] In a second aspect, the present invention provides a wind turbine data preprocessing device, comprising:

[0036] The data acquisition module is used to respond to the diagnostic commands of the wind turbine fault diagnosis model and obtain the target operating data associated with the diagnostic commands from the wind power operation database. The target operating data includes SCADA data reflecting the unit's operating conditions and CMS data reflecting the vibration of components.

[0037] The anomaly detection module is used to input CMS data into a pre-trained ensemble convolutional neural network model for anomaly detection and classification, identifying whether the CMS data belongs to normal signals or anomalies. The ensemble convolutional neural network model adopts a Bagging ensemble strategy and is composed of multiple independent convolutional neural network sub-models. Anomaly types include signal asymmetry, zero-point drift, and data jump anomalies.

[0038] The first data repair module is used to perform corresponding data repair processing on the CMS data according to the identified anomaly type to obtain the repaired CMS data; the data repair processing includes using different mathematical transformations or interpolation methods for signal asymmetry, zero-point drift and data jump anomalies respectively;

[0039] The second data repair module is used to remove outliers and fill in missing values ​​in the SCADA data to obtain repaired SCADA data. The repaired CMS data and the repaired SCADA data are used as target input data to input the fault diagnosis model.

[0040] Thirdly, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0041] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0042] The present invention has at least the following beneficial effects:

[0043] By integrating SCADA operating condition data and CMS vibration data, the shortcomings of insufficient information from a single data source are overcome. An integrated convolutional neural network based on the Bagging strategy is used to achieve accurate identification and robust classification of complex wind power data anomalies. Furthermore, a novel mathematical repair method is adopted for specific anomalies such as signal asymmetry, drift, and jumps, replacing the traditional direct rejection method. This transforms failed data into usable data, significantly improving data utilization and integrity. It provides high-quality standard input for downstream fault diagnosis models, thereby greatly improving the accuracy and reliability of wind turbine fault diagnosis. Attached Figure Description

[0044] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0045] Figure 1 This is a flowchart of a wind turbine data preprocessing method according to one embodiment of this application;

[0046] Figure 2 This is a time-domain waveform diagram of the normal vibration signal of a wind turbine generator in one embodiment of this application;

[0047] Figure 3 This is a time-domain waveform diagram of the positive and negative imbalance of the vibration signal of a wind turbine generator in one embodiment of this application;

[0048] Figure 4 This is a time-domain waveform diagram of the vibration signal data jump of a wind turbine generator in one embodiment of this application;

[0049] Figure 5 This is a time-domain waveform diagram of the zero-point drift of the vibration signal of a wind turbine generator in one embodiment of this application;

[0050] Figure 6 This is a structural diagram of a wind turbine data preprocessing device according to one embodiment of this application;

[0051] Figure 7This is a structural diagram of a terminal device in one embodiment of this application. Detailed Implementation

[0052] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. 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 embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0054] Existing technologies for wind turbine data preprocessing typically involve independent cleaning and analysis of a single data source (such as SCADA or CMS data only), with relatively simplistic processing methods. Common statistical methods are often used to directly discard data deviating from the normal range as outliers. The main drawbacks of this approach are: first, it fails to integrate features from multiple data sources, leading to insufficient information utilization and an inability to fully reflect the true state of the turbine under complex operating conditions; second, the traditional "one-size-fits-all" rejection strategy ignores the specific physical forms of data anomalies (such as zero-point drift and signal jumps), resulting in the direct discarding of a large amount of data that essentially contains valid information but is contaminated only by slight sensor distortion, leading to low data utilization; finally, traditional time-series classification algorithms suffer from long computation times and weak generalization ability when dealing with massive amounts of wind power data accompanied by strong noise, making it difficult to remove interference while retaining valid features, ultimately resulting in insufficient accuracy and robustness of downstream fault diagnosis models. To address these issues, this application proposes a wind turbine data preprocessing method based on an ensemble convolutional neural network (Ensemble CNN). This solution first establishes a cross-data source correlation mechanism, synchronously acquiring SCADA operating condition data and CMS vibration data based on fault diagnosis instructions. The core lies in constructing a convolutional neural network model based on a Bagging ensemble strategy. This model uses multiple sub-models trained in parallel to perform refined anomaly type identification on the CMS data (e.g., distinguishing signal asymmetry, zero-point drift, data jumps, etc.). Then, based on the identified specific anomaly type, it matches and executes targeted mathematical repair strategies (e.g., detrending processing, cubic spline interpolation, etc.) and performs field-group collaborative completion on the SCADA data, ultimately outputting high-quality repair data or triggering an alarm when the sensor completely fails. This application effectively overcomes the shortcomings of existing technologies through a dual mechanism of "multi-source fusion + intelligent repair." First, by leveraging the powerful feature extraction and generalization capabilities of the integrated CNN model, the problem of traditional algorithms' inability to accurately distinguish between "data noise" and "physical faults" is solved, significantly improving the accuracy of anomaly identification. Second, this application changes the previous approach of simply "removing" anomalous data, instead performing "mathematical repair" on specific types of anomalous data. For example, it repairs abrupt data through interpolation and drifting data through detrending, which significantly improves data utilization and transforms previously unusable dirty data into high-quality samples that can be used for model training. Finally, by fusing SCADA and CMS data, a comprehensive assessment of the unit's status is achieved, avoiding diagnostic blind spots caused by a single data source, thereby fundamentally improving the timeliness of wind turbine fault warnings and the reliability of diagnostic results.

[0055] Example 1

[0056] This embodiment details the wind turbine data preprocessing method provided by the present invention. This method is executed by a data preprocessing system deployed on an edge computing server or a central monitoring system within a wind farm. Figure 1 As shown, the method specifically includes steps 11 to 14.

[0057] Step 11: In response to the diagnostic command from the wind turbine fault diagnosis model, retrieve the target operating data associated with the diagnostic command from the wind power operation database.

[0058] In this embodiment of the invention, the target operating data includes SCADA data reflecting the unit's operating conditions and CMS data reflecting the vibration of components.

[0059] Specifically, step 11 includes steps 11.1 to 11.4.

[0060] Step 11.1: Analyze the diagnostic instructions of the wind turbine fault diagnosis model to determine the target component to be diagnosed and the corresponding fault type.

[0061] In one feasible implementation, assume that the wind farm central monitoring system sends a fault diagnosis model instruction for "generator drive end bearing wear detection" to the data preprocessing system for wind turbine unit A. The data preprocessing system first parses the received "generator drive end bearing wear detection" instruction, determines that the target component is "generator bearing", and the fault type is "abnormal vibration".

[0062] Step 11.2: Establish a data index based on the fault type, retrieve candidate SCADA data associated with the target component from the wind power SCADA database, and retrieve candidate CMS data associated with the target component from the CMS database.

[0063] In one feasible implementation, the data preprocessing system maintains a component data source mapping table in advance. This table is queried to retrieve SCADA and CMS parameters associated with abnormal generator bearing vibration. The SCADA parameters include the temperature of the generator drive-end bearing, the temperature of the generator non-drive-end bearing, the nacelle wind speed, and the generator active power. The CMS parameters include the radial vibration acceleration waveforms of the generator drive-end bearing and the generator non-drive-end bearing. Based on the time window specified in the instruction, the data preprocessing system sends a query request to the wind power SCADA historical database to retrieve the time-series data of the above four SCADA parameters within the specified time window, with a time resolution typically of 1 or 2 seconds per point. Simultaneously, a request is sent to the CMS waveform database to retrieve the high-frequency waveform data of the two vibration acceleration channels within the same time window.

[0064] Step 11.3: Extract wind speed and active power values ​​from candidate SCADA data, and identify the time interval in which the wind turbine is generating electricity based on the wind speed and active power values.

[0065] Step 11.4: Extract data segments from the candidate SCADA data and candidate CMS data that fall within the time interval, and combine them as the target running data.

[0066] To ensure the effectiveness of the diagnosis, data from units that are shut down or in standby mode must be excluded. The data preprocessing system filters data based on a preset logical threshold, which is set as: active power. 0kW and wind speed 3 m / s. In practical implementation, if SCADA data shows a wind speed of 2.5 m / s and an active power of -15 kW (self-consuming state), the unit is determined to be in standby mode. The CMS vibration data collected at this time mainly represents environmental noise and has no diagnostic value. Therefore, the vibration waveform and temperature data corresponding to this time period are discarded. If SCADA data shows a wind speed maintained at around 8.5 m / s and an active power stable at around 1200 kW, the unit is determined to be in active power generation mode. The vibration data at this time accurately reflects the operating conditions under bearing load, and the vibration waveform and temperature data corresponding to this time period are retained. The system packages the SCADA temperature data and CMS vibration waveform data within the selected and retained time period as target operating data and transmits it to the subsequent integrated convolutional neural network module for anomaly detection.

[0067] It is worth mentioning that by parsing diagnostic commands to accurately associate them with physical components (such as generators) and using wind speed and power parameters to filter out valid data under "generating state," invalid data interference from non-steady-state conditions such as shutdown and standby can be automatically eliminated. This not only reduces the amount of data processing and improves system operating efficiency, but also ensures the correlation between input data and fault characteristics from the perspective of physical operating conditions.

[0068] Step 12: Input the CMS data into the pre-trained ensemble convolutional neural network model for anomaly detection and classification, and identify whether the CMS data belongs to normal signals or anomalies.

[0069] In this embodiment of the invention, the integrated convolutional neural network model employs a Bagging ensemble strategy, consisting of multiple independent convolutional neural network sub-models. The goal of this step is to classify CMS vibration waveform data into "normal" or specific abnormal types (such as signal asymmetry, zero-point drift, and data jumps).

[0070] In practice, step 12 includes steps 12.1 to 12.4.

[0071] Step 12.1: Establish a raw dataset containing various types of vibration anomalies and label it.

[0072] In one feasible implementation, historical CMS vibration data is collected and initially labeled by experts or based on rules as: normal, signal asymmetry, zero-point drift, and data jump. Its time-domain waveform is shown below. Figures 2 to 5 As shown. From Figures 2 to 5 It can be seen that the difference between fault data and normal data is often not in common, single aspects such as the magnitude of peak values ​​or deviations. For example, in imbalanced abnormal data, the amplitude of the vibration signal is distributed in both positive and negative directions, but the distribution is not uniform, showing obvious imbalance. Data jump anomalies usually refer to significant fluctuations or abrupt changes in vibration data within a short period of time, typically caused by instantaneous fluctuations in the sensor or transportation system. Zero-point drift refers to the phenomenon where the output signal of a sensor or measurement system changes slowly over time in the absence of an input signal. This change may be caused by factors such as temperature changes, power fluctuations, and sensor aging. Zero-point drift can seriously affect the accuracy and reliability of measurements, leading to deviations in measurement results. Therefore, identifying and correcting zero-point drift is crucial for ensuring the performance of the measurement system.

[0073] Step 12.2: Use the Bootstrap self-sampling method to generate multiple different sub-training sets from the original dataset.

[0074] In one feasible implementation, in order to construct the ensemble model, the number of sub-models in the ensemble model is set to 1. (For example Assume the total number of samples in the original training dataset is... Regarding the first Sub-models ( The Bootstrap bootstrapping method was used to randomly sample data with replacement from the original dataset. The sub-model's dedicated training set consists of 10 samples. Because sampling with replacement is performed, each subset of the training set... The dataset contains approximately 63.2% unique samples from the original dataset, with the remaining approximately 36.8% of samples not selected. This randomness in data distribution ensures the diversity of the five sub-models on the training data, reducing model variance.

[0075] Step 12.3: Train multiple convolutional neural network sub-models with the same structure in parallel using multiple sub-training sets.

[0076] In one feasible implementation, each sub-model employs the same network structure, specifically designed for one-dimensional vibration time-series signals. The specific structure is as follows:

[0077] Input layer: Receives time series vectors with a length of 2048×1.

[0078] The first convolutional layer has a kernel size of 15×1, 16 kernels, a stride of 1, and an activation function of ReLU. It is used to extract low-frequency vibration features within a wide receptive field. In this embodiment of the invention, the convolutional layers in the integrated convolutional neural network model extract the physical features of vibration signals from CMS data through convolutional kernels. The feature extraction process satisfies the following formula:

[0079]

[0080] in, Indicates the first The first convolutional layer Each output feature map represents the deep features of the extracted vibration signal; Indicates the first The first convolutional layer Vibration data segment corresponding to each receptive field; Indicates the first The first convolutional layer The weight matrix of each convolution kernel corresponds to a specific vibration mode filter; This is the bias value; Represents a nonlinear activation function; This represents a convolution operation. The convolution kernel automatically extracts deep nonlinear features from vibration signals, replacing the tedious process of traditional manual feature extraction (such as peak value and kurtosis). This end-to-end feature extraction method can capture subtle fault signs that are difficult for the human eye to detect, giving the model stronger analytical capabilities when dealing with high-dimensional, highly dynamic wind power vibration data.

[0081] First pooling layer: Max Pooling, window size 2, stride 2.

[0082] The second convolutional layer has a kernel size of 3×1, a kernel count of 32, a stride of 1, and uses the ReLU activation function. It is used to extract high-frequency detail features.

[0083] Second pooling layer: max pooling, window size 2, stride 2.

[0084] Fully connected layer: Flatten the multidimensional features and connect them to a hidden layer containing 128 neurons (dropout rate is set to 0.5 to prevent overfitting).

[0085] Output layer: Contains 5 neurons, corresponding to 5 classification categories, and uses the Softmax function to output the probability value of belonging to each category.

[0086] Step 12.4: In the prediction phase, the CMS data to be detected is input into multiple convolutional neural network sub-models. The outputs of all convolutional neural network sub-models are integrated by voting or averaging to determine the final anomaly classification result.

[0087] In one feasible implementation, during the model deployment phase, the CMS vibration data to be detected is input into four pre-trained independent sub-models, and a soft voting mechanism is used to obtain the final result. Specifically, the sample to be tested is input into the first to fourth sub-models, and each sub-model outputs a probability vector. ,in, Indicates the first The model determines that the sample belongs to the first... The probability of a class. These correspond to normal, signal asymmetry, zero-point drift, and data jump, respectively. Then, the average vector of the output probabilities of all sub-models is calculated. Finally, take The index corresponding to the element with the highest probability is used as the final predicted category.

[0088] It is worth mentioning that by using Bootstrap sampling and parallel training of multiple sub-models for voting, the problem of overfitting or getting trapped in local optima by a single deep learning model when facing the complex noise environment of wind power sites is effectively overcome. This ensemble strategy significantly improves the model's generalization ability and stability in identifying abnormal data, and reduces the false negative rate and false positive rate.

[0089] Step 13: Perform corresponding data repair processing on the CMS data according to the identified anomaly type to obtain the repaired CMS data.

[0090] The data repair process includes using different mathematical transformations or interpolation methods to address signal asymmetry, zero-point drift, and data jump anomalies.

[0091] In practical implementation, the process for correcting signal asymmetry includes: when the model identifies a DC bias in the signal, i.e., the vibration waveform is shifted upwards or downwards, resulting in uneven distribution of amplitude on the positive and negative half-axis, zero-mean normalization is applied. First, the original vibration sequence is calculated. arithmetic mean Each data point in the original sequence Subtract the arithmetic mean The repaired data points were obtained. This eliminates the constant DC component caused by sensor zero-point calibration error or data acquisition card voltage bias, allowing the vibration signal to oscillate around the zero axis again, restoring its physical authenticity and facilitating subsequent calculation of effective values ​​and kurtosis.

[0092] The process for correcting zero-point drift includes: polynomial fitting detrending. First, the least squares method is used to process the original sequence. Perform low-order polynomial fitting (usually linear or quadratic polynomial) to fit the trend term. , Indicates the sampling time point. and This is obtained by minimizing the sum of squared errors. Then, the fitted trend term is subtracted from the original sequence to obtain the detrended sequence. . This process eliminates low-frequency interference components superimposed on the vibration signal, retains high-frequency impact components that reflect equipment malfunction characteristics, and prevents spectral energy leakage caused by drift.

[0093] The process for correcting data jump anomalies includes outlier removal and cubic spline interpolation. Specifically, the 3-sigma criterion or the first-order difference thresholding method is used to locate jump points. The difference between adjacent points is calculated. , .like If the fluctuation exceeds the threshold of the normal fluctuation range, then mark that location and the subsequently affected areas. Each point is an invalid point. Using the normal data points before and after the invalid point interval as known nodes, a piecewise cubic spline function is constructed. For each segment interval , The function must satisfy the following conditions: the function values ​​are equal at the nodes, the first derivative is continuous, and the second derivative is continuous. (Continuous). Substitute the time coordinates corresponding to the invalid points into the constructed spline function. Calculate the interpolation Replace invalid points in the original sequence with this value to obtain the repaired sequence. Compared to linear interpolation, cubic spline interpolation can maintain the smoothness and continuity of the vibration signal, restore the waveform change trend that may be lost during the transition to the greatest extent, and avoid introducing artificial high-frequency noise.

[0094] Step 14: Remove outliers and fill in missing values ​​in the SCADA data to obtain the repaired SCADA data.

[0095] In this embodiment of the invention, step 14 includes steps 14.1 to 14.3.

[0096] Step 14.1: Identify abnormal values ​​in SCADA data caused by communication interruption or sensor failure.

[0097] Outliers include NULL values, NA values, constant values, and consecutive 0 values. Identified outlier points are marked as missing.

[0098] Step 14.2: Interpolate outliers.

[0099] For common parameters (such as power and wind speed), linear interpolation is used for padding. For missing data edges, forward padding or backward padding is used.

[0100] Step 14.3: For the temperature data in the SCADA data, the method of aggregating and aligning the timelines of data collected from all units in the field is adopted, and the missing temperature parameters are filled in by the operating parameters of the same type of units on the same timeline.

[0101] Due to occasional malfunctions of the generator bearing temperature sensor, long-term data gaps occur, leading to distortion from simple interpolation. Therefore, this embodiment of the invention employs a cross-sectional comparison method across all wind turbine units. Specifically, this involves collecting operational data from normal wind turbine units within the same wind farm that are of the same model and geographically adjacent (in the same wind flow field) to the faulty unit during the same time period. The normal units are used as reference units, and their average value is used as the filler value for the faulty unit. This method leverages the clustering nature and similar operating conditions of wind turbine units to provide a physically reasonable and statistically reliable estimate of the missing temperature data. It is more accurate than time-series interpolation for a single unit, significantly improving the usability of temperature parameters in fault diagnosis (such as overheat warning).

[0102] Specifically, the repaired CMS waveform data and the cleaned and filled SCADA time series data are packaged by timestamp to form a high-quality, multi-source synchronized "target input data" which is then transmitted to the downstream wind turbine fault diagnosis model.

[0103] In one feasible implementation, the method further includes triggering a sensor fault alarm when the proportion of abnormal data exceeds a preset threshold; including steps A to C.

[0104] Step A: After processing the CMS data, calculate the ratio of abnormal data to normal data.

[0105] Step B: If the proportion exceeds a preset threshold, it is determined that the data acquisition sensor has suffered a physical fault. Data input to the fault diagnosis model is stopped, and a sensor maintenance alarm signal is output. This situation is highly likely not due to data anomalies, but rather a physical fault in the vibration sensor itself (such as looseness, damage, or broken wiring). In this case, data transmission to the fault diagnosis model is immediately stopped to avoid false alarms. Simultaneously, a higher-level "sensor maintenance alarm" is triggered, notifying maintenance personnel to inspect the specific vibration sensor of the specific unit.

[0106] Step C: If the proportion is less than or equal to the preset threshold, the repaired data is input into the fault diagnosis model to diagnose the faults of the unit components.

[0107] This mechanism enables intelligent inference from "data anomaly" to "equipment failure," upgrading the preprocessing system from a passive cleaning tool into an intelligent node with basic equipment health assessment capabilities. This prevents a chain of misjudgments caused by sensor hardware failures and improves the reliability and credibility of the entire monitoring system.

[0108] It should be noted that the core innovation of this application lies in proposing a refined preprocessing framework for multi-source wind turbine data based on a Bagging ensemble convolutional neural network (Ensemble CNN). Unlike conventional simple data cleaning, this application constructs a classifier specifically for "sensor anomaly patterns" in CMS vibration data. It utilizes an ensemble deep learning model to subdivide data anomalies into specific physical types such as signal asymmetry, zero-point drift, and data jumps. Combined with SCADA operating condition data, it performs specific mathematical repair techniques (such as detrending and cubic spline interpolation) for different types of anomalies, rather than simply removing data.

[0109] Traditional methods typically employ a single data source or are based on statistical thresholds (such as 3). The crude cleaning approach (based on existing standards) can easily misjudge sensor malfunctions (such as drift) as equipment malfunctions, or directly delete abnormal data caused by sensor jitter, resulting in low data utilization and a high false alarm rate. The innovation of this application lies in upgrading "data cleaning" from a passive filtering step to an active "feature recognition and repair" process. By introducing an ensemble learning strategy, it overcomes the problems of weak generalization ability and overfitting of single neural networks, accurately correcting sensor-introduced errors while preserving the physical characteristics of the data, significantly improving the data quality and diagnostic accuracy of subsequent fault diagnosis models. Those skilled in the art typically tend to directly apply deep learning algorithms to the final "fault diagnosis" stage, regarding preprocessing as a general signal processing step (such as filtering and noise reduction), and rarely consider introducing a complex ensemble deep learning model in the preprocessing stage to specifically identify the "ill-conditioned features of the data itself." Furthermore, existing technologies mostly focus on how to eliminate anomalies, while the approach of this application is to identify the physical causes of anomalies and perform targeted mathematical repairs. This "targeted repair strategy based on deep classification" is not a simple superposition of common knowledge, but requires a deep understanding of the characteristics of wind turbine data to propose, thus possessing outstanding substantive features.

[0110] Example 2

[0111] like Figure 6 As shown, the present invention discloses a wind turbine data preprocessing device, the device 600 comprising:

[0112] The data acquisition module 601 is used to respond to the diagnostic commands of the wind turbine fault diagnosis model and obtain the target operating data associated with the diagnostic commands from the wind power operation database; the target operating data includes SCADA data reflecting the operating conditions of the unit and CMS data reflecting the vibration of the components;

[0113] The anomaly detection module 602 is used to input CMS data into a pre-trained ensemble convolutional neural network model for anomaly detection and classification, and to identify whether the CMS data belongs to normal signals or anomaly types. The ensemble convolutional neural network model adopts the Bagging ensemble strategy and is composed of multiple independent convolutional neural network sub-models. Anomaly types include signal asymmetry, zero-point drift, and data jump anomalies.

[0114] The first data repair module 603 is used to perform corresponding data repair processing on the CMS data according to the identified anomaly type to obtain the repaired CMS data; wherein, the data repair processing includes using different mathematical transformations or interpolation methods for signal asymmetry, zero-point drift and data jump anomalies respectively;

[0115] The second data repair module 604 is used to remove outliers and fill in missing values ​​in the SCADA data to obtain repaired SCADA data; the repaired CMS data and the repaired SCADA data are used as target input data to input the fault diagnosis model.

[0116] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that, for the sake of convenience and brevity, the division of the above-mentioned functional units and modules is only used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0117] like Figure 7 As shown, embodiments of the present invention provide a terminal device, such as... Figure 7As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 7 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0118] Specifically, when the processor D100 executes the computer program D102, in response to the diagnostic instructions of the wind turbine fault diagnosis model, it obtains the target operating data associated with the diagnostic instructions from the wind power operation database; it inputs the CMS data into a pre-trained integrated convolutional neural network model for anomaly detection and classification, identifying whether the CMS data belongs to a normal signal or an abnormal type; it performs corresponding data repair processing on the CMS data according to the identified anomaly type to obtain repaired CMS data; it performs outlier removal and missing value filling on the SCADA data to obtain repaired SCADA data; it inputs the repaired CMS data and the repaired SCADA data as target input data into the fault diagnosis model, or triggers a sensor fault alarm when the proportion of abnormal data exceeds a preset threshold. By integrating SCADA operating condition data and CMS vibration data, the shortcomings of insufficient information from a single data source were overcome. An integrated convolutional neural network based on the Bagging strategy was used to achieve accurate identification and robust classification of complex wind power data anomalies. Furthermore, a mathematical repair method was innovatively adopted for specific anomalies such as signal asymmetry, drift, and jumps, replacing the traditional direct rejection method. This transformed failed data into usable data, significantly improving data utilization and integrity. It provided high-quality standard input for downstream fault diagnosis models, thereby greatly improving the accuracy and reliability of wind turbine fault diagnosis.

[0119] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0120] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0121] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0122] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0123] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0124] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for preprocessing wind turbine data, characterized in that, include: In response to the diagnostic command of the wind turbine fault diagnosis model, target operating data associated with the diagnostic command is obtained from the wind power operation database; The target operating data includes SCADA data reflecting the unit's operating conditions and CMS data reflecting component vibration. The CMS data is input into a pre-trained ensemble convolutional neural network model for anomaly detection and classification, identifying whether the CMS data belongs to normal signals or anomaly types. The ensemble convolutional neural network model adopts a Bagging ensemble strategy and is composed of multiple independent convolutional neural network sub-models. The anomaly types include signal asymmetry, zero-point drift, and data jump anomalies. Based on the identified anomaly type, corresponding data repair processing is performed on the CMS data to obtain repaired CMS data; wherein, the data repair processing includes applying different mathematical transformations or interpolation methods for signal asymmetry, zero-point drift, and data jump anomalies; the specific steps of performing corresponding data repair processing on the CMS data based on the identified anomaly type include: If the identified anomaly type is signal asymmetry anomaly, then zero-mean processing is performed on the CMS data to eliminate DC component deviation; If the identified anomaly type is zero-point drift anomaly, then detrending processing is performed on the CMS data to eliminate baseline drift that changes slowly over time; If the identified anomaly type is a data jump anomaly, then the jump point in the CMS data is identified, the jump point is marked as an invalid value, and the vibration value of the jump point is reconstructed based on the normal data before and after the jump point using a cubic spline interpolation algorithm, so that it is restored to normal data that conforms to physical continuity. The SCADA data is subjected to outlier removal and missing value filling to obtain repaired SCADA data; the repaired CMS data and the repaired SCADA data are used as target input data to the fault diagnosis model.

2. The wind turbine data preprocessing method according to claim 1, characterized in that, The step of obtaining the target operational data associated with the diagnostic command from the wind power operation database includes: The diagnostic instructions of the wind turbine fault diagnosis model are analyzed to determine the target component to be diagnosed and the corresponding fault type. A data index is established based on the fault type, and candidate SCADA data associated with the target component is retrieved from the wind power SCADA database, and candidate CMS data associated with the target component is retrieved from the CMS database. Extract wind speed and active power values ​​from the candidate SCADA data, and identify the time interval during which the wind turbine is generating electricity based on the wind speed and active power values. Extract data segments from the candidate SCADA data and the candidate CMS data that fall within the time interval, and combine them as the target running data.

3. The wind turbine data preprocessing method according to claim 2, characterized in that, The CMS data is input into a pre-trained ensemble convolutional neural network model for anomaly detection and classification, identifying whether the CMS data belongs to a normal signal or an anomaly type, including: A raw dataset containing various vibration anomaly types was created and labeled. Multiple distinct sub-training sets were generated from the original dataset using the Bootstrap self-sampling method; Multiple convolutional neural network sub-models with identical structures are trained in parallel using multiple sub-training sets; In the prediction phase, the CMS data to be detected is input into multiple convolutional neural network sub-models, and the outputs of all convolutional neural network sub-models are integrated by voting or averaging to determine the final anomaly classification result.

4. The wind turbine data preprocessing method according to claim 3, characterized in that, The convolutional layers in the integrated convolutional neural network model extract the physical features of vibration signals from CMS data through convolutional kernels. The feature extraction process satisfies the following formula: in, Indicates the first The first convolutional layer Each output feature map represents the deep features of the extracted vibration signal; Indicates the first The first convolutional layer Vibration data segment corresponding to each receptive field; Indicates the first The first convolutional layer The weight matrix of each convolutional kernel corresponds to the vibration mode filter; This is the bias value; Represents a nonlinear activation function; This represents the convolution operation.

5. The wind turbine data preprocessing method according to claim 4, characterized in that, The outlier removal and missing value filling of the SCADA data specifically includes: Identify abnormal values ​​in the SCADA data caused by communication interruption or sensor failure; the abnormal values ​​include NULL values, NA values, constant values, and consecutive 0 values; The outlier values ​​are then interpolated. For the temperature data in the SCADA data, a method of aggregating and aligning the timelines of data collected from all units in the field is adopted, and the missing temperature parameters are filled in by using the operating parameters of the same model of units on the same timeline.

6. The wind turbine data preprocessing method according to claim 1, characterized in that, The method also includes triggering a sensor fault alarm when the proportion of abnormal data exceeds a preset threshold, including: After processing the CMS data, the ratio of abnormal data to normal data is calculated. If the ratio is greater than a preset threshold, it is determined that the data acquisition sensor has a physical fault, the input of data to the fault diagnosis model is terminated, and a sensor maintenance alarm signal is output. If the ratio is less than or equal to the preset threshold, the repaired data will be input into the fault diagnosis model for fault diagnosis of unit components.

7. A wind turbine data preprocessing device, characterized in that, include: The data acquisition module is used to respond to the diagnostic command of the wind turbine fault diagnosis model and obtain the target operating data associated with the diagnostic command from the wind power operation database; the target operating data includes SCADA data reflecting the unit's operating condition and CMS data reflecting the vibration of components; An anomaly detection module is used to input the CMS data into a pre-trained ensemble convolutional neural network model for anomaly detection and classification, identifying whether the CMS data belongs to normal signals or anomaly types; the ensemble convolutional neural network model adopts a Bagging ensemble strategy and is composed of multiple independent convolutional neural network sub-models; the anomaly types include signal asymmetry, zero-point drift, and data jump anomalies. The first data repair module is used to perform corresponding data repair processing on the CMS data according to the identified anomaly type to obtain repaired CMS data; wherein, the data repair processing includes using different mathematical transformations or interpolation methods for signal asymmetry, zero-point drift, and data jump anomalies respectively; the step of performing corresponding data repair processing on the CMS data according to the identified anomaly type specifically includes: If the identified anomaly type is signal asymmetry anomaly, then zero-mean processing is performed on the CMS data to eliminate DC component deviation; If the identified anomaly type is zero-point drift anomaly, then detrending processing is performed on the CMS data to eliminate baseline drift that changes slowly over time; If the identified anomaly type is a data jump anomaly, then the jump point in the CMS data is identified, the jump point is marked as an invalid value, and the vibration value of the jump point is reconstructed based on the normal data before and after the jump point using a cubic spline interpolation algorithm, so that it is restored to normal data that conforms to physical continuity. The second data repair module is used to remove outliers and fill in missing values ​​in the SCADA data to obtain repaired SCADA data; the repaired CMS data and the repaired SCADA data are used as target input data to the fault diagnosis model.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images