Fault detection regulation and control method for electric control system of wind turbine generator

Abstract

The invention discloses a fault detection regulation and control method for an electric control system of a wind turbine generator set, and relates to the technical field of operation monitoring and electric control of the wind turbine generator set. The unit can obtain the abnormal representation related to the health state deviation before the vibration amplitude exceeds the limit, so that the dependence on a single threshold triggering condition is reduced; in order to solve the problem that an abnormal mode is unstable due to the fact that residual features are submerged by turbulent flow, control action and measurement noise, a reconstruction residual is focused in a time dimension through a time attention weight generated by a congelation risk index, so that residual information in a high congelation tendency time period obtains a higher contribution in an aggregation process; therefore, the interference of the noise in the low-risk period on the overall discrimination is reduced.

Description

Technical Field

[0001] This invention relates to the field of wind turbine generator operation monitoring and electrical control technology, and in particular to a fault detection and control method for wind turbine generator electrical control system. Background Technology

[0002] Megawatt-class wind turbines operating under cold and humid weather conditions such as low temperature and high humidity are prone to blade icing. Icing alters the blade airfoil shape and increases additional loads, resulting in decreased power generation, increased load and vibration levels, and potential safety risks such as ice blowout. In particular, when icing is unevenly distributed among different blades or along the blade span, it may cause rotor mass and aerodynamic imbalance, thereby generating periodic load components at the rotational frequency and its harmonics. Long-term effects may accelerate fatigue damage to transmission chain components such as the main shaft, bearings, and gearbox. When changes in operating conditions or temperature rise cause icing to detach / redistribute, the imbalance may abruptly occur, triggering load impacts and posing a risk to structural safety.

[0003] To improve the early detection capability of abnormal states, some existing technologies have gradually evolved from single threshold alarms to establishing normal behavior models (NBMs) using operational data such as SCADA, and realizing anomaly detection and early warning through the residual (prediction error) between real-time data and model predictions and its statistical characteristics.

[0004] However, under the condition of thin icing formed in the early stage of icing, the parameter changes caused by uneven icing are small and coupled with factors such as turbulence, wind shear, pitch / speed control actions and measurement noise, resulting in significant power curve dispersion and residual fluctuations. Conventional methods are unable to stably extract features strongly correlated with imbalance from high-dimensional time-series residuals, which may lead to problems such as delayed early warning or frequent false alarms. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a fault detection and control method for wind turbine electrical control systems to solve the problems of early-stage thin icing causing hidden imbalance, signal being submerged in turbulence and control noise, and existing methods having delayed early warning and many false alarms.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a fault detection and control method for a wind turbine electrical control system, comprising:

[0009] Step S1: Collect power, speed, pitch angle, generator electromagnetic torque and its fluctuation, ambient temperature and ambient humidity to form multi-dimensional time series data;

[0010] Step S2: Train the adversarial generative reconstruction model based on historical normal operation data, and fix the trained generator as the healthy baseline model;

[0011] Step S3: Input real-time multidimensional time series data into the health benchmark model to obtain reconstructed data and calculate the reconstructed residuals;

[0012] Step S4: Generate time attention weights based on the freezing risk index obtained from ambient temperature and humidity, and generate feature attention weights based on the similarity between the residual and the freezing prior feature template. Weight the reconstructed residuals to obtain focused residual features.

[0013] Among them, the freezing prior feature template is a template vector obtained by statistical analysis of the residuals of historically confirmed freezing samples;

[0014] Step S5: Input the focused residual features into the freezing discrimination network to output the confidence score of the hidden leaf imbalance.

[0015] Step S6: When the confidence score exceeds the adaptive warning threshold, an early warning is output and the corresponding operation control command is triggered.

[0016] As a preferred embodiment of the fault detection and control method for the wind turbine electrical control system described in this invention, the multi-dimensional time-series data further includes the frequency band energy characteristics corresponding to the blade rotation frequency and its harmonics obtained by spectrum analysis of the nacelle vibration signal or tower vibration signal.

[0017] As a preferred embodiment of the fault detection and control method for the wind turbine electrical control system of the present invention, the fluctuation of the generator electromagnetic torque is a statistical fluctuation characteristic calculated from the electromagnetic torque estimate or electromagnetic torque command value within a sliding time window, and the length of the sliding time window is related to the blade rotation period corresponding to the current speed.

[0018] As a preferred embodiment of the fault detection and control method for the wind turbine power control system of the present invention, the generator of the adversarial generative reconstruction model is an encoder-decoder structure, and both the encoder and the decoder contain recurrent neural network layers or temporal convolutional layers for learning the temporal dependencies of the multidimensional temporal data.

[0019] As a preferred embodiment of the fault detection and control method for the wind turbine electrical control system of the present invention, the freezing risk index is calculated by normalizing the ambient temperature and ambient humidity, and the freezing risk index is mapped to a time attention weight through monotonic mapping. The higher the freezing risk index, the greater the corresponding time attention weight.

[0020] As a preferred embodiment of the fault detection and control method for the wind turbine electrical control system described in this invention, the prior feature template for freezing is obtained by performing dimensionality reduction processing and feature filtering on the reconstruction residuals corresponding to historically confirmed initial freezing events.

[0021] As a preferred embodiment of the fault detection and control method for the wind turbine power control system of the present invention, the step of generating feature attention weights based on the similarity between the residual and the freezing prior feature template includes: calculating the similarity between each residual dimension and the freezing prior feature template and normalizing it, and assigning feature attention weights to the residual dimensions based on the normalization results.

[0022] As a preferred embodiment of the fault detection and control method of the wind turbine power control system described in this invention, the freezing discrimination network is obtained by pre-training using a training set containing real freezing samples and synthetic samples. The synthetic samples are generated by superimposing periodic perturbations related to the blade rotation frequency onto normal samples and simultaneously superimposing power deviation perturbations.

[0023] As a preferred embodiment of the fault detection and control method for the wind turbine electrical control system of the present invention, the adaptive warning threshold is adjusted according to the output power level, wind speed, or statistical quantity representing turbulence intensity. When the wind speed increases or the statistical quantity representing turbulence intensity increases, the adaptive warning threshold is increased accordingly. When the warning is triggered, the operation control command includes at least: power-limited operation, adjusting the pitch angle to the anti-icing angle, and / or starting the blade heating device.

[0024] As a preferred embodiment of the fault detection and control method for the wind turbine power control system of the present invention, it further includes a model update step: recording the multi-dimensional time series data, focusing residual features and event verification results corresponding to the early warning event, and incrementally updating the health benchmark model, attention weight generation rules and freezing discrimination network at a preset period.

[0025] Through the above technical solution, the present invention can achieve at least the following beneficial effects:

[0026] To address the issue that latent imbalances caused by early thin icing are difficult to trigger in a timely manner by traditional threshold protection, a health benchmark model is used to reconstruct multidimensional operating data and form reconstruction residuals. This allows the unit to obtain abnormal characteristics related to deviations from its health status before vibration amplitude exceeds limits, reducing reliance on single threshold triggering conditions.

[0027] To address the issue of unstable anomalous patterns caused by residual features being submerged by turbulence, control actions, and measurement noise, a time attention weight generated by the freezing risk index is used to focus the reconstructed residuals in the time dimension. This allows residual information from periods with high freezing tendency to contribute more to the aggregation process, thereby reducing the interference of noise from low-risk periods on the overall discrimination.

[0028] To address the difficulty of extracting a subset of features strongly correlated with early freezing from high-dimensional residuals using conventional anomaly detection, feature attention weights are generated by comparing the similarity between residuals and freezing prior feature templates. This enhances residual dimensions that are more consistent with the freezing prior morphology at the feature level, while suppressing weakly correlated dimensions, resulting in more specific focused residual features and improving the separability of early freezing-related imbalance signs.

[0029] To address the issues of low correlation between early warnings and actual faults and frequent false alarms, an adaptive warning threshold is determined by using the confidence score output by the freezing discrimination network and combining it with the output power level, wind speed, and turbulence intensity statistics. This allows the threshold to adapt to changes in operating conditions. Furthermore, continuous judgment and hysteresis logic reduce frequent triggering caused by threshold jitter, thereby improving the stability of the warning.

[0030] To address the issue of disconnect between early warning and control actions, which prevents rapid risk response, the system outputs operational control commands after an early warning is triggered and implements actions such as power limiting, pitch anti-icing angle adjustment, and blade heating based on the early warning level mapping table. This allows the detection results to be directly converted into an executable electronic control closed loop, and the control actions are ensured to meet the unit's safety logic through interlocking prohibition state constraints.

[0031] To address the problem of model mismatch caused by long-term drift between environment and unit status, this paper records multi-dimensional time-series data of early warning events, focuses on residual characteristics and event verification results, and performs incremental updates and version rollback management on a periodic basis. This enables the model to maintain its adaptability to changes in field operating conditions after new data is accumulated, while maintaining consistency and traceability of behavior before and after the update. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0033] Figure 1 This is a flowchart of the fault detection and control method of the wind turbine electrical control system in the embodiment. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0035] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0036] Example 1:

[0037] like Figure 1 As shown, this application proposes a fault detection and control method for a wind turbine electrical control system, comprising the following steps:

[0038] Step S1: Collect power, speed, pitch angle, generator electromagnetic torque and its fluctuation, ambient temperature and humidity to form multidimensional time-series data. The multidimensional time-series data is collected by the unit controller, converter control unit and environmental monitoring unit. All data undergoes unified time reference alignment processing before entering the model, is resampled according to the same sampling period, and missing samples are filled forward or linearly interpolated. Outliers that significantly exceed the allowable physical range of the unit are removed. Dimensionless and normalized processing is performed on the data of each dimension, and the reference statistics used for normalization are recorded for online restoration.

[0039] The electromagnetic torque of the generator is either an estimated value or a command value of electromagnetic torque output by the converter control unit; the fluctuation of the electromagnetic torque is a statistical fluctuation characteristic calculated from the electromagnetic torque sequence within a sliding time window, and the statistical fluctuation characteristic includes at least one of mean square fluctuation, peak-to-peak fluctuation, or first-order difference fluctuation; the sliding time window is updated with the real-time speed so that the time window covers several blade rotation cycles.

[0040] Step S2: Train the adversarial generative reconstruction model based on historical normal operation data, and fix the trained generator as the healthy baseline model;

[0041] Historical normal operation data consists of operation records of the unit in a stable grid-connected power generation state without triggering fault shutdowns or icing alarms; during the screening process, the historical normal operation data undergoes operating condition consistency verification to remove non-stable sections caused by start-stop transitions, strong power limiting interventions, or rapid pitch changes; sections where the power-wind speed relationship deviates significantly from the stable power curve distribution of the unit are removed; the screened dataset is divided into training and validation sets according to time order.

[0042] Step S3: Input real-time multidimensional time series data into the health benchmark model to obtain reconstructed data and calculate the reconstructed residuals;

[0043] The reconstructed residual is a sequence of differences between the real-time data and the reconstructed data at corresponding dimensions and time points. The reconstructed residual sequence is subjected to sliding statistical summarization to obtain a residual statistical feature sequence, which includes at least one of the following: mean drift feature, variance increment feature, or bandpass energy feature. The residual statistical feature sequence is aligned with the original reconstructed residual sequence on the same time axis.

[0044] Step S4: Generate time attention weights based on the freezing risk index obtained from ambient temperature and humidity, and generate feature attention weights based on the similarity between the residuals and the freezing prior feature templates. Weight the reconstructed residuals to obtain focused residual features. The focused residual features are weighted residual representations obtained by applying the time attention weights and feature attention weights to the reconstructed residuals. The output of the focused residual features is a fixed-dimensional vector or a fixed-dimensional time series segment. When the output is a fixed-dimensional vector, it is obtained by weighted statistical summarization within a sliding time window. When the output is a fixed-dimensional time series segment, the weighted time series is retained and kept aligned with the input time axis. The focused residual features contain saliency identification information of several key dimensions with the largest weights.

[0045] Among them, the freezing prior feature template is a template vector obtained by statistical analysis of the residuals of historically confirmed freezing samples;

[0046] The freezing risk index is calculated by combining the ambient temperature and ambient humidity after dimensionless conversion. The freezing risk index is a dimensionless scalar. The time attention weight is obtained by monotonic mapping of the freezing risk index and normalized within a sliding time window to ensure that the weight satisfies the summation and normalization condition. When the freezing risk index is in the low-risk range, the time attention weight remains at a low level. When the freezing risk index is in the high-risk range, the time attention weight is increased and maintained until the freezing risk index falls back to the relief range.

[0047] Step S5: Input the focused residual features into the freeze discrimination network and output the confidence score of the hidden leaf imbalance. The input of the freeze discrimination network is the focused residual features, and the output is the confidence score. The confidence score is obtained by performing a monotonic mapping on the network output and satisfies the relationship that it monotonically increases with the increase of the abnormality of the focused residual features. The freeze discrimination network adopts a supervised learning method during the training phase, and the training labels are generated by the freeze event verification results. After training, the network parameters are fixed and the distribution statistics of the output scores on the normal sample set are recorded for threshold setting.

[0048] Step S6: When the confidence score exceeds the adaptive warning threshold, an early warning is output and the corresponding operation control command is triggered;

[0049] In this embodiment, the multidimensional time-series data also includes the frequency band energy characteristics corresponding to the blade rotation frequency and its harmonics obtained by spectrum analysis of the nacelle vibration signal or tower vibration signal;

[0050] The blade rotation frequency is calculated from the rotational speed, and the center frequency of the harmonic band is updated synchronously with the real-time rotational speed. The frequency band energy characteristics are obtained by estimating the spectrum of the vibration signal within a sliding time window, and by integrating or accumulating the spectral amplitude within a preset bandwidth adjacent to the center frequency. The frequency band energy characteristics are timestamped with the multidimensional time series data within the same sliding time window.

[0051] In this embodiment, the fluctuation of the generator electromagnetic torque is a statistical fluctuation characteristic calculated from the electromagnetic torque estimate or electromagnetic torque command value within a sliding time window, and the length of the sliding time window is related to the blade rotation period corresponding to the current speed.

[0052] In this embodiment, the generator of the adversarial generative reconstruction model is an encoder-decoder structure. Both the encoder and the decoder contain recurrent neural network layers or temporal convolutional layers to learn the temporal dependencies of multidimensional temporal data.

[0053] The generator of the adversarial generative reconstruction model takes a multidimensional time series segment within a fixed-length time window as input and outputs a reconstructed time series segment with the same dimension and length as the input. The discriminator takes a real time series segment and a reconstructed time series segment as input and outputs the discrimination result. The training objective consists of a reconstruction error term and an adversarial constraint term. During training, the time to stop training is determined based on the convergence state of the reconstruction error on the validation set and the stable state of the discriminant output. After training, the generator parameters are fixed and its normalized baseline statistics are frozen.

[0054] In this embodiment, the freezing risk index is calculated by normalizing the ambient temperature and ambient humidity, and the freezing risk index is mapped to the time attention weight through monotonic mapping. The higher the freezing risk index, the greater the corresponding time attention weight.

[0055] In one implementation, the time attention weight is obtained by a monotonic mapping of the freezing risk index, and can form a coherent closed loop through a link of risk calculation - de-hysteresis - in-window normalization - upper and lower limit constraints; the specific steps are as follows:

[0056] When the sampling time is recorded as At that time, the ambient temperature and humidity are normalized into risk components of the same dimension, and then combined to obtain the freezing risk index:

[0057] ,

[0058] in, This represents the temperature risk component obtained by normalizing the ambient temperature at time n. This represents the humidity risk component obtained by normalizing the ambient humidity at time n. Indicates the first ambient temperature at all times Indicates the first Constant ambient humidity, Indicates the temperature risk reference point. This indicates the lower limit of temperature normalization. This indicates the lower limit of humidity normalization. This indicates the upper limit of humidity normalization. This indicates that the input will be truncated to... A saturation function for an interval;

[0059] Constructing a freezing risk index based on normalized components:

[0060] ,

[0061] in, Indicates the first Risk index of freezing at any time. This represents the weighting coefficient for the temperature component. This represents the weighting coefficient for the humidity component. Same meaning as before; command ;

[0062] To suppress the instantaneous rise caused by noise from temperature and humidity measurements, de-jittering and hysteresis are introduced into the risk index to obtain an effective risk level:

[0063] ,

[0064] in, This indicates the amount of risk exceeding the limit. This is an indicator function; it returns 1 if the condition is true, and 0 otherwise. This indicates raising the trigger threshold. Indicates the length of the continuous window count. Indicates a continuous over-limit sign. An effective risk index indicating the delay in de-shake performance. This represents the fallback threshold, and takes... , The index represents the continuous window product; 𝑘 represents the index of the normalized summation within the window; 𝑚 represents the index of the quadratic normalized summation.

[0065] The effective risk is converted into unnormalized time-based attention intensity using a monotonic mapping, and minimum / maximum intensities are set to avoid extreme values:

[0066] ,

[0067] in, Indicates the first Unnormalized time-based attention intensity This represents the lower limit coefficient of attention intensity. Represents the shape coefficient of a monotonic mapping. When the risk level increases, the relative weight of high-risk moments becomes more concentrated. Same meaning as before;

[0068] In length The intensity is normalized within the sliding window so that the sum of the weights within the window is 1:

[0069] ,

[0070] in, Indicates the first When the time is the end of the window, the first Normalized temporal attention weights corresponding to each time point within a window. Indicates the length of the sliding window. This represents the unnormalized temporal attention intensity at the corresponding moment within the window;

[0071] To avoid the weights being too large or too small at any single moment, upper and lower bounds are applied to the weights, followed by secondary normalization:

[0072] ,

[0073] in, This represents the temporal attention weight after amplitude limiting. This represents the temporal attention weights after double normalization. Indicates the lower limit of the weight. Indicates the upper limit of weight. This indicates that the input will be truncated to... A saturation function for an interval;

[0074] In step S4, it is possible to The reconstructed residual sequence obtained in step S3 is subjected to time focusing, for example, by weighting and converging the residual vectors within the window into time-focused residuals:

[0075] ,

[0076] in, Indicates the first The residual feature vector after time-focusing at each time step. This represents the reconstructed residual vector at the corresponding time point within the window. Same meaning as before;

[0077] Specifically, the mapping link in the above implementation constructs a freezing risk index using the normalized results of temperature and humidity, expressing the risk quantity within a uniform scale. Weight coefficients reflect the contribution ratio of the two types of environmental factors to the freezing tendency. After de-jittering and hysteresis processing, instantaneous fluctuations in the risk index are no longer directly amplified into changes in attention weights; a more significant weight increase only occurs when consecutive limit conditions are met. Simultaneously, threshold intervals are used to reduce frequent weight jumps during the decline phase. Monotonic mapping converts effective risk into attention intensity, and the mapping shape is used to adjust the degree of focus during high-risk moments, with the minimum intensity preventing complete information loss during low-risk phases. Normalization within the sliding window ensures the comparability of weights within the same time window and suppresses single-point dominance or near-failure caused by extreme values ​​through upper and lower bound constraints. This temporal attention weight, combined with the residual sequence, can enhance the residual contribution during periods of higher freezing risk, providing a more targeted input representation for subsequent feature attention and discrimination networks.

[0078] In this embodiment, the freezing prior feature template is obtained by dimensionality reduction and feature filtering of the reconstruction residuals corresponding to historically confirmed initial freezing events.

[0079] Historically confirmed initial freezing events are determined jointly by field operation records, meteorological records, and posterior verification results. For each freezing event, residual time window segments adjacent to the event's start time are extracted, and time alignment and amplitude normalization are performed on the residual segments of different events. Dimensionality reduction processing includes either principal component decomposition or autoencoder dimensionality reduction, and feature selection is determined based on the contribution ranking and stability selection rules after dimensionality reduction. The selection results of multiple freezing events are aggregated to obtain a priori feature template vector for freezing and its dimensional meaning is recorded.

[0080] In this embodiment, generating feature attention weights based on the similarity between the residual and the frozen prior feature template includes: calculating the similarity between each residual dimension and the frozen prior feature template and normalizing it, and assigning feature attention weights to the residual dimension based on the normalization result.

[0081] Similarity calculation takes the residual summary features formed within the sliding time window as input, and the residual summary features are obtained by truncating the residual statistical feature sequence. The similarity measure adopts either the relevance measure or the distance measure, and the similarity result is transformed into non-negative weights. The weights of all dimensions are normalized so that the weights satisfy the summation and normalization condition. When the similarity of a certain dimension is lower than the ignore threshold, the attention weight of the corresponding feature is reset to the suppression weight level.

[0082] In one implementation, the feature attention weights in step S4 are constructed around whether the residuals exhibit the morphology of a frozen prior feature template; the temporal attention weights are based on those determined in the aforementioned implementation. Within the same sliding window, residual summary values ​​are extracted for each residual dimension, and then similarity is calculated between these summaries and the corresponding dimensions of the frozen prior feature template. The similarity is then converted into non-negative and normalizable weights. The specific steps are as follows:

[0083] In the first The time is the end of the window, and the window length is Within the sliding window, for the first Construct mean-based and energy-based summary quantities using residual dimensions:

[0084] ,

[0085] in, Indicates the first Time of the first Time-weighted mean summation of residuals Indicates the first Time of the first Time-weighted energy-based summary of residuals Indicates the first in the window Attention weight at each moment Indicates the first The time step for reconstructing the residual vector is the first step. One portion, Indicates the length of the sliding window. Indicates the sampling time index. Indicates the residual dimension index;

[0086] The frozen prior feature template is obtained according to the aforementioned steps. The template vector is then split into template components that correspond one-to-one with the two types of summary quantities mentioned above, based on their dimensions. and ; Calculate the first based on scalar cosine form Dimensional similarity:

[0087] ,

[0088] in, Indicates the first Time of the first The similarity between the residual and the template in the mean-energy summarization space. Indicates the template vector with the first Template components corresponding to the mean-based summary size. Indicates the template vector with the first Template components corresponding to dimensional energy-based abstract quantities This represents the denominator stabilizing term, used to suppress numerical amplification when the template component approaches zero; the meanings of the other parameters are the same as before.

[0089] A negative similarity indicates an opposite trend to the template; therefore, it can be set to zero to achieve non-negativity. Then, Softmax is applied with a minimum weight constraint to obtain the feature attention weights.

[0090] ,

[0091] in, Indicates the first Time of the first Nonnegation of dimensional similarity results Indicates the first Time of the first Dimensional feature attention weights, This represents the minimum weight coefficient for feature attention. This represents the dimension of the residual vector. This represents the Softmax temperature coefficient, used to adjust the sensitivity of similarity differences to weight allocation. This indicates the dimension index used for summation; the other parameters have the same meaning as before.

[0092] In step S4, the feature attention weights are combined with the temporal focused residual components to form a dimensionally weighted result of the focused residual features:

[0093] ,

[0094] in, Indicates the first The time-focused residual eigenvector is the first One portion, Represents the time-focused residual vector The One portion, Obtained by focusing attention on residuals within the window over time. The meaning is the same as before.

[0095] The similarity metric can also be replaced by the Pearson correlation coefficient or distance-based similarity (such as negating and non-negating Mahalanobis distance). The weight generation still follows the similarity calculation-non-negation-normalization-minimum weight constraint link to keep the input-output relationship of step S4 consistent.

[0096] Specifically, in the above implementation, the feature attention construction establishes a computable correspondence between the residual dimensions and the frozen prior feature templates. Each residual dimension is first compressed into a physically meaningful summary within a sliding window, allowing the bias trend and fluctuation intensity of the time series to enter the comparison process simultaneously. The similarity is based on the degree of matching between the summary and the template components, resulting in a similarity result that can be positive or negative. Negative values ​​represent the opposite direction of change to the template and are no longer used to boost weights after non-negation processing, thereby reducing the impact of non-frozen perturbations on weight allocation. The similarity is then normalized to form dimensional weights, making the weights comparable across different residual dimensions and avoiding information loss due to weight collapse through minimum weight constraints. After this weight is combined with the time-focused residuals, the contribution of residual dimensions that better match the frozen template is amplified, while the contribution of dimensions with weaker template correlation is suppressed, providing a more concentrated input representation for the subsequent discriminative network.

[0097] In this embodiment, the freezing discrimination network is pre-trained using a training set containing real freezing samples and synthetic samples. The synthetic samples are generated by superimposing periodic perturbations related to the blade rotation frequency onto normal samples and simultaneously superimposing power deviation perturbations.

[0098] Periodic disturbances are generated using the blade rotation frequency and its harmonics as the fundamental frequency components, and the frequency position is adjusted synchronously with the real-time speed change. Periodic disturbances are superimposed on torque fluctuations, vibration frequency band energy characteristics, or both, and maintain phase consistency or a preset phase relationship between multidimensional disturbances. Power deviation disturbances are superimposed on the power dimension or power residual dimension and appear synchronously with periodic disturbances within the same time window. After the synthetic samples are generated, they are combined with real frozen samples to form a training set, and the training set samples are sampled evenly according to the working condition category.

[0099] In this embodiment, the adaptive warning threshold is adjusted according to the output power level, wind speed, or statistical quantity representing turbulence intensity. When the wind speed increases or the statistical quantity representing turbulence intensity increases, the adaptive warning threshold is increased accordingly. When the warning is triggered, the operation control command includes at least: power-limited operation, adjusting the blade pitch angle to the anti-icing angle, and / or starting the blade heating device.

[0100] The output power level is determined by segmenting the power level of the unit during grid-connected operation; the turbulence intensity statistic is calculated from the fluctuation characteristics of wind speed within the sliding time window; the adaptive warning threshold is determined by the confidence score distribution statistic of the normal sample set within the interval of the corresponding power level and the corresponding turbulence intensity statistic; the adaptive warning threshold has hysteresis logic and continuous judgment logic, and the warning triggering requires the confidence score to continuously exceed the adaptive warning threshold, and the warning cancellation requires the confidence score to continuously fall below the cancellation threshold.

[0101] Operational control commands are determined by the warning level mapping table, and the warning level is determined by the range of confidence scores. When the warning level is low, power-limited operation is performed and the pitch change rate is limited. When the warning level is medium, power-limited operation is performed and the target pitch angle is adjusted to the anti-icing angle range. When the warning level is high, a combined strategy of power-limited operation, target pitch angle adjustment, and blade heating device activation is implemented. When the heating device is in an interlocked prohibited state, the operational control commands do not include heating device activation commands. The exit condition for the control strategy is jointly determined by the warning release judgment and the unit safety interlock status.

[0102] Example 2

[0103] Based on Example 1, the model update step is also included: recording the multi-dimensional time series data, focusing residual features and event verification results corresponding to the early warning event, and incrementally updating the health benchmark model, attention weight generation rules and freezing discrimination network at a preset period;

[0104] The data corresponding to the early warning event is accompanied by an event verification result identifier after recording. The verification result includes three categories: confirmed freezing, confirmed non-freezing, or uncertain. Only confirmed freezing and confirmed non-freezing samples are included in the incremental update data pool. The incremental update is performed in an offline environment. The consistency assessment of the validation set is performed before and after the update, and a model version number is generated. When the false positive rate or false negative rate of the updated model on the validation set exceeds the allowable range, the model version is rolled back to the previous stable version. The updated normalized baseline statistics, freezing prior feature templates, and threshold statistics are updated synchronously and version consistency marking is completed.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0106] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for fault detection and regulation of an electrical control system of a wind turbine generator unit, characterized in that, The method comprises the following steps: Step S1, collecting power, rotating speed, pitch angle, generator electromagnetic torque and its fluctuation, environment temperature and environment humidity to form multi-dimensional time series data; Step S2, training an adversarial generative reconstruction model based on historical normal operation data, and fixing the trained generator as a health benchmark model; Step S3, inputting real-time multi-dimensional time series data into the health benchmark model to obtain reconstructed data and calculate reconstruction residual; Step S4, generating time attention weight according to the icing risk index obtained from the environment temperature and the environment humidity, and generating feature attention weight according to the similarity between the residual and the icing prior feature template, and weighting the reconstruction residual to obtain focused residual feature; Wherein, the icing prior feature template is a template vector obtained by statistical analysis of the residuals of the confirmed icing samples in history; Step S5, inputting the focused residual feature into the icing discrimination network to output the implicit blade imbalance confidence score; Step S6, when the confidence score exceeds the adaptive warning threshold, outputting a warning and triggering corresponding operation control instructions.

2. The method of claim 1, wherein the method further comprises: The multi-dimensional time series data further comprises frequency band energy features corresponding to the blade rotation frequency and its harmonics obtained by frequency spectrum analysis of the cabin vibration signal or the tower vibration signal.

3. The method of claim 1, wherein the method further comprises: The fluctuation of the generator electromagnetic torque is a statistical fluctuation feature calculated from the electromagnetic torque estimated value or the electromagnetic torque instruction value within a sliding time window, and the length of the sliding time window is associated with the blade rotation period corresponding to the current rotating speed.

4. The method of claim 1, wherein the method further comprises: The generator of the adversarial generative reconstruction model is an encoder-decoder structure, and the encoder and the decoder both contain recurrent neural network layers or time series convolution layers for learning the time series dependence of the multi-dimensional time series data.

5. The method of claim 1, wherein the method further comprises: The icing risk index is calculated after the environment temperature and the environment humidity are normalized, and the icing risk index is mapped to the time attention weight through monotonic mapping. The higher the icing risk index, the greater the corresponding time attention weight.

6. The method of claim 1, wherein the method further comprises: The icing prior feature template is obtained by dimension reduction processing and feature selection on the reconstruction residuals corresponding to the confirmed icing initial events in history.

7. The method of claim 1, wherein the method further comprises: The generation of the feature attention weight according to the similarity between the residual and the icing prior feature template comprises: calculating the similarity between each residual dimension and the icing prior feature template and normalizing it, and assigning feature attention weight to the residual dimension according to the normalization result.

8. The method of claim 1, wherein the method further comprises: The icing discrimination network is pre-trained using a training set containing real icing samples and synthetic samples, and the synthetic samples are generated by superimposing periodic disturbance related to the blade rotation frequency and power deviation on normal samples.

9. The method of claim 1, wherein, The adaptive warning threshold is adjusted according to the output power level, wind speed or statistical quantity representing the turbulence intensity. When the wind speed or the statistical quantity representing the turbulence intensity increases, the adaptive warning threshold increases correspondingly; When the warning is triggered, the operation control instructions at least include: limited power operation, adjusting the pitch angle to the anti-icing angle and / or starting the blade heating device.

10. The method of claim 1, wherein the method further comprises: The model updating step is further included: multi-dimensional time sequence data corresponding to the early warning event, focused residual features and event verification results are recorded, and the health benchmark model, attention weight generation rule and freeze judgment network are incrementally updated at a preset period.

Images