Wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold

Abstract

The application relates to the technical field of wind power intelligent operation and maintenance, and discloses a wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold values, wherein through a hybrid deep learning anomaly detection model, time sequence feature extraction and sequence memory enhancement are performed on synchronous time sequence data, wind turbine fusion features are obtained by fusing fine-grained short-term features and coarse-grained long-term features, the abnormal probability of each moment is predicted, the real-time working condition category is identified, the working condition specificity threshold value is dynamically output, the working condition stability index is combined, and a wind turbine temperature early warning signal is generated. Therefore, based on the hybrid deep learning and dynamic threshold value technology, the short-term peak detail features and the long-term dependence features are fused, the model peak fitting precision is improved, the abnormal probability can accurately reflect the temperature abnormality degree, meanwhile, the threshold value is adaptively and dynamically adjusted under different working conditions, the false positive rate of the strong turbulent flow condition is reduced, and accurate decision support is provided for operation and maintenance personnel.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for wind power, and in particular to a method for early warning of wind turbine temperature based on hybrid deep learning and dynamic threshold. Background Technology

[0002] During actual operation of wind turbines, strong turbulence and instantaneous overload conditions are prone to occur in scenarios such as coastal typhoon seasons and complex mountain wind farms. Under these conditions, wind speed fluctuates drastically within a millisecond to second timescale, causing sudden changes in turbine speed and large instantaneous changes in generator load. This leads to instantaneous spikes and rapid drops in the temperature of generator bearings and windings, forming a special temperature change pattern of spike-drop. The detection of this special temperature change pattern suffers from high latency and high false alarm / missed alarm rates, making it difficult to achieve rapid operation and maintenance response in scenarios requiring intelligent operation and maintenance of wind power.

[0003] Existing temperature warning methods for this temperature change pattern have significant technical shortcomings: First, the PID control, simple threshold judgment, or single LSTM model used are insufficiently sensitive to the time of instantaneous peaks, with response delays generally exceeding 10 seconds; second, the models do not consider the coupling relationship of multi-source data, relying only on temperature data, and cannot effectively distinguish between instantaneous anomalies caused by turbulence and temperature rises due to normal load changes; third, fixed threshold mechanisms are difficult to adapt to different operating states, resulting in a high false alarm rate; fourth, warning generation relies solely on comparing a single anomaly probability with a threshold, without integrating multi-source information such as operating condition stability, leading to insufficient warning reliability.

[0004] Therefore, how to capture instantaneous temperature spikes under strong turbulence conditions and provide high-precision, low-delay temperature early warning for wind turbine units is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] This invention provides a wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold, aiming to solve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, this invention provides a wind turbine temperature early warning method based on hybrid deep learning and dynamic thresholding, the method comprising the following steps:

[0007] S1: Acquire sensor operation monitoring data of the target wind turbine, perform preprocessing and integration on the sensor operation monitoring data, and obtain synchronous time series data;

[0008] S2: Construct a hybrid deep learning anomaly detection model, extract temporal features and enhance sequence memory on the synchronous time-series data, and output the anomaly probability at each time step;

[0009] The hybrid deep learning anomaly detection model includes:

[0010] Dilated Convolutional Network: Extracts short-term multi-timescale features containing temperature spikes from synchronous temporal data to obtain the first feature map; CBAM Attention Module: Extracts a second feature map from the first feature map, fusing multi-scale temperature and spike region details;

[0011] Bidirectional GRU: Extracts long-term trend and normal pattern features of temperature changes from the second feature map to obtain the third feature map; Customized local attention encoder: Extracts a fourth feature map from the third feature map that fuses long-term normal patterns and short-term spikes;

[0012] Feature fusion layer: fuses the second feature map and the fourth feature map to obtain fused features; Fully connected layer: outputs the anomaly probability at each time step based on the fused features;

[0013] S3: Extract the operating condition features at each moment from the synchronous time series data, input the real-time feature vector composed of the operating condition features into the trained SVM operating condition classifier, identify the real-time operating condition category of the target wind turbine, query the abnormal probability subset corresponding to the same operating condition category divided under the historical normal synchronous time series data, and take the abnormal probability value of the abnormal probability subset at the preset proportional quantile of the distribution density function as the operating condition specific threshold of the operating condition category.

[0014] S4: Based on the synchronous time series data, calculate the operating condition stability index at each moment, and combine the anomaly probability and the operating condition specific threshold to generate a wind turbine temperature early warning signal.

[0015] Optionally, in step S1, the sensor operation monitoring data is configured as a combination of front and rear bearing temperatures, stator and rotor winding temperatures, fan speed, generator load, nacelle wind speed, nacelle vibration, and grid voltage and current.

[0016] Optionally, in step S1, data synchronization is performed to obtain synchronized time-series data, specifically including: based on the timestamp of each set of data, the sensor operation monitoring data is synchronized and integrated to output the synchronized time-series data of the target wind turbine.

[0017] Optionally, in step S2, temporal feature extraction specifically includes:

[0018] S21: The synchronous temporal data is transmitted to a temporal feature extraction model that includes a dilated convolutional network and a CBAM attention module, and the feature map extracted by the temporal feature extraction model is output as a fine-grained short-term feature of the synchronous temporal data.

[0019] The dilated convolutional network comprises four layers of dilated convolutions, specifically expressed as follows:

[0020] ;

[0021] In the formula, Let t be the convolution output feature. The kernel size is [size]. The weights of the k-th convolutional kernel are... Input data for time td×k. The void ratio of the dilated convolution;

[0022] The CBAM attention module is configured to perform channel attention weighting by global average pooling on the feature map output by the dilated convolutional network and spatial attention weighting by convolutional compression on the channel attention-weighted feature map. The specific expression is as follows:

[0023] ;

[0024] ;

[0025] ;

[0026] In the formula, The attention weight for the Cth channel, It is the Sigmoid activation function. It is a two-layer perceptron. This is the global average pooling result for the C-th channel. This represents the global max pooling result for the C-th channel. The spatial attention weights are H rows and W columns. This represents the convolution operation. The result of global average pooling is H rows and W columns. The result of global max pooling is H rows and W columns. This is a concatenation of the global average pooling result and the global max pooling result. The final feature map after attention weighting. The feature map output by the dilated convolutional network. For channel multiplication, Attention weights that are weighted for channel attention. For space multiplication, Attention weights that are weighted for spatial attention.

[0027] Optionally, in step S2, sequence memory enhancement specifically includes:

[0028] S22: Transmit the temporal features obtained by temporal feature extraction to the sequence memory enhancement module, which includes a two-layer bidirectional GRU and a customized local attention encoder, and output the temporal features processed by the sequence memory enhancement module as coarse-grained long-term features of the synchronous temporal data.

[0029] The two-layer bidirectional GRU employs a two-layer Bi-GRU architecture, with 128 hidden units in each layer. It learns the long-term dependencies of the temperature sequence through update and reset gates, as specifically expressed in the following expression:

[0030] ;

[0031] In the formula, Let the hidden state be at time t. To update the door, To reset the door, The hidden state at time t-1 For the current input features, This is the weight matrix. For bias terms, Element-wise multiplication;

[0032] The customized local attention encoder is specifically expressed as follows:

[0033] ;

[0034] In the formula, Here, m represents the attention weight at time t, and m=3 represents the size of the attention window, designed to enhance local features that allow for a rapid drop in temperature peaks within 3 seconds. The cosine similarity function is used. Let be the local feature sequence m seconds before and after time t. Let be the feature vector of the i-th time point in the local feature sequence m seconds before and after time t.

[0035] Optionally, in step S2, the anomaly probability at each time step is output, specifically including:

[0036] S23: Construct a two-level feature pyramid by combining fine-grained short-term features and coarse-grained long-term features. Align the number of channels through 1×1 convolution, match the feature size using bilinear interpolation, and then achieve feature fusion through ReLU activation and BatchNorm normalization.

[0037] S24: Input the fused features into two fully connected layers, and output the anomaly probability at each time step using the Sigmoid function; the expression for the loss function is as follows:

[0038] ;

[0039] In the formula, For the total loss, For sequence length, For balance coefficients, outlier samples normal samples , For focusing coefficient, Let be the probability of an anomaly at time t. For real labels, 1 = abnormal, 0 = normal. For peak penalty weights, For the rate of temperature change, The threshold for determining the peak value. For indicator functions, This is the actual temperature value. To predict temperature values.

[0040] Optionally, in step S3, before extracting the operating condition features at each moment from the synchronous time-series data and inputting the real-time feature vector composed of the operating condition features into the trained SVM operating condition classifier to identify the real-time operating condition category of the target wind turbine, the method further includes:

[0041] S31: Extract operating condition features from historical normal synchronization time series data, and divide them into training sets by stratified sampling according to a preset ratio. and verification set ;

[0042] S32: Density peak and K-means clustering are used to cluster the extracted operating conditions. Initial cluster centers are selected by local density and relative distance, and then the sum of squares within the cluster is minimized iteratively to divide the operating conditions into four categories: strong turbulent overload, moderate fluctuation, normal operation, and shutdown.

[0043] S33: For the training set and verification set Assign cluster labels separately and This forms a labeled training dataset. and validation dataset ;

[0044] S34: Construct an SVM condition classifier using the training dataset. Train the SVM condition classifier on the validation dataset. The training performance of the model is evaluated, and the trained SVM condition classifier is saved after the performance meets the target.

[0045] Optionally, in step S3, the abnormal probability subset corresponding to the same working condition category divided under historical normal synchronized time series data is queried, and the abnormal probability value of the abnormal probability subset at the preset proportional quantile of the distribution density function is taken as the working condition specificity threshold of that working condition category, specifically including:

[0046] S35: Based on the four types of working condition labels formed by the division of working conditions, the abnormal probability values ​​obtained by feature extraction, fusion and abnormal probability calculation of historical normal synchronization time series data are divided into four abnormal probability subsets, and each abnormal probability subset corresponds to one type of working condition.

[0047] S36: Perform flow interruption-adaptive kernel density estimation on each subset of anomaly probabilities to calculate the distribution density of anomaly probabilities for various operating conditions;

[0048] S37: Take the probability density function of each type of abnormal working condition. The 99.5th percentile was used as the specific threshold for this type of working condition, forming a set of specific thresholds for four types of working conditions. .

[0049] Optionally, in step S4, based on the synchronized time series data, the operational stability index for each time moment is calculated, specifically including:

[0050] S41: Based on the sequence of nacelle wind speed and the sequence of generator load in the synchronous time series data, the operating condition stability index is obtained by weighted fusion calculation of wind speed variation coefficient and load fluctuation coefficient.

[0051] The expression for the operating condition stability index is as follows:

[0052] ;

[0053] In the formula, , These represent the standard deviation and mean of the real-time cabin wind speed, respectively. , These represent the standard deviation and mean of the real-time generator load, respectively. This is the coefficient of variation of wind speed. For load fluctuation coefficient, The larger the value, the more stable the operating conditions.

[0054] Optionally, in step S4, a wind turbine temperature warning signal is generated by combining the anomaly probability and the operating condition-specific threshold, specifically including:

[0055] S42: Based on the calculated operating condition stability index, combined with the anomaly probability and operating condition specific threshold, a weighted summation method is used to calculate the comprehensive anomaly score. The specific expression is as follows:

[0056] ;

[0057] In the formula, To calculate the overall abnormal score, This is the specific threshold for the c-th type of operating condition;

[0058] S43: Based on the comprehensive anomaly score and the preset mapping range between the wind turbine temperature warning level and the comprehensive anomaly score, generate a wind turbine temperature warning signal that includes the wind turbine temperature warning level.

[0059] The beneficial effects of this invention are as follows: It proposes a wind turbine temperature early warning method based on hybrid deep learning and dynamic thresholding. By acquiring synchronous time-series data of the target wind turbine, a constructed hybrid deep learning anomaly detection model is used to extract temporal features and enhance sequence memory in the synchronous time-series data, obtaining a fused feature of the wind turbine composed of fine-grained short-term features and coarse-grained long-term features. The fused feature is used to predict the anomaly probability at each moment. Then, by identifying the real-time operating condition category of the target wind turbine, a condition-specific threshold is dynamically output. Combined with the operating condition stability index, a wind turbine temperature early warning signal is generated. Thus, this invention, based on the introduction of hybrid deep learning and dynamic thresholding technology, retains the details of the peak position by fusing short-term peak detail features and long-term dependency features, increases the penalty for prediction errors in the peak region, forces the model to improve the peak fitting accuracy, and ensures that the anomaly probability can accurately reflect the degree of temperature anomaly. At the same time, it can consider the differences in the anomaly probability distribution under different operating conditions, realize adaptive dynamic adjustment of the threshold under different operating conditions, reduce the false alarm rate under strong turbulence conditions, and provide accurate decision support for operation and maintenance personnel. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the wind turbine temperature early warning method based on hybrid deep learning and dynamic thresholds according to an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram illustrating the principle of the wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold in an embodiment of the present invention.

[0062] Figure 3 This is a schematic diagram illustrating the principle of temporal feature extraction in this invention;

[0063] Figure 4 This is a schematic diagram illustrating the principle of dynamic threshold calculation in this invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0065] This invention provides a method for early warning of wind turbine temperature based on hybrid deep learning and dynamic thresholds, referring to... Figure 1 and Figure 2 .

[0066] In this embodiment, a wind turbine temperature early warning method based on hybrid deep learning and dynamic thresholds is provided, the method comprising the following steps:

[0067] S1: Acquire sensor operation monitoring data of the target wind turbine, perform preprocessing and integration on the sensor operation monitoring data, and obtain synchronous time series data;

[0068] S2: Construct a hybrid deep learning anomaly detection model, extract time-series features and enhance sequence memory on the synchronous time-series data, obtain the wind turbine fusion feature which is a fusion of fine-grained short-term features and coarse-grained long-term features, and output the anomaly probability at each time step after processing by a fully connected layer.

[0069] S3: Extract the operating condition features at each moment from the synchronous time series data, input the real-time feature vector composed of the operating condition features into the trained SVM operating condition classifier, identify the real-time operating condition category of the target wind turbine, and dynamically output the operating condition specific threshold corresponding to the real-time operating condition category.

[0070] S4: Based on the synchronous time series data, calculate the operating condition stability index at each moment, and combine the anomaly probability and the operating condition specific threshold to generate a wind turbine temperature early warning signal.

[0071] It should be noted that existing temperature warning methods have significant technical defects: First, the PID control, simple threshold judgment, or single LSTM model used are not sensitive enough to the time of instantaneous peaks, and the response delay generally exceeds 10 seconds; Second, the model does not consider the coupling relationship of multi-source data, and only relies on temperature data, which cannot effectively distinguish the temperature rise caused by instantaneous anomalies due to turbulence from that caused by normal load changes; Third, the fixed threshold mechanism is difficult to adapt to different operating states, resulting in a high false alarm rate; Fourth, the warning generation only relies on the comparison of a single anomaly probability with a threshold, without integrating multi-source information such as operating condition stability, resulting in insufficient reliability of the warning.

[0072] To address the aforementioned issues, this embodiment acquires synchronous time-series data of the target wind turbine and utilizes a constructed hybrid deep learning anomaly detection model to extract temporal features and enhance sequence memory in the synchronous time-series data. This yields a fused feature of the wind turbine, combining fine-grained short-term features with coarse-grained long-term features. The fused feature is then used to predict the anomaly probability at each moment. Furthermore, by identifying the real-time operating condition category of the target wind turbine, a condition-specific threshold is dynamically output. Combined with operating condition stability indicators, a wind turbine temperature warning signal is generated. Therefore, this invention, based on the introduction of hybrid deep learning and dynamic thresholding technology, fuses short-term peak detail features and long-term dependency features, preserving peak location details and increasing the penalty for prediction errors in peak areas. This forces the model to improve peak fitting accuracy, ensuring that the anomaly probability accurately reflects the degree of temperature anomaly. Simultaneously, it considers the differences in anomaly probability distribution under different operating conditions, enabling adaptive dynamic adjustment of thresholds under different conditions. This reduces the false alarm rate under strong turbulence conditions, providing accurate decision support for maintenance personnel.

[0073] In a preferred embodiment, in step S1, the sensor operation monitoring data is configured as a combination of front and rear bearing temperatures, stator and rotor winding temperatures, fan speed, generator load, nacelle wind speed, nacelle vibration, and grid voltage and current.

[0074] In this embodiment, considering that temperature anomalies are caused by the coupling of multiple factors, sensor data directly related to generator heating (front and rear bearing and winding temperatures), load changes (fan speed and generator load), external environment (nacelle wind speed), equipment status (nacelle vibration), and external power supply (grid voltage and current) are selected to ensure a high degree of matching between the data dimensions and the causes of temperature anomalies. This embodiment, through the complementarity of multi-source data, can solve the problem of traditional methods relying solely on single-dimensional temperature data, improving the ability to distinguish between instantaneous anomalies caused by strong turbulence and temperature rises due to normal load changes.

[0075] In a preferred embodiment, step S1 involves performing data synchronization to obtain synchronized time-series data, specifically including: integrating the sensor operation monitoring data based on the timestamp of each set of data, and outputting the synchronized time-series data of the target wind turbine.

[0076] In practical applications, the sensor acquisition end adds a precise timestamp to each sensor data source when collecting data. Using 1 second as the base period, the interpolation method is used to align data collected in non-1-second intervals to the same time axis, ensuring that all data has a corresponding value at each time step. This achieves time synchronization and integration of multi-source data, avoiding time misalignment caused by differences in acquisition frequency (such as cabin wind speed 0.5 seconds / time, temperature 1 second / time), and ensuring that the subsequent model processes multi-source data at the same time node.

[0077] In a preferred embodiment, step S2, temporal feature extraction, specifically includes:

[0078] S21: The synchronous temporal data is transmitted to a temporal feature extraction model that includes a dilated convolutional network and a CBAM attention module, and the feature map extracted by the temporal feature extraction model is output as a fine-grained short-term feature of the synchronous temporal data.

[0079] In practical applications, the dilated convolutional network contains four layers of dilated convolutions, specifically expressed as follows:

[0080] ;

[0081] In the formula, Let t be the convolution output feature. The kernel size is [size]. The weights of the k-th convolutional kernel are... Given input data at time td×k, the dilation rate of 4 layers of dilated convolution. The sensory field was set to 1, 2, 4, and 8 respectively, gradually expanding to capture temperature change patterns from 1 second to 8 seconds.

[0082] Specifically, considering that the duration of temperature spikes under strong turbulence conditions is on the order of seconds, and is usually accompanied by a rapid drop within 3 seconds, this embodiment designs a dilated convolutional network with four layers of gradient void ratios (d=1, 2, 4, 8). By capturing temperature change patterns across multiple timescales from 1 to 8 seconds, it ensures that no instantaneous anomalies are missed, thus adapting to the short-period characteristics of temperature spikes and drops under strong turbulence conditions. Simultaneously, by introducing void ratio d to insert holes between convolutional kernel elements, the receptive field can be expanded, covering a wider timescale without increasing the kernel size or the number of parameters.

[0083] In practical applications, such as Figure 3 As shown, the CBAM attention module is configured to perform channel attention weighting by global average pooling on the feature map output by the dilated convolutional network and spatial attention weighting by convolutional compression on the feature map after channel attention weighting. The specific expression is as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] In the formula, The attention weight for the Cth channel, It is the Sigmoid activation function. It is a two-layer perceptron, with the hidden layer dimension being half the number of channels. This is the global average pooling result for the C-th channel. This represents the global max pooling result for the C-th channel. The spatial attention weights are H rows and W columns. This represents the convolution operation. The result of global average pooling is H rows and W columns. The result of global max pooling is H rows and W columns. This is a concatenation of the global average pooling result and the global max pooling result. The final feature map after attention weighting. The feature map output by the dilated convolutional network. For channel multiplication, Attention weights that are weighted for channel attention. For space multiplication, Attention weights that are weighted for spatial attention.

[0088] In this embodiment, a CBAM module that integrates channel attention and spatial attention is used to enhance the feature weights of key data channels (such as temperature and load) and the spatial feature weights of temperature peak regions. Then, channel weights and spatial weights are applied to the original feature map through channel multiplication and spatial multiplication to obtain features enhanced in two dimensions. This achieves dual enhancement of key channels and peak regions, ensuring that the output features contain sufficient abnormal information and providing high-quality input for subsequent abnormality probability calculation.

[0089] Specifically, channel attention is introduced to weight the feature channels of different sensor data, highlighting the contributions of key channels such as temperature and generator load, while suppressing interference from secondary channels such as grid voltage and current. This is achieved by capturing statistical information of channel features through global average pooling and global max pooling, performing nonlinear transformation via MLP, and then outputting channel weights using the sigmoid function to enhance key channels. Spatial attention, on the other hand, is introduced to weight temperature abrupt change regions (peak locations) in the feature map, further amplifying peak features. This is achieved by performing global pooling of the spatial dimension of the feature map, concatenating the data, and then compressing it through convolution to generate spatial weights, thus strengthening the features of key spatial locations (such as peak regions). After processing with channel and spatial attention, the CBAM module finally outputs fine-grained short-term features that not only contain multi-scale temperature change patterns but also enhance the details of peak regions, improving the targeting and effectiveness of feature extraction and shortening the response time for subsequent anomaly identification.

[0090] In a preferred embodiment, step S2, sequence memory enhancement, specifically includes:

[0091] S22: Transmit the temporal features obtained by temporal feature extraction to the sequence memory enhancement module, which includes a two-layer bidirectional GRU and a customized local attention encoder, and output the temporal features processed by the sequence memory enhancement module as coarse-grained long-term features of the synchronous temporal data.

[0092] In practical applications, the two-layer bidirectional GRU employs two layers of Bi-GRU, with 128 hidden units in each layer. Long-term dependencies in the temperature sequence are learned through update and reset gates. The specific expression is as follows:

[0093] ;

[0094] In the formula, Let the hidden state be at time t. To update the door, To reset the door, The hidden state at time t-1 For the current input features, This is the weight matrix. For bias terms, This is element-wise multiplication.

[0095] Specifically, a two-layer bidirectional GRU network is employed to simultaneously learn the forward and reverse long-term dependencies of temperature sequences, enabling accurate memorization of historical normal temperature patterns. The bidirectional GRU adaptively retains key information from historical normal temperature patterns (such as the temperature fluctuation range during normal operation) through update and reset gates, while simultaneously learning dependencies in future sequences (such as recovery trends after a sudden temperature drop), providing a long-term basis for anomaly detection. Its operating principle is as follows: based on a gating mechanism, it adaptively adjusts the retention and forgetting of historical states; the update gate controls the retention ratio of historical states, and the reset gate controls the forgetting ratio; candidate states are generated through a tanh function, and finally, the current hidden state is obtained through fusion.

[0096] In practical applications, a customized local attention encoder is specifically expressed as follows:

[0097] ;

[0098] In the formula, Here, m represents the attention weight at time t, and m=3 represents the size of the attention window, designed to enhance local features that allow for a rapid drop in temperature peaks within 3 seconds. The cosine similarity function is used. Let be the local feature sequence m seconds before and after time t. Let be the feature vector of the i-th time point in the local feature sequence m seconds before and after time t.

[0099] Specifically, a customized 3-second local attention window (m=3) is used to focus on local features before and after temperature spikes, strengthening the feature representation of transient abrupt change regions and achieving dual optimization of long-term regularity memory and short-term spike focus. Considering that the features within 3 seconds (m=3) before and after a temperature spike contain complete information about the spike's rise, peak value, and fall, the local attention encoder fixes the attention window to 3 seconds before and after the spike (adapting to the rapid temperature drop). Feature weights within this window are calculated using cosine similarity, allowing the model to focus on temperature abrupt change regions and avoid diluting spike features with long-term normal data. Its operating principle is as follows: the correlation between the current feature and features within the local window is calculated based on cosine similarity, and attention weights are obtained through Softmax normalization. Higher weights indicate a greater impact of the feature on the current moment, allowing the model to focus on the core abnormal region and avoid interference from long-term stable features.

[0100] In this embodiment, a bidirectional GRU is used to distinguish between normal long-term temperature rises and abnormal instantaneous spikes. The spike features are amplified by a local attention window to improve the ability to identify temperature spikes and special patterns of 3-second sudden drops. The output coarse-grained long-term features integrate long-term normal pattern memory and short-term spike feature focus, realizing accurate switching between normal pattern modeling and abnormal pattern recognition. This can solve the problem that traditional memory models cannot simultaneously take into account long-term patterns and short-term mutations, and further reduce the false negative rate.

[0101] In a preferred embodiment, step S2 involves obtaining the fused features of the wind turbine generator, which are a fusion of fine-grained short-term features and coarse-grained long-term features, and then processing these features through a fully connected layer to output the anomaly probability at each time step. Specifically, this includes:

[0102] S23: Construct a two-level feature pyramid by combining fine-grained short-term features and coarse-grained long-term features. Align the number of channels through 1×1 convolution, match the feature size using bilinear interpolation, and then achieve feature fusion through ReLU activation and BatchNorm normalization.

[0103] In practical applications, the expression for feature fusion is as follows:

[0104] ;

[0105] In the formula, As a feature of fusion, It is a fine-grained short-term feature. This is a long-term characteristic of coarse-grained grains. A 1×1 convolution is used to compress the number of channels in P2 from 128 to 64. This is a bilinear interpolation used to enlarge the size of P2 to match that of P1;

[0106] S24: The fused features are input into two fully connected layers, and the anomaly probability at each time step is output through the Sigmoid function; in practical applications, the expression for the anomaly probability is specifically as follows:

[0107] ;

[0108] In the formula, Let be the probability of an anomaly at time t. , For the weights of the fully connected layer, , For bias terms, Use the Sigmoid activation function;

[0109] In practical applications, the expression for the loss function is as follows:

[0110] ;

[0111] In the formula, For the total loss, For sequence length, For balance coefficients, outlier samples normal samples , For focusing coefficient, Let be the probability of an anomaly at time t. For real labels, 1 = abnormal, 0 = normal. For peak penalty weights, For the rate of temperature change, The threshold for determining the peak value. For indicator functions, This is the actual temperature value. To predict temperature values.

[0112] In this embodiment, a feature pyramid structure is first used to fuse fine-grained short-term features (peak details) and coarse-grained long-term features (long-term patterns). Channel numbers are aligned using 1×1 convolution, and size is matched using bilinear interpolation to achieve complementarity between details and patterns. Then, an improved peak enhancement Focal loss function is designed, using a temperature change rate threshold (…). Triggering spike penalty solves the problem of positive and negative sample imbalance caused by a small number of instantaneous spike samples.

[0113] Specifically, during feature fusion, fine-grained short-term features (P1, containing sharp details) and coarse-grained long-term features (P2, containing long-term patterns) are aligned through a 1×1 convolution (compressing the number of channels from 128 to 64), matched in size through bilinear interpolation (scaling the size of P2 to match that of P1), and then normalized using ReLU activation (enhancing nonlinearity) and BatchNorm to output the fused features. This method preserves details such as peak location and amplitude while incorporating long-term temperature patterns. Subsequently, the fused features are input into two fully connected layers and mapped to anomaly probabilities in the [0,1] interval using the Sigmoid function. A higher probability indicates that the temperature deviates more from the normal pattern at the current moment.

[0114] It should be noted that, in this embodiment, during the training process, an improved Focal loss function is designed to handle temperature change rates exceeding [a certain threshold]. An additional squared penalty is applied to the peak region to force the model to improve the peak fitting accuracy. Specifically, this is done by considering the low proportion of instantaneous peak samples under strong turbulence conditions (imbalance between positive and negative samples), and using Focal loss through... The term reduces the weight of normal samples and uses a spike penalty term to achieve this. Item locks the peak area ( As an indicator function, it applies additional penalties only to the peak region (without affecting the fitting of the normal region), increasing the penalty for the prediction error in that region to ensure the model's fitting accuracy to the peak.

[0115] In a preferred embodiment, such as Figure 4 As shown, in step S3, before extracting the operating condition features for each moment from the synchronous time-series data and inputting the real-time feature vector composed of the operating condition features into the trained SVM operating condition classifier to identify the real-time operating condition category of the target wind turbine, the method further includes:

[0116] S31: Extract operating condition features from historical normal synchronization time series data, and divide them into training sets by stratified sampling according to a preset ratio. and verification set ;

[0117] S32: Density peak and K-means clustering are used to cluster the extracted operating conditions. Initial cluster centers are selected by local density and relative distance, and then the sum of squares within the cluster is minimized iteratively to divide the operating conditions into four categories: strong turbulent overload, moderate fluctuation, normal operation, and shutdown.

[0118] S33: For the training set and verification set Assign cluster labels separately and This forms a labeled training dataset. and validation dataset ;

[0119] S34: Construct an SVM condition classifier using the training dataset. Train the SVM condition classifier on the validation dataset. The training performance of the model is evaluated, and the trained SVM condition classifier is saved after the performance meets the target.

[0120] In this embodiment, multi-dimensional operating condition features are extracted from historical normal synchronous time-series data to comprehensively reflect the operating status of wind turbines (e.g., the average nacelle wind speed reflects environmental fluctuations, and the generator load peak factor reflects internal load changes). After dimensionality reduction and redundancy removal, a combined clustering strategy of density peak and K-means clustering is adopted. Initial cluster centers are quickly selected first, and then cluster compactness is optimized to transform unsupervised operating condition classification into supervised classification, improving the accuracy of operating condition identification. Finally, when performance meets the standards (e.g., overall accuracy ≥ 95%, strong turbulent overload condition recall ≥ 93%), the trained model is saved to ensure classification reliability. After this, operating condition features for each moment can be extracted from the synchronous time-series data, and the real-time feature vector composed of these features is input into the trained SVM operating condition classifier to identify the real-time operating condition category of the target wind turbine.

[0121] The combined clustering strategy of density peaks and K-means clustering first quickly selects initial cluster centers, then optimizes cluster compactness, transforming unsupervised classification into supervised classification. This process includes the following steps:

[0122] (1) First, perform density peak clustering initialization:

[0123] Calculate the local density for each sample and relative distance Select The first 5% of samples are used as initial cluster centers to preliminarily divide the work condition clusters. The specific expression is as follows:

[0124] ;

[0125] ;

[0126] In the formula, Let i be the Euclidean distance between samples i and j. To determine the cutoff distance, take the 2nd quantile of the sample distances. Screening for samples with high local density This is an indicator function; it returns 1 when the value is less than or equal to 0. The function ensures that the value is 1 only when the value is less than or equal to 0. When sample j is included in the local density of sample i, it highlights the local clustering of the operating condition cluster. Meanwhile, relative distance... The introduction of this method is used to calculate the minimum distance from sample i to all samples j with a density greater than its own. The density peak sample (with high local density and large relative distance) is used as the cluster center to balance the compactness within the cluster and the separation between the clusters, thus ensuring the accuracy of the working condition division.

[0127] (2) Perform K-means optimization again:

[0128] Using the initial cluster centers as input, iteratively minimize the sum of squares within each cluster. Ultimately, the operating condition was classified as strong turbulent overload. ), medium fluctuations ( ), normal operation ( ), shutdown ( There are 4 categories, and the specific expressions are as follows:

[0129] ;

[0130] In the formula, For the c-th type of working condition sample, For a single data point in the c-th type of working condition sample, is the center vector of the c-th type of working condition.

[0131] In this embodiment, the compactness of clustering is optimized by minimizing the sum of squared distances from samples within a class to the cluster center, making samples of the same working condition more clustered in the feature space. K-means iterative optimization is used to ensure the similarity of samples within each working condition, thereby improving the accuracy of subsequent SVM classification.

[0132] Therefore, this invention provides a precise basis for dynamic threshold adaptation by extracting and combining multi-dimensional operating condition features and clustering them together. It ensures the accuracy of the classification of four types of operating conditions (strong turbulent overload, moderate fluctuation, normal operation, and shutdown). After SVM training and verification, it achieves rapid and accurate identification of real-time operating conditions, providing a reliable basis for subsequent threshold switching.

[0133] In a preferred embodiment, step S3, dynamically outputting the condition-specific threshold corresponding to the real-time operating condition category, specifically includes:

[0134] S35: Based on the four types of working condition labels formed by the division of working conditions, the abnormal probability values ​​obtained by feature extraction, fusion and abnormal probability calculation of historical normal synchronization time series data are divided into four abnormal probability subsets, and each abnormal probability subset corresponds to one type of working condition.

[0135] S36: Perform flow interruption-adaptive kernel density estimation on each subset of anomaly probabilities to calculate the distribution density of anomaly probabilities for various operating conditions; wherein, the expression for the distribution density is specifically:

[0136] ;

[0137] In the formula, Let c be the probability density function of the abnormal probability of the c-th type of working condition. This represents the number of normal samples for condition c. For turbulence-adaptive bandwidth, Let be the wind speed variation coefficient for the c-th operating condition. For wind speed standard deviation, This is the average wind speed. For standard Gaussian kernel function, Let be any one of the anomaly probability values ​​to be evaluated under the c-th type of working condition. Let be the probability of an anomaly in the i-th normal sample under the c-th type of working condition;

[0138] S37: Take the probability density function of each type of abnormal working condition. The 99.5th percentile was used as the specific threshold for this type of working condition, forming a set of specific thresholds for four types of working conditions. .

[0139] In this embodiment, a turbulence-adaptive kernel density estimation is designed, incorporating the nacelle wind speed variation coefficient and a dynamically adjusted bandwidth to adapt to the discrete distribution of anomaly probabilities under strong turbulence conditions. Anomaly probabilities are divided into subsets according to operating condition categories, and a specific threshold with a 99.5% quantile is calculated for each category, achieving dynamic adaptation of a threshold for each operating condition. Specifically, considering the different fluctuations in normal temperature under different operating conditions (e.g., large fluctuations and more discrete anomaly probability distributions under strong turbulence conditions), a Gaussian kernel function is used to smooth the anomaly probability samples, estimating their probability density distribution. The quantile is determined based on this distribution, ensuring that the threshold adapts to the normal distribution characteristics of this type of operating condition. The turbulence-adaptive bandwidth... middle, The standard deviation of the subset of outlier probabilities. Used to balance the impact of sample size on bandwidth It can dynamically adapt to the degree of fluctuation in operating conditions, even under strong turbulent conditions. Larger bandwidth, adaptable to discrete distributions.

[0140] Therefore, considering that the probability distribution of anomalies is more discrete under strong turbulent conditions, this invention ensures accurate threshold calculation by dynamically adjusting the bandwidth, completely solving the problem of poor adaptability caused by the use of fixed thresholds in traditional methods, realizing threshold adaptation under different operating conditions, and avoiding the problem of high false alarm rate caused by using the same threshold to judge all operating conditions.

[0141] In a preferred embodiment, step S4, based on the synchronized time series data, calculates the operational stability index for each moment, specifically including:

[0142] S41: Based on the sequence of nacelle wind speed and the sequence of generator load in the synchronous time series data, the operating condition stability index is obtained by weighted fusion calculation of wind speed variation coefficient and load fluctuation coefficient.

[0143] The expression for the operating condition stability index is as follows:

[0144] ;

[0145] In the formula, , These represent the standard deviation and mean of the real-time cabin wind speed, respectively. , These represent the standard deviation and mean of the real-time generator load, respectively. The coefficient of variation of wind speed. For load fluctuation coefficient, The larger the value, the more stable the operating conditions.

[0146] In this embodiment, a weighted fusion method using the nacelle wind speed variation coefficient and the generator load fluctuation coefficient is employed to quantify the fluctuation level of the current operating condition, providing key evidence from the operating condition dimension for multi-source information fusion. Specifically, the nacelle wind speed sequence and generator load sequence are extracted from real-time synchronized time-series data, and the nacelle wind speed variation coefficient is calculated. The generator load fluctuation coefficient reflects the intensity of external wind field fluctuations. This reflects the intensity of load changes within the unit. By defining the weights of nacelle wind speed and generator load, and taking wind speed fluctuations as the main cause of temperature anomalies under strong turbulence conditions, a weighted stability index is obtained. , The larger the value, the more stable the operating conditions; conversely, the smaller the value, the more drastic the fluctuations.

[0147] In a preferred embodiment, step S4 involves generating a wind turbine temperature warning signal by combining the anomaly probability and the operating condition-specific threshold, specifically including:

[0148] S42: Based on the calculated operating condition stability index, combined with the anomaly probability and operating condition specific threshold, a weighted summation method is used to calculate the comprehensive anomaly score. The specific expression is as follows:

[0149] ;

[0150] In the formula, To calculate the overall abnormal score, This is the specific threshold for the c-th type of operating condition;

[0151] S43: Based on the comprehensive anomaly score and the preset mapping range between the wind turbine temperature warning level and the comprehensive anomaly score, generate a wind turbine temperature warning signal that includes the wind turbine temperature warning level.

[0152] In this embodiment, three core data points are integrated: the ratio of anomaly probability to a threshold, the inverse value of the operating condition stability index, and the relative strength of the anomaly probability. A weighted summation is used to fuse these three different dimensions of evidence into a single score, preserving the complementarity of multi-source information. Specifically, the ratio of anomaly probability to a threshold directly reflects the degree to which the anomaly exceeds the baseline (as the core); the inverse value of the operating condition stability index reflects the intensity of the anomaly's cause (as an auxiliary factor); and the relative strength of the anomaly probability highlights the degree to which the anomaly probability exceeds 0.5 (as a supplement). The fusion of these three factors ensures the comprehensiveness and reliability of the judgment.

[0153] In practical applications, for example, based on the comprehensive anomaly score... The range of values ​​is used to generate warning signals. When it falls within [0.6, 0.8), it is a Level 1 warning; when it falls within [0.8, 0.9), it is a Level 2 warning; and when it is ≥0.9, it is a Level 3 severe warning.

[0154] Therefore, this invention proposes a wind turbine temperature early warning method based on hybrid deep learning and dynamic thresholding. By introducing hybrid deep learning and dynamic thresholding technology, short-term peak detail features and long-term dependency features are fused to retain peak position details and increase the penalty for prediction errors in peak areas, forcing the model to improve peak fitting accuracy and ensuring that the anomaly probability can accurately reflect the degree of temperature anomaly. At the same time, it can consider the differences in anomaly probability distribution under different operating conditions, realize adaptive dynamic adjustment of thresholds under different operating conditions, reduce the false alarm rate under strong turbulence conditions, and provide accurate decision support for operation and maintenance personnel.

[0155] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0156] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0157] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for early warning of wind turbine temperature based on hybrid deep learning and dynamic thresholding, characterized in that, include: S1: Acquire sensor operation monitoring data of the target wind turbine, perform preprocessing and integration on the sensor operation monitoring data, and obtain synchronous time series data; S2: Construct a hybrid deep learning anomaly detection model, extract temporal features and enhance sequence memory on the synchronous time-series data, and output the anomaly probability at each time step; The hybrid deep learning anomaly detection model includes: Dilated Convolutional Network: Extracts short-term multi-timescale features containing temperature spikes from synchronous temporal data to obtain the first feature map; CBAM Attention Module: Extracts a second feature map from the first feature map, fusing multi-scale temperature and spike region details; Bidirectional GRU: Extracts long-term trend and normal pattern features of temperature changes from the second feature map to obtain the third feature map; Customized local attention encoder: Extracts a fourth feature map from the third feature map that fuses long-term normal patterns and short-term spikes; Feature fusion layer: fuses the second feature map and the fourth feature map to obtain fused features; Fully connected layer: outputs the anomaly probability at each time step based on the fused features; Temporal feature extraction specifically includes: S21: The synchronous temporal data is transmitted to a temporal feature extraction model that includes a dilated convolutional network and a CBAM attention module, and the feature map extracted by the temporal feature extraction model is output as a fine-grained short-term feature of the synchronous temporal data. The dilated convolutional network comprises four layers of dilated convolutions, specifically expressed as follows: ； In the formula, Let t be the convolution output feature. The kernel size is [size]. The weights of the k-th convolutional kernel are... Input data for time td×k. The void ratio of the dilated convolution; The CBAM attention module is configured to perform channel attention weighting by global average pooling on the feature map output by the dilated convolutional network and spatial attention weighting by convolutional compression on the channel attention-weighted feature map. The specific expression is as follows: ； ； ； In the formula, The attention weight for the C-th channel. It is the Sigmoid activation function. It is a two-layer perceptron. This represents the global average pooling result for the C-th channel. This represents the global max pooling result for the C-th channel. The spatial attention weights are H rows and W columns. This represents the convolution operation. The result of global average pooling is H rows and W columns. The result of global max pooling is H rows and W columns. This is a concatenation of the global average pooling result and the global max pooling result. The final feature map after attention weighting. The feature map output by the dilated convolutional network. For channel multiplication, Attention weights that are weighted for channel attention. For space multiplication, Attention weights that are weighted for spatial attention; Specifically, sequence memory enhancement includes: S22: Transmit the temporal features obtained by temporal feature extraction to the sequence memory enhancement module, which includes a two-layer bidirectional GRU and a customized local attention encoder, and output the temporal features processed by the sequence memory enhancement module as coarse-grained long-term features of the synchronous temporal data. The two-layer bidirectional GRU employs a two-layer Bi-GRU architecture, with 128 hidden units in each layer. It learns the long-term dependencies of the temperature sequence through update and reset gates, as specifically expressed in the following expression: ； In the formula, Let the hidden state be at time t. To update the door, To reset the door, The hidden state at time t-1 For the current input features, This is the weight matrix. For bias terms, Element-wise multiplication; The customized local attention encoder is specifically expressed as follows: ； In the formula, Here, m represents the attention weight at time t, and m=3 represents the size of the attention window, designed to enhance local features that allow for a rapid drop in temperature peaks within 3 seconds. The cosine similarity function is used. Let be the local feature sequence m seconds before and after time t. Let be the feature vector at the i-th time point in the local feature sequence m seconds before and after time t; S3: Extract the operating condition features at each moment from the synchronous time series data, input the real-time feature vector composed of the operating condition features into the trained SVM operating condition classifier, identify the real-time operating condition category of the target wind turbine, query the abnormal probability subset corresponding to the same operating condition category divided under the historical normal synchronous time series data, and take the abnormal probability value of the abnormal probability subset at the preset proportional quantile of the distribution density function as the operating condition specific threshold of the operating condition category. The generation of condition-specific thresholds specifically includes: S35: Based on the four types of condition labels formed by the condition classification, the abnormal probability values ​​obtained from feature extraction, fusion, and abnormal probability calculation of historical normal synchronization time series data are divided into four abnormal probability subsets, each corresponding to one type of condition; S36: Perform interruption-adaptive kernel density estimation on each abnormal probability subset to calculate the distribution density of abnormal probabilities for each type of condition; S37: Take the 99.5th percentile of the density function of the abnormal probability for each type of condition as the specific threshold for that type of condition, forming a set of specific thresholds for the four types of conditions. ; S4: Based on the synchronous time series data, calculate the operating condition stability index at each moment, and combine the anomaly probability and the operating condition specific threshold to generate a wind turbine temperature early warning signal.

2. The wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold as described in claim 1, characterized in that, In step S1, the sensor operation monitoring data is configured as a combination of front and rear bearing temperatures, stator and rotor winding temperatures, fan speed, generator load, nacelle wind speed, nacelle vibration, and grid voltage and current.

3. The wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold as described in claim 2, characterized in that, In step S1, the integration process after preprocessing the sensor operation monitoring data is performed, which specifically includes: based on the timestamp of each set of data, the sensor operation monitoring data is integrated and synchronized to output the synchronous timing data of the target wind turbine.

4. The wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold as described in claim 1, characterized in that, In step S2, the anomaly probability at each time step is output, specifically including: S23: Construct a two-level feature pyramid by combining fine-grained short-term features and coarse-grained long-term features. Align the number of channels through 1×1 convolution, match the feature size using bilinear interpolation, and then achieve feature fusion through ReLU activation and BatchNorm normalization. S24: Input the fused features into two fully connected layers, and output the anomaly probability at each time step using the Sigmoid function; the expression for the loss function is as follows: ； In the formula, For the total loss, For sequence length, For balance coefficients, outlier samples normal samples , For focusing coefficient, Let be the probability of an anomaly at time t. For real labels, 1 = abnormal, 0 = normal. For peak penalty weights, For the rate of temperature change, The threshold for determining the peak value. For indicator functions, This is the actual temperature value. To predict temperature values.

5. The wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold as described in claim 4, characterized in that, In step S3, before extracting the operating condition features for each moment from the synchronous time-series data and inputting the real-time feature vector composed of the operating condition features into the trained SVM operating condition classifier to identify the real-time operating condition category of the target wind turbine, the method further includes: S31: Extract operating condition features from historical normal synchronization time series data, and divide them into training sets by stratified sampling according to a preset ratio. and verification set ; S32: Density peak and K-means clustering are used to cluster the extracted operating conditions. Initial cluster centers are selected by local density and relative distance, and then the sum of squares within the cluster is minimized iteratively to divide the operating conditions into four categories: strong turbulent overload, moderate fluctuation, normal operation, and shutdown. S33: For the training set and verification set Assign cluster labels separately and This forms a labeled training dataset. , ) and validation dataset ( , ); S34: Construct an SVM condition classifier using the training dataset. Train the SVM condition classifier on the validation dataset. The training performance of the model is evaluated, and the trained SVM condition classifier is saved after the performance meets the target.

6. The wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold as described in claim 5, characterized in that, In step S4, based on the synchronized time series data, the operational stability index for each time moment is calculated, specifically including: S41: Based on the sequence of nacelle wind speed and the sequence of generator load in the synchronous time series data, the operating condition stability index is obtained by weighted fusion calculation of wind speed variation coefficient and load fluctuation coefficient. The expression for the operating condition stability index is as follows: ； In the formula, , These represent the standard deviation and mean of the real-time cabin wind speed, respectively. , These represent the standard deviation and mean of the real-time generator load, respectively. The coefficient of variation of wind speed. For load fluctuation coefficient, The larger the value, the more stable the operating conditions.

7. The wind turbine temperature early warning method based on hybrid deep learning and dynamic threshold as described in claim 6, characterized in that, In step S4, a wind turbine temperature warning signal is generated by combining the anomaly probability and the operating condition-specific threshold, specifically including: S42: Based on the calculated operating condition stability index, combined with the anomaly probability and operating condition specific threshold, a weighted summation method is used to calculate the comprehensive anomaly score. The specific expression is as follows: ； In the formula, To calculate the overall abnormal score, This is the specific threshold for the c-th type of operating condition; S43: Based on the comprehensive anomaly score and the preset mapping range between the wind turbine temperature warning level and the comprehensive anomaly score, generate a wind turbine temperature warning signal that includes the wind turbine temperature warning level.

Images