Wind turbine imbalance detection model training method, device, equipment and medium
By constructing a wind turbine imbalance detection model and using LSTM, skip connections, and attention mechanisms for feature extraction and weighting, the problem of poor adaptability of existing detection methods is solved, and more efficient imbalance detection is achieved.
Patent Information
- Application Number
- CN202310696902.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-13
- Publication Date
- 2026-02-13
- Estimated Expiration
- 2043-06-13
AI Technical Summary
Existing wind turbine imbalance detection methods are difficult to adapt to differences in different models, tower heights, and installation methods, resulting in poor adaptability of threshold alarm methods and neglect of time-domain changes, leading to frequent blade sweeping or tower collapse events.
By employing LSTM, skip connections, and two attention mechanisms, a wind turbine imbalance detection model is constructed through feature extraction and weighting operations on the training data. Features are extracted using channel and spatial attention mechanisms, and temporal correlation extraction and connection are performed using a long short-term memory network. The classification loss is then calculated to adjust the model parameters.
It improves the accuracy and efficiency of wind turbine imbalance detection, solves the problems of poor threshold adaptability and neglect of time domain changes, and achieves more accurate imbalance detection.
Smart Images
Figure CN116629313B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of wind power, in particular to a wind turbine imbalance detection model training method, device, equipment and medium. BACKGROUND
[0002] At present, the lower degree of electricity cost target drives wind turbines to develop continuously towards large-scale, and larger units face greater security challenges while showing superior power generation performance.
[0003] Due to the continuous increase of impeller diameter, the consequences of blade and impeller imbalance are becoming more and more serious, and the lack of effective operation health monitoring means also makes blade tower scanning and even tower overturning events occur frequently.
[0004] The existing wind turbine imbalance detection is mainly realized by comparing the amplitude value at 1P frequency in the amplitude-frequency curve of the unit operation. When the unit imbalance occurs, the amplitude value at 1P frequency can be observed to be much larger than the normal level. It is worth noting that due to the differences in unit type, tower height, installation method, installation site and other conditions, the basic amplitude value and imbalance amplitude value change rate of each unit at 1P frequency are different, which need to be combined with the change rule of the signal to be judged, which also makes it difficult for the simple threshold alarm method to be popularized. SUMMARY
[0005] Therefore, the purpose of the present application is to provide a wind turbine imbalance detection model training method, device, equipment and medium, which can effectively extract the features of the training data by using LSTM, skip connection and two attention mechanisms, so as to train a model that can better detect the imbalance of the wind turbine, thereby improving the accuracy of the wind turbine imbalance detection. The specific scheme is as follows:
[0006] In the first aspect, the present application provides a wind turbine imbalance detection model training method, comprising:
[0007] inputting the training data into the wind turbine imbalance detection model, and grouping the training data in time domain to obtain a preset number of groups of time domain data; the training data is a group of data containing multiple input signals collected in a preset time period;
[0008] performing feature extraction on each group of time domain data respectively, and performing channel dimension weighting operation on each group of extracted features respectively by using channel attention mechanism to obtain corresponding each group of weighted features;
[0009] performing time domain correlation extraction and connection operation on each group of weighted features by using long short-term memory network and skip connection to obtain connection features, and performing spatial dimension weighting operation on the connection features by using spatial attention mechanism to obtain weighted connection features;
[0010] A classification loss is calculated based on the predicted value of the weighted connection feature, and the model parameters are adjusted using the classification loss to complete the training operation of the wind turbine imbalance detection model.
[0011] Optionally, the training data is a set of data collected within a preset time period under normal power generation conditions, including the amplitude corresponding to the tower frequency, the 1P frequency, the 6P frequency, and the generator speed.
[0012] Optionally, the feature extraction of each group of time domain data includes:
[0013] Each group of time domain data is input into each feature extraction module to perform convolution, regularization and activation operations on each group of time domain data to obtain corresponding extracted features.
[0014] Optionally, the channel attention mechanism is used to perform weighting operation on the channel dimension of each group of extracted features to obtain corresponding weighted features, including:
[0015] The target group of extracted features is input into a channel attention module, and average pooling and maximum pooling operations are performed on the target group of extracted features to obtain corresponding first average pooling results and first maximum pooling results; the target group of extracted features is any one of the groups of extracted features.
[0016] The multi-layer perception is used to perform channel importance learning based on the first average pooling results and the first maximum pooling results to obtain channel importance weights.
[0017] Each channel importance weight is used to perform weighting operation on the corresponding feature in the target group of extracted features to obtain the corresponding target group of weighted features.
[0018] Optionally, the long short-term memory network and the skip connection are used to perform time domain correlation extraction and connection operation on each group of weighted features to obtain connection features, including:
[0019] Each group of weighted features is input into each long short-term memory network to perform time domain correlation extraction on each group of weighted features to obtain corresponding output results.
[0020] According to a preset skip connection rule, the output results corresponding to the first N long short-term memory networks are connected to obtain a connection result; the first N long short-term memory networks are other long short-term memory networks except the last long short-term memory network in the plurality of long short-term memory networks.
[0021] The connection result is dimensionally reduced by using a convolutional layer to obtain a reduced connection result, and the output result corresponding to the last long short-term memory network and the reduced connection result are connected to obtain a connection feature.
[0022] Optionally, the spatial dimension weighting operation on the connection feature by using the spatial attention mechanism comprises:
[0023] The connection feature is input into a spatial attention module to perform average pooling and maximum pooling operations on the connection feature respectively to obtain a corresponding second average pooling result and a second maximum pooling result.
[0024] The second average pooling result and the second maximum pooling result are spliced to obtain a spliced result.
[0025] The spliced result is activated by using an activation function to obtain a spatial attention weight, and the connection feature is weighted by using the spatial attention weight to obtain a weighted connection feature.
[0026] Optionally, the training operation of the wind turbine imbalance detection model is completed by using a loss function and calculating a classification loss based on a predicted value of the weighted connection feature, and adjusting model parameters by using the classification loss, comprising:
[0027] The weighted connection feature is fully connected by using a fully connected layer to obtain a fully connected feature.
[0028] A detection label corresponding to the fully connected feature and a corresponding predicted value are determined by using a softmax layer, and a classification loss is obtained by calculating the predicted value and a true value by using a cross-entropy loss function.
[0029] The training operation of the wind turbine imbalance detection model is completed by adjusting the model parameters by using the classification loss.
[0030] In a second aspect, the present application provides a wind turbine imbalance detection model training device, comprising:
[0031] A time domain grouping module is configured to input training data into a wind turbine imbalance detection model and group the training data in a time domain to obtain a preset number of groups of time domain data; the training data is a group of data containing multiple input signals collected in a preset time period.
[0032] A channel attention module is configured to extract features from each group of time domain data, and perform a channel dimension weighting operation on each group of extracted features by using a channel attention mechanism to obtain corresponding weighted features of each group.
[0033] a spatial attention module configured to perform time domain correlation extraction and connection operation on each group of the weighted features by using a long short-term memory network and a skip connection to obtain connection features, and perform weighting operation on the connection features in a spatial dimension by using a spatial attention mechanism to obtain weighted connection features;
[0034] a model training completion module configured to calculate a classification loss based on a predicted value of the weighted connection features by using a loss function, and adjust model parameters by using the classification loss to complete training of the wind turbine imbalance detection model.
[0035] In a third aspect, the present application provides an electronic device, comprising:
[0036] a memory configured to save a computer program;
[0037] a processor configured to execute the computer program to implement the wind turbine imbalance detection model training method described above.
[0038] In a fourth aspect, the present application provides a computer readable storage medium configured to save a computer program, wherein the computer program is executed by a processor to implement the wind turbine imbalance detection model training method described above.
[0039] In the present application, training data is input into a wind turbine imbalance detection model, and the training data is grouped in the time domain to obtain a preset number of groups of time domain data; the training data is a group of data collected within a preset time period and containing multiple input signals; feature extraction is performed on each group of time domain data, and a channel attention mechanism is used to perform weighting operation on the channel dimension of each group of extracted features to obtain the corresponding weighted features; a long short-term memory network and a skip connection are used to perform time domain correlation extraction and connection operation on each group of weighted features to obtain connection features, and a spatial attention mechanism is used to perform weighting operation on the spatial dimension of the connection features to obtain weighted connection features; a loss function is used to calculate a classification loss based on the predicted value of the weighted connection features, and the model parameters are adjusted using the classification loss to complete the training operation of the wind turbine imbalance detection model. As can be seen, the present application extracts features from the training data and learns the importance of the channel using the channel attention mechanism, thereby assigning a higher weight to the dominant feature through the weighting operation in the channel dimension; the long short-term memory network is used to mine the time domain correlation between features, and the skip connection is introduced to reduce the influence of noise on the output results of the long short-term memory network while dealing with the information ablation caused by the long short-term memory network; then the spatial feature learning is performed through the spatial attention mechanism, thereby fully extracting the dominant feature through the weighting operation in the spatial dimension, so that the prediction result of the weighted connection features is more accurate, effectively solving the problems of poor adaptability and ignoring the time domain changes caused by setting the threshold; finally, the model parameters are adjusted by calculating the classification loss, and a better model for detecting the imbalance of the wind turbine can be obtained, thereby improving the accuracy and efficiency of the wind turbine imbalance detection. BRIEF DESCRIPTION OF DRAWINGS
[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.
[0041] Figure 1 The present application discloses a wind turbine imbalance detection model training method flow chart;
[0042] Figure 2 The present application discloses a specific wind turbine imbalance detection model training method flow chart;
[0043] Figure 3 The present application discloses an LSTM input-output structure diagram;
[0044] Figure 4 This application discloses a flowchart for training a wind turbine imbalance detection model;
[0045] Figure 5 This application discloses a schematic diagram of a wind turbine unbalance detection model training device.
[0046] Figure 6 This application discloses a structural diagram of an electronic device. Detailed Implementation
[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0048] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for training a wind turbine imbalance detection model, including:
[0049] Step S11: Input the training data into the wind turbine imbalance detection model, and group the training data in the time domain to obtain a preset number of time domain data sets; the training data is a set of data containing multiple input signals collected within a preset time period.
[0050] In this embodiment, the dataset consists of data collected under conditions involving multiple wind turbine units and multiple sites, ensuring broad data coverage. The dataset includes a training set and a validation set, both of which contain normal operation groups, normal-unbalanced operation groups, and unbalanced operation groups. The data is labeled with 0 and 1, where label 0 represents normal operation results and label 1 represents unbalanced operation results. The training data in the training set consists of second-level data covering 200 minutes, collected under normal power generation conditions, and includes four input signals: amplitude corresponding to tower frequency, 1P frequency, 6P frequency, and generator speed.
[0051] In this embodiment, training data is input into the wind turbine imbalance detection model through four channels, and the training data is divided into four equal parts in the time domain. Referring to a data point length of 200*60, each group of data after four divisions has 200 / 4*60 = 3000 data points, resulting in four groups of time-domain data. The wind turbine imbalance detection model uses the Adam optimizer with a learning rate of 0.001, beta = (0.9, 0.999), eps = 1e-8, weight_decay = 0, and an input batch size of 8.
[0052] Step S12, feature extraction is performed on each group of time domain data respectively, and a channel attention mechanism is used to perform weighting operation on the channel dimension of each group of extracted features, to obtain corresponding weighted features of each group.
[0053] In this embodiment, the feature extraction performed on each group of time domain data can include inputting each group of time domain data into a respective feature extraction module, to perform convolution, regularization and activation operation on each group of time domain data, to obtain corresponding extracted features of each group. It can be understood that the four groups of time domain data are respectively input into four feature extraction modules, each of which includes three convolution layers, and each convolution layer sequentially performs 3x3 convolution, regularization and ReLU activation function operation, and the output channel of each feature extraction module remains four channels. In addition, the wind turbine imbalance feature will appear in each part of a long time domain signal sequence, so the four feature extraction modules are completely identical in structure and all have the function of extracting imbalance features. Further, the four groups of time domain data are respectively input into the four feature extraction modules, so that the four groups of time domain data are sequentially passed through the three convolution layers in the respective corresponding feature extraction modules, and sequentially perform convolution, regularization and activation operation, to obtain four groups of extracted features. The formula of each convolution layer is as follows:
[0054] f out =ReLU(Bn(conv(f in )));
[0055] Wherein, f in represents input data, Bn represents regularization operation, conv in this embodiment adopts 3x3 convolution, and ReLu represents activation function.
[0056] In this embodiment, the channel attention mechanism is used to weight the post-extraction features of each group in the channel dimension to obtain the corresponding weighted features of each group. This can include inputting the target group post-extraction feature into a channel attention module, and performing average pooling and maximum pooling operations on the target group post-extraction feature to obtain the corresponding first average pooling result and first maximum pooling result. The target group post-extraction feature is any one of the post-extraction features of the several groups. Channel importance learning is performed using a multi-layer perceptron based on the first average pooling result and the first maximum pooling result to obtain channel importance weights. The corresponding target group weighted feature is obtained by weighting the corresponding feature in the target group post-extraction feature using each channel importance weight. It can be understood that, considering that the influence of the features extracted from the four input signals on the wind turbine imbalance judgment is different, after the feature extraction is completed, the four groups of post-extraction features extracted by the four feature extraction modules are input into the same channel attention module for weighting operation, so that the network can learn to assign higher weights to the dominant signals. Since the features extracted by each feature extraction module have the same channel distribution rule, the same channel attention module can be used for processing. Specifically, for any one of the four groups of post-extraction features, i.e., the target group post-extraction feature, first perform average pooling and maximum pooling operations on the target group post-extraction feature to obtain the average pooling result and the maximum pooling result. Then, channel importance learning is performed on the average pooling result and the maximum pooling result using a multi-layer perceptron (MLP, Multi-Layer Perceptron), and then the two MLP output results are added and activated to obtain the final channel importance weights. Thus, the corresponding feature in the target group post-extraction feature is weighted using each channel importance weight to obtain the target group weighted feature. The formula is as follows:
[0057] F output =F input ×Att c
[0058] Att c =sigmoid(MLP(Avgpool(F input ))+MLP(Maxpool(F input )))
[0059] MLP=f(W T x+b);
[0060] wherein, F input represents the input feature; F output represents the output feature, and Att crepresents the channel importance weight; Avgpool and Maxpool represent average pooling and maximum pooling operations respectively; W represents the weight to be learned; b represents the bias; sigmoid represents the activation function; x represents the input of the MLP multi-layer perceptron; it should be noted that the initial weight to be learned and the initial bias are determined based on prior knowledge, and are adjusted through the subsequent classification loss calculated by the loss function.
[0061] In the embodiment, if the feature of the target group after extraction is CxHxW pixel numbers, C represents the number of channels, H represents the height of each channel, and W represents the width of each channel, the spatial size can be compressed through the pooling operation, which is convenient for subsequent learning of the channel feature, so that the feature of the target group after extraction becomes Cx1x1 format, then the importance of each channel is learned through the multi-layer perceptron, so as to weight each feature in the feature of the target group after extraction by using the importance weight of each channel, obtain the weighted feature of the target group, and input it to the LSTM (Long Short-Term Memory, Long Short-Term Memory) module.
[0062] In step S13, the weighted features of each group are subjected to time domain correlation extraction and connection operation by using the long short-term memory network and the skip connection to obtain the connection feature, and the connection feature is subjected to the weighting operation in the spatial dimension by using the spatial attention mechanism to obtain the weighted connection feature.
[0063] In the embodiment, the long short-term memory network is used to extract the time domain correlation of each group of weighted features, and the skip connection is used to connect the output results corresponding to each group of weighted features respectively to obtain the connection feature.
[0064] In this embodiment, the operation of weighting the connection features in the spatial dimension by using the spatial attention mechanism to obtain the weighted connection features can include inputting the connection features into a spatial attention module to perform average pooling and maximum pooling operations on the connection features respectively to obtain corresponding second average pooling results and second maximum pooling results; splicing the second average pooling results and the second maximum pooling results to obtain a splicing result; using an activation function to activate the splicing result to obtain a spatial attention weight, and using the spatial attention weight to weight the connection features to obtain the weighted connection features. It can be understood that, since the occurrence of the wind turbine imbalance feature in a length of feature data has uncertainty, the connection features are input into the spatial attention module, and the average pooling and maximum pooling operations are performed on the connection features to obtain the average pooling results and the maximum pooling results, that is, the channel size is compressed through the pooling operation to facilitate subsequent learning of spatial features. Then, the average pooling results and the maximum pooling results are spliced to obtain a splicing result with a channel number of 2; the splicing result is activated by an activation function to obtain a spatial attention weight, so that the connection features are weighted by using the spatial attention weight to obtain the weighted connection features. The formula involved is as follows:
[0065] F output =F input ×Att p
[0066] Att p =sigmoid(cat(Avgpool(F input ),Maxpool(F input )));
[0067] wherein, F input represents input features; F output represents output features, Att p represents a spatial attention weight; Avgpool and Maxpool represent average pooling and maximum pooling operations respectively; cat represents a splicing operation; sigmoid represents an activation function.
[0068] In step S14, a classification loss is calculated by using a loss function and based on a predicted value of the weighted connection features, and model parameters are adjusted by using the classification loss to complete the training operation of the wind turbine imbalance detection model.
[0069] In this embodiment, the use of the loss function and based on the predicted value of the weighted connection feature calculates the classification loss, and adjusts the model parameters using the classification loss to complete the training operation of the wind turbine imbalance detection model, which can include using a fully connected layer to fully connect the weighted connection feature to obtain a fully connected feature; determine the detection label corresponding to the fully connected feature and the corresponding predicted value through the softmax layer, calculate the predicted value and the true value using the cross-entropy loss function to obtain the classification loss; adjust the model parameters using the classification loss to complete the training operation of the wind turbine imbalance detection model. It can be understood that the weighted connection feature is first fully connected through the fully connected layer to obtain the fully connected feature; then the 0,1 label corresponding to the fully connected feature and the corresponding predicted value are determined through the softmax layer; the cross-entropy loss function loss bce The predicted value T(x) and the true value G(x) are calculated to obtain the classification loss, so that the classification loss is used for back propagation to complete the training operation of the wind turbine imbalance detection model. The loss function formula is as follows:
[0070]
[0071] Where, loss bce represents the cross-entropy loss function; T(x) represents the predicted value; G(x) represents the true value; x represents the input signal.
[0072] As can be seen, the present application extracts features from training data and uses channel attention mechanism to learn channel importance, so as to assign higher weights to dominant features through channel dimension weighting operation; use long short-term memory network to mine the time domain correlation between features, and introduce jump connection to reduce the influence of noise on long short-term memory network output results while dealing with the information ablation caused by long short-term memory network; then learn spatial features through spatial attention mechanism, so as to fully extract dominant features through spatial dimension weighting operation, so that the prediction result of the weighted connection feature is more accurate, effectively solving the problem of poor adaptability and ignoring time domain changes caused by setting threshold; finally, adjust the model parameters by calculating the classification loss, so as to obtain a model that can better detect the imbalance of the wind turbine, thereby improving the accuracy and efficiency of the wind turbine imbalance detection.
[0073] Based on the previous embodiment, the present application describes the overall training process of the wind turbine imbalance detection model. Next, the present application will describe in detail how to use long short-term memory network and jump connection to extract and connect the time domain correlation of each group of weighted features. Referring to Figure 2As shown, the embodiment of the present application discloses a process of time domain correlation extraction and skip connection, comprising:
[0074] Step S21, input each group of weighted features into each long short-term memory network respectively to extract time domain correlation of each group of weighted features respectively, and obtain corresponding output results.
[0075] In this embodiment, the input and output of the long short-term memory network are as shown in the formula (1). Figure 3 As shown, C t-1 , C t are input neuron state and output neuron state respectively, h t-1 , h t are neuron hidden layer input and neuron hidden layer output respectively, X t-1 is input feature, and t represents time. The neuron state is affected by the input feature and the neuron hidden layer input through the Sigmoid activation function and the tanh activation function inside the long short-term memory network, and significantly affects the feature gating between neurons, realizing the context correlation degree of the feature in the time domain. Specifically, the four groups of weighted features are input into four long short-term memory networks respectively, the first group of weighted features is extracted by the first long short-term memory network to obtain the first output result, and the first output result is input into the second long short-term memory network, and the second group of weighted features is extracted by the second long short-term memory network to obtain the second output result. The output of the former guides the feature mining of the latter, so as to obtain the second output result. In this way, four output results corresponding to the four long short-term memory networks respectively can be obtained.
[0076] Step S22, according to a preset skip connection rule, connect the output results corresponding to the first N long short-term memory networks respectively to obtain a connection result; the first N long short-term memory networks are other long short-term memory networks except the last long short-term memory network in the long short-term memory networks.
[0077] Step S23, use a convolution layer to perform dimension reduction processing on the connection result to obtain a reduced connection result, and connect the output result corresponding to the last long short-term memory network and the reduced connection result to obtain a connection feature.
[0078] In this embodiment, the skip connection refers to the stacking of the channel level. Specifically, first, the first output result, the second output result and the third output result corresponding to the first long short-term memory network, the second long short-term memory network and the third long short-term memory network respectively are connected to obtain a connection result. The connection result is reduced to four channels through a convolution layer to obtain a reduced connection result. Then, the fourth output result corresponding to the last long short-term memory network, i.e., the fourth long short-term memory network, is connected with the reduced connection result to obtain the final connection feature.
[0079] As can be seen, the application uses the long short-term memory network to mine the time domain correlation between features, and introduces the skip connection to reduce the influence of noise on the output result of the long short-term memory network while coping with the information ablation caused by the long short-term memory network. At the same time, the output result of the last long short-term memory network fuses the shallow feature information, effectively removes the noise caused by insufficient time domain information in the shallow layer, and avoids information loss.
[0080] Referring to Figure 4 The embodiment of the application discloses a wind turbine imbalance detection model training method, comprising:
[0081] A set of training data containing four input signals is input into the wind turbine imbalance detection model, and the input signal set, i.e., the training data is divided into four equal parts in the time domain to obtain four sets of time domain data. Four Feature extract modules (feature extraction modules) are used to extract features from the four sets of time domain data to obtain four sets of extracted features. The four sets of extracted features are input into the same Channel-wise attention module (channel attention module) to perform channel dimension weighting operation on the four sets of extracted features to obtain four sets of weighted features.
[0082] The four sets of weighted features are input into four LSTM modules (long short-term memory networks) respectively. The first LSTM module is used to extract the time domain correlation of the first set of weighted features to obtain a first output result, and the first output result is input into the second LSTM module. The output of the previous step guides the feature mining of the next step to better extract the time domain correlation of the second set of weighted features to obtain a second output result. In this way, four output results corresponding to the four LSTM modules can be obtained. The first three output results are concatenated to obtain a connection result, and a convolution layer is used to reduce the dimension of the connection result to obtain a reduced connection result with four channels. Then, the fourth output result is concatenated with the reduced connection result to obtain a connection feature.
[0083] The connection features are input into an FC layer (Fully connected layer) for full connection to obtain full connection features, and the full connection features are input into a softmax layer to determine 0, 1 detection labels and corresponding prediction values corresponding to the full connection features; a Loss Function loss function in classification is used to calculate the prediction values and true values to obtain a classification loss, so that the model parameters are adjusted according to the classification loss, and the training operation of the wind turbine imbalance detection model is completed.
[0084] It can be seen that, by feature extraction on training data and channel importance learning by using a channel attention mechanism, dominant features are given higher weights by weighting operation in the channel dimension; by using a long short-term memory network to mine time domain correlation between features and introducing a skip connection to reduce the influence of noise on long short-term memory network output results while coping with information ablation caused by the long short-term memory network; then, spatial feature learning is performed by using a spatial attention mechanism, so that dominant features are fully extracted by weighting operation in the spatial dimension, so that the prediction result of the weighted connection features is more accurate, and the problems of poor adaptability and ignoring time domain changes caused by setting a threshold are effectively solved; finally, the model parameters are adjusted by calculating the classification loss, so that a model for better detecting the imbalance of a wind turbine can be obtained, thereby improving the accuracy and efficiency of wind turbine imbalance detection.
[0085] Referring to Figure 5 The embodiment of the application discloses a wind turbine imbalance detection model training device, which comprises:
[0086] The time domain grouping module 11 is configured to input training data into the wind turbine imbalance detection model and group the training data in time domain to obtain a preset number of groups of time domain data; the training data is a group of data collected in a preset time period and containing a plurality of input signals;
[0087] The channel attention module 12 is configured to extract features from each group of time domain data respectively, and perform weighting operation in the channel dimension on each group of extracted features by using a channel attention mechanism to obtain corresponding weighted features of each group;
[0088] The spatial attention module 13 is configured to extract time domain correlation of each group of weighted features by using a long short-term memory network and a skip connection, and perform connection operation to obtain connection features, and perform weighting operation in the spatial dimension on the connection features by using a spatial attention mechanism to obtain weighted connection features;
[0089] The model training completion module 14 is configured to calculate a classification loss based on the predicted value of the weighted connection feature by using a loss function, and adjust the model parameters by using the classification loss, so as to complete the training of the wind turbine imbalance detection model.
[0090] Therefore, the training data is feature extracted, and the channel attention mechanism is used for channel importance learning, so that the dominant features are given higher weights through channel dimension weighting operation; the long short-term memory network is used to mine the time domain correlation between features, and the jump connection is introduced to reduce the influence of noise on the output result of the long short-term memory network while coping with the information ablation caused by the long short-term memory network; then the spatial attention mechanism is used for spatial feature learning, so that the dominant features are fully extracted through spatial dimension weighting operation, so that the prediction result of the weighted connection feature is more accurate, and the problems of poor adaptability and ignoring time domain changes caused by setting threshold are effectively solved; finally, the model parameters are adjusted by calculating the classification loss, so that a model for better detecting the imbalance of the wind turbine can be obtained, thereby improving the accuracy and efficiency of the wind turbine imbalance detection.
[0091] In some embodiments, the channel attention module 12 can specifically include:
[0092] The feature extraction unit is configured to input each group of the time domain data into each feature extraction module respectively, so as to perform convolution, regularization and activation operations on each group of the time domain data respectively, and obtain corresponding extracted features.
[0093] In some embodiments, the channel attention module 12 can specifically include:
[0094] The first pooling unit is configured to input the target group of extracted features into a channel attention module, and perform average pooling and maximum pooling operations on the target group of extracted features respectively, to obtain corresponding first average pooling result and first maximum pooling result; the target group of extracted features is any one group of features in the extracted features;
[0095] The channel importance learning unit is configured to perform channel importance learning based on the first average pooling result and the first maximum pooling result by using a multi-layer perception, to obtain channel importance weights;
[0096] The channel weighting unit is configured to perform weighting operation on corresponding features in the target group of extracted features by using each channel importance weight, to obtain corresponding target group of weighted features.
[0097] In some embodiments, the spatial attention module 13 can specifically include:
[0098] The time domain correlation extraction unit is configured to input each group of the weighted features into each long short-term memory network respectively, so as to extract time domain correlation of each group of the weighted features respectively, and obtain corresponding output results;
[0099] The skip connection unit is configured to connect the output results corresponding to the first N long short-term memory networks according to a preset skip connection rule, so as to obtain a connection result; the first N long short-term memory networks are long short-term memory networks other than the last long short-term memory network in the long short-term memory networks;
[0100] The dimension reduction connection unit is configured to perform dimension reduction processing on the connection result by using a convolution layer, so as to obtain a dimension-reduced connection result, and connect the output result corresponding to the last long short-term memory network and the dimension-reduced connection result, so as to obtain a connection feature.
[0101] In some embodiments, the spatial attention module 13 can specifically include:
[0102] The second pooling unit is configured to input the connection feature into the spatial attention module, so as to perform average pooling and maximum pooling operations on the connection feature respectively, and obtain corresponding second average pooling result and second maximum pooling result;
[0103] The concatenation unit is configured to concatenate the second average pooling result and the second maximum pooling result, so as to obtain a concatenation result;
[0104] The spatial weighting unit is configured to activate the concatenation result by using an activation function, so as to obtain a spatial attention weight, and perform weighting operation on the connection feature by using the spatial attention weight, so as to obtain a weighted connection feature.
[0105] In some embodiments, the model training completion module 14 can specifically include:
[0106] The full connection unit is configured to perform full connection on the weighted connection feature by using a full connection layer, so as to obtain a full connection feature;
[0107] The classification loss calculation unit is configured to determine a detection label and a corresponding prediction value corresponding to the full connection feature by using a softmax layer, and calculate the prediction value and a true value by using a cross-entropy loss function, so as to obtain a classification loss;
[0108] The model parameter adjustment unit is configured to adjust the model parameters by using the classification loss, so as to complete the training operation of the wind turbine imbalance detection model.
[0109] Further, the embodiments of the present application also disclose an electronic device,Figure 6 is an electronic device 20 structure diagram shown according to an exemplary embodiment, the contents of the figure cannot be considered as any limitation on the scope of use of the present application.
[0110] Figure 6 An electronic device 20 structure diagram provided for the embodiments of the present application. The electronic device 20, specifically can include: at least one processor 21, at least one memory 22, power supply 23, communication interface 24, input output interface 25 and communication bus 26. Wherein, the memory 22 is used to store computer programs, the computer programs are loaded and executed by the processor 21, to realize the related steps in the wind turbine imbalance detection model training method disclosed in any of the preceding embodiments. In addition, the electronic device 20 in the embodiment can be an electronic computer.
[0111] In the embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol followed is any communication protocol applicable to the technical solution of the present application, which is not specifically limited here; the input output interface 25 is used to obtain external input data or output data to the outside world, and the specific interface type can be selected according to the specific application needs, which is not specifically limited here.
[0112] In addition, the memory 22 as a resource storage carrier can be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc., and the resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage mode can be temporary storage or permanent storage.
[0113] Wherein, the operating system 221 is used to manage and control each hardware device on the electronic device 20 and the computer program 222, which can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of completing the wind turbine imbalance detection model training method executed by the electronic device 20 disclosed in any of the preceding embodiments, the computer program 222 can further include a computer program capable of completing other specific work.
[0114] Further, the present application also discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by the processor to realize the wind turbine imbalance detection model training method disclosed in the preceding embodiments. The specific steps of the method can refer to the corresponding contents disclosed in the preceding embodiments, which will not be repeated here.
[0115] The various embodiments described in the specification are progressive in nature, and each embodiment highlights the differences from other embodiments. The same or similar parts among the various embodiments can be mutually referred to. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method.
[0116] Those skilled in the art will further appreciate that the individual steps of the examples described in connection with the embodiments disclosed herein can be embodied in electronic hardware, computer software, or combinations of both. The various examples have been described in relation to the described embodiments, as a means of generalizing the interchangeability of hardware and software. Whether employing hardware or software, the described functionality is implemented as desired by the particular application and design constraints. Skilled artisans appreciate that the replacement of one part by a different part, the addition of new parts, or the removal of existing parts from a design, do not alter the basic underlying function of the inventive concept.
[0117] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in random access memory (RAM), flash memory, read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
[0118] Finally, it is to be noted that the terms such as first and second, etc., are used herein only to differentiate one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between such entities or operations. Also, the terms "comprising", "containing", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements in the list, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the phrase "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.
[0119] The above describes the technical solutions provided by the present application in detail, and the principles and implementation manners of the present application are described by using specific examples. The above description of the embodiments is only to help understand the method and core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range can be changed, and the above description of the specification should not be understood as limiting the present application.
Claims
1. A wind turbine imbalance detection model training method, characterized in that, The method comprises the following steps: inputting training data into a wind turbine imbalance detection model, and grouping the training data in time domain to obtain a preset number of groups of time domain data; the training data is a group of data collected within a preset time period and containing multiple input signals; performing feature extraction on each group of time domain data respectively, and performing weighting operation on the channel dimension of each group of extracted features by using channel attention mechanism to obtain corresponding weighted features of each group; performing time domain correlation extraction and connection operation on each group of weighted features by using long short-term memory network and skip connection to obtain connection features, and performing weighting operation on the spatial dimension of the connection features by using spatial attention mechanism to obtain weighted connection features; calculating classification loss based on the predicted value of the weighted connection features by using a loss function, and adjusting the model parameters by using the classification loss to complete the training operation of the wind turbine imbalance detection model; wherein, the step of performing time domain correlation extraction and connection operation on each group of weighted features by using long short-term memory network and skip connection to obtain connection features comprises: inputting each group of weighted features into each long short-term memory network to perform time domain correlation extraction on each group of weighted features respectively to obtain corresponding output results; connecting the output results corresponding to the first N long short-term memory networks according to a preset skip connection rule to obtain a connection result; the first N long short-term memory networks are other long short-term memory networks except the last long short-term memory network in the long short-term memory networks; performing dimension reduction processing on the connection result by using a convolution layer to obtain a reduced connection result, and connecting the output result corresponding to the last long short-term memory network and the reduced connection result to obtain connection features.
2. The wind turbine imbalance detection model training method of claim 1, wherein, The training data is a group of data collected within a preset time period under normal power generation conditions, containing the amplitude corresponding to the tower frequency, 1P frequency, 6P frequency and generator speed.
3. The wind turbine imbalance detection model training method of claim 1, wherein, The step of performing feature extraction on each group of time domain data respectively comprises: inputting each group of time domain data into each feature extraction module to perform convolution, regularization and activation operation on each group of time domain data respectively to obtain corresponding extracted features of each group.
4. The wind turbine imbalance detection model training method of claim 1, wherein, The step of performing weighting operation on the channel dimension of each group of extracted features by using channel attention mechanism to obtain corresponding weighted features of each group comprises: inputting target group extracted features into a channel attention module, and performing average pooling and maximum pooling operation on the target group extracted features respectively to obtain corresponding first average pooling result and first maximum pooling result; the target group extracted features are any one group of features in the extracted features; performing channel importance learning based on the first average pooling result and the first maximum pooling result by using a multilayer perceptron to obtain channel importance weights; performing weighting operation on the corresponding features in the target group extracted features by using each channel importance weight respectively to obtain corresponding target group weighted features.
5. The wind turbine imbalance detection model training method of claim 1, wherein, The spatial dimension weighting operation on the connection feature is performed by using a spatial attention mechanism to obtain a weighted connection feature, including: The connection feature is input into a spatial attention module to perform average pooling and maximum pooling operations on the connection feature respectively to obtain a corresponding second average pooling result and a second maximum pooling result; The second average pooling result and the second maximum pooling result are spliced to obtain a spliced result; An activation function is used to activate the spliced result to obtain a spatial attention weight, and the connection feature is weighted by using the spatial attention weight to obtain a weighted connection feature.
6. The wind turbine imbalance detection model training method of any one of claims 1-5, wherein, The classification loss is calculated based on the predicted value of the weighted connection feature by using a loss function, and the model parameters are adjusted by using the classification loss to complete the training operation of the wind turbine imbalance detection model, including: The weighted connection feature is fully connected by using a fully connected layer to obtain a fully connected feature; The detection label corresponding to the fully connected feature and the corresponding predicted value are determined by using a softmax layer, and the classification loss is obtained by calculating the predicted value and the true value by using a cross-entropy loss function; The model parameters are adjusted by using the classification loss to complete the training operation of the wind turbine imbalance detection model.
7. A wind turbine imbalance detection model training apparatus, characterized by, The device is used to implement the wind turbine imbalance detection model training method according to any one of claims 1 to 6, and the device includes: A time domain grouping module is configured to input training data into a wind turbine imbalance detection model and group the training data in a time domain to obtain a preset number of groups of time domain data; the training data is a group of data collected in a preset time period and containing multiple input signals; A channel attention module is configured to extract features from each group of time domain data, and perform channel dimension weighting operation on each group of extracted features by using a channel attention mechanism to obtain corresponding weighted features of each group; A spatial attention module is configured to extract time domain correlation and connection of each group of weighted features by using a long short-term memory network and a skip connection to obtain connection features, and perform spatial dimension weighting operation on the connection features by using a spatial attention mechanism to obtain weighted connection features; A model training completion module is configured to calculate a classification loss based on a predicted value of the weighted connection feature by using a loss function, and adjust the model parameters by using the classification loss to complete the training operation of the wind turbine imbalance detection model.
8. An electronic device, comprising: It includes: A memory is configured to save a computer program; A processor is configured to execute the computer program to implement the wind turbine imbalance detection model training method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, A computer program is saved, and the computer program is executed by a processor to implement the wind turbine imbalance detection model training method according to any one of claims 1 to 6.