Wind turbine generator gearbox bearing temperature estimation method

Abstract

The invention belongs to the field of wind power, and discloses a wind turbine generator gearbox bearing temperature estimation method, which comprises the following steps of: acquiring a multi-dimensional original signal in an acquisition and monitoring control system; designing a wind turbine generator bearing temperature estimation model; respectively training, verifying and testing the wind turbine generator bearing temperature estimation model through the training set, the verification set and the test set to obtain an optimal wind turbine generator bearing temperature estimation model; the optimal wind turbine generator bearing temperature estimation model is used for conducting temperature estimation on new wind turbine generator SCADA data, and the wind turbine generator bearing temperature estimation model can output corresponding temperature estimation values according to the learned features and relations. According to the packet convolution method adopted by the encoder, the number of model parameters can be reduced, the calculation efficiency of the network is improved, and offline deployment of the model is facilitated.

Description

Technical Field

[0001] The present invention relates to the field of wind power, and in particular to a method for estimating the temperature of a gearbox bearing of a wind turbine generator set. Background Art

[0002] In a wind turbine, the gearbox is a critical subsystem connecting the blades to the generator. Gearbox failures can directly impact the operational status of other key components within the turbine, causing significant economic losses and even safety incidents. Gearbox bearing temperature directly reflects the condition of the gearbox, and many potential faults manifest as abnormally high bearing temperatures. Therefore, monitoring wind turbine gearbox bearing temperature is crucial for improving turbine reliability and reducing operation and maintenance costs.

[0003] Due to the complex structure of wind turbine gearboxes, accurate real-time bearing temperature monitoring is difficult, necessitating an indirect estimation of bearing temperature. Existing methods for estimating wind turbine gearbox bearing temperature are primarily based on data-driven approaches. Supervisory control and data acquisition (SCADA) systems and condition monitoring systems (CMS) are widely used in wind turbines, providing a reliable data foundation for data-driven approaches. Data-driven approaches can describe the operating state of wind turbines based on this data without relying on precise mathematical models, making them the mainstream approach favored in relevant literature. However, the complex and variable wind conditions lead to variable wind turbine operating states, particularly those of the gearbox. Data-driven models must account for these varying states. In particular, different eigenvectors in data-driven models may interfere with each other, affecting the model's temperature estimation. Furthermore, data-driven models typically require high computational complexity and resource usage. In real-world scenarios, offline deployment on devices requires minimal computational effort and storage space to improve computational speed. Therefore, offline deployment of such models presents significant challenges.

[0004] Based on the above problems, the applicant proposes a method for estimating the bearing temperature of a wind turbine gearbox. Summary of the Invention

[0005] The object of the present invention is to provide a method for estimating the bearing temperature of a wind turbine gearbox to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for estimating the bearing temperature of a wind turbine gearbox includes:

[0008] Step 1: Collect multi-dimensional original signals from the data acquisition and monitoring control system of the wind turbine generator system and divide them into training set, verification set and test set;

[0009] Step 2: Design a wind turbine bearing temperature estimation model based on an encoder-decoder structure. This model extracts features through an encoder and performs temperature estimation through a decoder, capturing the complex nonlinear relationship between the input signal and the output temperature to improve the accuracy of temperature estimation.

[0010] Step 3: training, verifying, and testing the wind turbine bearing temperature estimation model using the training set, validation set, and test set to obtain the optimal wind turbine bearing temperature estimation model;

[0011] Step 4: Use the optimal wind turbine bearing temperature estimation model to estimate the temperature of the new wind turbine SCADA data. The wind turbine bearing temperature estimation model can output the corresponding temperature estimation value based on the learned features and relationships.

[0012] Furthermore, in step 1, required SCADA variables are selected as input according to actual conditions. The SCADA variables include multiple ones of wind speed, generator speed, output active power, active power generation, converter motor speed, main shaft speed, current value of each blade motor, temperature value of each blade motor, main shaft overspeed trigger value, high-speed shaft braking state, generator drive end bearing temperature, generator non-drive end bearing temperature, and gearbox oil temperature, thereby constituting multidimensional SCADA time series data.

[0013] Furthermore, the step 2 includes:

[0014] Step 2.1, split the multi-dimensional SCADA time series data selected in step 1 into segment samples according to a certain interval and step size;

[0015] In step 2.2, the segmented SCADA data samples are input into the encoder of the wind turbine bearing temperature estimation model. The encoder uses a one-dimensional convolutional neural network based on the grouped convolution method to extract the time series features by performing grouped convolution on the samples to obtain a feature matrix.

[0016] In step 2.3, the feature matrix is ​​input into the decoder of the wind turbine bearing temperature estimation model. The decoder uses LSTM-RNN as a sequence information capturer to capture the complex nonlinear relationship between input and output from the features and give a specific temperature estimation value.

[0017] Furthermore, the encoder of the wind turbine bearing temperature estimation model includes two one-dimensional convolution layers and a pooling layer. The two one-dimensional convolution layers of the convolution network adopt the group convolution method to prevent the mutual interference of different feature vectors. The size of the input time series data X is set to L×C, where L is the time length of the input time series data and C is the number of channels. The size of the output X' of the convolution layer is L'×C. out , L′ is the length of the output, C out is the number of channels, L′ depends on the step size and filling method of the convolution; after the input time series data is divided into G groups according to the channels, the number of channels of each group of input time series data is Therefore, the number of channels of each convolution kernel is also reduced to C', so each group requires at least one convolution kernel of length k is a set of real numbers, and convolution is performed in the corresponding group. The number of output channels of the grouped convolution is at least G;

[0018] For a one-dimensional convolutional layer, the mathematical expression of the i-th group convolution is as follows:

[0019]

[0020] in, is the jth element of the cth channel of the i-th group of the input sequence X, is the convolution kernel K i The weight of the j-th position of the c-th channel, * represents the convolution operation; is the convolution output of group i;

[0021] Finally, the convolution results of all groups are concatenated along the channel dimension to obtain the output of the one-dimensional convolution layer.

[0022] X'=concat(X1′,X2′,…,X G ′)

[0023] Among them, concat represents the splicing operation, and the pooling layer adopts the average sampling method to the output X of the second convolutional layer. (2) Sampling along the time dimension, we get the output X of the one-dimensional convolutional neural network (p) ;

[0024] The encoder is expressed as follows:

[0025] enter:

[0026] Convolutional layer 1: X (1) =Conv1(X)

[0027] Convolutional layer 2: X (2)=Conv2(X (1) )

[0028] Pooling layer: X (p) =Pooling(X (2) )

[0029] Output:

[0030] in, and are the outputs of the input X after passing through convolution layer 1, convolution layer 2, and pooling layer respectively; L1 and L2 depend on the step size and padding method of the corresponding convolution layer, C1 and C2 are the hyperparameters of convolution layer 1 and convolution layer 2 respectively; L p Depends on the pooling window size of the pooling layer.

[0031] Furthermore, the decoder of the wind turbine bearing temperature estimation model includes an LSTM recurrent layer and a fully connected layer; the output of the encoder is input into the LSTM recurrent layer, and then the sequence output of the LSTM recurrent layer is converted into a specific temperature estimation value y through the fully connected layer;

[0032] The LSTM recurrent layer has H LSTM units and a hidden state. h t The current input X t and the hidden state h at the previous moment t-1 Decision, subject to input gate i t 、Forget Gate t , output gate o t Regulation; at time t, memory unit c t and hidden state h t The calculation formula is:

[0033] i t =σ(W xi X t +W hi h t-1 )

[0034] f t =σ(W xf X t +W hf h t-1 )

[0035] o t =σ(W xo X t +W ho h t-1 )

[0036] g t =tanh(W xgX t +W hg h t-1 )

[0037] c t =f t ⊙c t-1 +i t ⊙g t

[0038] h t =o t ⊙tanh(c t )

[0039] Among them, g t is a candidate memory unit, c t is the memory unit; W xi 、W hi are the input weight matrix of the input gate and the hidden state weight matrix, W xf 、W hf are the input weight matrix and hidden state weight matrix of the forget gate, W xo 、W ho are the input weight matrix of the output gate and the hidden state weight matrix, W xg 、W hg are the input weight matrix and hidden state weight matrix of the candidate memory unit, W xi 、W hi 、W xf 、W hf 、W xo 、W ho 、W xg 、W hg are all learnable parameters; σ(·) and tanh(·) denote sigmoid and tanh activation functions respectively; ⊙ denotes element-wise multiplication;

[0040] The decoder is expressed as follows:

[0041] enter:

[0042] LSTM recurrent layer: h t ,c t =LSTM(X (p) (t),h t-1 ,c t-1 ),t=1,2,…,L p

[0043] Fully connected layer:

[0044] Output: y

[0045] Among them, ht and c t are the hidden state and memory unit state of the LSTM recurrent layer at the tth moment respectively; For the LSTM hidden layer at the last time step L p The hidden state of and b d are the weights and bias terms of the fully connected layer, respectively.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1) This paper proposes a wind turbine bearing temperature estimation method based on an encoder-decoder structure and grouped convolution. The encoder uses a one-dimensional convolutional neural network (1D-CNN) based on a grouped convolution method as a feature extractor to extract features from the input time series data. The decoder uses a long short-term memory recurrent neural network (LSTM-RNN) as a sequence information capturer to capture the complex nonlinear relationship between input and output from the features and provide an output temperature estimate. When processing input data with multiple channels or features, different groups of convolution kernels can capture the relationship between different channels or features, isolate information exchange between different groups, and avoid interference from different signals.

[0048] 2) The group convolution method adopted by the encoder of the present invention can reduce the number of model parameters, improve the computational efficiency of the network, and facilitate the offline deployment of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the present invention.

[0050] Figure 2 This is a schematic diagram of the structure of the wind turbine bearing temperature estimation model in the present invention.

[0051] Figure 3 Schematic diagram of group convolution in the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] See also Figure 1-Figure 3 The present invention proposes a method for estimating the bearing temperature of a wind turbine gearbox. The core of the method is the wind turbine bearing temperature estimation model based on the encoding-decoding structure and group convolution proposed by the present invention. Figure 2As shown in the figure, the encoder uses a one-dimensional convolutional neural network (1D-CNN) based on the grouped convolution method as a feature extractor to extract the features of the input time series data; the decoder uses a long short-term memory recurrent neural network (LSTM-RNN) as a sequence information capturer to capture the complex nonlinear relationship between input and output from the features and provide an output temperature estimate. Specifically, it includes the following steps:

[0054] Step 1: Collect multi-dimensional original signals from the wind turbine generator system's supervisory control and data acquisition (SCADA) system and divide them into training set, validation set and test set.

[0055] Among them, the required SCADA variables can be selected as input according to actual conditions, such as wind speed, generator speed, output active power, active power generation, converter motor speed, main shaft speed, current value of each blade motor, temperature value of each blade motor, main shaft overspeed trigger value, high-speed shaft braking status, generator drive end bearing temperature, generator non-drive end bearing temperature, gearbox oil temperature and other 17 variables to form multi-dimensional SCADA time series data for subsequent model training and testing, ensuring that the model can learn features related to temperature changes.

[0056] Step 2: Design a wind turbine bearing temperature estimation model based on an encoder-decoder structure. This model extracts features through an encoder and performs temperature estimation through a decoder, capturing the complex nonlinear relationship between the input signal and the output temperature to improve the accuracy of temperature estimation. Specifically, the model includes the following steps:

[0057] Step 2.1: Divide the multi-dimensional SCADA time series data selected in step 1 into segment samples according to certain intervals and step sizes.

[0058] In step 2.2, the segmented SCADA data samples are input into the encoder of the wind turbine bearing temperature estimation model. The encoder uses a one-dimensional convolutional neural network (1D-CNN) based on the group convolution method to extract the time series features by performing group convolution on the samples to obtain the feature matrix.

[0059] In step 2.3, the feature matrix is ​​input into the decoder of the wind turbine bearing temperature estimation model. The decoder uses LSTM-RNN as a sequence information capturer to capture the complex nonlinear relationship between input and output from the features and give a specific temperature estimation value.

[0060] Among them, the encoder structure of the wind turbine bearing temperature estimation model is as follows: Figure 2 As shown in the figure, it contains two one-dimensional convolution layers (Conv1, Conv2) and a pooling layer (Pooling). The convolution network uses the group convolution method to prevent the mutual interference of different feature vectors. The diagram of group convolution is shown in the attached figure. Figure 3 As shown. Set the size of the input time series data X to L×C, L is the time length of the input time series data, C is the number of channels (number of variables); the size of the output X' of the convolutional layer is L'×C out , L' is the length of the output, C out is the number of channels, and L′ depends on the step size and filling method of the convolution. After the input time series data is divided into G groups according to the channels, the number of channels of each group of input time series data is Therefore, the number of channels of each convolution kernel is also reduced to C′, so each group requires at least one convolution kernel of length k. is a set of real numbers, and convolution is performed in the corresponding group. The number of group convolution output channels is at least G.

[0061] For a one-dimensional convolutional layer, the mathematical expression of the i-th group convolution is as follows:

[0062]

[0063] in, is the jth element of the cth channel of the i-th group of the input sequence X, is the convolution kernel K i The weight of the j-th position of the c-th channel, * represents the convolution operation; is the convolution output of group i.

[0064] Finally, the convolution results of all groups are concatenated along the channel dimension to obtain the output of the one-dimensional convolution layer.

[0065] X'=concat(X1′,X2′,…,X G ′)

[0066] Here, "concat" represents a concatenation operation. This shows that using grouped convolution can reduce the number of model parameters, reducing the total number of parameters to 1 / G of the original, thereby improving the network's computational efficiency. When processing input data with multiple channels or features, different groups of convolution kernels can capture the relationship between different channels or features, isolate information exchange between different groups, and avoid interference between different signals.

[0067] The pooling layer uses an average sampling method to sample the output X of the second convolutional layer. (2) Sampling along the time dimension, we get the output X of the one-dimensional convolutional neural network (p) .

[0068] The encoder is expressed as follows:

[0069] enter:

[0070] Convolutional layer 1: X (1) =Conv1(X)

[0071] Convolutional layer 2: X (2) =Conv2(X (1) )

[0072] Pooling layer: X (p) =Pooling(X (2) )

[0073] Output:

[0074] Among them, Conv represents the convolution layer, Pooling represents the pooling layer, and are the outputs of the input X after passing through convolution layer 1, convolution layer 2, and pooling layer respectively; L1 and L2 depend on the step size and padding method of the convolution layer, C1 and C2 are the hyperparameters of convolution layer 1 and convolution layer 2 respectively; L p Depends on the pooling window size of the pooling layer.

[0075] Among them, the decoder structure of the wind turbine bearing temperature estimation model is as follows: Figure 2 As shown in the figure, it contains an LSTM recurrent layer and a fully connected layer (Dense). The encoder output (extracted features) is input into the LSTM recurrent layer, and then the fully connected layer converts the sequence output of the LSTM recurrent layer into a specific temperature estimate y.

[0076] The LSTM recurrent layer has H LSTM units and a hidden state. h t The current input X t and the hidden state h at the previous moment t-1 Decision, subject to input gate i t 、Forget Gate t , output gate o t At time t, memory unit c t and hidden state h t The calculation formula is:

[0077] i t =σ(W xi X t +W hi h t-1 )

[0078] f t =σ(W xf X t +W hf h t-1 )

[0079] o t =σ(W xo X t +W ho h t-1 )

[0080] g t =tanh(W xg X t +W hg h t-1 )

[0081] c t =f t ⊙c t-1 +i t ⊙g t

[0082] h t =o t ⊙tanh(c t )

[0083] Among them, g t is a candidate memory unit, c t is the memory unit; W xi 、W hi are the input weight matrix of the input gate and the hidden state weight matrix, W xf 、W hf are the input weight matrix and hidden state weight matrix of the forget gate, W xo 、W ho are the input weight matrix of the output gate and the hidden state weight matrix, W xg 、W hg are the input weight matrix and hidden state weight matrix of the candidate memory unit, W xi 、W hi 、W xf 、W hf 、W xo 、W ho 、W xg 、W hg are all learnable parameters; σ(·) and tanh(·) denote sigmoid and tanh activation functions, respectively; ⊙ denotes element-wise multiplication.

[0084] The decoder is expressed as follows:

[0085] enter:

[0086] LSTM recurrent layer: h t ,c t =LSTM(X (p) (t),h t-1 ,ct-1 ),t=1,2,…,L p

[0087] Fully connected layer:

[0088] Output: y

[0089] Among them, LSTM represents the cycle layer, h t and c t are the hidden state and memory unit state of the LSTM recurrent layer at the tth moment respectively; For the LSTM hidden layer at the last time step L p The hidden state of and b d are the weights and bias terms of the fully connected layer, respectively.

[0090] Step 3: Train, verify, and test the wind turbine bearing temperature estimation model using the training set, validation set, and test set to obtain the optimal wind turbine bearing temperature estimation model.

[0091] By training the model using the training and validation sets and evaluating it on the test set, we can effectively assess the model's performance at different stages. Parameter optimization during training and adjustments during validation can help the model better adapt to real-world application scenarios. Model hyperparameters can also be adjusted during training based on data characteristics and validation test results. For example, the convolution kernel size of a 1D-CNN can be adjusted, such as using a kernel size of n×10 for convolution, where n is the dimension of the input variable.

[0092] Step 4: Use the trained and validated model to estimate the temperature of the new wind turbine SCADA data. The model can output the corresponding temperature estimate based on the previously learned features and relationships.

[0093] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for estimating the bearing temperature of a wind turbine gearbox, characterized in that: include: Step 1: Collect multi-dimensional original signals from the data acquisition and monitoring control system of the wind turbine generator system and divide them into training set, verification set and test set; Step 2: Design a wind turbine bearing temperature estimation model based on an encoder-decoder structure. This model extracts features through an encoder and performs temperature estimation through a decoder, capturing the complex nonlinear relationship between the input signal and the output temperature to improve the accuracy of temperature estimation. Step 3: training, verifying, and testing the wind turbine bearing temperature estimation model using the training set, validation set, and test set to obtain the optimal wind turbine bearing temperature estimation model; Step 4: Use the optimal wind turbine bearing temperature estimation model to estimate the temperature of the new wind turbine SCADA data. The wind turbine bearing temperature estimation model can output the corresponding temperature estimation value based on the learned features and relationships.

2. A method for estimating the bearing temperature of a wind turbine gearbox according to claim 1, characterized in that: In step 1, required SCADA variables are selected as input according to actual conditions. The SCADA variables include multiple ones of wind speed, generator speed, output active power, active power generation, converter motor speed, main shaft speed, current value of each blade motor, temperature value of each blade motor, main shaft overspeed trigger value, high-speed shaft braking state, generator drive end bearing temperature, generator non-drive end bearing temperature, and gearbox oil temperature, thereby constituting multidimensional SCADA time series data.

3. The method for estimating the bearing temperature of a wind turbine gearbox according to claim 1, wherein: The step 2 includes: Step 2.1, split the multi-dimensional SCADA time series data selected in step 1 into segment samples according to a certain interval and step size; In step 2.2, the segmented SCADA data samples are input into the encoder of the wind turbine bearing temperature estimation model. The encoder uses a one-dimensional convolutional neural network based on the grouped convolution method to extract the time series features by performing grouped convolution on the samples to obtain a feature matrix. In step 2.3, the feature matrix is ​​input into the decoder of the wind turbine bearing temperature estimation model. The decoder uses LSTM-RNN as a sequence information capturer to capture the complex nonlinear relationship between input and output from the features and give a specific temperature estimation value.

4. A method for estimating the bearing temperature of a wind turbine gearbox according to claim 3, characterized in that: The encoder of the wind turbine bearing temperature estimation model includes two one-dimensional convolution layers and a pooling layer. The two one-dimensional convolution layers of the convolution network adopt the group convolution method to prevent the mutual interference of different feature vectors. The size of the input time series data X is set to L×C, where L is the time length of the input time series data and C is the number of channels. The size of the output X' of the convolution layer is L'×C. out , L′ is the length of the output, C out is the number of channels, L′ depends on the step size and filling method of the convolution; after the input time series data is divided into G groups according to the channels, the number of channels of each group of input time series data is Therefore, the number of channels of each convolution kernel is also reduced to C', so each group requires at least one convolution kernel of length k is a set of real numbers, and convolution is performed in the corresponding group. The number of output channels of the grouped convolution is at least G; For a one-dimensional convolutional layer, the mathematical expression of the i-th group convolution is as follows: in, is the jth element of the cth channel of the i-th group of the input sequence X, is the convolution kernel K i The weight of the j-th position of the c-th channel, * represents the convolution operation; is the convolution output of group i; Finally, the convolution results of all groups are concatenated along the channel dimension to obtain the output of the one-dimensional convolution layer. X'=concat(x1′,X2′,…,X G ′) Among them, concat represents the splicing operation, and the pooling layer adopts the average sampling method to the output X of the second convolutional layer. (2) Sampling along the time dimension, we get the output X of the one-dimensional convolutional neural network (p) ; The encoder is expressed as follows: enter: Convolutional layer 1: X (1) =Conv1(X) Convolutional layer 2: X (2) =Conv2(X (1) ) Pooling layer: X (p) =Pooling(X (2) ) Output: in, and are the outputs of the input X after passing through convolution layer 1, convolution layer 2, and pooling layer respectively; L1 and L2 depend on the step size and padding method of the corresponding convolution layer, C1 and C2 are the hyperparameters of convolution layer 1 and convolution layer 2 respectively; L p Depends on the pooling window size of the pooling layer.

5. The method for estimating the bearing temperature of a wind turbine gearbox according to claim 3, wherein: The decoder of the wind turbine bearing temperature estimation model includes an LSTM recurrent layer and a fully connected layer; the output of the encoder is input into the LSTM recurrent layer, and then the sequence output of the LSTM recurrent layer is converted into a specific temperature estimation value y through the fully connected layer; The LSTM recurrent layer has H LSTM units and a hidden state. h t The current input X t and the hidden state h at the previous moment t-1 Decision, subject to input gate i t 、Forget Gate t , output gate o t Regulation; at time t, memory unit c t and hidden state h t The calculation formula is: i t =σ(W xi X t +W hi h t-1 ) f t =σ(W xf X t +W hf h t-1 ) o t =σ(W xo X t +W ho h t-1 ) g t =tanh(W xg X t +W hg h t-1 ) c t =f t ⊙c t-1 +i t ⊙g t h t =o t ⊙tanh(c t ) Among them, g t is a candidate memory unit, c t is the memory unit; W xi 、W hi are the input weight matrix of the input gate and the hidden state weight matrix, W xf 、W hf are the input weight matrix and hidden state weight matrix of the forget gate, W xo 、W ho are the input weight matrix of the output gate and the hidden state weight matrix, W xg 、W hg are the input weight matrix and hidden state weight matrix of the candidate memory unit, W xi 、W hi 、W xf 、W hf 、W xo 、W ho 、W xg 、W hg are all learnable parameters; σ(·) and tanh(·) denote sigmoid and tanh activation functions respectively; ⊙ denotes element-wise multiplication; The decoder is expressed as follows: enter: LSTM recurrent layer: h t ,c t =LSTM(X (p) (t),h t-1 ,c t-1 ),t=1,2,…,L p Fully connected layer: Output: y Among them, h t and c t are the hidden state and memory unit state of the LSTM recurrent layer at the tth moment respectively; For the LSTM hidden layer at the last time step L p The hidden state of and b d are the weights and bias terms of the fully connected layer, respectively.

Images