Wind power bearing health assessment method of lightweight coarse and fine granularity feature fusion model

Abstract

The invention discloses a wind power bearing health assessment method for a lightweight coarse and fine granularity feature fusion model, and the method comprises the steps: 1, collecting the vibration signals of an insulating bearing at a transmission end and a non-transmission end of a wind power generator through two three-way acceleration sensors, and obtaining the real fault data of the insulating bearing; 2, performing normalization preprocessing on all the insulation bearing vibration signals, constructing an insulation bearing data set, and randomly dividing the data set into a 80% training set and a 20% test set; step 3, constructing a lightweight coarse and fine granularity feature fusion model BSTA-Net; 4, taking the training set data as the input of a BSTA-Net model, taking the fault category as the output, and carrying out model training; and step 5, inputting test set data into the trained BSTA-Net model to obtain an insulation bearing fault classification result and test accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of wind turbine bearing condition monitoring, and relates to a wind turbine bearing health assessment method based on a lightweight coarse-grained and fine-grained feature fusion model. Background Art

[0002] In modern industrial systems, bearings are the key factor in determining the life of large rotating equipment. Bearings are like the heart of high-end rotating systems and are indispensable in rotating machinery fields such as wind turbines, aerospace, etc. With the development of high-power semiconductor devices and the advancement of pulse width modulation technology, motors have been faced with serious bearing electrocorrosion problems. Insulated bearings can effectively suppress bearing electrocorrosion problems, but as the service environment becomes increasingly harsh, the insulation performance of insulated bearings will decay rapidly during service, making insulated bearings very susceptible to shaft current micro-damage failures. According to statistics, almost all bearings of variable frequency controlled motors have this kind of failure. Insulated bearing electrocorrosion failures not only lead to catastrophic safety accidents, but also cause huge economic losses. Therefore, the study of effective intelligent monitoring and diagnosis technology for insulated bearings is of great significance to ensure the safe service of mechanical equipment.

[0003] Deep learning has shown great potential in the field of fault diagnosis. However, considering the complexity of modern industrial systems and the variability of mechanical working conditions, especially when facing the fault identification of insulated bearings of high-power wind turbines, the existing models still have the following shortcomings: (1) Due to the complexity of the structure of the insulated bearings of wind turbines, the fault characteristics of the insulated bearings are difficult to detect, especially in harsh and demanding service environments such as alternating voltage, alternating temperature and high load, which makes the fault information components of the insulated bearings very complex. Therefore, accurate identification of insulated bearing faults has become a recognized technical problem. (2) The performance degradation data of insulated bearings often have the characteristics of time series. Under the interaction of shaft current micro-damage and fatigue wear, the degradation rate of bearing insulation performance and mechanical structure damage have strong time-varying characteristics. Therefore, considering how to fully exploit the time series characteristics of insulated bearing fault data has become a key problem that needs to be overcome. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a wind turbine bearing health assessment method based on a lightweight coarse-grained and fine-grained feature fusion model.

[0005] The present invention provides a wind turbine bearing health assessment method based on a lightweight coarse-grained and fine-grained feature fusion model, comprising:

[0006] Step 1: Use two three-axis acceleration sensors to collect the insulation bearing vibration signals at the transmission end and non-transmission end of the wind turbine respectively to obtain the real insulation bearing fault data;

[0007] Step 2: Normalize and preprocess all the vibration signals of the insulating bearings, construct an insulating bearing dataset, and randomly divide the dataset into 80% training set and 20% test set;

[0008] Step 3: Construct a lightweight coarse-grained and fine-grained feature fusion model BSTA-Net;

[0009] Step 4: Use the training set data as the input of the BSTA-Net model and the fault category as the output to perform model training;

[0010] Step 5: Input the test set data into the trained BSTA-Net model to obtain the insulated bearing fault classification results and test accuracy.

[0011] The wind turbine bearing health assessment method of the present invention based on a lightweight coarse-grained and fine-grained feature fusion model has the following beneficial effects:

[0012] (1) Considering the complexity of modern high-power variable-frequency industrial systems and the variability of mechanical working conditions, the correlation of coupling, and the multi-source of excitation, the proposed lightweight coarse-grained and fine-grained feature fusion model BSTA-Net is used to train the measured insulating bearing fault data, so that the insulating bearing fault of wind turbines can be intelligently identified for the first time, providing an innovative solution for wind turbine bearing fault identification.

[0013] (2) Under the interaction of shaft current micro-damage and fatigue wear, the degradation rate of bearing insulation performance and mechanical structure damage have strong time-varying characteristics, which makes the existing model have poor global convergence and complex calculation problems during training. The present invention innovatively designs a bidirectional time series information feature fusion structure, namely the Bi-SRU layer, and cleverly applies this strategy to the coarse-grained feature extraction module of the proposed BSTA-Net model for the first time, thereby fully extracting the time series information of the insulating bearing in the process of insulation performance and mechanical fatigue comprehensive failure, providing an innovative and practical contribution to the development of time series information feature extraction.

[0014] (3) Based on the same dataset, the diagnostic performance of the proposed BSTA-Net model was evaluated from multiple dimensions. The results show that the BSTA-Net model can extract effective features from the original signal to realize the intelligent diagnosis of insulated bearings. Compared with previous studies, the proposed framework exhibits excellent performance with high superiority, generalization and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a wind turbine bearing health assessment method based on a lightweight coarse-grained and fine-grained feature fusion model of the present invention.

[0016] Figure 2It is a schematic diagram of a simple recurrent unit;

[0017] Figure 3 is a schematic diagram of a simple recurrent unit network with two directions. DETAILED DESCRIPTION

[0018] like Figure 1 As shown, a wind turbine bearing health assessment method based on a lightweight coarse-grained and fine-grained feature fusion model of the present invention comprises:

[0019] Step 1: Use two three-axis acceleration sensors to collect the insulation bearing vibration signals at the transmission end and non-transmission end of the wind turbine respectively to obtain the real insulation bearing fault data.

[0020] Step 2: Normalize and preprocess all the insulating bearing vibration signals, construct an insulating bearing dataset, and randomly divide the dataset into 80% training set and 20% test set.

[0021] Step 3: Build a lightweight coarse-grained and fine-grained feature fusion model BSTA-Net.

[0022] In specific implementation, the lightweight coarse-grained and fine-grained feature fusion model BSTA-Net includes: a fine-grained feature extraction module, a coarse-grained feature extraction module and a feature fusion recognition module. The vibration signal of the insulating bearing is first input into the fine-grained feature extraction module to extract local features, and the extracted local features are then input into the coarse-grained feature extraction module to extract the global feature information containing the spatial features and temporal features of the measured original vibration signal of the insulating bearing, and the global feature information is input into the feature fusion recognition module for fault classification.

[0023] In a specific implementation, the fine-grained feature extraction module includes: an input layer, a first convolutional layer, a first pooling layer, 4 groups of Fire units, a maximum pooling layer and a discard layer.

[0024] The vibration signal of the insulating bearing is input into the BSTA-Net model. In order to adapt to the structural characteristics of the time series of the one-dimensional time domain vibration signal of the insulating bearing, the convolution kernel of the first convolution layer and the pooling kernel of the first pooling layer are both designed as one-dimensional structures.

[0025] Four groups of sequentially connected Fire units are used to replace the deep convolutional neural network, so that the size of the convolution kernel in the network is simplified from 3×3 to 1×1, and the number of input channels is reduced to 3×3.

[0026] In order to adaptively pick up the nonlinear feature expression hidden in the fault state of the insulating bearing and the topological structure characteristics inside the data, the SELU function is introduced as an activation unit to perform nonlinear processing on the feature mapping. Its function expression is:

[0027]

[0028] Where x is the output of the last group of Fire units; scale and α are constants;

[0029] A maximum pooling layer is added after the activation function to reduce the dimension of the input feature map and the parameters of network training, thereby preventing overfitting and reducing the amount of computation.

[0030] In specific implementation, the coarse-grained feature extraction module includes: a second convolution layer, a BN layer and a Bi-SRU layer. The local features extracted by the fine-grained feature extraction module are sequentially input into the second convolution layer, the BN layer and the Bi-SRU layer, and the global feature information containing the spatial features and temporal features of the measured original vibration signal of the insulating bearing is output.

[0031] like Figure 2 and Figure 3 As shown, the Bi-SRU layer includes a forward simple recurrent unit network and a backward simple recurrent unit network, and the forward simple recurrent unit network and the backward simple recurrent unit network are both composed of multiple simple recurrent units.

[0032] The gated recurrent unit is a variant of the recurrent neural network, which can better capture the dependencies with large time step distances in the time series. However, although this type of recurrent neural network with a gating mechanism overcomes the defects of ordinary recurrent neural networks, its internal structure is more complex, and its internal computing units still have time-series dependencies, so they cannot be calculated in parallel. In order to improve the training speed of the recurrent neural network, a simple recurrent neural network is used.

[0033] The simple recurrent unit includes input gate, forget gate, reset gate, output gate, internal state and residual connection. The simple recurrent unit removes the dependence of internal calculation on the hidden state of the previous moment and improves the overall gate structure to ensure that the network has the performance of LSTM. The bidirectional simple recurrent unit network of the Bi-SRU layer performs sequential and reverse propagation to extract global features.

[0034] The calculation process of the simple cycle unit is as follows:

[0035] f t =σ(W f x t +b f )

[0036] r t =σ(W r x t +b r )

[0037] y t =σ(W * xt )

[0038] A t =f t ·e·A t-1 +(1-f t )·e·y t

[0039] h t =r t ·e·g(A t )+(1-r t )·e·T t

[0040] Among them, f t is the hidden state of the forget gate, r t To reset the hidden state of the door, W f , W r and W * is the parameter matrix of the simple recurrent unit, b f 、b r is the bias of the simple recurrent unit, which is the parameter vector to be learned during training. t represents the input features at time t, y t represents the output of the Bi-SRU layer at time t, h t represents the output of the simple recurrent unit at the current time t, h t-1 represents the output of the simple recurrent unit at the previous time t-1, A t represents the hidden state of the simple recurrent unit at the current time t, and e represents the multiplication operation of the corresponding elements of the matrix.

[0041] Since the simple recurrent unit extracts long-distance dependencies, the more distant the input is, the less influence it has on the current moment, and the closer the input is, the greater the influence is. Therefore, the sequence features extracted by the simple recurrent unit in a single direction will be more affected by the input at the current moment. In order to better model the temporal relationship, a bidirectional simple recurrent unit network encoder is used to capture the temporal features of both directions at the same time, and then the features at the same moment are fused to obtain the final feature expression at each moment. The feature extraction layer is based on the flow matrix x = [x1, x2, ..., x n ] is the input, the forward simple recurrent unit network of the bidirectional simple recurrent unit network From x1 to x n Read input feature x; backward simple recurrent unit network From x n To x1 read the input feature x:

[0042]

[0043] in, is the forward hidden state, collecting x in the stream sequence t Previous information; For the backward hidden state, collect x t The information after that; the initial hidden state vector and are all zero vectors, and Spliced ​​together, we get the summary information H of the entire flow at time step t t ,Right now

[0044] In order to further obtain the high-order features of the traffic data, the bidirectional simple recurrent unit network is stacked into M layers, and the input of the i-th layer at time t is the output of the i-1th layer at that time As shown below:

[0045]

[0046] The number of bytes per stream, N, and the number of simple cyclic unit network stacking layers, M, will be determined experimentally. The summary information of the entire stream at the i-th layer at time t is expressed as Concatenate the hidden states generated at each time step of the last layer to get the output of the feature extraction layer:

[0047]

[0048] After the feature extraction process of M-layer bidirectional simple recurrent unit network, the feature vector H N Highly aggregated bidirectional information of insulated bearings.

[0049] In a specific implementation, the feature fusion recognition module includes: an exchange axis layer, a self-attention layer, a data flattening layer and an output layer.

[0050] The global feature information containing the spatial features and temporal features of the measured original vibration signal of the insulating bearing output by the coarse-grained feature extraction module is used as the input of this module. The internal feature representation is obtained by exchanging the axial layer, and then the self-attention layer is used to calculate the weights of different channels and output them through the Sigmoid function. The data flattening layer is used to replace the original fully connected layer, which can reduce model parameters and computational complexity. Finally, the Softmax function converts the neuron output into a probability distribution of 10 types of insulating bearing faults, and the features with different weights are input into the classifier for fault classification, thereby realizing the classification of insulating bearing faults.

[0051] In specific implementation, the self-attention layer can suppress redundant information irrelevant to the target result from complex and numerous information, and its input time series is:

[0052] For each input X i Linearly map it to three different spaces to obtain the query vector Key Vector Sum value vector The output time series is For the entire input time series X, the linear mapping process is expressed as:

[0053]

[0054] in, are the parameter matrices of the linear mapping, Q = [q1, q2, …, q n ], K=[k1,k2,…,k n ]、V=[v1,v2,…,v n ] are respectively composed of the query vector q i , key vector k i Sum value vector v i The matrix D k and N represent the matrix dimensions respectively;

[0055] For each query vector q in the input time series n ∈Q, get the output vector Y n :

[0056]

[0057] Among them, n∈[1,2,...,N] is the position of the output vector sequence, j∈[1,2,...,N] is the position of the input vector sequence, and a nj Indicates the weight of the n-th output paying attention to the j-th input.

[0058] Step 4: Use the training set data as the input of the BSTA-Net model and the fault category as the output to train the model.

[0059] First, the training set data is input into the BSTA-Net model for pre-training, the model parameters are trained through the back-propagation update strategy, and the hyperparameters of the framework such as learning rate and number of iterations are selected according to the iteration curve. Then, the training set data is input into the BSTA-Net model for forward propagation, and the BSTA-Net model is used to learn the fault characteristics of the insulated bearing. Finally, the weight parameters of the BSTA-Net model are adjusted according to the output results, and the update process is repeated to reduce the loss function value and improve the diagnostic accuracy until the optimal BSTA-Net model is obtained.

[0060] Step 5: Input the test set data into the trained BSTA-Net model to obtain the insulated bearing fault classification results and test accuracy.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind turbine bearing health assessment method based on a lightweight coarse-grained and fine-grained feature fusion model, characterized in that: include: Step 1: Use two three-axis acceleration sensors to collect the insulation bearing vibration signals at the transmission end and non-transmission end of the wind turbine respectively to obtain the real insulation bearing fault data; Step 2: Normalize and preprocess all the vibration signals of the insulating bearings, construct an insulating bearing dataset, and randomly divide the dataset into 80% training set and 20% test set; Step 3: Construct a lightweight coarse-grained and fine-grained feature fusion model BSTA-Net; Step 4: Use the training set data as the input of the BSTA-Net model and the fault category as the output to train the model; Step 5: Input the test set data into the trained BSTA-Net model to obtain the insulated bearing fault classification results and test accuracy.

2. The wind turbine bearing health assessment method based on the lightweight coarse-grained and fine-grained feature fusion model according to claim 1 is characterized in that: The lightweight coarse-grained and fine-grained feature fusion model BSTA-Net includes: a fine-grained feature extraction module, a coarse-grained feature extraction module and a feature fusion recognition module; the vibration signal of the insulating bearing is first input into the fine-grained feature extraction module to extract local features, and the extracted local features are then input into the coarse-grained feature extraction module to extract global feature information containing the spatial features and temporal features of the measured original vibration signal of the insulating bearing, and the global feature information is input into the feature fusion recognition module for fault classification.

3. The wind turbine bearing health assessment method based on the lightweight coarse-grained and fine-grained feature fusion model according to claim 1 is characterized in that: The fine-grained feature extraction module includes: an input layer, a first convolutional layer, a first pooling layer, 4 groups of Fire units, a maximum pooling layer and a discard layer; The vibration signal of the insulating bearing is input into the BSTA-Net model. To adapt to the structural characteristics of the time series of the one-dimensional time domain vibration signal of the insulating bearing, the convolution kernel of the first convolution layer and the pooling kernel of the first pooling layer are designed as one-dimensional structures. Four groups of Fire units connected in sequence are used to replace the deep convolutional neural network, so that the size of the convolution kernel in the network is simplified from 3×3 to 1×1, and the number of input channels is reduced to 3×3; In order to adaptively pick up the nonlinear feature expression hidden in the fault state of the insulating bearing and the topological structure characteristics inside the data, the SELU function is introduced as an activation unit to perform nonlinear processing on the feature mapping. Its function expression is: Where x is the output of the last group of Fire units; scale and α are constants; Add a maximum pooling layer after the activation function to reduce the dimension of the input feature map and reduce the parameters of network training, prevent overfitting and reduce the amount of calculation.

4. The wind turbine bearing health assessment method based on the lightweight coarse-grained and fine-grained feature fusion model according to claim 1 is characterized in that: The coarse-grained feature extraction module includes: a second convolution layer, a BN layer and a Bi-SRU layer; the local features extracted by the fine-grained feature extraction module are sequentially input into the second convolution layer, the BN layer and the Bi-SRU layer, and the global feature information containing the spatial features and the temporal features of the measured original vibration signal of the insulating bearing is output; The Bi-SRU layer includes a forward simple recurrent unit network and a backward simple recurrent unit network, each of which is composed of a plurality of simple recurrent units; the simple recurrent unit includes an input gate, a forget gate, a reset gate, an output gate, an internal state, and a residual connection; sequential and reverse propagation is performed through the bidirectional simple recurrent unit network of the Bi-SRU layer to extract global features; The calculation process of the simple cycle unit is as follows: f t =σ(W f x t +b f ) r t =σ(W r x t +b r ) y t =σ(W * x t ) A t =f t ·e·A t-1 +(1-f t )·e·y t h t =r t ·e·g(A t )+(1-r t )·e·T t Among them, f t is the hidden state of the forget gate, r t To reset the hidden state of the door, W f , W r and W * is the parameter matrix of the simple recurrent unit, b f 、b r is the bias of the simple recurrent unit, which is the parameter vector to be learned during training. t represents the input features at time t, y t represents the output of the Bi-SRU layer at time t, h t represents the output of the simple recurrent unit at the current time t, h t-1 represents the output of the simple recurrent unit at the previous time t-1, A t represents the hidden state of the simple recurrent unit at the current time t, and e represents the multiplication operation of the corresponding elements of the matrix; A bidirectional simple recurrent unit network encoder is used to capture the temporal features of both directions at the same time, and then the features at the same time are fused to obtain the final feature expression at each time. The feature extraction layer is based on the flow matrix x = [x1, x2, ..., x n ] is the input, the forward simple recurrent unit network of the bidirectional simple recurrent unit network From x1 to x n Read input feature x; backward simple recurrent unit network From x n To x1 read the input feature x: in, is the forward hidden state, collecting x in the stream sequence t Previous information; For the backward hidden state, collect x t The information after that; the initial hidden state vector and are all zero vectors, and Spliced ​​together, we get the summary information H of the entire flow at time step t t ,Right now In order to further obtain the high-order features of the traffic data, the bidirectional simple recurrent unit network is stacked into M layers, and the input of the i-th layer at time t is the output of the i-1th layer at that time As shown below: The number of bytes per stream, N, and the number of simple cyclic unit network stacking layers, M, will be determined experimentally. The summary information of the entire stream at the i-th layer at time t is expressed as Concatenate the hidden states generated at each time step of the last layer to get the output of the feature extraction layer: After the feature extraction process of M-layer bidirectional simple recurrent unit network, the feature vector H N Highly aggregated bidirectional information of insulated bearings.

5. The wind turbine bearing health assessment method based on the lightweight coarse-grained and fine-grained feature fusion model according to claim 1 is characterized in that: The feature fusion recognition module includes: an exchange axis layer, a self-attention layer, a data flattening layer and an output layer; The global feature information containing the spatial features and temporal features of the measured original vibration signal of the insulating bearing output by the coarse-grained feature extraction module is used as the input of this module. The internal feature representation is obtained by exchanging the axial layer, and then the self-attention layer is used to calculate the weights of different channels and output them through the Sigmoid function. The data flattening layer is used to replace the original fully connected layer, which can reduce model parameters and computational complexity. Finally, the Softmax function converts the neuron output into a probability distribution of 10 types of insulating bearing faults, and the features with different weights are input into the classifier for fault classification, thereby realizing the classification of insulating bearing faults.

6. The wind turbine bearing health assessment method based on the lightweight coarse-grained and fine-grained feature fusion model according to claim 2 is characterized in that: The self-attention layer can suppress redundant information irrelevant to the target result from complex and numerous information. The input time series is: For each input X i Linearly map it to three different spaces to obtain the query vector Key Vector Sum value vector The output time series is For the entire input time series X, the linear mapping process is expressed as: in, are the parameter matrices of the linear mapping, Q = [q1, q2, …, q n ], K=[k1,k2,…,k n ]、V=[v1,v2,…,v n ] are respectively composed of the query vector q i , key vector k i Sum value vector v i The matrix D k and N represent the matrix dimensions respectively; For each query vector q in the input time series n ∈Q, get the output vector Y n : Among them, n∈[1,2,…,N] is the position of the output vector sequence, j∈[1,2,…,N] is the position of the input vector sequence, and a nj Indicates the weight of the n-th output paying attention to the j-th input.

Images