A small sample fault classification method and system for a gearbox
By introducing a deep convolutional generative adversarial network with a self-attention mechanism and transfer learning techniques, high-quality sample data is generated, which solves the problem of poor classification performance in small sample fault diagnosis of gearboxes and improves diagnostic accuracy and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2023-05-29
- Publication Date
- 2026-05-01
AI Technical Summary
Existing gearbox fault diagnosis methods suffer from poor fault classification performance under small sample data conditions. In particular, convolutional neural networks do not perform well in training small sample datasets, and the generated pseudo samples are of low quality, affecting the accuracy of diagnosis.
A deep convolutional generative adversarial network (SA-DCGAN) with an improved self-attention mechanism is used to generate sufficient sample data. The parameters of the pre-trained source domain model are transferred to the target domain model through transfer learning techniques, and fault classification is performed by combining it with a dense convolutional network.
It improves the quality of fault sample data, saves training time and cost for the target domain model, and enhances the generalization ability and classification accuracy of the fault classification model.
Smart Images
Figure CN116578907B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine fault diagnosis, and specifically to a small sample fault classification method for gearboxes. Background Technology
[0002] In recent years, with increasing global environmental awareness and energy demand, pollution-free renewable clean energy has seen significant development. According to data from the Global Wind Energy Council, wind energy, as one of the pollution-free renewable clean energy sources, is one of the most reliable renewable energy sources, and with the vigorous development of the wind energy industry, the cumulative installed capacity has increased substantially. However, many wind turbines already in operation are built in oceans, mountains, and remote areas due to abundant wind resources. At the same time, these wind turbines also face harsh operating and environmental conditions, aging mechanical components, and other problems, seriously affecting the reliability and safety of wind turbine units, and potentially causing significant economic losses. Statistics show that the cumulative failure rate of wind turbine gearboxes in existing wind farms across China is as high as 40%-50%, indicating that the gearbox is the mechanical component with the highest failure rate in wind turbines. Wind turbine gearboxes are often referred to as gearboxes because they consist of a series of gears used to convert the wind turbine's rotational speed into the appropriate speed required by the generator. Gearboxes are typically made of high-strength alloys to withstand high loads and frequent operation. Common gearbox failures include bearing wear, pitting, and broken teeth. These failures gradually worsen, leading to prolonged downtime of wind turbines and high maintenance costs. Therefore, implementing advanced fault diagnosis strategies before gearbox failures escalate is crucial for effectively reducing maintenance difficulty, minimizing economic losses, and improving the operational safety of wind turbines.
[0003] Currently, methods applied to gearbox fault diagnosis mainly fall into two categories: model-based and model-free methods. Model-based methods primarily establish state-space models using techniques such as Kalman filters, intermediate observers, and extended state observers, but this requires an accurate system model. However, obtaining an accurate system model is difficult in actual wind turbines. Model-free methods do not require an accurate system model; they are defined from a pattern recognition perspective. Model-free methods mainly utilize artificial intelligence methods such as backpropagation neural network algorithms and convolutional neural network algorithms to automatically learn and extract features from gearbox vibration data, thereby determining whether a fault exists. Among these, the backpropagation neural network algorithm has self-learning capabilities and can implement complex nonlinear mappings; however, this algorithm requires significant computational resources and time for training and is prone to getting trapped in local optima, exhibiting "overfitting" and affecting its predictive ability. Convolutional neural networks (CNNs) have powerful classification capabilities, but they also require significant computational resources and time for training. The performance of CNNs is highly dependent on the quality and size of the sample dataset, and they perform poorly on small-scale sample datasets.
[0004] However, wind turbines are highly reliable devices during wind farm operation and maintenance, operating normally most of the time. The probability of gearbox failure is low, resulting in limited real-world vibration data. With only a small amount of real-world vibration data, it's difficult to train an effective fault diagnosis model. Specifically, directly using this limited vibration data for deep learning training can lead to overfitting, poor generalization, and a low fault identification rate for the target gearbox, making it unsuitable for effective wind turbine gearbox fault identification.
[0005] To address this, existing technologies typically mix pseudo-samples generated by Generative Adversarial Networks (GANs) with real fault samples to compensate for the small sample size problem caused by a lack of real data. However, in practice, due to limitations in GAN network structure and computational cost, the similarity between GAN-generated data and real data is low. Moreover, to solve the problem of insufficient fault sample data, the number of pseudo-samples generated by GANs often far exceeds the number of real samples. This results in a large number of low-quality pseudo-samples affecting the classification boundary of the fault diagnosis model, thus making it difficult to distinguish the fault type of the gearbox vibration signal that needs to be diagnosed. Summary of the Invention
[0006] The technical problem to be solved by this invention is that current fault diagnosis methods for gearboxes suffer from a lack of real fault sample data and poor fault classification results. This invention proposes a small-sample fault classification method and system for gearboxes to solve the above-mentioned technical problems.
[0007] This invention adopts the following technical solution: a small sample fault classification method for gearboxes, comprising the following steps:
[0008] Step S1: Collect gearbox vibration signal data X for different fault types, and preprocess the gearbox vibration signal data X to obtain dataset X. i ;
[0009] Step S2: Obtain the network structure of DCGAN, introduce the self-attention mechanism into the network structure of DCGAN, and denote the improved DCGAN as SA-DCGAN. The SA-DCGAN includes a generator and a discriminator.
[0010] Step S3, using dataset X i The SA-DCGAN is trained, and a noise vector is input into the trained SA-DCGAN to output several samples Y of different fault types. These samples Y constitute the sample dataset Y. i ;
[0011] Step S4: Construct a dense convolutional network model using the sample dataset Y. i The dense convolutional network model is pre-trained, and the pre-trained dense convolutional network model is used as the source domain model;
[0012] Step S5: Construct a target domain model, transfer the parameters of the source domain model to the target domain model, and fine-tune the target domain model using gearbox vibration signal data X. Use the fine-tuned target domain model as the fault classification model.
[0013] Step S6: Input the gearbox vibration signal for which the fault category needs to be diagnosed into the fault classification model to obtain the corresponding gearbox small sample fault classification result.
[0014] DCGAN (Deep Convolutional Generative Adversarial Networks) is a type of Generative Adversarial Network (GAN). DCGAN uses deep convolutional neural networks and adversarial training to learn how to generate realistic and sharp images. A Dense Convolutional Network is a type of convolutional neural network whose main characteristic is the use of densely connected convolutional structures in the deep layers of the network, connecting multiple locations on each feature map rather than using local connections.
[0015] By introducing a self-attention mechanism into the DCGAN network structure, SA-DCGAN is formed, and the dataset X is used. iTraining SA-DCGAN enables the generation of faulty sample data that is closer to real sample data while generating sufficient sample data, thus improving the quality of faulty sample data. This is because the self-attention mechanism can model long-range dependencies between pixels. This fully connected characteristic helps the generator to better utilize the global contextual information of the input image and generate images with more coherent structures. Therefore, introducing the self-attention mechanism into the DCGAN network structure helps DCGAN generate more realistic details, thereby improving the quality of the generated sample data.
[0016] By using sufficient sample data generated by SA-DCGAN to pre-train a dense convolutional network model, the technical problem of poor training performance of dense convolutional networks on small-scale sample datasets due to the scarcity of real fault sample data is solved.
[0017] By using transfer learning techniques to transfer the parameters of a pre-trained source domain model to the target domain model, the training time and cost of the target domain model can be saved, while improving the generalization ability of the target domain model, since the source domain model has already been trained with sufficient sample data.
[0018] By fine-tuning the target domain model using a small amount of real gearbox sample data, the influence of a large number of generated samples on the boundary of the fault classification model is mitigated, thereby improving the classification accuracy of the fault classification model.
[0019] Preferably, in step S2, the method for introducing the self-attention mechanism into the DCGAN network structure includes:
[0020] A self-attention module is embedded after the fourth deconvolution layer of the generator and the fourth convolution layer of the discriminator. The self-attention module is used to learn global feature information.
[0021] The DCGAN network structure includes a generator and a discriminator. The generator consists of five deconvolutional layers and one fully connected layer, while the discriminator consists of five convolutional layers and one fully connected layer. A self-attention module is an attention mechanism used in deep learning.
[0022] By embedding a self-attention module after the fourth deconvolution layer of the generator, the cross-space dependencies in the feature map can be strengthened, thereby enhancing the generation capability of DCGAN.
[0023] By embedding a self-attention module after the fourth convolutional layer of the discriminator, long-range dependencies in the noise vector sequence can be learned, which significantly enhances the representation and learning capabilities of DCGAN, enabling the discriminator to distinguish between true and false signals more precisely.
[0024] Embedding the self-attention module after the fourth deconvolutional layer or convolutional layer is a reasonable position. This is because at this position, the discriminator has already extracted higher-level image features from pixel-level features. Therefore, the self-attention module can further extract the relationships between these higher-level image features, learn richer high-dimensional feature representations, and thus exert its powerful capabilities.
[0025] Preferably, the method for the self-attention module to learn global feature information includes:
[0026] The noise vector sequence after passing through the DCGAN convolutional or deconvolutional layer is split into a query matrix Q, a key matrix K, and a value matrix V;
[0027] Calculate the dot product between the query matrix Q and the key matrix K, and scale the calculated dot product to obtain the weight W of the values. V ;
[0028] Weight W of the value V Multiplying with the value matrix V yields the feature information between the noise vector sequences.
[0029] Preferably, in step S4, the sample dataset Y is used. i Methods for pre-training the dense convolutional network model include:
[0030] The sample dataset Y i The samples Y in the model are input into the initial convolutional layer to generate sample features;
[0031] The sample features are input into the first dense block to extract primary features;
[0032] The primary features are input into the transition layer for compression, and the compressed primary features are input into subsequent dense blocks until the final dense block is input, thereby extracting high-level abstract features.
[0033] After the high-level abstract features are processed through global pooling and classification layers, the fault type corresponding to each sample Y is output.
[0034] The dense convolutional network model consists of several convolutional layers, each followed by a dense block. The first dense block is followed by a transition layer, and the last dense block is followed by a global pooling and classification layer.
[0035] Preferably, in step S5, the method for transferring the parameters of the source domain model to the target domain model includes:
[0036] The parameters of the source domain model are directly used as the parameters of the target domain model.
[0037] Preferably, in step S5, the method for fine-tuning the target domain model using gearbox vibration signal data X includes:
[0038] The gearbox vibration signal data X is input into the initial convolutional layer to generate real sample features;
[0039] The real sample features are input into the first dense block to extract primary features;
[0040] The primary features are input into the transition layer for compression, and the compressed primary features are input into subsequent dense blocks until the final dense block is input, thereby extracting high-level abstract features.
[0041] After the high-level abstract features are processed through global pooling and classification layers, the fault type corresponding to each gearbox vibration signal data X is output.
[0042] Preferably, in step S1, the gearbox vibration signal data X is preprocessed to obtain dataset X. i The methods include:
[0043] The gearbox vibration signal data X is converted into two-dimensional time-frequency image data using the short-time Fourier transform method;
[0044] The two-dimensional time-frequency image data is smoothed and filtered.
[0045] Geometric transformation is performed on the processed two-dimensional time-frequency image data to obtain the augmented image data;
[0046] The expanded image data and the two-dimensional time-frequency image data constitute dataset X. i .
[0047] A small-sample fault classification system for gearboxes includes:
[0048] The signal acquisition module is used to acquire gearbox vibration signal data X for different fault types;
[0049] The preprocessing module is used to preprocess the gearbox vibration signal data X to obtain dataset X. i ;
[0050] The sample generation module is used to obtain the network structure of DCGAN, introduce a self-attention mechanism into the DCGAN network structure, and denote the improved DCGAN as SA-DCGAN. SA-DCGAN includes a generator and a discriminator, using dataset X. i The SA-DCGAN is trained, and a noise vector is input into the trained SA-DCGAN to output several samples Y of different fault types. These samples Y constitute the sample dataset Y. i ;
[0051] The model building module is used to build dense convolutional network models using the sample dataset Y. i The dense convolutional network model is pre-trained, and the pre-trained dense convolutional network model is used as the source domain model to construct the target domain model. The parameters of the source domain model are transferred to the target domain model, and the target domain model is fine-tuned using gearbox vibration signal data X. The fine-tuned target domain model is used as the fault classification model.
[0052] The fault classification module is used to input the gearbox vibration signal that needs to be diagnosed into the fault classification model to obtain the corresponding gearbox small sample fault classification results.
[0053] The beneficial technical effects of this invention include: employing a small-sample fault classification method and system for gearboxes, forming SA-DCGAN by introducing a self-attention mechanism into the DCGAN network structure, and using dataset X. i Training SA-DCGAN enables the generation of fault sample data that more closely resembles real sample data while generating sufficient sample data, thus improving the quality of fault sample data. Pre-training a dense convolutional network model using sufficient sample data generated by SA-DCGAN addresses the technical problem of poor training performance of dense convolutional networks on small-scale datasets due to limited real fault sample data. Transfer learning techniques are used to transfer the parameters of the pre-trained source domain model to the target domain model, saving training time and cost while improving the generalization ability of the target domain model, since the source domain model has already been trained with sufficient sample data. Fine-tuning the target domain model using a small amount of real gearbox sample data mitigates the impact of a large number of generated samples on the boundary of the fault classification model, thereby improving the classification accuracy of the fault classification model.
[0054] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description
[0055] The invention will be further described below with reference to the accompanying drawings:
[0056] Figure 1 This is a flowchart of a small-sample fault classification method for gearboxes according to an embodiment of the present invention.
[0057] Figure 2 This is a schematic diagram of the structure of SA-DCGAN according to an embodiment of the present invention.
[0058] Figure 3 This is a flowchart illustrating the method for learning global feature information using a self-attention module according to an embodiment of the present invention.
[0059] Figure 4 This is a schematic diagram illustrating the principle of the SA-DCGAN output sample Y after training in an embodiment of the present invention.
[0060] Figure 5 This is a schematic diagram of the small sample fault classification system for a gearbox according to an embodiment of the present invention.
[0061] The module consists of: 1. Signal acquisition module, 2. Preprocessing module, 3. Sample generation module, 4. Model building module, and 5. Fault classification module. Detailed Implementation
[0062] The technical solutions of the embodiments of the present invention will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of the present invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of the present invention.
[0063] In the following description, terms such as “inner,” “outer,” “upper,” “lower,” “left,” and “right” are used only to indicate orientation or positional relationship for the convenience of describing the embodiments and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.
[0064] This application provides a small-sample fault classification method for gearboxes. Please refer to the appendix. Figure 1 This includes the following steps:
[0065] Step S1: Collect gearbox vibration signal data X for different fault types, preprocess the gearbox vibration signal data X to obtain dataset X. i .
[0066] The main types of gearbox failures include gear wear, gear breakage, bearing wear, rolling element breakage, lubricating oil contamination, and rolling element detachment.
[0067] Specifically, the gearbox vibration signal data X is preprocessed to obtain the dataset X. i The methods include:
[0068] The gearbox vibration signal data X is converted into two-dimensional time-frequency image data using the short-time Fourier transform method;
[0069] Smoothing and filtering processes are applied to two-dimensional time-frequency image data.
[0070] Geometric transformation is performed on the processed two-dimensional time-frequency image data to obtain the augmented image data;
[0071] The augmented image data and the two-dimensional time-frequency image data constitute dataset X. i .
[0072] Since SA-DCGAN can only perform data augmentation on two-dimensional images, it is necessary to preprocess the gearbox vibration signal data X to convert the one-dimensional vibration signal data into two-dimensional time-frequency image data.
[0073] For example, in order to simulate the case of small sample size, 20 gearbox vibration signal data X were selected for each type of fault. In the experiment, it was found that due to the small number of samples, the learning ability of SA-DCGAN was insufficient to fit high-quality time-frequency image samples. Therefore, geometric transformations such as rotation, translation, and scaling were used to expand the time-frequency image samples of the two-dimensional time-frequency image data converted from the gearbox vibration signal data X.
[0074] Step S2: Obtain the network structure of DCGAN, introduce the self-attention mechanism into the network structure of DCGAN, and denote the improved DCGAN as SA-DCGAN. SA-DCGAN includes a generator and a discriminator.
[0075] Among them, DCGAN (Deep Convolutional Generative Adversarial Networks) is a type of Generative Adversarial Network (GAN). DCGAN uses deep convolutional neural networks and adversarial training to learn to generate realistic and clear images.
[0076] Step S3, using dataset X i Train the SA-DCGAN, input the noise vector into the trained SA-DCGAN, and output several samples Y of different fault types. These samples Y constitute the sample dataset Y. i .
[0077] For example, using dataset X i The specific implementation method for training SA-DCGAN is as follows:
[0078] 1) Before training: A. Select the loss functions for the discriminator and generator. In this embodiment, the binary cross-entropy loss function is used for the discriminator, and the least squares loss function is used for the generator. B. Define the optimizers for the discriminator and generator respectively. In this embodiment, the Adam optimizer is selected, and appropriate learning rate parameters are chosen for the discriminator and generator. C. Initialize the network parameters, such as using a Gaussian distribution to initialize the weights.
[0079] 2) During training: The learning rates of the discriminator and generator are dynamically adjusted to achieve a relative balance between their loss functions (i.e., Nash equilibrium). Specifically, if the discriminator loss function decreases rapidly, the discriminator's learning rate needs to be reduced; if the generator loss function decreases slowly, the generator's learning rate needs to be increased. SA-DCGAN training is most effective only when the discriminator and generator loss functions decrease synchronously.
[0080] Step S4: Construct a dense convolutional network model using the sample dataset Y. i A pre-trained dense convolutional network model is used as the source domain model.
[0081] Among them, a dense convolutional network is a type of convolutional neural network. Its main feature is that it uses densely connected convolutional structures in the deep layers of the network, that is, connecting multiple locations on each feature map, rather than local connections.
[0082] Step S5: Construct the target domain model, transfer the parameters of the source domain model to the target domain model, and fine-tune the target domain model using gearbox vibration signal data X. Use the fine-tuned target domain model as the fault classification model.
[0083] The network structure of the target domain model is consistent with that of the source domain model.
[0084] Step S6: Input the gearbox vibration signal for which the fault category needs to be diagnosed into the fault classification model to obtain the corresponding gearbox small sample fault classification results.
[0085] By introducing a self-attention mechanism into the DCGAN network structure, SA-DCGAN is formed, and the dataset X is used. i Training SA-DCGAN enables the generation of faulty sample data that is closer to real sample data while generating sufficient sample data, thus improving the quality of faulty sample data. This is because the self-attention mechanism can model long-range dependencies between pixels. This fully connected characteristic helps the generator to better utilize the global contextual information of the input image and generate images with more coherent structures. Therefore, introducing the self-attention mechanism into the DCGAN network structure helps DCGAN generate more realistic details, thereby improving the quality of the generated sample data.
[0086] By using sufficient sample data generated by SA-DCGAN to pre-train a dense convolutional network model, the technical problem of poor training performance of dense convolutional networks on small-scale sample datasets due to the scarcity of real fault sample data is solved.
[0087] By using transfer learning techniques to transfer the parameters of a pre-trained source domain model to a target domain model, the training time and cost of the target domain model can be saved, while improving the generalization ability of the target domain model, since the source domain model has already been trained with sufficient sample data.
[0088] By fine-tuning the target domain model using a small amount of real gearbox sample data, the influence of a large number of generated samples on the boundary of the fault classification model is mitigated, thereby improving the classification accuracy of the fault classification model.
[0089] On the other hand, in this embodiment, the method of introducing the self-attention mechanism into the DCGAN network structure in step S2 includes:
[0090] Self-attention modules are embedded after the fourth deconvolution layer of the generator and the fourth convolution layer of the discriminator. These self-attention modules are used to learn global feature information.
[0091] The DCGAN network structure includes a generator and a discriminator. The generator consists of five deconvolutional layers and one fully connected layer, while the discriminator consists of five convolutional layers and one fully connected layer. A self-attention module is an attention mechanism used in deep learning.
[0092] Please refer to the attached diagram for the structure of the improved DCGAN, namely SA-DCGAN. Figure 2 The left side shows the structure of the generator, and the right side shows the structure of the discriminator.
[0093] By embedding a self-attention module after the fourth deconvolution layer of the generator, the cross-space dependencies in the feature map can be strengthened, thereby enhancing the generation capability of DCGAN.
[0094] By embedding a self-attention module after the fourth convolutional layer of the discriminator, long-range dependencies in the noise vector sequence can be learned, which significantly enhances the representation and learning capabilities of DCGAN, enabling the discriminator to distinguish between true and false signals more precisely.
[0095] Embedding the self-attention module after the fourth deconvolutional layer or convolutional layer is a reasonable position. This is because at this position, the discriminator has already extracted higher-level image features from pixel-level features. Therefore, the self-attention module can further extract the relationships between these higher-level image features, learn richer high-dimensional feature representations, and thus exert its powerful capabilities.
[0096] On the other hand, in this embodiment, please refer to the appendix. Figure 3 Methods for self-attention modules to learn global feature information include:
[0097] Step S201: The noise vector sequence that has passed through the DCGAN convolutional layer or deconvolutional layer is split into a query matrix Q, a key matrix K, and a value matrix V.
[0098] Step S202: Calculate the inner product of the query matrix Q and the key matrix K, and scale the calculated inner product to obtain the weight W of the value. V ;
[0099] Step S203, assign weights W to the values. V Multiplying with the value matrix V yields the feature information between the noise vector sequences.
[0100] Here, the query matrix Q represents the object to be queried, the key matrix K corresponds to a certain feature in the input, and the value matrix V corresponds to the descriptor of the feature.
[0101] For example, in this embodiment, the specific implementation method of step S201 is as follows: convert the noise vector sequence into a two-dimensional matrix, let Q = K = V, that is, select the first n rows of the matrix to form the query matrix Q, select the middle part of the matrix to form the key matrix K, and select the last n rows of the matrix to form the value matrix V.
[0102] For example, the weight W of the value in step S202 V It can be obtained through the following formula:
[0103]
[0104] Where, d k is the length of the noise vector sequence.
[0105] By calculating the inner product of the query matrix Q and the key matrix K, the relevance between Q and K can be determined, and the weight of each value can be assigned accordingly. The higher the weight of a value, the more relevant it is.
[0106] This embodiment obtains feature information between noise vector sequences through a self-attention module. This feature information fully considers the dependencies between noise vectors in the sequence. Compared with directly concatenating noise vectors together, it has a stronger feature expression capability. Using the self-attention feature information of the noise vector sequence as the input of the generator in SA-DCGAN can yield richer generated features.
[0107] SA-DCGAN is formed by embedding self-attention modules into the DCGAN network structure, and the dataset X is used. i By training SA-DCGAN, we were able to generate fault sample data that is closer to real sample data while SA-DCGAN generates a sufficient amount of sample data, thereby improving the quality of fault sample data.
[0108] On the other hand, in this embodiment, in step S3, the noise vector is input into the trained SA-DCGAN, and several samples Y of different fault types are output. Please refer to the attached diagram for the principle. Figure 4 .
[0109] On the other hand, in this embodiment, in step S4, the sample dataset Y is used. i Methods for pre-training dense convolutional network models include:
[0110] The sample dataset Y i The samples Y in the model are input into the initial convolutional layer to generate sample features;
[0111] Input the sample features into the first dense block to extract the primary features;
[0112] The primary features are input into the transition layer for compression, and the compressed primary features are input into subsequent dense blocks until the final dense block is input, in which high-level abstract features are extracted.
[0113] After passing the high-level abstract features through global pooling and classification layers, the fault type corresponding to each sample Y is output.
[0114] The dense convolutional network model consists of several convolutional layers, each followed by a dense block. The first dense block is followed by a transition layer, and the last dense block is followed by a global pooling and classification layer.
[0115] Specifically, using the sample dataset Y i The method for pre-training dense convolutional network models is as follows: the input sample dataset Y i The sample Y in the model first generates sample features through an initial convolutional layer. These sample features are then input into the first dense block to extract primary features. After compression through a transition layer, the features are then transmitted to subsequent dense blocks to continuously extract features until the final dense block extracts high-level abstract features. Finally, global pooling and a classification layer are used to achieve the final classification, which outputs the fault type corresponding to sample Y, thus completing the pre-training of the dense convolutional network model.
[0116] By using sufficient sample data generated by SA-DCGAN to pre-train a dense convolutional network model, the technical problem of poor training performance of dense convolutional networks on small-scale sample datasets due to the scarcity of real fault sample data is solved.
[0117] On the other hand, in this embodiment, step S5, the method for transferring the parameters of the source domain model to the target domain model includes:
[0118] The parameters of the source domain model are directly used as the parameters of the target domain model.
[0119] Since the source domain model and the target domain model have the same network structure, the parameters of the target domain model can be obtained by direct transfer learning. By using transfer learning techniques to transfer the parameters of the pre-trained source domain model to the target domain model, the training time and cost of the target domain model can be saved, while improving the generalization ability of the target domain model, because the source domain model has already been trained with sufficient sample data.
[0120] On the other hand, in this embodiment, the method for fine-tuning the target domain model using gearbox vibration signal data X in step S5 includes:
[0121] The gearbox vibration signal data X is input into the initial convolutional layer to generate real sample features;
[0122] Input the features of real samples into the first dense block to extract primary features;
[0123] The primary features are input into the transition layer for compression, and the compressed primary features are input into subsequent dense blocks until the final dense block is input, in which high-level abstract features are extracted.
[0124] After passing the high-level abstract features through global pooling and classification layers, the fault type corresponding to each gearbox vibration signal data X is output.
[0125] By fine-tuning the target domain model using a small amount of real gearbox sample data, the influence of a large number of generated samples on the boundary of the fault classification model is mitigated, thereby improving the classification accuracy of the fault classification model.
[0126] Under the same noise environment, the small sample fault classification method (DATL) for gearboxes proposed in this invention was compared with other different comparative methods. The small sample fault diagnosis performance (i.e., the accuracy of obtaining the corresponding small sample fault classification results) was compared when the number of gearbox vibration signals requiring fault classification was 2, 5, and 10. The experimental results are shown in the table below:
[0127]
[0128] This experiment used four comparison methods:
[0129] MLP-based fault diagnosis method (see Waqar T, Demetgul M. Thermal analysis MLP neural network based fault diagnosis on worm gears[J]. Measurement, 2016, 86:56-66.);
[0130] A fault diagnosis method based on CNN (see the literature Gou L, Li H, Zheng H, et al. Aeroengine control system sensor fault diagnosis based on CWT and CNN[J]. Mathematical Problems in Engineering, 2020).
[0131] A fault diagnosis method based on MSCNN (see the literature Jiang G, He H, Yan J, et al. Multiscale convolutional neural networks for fault diagnosis of wind turbine gearbox[J]. IEEE Transactions on Industrial Electronics, 2018, 66(4):3196-3207.);
[0132] A fault diagnosis method based on ResNet-50 (see Wen L, Li X, Gao LA. Transfer convolutional neural network for fault diagnosis based on ResNet-50[J].Neural Computing and Applications, 2020, 32:6111-6124.).
[0133] The table shows the accuracy of different fault classification methods for each type of gearbox vibration signal requiring fault diagnosis, from left to right: MLP, CNN, MSCNN, ResNet50, the improved SA-DCGAN of this invention, and DATL proposed in this invention. As can be seen from the table, compared to the other methods, the accuracy of the small-sample fault classification method (DATL) for gearboxes proposed in this invention is significantly higher. This indicates that the method of this invention can accurately diagnose gearbox fault types under small sample conditions and has better fault classification performance.
[0134] On the other hand, embodiments of this application also provide a small-sample fault classification system for gearboxes; please refer to the appendix. Figure 5 ,include:
[0135] Signal acquisition module 1 is used to acquire gearbox vibration signal data X for different fault types;
[0136] Preprocessing module 2 is used to preprocess the gearbox vibration signal data X to obtain dataset X.i ;
[0137] Sample generation module 3 is used to obtain the network structure of DCGAN. It introduces a self-attention mechanism into the DCGAN network structure, and the improved DCGAN is denoted as SA-DCGAN. SA-DCGAN includes a generator and a discriminator, using dataset X. i Train the SA-DCGAN, input the noise vector into the trained SA-DCGAN, and output several samples Y of different fault types. These samples Y constitute the sample dataset Y. i ;
[0138] Model building module 4 is used to build dense convolutional network models, using the sample dataset Y. i A pre-trained dense convolutional network model is used as the source domain model to construct a target domain model. The parameters of the source domain model are transferred to the target domain model, and the target domain model is fine-tuned using gearbox vibration signal data X. The fine-tuned target domain model is then used as a fault classification model.
[0139] The fault classification module 5 is used to input the gearbox vibration signal that needs to be diagnosed into the fault classification model to obtain the corresponding gearbox small sample fault classification results.
[0140] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments above. Any modifications that do not depart from the functional and structural principles of the present invention will be included within the scope of the claims.
Claims
1. A small-sample fault classification method for gearboxes, characterized in that, Includes the following steps: Step S1: Collect gearbox vibration signal data X for different fault types, and preprocess the gearbox vibration signal data X to obtain a dataset. ; Step S2: Obtain the network structure of DCGAN, introduce the self-attention mechanism into the network structure of DCGAN, and denote the improved DCGAN as SA-DCGAN. The SA-DCGAN includes a generator and a discriminator. Step S3, using the dataset The SA-DCGAN is trained, and a noise vector is input into the trained SA-DCGAN to output several samples Y of different fault types. These samples Y constitute the sample dataset. ; Step S4: Construct a dense convolutional network model. The network structure of the dense convolutional network model includes several convolutional layers, each followed by a dense block. A transition layer follows the first dense block, and a global pooling and classification layer follows the last dense block. The sample dataset is used. The dense convolutional network model is pre-trained, and the pre-trained dense convolutional network model is used as the source domain model; Step S5: Construct a target domain model, transfer the parameters of the source domain model to the target domain model, and fine-tune the target domain model using gearbox vibration signal data X. Use the fine-tuned target domain model as the fault classification model. Step S6: Input the gearbox vibration signal for which the fault category needs to be diagnosed into the fault classification model to obtain the corresponding gearbox small sample fault classification result; In step S2, the method of introducing the self-attention mechanism into the DCGAN network structure includes: A self-attention module is embedded after the fourth deconvolution layer of the generator and the fourth convolution layer of the discriminator. This self-attention module is used to learn global feature information. The method by which the self-attention module learns global feature information includes: The noise vector sequence after passing through the DCGAN convolutional or deconvolutional layer is split into a query matrix Q, a key matrix K, and a value matrix V; Calculate the dot product between the query matrix Q and the key matrix K, and scale the calculated dot product to obtain the weights of the values. ; Weight of values Multiplying by the value matrix V yields the feature information between the noise vector sequences; In step S5, the method for fine-tuning the target domain model using gearbox vibration signal data X includes: The gearbox vibration signal data X is input into the initial convolutional layer to generate real sample features; The real sample features are input into the first dense block to extract primary features; The primary features are input into the transition layer for compression, and the compressed primary features are input into subsequent dense blocks until the final dense block is input, thereby extracting high-level abstract features. After the high-level abstract features are processed through global pooling and classification layers, the fault type corresponding to each gearbox vibration signal data X is output.
2. The small-sample fault classification method for gearboxes as described in claim 1, characterized in that, In step S4, the sample dataset is used. Methods for pre-training the dense convolutional network model include: The sample dataset The samples Y in the model are input into the initial convolutional layer to generate sample features; The sample features are input into the first dense block to extract primary features; The primary features are input into the transition layer for compression, and the compressed primary features are input into subsequent dense blocks until the final dense block is input, thereby extracting high-level abstract features. After the high-level abstract features are processed through global pooling and classification layers, the fault type corresponding to each sample Y is output.
3. The small-sample fault classification method for gearboxes as described in claim 1, characterized in that, In step S5, the method for transferring the parameters of the source domain model to the target domain model includes: The parameters of the source domain model are directly used as the parameters of the target domain model.
4. A small-sample fault classification method for gearboxes as described in any one of claims 1 to 3, characterized in that, In step S1, the gearbox vibration signal data X is preprocessed to obtain a dataset. The methods include: The gearbox vibration signal data X is converted into two-dimensional time-frequency image data using the short-time Fourier transform method; The two-dimensional time-frequency image data is smoothed and filtered. Geometric transformation is performed on the processed two-dimensional time-frequency image data to obtain the augmented image data; The expanded image data and the two-dimensional time-frequency image data constitute a dataset. .
5. A small-sample fault classification system for gearboxes, characterized in that, include: The signal acquisition module is used to acquire gearbox vibration signal data X for different fault types; The preprocessing module is used to preprocess the gearbox vibration signal data X to obtain a dataset. ; The sample generation module is used to obtain the network structure of DCGAN, introduce a self-attention mechanism into the DCGAN network structure, and denote the improved DCGAN as SA-DCGAN. SA-DCGAN includes a generator and a discriminator, and uses a dataset... The SA-DCGAN is trained, and a noise vector is input into the trained SA-DCGAN to output several samples Y of different fault types. These samples Y constitute the sample dataset. ; The model building module is used to build dense convolutional network models using sample datasets. The dense convolutional network model is pre-trained, and the pre-trained dense convolutional network model is used as the source domain model to construct the target domain model. The parameters of the source domain model are transferred to the target domain model, and the target domain model is fine-tuned using gearbox vibration signal data X. The fine-tuned target domain model is used as the fault classification model. The fault classification module is used to input the gearbox vibration signal that needs to be diagnosed into the fault classification model to obtain the corresponding gearbox small sample fault classification result; The sample generation module is also used to perform the following steps: A self-attention module is embedded after the fourth deconvolution layer of the generator and the fourth convolution layer of the discriminator. This self-attention module is used to learn global feature information. The method by which the self-attention module learns global feature information includes: The noise vector sequence after passing through the DCGAN convolutional or deconvolutional layer is split into a query matrix Q, a key matrix K, and a value matrix V; Calculate the dot product between the query matrix Q and the key matrix K, and scale the calculated dot product to obtain the weights of the values. ; Weight of values Multiplying by the value matrix V yields the feature information between the noise vector sequences; The model building module is also used to perform the following steps: The gearbox vibration signal data X is input into the initial convolutional layer to generate real sample features; The real sample features are input into the first dense block to extract primary features; The primary features are input into the transition layer for compression, and the compressed primary features are input into subsequent dense blocks until the final dense block is input, thereby extracting high-level abstract features. After the high-level abstract features are processed through global pooling and classification layers, the fault type corresponding to each gearbox vibration signal data X is output.
Citation Information
Patent Citations
Gearbox fault diagnosis model construction method based on CNN and transfer learning
CN115130521A