A fault data generation method for deep SMOTE integrated with HCFF
Through the integrated HCFF deep SMOTE method, the problem of insufficient sample authenticity and diversity in intelligent fault diagnosis is solved, and samples with rich fault characteristics are generated, which significantly improves the recognition ability and robustness of the fault diagnosis model, especially in rare faults. The performance was significant in the case.
Patent Information
- Application Number
- CN202411415209.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-11
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-10-11
AI Technical Summary
The prior art has problems with insufficient sample authenticity and diversity in intelligent fault diagnosis, which leads to low ability to identify a few types of faults, especially in rare fault situations, which is difficult to generate samples with sufficient fault characteristics.
A deep SMOTE method integrating HCFF is proposed to generate samples rich in fault characteristics through multi-level feature extraction and fusion mechanisms, enhancing the recognition ability and robustness of fault diagnosis models. The method includes using an acceleration sensor to acquire vibration data, constructing an encoded-decoding structural network, generating new deep feature vectors through the SMOTE algorithm and the HCFF mechanism, and reconstructing the samples through the decoder to ensure the diversity and authenticity of the generated samples.
It effectively improves the ability of the fault diagnosis model to identify a few types of samples, significantly improves the quality and diversity of generated samples, especially in rare faults and complex operating conditions, and enhances the generalization performance of the model.
Smart Images

Figure CN118940124B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent mechanical fault diagnosis, and in particular to a fault data generation method of a deep SMOTE integrated with HCFF. Background Art
[0002] In recent years, with the rapid development of Industry 4.0 and intelligent manufacturing technology, machine health management (PHM) and predictive maintenance (PdM) have gradually become research hotspots in the industrial field. In this context, intelligent fault diagnosis technology has received widespread attention. Deep learning models have gradually become the core tools in fault diagnosis due to their excellent feature extraction and classification capabilities. However, the performance of deep learning models depends largely on high-quality training data sets, which places higher requirements on the diversity and balance of data.
[0003] In actual industrial applications, large-scale mechanical equipment usually runs stably for a long time in a healthy state, and the probability of failure is low and unevenly distributed. This characteristic leads to the problem of class imbalance in the training data set of the fault diagnosis model, that is, the number of healthy samples is far more than the number of fault samples, especially the number of samples of some rare faults is extremely limited. This data imbalance seriously affects the generalization ability of the fault diagnosis model, resulting in low accuracy of the model in identifying minority class faults. In addition, the method of artificially implanting fault data is costly, and the short effective operation time of the equipment under fault conditions further limits the scope and depth of data collection, which has become one of the main obstacles to the widespread application of intelligent fault diagnosis models.
[0004] In order to solve the problems of small samples and data imbalance, data enhancement technology came into being and is widely used in various machine learning tasks. Generative adversarial networks (GAN), variational autoencoders (VAE), and synthetic minority oversampling technology (SMOTE) are the mainstream methods in the current field of data enhancement. Specifically, GAN can generate samples with a distribution similar to the original data through adversarial training of the generator and the discriminator; VAE compresses the data into the latent space through the encoder, and then generates new samples through the decoder; and SMOTE generates new minority class samples in the feature space through interpolation. These methods have alleviated the small sample problem to a certain extent and improved the model's ability to recognize minority class samples.
[0005] However, these traditional data augmentation methods still have many shortcomings when applied to intelligent fault diagnosis. For example, GAN is prone to mode collapse during training, resulting in a lack of diversity in the generated samples. When generating new samples, VAE and GAN usually inject Gaussian noise into the latent space, which may reduce the authenticity of the generated samples. In addition, although the SMOTE method can generate continuous samples in the feature space when the fault features are significant, the samples it generates still lack sufficient diversity and it is difficult to fully cover the fault sample space.
[0006] In fault diagnosis tasks, the generative model needs to generate not only samples with high authenticity, but also diversified fault samples. In particular, in the case of rare and complex faults, the generated samples should have obvious fault characteristics to improve the fault diagnosis model's ability to identify minority faults. Therefore, a new fault sample generation method is urgently needed that can ensure the authenticity of the samples while enhancing the diversity of the samples, thereby effectively improving the performance and robustness of the intelligent fault diagnosis model. Summary of the invention
[0007] In order to solve the problem of insufficient sample authenticity and diversity in the prior art, this paper proposes a deep synthetic minority oversampling technique (SMOTE) data generation method based on hierarchical contextual feature fusion (HCFF). This method effectively generates samples with richer fault features while ensuring sample authenticity through multi-level feature extraction and fusion mechanisms, thereby enhancing the recognition ability and robustness of the fault diagnosis model for minority class samples.
[0008] The technical solution of the present invention mainly comprises the following steps:
[0009] A fault data generation method of deep SMOTE integrated with HCFF, the steps are as follows:
[0010] Step 1: Use the accelerometer installed on the equipment to collect vibration data. The data sampling rate is determined according to the equipment type and fault characteristics, usually several thousand to tens of thousands of hertz, to capture subtle vibration changes in the operation of mechanical equipment. The collected continuous vibration signal is segmented according to a fixed time window to obtain multiple data segments. Each segment is used as a sample for subsequent feature extraction and model training. Based on actual conditions and expert experience, each sample is labeled with the corresponding fault type label, such as normal, gear wear, bearing failure, etc., to construct an unbalanced data set with clear category labels.
[0011] Step 2: Build an encoding-decoding structure network integrating HCFF. Divide the imbalanced data set into multiple subsets according to the fault type label, and each subset contains data of one fault type. Train the encoding-decoding structure network model integrating HCFF for each subset until the model converges to ensure that each type of fault sample can be effectively represented by the model. After training, save the encoding-decoding structure network model corresponding to each type of fault sample.
[0012] Step 3: In the sample generation process, first call the trained encoder-decoder network model integrating HCFF, input the minority class sample, and extract the shallow features and deep features of the sample through the encoder. Based on the deep features of the minority class samples, select a deep feature from the deep features of the minority class samples. Secondly, randomly select N samples from the K nearest neighbors of the deep feature according to the sampling rate. Finally, randomly synthesize new deep feature vectors between these N samples and the deep feature in turn.
[0013] Step 4: In the decoding stage, the new deep feature vector is enlarged by upsampling operation to match the input vector dimension of HCFF. The generated deep features are fused with the original shallow features through the HCFF mechanism. The fused feature map is input into the data reconstruction layer, and the feature map is adjusted through multiple convolution operations and fully connected layers, and finally restored to the size of the original data. The final generated sample contains both the shallow local information of the original sample and the deep global information generated and fused by SMOTE. These samples can be used to expand the training data set and significantly improve the fault diagnosis model's ability to identify minority faults.
[0014] In one implementation, the encoding-decoding structure network integrating HCFF includes an encoder part and a decoder part. The encoder part is composed of multiple feature extraction layers, and features of different scales are extracted through multiple feature extraction layers. Each feature extraction layer is grouped into two convolution layers. The convolution layer first extracts features through a one-dimensional convolution operation, and then a DropOut operation is performed to prevent overfitting, and nonlinear features are introduced through the LeakyReLU activation function. Each feature extraction layer is connected through a downsampling layer, and downsampling is performed using a convolution operation to reduce the size of the feature map but keep the number of channels unchanged, thereby extracting more abstract global features. The decoder part is composed of multiple data reconstruction layers, wherein the first data reconstruction layer is responsible for adjusting the shape of the input data, and the second and third data reconstruction layers have similar structures. First, the deep and shallow feature maps are fused through HCFF, and then two consecutive one-dimensional convolutions, DropOut and LeakyReLU are used to realize feature extraction. The feature map size is enlarged by interpolation between each data reconstruction layer, and the number of channels is adjusted by convolution to achieve size matching of the deep and shallow feature maps. Finally, the output data shape is adjusted through convolution-fully connected operations to generate the final high-quality samples.
[0015] In one implementation, the HCFF mechanism includes: adding shallow and deep features of the sample to form a preliminary feature fusion unit. Through the channel attention mechanism, the importance weight of each channel is calculated to highlight the feature channels that are critical to fault diagnosis. Then, the spatial attention mechanism is applied to calculate the importance of each position in the feature map to identify the key areas of the fault signal. The calculation of channel attention and spatial attention is implemented through parallel operations to generate a channel weight matrix and a spatial weight matrix respectively. The two matrices are broadcast and added to obtain a comprehensive channel-space importance weight matrix. Subsequently, the information interaction between channel attention and spatial attention is realized through the intermediate fusion attention mechanism. The mechanism further adjusts the importance distribution of channels and spaces through convolution operations, optimizes the feature fusion effect, and ensures that the generated features are more representative. The intermediate fusion attention mechanism refines and adjusts the channel-space importance weight matrix, enhances the synergy between different attention mechanisms, and ensures that the fused feature map has higher recognition ability. The optimized channel-space importance weight matrix is multiplied with the initial feature fusion unit to generate the final HCFF fused feature map. This feature map combines shallow and deep feature information, and has higher expressive power after channel and spatial dual attention optimization.
[0016] Through the above steps, the present invention effectively solves the requirements of the generative model for the authenticity and diversity of generated samples in fault diagnosis tasks, and improves the quality of generated samples. In addition, the samples generated by the method described in the present invention can enhance the generalization performance of the fault diagnosis model, especially in rare faults and complex working conditions.
[0017] Compared with the prior art, the present invention has the following beneficial effects:
[0018] Improved diversity: The HCFF mechanism effectively integrates shallow and deep features. The generated samples not only have significantly enhanced diversity, but also reflect the characteristic differences of different fault types.
[0019] Accuracy guarantee: Combining the SMOTE algorithm with the HCFF mechanism ensures that the generated samples have clear category boundaries in the deep feature space, thus avoiding the sample confusion problem in traditional methods.
[0020] Model performance enhancement: The generated high-quality samples not only expand the training dataset, but also improve the recognition ability of the fault diagnosis model on minority class samples by optimizing the feature distribution. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is the overall flow chart of the present invention;
[0022] Figure 2 It is the overall structure diagram of the encoding-decoding network integrated with HCFF described in the present invention;
[0023] Figure 3 It is a detailed diagram of the implementation of the encoding-decoding network feature extraction layer of the integrated HCFF described in the present invention;
[0024] Figure 4 It is a detailed diagram of the implementation of the downsampling operation of the encoding-decoding network integrated with HCFF according to the present invention;
[0025] Figure 5 It is a detailed diagram of the implementation of the sampling operation on the encoding-decoding network integrated with HCFF according to the present invention;
[0026] Figure 6 It is a detailed diagram of the implementation of the encoding-decoding network data reconstruction layer of the integrated HCFF of the present invention;
[0027] Figure 7 is the HCFF feature fusion flow chart of the present invention;
[0028] Figure 8 is a comparison diagram of real sample frequency domain signals in an embodiment of the present invention;
[0029] Fig. 9In an embodiment of the present invention, a sample frequency domain signal comparison diagram is generated. DETAILED DESCRIPTION
[0030] The invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0031] like Figure 1 As shown, the present invention provides a method for generating data of a deep synthetic minority oversampling technique (SMOTE) based on hierarchical contextual feature fusion (HCFF), specifically comprising:
[0032] Step 1: Collect real fault data of the aircraft engine accessory transmission gearbox, make samples of the real fault data, and establish a fault data set.
[0033] Step 1.1 First, collect the real operating data of the aircraft engine accessory transmission gearbox. The fault data is collected and recorded when the aircraft engine accessory transmission gearbox is running by artificially implanting faults. To ensure the complete capture of high-frequency fault characteristics, set the sampling frequency to 10240 kHz. This step aims to obtain high-quality fault data covering a variety of operating conditions as the basis for subsequent data processing and model training.
[0034] Step 1.2 After obtaining the real fault data, this embodiment divides the real fault data into samples and creates a fault imbalance data set based on the obtained samples. In this embodiment, the sample length is 1024 data points, the data of healthy conditions in the training set contains 480 samples, each type of fault data contains 3, 4, and 5 samples, and the data of healthy and fault conditions in the test set contains 120 samples each. The samples in the training set are obtained by random sampling.
[0035] Step 1.3 Apply Fast Fourier Transform (FFT) to each sample to convert the time domain vibration data into frequency domain data. Since the FFT result is symmetrical, only the positive frequency data is retained. At this time, the length of each sample is 512 data points for subsequent model training.
[0036]
[0037] in , represents the sampled time domain vibration data, Indicates the frequency domain data obtained by fast Fourier transform. In this example, 9 experimental groups are constructed to comprehensively characterize the practicality of the method from the perspective of imbalance degree and signal-to-noise ratio. The data sets used in each experimental group are shown in Table 1.
[0038] Table 1 Imbalanced datasets used in the examples
[0039]
[0040] Step 2: Use the unbalanced data set to train the encoding-decoding network structure of the integrated HCFF, and train a dedicated encoding-decoding network structure of the integrated HCFF for each type of fault sample. Figure 2 As shown, the encoder includes a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a feature output layer; the decoder includes a first data reconstruction layer, a second data reconstruction layer, a third data reconstruction layer, and a reconstructed data output layer.
[0041] Step 2.1 Input the training sample into the encoder, and then Figure 3 The Conv feature extraction layer shown can be obtained ,right Proceed as Figure 4 The Downsample operation is shown, and then feature extraction is performed to obtain ,right Repeat the above steps to get This process uses multiple convolution and downsampling operations to enable the feature layer to gradually extract multi-scale context information from the input data, thereby effectively capturing and expressing fault features.
[0042]
[0043] in , , .
[0044] Step 2.2 As decoder input, to decode, first use Figure 5 The interpolation operation shown makes and Size matching, then Figure 6 As shown, and The first-level reconstruction features are obtained by feature fusion as the input of HCFF, and the above operation is repeated to obtain the second-level reconstruction features.
[0045]
[0046] in, , , , .
[0047] Step 2.3 The decoder adjusts the output data shape through convolution-full connection operations to generate reconstructed data .
[0048]
[0049] in .
[0050] Step 2.4 Calculation and The MSE loss is calculated and the gradient is updated using the Adam algorithm until the network converges. The encoding-decoding network parameters of the integrated HCFF for each type of fault are saved.
[0051] like Figure 7 As shown in the figure, the HCFF includes: channel attention and spatial attention mechanisms, which add the shallow and deep features of the sample to form a preliminary feature fusion unit. Through the channel attention mechanism, the importance weight of each channel is calculated to highlight the feature channels that are critical to fault diagnosis. Then, the spatial attention mechanism is applied to calculate the importance of each position in the feature map and identify the key areas of the fault signal. The calculation of channel attention and spatial attention is completed through parallel operations, and the channel weight matrix and the spatial weight matrix are generated respectively. The two matrices are broadcast and added to obtain a comprehensive channel-space importance weight matrix. The intermediate fusion attention (IFA) mechanism realizes the information interaction between channel attention and spatial attention through the intermediate fusion attention mechanism. This mechanism further adjusts the importance distribution of channels and spaces through convolution operations, optimizes the feature fusion effect, and ensures that the generated features are more representative. The intermediate fusion attention mechanism refines the channel-space importance weight matrix, enhances the synergy between different attention mechanisms, and ensures that the fused feature map has higher recognition ability. The optimized channel-space importance weight matrix is multiplied with the initial feature fusion unit to generate the final HCFF fusion feature map. This feature map combines shallow and deep feature information, and has higher expressive power after channel and spatial dual attention optimization. The feature map after HCFF fusion is further subjected to two layers of convolution and activation operations to further extract high-dimensional features and ensure that the feature distribution of the generated samples is consistent with the actual fault situation.
[0052] Step 3: Call the integrated HCFF encoding-decoding network model of each type of fault sample trained in step 2, input the sample to be processed into the encoder part, and extract the deep features of the sample ( ). The SMOTE algorithm is used to oversample minority class samples in the deep feature space to generate new deep fault features. Specifically, the deep feature representation of the new sample is generated by randomly interpolating the feature vectors between adjacent samples.
[0053]
[0054] Step 4: The generated deep fault features are used as the input of the decoder to reconstruct new sample data. The decoding process combines the newly generated deep fault features with the shallow features of the real data to ensure that the synthetic data contains the high-dimensional features of the fault samples and the key information of the original data. Finally, a mixed sample including the newly sampled deep features and the shallow features of the real samples is generated.
[0055] The synthetic samples are mixed with the real samples in the original dataset to form an enhanced training dataset. Then we use a simple one-dimensional convolutional fault classification network to train with our enhanced dataset, and use an unbalanced training set and a training set enhanced by four data augmentation methods for comparison. The results of the nine groups of experiments are shown in Table 2. The spectrum of the real sample is compared with the spectrum of the synthetic sample. Figure 8 , Fig. 9 shown.
[0056] Table 2 Comparison of experimental results of each experimental group in the example
[0057]
[0058] HCFF-SMOTE generally shows advantages in terms of average fault classification accuracy, which is 2%-9% higher than the unbalanced training set, indicating that the diversity of fault data generated using HCFF-SMOTE can effectively expand the decision domain of fault data. Secondly, the variance of HCFF-SMOTE's fault classification accuracy is significantly smaller than the variance of the comparative experiment, indicating that the fault data generated by HCFF-SMOTE is more stable. Thirdly, the confidence lower limit of HCFF-SMOTE's fault classification accuracy is higher than that of the unbalanced training set, indicating that the fault data generated by HCFF-SMOTE is more realistic, especially under the condition that the unbalanced training sets in data sets H and I achieve good fault classification accuracy, and the samples generated by the HCFF-SMOTE method do not have a negative impact on the fault diagnosis model.
[0059] In summary, the method proposed in the present invention has achieved remarkable results in the one-dimensional vibration signal analysis of the scaled test bench for the transmission case of aircraft engine accessories, effectively solved the fault diagnosis challenges under various sample imbalance ratio conditions, and fully verified the practicality and feasibility of the method.
[0060] Although the present invention has been described in detail with the accompanying drawings, it should be clear that the application scope of the present invention is not limited thereto. A technician with basic knowledge in this field can make various flexible innovations and adjustments to adapt to different practical needs while keeping the core concept of the present invention unchanged.
Claims
1. A fault data generation method of deep SMOTE integrated with HCFF, characterized in that: The following steps are involved: Step 1: Based on the acceleration data collected by the sensor, the data is cut into fixed lengths as samples, and the fault type is marked as a label for each sample to construct an unbalanced data set; Step 2: Build an encoding-decoding structure network integrating HCFF, and train the fault samples of the imbalanced data set according to the fault labels. Train an independent encoding-decoding structure network integrating HCFF for each type of fault sample, and save the models after the training converges. The encoding-decoding structure network integrating HCFF includes an encoder and a decoder, the encoder includes a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a feature output layer; the decoder includes a first data reconstruction layer, a second data reconstruction layer, a third data reconstruction layer, and a reconstructed data output layer; the feature extraction layer includes two consecutive convolution-DropOut operations, and includes a downsampling operation between adjacent levels of feature extraction layers; the data reconstruction layer includes two consecutive convolution-DropOut operations, and includes an upsampling operation and an HCFF operation between adjacent levels of data reconstruction layers; Step 3: Call the trained encoding-decoding structure network model of each type, take each type of sample as input, use the encoding part to retain the shallow features extracted from the first two levels of features, and use synthetic minority oversampling to generate deep features for the third level deep features of this type of fault samples; Step 4: Call the trained encoding-decoding structure network model of each type, and use the decoding part to reconstruct the data of the newly sampled deep features. During the reconstruction process, HCFF is used to fuse the shallow features of the real samples with the deep features obtained by oversampling, and finally a new sample is obtained.
2. The method for generating fault data of deep SMOTE integrated with HCFF as claimed in claim 1, characterized in that: The HCFF operation specifically includes the following steps: 1.1 Add deep features to shallow features to form a preliminary feature fusion unit; 1.2 The channel weight matrix and the spatial weight matrix are obtained through the parallel operation of the channel attention mechanism and the spatial attention mechanism; 1.3 Add these two weight matrices by broadcasting mechanism to obtain a comprehensive channel-space importance weight matrix; 1.4 Use the intermediate fusion attention mechanism to refine the channel-space importance weight matrix, and then multiply the optimized channel-space importance weight matrix with the initial feature fusion unit to obtain the fused feature.
3. The method for generating fault data of a deep SMOTE integrated with HCFF according to claim 1 or 2, characterized in that: In the second step, the unbalanced data set is used to train the encoding-decoding network by category, the MSE loss is used to calculate the loss of reconstructed data and real data, and the Adam algorithm gradient is used to update the network parameters.
4. The method for generating fault data of a deep SMOTE integrated with HCFF according to claim 1 or 2, characterized in that: In the third step, the specific operations are as follows: In the sample generation process, the trained HCFF-integrated encoder-decoder network model is first called, and the minority class sample is input. The shallow and deep features of the sample are extracted through the encoder. Based on the deep features of the minority class samples, a deep feature is selected from the deep features of the minority class samples; secondly, N samples are randomly selected from the K nearest neighbors of the deep feature according to the sampling ratio; finally, a new deep feature vector is randomly synthesized between these N samples and the deep feature in turn.
5. The method for generating fault data of deep SMOTE integrated with HCFF as claimed in claim 1 or 2, characterized in that: In the fourth step, the specific operations are as follows: The new deep feature vector is enlarged through upsampling operation to match the input vector dimension of HCFF; the generated deep features are fused with the original shallow features through the HCFF mechanism; the fused feature map is input into the data reconstruction layer, and the feature map is adjusted through multiple convolution operations and fully connected layers, and finally restored to the size of the original data; the final generated sample contains both the shallow local information of the original sample and the deep global information generated and fused by SMOTE.
