A data enhancement method based on prior knowledge of parkinson disease

Abstract

The application provides a data enhancement method based on prior knowledge of Parkinson's disease, comprising: preprocessing original data to obtain preprocessed data; based on Parkinson's medical prior knowledge, decomposing signals in the preprocessed data into different frequency bands; based on a U-net model as a basic diffusion model, obtaining enhanced data; extracting multi-scale features in the enhanced data to obtain feature data; based on the feature data, learning the previous feature representation through a classifier to obtain the final classification result. The application integrates the medical prior knowledge that Parkinson's disease has obvious tremor characteristics on (3-7) Hz, so that the generated specific disease samples have the core discriminant features of the original samples, and the level of the downstream task is improved.

Description

Technical Field

[0001] This application belongs to the field of data augmentation technology, specifically relating to a data augmentation method based on incorporating prior knowledge of Parkinson's disease. Background Technology

[0002] Parkinson's disease (PD), a common neurodegenerative disease, severely impacts patients' quality of life as it progresses. In recent years, with the development of artificial intelligence (AI) technology, it has become possible to assist doctors in diagnosing Parkinson's using techniques such as deep learning. CNNs' powerful feature extraction capabilities capture subtle local features of patients, while LSTMs demonstrate excellent sensitivity to time-series related data signals through their superior temporal modeling capabilities. However, the limited number of Parkinson's disease samples and the prevalence of symptoms similar to other diseases pose significant obstacles to training these models.

[0003] Data augmentation techniques, as an effective way to alleviate sample scarcity, have shown great promise in the classification of Parkinson's disease. Traditional data augmentation uses methods such as random flipping and pruning to modify the original data to obtain new generated data, providing the model with more supervisory signals and improving its performance on downstream tasks. To obtain generated signals with better data quality, data augmentation using deep learning models has become a new trend. GANs and their variants use an adversarial strategy to gradually generate augmented data that is extremely similar to the original signal. Diffusion models learn the process of adding noise to the original signal and then denoising it, gradually denoising a noisy input to obtain samples that conform to the data distribution of the original signal.

[0004] However, these deep learning-based data augmentation techniques currently face several shortcomings. In GAN-like methods, if either the generator or discriminator becomes too powerful, the other cannot be trained effectively and eventually collapses. While diffusion models are stable during training, the generated data may lose some important original features.

[0005] To further improve the effectiveness of data augmentation on the Parkinson's dataset, it is urgent to construct an augmentation method that is both stable and retains the discriminative features of the original data while generating new data, so as to improve the clinical diagnosis of Parkinson's disease. Summary of the Invention

[0006] This application provides a data augmentation method based on incorporating prior knowledge of Parkinson's disease to solve the aforementioned technical problems.

[0007] To address the aforementioned technical problems, this application adopts the following technical solution: a data augmentation method based on prior knowledge of Parkinson's disease, comprising:

[0008] S1. Preprocess the raw data to obtain preprocessed data;

[0009] S2. Based on prior knowledge of Parkinson's disease, the signals in the preprocessed data are decomposed into different frequency bands to facilitate further analysis of each frequency band.

[0010] S3. Based on the U-net model as the basic diffusion model, random Gaussian noise is added to the preprocessed data using the DDPM noise addition formula, allowing the diffusion model to learn the denoising process, thereby generating new augmented samples that are highly similar to the original data distribution. At the same time, in order to enhance the model's perception of the current time step and the distribution of augmented sample data, augmented data is obtained through temporal embedding and class embedding.

[0011] S4. Extract multi-scale features from the augmented data, combine wavelet position coding to capture the position-aware representation fused with the input features, and use the attention mechanism to assign attention weights to the features to obtain feature data;

[0012] S5. Based on the feature data, the classifier learns the previous feature representations to obtain the final classification result.

[0013] Furthermore, the method in step S1 includes:

[0014] S11. During each recording, the data acquisition unit will send a vibration notification, thus deleting the first 0.5 seconds of data for each time series of each sample to obtain the raw data;

[0015] S12. In specific data measurement, the gravity component contained in the measured acceleration data will interfere with the final result and affect the final classification effect. The L1 trend filter method is used to filter the gravity component in the original signal to obtain preprocessed data.

[0016] Furthermore, the method in step S2 includes:

[0017] S21. Based on prior knowledge of Parkinson's disease, the original signal in each preprocessed data is decomposed into three frequency bands: (0.5-3) Hz, (3-7) Hz, and (7-12) Hz, as shown in formula (1):

[0018] For frequency range ,

[0019] (1);

[0020] in, This represents the frequency band mask, where n represents the frequency point index;

[0021] S23. Based on formula (2), obtain the actual frequency corresponding to the frequency point. ; where formula (2) is:

[0022] (2);

[0023] in, Indicates the actual signal sampling rate. Indicates the total length of the signal;

[0024] S23. Based on formula (3), obtain the final transformation back to the time domain; where formula (3) is:

[0025] (3);

[0026] in, This represents the original signal.

[0027] Furthermore, the method in step S3 includes:

[0028] S31. Based on the power spectral density (PSD), calculate the spectral difference between the generated sample and the original sample in a specific frequency band, and obtain the first loss function;

[0029] S32. Based on the spectral differences, calculate the difference between the generated signal and the original signal, and obtain the second loss function in the synthesis process;

[0030] S33. Based on the DDPM noise addition formula, noise is added to the motion data, and then the noise at any time step is removed to obtain the third loss function;

[0031] S34. Based on the first loss function, the second loss function, and the third loss function, obtain the total loss function;

[0032] S35. Based on the U-net model, the original class labels are embedded according to the class of the current augmented sample through the total loss function and the class embedding mechanism to obtain augmented data.

[0033] Furthermore, the method in step S31 includes:

[0034] Based on formula (4), obtain the first loss function. ; where formula (4) is:

[0035] (4);

[0036] in, Indicates the size of the current batch. and These refer to the PSDs of the generated sample and the original sample, respectively.

[0037] Furthermore, the method in step S32 includes:

[0038] Based on formula (5), obtain the second loss function. ; where formula (3) is:

[0039] (5);

[0040] in, This represents the reconstructed signal. This represents the original signal.

[0041] Furthermore, the method in step S33 includes:

[0042] Based on formula (6), obtain the third loss function. ; where formula (6) is:

[0043] (6);

[0044] in, As input to the diffusion model, This represents the noise in the model's predictions. Indicates the current denoising time step. Indicates the current sample category.

[0045] Furthermore, the method in step S4 includes:

[0046] S41. Based on augmented data, use InceptionTime as a feature extractor to extract multi-scale features from the data;

[0047] S42. Based on multi-scale features, the Mexican hat wavelet is used as the mother wavelet, and scale and shift are added to each wavelet as learnable parameters to generate a position-aware vector fused with the input features.

[0048] S43. Based on the channel-time dual attention mechanism and position-aware vector, differentiated attention weights are assigned to the original features, channel attention is applied to the channels, and feature data is obtained.

[0049] The beneficial effects of this application are: it incorporates prior medical knowledge about the significant tremor characteristics of Parkinson's disease at (3-7) Hz, allowing the generated disease-specific samples to possess the core discriminative features of the original samples, thus improving the performance of downstream tasks. This application uses reconstruction loss to ensure that the generated samples are generally similar to the original data, avoiding distortions caused by decomposition and synthesis, as well as temporal inconsistencies. This effectively alleviates the scarcity of clinical Parkinson's samples and improves the accuracy of downstream disease diagnosis tasks. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating an embodiment of the data augmentation method based on prior knowledge of Parkinson's disease according to this application;

[0051] Figure 2 yes Figure 1 A flowchart illustrating step S3 in one embodiment;

[0052] Figure 3 yes Figure 1 A flowchart illustrating step S4 in one embodiment;

[0053] Figure 4 This is a flowchart of an embodiment of the data augmentation method based on prior knowledge of Parkinson's disease in this application;

[0054] Figure 5 This is a result confusion matrix for downstream classification in an embodiment of the data augmentation method based on prior knowledge of Parkinson's disease in this application;

[0055] Figure 6 This is a result confusion matrix for downstream classification in an embodiment of the data augmentation method based on prior knowledge of Parkinson's disease in this application;

[0056] Figure 7 This is a waveform diagram of three sub-bands obtained by applying Fast Fourier Transform to Parkinson's samples in one embodiment of the data augmentation method based on prior knowledge of Parkinson's disease in this application.

[0057] Figure 8 This is a PSD comparison diagram of the data augmentation sample obtained by data augmentation and the original sample in one embodiment of the data augmentation method based on prior knowledge of Parkinson's disease in this application;

[0058] Figure 9 This is a heatmap showing the differences between the enhanced sample and the original sample at different time points in different channels, in one embodiment of the data augmentation method based on prior knowledge of Parkinson's disease in this application. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0060] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0061] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the data augmentation method based on prior knowledge of Parkinson's disease according to this application. The method includes:

[0062] S1. Preprocess the raw data to obtain preprocessed data.

[0063] Specifically, the method in step S1 includes:

[0064] S11. Manually remove the first 0.5 seconds of data from each time series for each sample. In actual data measurement, the gravitational component contained in the measured acceleration data will interfere with the final result and affect the final classification effect; obtain the original data.

[0065] S12. The L1 trend filter method is used to filter the gravity component in the original signal. Since the original data is divided into 11 different tasks, each task is on the x, y, z axes of the two arms, with a total of 132 data channels and 976 data sampling points in each channel, the data is reshaped to a scale of 132×976 to obtain preprocessed data.

[0066] S2. Based on prior knowledge of Parkinson's disease, the signals in the preprocessed data are decomposed into different frequency bands to facilitate further analysis of each frequency band.

[0067] Specifically, firstly, a Fast Fourier Transform (FFT) is used to convert the original time-domain signal to the frequency domain. Then, frequency domain signals specific to each of the three frequency bands (0.5-3) Hz, (3-7) Hz, and (7-12) Hz are filtered out using frequency band masks. Finally, an Inverse Fast Fourier Transform (IFFT) is used to convert the filtered frequency domain signals back to the time domain, thus achieving the filtering of signals in different frequency bands. This is illustrated in the following diagram:

[0068] For frequency range ,

[0069] (1);

[0070] in, This represents the frequency band mask, where n represents the frequency point index. This represents the actual frequency corresponding to that point, and its specific calculation is as follows:

[0071] (2);

[0072] in, Indicates the actual signal sampling rate. This represents the total signal length. It is then multiplied by the transformed frequency domain signal to obtain the filtered frequency domain signal representation, which is finally transformed back to the time domain.

[0073] (3);

[0074] in, This represents the original signal.

[0075] Since only Parkinson's disease has prior medical knowledge, the above frequency band decomposition is only applicable to Parkinson's disease samples. The final decomposition diagram is shown below. Figure 6 As shown.

[0076] S3. Based on the U-net model as the basic diffusion model, random Gaussian noise is added to the preprocessed data using the DDPM noise addition formula, allowing the diffusion model to learn the denoising process and generate new augmented samples that are highly similar to the original data distribution. At the same time, in order to enhance the model's perception of the current time step and the distribution of augmented sample data, augmented data is obtained through temporal embedding and class embedding.

[0077] Specifically, the method in step S3 includes:

[0078] S31. Based on the power spectral density (PSD), calculate the spectral difference between the generated sample and the original sample in a specific frequency band, and obtain the first loss function. .

[0079] First, the PSD in the frequency band is calculated using Python's built-in Welch method. By constraining this loss, it is gradually reduced so that the generated signal retains the jitter characteristics. For each batch in the training, the specific representation is as follows:

[0080] (4);

[0081] in, Indicates the size of the current batch. and These refer to the PSDs of the generated sample and the original sample, respectively.

[0082] S32. Calculate the difference between the generated signal and the original signal. By constraining this loss to gradually reduce it, the generated signal is made as similar as possible to the original signal as possible overall, preventing signal distortion during decomposition and synthesis. Obtain the second loss function, which is specifically represented as follows for the i-th sample:

[0083] (5);

[0084] in, This represents the reconstructed signal. This represents the original signal.

[0085] S33. Using the DDPM noise addition formula, noise is added to the motion data to obtain the noise representation at any time step. Specifically, it is expressed as follows:

[0086] ;

[0087] in, Indicates standard Gaussian noise. It is the percentage of the original signal retained at the current time step. Represents all values ​​on [0,t]. The product of Indicates the signal attenuation coefficient. Indicates the noise enhancement factor. This represents the original data.

[0088] Then As input to the diffusion model, the model predicts the added noise through progressive denoising. Calculate the third loss function Specifically, it is expressed as follows:

[0089] (6);

[0090] in, As input to the diffusion model, This represents the noise in the model's predictions. Indicates the current denoising time step. Indicates the current sample category.

[0091] S34. By linearly integrating the three loss functions, further constraints are provided to the model. At the same time, by introducing appropriate hyperparameters, it is prevented that external constraints overly restrict the model and affect the main optimization objective. This is specifically expressed as follows:

[0092] ;

[0093] in, and It's a hyperparameter.

[0094] S35. Based on the U-net model, the original class labels are embedded according to the class of the current augmented sample through the total loss function and the class embedding mechanism to obtain augmented data.

[0095] Specifically, to enable the model to have a clearer understanding of the current denoising stage and adaptively adjust the denoising intensity of the current time step, this paper adopts a continuous-time embedding strategy, mapping the original time step into a continuous vector representation. First, for the current time step... Normalize:

[0096] ;

[0097] in, =100 represents the total number of diffusion steps. The obtained normalized time information is encoded using a feedforward neural network (FNN) to obtain the time step embedding representation:

[0098] ;

[0099] Temporal embedding provides crucial temporal guidance signals for the model, ensuring both the model's ability to perceive the denoising process and the smoothness and continuity of the denoising process. Data augmentation based on diffusion models often learns the data distribution of each category and the correlation features between categories through a single model, making the generated samples more closely match the overall task distribution. To better guide the generation of signals for specific categories, this paper also introduces a category embedding mechanism. The samples in this paper are divided into Parkinson's patients (y=0), healthy controls (y=1), and others (y=2), and a category embedding matrix is ​​defined as follows:

[0100] ;

[0101] in, =3 indicates the number of categories. =128 indicates the hidden layer dimension. For each batch of samples... If a category label is given Then its corresponding category label is:

[0102] ;

[0103] The final embedding representation is a concatenation of the temporal embedding and the category embedding, followed by a linear projection layer (Proj).

[0104] ;

[0105] The obtained conditional information is injected into the bottleneck layer of U-Net to guide the diffusion model to finally generate signals with specific category features.

[0106] The gradient is calculated through backpropagation to update the model parameters. The finally trained model then progressively denoises a noisy input, restoring it to a data distribution highly similar to the original training samples while preserving the original tremor characteristics. This process expands the data sample, alleviates data imbalance, and enhances the learning of discriminative features. For details on training the above data augmentation model, see [link to training documentation]. Figure 4 As shown.

[0107] like Figure 8 The image shows a single channel of data from an original Parkinson's disease sample and PSD values ​​of augmented data based on that sample at different frequencies. The yellow boxes in the image show distinct tremor peaks at 3 Hz and 7 Hz. Due to the addition of frequency loss, the augmented data retains the peak characteristics of the original data very well.

[0108] like Figure 9The figure shows a heatmap of the differences between the final augmented sample and the corresponding original sample at different channels and time points. As can be seen from the figure, the proposed data augmentation method only produces large data deviations in some channels, and the overall distribution is close to that of the original data.

[0109] This application embodiment presents a Parkinson's disease sensor dataset with imbalanced data. The original data distribution for the three sample classes is as follows: 276 Parkinson's disease samples, 79 healthy controls, and 114 other disease symptoms. Through the aforementioned data augmentation techniques, the number of Parkinson's disease samples is doubled, the number of healthy controls is seven times, and the number of other disease symptoms is five times. The final data distribution is balanced, which avoids the local optimum situation where the model predicts all samples as the class with the larger number of samples. At the same time, it can enhance the learning of the discriminative features of Parkinson's disease samples and improve downstream classification tasks.

[0110] S4. Extract multi-scale features from the augmented data, combine wavelet position coding to capture the position-aware representation fused with the input features, and use the attention mechanism to assign differentiated attention weights to the features to obtain feature data.

[0111] Specifically, the method in step S4 includes:

[0112] S41. Based on augmented data, InceptionTime is used as a feature extractor to extract multi-scale features from the data.

[0113] Specifically, InceptionTime is first used as the feature extractor, its structure consisting of two stacked InceptionBlocks, to extract multi-scale features from the data. Small-scale features tell the model detailed information about the data, while large-scale features represent the overall global distribution of the data. Feature compression is first achieved through a common bottleneck layer to save subsequent computation and retain important feature patterns. Then, three convolutional blocks of different scales (10, 20, and 40) are used to obtain features under different receptive fields. Pooling is then used to obtain the global feature representation. The output features from the three scales and the global features are concatenated and then activated by ReLU to obtain the final representation. The output also mitigates the gradient vanishing problem, which can occur when training deep networks, through residual connection techniques. Specifically, it is represented as follows:

[0114] ;

[0115] ;

[0116] As the output of InceptionBlock Indicates the input sample

[0117] S42. Based on multi-scale features, the Mexican hat wavelet is used as the mother wavelet, and scale and shift are added to each wavelet as learnable parameters to generate a position-aware vector fused with the input features.

[0118] Specifically, a self-learning wavelet positional encoding module is used, employing the Mexican hat wavelet as the parent wavelet. Its shape closely resembles a hat, naturally matching the tremor pattern exhibited in Parkinson's disease over time, thus better capturing sudden tremors in the time series. Convolution is introduced as a parameterized positional encoder, applying the learnable wavelet to the time dimension to generate a position-aware representation fused with input features. Attention mechanisms themselves struggle to perceive sequential positional relationships within a time series. Here, convolutional operations compensate for the insensitivity of subsequent attention mechanisms to position. Wavelet positional encoding is applied independently to each channel dimension, maintaining channel independence. This further positions the original time series, allowing the model to better capture the significant temporal tremor effect of Parkinson's disease and enhancing long-term modeling capabilities.

[0119] S43. Based on the channel-temporal dual attention mechanism and position-aware vector, differentiated attention weights are assigned to the original features, channel attention is applied to the channels, and feature data is obtained.

[0120] Specifically, a channel-temporal dual attention mechanism is used to assign differentiated attention weights to the original features. The original data has 132 channels, and inevitably some channels with small differences between samples will become noise data for subsequent classification, interfering with the model's classification performance. Channel attention is applied to different channels. First, average-max pooling is performed on the input data to extract the overall feature representation. An MLP is used to calculate the attention scores for different channels. To facilitate training, the attention scores are normalized to the [0,1] range using a sigmoid function, resulting in channel-level attention weights. This allows the model to perceive which channel features are beneficial for the final classification, as shown below:

[0121] ;

[0122] ;

[0123] ;

[0124] in, This represents the time series features of the input. This represents the Sigmoid activation function. These are the calculated attention weights.

[0125] by As weighted channel features, different time slices within the same channel also affect the final classification. A gated attention unit assigns attention weights to different time slices, mitigating potential noise interference from within the channel. Here, the V path learns the feature importance at each time step, and the U path learns the gating signal to determine which features should be focused. Both are initialized with a linear layer. The calculated V path and U path are then multiplied element-wise, as shown below:

[0126] ;

[0127] ;

[0128] ;

[0129] in, , and All are linear layer weights;

[0130] Attention weights at the temporal slice level are obtained through normalization. Then, through feature aggregation, a weighted time-slice feature representation is obtained. Specifically, it is expressed as follows:

[0131] ;

[0132] Finally, the channel-time weighted features are integrated to represent the important global features.

[0133] S5. Based on the feature data, the classifier learns the previous feature representations to obtain the final classification result.

[0134] Specifically, a fully connected layer is used as the downstream classifier to learn the core feature representation of the input, thereby achieving the classification and diagnosis of Parkinson's disease. The final classification results are shown in Table 1.

[0135] Table 1 shows the classification results metrics of this application applied to real-world datasets.

[0136]

[0137] To more clearly demonstrate the improvement in classification performance by data augmentation methods, the final classification results without data augmentation are shown in Table 2.

[0138] Table 2 shows the classification result metrics without data augmentation.

[0139]

[0140] Figure 5-6The images show confusion matrices with and without data augmentation. Further observation reveals that the accuracy of Parkinson's disease classification increases after data augmentation is added, reducing confusion with other diseases.

[0141] In one embodiment, the comparative experimental evaluation specifically includes the following steps:

[0142] To verify the improvement of the proposed method in the diagnosis and classification of Parkinson's disease, this invention conducted an experimental evaluation based on the dataset in Example 1, classifying the disease using different data augmentation methods. The results are shown in Table 3.

[0143] Table 3 compares the classification results metrics of different data augmentation methods applied to the same dataset.

[0144]

[0145] In summary, this application provides a data augmentation method based on the incorporation of prior knowledge about Parkinson's disease.

[0146] This application incorporates prior medical knowledge about the significant tremor characteristic of Parkinson's disease at (3-7) Hz, allowing the generated disease-specific samples to possess the core discriminative features of the original samples, thus improving the performance of downstream tasks. This application uses reconstruction loss to ensure that the generated samples are generally similar to the original data, avoiding distortions caused by decomposition and synthesis, as well as temporal inconsistencies. This effectively alleviates the scarcity of clinical Parkinson's samples and improves the accuracy of downstream disease diagnosis tasks.

[0147] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A data augmentation method based on prior knowledge of Parkinson's disease, characterized in that, include: S1. Preprocess the raw data to obtain preprocessed data; S2. Based on prior knowledge of Parkinson's disease, the signals in the preprocessed data are decomposed into different frequency bands to facilitate further analysis of each frequency band. S3. Based on the U-net model as the basic diffusion model, random Gaussian noise is added to the preprocessed data using the DDPM noise addition formula, allowing the diffusion model to learn the denoising process, thereby generating new enhanced samples that are highly similar to the original data distribution. At the same time, in order to enhance the model's perception of the current time step and the distribution of enhanced sample data, enhanced data is obtained through time embedding and category embedding. S4. Extract multi-scale features from the enhanced data, combine wavelet position coding to capture the position-aware representation fused with input features, and use an attention mechanism to assign attention weights to the features to obtain feature data; S5. Based on the feature data, the classifier learns the previous feature representations to obtain the final classification result.

2. The method according to claim 1, characterized in that, The method of step S1 includes: S11. During each recording, the data acquisition unit will send a vibration notification, thus deleting the first 0.5 seconds of data for each time series of each sample to obtain the raw data; S12. In specific data measurement, the gravity component contained in the measured acceleration data will interfere with the final result and affect the final classification effect. The L1 trend filter method is used to filter the gravity component in the original signal to obtain preprocessed data.

3. The method according to claim 1, characterized in that, The method of step S2 includes: S21. Based on prior knowledge of Parkinson's disease, the original signal in each preprocessed data is decomposed into three frequency bands: (0.5-3) Hz, (3-7) Hz, and (7-12) Hz, as shown in formula (1): For frequency range , （1）； in, This represents the frequency band mask, where n represents the frequency point index; S23. Based on formula (2), obtain the actual frequency corresponding to the frequency point. ; where formula (2) is: （2）； in, Indicates the actual signal sampling rate. Indicates the total length of the signal; S23. Based on formula (3), obtain the final transformation back to the time domain; wherein, formula (3) is: （3）； in, This represents the original signal.

4. The method according to claim 1, characterized in that, The method of step S3 includes: S31. Based on the power spectral density, calculate the spectral difference between the generated sample and the original sample in a specific frequency band, and obtain the first loss function; S32. Based on the spectral differences, calculate the difference between the generated signal and the original signal, and obtain the second loss function during the synthesis process; S33. Based on the DDPM noise addition formula, noise is added to the motion data, and then the noise at any time step is removed to obtain the third loss function; S34. Based on the first loss function, the second loss function, and the third loss function, obtain the total loss function; S35. Based on the U-net model, the original category labels are embedded according to the category of the current augmented sample through the total loss function and the category embedding mechanism to obtain the augmented data.

5. The method according to claim 4, characterized in that, The method of step S31 includes: Based on formula (4), the first loss function is obtained. ; where formula (4) is: （4）； in, Indicates the size of the current batch. and These refer to the PSDs of the generated sample and the original sample, respectively.

6. The method according to claim 4, characterized in that, The method of step S32 includes: Based on formula (5), the second loss function is obtained. ; where formula (5) is: （5）； in, This represents the reconstructed signal. This represents the original signal.

7. The method according to claim 1, characterized in that, The method of step S33 includes: Based on formula (6), the third loss function is obtained. ; where formula (6) is: （6）； in, As input to the diffusion model, This represents the noise in the model's predictions. Indicates the current denoising time step. Indicates the current sample category.

8. The method according to claim 4, characterized in that, The method of step S4 includes: S41. Based on the enhanced data, use InceptionTime as a feature extractor to extract multi-scale features from the data; S42. Based on the multi-scale features, using the Mexican hat wavelet as the mother wavelet, and adding scale and shift as learnable parameters to each wavelet, a position-aware vector fused with the input features is generated. S43. Based on the channel-temporal dual attention mechanism and the location-aware vector, assign differentiated attention weights to the original features, apply channel attention to the channels, and obtain feature data.

Images