Pulse signal classification method fusing diffusion model
The ipluseDiff network improved by the labeled index method and diffusion model, combined with the ResNet50 model, solves the data scarcity and segmentation problems in pulse signal classification, generates high-quality pulse signals, and improves the accuracy and generalization ability of the classification model.
Patent Information
- Application Number
- CN202510804816.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-09-16
AI Technical Summary
Existing pulse signal classification technology is difficult to adapt to dynamic changes, resulting in signal truncation or morphological distortion. The small size of medical data sets leads to model overfitting and insufficient generalization ability. The generated model is difficult to take into account the overall waveform structure and local high-frequency details, affecting the data enhancement effect.
The labeled index method is used for period segmentation. The ipluseDiff network and ResNet50 model improved by the diffusion model are combined. By introducing morphological constraints and frequency domain feature guidance mechanisms, high-quality pulse signals are generated, and the training dataset is expanded to improve the robustness and generalization ability of the classification model.
It achieves high-quality pulse signal generation and classification, improves the accuracy and generalization ability of the model, alleviates the problem of data scarcity, and provides new ideas for intelligent analysis of pulse signals.
Smart Images
Figure CN120653948A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of artificial intelligence and signal generation technology, and in particular to a pulse signal classification method based on a fusion diffusion model. Background Art
[0002] With the increasing incidence of chronic diseases such as cardiovascular and metabolic diseases, the public's demand for daily home health monitoring is also growing. As a simple, non-invasive, and low-cost physiological signal, the pulse signal contains rich physiological and pathological information, reflecting changes in the cardiovascular system, nervous system, and even overall health status, and has extremely high application value. Due to its periodicity and easy acquisition, the pulse signal has been widely used in health assessment and early disease screening. It has become one of the most important indicators in home health monitoring systems, providing strong data support and auxiliary decision-making basis for clinical and personal health management.
[0003] Existing pulse classification technologies have the following limitations: 1: As a typical nonlinear signal with complex physiological rhythms, the accurate segmentation of single-cycle pulses is crucial for subsequent analysis. However, traditional preprocessing and segmentation algorithms (such as fixed threshold method and template matching method) are often difficult to fully adapt to the dynamic characteristics of pulse signals, which can easily lead to signal truncation or morphological distortion, and some valuable signal segments are mistakenly deleted or noise remains, affecting the overall classification performance; 2: Since clinical pulse data collection depends on professional equipment and medical scenarios, and is limited by factors such as patient privacy protection and long sample acquisition cycle, the existing pulse signal public data sets are generally small in size and unevenly distributed, making it difficult to meet the needs of deep neural network models for large-scale, high-quality training data, which can easily lead to model overfitting and insufficient generalization ability; 3: In recent years, generative models (such as Although GANs and VAEs have been initially applied in the field of pulse signal data enhancement, the pulse signals they generate often struggle to balance both the overall waveform structure and local high-frequency details. Problems such as excessive smoothing, missing details, and morphological distortion affect the generated data's ability to fit the original signal distribution, and thus the effectiveness of data enhancement. To address these issues, this paper proposes a pulse signal classification method that incorporates a diffusion model. Combined with a labeled indexing method, it can accurately extract peak and trough locations. By introducing morphological constraints and a frequency domain feature guidance mechanism, it achieves coordinated optimization of the overall morphology and local details of the pulse signal, improving the quality and diversity of the generated pulse signal. Furthermore, by enhancing a limited pulse dataset with generated signals, the robustness and generalization capabilities of the classification model are significantly improved, effectively overcoming the challenges of data scarcity and nonlinear segmentation, and providing new ideas and technical paths for intelligent pulse signal analysis. Summary of the Invention
[0004] () Technical problems solved
[0005] In view of the shortcomings of the prior art, the present invention provides a pulse signal classification method integrating a diffusion model, which solves the problems raised in the above background technology.
[0006] (2) Technical solution
[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:
[0008] A pulse signal classification method based on a fusion diffusion model comprises the following steps:
[0009] Step 1: Construct a pulse pressure signal dataset. The collected long-time pulse pressure signal is denoised by EEMD, periodically segmented by the label index method, and averaged and normalized before being divided into a training dataset and a test dataset.
[0010] Step 2: Input the training data set in step 1 into the ipluseDiff network improved based on the diffusion model. The improved ipluseDiff network includes a forward process and a backward process. The network is trained through the forward process, and a high-quality pulse signal is generated through the backward process.
[0011] Step 3: Use the ipluseDiff network trained in step 2 to generate pulse signals. Then, manually remove the variant signals in the generated signals that do not conform to the pulse signal morphology. Finally, add the selected pulse signals, which are equivalent in number to the original training data set, to the original training data set in step 1 to form the enhanced training data.
[0012] Step 4: Input the original training dataset and enhanced training dataset from steps 1 and 3 into the ResNet50 network model for training; finally, use the test dataset to detect the performance improvement brought by this method.
[0013] Furthermore, in step 1, the tag index method is to collect the long time series pulse pressure signal Add random Gaussian noise , get the noise signal :
[0014]
[0015] Then add the noise signal Perform empirical mode decomposition respectively to obtain a series of IMF component sets;
[0016]
[0017] Where, for No. The order IMF component, the remainder after decomposition;
[0018] The IMFs of each order are averaged across all experiments to obtain a stable IMF decomposition:
[0019]
[0020] Finally, the front The IMF representing the main signal components is reconstructed to obtain the denoised signal and index signal :
[0021] ;
[0022] :
[0023] The index signal is used as a guide to extract the peak and trough positions of the denoised signal, and the peak and trough positions are used to perform period segmentation to obtain a single-cycle pulse signal set;
[0024] Then, average normalization is performed to convert the single-cycle pulse signal set into a standard normalized pulse signal with a length of 200 and an amplitude between 0 and 1. After performing the above operation on all pulse signals in the training data set, these standard normalized pulse signals are randomly divided into a training data set and a test data set in a ratio of 8:2.
[0025] Furthermore, the specific content of step 2 is: inputting the training data set in step 1 into the ipluseDiff network improved based on the diffusion model, wherein the ipluseDiff network combines the self-attention mechanism with the convolutional neural network of the U-Net structure and introduces a pulse morphology-specific loss function;
[0026] The pulse morphology-specific loss function is specifically designed as follows:
[0027] set up For the model at time step The predicted value of the noise, is the actual additive noise, then the noise prediction loss is defined as as follows:
[0028]
[0029] in for:
[0030]
[0031] Logarithmic variance Variational loss function for prediction as follows:
[0032]
[0033] in The model predicts With predefined and Linear weighting of:
[0034]
[0035] Where,
[0036] Loss of pulse characteristics It further uses domain knowledge to guide the generated pulse signal to conform to the real physiological characteristics, and adopts a loss function based on pulse characteristics to guide the model to generate a signal that is more consistent with physiological characteristics:
[0037]
[0038] is the gradient loss term based on the first-order difference:
[0039]
[0040] in, Represents the gradient of a one-dimensional signal along the time axis.
[0041] To introduce time-step weighted Loss reconstruction error:
[0042]
[0043] For the time step The relevant weighting coefficients are: .
[0044] Final training loss function It consists of multiple losses and adopts a dynamic weight adjustment mechanism. The specific form is as follows:
[0045]
[0046] The ipluseDiff network includes a forward process and a backward process;
[0047] The forward process takes as input the pulse signal of the training dataset in step 1. Through the Markov perturbation mechanism, Gaussian noise is gradually added to the original input signal at each time step, causing its distribution to gradually deviate from the original form and eventually transform into pure noise. This process controls the perturbation intensity through a preset noise scheduling strategy to ensure the stability and reversibility of the generated path.
[0048] The backward process is used to gradually recover the original pulse signal from the fully perturbed signal. This process is based on the neural network model trained by the forward process, gradually predicting and removing the noise in different time steps, restoring the signals in earlier states in turn, and finally generating a reconstruction result with a structure consistent with the input pulse signal. The ipluseDiff network is trained by comparing the real noise in the forward process with the noise predicted by the model, and using this to construct the loss to modify the network parameters to achieve the training effect.
[0049] Furthermore, the specific content of step 4 is: the original training data set obtained in step 1 and the enhanced training data set obtained in step 3 are respectively input into the ResNet50 model for training, and the trained ResNet50 model trained with the enhanced data set and the ResNet50 model not trained with the enhanced data set are used to classify the test data set obtained in step 1, and the accuracy, recall rate, sensitivity and specificity of the two models on the test set are used to judge the enhancement effect.
[0050] (3) Beneficial effects
[0051] Compared with the prior art, the present invention provides a pulse signal classification method based on a fusion diffusion model, which has the following beneficial effects:
[0052] 1. This application introduces a marking index cycle segmentation method, which can more accurately mark the peak and trough positions of the pulse signal through the index signal, and obtain a high-quality single-cycle pulse signal in the subsequent cycle segmentation link.
[0053] 2. The ipluseDiff network proposed in this application integrates convolution and self-attention mechanisms, and introduces specificity loss based on pulse morphology to generate more realistic pulse signals.
[0054] 3. This application introduces the ipluseDiff network to generate signal augmented data sets, which can alleviate the problem of poor model performance caused by too small medical data sets and improve the generalization ability of the classification model.
[0055] 3. This application also provides a reference for other problems in the field of classification using pulse signals. The ideas of this application can be applied to the classification of other diseases and health problems, and has very broad application prospects and in-depth research value. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 The following are the steps of the method of the present invention;
[0057] Figure 2 This is a schematic diagram of the marking indexing method of the present invention;
[0058] Figure 3 A comparison diagram of the present invention using the marked index method and not using the marked index method;
[0059] Figure 4 Establish an ipluseDiff network structure diagram for the present invention;
[0060] Figure 5 This is a schematic diagram of the ipluseDiff diffusion model of the present invention;
[0061] Figure 6 A comparison diagram of the real pulse signal of the present invention and the pulse signal generated using ipluseDiff;
[0062] Figure 7 This is a comparison chart of the confusion matrices of the ResNet50 network without enhancement and with ipluseDiff enhancement on the test set. DETAILED DESCRIPTION
[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0064] Example
[0065] like Figure 1-7 As shown, a pulse signal classification method based on a fusion diffusion model proposed in one embodiment of the present invention includes the following steps:
[0066] Step 1: Construct a pulse pressure signal dataset: The collected long-time pulse pressure signal is subjected to EEMD denoising and period segmentation using the label index method. The specific operations of the label index method period segmentation are as follows:
[0067]
[0068] Where, is the original long-time pulse signal, is random Gaussian noise.
[0069] Then for each noisy signal Perform empirical mode decomposition separately to obtain a series of IMF component sets.
[0070]
[0071] Where, for No. IMF components, the remainder after decomposition;
[0072] The IMF of each order is averaged over all experiments to obtain Figure 3 The final stable IMF decomposition:
[0073]
[0074] Finally, the front The IMF representing the main signal components is reconstructed to obtain the denoised signal and index signal :
[0075] ;
[0076] :
[0077] The index signal is used as a guide to extract the peak and trough positions of the denoised signal, and the peak and trough positions are used to perform period segmentation to obtain a single-cycle pulse signal set;
[0078] Then, average normalization is performed to convert the single-cycle pulse signal set into a standard normalized pulse signal with a length of 200 and an amplitude between 0 and 1. After performing the above operation on all pulse signals in the training data set, these standard normalized pulse signals are randomly divided into a training data set and a test data set in a ratio of 8:2.
[0079] Step 2: Input the training data set in step 1 into Figure 4 The ipluseDiff network shown in FIG. Figure 5 The two Markov processes shown are forward diffusion process and reverse diffusion process. In the forward process, The variance table will be used stepwise Generate Gaussian noise and add it to the data , so that it degenerates into a standard normal distribution:
[0080]
[0081] in , is a standard matrix, the sampling signal ;in , , .
[0082] Specifically, the signal and the time information after cosine time embedding are input into the self-attention module. After fusing the time series information, the multi-scale semantic features are extracted through the 1D convolution and downsampling modules. , we map the time step into a high-dimensional vector through the cosine embedding function; assuming the dimension of the time embedding is , then the embedded time variable for:
[0083]
[0084] in is the time step, 10000 is the scaling factor; the time feature after embedding The normalized pulse signal will be input Add together to get the joint features , input to the subsequent attention module; in the attention module, the joint features After linear transformation, query, key and value matrices are generated:
[0085]
[0086] Then, the attention weight matrix A is obtained by calculation:
[0087]
[0088] in is the key vector The dimension is then calculated and the weighted sum is used to get the final attention output:
[0089]
[0090] The output of the self-attention mechanism is passed through 1D convolution and downsampling for 3 cycles to obtain multi-scale semantic features, which are gradually upsampled and the high-resolution features of the signal are restored. At the same time, the low-level features of the encoder are connected through jumps to fuse the low-level features with the high-level features of the same scale, retaining local detail information and outputting the predicted noise.
[0091] Then, the actual noise in the forward process is compared with the noise predicted by the model, and the loss function is constructed to modify the network parameters to achieve the training effect. The specific loss function constructed for pulse morphology is designed as follows:
[0092] set up For the model at time step The predicted value of the noise, is the actual additive noise, then the noise prediction loss is defined as as follows:
[0093]
[0094] in for:
[0095]
[0096] Logarithmic variance Variational loss function for prediction as follows:
[0097]
[0098] in The model predicts With predefined and Linear weighting of:
[0099]
[0100] Where,
[0101] Loss of pulse characteristics It further uses domain knowledge to guide the generated pulse signal to conform to the real physiological characteristics, and adopts a loss function based on pulse characteristics to guide the model to generate a signal that is more consistent with physiological characteristics:
[0102]
[0103] is the gradient loss term based on the first-order difference:
[0104]
[0105] in, Represents the gradient of a one-dimensional signal along the time axis.
[0106] To introduce time-step weighted Loss reconstruction error:
[0107]
[0108] For the time step The relevant weighting coefficients are: .
[0109] Final training loss function It consists of multiple losses and adopts a dynamic weight adjustment mechanism. The specific form is as follows:
[0110]
[0111] In the reverse process, we use the trained network to predict the noise to be removed at each time step, and gradually remove the noise to obtain the final generated pulse signal.
[0112] Step 3: Use the ipluseDiff network trained in step 2 to generate a pulse signal. Then, manually remove the variant signals in the generated signal that are obviously inconsistent with the pulse signal morphology. Finally, add the selected generated pulse signals, which are equivalent in number to the original training data set, to the original training data set in step 1 to form the enhanced training data.
[0113] Step 4: Input the original training dataset obtained in step 1 and the enhanced training dataset obtained in step 3 into the ResNet50 model for training respectively, and use the trained ResNet50 model to classify the test dataset obtained in step 1. Calculate the accuracy, recall, sensitivity, and specificity of the test dataset to evaluate the performance of the network after using this method. The calculation method of the parameters is as follows:
[0114]
[0115]
[0116]
[0117] Where, For real, For true negative, For false positive, False negative.
[0118] The above four parameters are used as reference to modify the relevant parameters of the network in this method. The final network classification performance improvement on the test set is as follows: Figure 7 shown.
[0119] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.
Claims
1. A pulse signal classification method based on a fusion diffusion model, characterized by: The steps include: Step 1: Construct a pulse pressure signal dataset. The collected long-time pulse pressure signal is denoised by ensemble empirical mode decomposition (EEMD), periodically segmented by the label index method, and averaged and normalized before being divided into a training dataset and a test dataset. Step 2: Input the training data set in step 1 into the ipluseDiff network improved based on the diffusion model. The ipluseDiff network includes a forward process and a backward process. The network is trained through the forward process, and a high-quality pulse signal is generated through the backward process. Step 3: Use the ipluseDiff network trained in step 2 to generate pulse signals. Then, manually remove the variant signals in the generated signals that do not conform to the pulse signal morphology. Finally, add the manually selected generated pulse signals, which are equivalent in number to the original training data set, to the original training data set in step 1 to form the enhanced training data. Step 4: Input the original training dataset and enhanced training dataset from steps 1 and 3 into the ResNet50 network model for training; finally, use the test dataset to detect the performance improvement brought by this method.
2. The pulse signal classification method based on the fusion diffusion model according to claim 1, characterized in that: In step 1, the tag index method is to collect the long time series pulse pressure signal Add random Gaussian noise , get the noise signal : ； Then for each noisy signal Perform empirical mode decomposition respectively to obtain a series of IMF component sets; ； Where, for No. The order IMF component, the remainder after decomposition; The IMFs of each order are averaged across all experiments to obtain a stable IMF decomposition: ； Finally, the front The IMF representing the main signal components is reconstructed to obtain the denoised signal and index signal : ； ： ； Extract the denoised signal using the index signal as a guide The peak and trough positions of the pulse signal are obtained by period segmentation using the peak and trough positions. Then, average normalization is performed to convert the single-cycle pulse signal set into a standard normalized pulse signal with a length of 200 and an amplitude between 0 and 1. After performing the above operation on all pulse signals in the training data set, these standard normalized pulse signals are randomly divided into a training data set and a test data set in a ratio of 8:
2.
3. The pulse signal classification method based on the fusion diffusion model according to claim 1, characterized in that: The specific content of step 2 is: inputting the training data set in step 1 into the ipluseDiff network improved based on the diffusion model, wherein the ipluseDiff network combines the self-attention mechanism with the convolutional neural network of the U-Net structure and introduces a pulse morphology-specific loss function; The pulse morphology-specific loss function is specifically designed as follows: set up For the model at time step The predicted value of the noise, is the actual additive noise, then the noise prediction loss is defined as as follows: ； in for: ； Logarithmic variance Variational loss function for prediction as follows: ； in The model predicts With predefined and Linear weighting of: ； Where, ; Loss of pulse characteristics It further uses domain knowledge to guide the generated pulse signal to conform to the real physiological characteristics, and adopts a loss function based on pulse characteristics to guide the model to generate a signal that is more consistent with physiological characteristics: ； is the gradient loss term based on the first-order difference: ； in, Represents the gradient of a one-dimensional signal along the time axis; To introduce time-step weighted Loss reconstruction error: ； For the time step The relevant weighting coefficients are: ; Final training loss function It consists of multiple losses and adopts a dynamic weight adjustment mechanism. The specific form is as follows: ； The ipluseDiff network includes a forward process and a backward process; The forward process takes as input the pulse signal of the training dataset in step 1. Through the Markov perturbation mechanism, Gaussian noise is gradually added to the original input signal at each time step, causing its distribution to gradually deviate from the original form and eventually transform into pure noise. This process controls the perturbation intensity through a preset noise scheduling strategy to ensure the stability and reversibility of the generated path. The backward process is used to gradually recover the original pulse signal from the fully perturbed signal. This process is based on the neural network model trained by the forward process, gradually predicting and removing the noise in different time steps, restoring the signals in earlier states in turn, and finally generating a reconstruction result with a structure consistent with the input pulse signal. The ipluseDiff network is trained by comparing the real noise in the forward process with the noise predicted by the model, and using this to construct the loss to modify the network parameters to achieve the training effect.
4. The pulse signal classification method based on the fusion diffusion model according to claim 1, characterized in that: The specific content of step 4 is: the original training data set obtained in step 1 and the enhanced training data set obtained in step 3 are respectively input into the ResNet50 model for training, and the trained ResNet50 model trained with the enhanced data set and the ResNet50 model not trained with the enhanced data set are used to classify the test data set obtained in step 1, and the accuracy, recall rate, sensitivity and specificity of the two models on the test set are used to judge the enhancement effect.