A pulse position recognition method and system fusing an interpretable module

Abstract

The present application belongs to the technical field of intelligent pulse diagnosis of traditional Chinese medicine, and specifically discloses a pulse position recognition method and system fusing an interpretable module, the method comprising: obtaining a radial artery multi-channel pulse pressure signal, preprocessing and time-frequency transforming the signal to generate a time-frequency scale diagram; inputting the time-frequency scale diagram into a pre-trained PulseMed model, the model comprising a classification recognition module and a parallelly arranged interpretability analysis module, the classification recognition module obtaining a latent representation vector representing the pulse signal according to the time-frequency scale diagram and obtaining a pulse position classification result according to the latent representation vector; and the interpretability analysis module being used to generate attribution information corresponding to the pulse position classification result according to the time-frequency scale diagram. The present application realizes high-accuracy pulse position recognition, provides interpretable attribution information, and enhances the transparency and trustworthiness of clinical decision-making.

Description

Technical Field

[0001] This invention belongs to the field of intelligent pulse diagnosis technology in traditional Chinese medicine, and specifically relates to a pulse position recognition method and system that integrates an interpretable module. Background Technology

[0002] Pulse diagnosis is one of the most distinctive diagnostic methods in Traditional Chinese Medicine (TCM). Physicians assess the strength, rhythm, and texture of the pulse by palpating three points on the radial artery: cun, guan, and chi. This assessment helps determine the functional status of the body's internal organs. Accurate differentiation of these three points is fundamental to TCM pulse diagnosis, as each point corresponds to a different organ system: cun reflects the heart and lungs, guan corresponds to the liver, gallbladder, and spleen, and chi reflects the kidneys and bladder. This spatial differentiation allows pulse diagnosis to provide a comprehensive assessment of overall health.

[0003] However, traditional pulse diagnosis relies heavily on the physician's experience and subjective judgment, lacking objective and quantifiable standards. To address this issue, researchers have developed pulse acquisition devices based on pressure sensors and combined them with machine learning methods for automatic classification. In existing technologies, feature extraction methods based on Fourier transform and wavelet transform, as well as shallow models such as support vector machines and random forests, have been applied to pulse signal analysis. In recent years, convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have also been used for pulse waveform classification, achieving some success.

[0004] However, existing methods have two main drawbacks: First, traditional deep learning models (such as CNNs) struggle to simultaneously capture both local time-frequency features and long-range global dependencies in pulse signals, especially when processing high-dimensional spatiotemporal data generated by multi-channel sensor arrays, resulting in limited classification accuracy. Second, most models employ a "black box" structure, lacking interpretability in their decision-making process and failing to provide a basis for classification decisions, making them difficult to trust and adopt in practical applications. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of limited classification accuracy and lack of interpretability in existing intelligent pulse position recognition technology, and to propose a pulse position recognition method and system that integrates interpretable modules.

[0006] The technical solution of this invention is as follows: This invention provides a pulse position recognition method integrating an interpretable module, comprising: Step 1: Acquire the multi-channel pulse pressure signal of the radial artery of the object to be identified; Step 2: Preprocess and perform time-frequency transformation on the multi-channel pulse pressure signal to generate at least one time-frequency scaling map; Step 3: Input the time-frequency scaling map into the pre-trained PulseMed model and output the corresponding pulse position classification results and the attribution information corresponding to the pulse position classification results; The PulseMed model includes a classification and identification module and an interpretability analysis module. The classification and identification module obtains a latent representation vector representing the pulse signal based on the time-frequency scaling map, and obtains the pulse position classification result based on the latent representation vector. The interpretability analysis module is used to generate attribution information corresponding to the pulse position classification result based on the time-frequency scaling map.

[0007] Preferably, step 2 includes: Step 2.1: The multi-channel pulse pressure signal is filtered using a fourth-order Butterworth bandpass filter with a filtering range of 0.7Hz-10Hz; Step 2.2: Perform continuous wavelet transform on the filtered multi-channel pulse pressure signal using complex Morlet wavelets to obtain the initial time-frequency scale map; Step 2.3: Scale the initial time-frequency scale map to a uniform size of 224×224 pixels to obtain the time-frequency scale map.

[0008] Preferably, the pulse position categories include the cun pulse position, the guan pulse position, and the chi pulse position; the attribution information includes at least one of the following: attribution mask, interpretation latent representation, time importance curve, and representativeness waveform.

[0009] Preferably, the classification and recognition module adopts a Swing Transformer-based architecture, comprising: a sequentially cascaded image block partitioning unit, a linear embedding unit, a multi-layer Transformer coding unit, a global pooling unit, and a classification output unit, wherein... The image block division unit is used to divide the time-frequency scale map into multiple non-overlapping image blocks; The linear embedding unit is used to map each image block to a token embedding vector and add position encoding to preserve the time-frequency spatial arrangement information; In the multi-layer Transformer coding unit, each layer of Transformer coding unit includes a first sub-coding unit and a second sub-coding unit connected in sequence. The first sub-coding unit captures the correlation of the time-frequency region within a local window based on window multi-head self-attention to form a first intermediate representation. The second sub-coding unit realizes information interaction between adjacent windows based on shift window multi-head self-attention, and fuses local features with global context to form the output of the Transformer coding unit of this layer. The global pooling unit is used to perform global average pooling on the feature map output by the multi-layer Transformer coding unit to obtain the latent representation vector. The classification output unit is used to output the predicted probabilities of the three pulse position categories of Cun, Guan, and Chi based on the latent representation vector, so as to obtain the pulse position classification result.

[0010] Preferably, the first sub-encoding unit includes: a first normalization layer and a window multi-head self-attention layer. The first normalization layer performs layer normalization processing on the input token sequence of the Transformer encoding unit at this layer and then inputs it to the window multi-head self-attention layer. The output of the window multi-head self-attention layer is added to the input token sequence of the Transformer encoding unit at this layer through a first residual connection to obtain the first intermediate representation. The second sub-encoding unit includes: a second normalization layer, a shift-window multi-head self-attention layer, and a multilayer perceptron. The second normalization layer performs layer normalization processing on the first intermediate representation and then inputs it to the shift-window multi-head self-attention layer. The output of the shift-window multi-head self-attention layer is added to the first intermediate representation through a second residual connection to obtain a second intermediate representation. The multilayer perceptron performs a nonlinear transformation on the second intermediate representation, and its output is added to the second intermediate representation through a third residual connection as the output of the Transformer encoding unit of this layer.

[0011] Preferably, the interpretability analysis module includes: an integral gradient unit and a self-supervised proxy interpreter; wherein, The integral gradient unit is used to calculate the gradient integral of the output of the classification and recognition module relative to each time point of the time-frequency scale map, obtain the importance score of the time dimension, and map the importance score back to the original pulse waveform to generate a time importance curve. The self-supervised proxy interpreter is used to generate an attribution mask based on the time-frequency scale map, and then use the attribution mask to perform a masking operation on the time-frequency scale map and map it to the interpretation latent space to obtain the interpretation latent representation. The interpretation latent representation is then compared with the interpretation latent representation corresponding to the prototype flag waveform pre-stored in the training phase to calculate the cosine similarity, and the most matching prototype flag waveform is determined as the representative flag waveform output.

[0012] Preferably, the self-supervised proxy interpreter includes: an interpretation generator, an interpretation encoder, an interpretation predictor, and a prototype extraction subunit; wherein, The interpretation generator is used to generate the attribution mask based on the time-frequency scale map, and the attribution mask is used to highlight the discriminative regions in the time-frequency scale map; The interpretive encoder is used to perform a masking operation between the attribution mask and the time-frequency scaling map and map it to the interpretive latent space to obtain the interpretive latent representation. At the same time, it receives the latent representation vector generated by the classification and recognition module as a supervision signal for model behavior consistency constraints during the model training phase. The interpretation predictor is used to output an auxiliary pulse position classification result based on the interpretation latent representation, and during the model training phase, the auxiliary pulse position classification result and the pulse position classification result output by the classification and recognition module are used to perform label consistency constraints. The prototype extraction subunit is used to perform cosine similarity calculation between the explanatory latent representation and the explanatory latent representation corresponding to the prototype marker waveform pre-stored in the training phase, determine the most matching prototype marker waveform as the representative marker waveform output, and cluster the explanatory latent representations of the same pulse position in the model training phase, and extract the cluster center as the prototype marker waveform.

[0013] Preferably, the training process of the PulseMed model includes: S1: Construct a PulseMed model, and train the classification and recognition module of the PulseMed model using the acquired dataset to obtain a trained classification and recognition module. The dataset includes time-frequency scaling maps of multiple different pulse diagnosis locations. The classification module adopts a first total loss function including cross-entropy loss and contrast loss during the training process. S2: After the classification and recognition module is trained, its network parameters are fixed, the dataset is input into the PulseMed model, and the latent representation vector and pulse position classification result obtained by the classification and recognition module are used as supervision signals to train the self-supervised agent interpreter to obtain the trained self-supervised agent interpreter; the self-supervised agent interpreter adopts a second total loss function including model behavior consistency loss and label consistency loss during the training process.

[0014] Preferably, the first total loss function is expressed as: ; ; ; In the formula, The first total loss, For cross-entropy loss, To compare the losses, The number of training samples, For the number of categories, For real category labels, To predict class probabilities, Indicates the first The latent representation vector obtained by the classification and recognition module for each sample; Indicates and The latent representation vector that constitutes a positive sample pair Represents cosine similarity. For temperature parameters, Indicates the first The latent representation vectors obtained by the classification and recognition module for each sample Weighting coefficients for comparison losses; The second total loss function is expressed as: ; ; ; In the formula, This is the second total loss. For model behavior consistency loss, For label consistency loss, The weighting coefficients for label consistency loss are: They represent the first The first sample and the first The latent representation vectors obtained by the classification and recognition module for each sample The outputs of the self-supervised agent interpreter are respectively the first... The first sample and the first The latent representation of each sample, Represents the cross-entropy function. For the classification and recognition module, the first The predicted probability distribution of each sample output. To explain the predictor's effect on the first The auxiliary prediction probability distribution of each sample output. This represents the distance metric between samples in the latent space of the classification and recognition module. This represents the distance metric between samples in the latent space.

[0015] This invention provides a pulse position recognition system with an integrated interpretable module, applicable to the pulse position recognition method with an integrated interpretable module described in any of the above embodiments, comprising: The signal acquisition module acquires multi-channel pulse pressure signals from the radial artery of the object to be identified; The signal preprocessing and time-frequency transformation module is used to preprocess and perform time-frequency transformation on the multi-channel pulse pressure signal to generate at least one time-frequency scale map. The data processing module is used to input the time-frequency scaling map into the pre-trained PulseMed model and output the corresponding pulse position classification results and the attribution information corresponding to the pulse position classification results; The PulseMed model includes a classification and identification module and an interpretability analysis module. The classification and identification module obtains a latent representation vector representing the pulse signal based on the time-frequency scaling map, and obtains the pulse position classification result based on the latent representation vector. The interpretability analysis module is used to generate attribution information corresponding to the pulse position classification result based on the time-frequency scaling map.

[0016] The beneficial effects of this invention are: This invention presents a pulse position recognition method with an integrated interpretable module, forming a complete end-to-end process from multi-channel pulse signal acquisition, preprocessing, time-frequency transformation, Transformer classification to interpretable output. The integrated interpretable analysis module quantifies the contribution of each time point in the input waveform to the classification result, generates attribution masks to highlight discriminative regions in the time-frequency scaling map, and extracts prototype marker waveforms from the interpretable latent space, providing typical pulse morphology templates for each pulse position. These interpretation results are highly consistent with the theories of traditional Chinese medicine's viscera and bowels, enhancing trust in the model's decisions. This invention is applicable to scenarios such as TCM-assisted diagnosis, telemedicine, and intelligent health monitoring. Attached Figure Description

[0017] Figure 1 The diagram shows a flowchart of a pulse position recognition method that integrates interpretable modules; Figure 2 The diagram shown illustrates the data processing of the classification and recognition module. Figure 3 The diagram shown is a structural block diagram of a pulse position recognition system that integrates an interpretable module. Figure 4 The diagram shows the confusion matrix and ROC curves for comparing PulseMed models on the development and validation datasets. Figure 5 The results of the interpretability analysis of the model using integral gradients are shown below. Figure 6 The diagram shows the interpretation results of a self-supervised agent interpreter learning the embedding space of the interpretation and identifying representative symbols. Detailed Implementation

[0018] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the invention, and are not intended to limit the scope of the invention.

[0019] In a first aspect, embodiments of the present invention provide a pulse position recognition method incorporating an interpretable module. Please refer to... Figure 1 , Figure 1 The diagram shows a flowchart of a pulse position recognition method that integrates an interpretable module.

[0020] like Figure 1 As shown, the pulse position recognition method of the fusion interpretable module in this embodiment includes the following steps: Step 1: Obtain the multi-channel pulse pressure signal of the radial artery of the object to be identified.

[0021] In this embodiment, the UBenHealth pulse diagnosis device (UBJQR-002A) was used for signal acquisition. This device is equipped with a 96-channel flexible pressure sensor array. The array is divided according to three different pulse diagnosis locations defined in Traditional Chinese Medicine: cun, guan, and chi, with 32 sensors assigned to each location. The device uses an automatically inflating cuff to linearly increase pressure for 60 seconds, simulating the palpation process of a TCM practitioner from light to heavy pressure. The sampling frequency is 50Hz, and the subject records 90 seconds of continuous pulse signal data. The initial 80-second signal segment (approximately 4000 sampling points) is extracted for analysis. The subject rests for at least 10 minutes before measurement, and the ambient temperature is controlled at 20-25℃.

[0022] Step 2: Preprocess and perform time-frequency transformation on the multi-channel pulse pressure signal to generate at least one time-frequency scale map.

[0023] In this embodiment, step 2 includes: Step 2.1: Use a fourth-order Butterworth bandpass filter to filter the multi-channel pulse pressure signal, with a filtering range of 0.7Hz-10Hz; Step 2.2: Perform continuous wavelet transform on the filtered multi-channel pulse pressure signal using complex Morlet wavelets to obtain the initial time-frequency scale map; Step 2.3: Scale the initial time-frequency scale map to a uniform size of 224×224 pixels to obtain the time-frequency scale map.

[0024] Specifically, to separate physiologically relevant frequency content, a fourth-order Butterworth bandpass filter with a filtering range of 0.7Hz-10Hz was first applied to the multi-channel pulse pressure signal to remove baseline drift (<0.7Hz) and high-frequency noise (>10Hz) while preserving heart rate dynamics. Then, a continuous wavelet transform was performed on the filtered signal using complex Morlet wavelets to obtain an initial time-frequency scale map. This scale map represents the energy distribution in the time and frequency dimensions as a two-dimensional image. This transform captures time and spectral features in a compact, image-like format, suitable for subsequent visual Transformer analysis in the PulseMed model. Finally, the initial time-frequency scale map was adjusted to 224×224 pixels to obtain the final time-frequency scale map, meeting the input requirements of the subsequent PulseMed model.

[0025] Understandably, in this embodiment, traditional pulse features, including key reference points (start point, systolic peak, dicrotic notch, diastolic peak) and pulse-derived features (amplitude, peak time, peak width, reflection index), are also extracted from the filtered signal for statistical characterization and comparison across pulse diagnosis locations, and serve as auxiliary verification criteria for model classification and interpretability results. It should be noted that these traditional pulse features are not used as direct input to the PulseMed model classification and recognition module, but are used in experimental analysis to verify whether the waveform differences between different pulse diagnosis locations are consistent with the model's attribution region and representative marker waveform.

[0026] Step 3: Input the time-frequency scaling map into the pre-trained PulseMed model and output the corresponding pulse position classification results and the attribution information corresponding to the pulse position classification results.

[0027] In this embodiment, the PulseMed model includes a classification and identification module and an interpretability analysis module. The classification and identification module obtains a latent representation vector representing the pulse signal based on the time-frequency scaling graph, and obtains the pulse position classification result based on the latent representation vector. The pulse position categories include cun pulse position, guan pulse position, and chi pulse position. The interpretability analysis module is used to generate attribution information corresponding to the pulse position classification result based on the time-frequency scaling graph. The attribution information includes at least one of the following: attribution mask, interpretable latent representation, time importance curve, and representative marker waveform.

[0028] Specifically, the classification and recognition module of the PulseMed model is responsible for learning the discriminative representations of the three pulse positions: cun, guan, and chi; while the interpretability analysis module further identifies the key time periods, time-frequency regions, and representative waveform patterns on which the model's decisions depend.

[0029] Furthermore, the structure, data processing, and training process of the PulseMed model in this embodiment will be explained.

[0030] In this embodiment, the classification and recognition module adopts a visual Transformer based on the Swing Transformer architecture. Specifically, the classification and recognition module includes: an image block segmentation unit, a linear embedding unit, a multi-layer Transformer coding unit, a global pooling unit, and a classification output unit, which are cascaded in sequence.

[0031] Please see Figure 2 , Figure 2 The diagram shown illustrates the data processing of the classification and recognition module. Figure 2As shown, in this embodiment, the image patch partitioning unit is used to divide the time-frequency scale map into multiple non-overlapping image patches. The linear embedding unit is used to map each image patch into a token embedding vector and add positional encoding to preserve the time-frequency spatial arrangement information. In the multi-layer Transformer encoding unit, each layer of Transformer encoding unit includes a first sub-encoding unit and a second sub-encoding unit connected in sequence. The first sub-encoding unit captures the correlation of the time-frequency region within a local window based on window multi-head self-attention, forming a first intermediate representation; the second sub-encoding unit realizes information interaction between adjacent windows based on shift window multi-head self-attention, fusing local features and global context to form the output of this layer of Transformer encoding unit. The global pooling unit is used to perform global average pooling on the feature map output by the multi-layer Transformer encoding unit to obtain a latent representation vector. The classification output unit is used to output the predicted probabilities of the three pulse position categories of Cun, Guan, and Chi based on the latent representation vector to obtain the pulse position classification result.

[0032] Specifically, the first sub-encoding unit includes: a first normalization layer and a window multi-head self-attention layer. The first normalization layer performs layer normalization processing on the input token sequence of the Transformer encoding unit at this layer before inputting it to the window multi-head self-attention layer. The output of the window multi-head self-attention layer is added to the input token sequence of the Transformer encoding unit at this layer through a first residual connection to obtain a first intermediate representation. The second sub-encoding unit includes: a second normalization layer, a shifted window multi-head self-attention layer, and a multilayer perceptron. The second normalization layer performs layer normalization processing on the first intermediate representation before inputting it to the shifted window multi-head self-attention layer. The output of the shifted window multi-head self-attention layer is added to the first intermediate representation through a second residual connection to obtain a second intermediate representation. The multilayer perceptron performs a nonlinear transformation on the second intermediate representation, and its output is added to the second intermediate representation through a third residual connection as the output of the Transformer encoding unit at this layer.

[0033] It should be noted that in the multi-layer Transformer coding unit, the input of the first normalization layer of the first layer of the Transformer coding unit is the token embedding vector output by the linear embedding unit, and the input of the first normalization layer of the remaining Transformer coding units is the token sequence output by the previous layer of the Transformer coding unit.

[0034] Specifically, the core of the classification and recognition module adopts a visual Transformer (a variant of the Swin-Tiny, pre-trained on ImageNet) based on the Swin Transformer architecture. It takes a time-frequency scale map as input and uses a hierarchical Swin Transformer structure to extract discriminative features. The internal connections of the visual Transformer architecture model can be divided into four stages: image patch partitioning, linear embedding, hierarchical Transformer encoding, and classification prediction. First, the input 224×224 scale map is divided into a series of non-overlapping image patches, each corresponding to the energy distribution of the pulse signal in the local time-frequency region. Then, the image patches are linearly projected into token embedding vectors, preserving their relative spatial arrangement in the time-frequency map. These token embedding vectors are then input into the Swin Transformer encoder. The encoder consists of multiple repetitive modules, each mainly including a window multi-head self-attention mechanism, a shifted window multi-head self-attention mechanism, layer normalization, and a multilayer perceptron. Window attention is used to capture correlations within local time-frequency regions, while shift window attention promotes information interaction between adjacent windows, thereby achieving the fusion of local features and global context.

[0035] After multi-layer Transformer encoding, a high-dimensional latent representation vector is obtained to characterize the overall time-frequency pattern of the input pulse signal. This latent representation vector serves two purposes: firstly, it is input into the classification head, which outputs the predicted probabilities of the three pulse position categories (cun, guan, and chi) after pooling and fully connected layers; secondly, it serves as the input basis for subsequent contrastive learning and interpretability analysis modules. Through this design, the PulseMed model can not only perform pulse position classification but also form a class-separable pulse representation in the latent space.

[0036] Understandably, the time-frequency scaling plot enables the PulseMed model to simultaneously utilize the temporal dynamics, spectral distribution, and cross-pulse spatial differences of the pulse signal. Compared to directly inputting a one-dimensional waveform, the time-frequency scaling plot can more completely preserve the energy changes of the pulse wave across different time windows and frequency ranges, thus providing a richer information basis for the Transformer to capture long-range dependencies and local discriminative features.

[0037] In this embodiment, the interpretability analysis module includes an integral gradient unit and a self-supervised surrogate interpreter. The integral gradient unit calculates the gradient integral of the output of the classification and recognition module relative to each time point of the time-frequency scale map, obtaining a time-dimensional importance score. This importance score is then mapped back to the original pulse waveform to generate a time-importance curve. The self-supervised surrogate interpreter generates an attribution mask based on the time-frequency scale map. After masking the time-frequency scale map using the attribution mask, it maps it to the interpretive latent space to obtain an interpretive latent representation. The cosine similarity between this interpretive latent representation and the interpretive latent representations corresponding to the prototype marker waveforms pre-stored during the training phase is calculated, and the best-matching prototype marker waveform is determined as the representative marker waveform output.

[0038] Specifically, the self-supervised surrogate interpreter includes an interpretation generator, an interpretation encoder, an interpretation predictor, and a prototype extraction subunit. The interpretation generator generates an attribution mask based on the time-frequency scale map, highlighting discriminative regions within the map. The interpretation encoder maps the attribution mask to the time-frequency scale map and then onto the interpretation latent space to obtain the interpretation latent representation. It also receives the latent representation vector generated by the classification module as a supervisory signal for model behavior consistency constraints during model training. The interpretation predictor outputs auxiliary pulse position classification results based on the interpretation latent representation. During model training, it uses these auxiliary pulse position classification results to perform label consistency constraints with the pulse position classification results output by the classification module. The prototype extraction subunit calculates the cosine similarity between the interpretation latent representation and the interpretation latent representations corresponding to prototype marker waveforms pre-stored during training. It determines the best-matching prototype marker waveform as the representative marker waveform output. During model training, it clusters the interpretation latent representations for the same pulse position and extracts the cluster centers as prototype marker waveforms.

[0039] Specifically, to improve the transparency of the PulseMed model's decision-making process, this embodiment conducts interpretability analysis on the model from two levels: local attribution and proxy explanation. The latent representation vectors obtained by the classification and identification module and the pulse position classification results together form the basis of the interpretability analysis in the PulseMed model.

[0040] At the local attribution level, the Integrated Gradients (IG) method is used to evaluate the contribution of different time points or time-frequency regions in the input signal to the model's prediction results. Specifically, the IG method uses a baseline signal lacking effective pulse information as a reference. An interpolation path is constructed between the baseline signal and the actual observed signal, and the gradient of the model output relative to the input features is accumulated along this path to calculate the attribution score for each input feature. This score reflects the importance of different segments in the input waveform to the model's decision. The resulting attribution scores can be further converted into attribution masks to highlight key time-frequency regions that contribute significantly to classification decisions, while suppressing regions with low information content or weak classification relationships. The resulting temporal importance score reflects which segments of the input waveform have the greatest influence on the model's decision-making process.

[0041] For example, let F : Denotes the differentiable function of the model, where It corresponds to the input vector Scalar output of ). Let The baseline input represents a reference point for signals lacking information (e.g., a zero vector or the average waveform in the signal data). Along the... i The integral gradient of dimension is defined as: ; The formula is input from... Transition to The time-cumulative model outputs information about the input features. The gradient of the model output. Intuitively, it measures the contribution of each feature to the change in the model output relative to the baseline. In practice, it is used along the interpolation path. The Riemann sum of discrete steps is used to approximate the integral: ; Attribution vector obtained Satisfying the completeness property ensures that the sum of attributions equals the difference between the model's output at the input and the baseline. .in, This represents the number of interpolation steps during discrete approximate integration. For the first The index of a discrete interpolation point.

[0042] At the surrogate interpretation level, the interpretation generator generates an attribution mask based on the input time-frequency scale map and applies the mask to the original input time-frequency scale map, forming a masked input with emphasized key regions. This masked input is then fed into the interpretation encoder to obtain a latent representation in the interpretation space. The interpretation predictor then outputs an auxiliary pulse position classification result corresponding to the classification and identification module based on this latent representation. By constraining the interpretability analysis module to maintain consistency with the classification and identification module of the PulseMed model in both predicted classification results and latent space structure, the self-supervised surrogate interpreter can learn an interpretation representation that matches the decision logic of the classification and identification module, thereby enhancing the credibility of the attribution results.

[0043] Finally, the attribution results are visualized at both the individual sample and population levels. At the individual signal level, the attribution scores are mapped back to the original pulse waveform, and the importance of different time points is shown through amplitude-time scatter plots, with warmer colors indicating a higher contribution to the classification result at that time point. At the population level, the attribution scores of all samples at the same pulse position are averaged to obtain the average importance profile, which is used to reveal the key time windows and discriminative waveform patterns that stably appear at the cun, guan, and chi pulse positions in the model's decision-making process.

[0044] In this embodiment, the training process of the PulseMed model includes two parts: training of the classification and recognition module and training of the self-supervised agent interpreter. Specifically, it includes the following steps: S1: Construct the PulseMed model and train the classification and recognition module of the PulseMed model using the acquired dataset to obtain the trained classification and recognition module. The dataset includes time-frequency scale maps of multiple different pulse diagnosis locations. The classification module adopts a first total loss function that includes cross-entropy loss and contrast loss during the training process.

[0045] Specifically, the UBenHealth pulse diagnosis device (UBJQR-002A) was used to collect wrist pulse signal data from recruited participants. Each participant recorded 90 seconds of continuous pulse signal data at a sampling frequency of 50Hz, resulting in a development dataset of 300 records (cun: 100, guan: 100, chi: 100) for model training and a validation dataset of 150 records (cun: 50, guan: 50, chi: 50) for model validation. Each data record underwent preprocessing and time-frequency transformation to obtain the corresponding time-frequency scaling map, forming the final dataset input to the PulseMed model.

[0046] In the training process of the classification and recognition module in this embodiment, the standard cross-entropy loss is used for classification, and contrastive loss is combined to enhance the discriminative representation between pulse locations.

[0047] Specifically, the first total loss function is expressed as: ; ; ; In the formula, The first total loss, For cross-entropy loss, To compare the losses, The number of training samples, For the number of categories, For real category labels, To predict class probabilities, Indicates the first The latent representation vector obtained by the classification and recognition module for each sample; Indicates and The latent representation vector that constitutes a positive sample pair Represents cosine similarity. For temperature parameters, Indicates the first The latent representation vectors obtained by the classification and recognition module for each sample The weighting coefficients are used to compare the losses.

[0048] For example, the classification module is trained for 25 epochs with a batch size of 64 and a learning rate of 1e-3 (reduced to 1e-4 at the 1000th iteration). The weight coefficients in the first total loss function... The value is 0.5.

[0049] S2: After the classification and recognition module is trained, its network parameters are fixed, the dataset is input into the PulseMed model, and the latent representation vector and pulse position classification results obtained by the classification and recognition module are used as supervision signals to train the self-supervised agent interpreter, thus obtaining the trained self-supervised agent interpreter; the self-supervised agent interpreter adopts a second total loss function including model behavior consistency loss and label consistency loss during the training process.

[0050] Understandably, the self-supervised surrogate interpreter model is used to learn representative explanatory patterns behind the classification decisions of the classification module of the PulseMed model. This self-supervised surrogate interpreter and the classification module of the PulseMed model form two corresponding paths: the classification module path provides the pulse position classification results and latent representation vectors from the original input, while the self-supervised surrogate interpreter path learns explanatory representations that can reproduce the behavior of the classification module. Specifically, the classification module first obtains the latent representation vector of the input time-frequency scale map through the Transformer encoder and outputs the predicted probabilities of cun, guan, and chi. Subsequently, the explanation generator in the self-supervised surrogate interpreter generates an attribution mask based on the input time-frequency scale map. This mask highlights the time-frequency regions most discriminative for classification while suppressing regions with low information content or irrelevant to classification. The masked time-frequency scale map is further fed into the explanation encoder to obtain the representation in the explanatory latent space, and finally, the explanation predictor outputs the corresponding prediction results as auxiliary pulse position classification results.

[0051] To ensure that the attribution masks and interpretive representations generated by the self-supervised surrogate interpreter accurately reflect the decision-making logic of the classification and recognition module, rather than learning pseudo-patterns irrelevant to classification, two types of consistency constraints connect the classification and recognition module path and the interpreter path. First, the model behavior consistency loss requires that the inter-sample relationships in the interpreter's latent space remain consistent with those in the classification and recognition module's latent space. That is, if two samples are similar in the classification and recognition module's representation space, they should also remain similar in the interpretive space; if two samples differ significantly in the classification and recognition module's representation space, this difference should also be reflected in the interpretive space. This constraint maintains the fidelity of the interpreter's representation of the internal structure of the classification and recognition module. Second, the label consistency loss requires that the interpreter's predictions remain consistent with the classification and recognition module's predictions, thereby ensuring that the masks and label waveforms generated by the interpreter truly serve the classification decision of the classification and recognition module.

[0052] In the training process of the self-supervised agent interpreter in this embodiment, the distance relationship between samples in the latent space is explained by the model behavior consistency loss constraint to ensure that the distance relationship between samples in the latent space of the classification and recognition module is consistent, and the output of the predictor is explained by the label consistency loss constraint to ensure that the output of the classification and recognition module is consistent.

[0053] Specifically, the second total loss function is expressed as: ; ; ; In the formula, This is the second total loss. For model behavior consistency loss, For label consistency loss, The weighting coefficients for label consistency loss are: They represent the first The first sample and the first The latent representation vectors obtained by the classification and recognition module for each sample The outputs of the self-supervised agent interpreter are respectively the first... The first sample and the first The latent representation of each sample, Represents the cross-entropy function. For the classification and recognition module, the first The predicted probability distribution of each sample output. To explain the predictor's effect on the first The auxiliary prediction probability distribution of each sample output. This represents the distance metric between samples in the latent space of the classification and recognition module. This represents the distance metric between samples in the latent space.

[0054] After training, the self-supervised surrogate interpreter can generate multi-level interpretability outputs. The first type is a saliency map or attribution mask, used to display the key time-frequency regions the model focuses on; the second type is an interpretability latent space, used to visualize the clustering structure of the cun, guan, and chi samples in the interpretability space; the third type is representative waveform markers, used to summarize the most representative prototypical patterns for each pulse position. Combined with the time importance curves obtained from the aforementioned integral gradient unit, these results together constitute the interpretability evidence chain of the PulseMed model, enabling the model not only to output pulse position classification results but also to explain which time periods, frequency regions, and waveform morphologies it primarily relies on for its judgment.

[0055] This invention presents a pulse position recognition method integrating an interpretable module, utilizing the PulseMed model for pulse position classification prediction and interpretability analysis. The PulseMed model employs a Swin-Tiny visual Transformer architecture, effectively capturing local details and global contextual information in the pulse signal time-frequency scale map through window attention and shifted window attention mechanisms. Simultaneously, a contrastive learning loss function is introduced to narrow the distance between similar pulse position samples in the latent space and widen the distance between dissimilar samples, significantly improving feature discriminative power. The interpretability analysis module, integrating an integral gradient unit and a self-supervised surrogate interpreter, quantifies the contribution of each time point in the input waveform to the classification result, generating a time importance curve that visually displays key discrimination periods in color-coded form. Furthermore, it learns to generate attribution masks (highlighting discriminative regions in the time-frequency scale map) and extracts prototype marker waveforms from the interpretable latent space, providing a typical pulse morphology template for each pulse position. These interpretation results are highly consistent with traditional Chinese medicine's organ theory, enhancing trust in the model's decisions.

[0056] Secondly, embodiments of the present invention provide a pulse position recognition system with an integrated interpretable module, which can be used to implement the pulse position recognition method with an integrated interpretable module as described in the first aspect above.

[0057] Please see Figure 3 , Figure 3 The diagram shown is a structural block diagram of a pulse position recognition system that integrates interpretable modules. Figure 3 As shown, the pulse position recognition system with a fusion interpretable module in this embodiment includes a signal acquisition module, a signal preprocessing and time-frequency transformation module, and a data processing module. The signal acquisition module acquires multi-channel pulse pressure signals from the radial artery of the object to be identified; the signal preprocessing and time-frequency transformation module preprocesses and transforms the multi-channel pulse pressure signals to generate at least one time-frequency scaling map; the data processing module inputs the time-frequency scaling map into a pre-trained PulseMed model and outputs the corresponding pulse position classification result and the attribution information corresponding to the pulse position classification result.

[0058] In this embodiment, the signal preprocessing and time-frequency transformation module includes a signal preprocessing module and a time-frequency transformation module. The signal preprocessing module uses a fourth-order Butterworth bandpass filter to filter the multi-channel pulse pressure signal, with a filtering range of 0.7Hz-10Hz. The time-frequency transformation module uses a complex Morlet wavelet to perform a continuous wavelet transform on the filtered multi-channel pulse pressure signal to obtain an initial time-frequency scale map. This initial time-frequency scale map is then uniformly scaled to a size of 224×224 pixels to obtain the final time-frequency scale map.

[0059] In this embodiment, the PulseMed model includes a classification and identification module and an interpretability analysis module. The classification and identification module obtains a latent representation vector representing the pulse signal based on the time-frequency scaling graph, and obtains the pulse position classification result based on the latent representation vector. The interpretability analysis module is used to generate attribution information corresponding to the pulse position classification result based on the time-frequency scaling graph. The attribution information includes at least one of the following: attribution mask, interpretable latent representation, time importance curve, and representative marker waveform.

[0060] For details regarding the pulse position recognition system with fused interpretable modules and its corresponding beneficial effects, please refer to the relevant content of the pulse position recognition method with fused interpretable modules provided in the first aspect, which will not be repeated here.

[0061] Furthermore, the beneficial effects of the pulse position recognition method with integrated interpretable modules in this embodiment are illustrated through simulation experiments.

[0062] Comparative experiments were conducted on the proposed PulseMed model, ResNet-50, and CNN (baseline) using the development and validation datasets. The performance results of the models using five-fold cross-validation on the development and validation datasets are shown in Table 1, with the results reported as mean + / - standard deviation.

[0063] Table 1

[0064] As shown in Table 1, the PulseMed model proposed in this invention achieves the highest overall performance of 83.05+ / -2.0% accuracy, which is better than ResNet-50 (76.9+ / -2.1% accuracy) and CNN (72.5+ / -2.4% accuracy).

[0065] Please see Figure 4 , Figure 4 The diagram shows the confusion matrix and ROC curves for comparing PulseMed models on the development and validation datasets.

[0066] Figure 4Figures a, b, and c in the table show the confusion matrices of the three models on the development dataset, revealing their detailed classification patterns. The PulseMed model (Figure a) exhibits the strongest diagonal advantage, correctly predicting 14,918 incisions, 12,739 gates, and 14,627 feet. The most common misclassifications occur between gates and adjacent locations (1,612 gate samples were misclassified as incisions, and 1,629 as feet), consistent with the intermediate anatomical location of gates and overlapping hemodynamic features. In contrast, CNN (Figure b) shows significantly weaker diagonal values ​​and more scattered errors, correctly predicting only 11,028 gates. ResNet-50 (Figure c) achieves moderate performance, correctly predicting 11,755 gates, confirming that its deeper residual architecture partially bridges the gap between simple CNNs and Transformer-based models.

[0067] Figure 4 Figures d, e, and f in the diagram show the ROC curves for the three models at the cun, guan, and chi pulse locations on the development dataset, demonstrating their discrimination performance at each location. For cun classification (Figure d), the PulseMed model achieves an AUC of 0.889, significantly outperforming CNN and ResNet-5. Similar superiority is observed for chi classification (Figure f), with PulseMed achieving an AUC of 0.891, CNN achieving 0.659, and ResNet-50 achieving 0.805. Guan discrimination (Figure e) remains challenging for all models, but PulseMed's AUC of 0.836 still surpasses CNN's AUC of 0.633 and ResNet-50's AUC of 0.775. The PulseMed model demonstrates consistent superiority across all locations and metrics, with AUCs exceeding 88% for cun and chi.

[0068] Figure 4 The g, h, and i plots in the figure show the confusion matrices of the three models on the validation dataset, confirming a consistent generalization pattern across models. PulseMed (g plot) correctly classified 6457 inch, 5507 door, and 6453 foot, maintaining a strong diagonal advantage despite independent participant pools. The misclassification pattern is similar to that observed in the development set, with door showing the greatest confusion (1211 misclassified as inch, 1282 as foot). In contrast, CNN (h plot) shows significant degradation on external data, correctly predicting only 4572 doors with widespread errors across categories. ResNet-50 (i plot) shows moderate robustness, correctly identifying 5082 doors, but still lower than the PulseMed model.

[0069] Figure 4Figures j, k, and l in the table show the ROC curves for the three models at the cun, guan, and chi pulse positions on the validation dataset. Their position-specific accuracy follows the same pattern as the development data: cun (80.38 ± 3.4%), chi (81.26 ± 1.6%), and guan (68.69 ± 2.3%). The AUC for cun and chi remained above 84% across all tradeoffs. The modest generalization gap between the development and validation datasets (average AUC reduction of 5-7%) highlights the robustness of the PulseMed model to subject-level variability.

[0070] Please see Figure 5 , Figure 5 The results of the model interpretability analysis using integral gradients are shown. Figure a shows the global time-attributed average for all samples of the Cun, Guan, and Chi pulses. Among the three, Cun consistently shows the highest importance, with all positions showing an increase in attribution towards later time windows. However, the early window of Cun is still more informative than Chi and Guan. Figure b shows the local attribution of the Cun pulse signal, with importance concentrated in the early segment (<20 seconds). Figure c shows the local attribution of the Guan pulse signal, highlighting the moderate importance of the mid-segment (30-50 seconds). Figure d shows the local attribution of the Chi pulse signal, demonstrating a more dispersed contribution extending to the later window (40-80 seconds).

[0071] Please see Figure 6 , Figure 6 The diagram shows the self-supervised surrogate interpreter's learned embedding space and the identification of representative markers in the interpretation results. By computing the attribution plot of the time-sensor pairs, the interpreter captures characteristic temporal patterns that help reveal broader trends in the dataset. Representative marker waveforms derived from the surrogate interpreter's latent space illustrate the characteristic temporal patterns the model uses to distinguish the three pulse locations. Markers 1-3 correspond to the cun pulse, showing higher-frequency oscillations and a pronounced early contraction component; markers 4-6 represent the guan pulse, exhibiting a wider, asymmetrical waveform and increased mid-segment amplitude; markers 7-9 depict the chi pulse, characterized by slower, smoother oscillations and decaying late-stage activity.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0073] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0074] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A pulse position recognition method integrating an interpretable module, characterized in that, include: Step 1: Acquire the multi-channel pulse pressure signal of the radial artery of the object to be identified; Step 2: Preprocess and perform time-frequency transformation on the multi-channel pulse pressure signal to generate at least one time-frequency scaling map; Step 3: Input the time-frequency scaling map into the pre-trained PulseMed model and output the corresponding pulse position classification results and the attribution information corresponding to the pulse position classification results; The PulseMed model includes a classification and identification module and an interpretability analysis module. The classification and identification module obtains a latent representation vector representing the pulse signal based on the time-frequency scaling map, and obtains the pulse position classification result based on the latent representation vector. The interpretability analysis module is used to generate attribution information corresponding to the pulse position classification result based on the time-frequency scaling map.

2. The pulse position recognition method with fused interpretable modules according to claim 1, characterized in that, Step 2 includes: Step 2.1: The multi-channel pulse pressure signal is filtered using a fourth-order Butterworth bandpass filter with a filtering range of 0.7Hz-10Hz; Step 2.2: Perform continuous wavelet transform on the filtered multi-channel pulse pressure signal using complex Morlet wavelets to obtain the initial time-frequency scale map; Step 2.3: Scale the initial time-frequency scale map to a uniform size of 224×224 pixels to obtain the time-frequency scale map.

3. The pulse position recognition method with integrated interpretable modules according to claim 1, characterized in that, The pulse position categories include the cun pulse position, guan pulse position, and chi pulse position; the attribution information includes at least one of the following: attribution mask, interpretation latent representation, time importance curve, and representativeness waveform.

4. The pulse position recognition method with fused interpretable modules according to claim 1, characterized in that, The classification and recognition module adopts a Swing Transformer-based architecture, comprising: a cascaded image block partitioning unit, a linear embedding unit, a multi-layer Transformer coding unit, a global pooling unit, and a classification output unit. The image block division unit is used to divide the time-frequency scale map into multiple non-overlapping image blocks; The linear embedding unit is used to map each image block to a token embedding vector and add position encoding to preserve the time-frequency spatial arrangement information; In the multi-layer Transformer coding unit, each layer of Transformer coding unit includes a first sub-coding unit and a second sub-coding unit connected in sequence. The first sub-coding unit captures the correlation of the time-frequency region within a local window based on window multi-head self-attention to form a first intermediate representation. The second sub-coding unit realizes information interaction between adjacent windows based on shift window multi-head self-attention, and fuses local features with global context to form the output of the Transformer coding unit of this layer. The global pooling unit is used to perform global average pooling on the feature map output by the multi-layer Transformer coding unit to obtain the latent representation vector. The classification output unit is used to output the predicted probabilities of the three pulse position categories of Cun, Guan, and Chi based on the latent representation vector, so as to obtain the pulse position classification result.

5. The pulse position recognition method with fused interpretable modules according to claim 4, characterized in that, The first sub-encoding unit includes: a first normalization layer and a window multi-head self-attention layer. The first normalization layer performs layer normalization processing on the input token sequence of the Transformer encoding unit at this layer and then inputs it to the window multi-head self-attention layer. The output of the window multi-head self-attention layer is added to the input token sequence of the Transformer encoding unit at this layer through a first residual connection to obtain the first intermediate representation. The second sub-encoding unit includes: a second normalization layer, a shift-window multi-head self-attention layer, and a multilayer perceptron. The second normalization layer performs layer normalization processing on the first intermediate representation and then inputs it to the shift-window multi-head self-attention layer. The output of the shift-window multi-head self-attention layer is added to the first intermediate representation through a second residual connection to obtain a second intermediate representation. The multilayer perceptron performs a nonlinear transformation on the second intermediate representation, and its output is added to the second intermediate representation through a third residual connection as the output of the Transformer encoding unit of this layer.

6. The pulse position recognition method with fused interpretable modules according to claim 1, characterized in that, The interpretability analysis module includes: an integral gradient unit and a self-supervised proxy interpreter; wherein... The integral gradient unit is used to calculate the gradient integral of the output of the classification and recognition module relative to each time point of the time-frequency scale map, obtain the importance score of the time dimension, and map the importance score back to the original pulse waveform to generate a time importance curve. The self-supervised proxy interpreter is used to generate an attribution mask based on the time-frequency scale map, and then use the attribution mask to perform a masking operation on the time-frequency scale map and map it to the interpretation latent space to obtain the interpretation latent representation. The interpretation latent representation is then compared with the interpretation latent representation corresponding to the prototype flag waveform pre-stored in the training phase to calculate the cosine similarity, and the most matching prototype flag waveform is determined as the representative flag waveform output.

7. The pulse position recognition method with integrated interpretable modules according to claim 6, characterized in that, The self-supervised proxy interpreter includes: an interpretation generator, an interpretation encoder, an interpretation predictor, and a prototype extraction subunit; wherein, The interpretation generator is used to generate the attribution mask based on the time-frequency scale map, and the attribution mask is used to highlight the discriminative regions in the time-frequency scale map; The interpretive encoder is used to perform a masking operation between the attribution mask and the time-frequency scaling map and map it to the interpretive latent space to obtain the interpretive latent representation. At the same time, it receives the latent representation vector generated by the classification and recognition module as a supervision signal for model behavior consistency constraints during the model training phase. The interpretation predictor is used to output an auxiliary pulse position classification result based on the interpretation latent representation, and during the model training phase, the auxiliary pulse position classification result and the pulse position classification result output by the classification and recognition module are used to perform label consistency constraints. The prototype extraction subunit is used to perform cosine similarity calculation between the explanatory latent representation and the explanatory latent representation corresponding to the prototype marker waveform pre-stored in the training phase, determine the most matching prototype marker waveform as the representative marker waveform output, and cluster the explanatory latent representations of the same pulse position in the model training phase, and extract the cluster center as the prototype marker waveform.

8. The pulse position recognition method with fused interpretable modules according to claim 6, characterized in that, The training process of the PulseMed model includes: S1: Construct a PulseMed model, and train the classification and recognition module of the PulseMed model using the acquired dataset to obtain a trained classification and recognition module. The dataset includes time-frequency scaling maps of multiple different pulse diagnosis locations. The classification module adopts a first total loss function including cross-entropy loss and contrast loss during the training process. S2: After the classification and recognition module is trained, its network parameters are fixed, the dataset is input into the PulseMed model, and the latent representation vector and pulse position classification result obtained by the classification and recognition module are used as supervision signals to train the self-supervised agent interpreter to obtain the trained self-supervised agent interpreter; the self-supervised agent interpreter adopts a second total loss function including model behavior consistency loss and label consistency loss during the training process.

9. The pulse position recognition method with fused interpretable modules according to claim 8, characterized in that, The first total loss function is expressed as: ； ； ； In the formula, The first total loss, For cross-entropy loss, To compare the losses, The number of training samples, For the number of categories, For real category labels, To predict class probabilities, Indicates the first The latent representation vector obtained by the classification and recognition module for each sample; Indicates and The latent representation vector that constitutes a positive sample pair Represents cosine similarity. For temperature parameters, Indicates the first The latent representation vectors obtained by the classification and recognition module for each sample Weighting coefficients for comparison losses; The second total loss function is expressed as: ； ； ； In the formula, This is the second total loss. For model behavior consistency loss, For label consistency loss, The weighting coefficients for label consistency loss are: They represent the first The first sample and the first The latent representation vectors obtained by the classification and recognition module for each sample The outputs of the self-supervised agent interpreter are respectively the first... The first sample and the first The latent representation of each sample, Represents the cross-entropy function. For the classification and recognition module, the first The predicted probability distribution of each sample output. To explain the predictor's effect on the first The auxiliary prediction probability distribution of each sample output. This represents the distance metric between samples in the latent space of the classification and recognition module. This represents the distance metric between samples in the latent space.

10. A pulse position recognition system integrating an interpretable module, characterized in that, The pulse position recognition method applicable to the fusion interpretable module according to any one of claims 1-9 includes: The signal acquisition module acquires multi-channel pulse pressure signals from the radial artery of the object to be identified; The signal preprocessing and time-frequency transformation module is used to preprocess and perform time-frequency transformation on the multi-channel pulse pressure signal to generate at least one time-frequency scale map. The data processing module is used to input the time-frequency scaling map into the pre-trained PulseMed model and output the corresponding pulse position classification results and the attribution information corresponding to the pulse position classification results; The PulseMed model includes a classification and identification module and an interpretability analysis module. The classification and identification module obtains a latent representation vector representing the pulse signal based on the time-frequency scaling map, and obtains the pulse position classification result based on the latent representation vector. The interpretability analysis module is used to generate attribution information corresponding to the pulse position classification result based on the time-frequency scaling map.

Images