A small-sample exhaled breath detection method and system based on channel attention mechanism

Abstract

This invention discloses a small-sample exhaled air detection method and system based on a channel attention mechanism. It relates to the fields of machine learning and gas sensor technology. The method includes: Step 1: Acquiring and preprocessing exhaled air data to obtain exhalation time-series data; Step 2: Constructing a 1D-CNN feature extraction module to extract local key features from the exhalation time-series data, obtaining local dynamic features; Step 3: Adaptively enhancing the local dynamic features through a channel attention module to obtain enhanced features; Step 4: Compressing and abstracting the enhanced features to output disease classification probabilities. This invention can effectively improve the accuracy of gas detection.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and gas sensor technology, and more specifically to a small-sample exhaled breath detection method and system based on a channel attention mechanism. Background Technology

[0002] Gaseous signaling molecules are endogenous molecules that play a bioindicative and regulatory role in physiological and pathological changes in the body. They are usually produced by cells or bacteria within the body. By measuring changes in the concentration of these gaseous signaling molecules exhaled from the body, pathological processes such as inflammation, bacterial overgrowth, or dysbiosis in the respiratory and digestive tracts can be monitored.

[0003] Currently, breath tests for NO and CO are widely used in the diagnosis and treatment of inflammation in respiratory diseases such as asthma and COPD, while breath tests for H2 and CH4 are used for the diagnosis and assessment of intestinal bacterial overgrowth or flora imbalance in irritable bowel syndrome and inflammatory bowel disease. For example, a 1977 study in *The Lancet* reported on the use of breath tests for CH4 in colorectal cancer, and a 2003 study in *British Journal of Respiratory Medicine* reported on the use of breath tests for NO in lung cancer.

[0004] With the development of new-generation computer information technology, machine learning algorithms based on neural network structures, after being trained on a large amount of data, can accurately analyze the hidden features of exhalation signals and quickly provide highly sensitive and specific pattern recognition results.

[0005] However, in the medical field, clinical cases are scarce and samples are scattered, making it difficult to collect large-scale breath test data in a centralized manner. This makes it difficult to apply traditional data processing methods and machine learning models in the medical field.

[0006] Therefore, how to combine limited clinical data with deep learning algorithms to improve the accuracy of disease diagnosis has become an urgent problem to be solved in the medical field. Summary of the Invention

[0007] In view of this, the present invention provides a small-sample exhaled gas detection method and system based on channel attention mechanism, which can effectively improve the accuracy of gas detection.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A small-sample exhaled breath detection method based on channel attention mechanism includes: Step 1: Acquire exhaled air data and preprocess it to obtain exhalation timing data; Step 2: Construct a 1D-CNN feature extraction module to extract local key features from the exhalation time series data and obtain local dynamic features; Step 3: Adaptively enhance local dynamic features using the channel attention module to obtain enhanced features; Step 4: Perform feature compression and abstraction on the enhanced features and output the disease classification probability.

[0009] Preferably, the preprocessing process in step 1 includes: baseline subtraction, physiological range clipping, and channel-level standardization, resulting in a tensor of batch size B × time step T × number of channels C.

[0010] Preferably, the 1D-CNN feature extraction module is used to extract disease-related local dynamic features from standardized expiratory time-series data, including two cascaded convolution-pooling-batch normalization sub-blocks.

[0011] Preferably, the channel attention module includes: a global information aggregation layer, a shared fully connected layer, an attention weight generation layer, and a feature weighting layer; The global information aggregation layer is used to receive local dynamic features and aggregate global information of each virtual channel from different dimensions through global average pooling (GAP) and global max pooling (GMP). The shared fully connected layer feeds the two feature vectors generated by global average pooling (GAP) and global max pooling (GMP) into a shared two-layer fully connected network, and strengthens the virtual channel features that are strongly correlated with the disease through compression-restore operations. The attention weight generation layer performs feature fusion by adding the two feature vectors after passing through the shared fully connected layer element by element, and then using the sigmoid activation function to map the fused feature vector to the [0,1] interval to generate channel attention weights. The feature weighting layer strengthens key channels and suppresses redundant channels based on channel attention weights.

[0012] Preferably, step 4 utilizes the classification module to transform the enhanced features output by the channel attention module into disease classification probabilities through a flattening-fully connected-classification process, specifically including: The flattening layer flattens the enhanced features into a one-dimensional vector, transforming the temporal-channel structured features into one-dimensional features that can be processed by the fully connected layer. Fully connected layers extract abstract features across time and channels from one-dimensional features; The output layer sets the number of neurons and outputs a probability vector based on abstract features.

[0013] Preferably, it also includes: small sample adaptation training, using the Adam optimizer, sparse classification cross-entropy loss function, combined with early stopping strategy and learning rate decay strategy to train the model, to ensure the convergence and generalization ability of the model in small sample scenarios.

[0014] Preferably, a small-sample exhaled breath detection system based on a channel attention mechanism includes: Data acquisition module: used to acquire exhaled air data and preprocess it to obtain exhalation timing data; Feature extraction module: used to extract local key features from expiratory time-series data to obtain local dynamic features; Channel attention module: used to adaptively enhance local dynamic features to obtain enhanced features; Output module: Used to compress and abstract the enhanced features and output the disease classification probability.

[0015] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a small sample exhaled breath detection method and system based on channel attention mechanism to solve the problem of model underfitting caused by the scarcity of clinical samples and improve the accuracy of diagnosing diseases through exhaled breath concentration. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 The method flowchart provided by the present invention.

[0018] Figure 2 A schematic diagram of the channel attention combined with the 1D-CNN algorithm provided by this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The core of this invention is to address the unique characteristics of multi-channel time-series data of exhaled breath by capturing key features in the time dimension using 1D-CNN and combining this with a channel attention mechanism to filter core gas signals in multiple channels, ultimately achieving accurate disease classification.

[0021] like Figure 1 As shown, this embodiment of the invention discloses a small-sample exhaled breath detection method based on a channel attention mechanism, comprising: Step 1: Acquire exhaled air data and preprocess it to obtain exhalation timing data; Step 2: Construct a 1D-CNN feature extraction module to extract local key features from the exhalation time series data and obtain local dynamic features; Step 3: Adaptively enhance local dynamic features using the channel attention module to obtain enhanced features; Step 4: Perform feature compression and abstraction on the enhanced features and output the disease classification probability.

[0022] In a specific embodiment of the present invention, the adaptation of the algorithm application background and data characteristics includes: The acquisition process for exhaled breath data has clear clinical constraints: it typically involves continuous acquisition at a frequency of 10-20 Hz for 30-40 seconds (corresponding to 300-800 time steps), with the concentration values ​​of 4-6 core gases (such as NO, H2S, H2, CH4) recorded simultaneously at each time step, forming two-dimensional time-series data in the shape of "time step (T) × number of channels (C)". Among these: Time step (T): corresponds to the complete physiological process of exhalation (e.g., T=342, representing a 34.2-second exhalation sequence, covering the three stages of "initial exhalation (baseline) → mid-expiration (peak concentration) → end-expiration (concentration decline)"). Channel number (C): corresponds to different gaseous signal molecules (e.g., when C=4, channel 1=NO, channel 2=H2S, channel 3=H2, channel 4=CH4), and the association strength between different gases and diseases varies significantly (e.g., NO is a strong indicator of asthma, while CH4 is a strong indicator of intestinal adenoma).

[0023] While traditional 1D-CNN can extract temporal features, it treats all channels "equally" and is easily interfered with by noise from irrelevant gases. In contrast, channel attention mechanisms can enhance the feature contribution of key gases through adaptive weight allocation. The combination of the two can perfectly adapt to the characteristics of exhaled breath data.

[0024] Specifically, the preprocessing process in step 1 includes baseline subtraction, physiological range clipping, and channel-level standardization, resulting in a tensor of batch size B × time step T × number of channels C.

[0025] Specifically, the 1D-CNN feature extraction module is used to extract disease-related local dynamic features from standardized expiratory time-series data, including two cascaded convolution-pooling-batch normalization sub-blocks.

[0026] In a specific embodiment of the present invention, such as Figure 2 As shown, the complete architecture and module details of the CA+1D-CNN algorithm include: (a) Input layer: Standardization processing of multi-channel time series data The algorithm receives a tensor of "batch size (B) × time step (T) × number of channels (C)" (e.g., in a small clinical sample scenario, B=16, T=342, C=4), but the raw data must first be standardized by physiological constraints to avoid interference from sensor noise and individual differences. Baseline subtraction: Record the ambient air signal for 10 seconds before collecting exhaled breath, calculate the background mean of each channel, and subtract the background value from the original signal (e.g., original NO concentration = 5ppb, background mean = 2ppb, corrected = 3ppb) to eliminate ambient gas interference. Physiological range trimming: Based on clinical consensus, reasonable concentration ranges for each gas are set (NO: 0-50ppb, H2S: 0-20ppb, H2: 0-200ppm, CH4: 0-150ppm). Abnormal values ​​outside the range (such as 100 ppb NO caused by sensor drift) are trimmed to the boundary value to ensure the physiological rationality of the data. Channel-level standardization: Z-score standardization is performed separately for each channel to avoid bias towards higher-level channels due to differences in gas concentration magnitudes (e.g., H2 in ppm, H2S in ppb). The standardized input data retains the dynamic changes in the exhalation sequence while eliminating non-disease-related interference, laying the foundation for subsequent feature extraction.

[0027] (ii) 1D-CNN Feature Extraction Module: Capturing Local Key Features of the Exhalation Timing

[0028] This module contains two concatenated "convolution-pooling-batch normalization" sub-blocks. Its core function is to extract disease-related local dynamic features (such as mid-expiratory concentration peaks, concentration rise / fall slopes, etc.) from standardized time-series data. Specific parameters and design logic are as follows: 1. First convolutional sub-block (shallow feature extraction) 1D convolutional layer: Number of filters: 64 (to balance the extraction capability and computational load of shallow temporal features; 64 filters can initially capture different types of concentration change patterns, such as slow rise, rapid peak, plateau, etc.). Kernel size: 3 (corresponding to the temporal window length, such as 3 time steps = 300ms, which matches the smallest time scale of concentration change during exhalation - the response time of concentration change during human exhalation is about 200-500ms, and 3 time steps can completely cover a small fluctuation). Filling method: "Same" (keep the output timing length consistent with the input to avoid losing edge features of the beginning / end of expiration, such as the rapid rise in NO concentration at the beginning of expiration). Activation function: ReLU (alleviates the gradient vanishing problem and has a natural inhibitory effect on negative concentration values ​​(which have been eliminated through physiological pruning), suitable for nonlinear mapping of temporal features); Regularization: L2 regularization (weight decay coefficient = 1e) -5 To avoid overfitting in small sample scenarios and prevent the model from becoming overly reliant on a particular local temporal pattern.

[0029] Max pooling layer: Pooling window size: 2, stride: 2 (compresses the time series length to half, e.g., T=342→171). Max pooling can retain the "maximum concentration" within each time series window, corresponding to key peaks during exhalation (such as the peak concentration of CH4 in the mid-exhalation phase of a patient's breath, which is an important biomarker for the disease), while reducing the computational load by half and improving model training efficiency.

[0030] Batch normalization layer: Standardizes the feature map output by convolution (mean=0, variance=1), accelerates model convergence, and avoids training stagnation caused by initial weight bias. It is especially suitable for fast and stable training in small sample scenarios.

[0031] 2. Second convolutional sub-block (deep feature extraction)

[0032] 1D convolutional layer: Number of filters: 128 (based on the 64 shallow features, the feature dimension is further improved to capture more complex temporal combination patterns, such as features related to complex diseases, such as "delayed NO peak + prolonged CH4 plateau phase"). Other parameters: kernel=3, padding="Same", ReLU activation function, L2 regularization (same as the first convolutional layer); Design logic: Deep convolution can integrate local features extracted from shallow layers (such as a single concentration peak) to form a more abstract temporal pattern (such as a complete concentration curve feature of "rising phase-plateau phase-falling phase"), which is more in line with the overall impact of the disease on the exhalation process.

[0033] Max pooling layer: Pooling window = 2, stride = 2 (timing length further compressed to 85, e.g., T = 171 → 85); Clinical significance: The 85 time steps that are ultimately retained correspond to the core phases of the expiratory process (such as the effective expiratory segment from 5 to 35 seconds), filtering out the unstable segment at the beginning of exhalation and the residual segment at the end, focusing on key time sequence information related to the disease.

[0034] Batch normalization layer: Same as the first sub-block, it ensures the stable distribution of deep features and provides high-quality input for subsequent attention modules.

[0035] Specifically, the channel attention module includes: a global information aggregation layer, a shared fully connected layer, an attention weight generation layer, and a feature weighting layer; The global information aggregation layer is used to receive local dynamic features and aggregate global information of each virtual channel from different dimensions through global average pooling (GAP) and global max pooling (GMP). The shared fully connected layer feeds the two feature vectors generated by global average pooling (GAP) and global max pooling (GMP) into a shared two-layer fully connected network, and strengthens the virtual channel features that are strongly correlated with the disease through compression-restore operations. The attention weight generation layer performs feature fusion by adding the two feature vectors after passing through the shared fully connected layer element by element, and then using the sigmoid activation function to map the fused feature vector to the [0,1] interval to generate channel attention weights. The feature weighting layer strengthens key channels and suppresses redundant channels based on channel attention weights.

[0036] In a specific embodiment of the present invention, the channel attention module is used to adaptively enhance the feature contribution of key gases. This module is the core innovation of the algorithm. It distinguishes the importance of different gas channels through four steps: "global information aggregation → channel feature compression → weight generation → feature weighting". The specific process and technical details are as follows: 1. Global Information Aggregation: Capturing the global temporal characteristics of the channel. The module receives the output of the second convolutional sub-block (with a shape of "B×85×128", where 128 represents the number of convolutional filters, which can be considered as "virtual channels" corresponding to different temporal feature patterns). It then aggregates the global information of each virtual channel from different dimensions through two operations: Global Average Pooling (GAP) and Global Max Pooling (GMP). Global Average Pooling (GAP): The average value is calculated over 85 time steps for each virtual channel to generate a B×128 feature vector. The formula is as follows:

[0037] in Let be the feature value of the c-th virtual channel at the t-th time step.

[0038] GAP reflects the "average response strength" of the virtual channel, corresponding to the overall impact of disease on the exhalation process (such as chronic inflammation leading to an increase in the average characteristic value of the NO channel).

[0039] Global Max Pooling (GMP): The maximum value is taken over 85 time steps for each virtual channel to generate a B×128 feature vector. The formula is as follows:

[0040] GMP reflects the "peak response intensity" of the virtual channel, corresponding to the drastic changes at a certain moment during the exhalation process caused by disease (such as a sudden increase in the peak characteristic value of the H2S channel due to acute infection).

[0041] 2. Shared fully connected layer: Compresses channel features and enhances abstract expression.

[0042] The two B×128 feature vectors generated by GAP and GMP are fed into a shared two-layer fully connected network. Through a "compression-restoration" operation, the number of parameters is reduced and the feature abstraction capability is improved. Compression layer (first fully connected layer): Output dimension: 128 / ratio (ratio is the compression ratio, usually taken as 8, i.e., output dimension = 16). Activation function: ReLU By compressing the 128-dimensional channel features to 16 dimensions, the model is forced to discard redundant information and focus on the core correlations between channels (such as the synergistic changes of NO and H2S features), while reducing the number of parameters (128×16=2048, far fewer than directly processing 128 dimensions) to avoid overfitting.

[0043] Recovery layer (second fully connected layer): Output dimension: 128 (restored to the original number of virtual channels); Activation function: None (preserves the original feature scale for easier subsequent fusion); The compressed 16-dimensional abstract features are restored to 128 dimensions, realizing the "re-encoding" of channel features and strengthening valuable feature patterns (such as virtual channel features that are strongly correlated with diseases).

[0044] 3. Attention Weight Generation: Quantifying Channel Importance

[0045] Feature fusion: The two B×128 feature vectors of GAP and GMP, after passing through a shared fully connected layer, are added element-wise, using the following formula:

[0046] By combining information from "average response" and "peak response", we can avoid feature loss caused by a single pooling operation (e.g., using only GAP may miss the peak features of acute lesions, while using only GMP may ignore the average features of chronic lesions).

[0047] Weight normalization: The fused feature vector is mapped to the "0,1" interval using the Sigmoid activation function to generate the channel attention weight W(c), as shown in the formula:

[0048] The value of W(c) represents the importance of the c-th virtual channel. For example, in the diagnosis of adenoma, the weight of the virtual channel corresponding to the CH4 peak feature is W(c) = 0.85 (important), and the weight of the virtual channel corresponding to the H2 unrelated fluctuation is W(c) = 0.12 (not important).

[0049] Feature weighting: Key channels are strengthened, redundant channels are suppressed. The generated channel attention weight W(c) is multiplied channel-by-channel by the output feature map "B×85×128" of the second convolutional sub-block, as follows:

[0050] in The weighted eigenvalues These are the original convolutional feature values. After weighting, virtual channel features strongly correlated with the disease (such as the CH4 peak) are amplified (0.85×). Irrelevant channel features (such as H2 noise) are suppressed (0.12×). The model can focus on core disease characteristics and reduce noise interference.

[0051] Specifically, step 4 utilizes the classification module to transform the enhanced features output by the channel attention module into disease classification probabilities through a flattening-fully connected-classification process, specifically including: The flattening layer flattens the enhanced features into a one-dimensional vector, transforming the temporal-channel structured features into one-dimensional features that can be processed by the fully connected layer. Fully connected layers extract abstract features across time and channels from one-dimensional features; The output layer sets the number of neurons and outputs a probability vector based on abstract features.

[0052] In a specific embodiment of the present invention, the classification module is used for feature fusion and disease probability output, and the output of the channel attention module ("B×85×128") needs to go through the "flattening-fully connected-classification" process to transform the temporal-channel features into disease classification probabilities: 1. Flattened layer: The two-dimensional feature map of "B×85×128" is flattened into a one-dimensional vector of "B×10880" (10880=85×128). The purpose is to transform the structured features of time series and channels into one-dimensional features that can be processed by the fully connected layer, so as to achieve feature fusion across time series and channels.

[0053] 2. Fully Connected Layer (Feature Compression and Abstraction)

[0054] First fully connected layer: Number of neurons: 128; Activation function: ReLU; Dropout layer: dropout rate=0.5 (randomly deactivates 50% of neurons to avoid overfitting in small sample scenarios and prevent the model from memorizing noise in the training data); The high-dimensional features of 10,880 dimensions are compressed to 128 dimensions to extract abstract features across time series and channels (such as complex disease features like "average NO increase + CH4 peak delay").

[0055] Output layer (disease classification probability)

[0056] Number of neurons: N (N is the number of disease categories, such as N=2 for binary classification "health / adenoma" and N=3 for multi-category classification "health / adenoma / inflammation"). Activation function: Softmax; Output: A probability vector of "B×N", where each element represents the probability that the sample belongs to the corresponding category (e.g., "0.12, 0.88" represents a 12% probability of being healthy and an 88% probability of being an adenoma). Specifically, this also includes: small sample adaptation training, using the Adam optimizer, sparse classification cross-entropy loss function, combined with early stopping strategy and learning rate decay strategy to train the model, ensuring the convergence and generalization ability of the model in small sample scenarios.

[0057] In a specific embodiment of the present invention, the model training strategy needs to be adapted to small sample clinical data. Considering the scarcity of clinical exhaled breath samples (e.g., only 72 cases), the following training strategy is adopted to ensure model convergence and strong generalization ability: Optimizer: Adam optimizer, parameters set as follows: Initial learning rate: 0.001 (balancing convergence speed and stability; a small learning rate avoids oscillations and is suitable for small samples); beta1=0.9, beta2=0.999 (momentum parameters, accelerating gradient descent and avoiding local optima); epsilon=1e -7 (To prevent the denominator from being zero and improve numerical stability).

[0058] Loss function: Sparse Categorical Crossentropy

[0059] Applicable scenarios: When the tags are integer codes (e.g., 0=health, 1=adenoma, 2=inflammation), one-hot encoding is not required, reducing the amount of computation; Loss calculation: The probability that the i-th sample corresponds to the true class

[0060] Regularization and early stopping: Early stopping strategy: Monitor the validation set loss; if there is no decrease after 8 consecutive epochs, stop training, restore the optimal weights, and avoid overfitting. Learning rate decay: When the accuracy of the validation set does not improve for 4 consecutive epochs, the learning rate is multiplied by 0.9 to refine the search for the optimal parameters.

[0061] Specifically, a small-sample exhaled breath detection system based on a channel attention mechanism includes: Data acquisition module: used to acquire exhaled air data and preprocess it to obtain exhalation timing data; Feature extraction module: used to extract local key features from expiratory time-series data to obtain local dynamic features; Channel attention module: used to adaptively enhance local dynamic features to obtain enhanced features; Output module: Used to compress and abstract the enhanced features and output the disease classification probability.

[0062] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0063] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A small-sample exhaled breath detection method based on channel attention mechanism, characterized in that, include: Step 1: Acquire exhaled air data and preprocess it to obtain exhalation timing data; Step 2: Construct a 1D-CNN feature extraction module to extract local key features from the exhalation time series data and obtain local dynamic features; Step 3: Adaptively enhance local dynamic features using the channel attention module to obtain enhanced features; Step 4: Perform feature compression and abstraction on the enhanced features and output the disease classification probability.

2. The method for detecting small-sample exhaled air based on channel attention mechanism according to claim 1, characterized in that, The preprocessing process in step 1 includes baseline subtraction, physiological range clipping, and channel-level standardization. The resulting expiratory time series data is a tensor of batch size B × time step T × number of channels C.

3. The method for detecting small-sample exhaled air based on channel attention mechanism according to claim 1, characterized in that, The 1D-CNN feature extraction module is used to extract disease-related local dynamic features from standardized expiratory time-series data, including two cascaded convolution-pooling-batch normalization sub-blocks.

4. The method for detecting small-sample exhaled air based on channel attention mechanism according to claim 1, characterized in that, The channel attention module includes: a global information aggregation layer, a shared fully connected layer, an attention weight generation layer, and a feature weighting layer; The global information aggregation layer is used to receive local dynamic features and aggregate global information of each virtual channel from different dimensions through global average pooling (GAP) and global max pooling (GMP). The shared fully connected layer feeds the two feature vectors generated by global average pooling (GAP) and global max pooling (GMP) into a shared two-layer fully connected network, and strengthens the virtual channel features that are strongly correlated with the disease through compression-restore operations. The attention weight generation layer performs feature fusion by adding the two feature vectors after passing through the shared fully connected layer element by element, and then using the sigmoid activation function to map the fused feature vector to the [0,1] interval to generate channel attention weights. The feature weighting layer strengthens key channels and suppresses redundant channels based on channel attention weights.

5. The method for detecting small-sample exhaled air based on channel attention mechanism according to claim 1, characterized in that, Step 4 utilizes the classification module to transform the enhanced features output by the channel attention module into disease classification probabilities through a flattening-fully connected-classification process, specifically including: The flattening layer flattens the enhanced features into a one-dimensional vector, transforming the temporal-channel structured features into one-dimensional features that can be processed by the fully connected layer. Fully connected layers extract abstract features across time and channels from one-dimensional features; The output layer sets the number of neurons and outputs a probability vector based on abstract features.

6. The method for detecting small-sample exhaled air based on channel attention mechanism according to claim 1, characterized in that, Also includes: We employ a few-shot adaptation training method, using the Adam optimizer, sparse classification cross-entropy loss function, and combining early stopping and learning rate decay strategies to train the model, ensuring the convergence and generalization ability of the model in small-shot scenarios.

7. A small-sample exhaled breath detection system based on channel attention mechanism, characterized in that, include: Data acquisition module: used to acquire exhaled air data and preprocess it to obtain exhalation timing data; Feature extraction module: used to extract local key features from expiratory time-series data to obtain local dynamic features; Channel attention module: used to adaptively enhance local dynamic features to obtain enhanced features; Output module: Used to compress and abstract the enhanced features and output the disease classification probability.

Images