Human body exhaled air detection method and device based on deep learning and semiconductor sensor array

Abstract

The invention provides a human body expired gas detection method and device based on deep learning and a semiconductor sensor array, and belongs to the technical field of gas detection. The problem that an existing semiconductor gas sensor is low in mixed gas recognition accuracy is solved. Comprising the following steps: acquiring expired gas R-t response curve data of a patient as sample data; constructing mixed gas, and obtaining and storing a mixed gas R-t response curve group; the collected patient exhaled gas R-t response curve data and the mixed gas R-t response curve set are preprocessed, and a patient exhaled gas gas response curve data set is constructed through the preprocessed patient exhaled gas R-t response curve set; constructing an expired gas disease marker fitting data set by utilizing the preprocessed mixed gas R-t response curve group; fitting a data set by using the exhaled air disease markers to construct an evaluation system; constructing a recognition model, and performing learning training on the recognition model by using the data set; the method is applied to human body exhaled air detection.

Description

Technical Field

[0001] The present application relates to the field of gas detection technology, and in particular to a method and device for detecting human exhaled breath based on deep learning and a semiconductor sensor array. Background Art

[0002] Exhaled breath contains volatile organic compounds (VOCs), highly volatile gaseous organic molecules. VOCs are derived from a variety of metabolic pathways in the body and are released through the skin, feces, urine, and breath. Because disease alters cellular metabolism, changes in VOCs can serve as biomarkers for specific pathophysiological conditions. In infectious diseases, pathogenic microorganisms, after infecting the body, metabolize these substances to produce a wide range of VOCs, including hydrocarbons, alcohols, ketones, and nitrogen- and sulfur-containing compounds.

[0003] Currently, the use of VOCs as biomarkers for early disease detection has received widespread attention. For example, in respiratory diseases, exhaled breath is of particular interest due to its close contact with the respiratory tract. In recent years, progress has been made in the diagnosis of both non-infectious diseases (such as bronchial asthma, lung cancer, breast cancer, and colorectal cancer) and infectious diseases (such as tuberculosis) using exhaled breath testing.

[0004] A semiconductor gas sensor is a sensor based on the electrical properties of semiconductor materials and the interaction between gas and semiconductor surfaces. It detects the presence and concentration of gas by using changes in physical properties such as conductivity produced when the gas to be measured comes into contact with the surface of a semiconductor (primarily metal oxide). Since the conductivity of the semiconductor material changes when in contact with the gas to be measured, the resistance of the semiconductor material also changes, thus forming the resistance-time (Rt) response curve of the semiconductor material. Semiconductor gas sensors have been widely used in the field of gas detection due to their high sensitivity, portability, and low price. Therefore, the use of semiconductor gas sensors to assess the types of exhaled breath diseases has great potential. However, in actual application environments, semiconductor gas sensors still have the problem of low accuracy in identifying mixed gases. Summary of the Invention

[0005] In order to solve the above technical problems, this application proposes a method and device for detecting human exhaled breath based on deep learning and semiconductor sensor arrays.

[0006] The technical solution adopted in this application is: a method for detecting human exhaled breath based on deep learning and a semiconductor sensor array, comprising the following steps:

[0007] Step 1: Acquire the patient's exhaled breath Rt response curve data as sample data; wherein the semiconductor gas sensor array uses a variety of different gas sensors;

[0008] Step 2: Use different concentrations of standard NH3, TMA, and NO x The gas mixture is constructed, and the mixture Rt response curve group is obtained and stored;

[0009] Step 3: Data preprocessing and dataset construction: Preprocess the patient exhaled breath Rt response curve data and the mixed gas Rt response curve group collected in steps 1 and 2, and use the preprocessed patient exhaled breath Rt response curve group to construct a patient exhaled breath gas response curve dataset, and use the preprocessed mixed gas Rt response curve group to construct an exhaled breath disease marker fitting dataset;

[0010] Step 4: Construct an evaluation system using the exhaled breath disease marker fitting dataset;

[0011] Step 5: Build a recognition model and use the dataset to train the recognition model:

[0012] Step 6: Use the trained model to analyze the patient's exhaled breath data and output the gas type judgment result.

[0013] In both steps 1 and 2, an acquisition system based on a semiconductor gas sensor array is used to obtain the exhaled gas Rt response curve data and the mixed gas Rt response curve data. The acquisition system includes a closed gas chamber, an air pump device and an electronic device, wherein a multi-channel semiconductor sensor array is provided in the closed gas chamber to provide the patient's exhaled gas to react with the multi-channel semiconductor sensor array; the air pump device is used to draw the gas to be tested into the closed gas chamber; the electronic device includes an acquisition device, an analog-to-digital converter, a processor and a display device.

[0014] The sensing units in the multi-channel semiconductor sensor array use a variety of different metal oxide semiconductor nanomaterials.

[0015] Step 3 also includes creating a one-hot dataset of disease labels corresponding to the disease conditions corresponding to the patient's exhaled gas response curve dataset; the labels include: asthma, chronic obstructive pulmonary disease, pulmonary fibrosis, chronic kidney disease, cirrhosis, and health. Subsequent model training uses the patient's exhaled gas Rt response curve group as input and the corresponding disease label one-hot label as model output.

[0016] In step 4, the patient's exhaled breath gas response curve dataset is compared with the exhaled breath disease marker fitting dataset to obtain the exhaled breath components of different diseases, and the disease type-exhaled breath disease marker-semiconductor sensor array responsivity relationship is established to evaluate the accuracy of the recognition model training.

[0017] The recognition model is a prediction model based on a one-dimensional convolutional neural network and a multi-layer perceptron using an attention mechanism. It includes a convolutional neural sub-network with three convolutional layers and a multi-layer perceptron sub-network using an attention mechanism. The one-dimensional convolutional neural sub-network consists of three convolutional layers and two fully connected layers. Each convolutional layer is connected to a maximum pooling layer. The filter sizes of the three convolutional layers are 5, 4, and 3 respectively. Each convolutional layer is configured with 64 convolution kernels, a step size of 1, and a padding of 1.

[0018] The cross entropy loss function is used as the loss function for model training, and the Adam algorithm is used to achieve iterative optimization of model parameters.

[0019] A human exhaled breath detection device based on a semiconductor sensor array using deep learning, comprising:

[0020] The first acquisition module uses an acquisition system based on a semiconductor gas sensor array to collect exhaled gas Rt response curve data of different patients to form a patient exhaled gas response curve data set;

[0021] The second acquisition module uses different concentrations of standard NH3, TMA, and NO x The gas mixture is constructed, and the Rt response curve data of the mixed gas with different components are collected using an acquisition system based on a semiconductor gas sensor array to form an exhaled breath disease marker fitting data set;

[0022] The third acquisition module pre-processes the collected data and creates a disease type label dataset corresponding to the patient's exhaled gas response curve dataset to form a training dataset;

[0023] The evaluation system module compares the patient's exhaled breath gas response curve dataset with the exhaled breath disease marker fitting dataset to obtain the exhaled breath markers of different diseases. It then establishes the disease type-exhaled breath disease marker-semiconductor sensor array responsivity relationship to evaluate the accuracy of the recognition model training.

[0024] The training module uses the training data set constructed above to train the constructed recognition model;

[0025] The evaluation module compares the prediction results obtained from the model analysis with the disease type-exhaled breath disease marker-semiconductor sensor array responsiveness evaluation system constructed by standard gas fitting to verify the accuracy of the model.

[0026] A computer device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0027] A computer-readable storage medium stores a computer program / instruction thereon, which implements the steps of the method when executed by a processor.

[0028] The beneficial effects of the present application compared to the prior art are as follows: different semiconductor sensors will produce different responses after coming into contact with human exhaled breath due to the differences in gas sensitivity of different sensors. Therefore, the present application can effectively solve the problem of low accuracy of mixed gas detection by a single sensor by constructing a sensor array to obtain a multi-channel signal of the gas to be tested. In the detection of exhaled breath diseases, a mixed gas Rt response curve group is obtained by designing a multi-channel semiconductor gas sensor array, and a corresponding data set is constructed, and a corresponding fingerprint spectrum of the patient's exhaled breath is established at the same time. Since the collected fingerprint spectrum cannot intuitively reflect the disease condition, the detection of disease types is achieved by developing a deep learning model applied to human exhaled breath disease detection and learning the collected sensor array data. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present application will be further described below with reference to the accompanying drawings:

[0030] Figure 1 A flowchart of a method for detecting human exhaled breath diseases based on deep learning provided in an embodiment of the present application;

[0031] Figure 2 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present disclosure is shown;

[0032] Figure 3 A flow chart of human exhaled breath detection provided in an embodiment of the present application;

[0033] Figure 4 A schematic diagram of the deep learning model used in the embodiments of the present application;

[0034] Figure 5 Shown are test results based on exhaled breath from different patients. DETAILED DESCRIPTION

[0035] like Figures 1 to 5As shown, the present application provides an acquisition system based on a semiconductor gas sensor array, comprising a closed gas chamber, an air pump device, and an electronic device. The closed gas chamber is provided with a multi-channel semiconductor sensor array for providing a patient's exhaled breath to react with the multi-channel semiconductor sensor array; the air pump device is used to draw the gas to be measured into the closed gas chamber; the electronic device includes an acquisition device, an analog-to-digital converter, a processor, and a display device. The acquisition device is a resistance acquisition device that uses a voltage divider ratio measurement method and a multi-channel reference resistor to achieve proportional resistance measurement. A voltage follower and a 12-bit ADC are used to acquire voltage signals, achieving an ultra-wide acquisition range from Ω to GΩ while maintaining acquisition accuracy, and realizing resistance measurement of semiconductor gas sensors with resistance values of MΩ and above. The multi-channel semiconductor sensor array reacts with the patient's exhaled breath, and its resistance value changes accordingly. The acquisition device converts the resistance into a voltage, then performs an analog-to-digital conversion to obtain the corresponding resistance value, and sends the resistance value to the processor. The processor is configured to read a computer program in a storage unit and perform the following operations: controlling the acquisition device to obtain a patient's exhaled breath gas response curve dataset, preprocessing and storing the dataset.

[0036] The closed gas chamber is made of polytetrafluoroethylene to reduce the interference and adsorption of the gas to be measured.

[0037] The air pump assembly, which includes an air pump, nitrogen cylinder, and associated air circuitry, is used to pump the gas to be tested into a sealed chamber. Nitrogen is also used to purge the sealed chamber before and after testing. The associated air circuitry includes a connection for a mouthpiece.

[0038] The present application also provides a method for detecting human exhaled breath using a semiconductor sensor array based on deep learning, comprising the following steps:

[0039] Step 1: Use Figure 2 The semiconductor gas sensor array-based acquisition system captures exhaled breath Rt response curve data as sample data. This embodiment uses metal oxide semiconductor nanomaterials as sensor elements. Six sensor types are provided: a ZnO sensor, a ZnO / rGO (ZnO-loaded graphene oxide) sensor, an In2O3 sensor, an In2O3 / rGO (In2O3-loaded graphene oxide) sensor, and a CuO sensor, a CuO / rGO (CuO-loaded graphene oxide) sensor.

[0040] See also Figure 3The patients to be tested will take the breath test on an empty stomach in the morning. After sitting quietly for ten minutes, the subjects will rinse their mouths with room temperature water in a unified manner and then take breath samples. Before sampling, the closed air chamber is cleaned with nitrogen to remove residual gas and air from the previous test. The patient to be tested blows into the air path through a disposable mouthpiece. Each blowing should not be less than six seconds. The patient's exhaled air is then pumped into the closed air chamber for testing through an air pump. The acquisition device is used to collect the Rt response curve group of the semiconductor sensor array for the current gas to be tested and store it. This is used to subsequently construct a patient's exhaled gas response curve dataset.

[0041] Before testing and after completing a single response, the closed gas chamber was purged with pure nitrogen to remove residual gas and air.

[0042] Step 2: Use different concentrations of standard NH3, TMA, and NO x To construct a gas mixture, a gas pump is used to pump the test gas mixture into a closed gas chamber for testing, with a single pumping time of 6 seconds. An acquisition device collects and stores the Rt response curves of the semiconductor sensor array for the current test gas. This data is then used to construct a fitting dataset for exhaled breath disease markers.

[0043] Before testing and after completing a single response, the closed gas chamber was purged with pure nitrogen to remove residual gas and air.

[0044] Step 3: Data preprocessing and dataset construction:

[0045] The collected Rt response curve data of the patient's exhaled breath and mixed gas were preprocessed, using a limiting filter algorithm and a Kalman filter to remove outliers and smooth the data. After removing outliers from the dataset, a model training dataset was constructed.

[0046] The patient exhaled gas response curve dataset is constructed using the pre-processed patient exhaled breath Rt response curve group, and the exhaled breath disease marker fitting dataset is constructed using the pre-processed mixed gas Rt response curve group.

[0047] A preprocessed patient exhaled breath gas response curve dataset was used to construct a fingerprint of the patient's exhaled breath. Simultaneously, a preprocessed exhaled breath disease marker fitting dataset was used to construct a disease type-exhaled breath disease marker-semiconductor sensor array responsivity relationship. This was used to reflect the composition of different exhaled breath disease markers and to verify the accuracy of the subsequently constructed model.

[0048] A one-hot dataset of disease labels was also created based on the patient's exhaled breath response curve dataset. Labels included: asthma, COPD, pulmonary fibrosis, chronic kidney disease, cirrhosis, and healthy. Subsequent model training used the patient's exhaled breath Rt response curve dataset as input and the corresponding one-hot disease labels as output.

[0049] Step 4: Use the exhaled breath disease marker fitting dataset to build an evaluation system:

[0050] By comparing a patient's exhaled breath gas response curve dataset with a fitted dataset of exhaled breath disease markers, we can determine the composition of exhaled breath for different diseases. We then establish a relationship between disease type, exhaled breath disease marker, and semiconductor sensor array responsivity. This is used to evaluate the accuracy of recognition model training.

[0051] Step 5: Build a recognition model and use the dataset to train the recognition model:

[0052] The recognition model constructed in the embodiment of the present application is a prediction model based on a one-dimensional convolutional neural network and a multi-layer perceptron using an attention mechanism. The model is trained using the training data set constructed in step 2. The model includes a one-dimensional convolutional neural network sub-network with three convolutional layers and a multi-layer perceptron sub-network using an attention mechanism. The schematic diagram of the model structure is shown in FIG. Figure 4 .

[0053] The recognition model is trained using a training data set to obtain an effective recognition model. The embodiment of the present application uses Pytorch to build a deep learning model, and continuously optimizes the model by adjusting parameters including the learning rate and the number of iterations so that the model can converge and achieve target recognition.

[0054] The one-dimensional convolutional network used in the embodiment of the present application can effectively capture low- and medium-frequency features and local data in sequence data. The multilayer perceptron network based on the attention mechanism used in the embodiment of the present application can focus more on key information when extracting features. The neural network constructed in this embodiment fuses the local information extracted by the convolutional network with the global key feature information extracted by the multilayer perceptron to obtain the prediction result.

[0055] The model training process is described in detail below.

[0056] Step 5.1: Input the training data into the initialized model. The embodiment of the present application uses the semiconductor sensor array Rt response curve group collected by the acquisition device and pre-processed as the model input. In the embodiment of the present application, the fully connected layer in the model uses the standard random initialization method to initialize the parameters, so that the initialization parameters have a mean of 0 and obey a random distribution. The CNN network uses the normalized Glorot initialization method to initialize the parameters, so that the model initialization parameters have a mean of 0 and a standard deviation of Among them fan in and fan out is the number of input and output units of the weight tensor:

[0057] fan in =channels in *kernels width *kernels height ;

[0058] fan out =channels out *kernels width *kernels height ;

[0059] Among them, channels in is the number of input channels, channels out is the number of output channels, kernels width is the width of the convolution kernel, kernels height is the height of the convolution kernel.

[0060] Step 5.2.1: The one-dimensional convolutional neural network consists of three convolutional layers and two fully connected layers, each followed by a max pooling layer. The filter sizes of the three convolutional layers are 5, 4, and 3, respectively. Each convolutional layer is configured with 64 convolution kernels, a stride of 1, and padding of 1.

[0061] Step 5.2.2: Use the attention mechanism in the multilayer perceptron subnetwork to process the Rt response curve of the semiconductor sensor array. The attention mechanism enables the multilayer perceptron network to focus more on information that is more critical to the task objectives when extracting features, helping the network to more accurately capture key features in the input data and enhance the model's expressiveness and generalization capabilities.

[0062] Step 5.3: Fuse the two sub-network results. The model constructed in this embodiment uses weighted connections. The fused result is passed through two fully connected layers to obtain the result. Leaky ReLU is used as the activation function, which is defined as follows:

[0063]

[0064] Where α is the slope of the negative region and is a small positive hyperparameter. In this embodiment, α = 0.01. The model constructed in this application outputs a 6×1 matrix label, where label[i] represents the predicted probability of the i-th disease condition. Softmax is used as the activation function of the model output layer so that the output satisfies 0≤label[i]≤1 and The highest probability (N = max(Label[i]), i = 0, 1, 2, 3, 4, 5) is selected as the final predicted disease probability Output (asthma, COPD, pulmonary fibrosis, chronic kidney disease, cirrhosis, healthy) as the doctor's reference data.

[0065] Step 5.4: Continuously iterate to obtain the optimized model, and use the cross entropy loss function (Cross EntropyLoss) as the loss function for model training. The cross entropy function is specifically defined as follows:

[0066] H(P,Q)=-∑ i P(i)log(Q(i));

[0067] Where P is the one-hot label of the actual disease condition, and Q represents the predicted label result output by the model. In this embodiment, the cross entropy loss function formula is used:

[0068] L(p,q)=-∑ x (p(x)logq(x)+(1-p(x))log(1-q(x))).

[0069] Among them, p is the one-hot label of the actual disease condition, and q is the predicted label result output by the model.

[0070] At the same time, the prediction results obtained by model analysis are compared with the evaluation system constructed by standard gas fitting to obtain the accuracy evaluation of the model.

[0071] The hyperparameters for model training in this embodiment include a batch size of 64, a learning rate (lr) of 1e-3, and a maximum number of epochs of 300.

[0072] The embodiment of this application uses the Adam algorithm to achieve iterative optimization of model parameters. Let θ be the parameter to be updated, L(θ) be the cross entropy loss function used by the model, g be the gradient of L(θ), m k ,v kare the first and second moments of g, respectively. β1 and β2 are the exponential decay rates of the moment estimates. In this implementation example, β1 = 0.9 and β2 = 0.999. The Adam update rule can be expressed as:

[0073]

[0074] m k =β1m k-1 +(1-β1)g;

[0075] v k =β2v k-1 +(1-β2)g⊙g;

[0076]

[0077] in, For m k The deviation correction, For v k , η is the initial learning rate, and ε is a small constant to prevent division by zero.

[0078] The Adam algorithm can adaptively adjust the learning rate for each parameter by calculating the first-order and second-order moment estimates. At the same time, since the gradient is low in the initial stage, bias correction can accelerate the initial learning rate.

[0079] Step 6: Use the trained model to analyze the patient's exhaled breath data and determine the type of exhaled gas.

[0080] The present embodiment collects multiple exhaled breath sample data from six patients with different diseases and six healthy people, and uses the present model for training and testing. Figure 5 As shown in the figure, after 300 iterations, the model converged. According to the confusion matrix obtained from the test, the model can effectively identify the types of exhaled breath in different diseases and make judgments. The model's recognition accuracy for the patient's exhaled breath is 95%.

[0081] Based on the above method, the present application also proposes a human exhaled breath detection device based on a semiconductor sensor array based on deep learning, comprising:

[0082] The first acquisition module uses an acquisition system based on a semiconductor gas sensor array to collect exhaled gas Rt response curve data of different patients to form a patient exhaled gas response curve data set.

[0083] The second acquisition module uses standard NH3, TMA, and NOx gases of different concentrations to construct a mixed gas, and uses an acquisition system based on a semiconductor gas sensor array to collect the Rt response curve data of the mixed gas with different components to form an exhaled breath disease marker fitting data set.

[0084] The third acquisition module pre-processes the collected data and creates a disease category label data set corresponding to the patient's exhaled gas response curve data set to form a training data set.

[0085] The evaluation system module compares the patient's exhaled breath gas response curve dataset with the exhaled breath disease marker fitting dataset to determine the exhaled breath marker profiles for different diseases. This module establishes a relationship between disease type, exhaled breath disease marker, and semiconductor sensor array responsivity. This module is used to evaluate the accuracy of recognition model training.

[0086] The training module uses the training data set constructed above to train the constructed recognition model, in order to obtain a deep learning model that can effectively distinguish the types of human exhaled breath corresponding to different diseases.

[0087] The evaluation module compares the prediction results obtained from the model analysis with the disease type-exhaled breath disease marker-semiconductor sensor array responsiveness evaluation system constructed by standard gas fitting to verify the accuracy of the model.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting human exhaled breath based on deep learning and a semiconductor sensor array, characterized by: The following steps are involved: Step 1: Obtain the patient's exhaled air Rt response curve data as sample data; Among them, the semiconductor gas sensor array uses a variety of different gas sensors; Step 2: Use different concentrations of standard NH3, TMA, and NO x The gas mixture is constructed, and the mixture Rt response curve group is obtained and stored; Step 3: Data preprocessing and dataset construction: Preprocess the patient exhaled breath Rt response curve data and the mixed gas Rt response curve group collected in steps 1 and 2, and use the preprocessed patient exhaled breath Rt response curve group to construct a patient exhaled breath gas response curve dataset, and use the preprocessed mixed gas Rt response curve group to construct an exhaled breath disease marker fitting dataset; Step 4: Construct an evaluation system using the exhaled breath disease marker fitting dataset; Step 5: Build a recognition model and use the dataset to train the recognition model: Step 6: Use the trained model to analyze the patient's exhaled breath data and output the gas type judgment result.

2. The method for detecting human exhaled breath based on deep learning and a semiconductor sensor array according to claim 1, wherein: In both steps 1 and 2, an acquisition system based on a semiconductor gas sensor array is used to obtain the exhaled gas Rt response curve data and the mixed gas Rt response curve data. The acquisition system includes a closed gas chamber, an air pump device and an electronic device, wherein a multi-channel semiconductor sensor array is provided in the closed gas chamber to provide the patient's exhaled gas to react with the multi-channel semiconductor sensor array; the air pump device is used to draw the gas to be tested into the closed gas chamber; the electronic device includes an acquisition device, an analog-to-digital converter, a processor and a display device.

3. The method for detecting human exhaled breath based on deep learning and a semiconductor sensor array according to claim 2, wherein: The sensing units in the multi-channel semiconductor sensor array use a variety of different metal oxide semiconductor nanomaterials.

4. The method for detecting human exhaled breath based on deep learning and a semiconductor sensor array according to claim 1, wherein: Step 3 also includes creating a one-hot dataset of disease labels corresponding to the disease conditions corresponding to the patient's exhaled gas response curve dataset; the labels include: asthma, chronic obstructive pulmonary disease, pulmonary fibrosis, chronic kidney disease, cirrhosis, and health. Subsequent model training uses the patient's exhaled gas Rt response curve group as input and the corresponding disease label one-hot label as model output.

5. The method for detecting human exhaled breath based on deep learning and a semiconductor sensor array according to claim 1, wherein: In step 4, the patient's exhaled breath gas response curve dataset is compared with the exhaled breath disease marker fitting dataset to obtain the exhaled breath components of different diseases, and the disease type-exhaled breath disease marker-semiconductor sensor array responsivity relationship is established to evaluate the accuracy of the recognition model training.

6. The method for detecting human exhaled breath based on deep learning and a semiconductor sensor array according to claim 1, wherein: The recognition model is a prediction model based on a one-dimensional convolutional neural network and a multi-layer perceptron using an attention mechanism. It includes a convolutional neural sub-network with three convolutional layers and a multi-layer perceptron sub-network using an attention mechanism. The one-dimensional convolutional neural sub-network consists of three convolutional layers and two fully connected layers. Each convolutional layer is connected to a maximum pooling layer. The filter sizes of the three convolutional layers are 5, 4, and 3 respectively. Each convolutional layer is configured with 64 convolution kernels, a step size of 1, and a padding of 1.

7. The method for detecting human exhaled breath based on deep learning and a semiconductor sensor array according to claim 1, wherein: The cross entropy loss function is used as the loss function for model training, and the Adam algorithm is used to achieve iterative optimization of model parameters.

8. A human exhaled breath detection device using a semiconductor sensor array based on deep learning, characterized in that: include: The first acquisition module uses an acquisition system based on a semiconductor gas sensor array to collect exhaled gas Rt response curve data of different patients to form a patient exhaled gas response curve data set; The second acquisition module uses different concentrations of standard NH3, TMA, and NO x The gas mixture is constructed, and the Rt response curve data of the mixed gas with different components are collected using an acquisition system based on a semiconductor gas sensor array to form an exhaled breath disease marker fitting data set; The third acquisition module pre-processes the collected data and creates a disease type label dataset corresponding to the patient's exhaled gas response curve dataset to form a training dataset; The evaluation system module compares the patient's exhaled breath gas response curve dataset with the exhaled breath disease marker fitting dataset to obtain the exhaled breath markers of different diseases. It then establishes the disease type-exhaled breath disease marker-semiconductor sensor array responsivity relationship to evaluate the accuracy of the recognition model training. The training module uses the training data set constructed above to train the constructed recognition model; The evaluation module compares the prediction results obtained from the model analysis with the disease type-exhaled breath disease marker-semiconductor sensor array responsiveness evaluation system constructed by standard gas fitting to verify the accuracy of the model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images