Gas identification method and system based on physical-data dual-drive framework

Abstract

The invention relates to the technical field of artificial olfaction and intelligent sensing, and discloses a gas recognition method and system based on a physical-data dual-drive framework. The method comprises the following steps: acquiring a macroscopic dynamic sensing signal triggered by target gas on a sensor array containing atomic-scale catalytic active sites, and a microscopic physical descriptor for representing an atomic-scale electron interaction rule between gas molecules and the catalytic active sites; and the data branch network and the physical branch network are respectively input into the data branch network and the physical branch network for feature extraction and mapping, cross-modal fusion is carried out through a feature fusion sub-module to obtain joint feature representation constrained by a physical rule, and accordingly, a gas category identification result is output. According to the method, the microcosmic physical descriptors are fused into the model reasoning logic, so that the recognition precision of the structural similar object is effectively improved.

Description

Technical Field

[0001] This application relates to the fields of artificial olfaction and intelligent sensing technology, and in particular to gas recognition technology. Background Technology

[0002] The accurate identification of volatile organic compounds (VOCs) has significant application value in non-invasive medical diagnostics, environmental monitoring, and industrial safety. For example, changes in the concentration of specific aldehyde markers in human exhaled breath are associated with the early occurrence of diseases such as lung cancer. Highly sensitive detection of exhaled breath components holds promise for early, non-invasive screening of these diseases. Artificial olfactory systems (commonly known as "electronic noses") are intelligent detection devices that mimic the biological olfactory perception mechanism. Their basic working principle involves using a sensor array to capture the chemical information of gas molecules and convert it into electrical signals. Subsequently, pattern recognition algorithms are used to extract and classify features from these electrical signals to achieve automated identification of target gas types.

[0003] At the level of gas-sensitive sensing materials, the mainstream sensing interface materials currently include metal oxide semiconductors and carbon-based nanomaterials. When gas molecules are adsorbed on the surface of the sensing material, electron transfer occurs between the gas molecules and the material, causing a change in the carrier concentration of the material. Macroscopically, this manifests as a change in the material's resistance value, i.e., generating a resistance response signal. However, in practical applications, the gases to be detected often contain a large number of structurally similar homologues, such as methanol and ethanol, propionaldehyde and butyraldehyde, etc., and the physicochemical properties of these homologues are extremely similar. Traditional sensor materials lack precise recognition sites for specific functional groups, and when faced with such structurally similar substances, their resistance response curves are highly similar with extremely weak feature differences, making it difficult for downstream classification algorithms to effectively distinguish them, thus forming a performance bottleneck in the selectivity of artificial olfactory systems.

[0004] At the algorithmic level, existing intelligent gas identification solutions primarily employ purely data-driven deep learning models. These solutions typically use resistance-time response curves acquired by sensor arrays as input, employing deep neural networks such as 1-Dimensional Convolutional Neural Networks (1D-CNN) or Long Short-Term Memory (LSTM) networks to extract features and classify the raw signals. However, these purely data-driven models treat sensor signals merely as abstract time series, classifying them by learning the mathematical statistical fluctuations of the signal curves. They do not involve understanding the physicochemical mechanisms behind the gas-sensitive response, such as electron transfer and orbital hybridization. Their decision-making process is essentially a "black box" operation, lacking physical interpretability. Furthermore, these models heavily rely on the coverage of labels in the training set. When novel chemical homologues outside the training set appear in the detection environment, the model often produces incorrect predictions because it has not learned the relevant features; that is, the model's extrapolation and generalization capabilities are severely insufficient. Summary of the Invention

[0005] This application provides a gas identification method and system based on a physics-data dual-driven framework, which solves the technical problem of insufficient identification accuracy for structural homologues with extremely similar physicochemical properties.

[0006] This application discloses a gas identification method based on a physical-data dual-driven framework, including: Acquire macroscopic dynamic sensing signals triggered by the target gas on a sensor array, wherein the sensing interface of the sensor array includes a conductive carrier and catalytic active sites dispersed on the conductive carrier at the atomic or sub-nanometer level. Obtain a microphysical descriptor corresponding to a preset set of candidate gases. The microphysical descriptor is used to characterize the atomic-level electronic interaction between gas molecules in the set of candidate gases and the catalytic active site. The macroscopic dynamic sensing signal is input into a preset data branch network for feature extraction to obtain macroscopic sensing features; The microscopic physical descriptor is input into a preset physical branch network for feature mapping to obtain microscopic physical embedding features; The macroscopic sensing features and the microscopic physical embedded features are fused across modes by the feature fusion submodule to obtain a joint feature representation constrained by the atomic-level electron interaction law; Based on the joint feature representation, classification calculations are performed, and the category identification result of the target gas is output.

[0007] In a preferred embodiment, the microphysical descriptor is a quantized parameter obtained through first-principles calculations of quantum mechanics, molecular dynamics simulations, or semi-empirical methods of quantum chemistry; The microscopic physical descriptor characterizing the atomic-level electronic interaction law includes at least one of the following microscopic parameters: adsorption energy, charge transfer amount, Fermi level shift, system band gap change, d-band center eigenvalue, energy level gap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital, dipole moment, molecular polarizability, and ionization energy.

[0008] In a preferred embodiment, the data branch network includes at least one of a one-dimensional convolutional neural network, a long short-term memory network, a gated recurrent unit, an attention mechanism network, or a temporal convolutional network; The physical branch network includes a multilayer perceptron network; The cross-modal feature fusion of the macroscopic sensing features and the microscopic physical embedding features through the feature fusion submodule includes: fusing the macroscopic sensing features and the microscopic physical embedding features through splicing operators, element-level multiplication operators, weighted summation mechanisms, or attention-based cross-alignment operations; wherein, when the dimensions of the macroscopic sensing features and the microscopic physical embedding features are inconsistent, they are mapped to a unified feature space dimension through linear transformation layers before fusion.

[0009] In a preferred embodiment, the method supports generalized identification of unknown homologues to be tested, and the method further includes: During the model training or initialization phase, the physical branch network is used to learn and extract the common physical laws and distribution mapping intervals of known homologs belonging to the same homologous series on the corresponding microscopic physical descriptors. Members of the same homologous series share the same core chemical functional groups. The output of the target gas category identification result specifically includes: when the target gas is an unknown homologue not included in the model training set, based on the structural characteristics that the unknown homologue belongs to the same homologue series and has the same core chemical functional group as the known homologue, the distribution mapping interval in the feature space where the microscopic physical embedding feature corresponding to the unknown homologue falls, and the output of the generalized category result of the unknown homologue through logical transfer inference.

[0010] In a preferred embodiment, the conductive carrier is a material with a high specific surface area; The catalytic active sites are composed of transition metals or noble metals and are discretely distributed in the form of single atoms or ultrafine nanoclusters with a particle size of less than 2 nanometers. They are used to provide a uniform atomic-level coordination environment for the adsorption of chemical gases and to convert the microscopic differences of the functional groups of the target gas into identifiable electrical signal characteristics.

[0011] In a preferred embodiment, the conductive support is reduced graphene oxide, and the catalytically active site is a single palladium atom anchored in a defect site on the surface of the reduced graphene oxide. The palladium single atom generates a specific electronic fingerprint signal by undergoing differentiated orbital hybridization with functional groups of different types of target gases through its unsaturated d orbitals.

[0012] In a preferred embodiment, after outputting the category identification result of the target gas, the following step is further included: An attribution analysis algorithm based on Shapley additive interpretation is introduced to calculate the contribution weight of each input microphysical descriptor for the final category identification result. Based on the ranking analysis of the contribution weights, it is verified whether the classification decision logic of the joint feature representation is dominated by the real microscopic electronic interaction law, so as to eliminate the interference of environmental noise and signal artifacts.

[0013] In a preferred embodiment, after acquiring the macroscopic dynamic sensing signal triggered by the target gas on the sensor array and before inputting the macroscopic dynamic sensing signal to a preset data branch network, a signal preprocessing step is further included: The time series data of the acquired macroscopic dynamic sensing signals are extracted using the sliding window technique; Extract multidimensional statistical features from each captured signal window, wherein the multidimensional statistical features include at least one of the signal sequence mean, standard deviation, peak-to-peak value, skewness, or kurtosis; The multidimensional statistical features are input into the data branch network as a representation of the macroscopic dynamic sensing signal.

[0014] This application also discloses a gas identification system based on a physical-data dual-drive framework, including: The sensing signal acquisition module is used to acquire the macroscopic dynamic sensing signal triggered by the target gas on the sensor array. The sensing interface of the sensor array includes a conductive carrier and catalytic active sites dispersed on the conductive carrier at the atomic or sub-nanometer level. The microscopic feature retrieval module is used to obtain microscopic physical descriptors corresponding to a preset set of candidate gases. The microscopic physical descriptors are used to characterize the atomic-level electronic interaction rules between gas molecules in the set of candidate gases and the catalytic active sites. A dual-drive deep learning computing module is connected to the sensing signal acquisition module and the microscopic feature retrieval module, respectively. Internally, it includes a data branch network, a physical branch network, and a feature fusion submodule. The data branch network extracts the macroscopic dynamic sensing signals to obtain macroscopic sensing features; the physical branch network maps the microscopic physical descriptors to obtain microscopic physical embedded features; and the feature fusion submodule performs cross-modal feature fusion of the macroscopic sensing features and the microscopic physical embedded features to obtain a joint feature representation constrained by the atomic-level electron interaction laws. The classification output module is used to perform classification calculations based on the joint feature representation and output the category identification result of the target gas.

[0015] This application also discloses a non-volatile computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the gas identification method as described above.

[0016] In the embodiments of this application, a physical-data dual-driven gas identification framework is constructed. The macroscopic dynamic sensing signal triggered by the target gas on a sensor array containing atomic or sub-nanometer catalytic active sites, and the microscopic physical descriptor characterizing the atomic-level electronic interaction between gas molecules and catalytic active sites are respectively input into the data branch network and the physical branch network for feature extraction and mapping. Then, through cross-modal feature fusion, a joint feature representation constrained by the atomic-level electronic interaction law is obtained, and classification output is performed accordingly. This allows the decision logic of the gas identification model to be strictly constrained by the microscopic physical law, fundamentally overcoming the "black box" defect of pure data-driven models that only learn the mathematical fluctuations of signals without understanding the chemical essence. This significantly improves the classification accuracy of homologous structures with extremely similar physicochemical properties. At the same time, it realizes the real-time correlation and synergistic driving of microscopic electronic structure parameters and macroscopic resistance response curves, greatly enhancing the robustness, interpretability, and intelligence of the detection system in complex chemical environments.

[0017] Furthermore, by expanding the acquisition methods of microscopic physical descriptors to first-principles calculations in quantum mechanics, molecular dynamics simulations, or semi-empirical methods in quantum chemistry, and by expanding the types of descriptors to include various parameters such as adsorption energy, charge transfer, Fermi level shift, band gap change, d-band center eigenvalues, HOMO-LUMO level gap, dipole moment, molecular polarizability, and ionization energy, the electronic interaction characteristics between gas molecules and active sites can be comprehensively characterized from multiple physical dimensions. This enhances the richness and completeness of prior physical knowledge, thereby strengthening the model's ability to identify subtle differences among different types of gas molecules.

[0018] Furthermore, by setting the data branch network as at least one of a one-dimensional convolutional neural network, a long short-term memory network, a gated recurrent unit, an attention mechanism network, or a temporal convolutional network, and setting the physical branch network as a multilayer perceptron network, and using splicing operators, element-wise multiplication operators, weighted summation mechanisms, or attention-based cross-alignment operations for cross-modal feature fusion, optimal processing paths adapted to the data characteristics of signals from different modalities can be provided. Through flexible fusion strategies, heterogeneous features can be effectively integrated in a unified feature space, improving the applicability of the dual-branch network architecture and the quality of feature fusion.

[0019] Furthermore, by learning the common physical laws and distribution mapping intervals of known homologs on microscopic physical descriptors through a physical branch network during the training phase, and by performing logical transfer inference based on the distribution mapping intervals in the feature space where the specific functional groups shared by unknown and known homologs and their microscopic physical embedding features fall, the limitations of traditional models that heavily rely on the coverage of training labels can be overcome, giving the system a reliable generalization recognition capability for unknown homologs outside the training set.

[0020] Furthermore, by setting the conductive support as a material with a high specific surface area and setting the catalytic active sites as transition metals or noble metals discretely distributed in the form of single atoms or ultrafine nanoclusters with a particle size of less than 2 nanometers, a uniform atomic-level coordination environment can be provided for gas adsorption, and the microscopic differences of different gas functional groups can be efficiently converted into identifiable electrical signal features, thereby improving the selectivity for structurally similar objects at the sensing interface level.

[0021] Furthermore, by using reduced graphene oxide as a conductive carrier and anchoring palladium single atoms in its surface defect sites, specific electronic fingerprint signals are generated by utilizing the differential orbital hybridization between the unsaturated d orbitals of palladium single atoms and functional groups of different target gases. This allows for precise differentiation of the electron cloud distribution differences of different types of gas molecules at the atomic scale, providing highly recognizable source signals for subsequent deep learning models.

[0022] Furthermore, by introducing an attribution analysis algorithm based on Shapley additive interpretation after the output category recognition results, the contribution weight of each microscopic physical descriptor to the classification decision is calculated and ranked for verification. This makes the model's decision-making process transparent, confirming that the classification logic is indeed dominated by the real laws of microscopic electronic interactions rather than by environmental noise or signal artifacts, thereby improving the credibility of the detection results in high-reliability application scenarios.

[0023] Furthermore, by using the sliding window technique to extract macroscopic dynamic sensing signals and extracting multidimensional statistical features such as mean, standard deviation, peak-to-peak value, skewness, or kurtosis from each window as input representations for the data branch network, the dynamic fluctuation characteristics of the sensing signals can be effectively captured and the influence of random noise in the original time series can be reduced, thereby improving the feature extraction quality of the data branch network for macroscopic sensing signals. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of a gas identification method based on a physical-data dual-drive framework according to an embodiment of this application; Figure 2 This is a schematic diagram of a gas identification system based on a physical-data dual-drive framework according to an embodiment of this application; Figure 3 The experimental results of the embodiments of this application are presented from five dimensions, including: sub-figure (a) is a feature correlation matrix; sub-figure (b) is a feature-category Sankey diagram; sub-figure (c) is a global feature importance ranking histogram; sub-figure (d) is a SHAP beehive diagram; and sub-figure (e) is a receiver operating characteristic curve (ROC curve). Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the application will be described in further detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0026] Example 1 This embodiment provides a gas identification method based on a physical-data dual-driven framework. The overall process of this method is as follows: Figure 1 As shown, the main steps include acquiring macroscopic dynamic sensing signals, acquiring microscopic physical descriptors, extracting features from data branch networks, mapping features from physical branch networks, fusing cross-modal features, and performing classification calculations and outputting results. This method constructs a dual-branch heterogeneous neural network architecture to deeply couple macroscopic electrical signals from the sensor array with theoretically calculated microscopic electronic structure parameters, thereby achieving accurate identification of volatile organic compounds. Each step is described in detail below.

[0027] Step 101: Acquire the macroscopic dynamic sensing signal triggered by the target gas on the sensor array. In this application, the "macroscopic dynamic sensing signal" refers to the electrical signal that changes continuously with time at the sensor output terminal when the target gas molecules adsorb onto the sensing interface of the sensor array, causing electron transfer between the gas molecules and the sensing material, resulting in a change in the carrier concentration of the material. A typical manifestation is a resistance-time response curve. "Macroscopic" means that this signal is a device-level electrical output that can be directly acquired and measured, unlike the microscopic physical descriptors describing the atomic-level interactions between molecules and active sites. "Dynamic" means that this signal continuously evolves as the gas adsorption / desorption process proceeds, carrying the temporal process information of the interaction between the gas and the sensing interface. In this embodiment, the sensing interface of the sensor array includes a conductive carrier and catalytic active sites dispersed on the conductive carrier at the atomic or sub-nanometer level. When the target gas molecules contact the sensing interface, chemisorption occurs between the gas molecules and the catalytic active sites, causing a change in the carrier concentration at the sensing interface, which in turn leads to a change in the material resistance. The sensor array monitors the resistance changes of each channel in real time through a constant current source drive circuit, converting the chemical information caused by gas adsorption into electrical signals. Specifically, this embodiment uses a sensor array with eight channels, and the sensing interface of each channel adopts a composite material system with conductive carriers supporting catalytic active sites. During the detection process, the target gas is introduced into the controlled detection chamber, and each channel in the sensor array synchronously acquires the dynamic response curve of resistance changing over time. Let the... Each channel at time The resistance value is Then the macroscopic dynamic sensing signal output by the sensor array can be expressed as: 3D matrix ,in The number of time sampling points within the detection period, the matrix's _____ List Representing the The resistance time series of each channel. It should be noted that the number of channels in the sensor array is not limited to 8. Those skilled in the art can choose 4 to 16 channels or other suitable numbers depending on the complexity of the detection task. The specific material of the conductive carrier can be a material with high specific surface area and good conductivity, such as reduced graphene oxide, carbon nanotubes, MXene, or conductive metal-organic frameworks. The catalytic active sites can be composed of transition metals or noble metals such as palladium, platinum, gold, ruthenium, cobalt, iron, or nickel. Its dispersion form can be a single atom or an ultrafine nanocluster with a particle size of less than 2 nanometers, as long as it can provide a uniform atomic-level coordination environment for the adsorption of gas molecules.

[0028] Step 102: Obtain the microphysical descriptor corresponding to the preset candidate gas set. The microphysical descriptor is used to characterize the atomic-level electronic interaction between each gas molecule in the candidate gas set and the catalytic active site on the sensing interface. In this embodiment, for the candidate gas types covered by the detection target (e.g., various volatile organic compounds including alkanes, alcohols, ketones, aldehydes, esters and their structural homologues), the adsorption behavior of each gas molecule on the sensing interface is simulated by theoretical calculations to extract quantitative parameters that reflect the essence of the electronic interaction between molecules and active sites. Each candidate gas corresponds to a set of... 3D physical descriptor vector ,in This is an index for the candidate gas. In this embodiment... The system uses a 10-dimensional descriptor, meaning each gas molecule corresponds to a 10-dimensional microscopic physical descriptor. These descriptors are calculated and stored in a physical descriptor database during system initialization, and retrieved during actual detection using table lookups or indexes. It should be noted that the specific methods of theoretical calculations and the types of parameters for the descriptors can be flexibly selected according to the needs of the research system. For example, calculation methods can employ density functional theory, molecular dynamics simulations, or semi-empirical quantum chemical methods. Descriptor parameters can include various microscopic quantitative parameters reflecting the essential properties of molecules, such as adsorption energy, charge transfer, Fermi level shift, band gap change, and d-band center eigenvalues. The dimensions are not limited to 10 and can be increased or decreased as needed.

[0029] Step 103: The macroscopic dynamic sensing signal is input into a preset data branch network for feature extraction to obtain macroscopic sensing features. The data branch network is used to extract discriminative high-dimensional feature representations from the dynamic electrical signals collected by the sensor array. In this embodiment, the data branch network adopts a one-dimensional convolutional neural network (1D-CNN) architecture. Specifically, the 1D-CNN contains three convolutional layers stacked sequentially. The first convolutional layer has a kernel size of 5 and 32 output channels, the second convolutional layer has a kernel size of 5 and 64 output channels, and the third convolutional layer has a kernel size of 3 and 128 output channels. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer, defined as... A global average pooling layer is applied after three convolutional layers to compress the variable-length feature map into a fixed-length feature vector. The output of the pooling layer is then mapped to a 128-dimensional macroscopic sensing feature vector through a fully connected layer, denoted as . To prevent overfitting, a dropout layer is placed before the fully connected layer, with a dropout rate of 0.3. It should be noted that the specific type of data branch network is not limited to one-dimensional convolutional neural networks. Those skilled in the art can choose other network architectures suitable for processing time series data, such as long short-term memory networks, gated recurrent units, attention mechanism networks, or temporal convolutional networks, based on the signal characteristics.

[0030] Step 104: The microscopic physical descriptor is input into a preset physical branch network for feature mapping to obtain microscopic physical embedding features. The physical branch network is used to map low-dimensional physical descriptors to a high-dimensional feature space, enabling them to be fused with the macroscopic sensing features extracted from the data branch in a unified feature space. In this embodiment, the physical branch network adopts a multilayer perceptron (MLP) architecture. This MLP contains three fully connected layers with an input dimension of 10 (corresponding to a 10-dimensional physical descriptor), 32 neurons in the first hidden layer, 64 neurons in the second hidden layer, and a 64-dimensional output layer. Each hidden layer is followed by a ReLU activation function and a batch normalization layer. After feature mapping by the physical branch network, the microscopic physical descriptor of each candidate gas is encoded into a 64-dimensional microscopic physical embedding feature vector, denoted as... This embedded feature vector contains a high-level abstract representation of the electronic interactions between gas molecules and catalytically active sites, providing physical constraint information for subsequent cross-modal fusion. It should be noted that the number of layers and neurons in each layer of the MLP can be adjusted according to the dimension and complexity of the descriptor; for example, it can be set to a 2-layer or 4-layer structure, and the hidden layer dimension can be selected between 32 and 128.

[0031] Step 105: The macroscopic sensing features and microscopic physical embedded features are fused across modes using the feature fusion submodule to obtain a joint feature representation constrained by atomic-level electron interaction laws. In this embodiment, the feature fusion submodule uses a splicing operator to achieve cross-modal fusion. Specifically, the 128-dimensional macroscopic sensing feature vector obtained in step 103 is... The 64-dimensional microscopic physical embedding feature vector obtained in step 104 The features are concatenated along the feature dimension to form a 192-dimensional joint feature vector. ,in This represents a vector concatenation operation. The joint feature representation simultaneously carries macroscopic dynamic response information from the sensor array and microscopic electronic structure information from theoretical calculations, thus subjecting subsequent classification decisions to explicit constraints from physical laws. It should be noted that cross-modal fusion methods are not limited to concatenation operations; other fusion strategies such as element-wise multiplication, weighted summation, or attention-based cross-alignment operations can also be used. When the dimensions of the two feature paths are inconsistent, they can be reduced to a unified feature space dimension through linear transformation layers before fusion.

[0032] Step 106: Perform classification calculation based on the joint feature representation and output the category identification result of the target gas. In this embodiment, the joint feature vector... The input is fed into the classifier module. The classifier module consists of two fully connected layers. The first fully connected layer maps the 192-dimensional joint features to a 96-dimensional intermediate representation, and the second fully connected layer maps the 96-dimensional intermediate representation to... The output vector of dimension, where This represents the total number of candidate gas categories. The output vector is then transformed into a probability distribution using the Softmax function. in For the first The logit value corresponds to each category. The category with the highest probability value is taken as the final category identification result, i.e. The model is trained using the cross-entropy loss function, defined as follows: ,in One-hot encoding of the true labels was used. The optimizer employed the Adam algorithm with an initial learning rate of 0.001 and a training duration of 200 epochs. Experimental results showed that, using the aforementioned physical-data dual-drive method, the sensor array achieved a classification accuracy of over 99% for various volatile organic compounds, including alcohols, aldehydes, ketones, esters, and alkanes, representing an improvement of approximately 7 percentage points compared to the single-modal approach using only a data branch network.

[0033] Steps 101 to 106 described above constitute the complete technical solution of this embodiment. Those skilled in the art can implement gas identification based on a physical-data dual-drive framework by following these six steps. Several optional improvements to this solution are described below.

[0034] In some alternative implementations, the microscopic physical descriptor in step 102 can be obtained through calculations based on first principles of quantum mechanics. Specifically, density functional theory can be used to optimize the adsorption configurations and calculate the electronic structures of each candidate gas molecule at the sensing interface. For example, the adsorption energy is defined as... ,in This represents the total energy of the entire system after the adsorption of gas molecules. Energy for cleaning the sensing interface, The adsorption energy represents the energy of an isolated gas molecule. Charge transfer is obtained through Bader charge analysis, characterizing the gain and loss of electrons between the gas molecule and the sensing interface. In addition to adsorption energy and charge transfer, the descriptor may further include at least one of the following parameters reflecting the essential properties of the molecule: Fermi level shift, band gap change, d-band center eigenvalue, energy level gap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital, dipole moment, molecular polarizability, and ionization energy. When density functional theory is not used, the above physical parameters can also be obtained through molecular dynamics simulations, Monte Carlo simulations, or semi-empirical quantum chemical methods; those skilled in the art can flexibly choose according to the requirements of computational accuracy and efficiency.

[0035] As a preferred but not mandatory improvement, the data branch network in step 103, in addition to the one-dimensional convolutional neural network, can be replaced with a Long Short-Term Memory (LSTM) network to better capture the long-range temporal dependence features of the sensing signal, or with a Gated Recurrent Unit (GRU) to reduce computational overhead while maintaining temporal modeling capabilities, or with an attention-based Transformer network to enhance selective attention to key time periods in the signal, or with a Temporal Convolutional Network (TCN) to process temporal signals using causal convolutional structures. In this embodiment, the physical branch network in step 104 uses a multilayer perceptron, a choice suitable for processing low-dimensional and semantically clear physical descriptor data. In step 105, the feature fusion method, besides the concatenation operator, can also employ an element-wise multiplication operator to multiply corresponding elements of the two features point-by-point to capture interactive information, or a weighted summation mechanism to weight and superimpose the two features using learnable weight coefficients, or an attention-based cross-alignment operation to allow the two features to reference each other during the fusion process. When the dimensions of the two features before fusion do not match, they can be reduced to a unified feature space dimension through a linear transformation layer before performing the fusion operation. Even without the above improvement, the technical objective of this method can be achieved by using the basic network architecture and splicing fusion method described in steps 103 to 105 of this embodiment.

[0036] In a variation of this embodiment, the material system of the sensing interface can be further refined. The conductive support can be a material with a high specific surface area, and the catalytic active sites are composed of transition metals or noble metals, discretely distributed on the surface of the conductive support in the form of single atoms or ultrafine nanoclusters with a particle size of less than 2 nanometers. This provides a uniform atomic-level coordination environment for the adsorption of chemical gases and converts the microscopic differences in the functional groups of the target gas into identifiable electrical signal characteristics. Further, as a preferred material combination, the conductive support can be reduced graphene oxide, and the catalytic active sites can be palladium single atoms anchored in defect sites on the surface of reduced graphene oxide. Reduced graphene oxide is obtained by chemically reducing graphene oxide; the density of oxygen-containing functional groups and defect density on its surface can be precisely controlled by adjusting the degree of reduction. Palladium single atoms are anchored in oxygen vacancies on the surface of reduced graphene oxide through low-temperature electrodeposition and coordination-induced methods. The atomic percentage of palladium is controlled between approximately 1% and 2% to ensure high dispersion and high selectivity of the active sites. Due to their unsaturated d-orbital electronic configuration, palladium single atoms can undergo differentiated orbital hybridization with functional groups (such as hydroxyl, carbonyl, and ester groups) of different target gases, generating specific electronic fingerprint signals. This differentiated interaction allows structurally similar homologs to produce highly recognizable response curves on the sensor array, providing a hardware foundation for accurate classification by subsequent algorithms. Of course, in addition to the aforementioned reduced graphene oxide-supported palladium single-atom system, the conductive support can also be replaced with other high specific surface area conductive materials such as carbon nanotubes, micene, or conductive metal-organic frameworks. The catalytically active sites can also be replaced with other catalytically active metal single atoms or ultrafine nanoclusters such as platinum, gold, ruthenium, cobalt, iron, or nickel. Those skilled in the art can select the appropriate method based on the characteristics of the target detection system.

[0037] To further improve the representation quality of the sensing signals, a signal preprocessing step can be added after acquiring the macroscopic dynamic sensing signals and before inputting them into the data branch network. Specifically, a sliding window technique is used to extract data from the time series of the acquired macroscopic dynamic sensing signals. The window length can be set to 50 to 200 sampling points, and the window sliding step size can be set to 25% to 50% of the window length. Let the signal window contain... The sampling point, the first The signal value of each sampling point is denoted as ( The maximum and minimum values ​​of the signal within the signal window are denoted as follows: and Multidimensional statistical features, including the mean of the signal sequence, are extracted from each captured signal window. Standard deviation Peak-to-peak value skewness and kurtosis In this embodiment, various statistical features are extracted to characterize the dynamic fluctuation characteristics of the signal. These statistical features are input into the data branch network as a representation of the macroscopic dynamic sensing signal. It should be noted that the above signal preprocessing steps are not a necessary condition for implementing this method. Without setting this preprocessing step, the original resistance time series can be directly input into the data branch network, and this method can still operate normally. In addition to the sliding window statistical features mentioned above, the preprocessing method can also be replaced by Fast Fourier Transform to extract frequency domain features, or wavelet transform to extract time-frequency joint features.

[0038] Example 2 Based on Example 1, this embodiment further describes a generalized identification method for unknown homologues, that is, a method in which the system can still reliably classify and identify a target gas that is a novel chemical homologue not included in the model training set.

[0039] Chemical homologues are groups of chemical compounds with similar structures that differ only in carbon chain length or the number of methylene groups. For example, methanol, ethanol, propanol, and butanol belong to the alcohol homologues, sharing the core structural feature of the hydroxyl (-OH) functional group. In traditional data-driven approaches, if the training set only contains experimental data for methanol and propanol, the model cannot identify ethanol or butanol, which are not present in the training set, because the data-driven model learns the mathematical characteristics of the signal curve rather than its chemical essence. This embodiment, through a feature learning mechanism of physical branch networks, endows the system with the cognitive ability to logically transfer from known homologues to unknown homologues.

[0040] Specifically, during the model training or initialization phase, the physical branch network uses a multilayer perceptron to perform feature mapping on the microscopic physical descriptors of known homologues. It learns and extracts the common physical laws and distribution mapping intervals of known homologues containing specific shared chemical functional groups on the corresponding microscopic physical descriptors. Taking alcohol homologues as an example, although methanol, propanol, and pentanol have different carbon chain lengths, they share the hydroxyl functional group. Therefore, when they undergo chemisorption with palladium single-atom active sites, their microscopic parameters, such as adsorption energy, charge transfer, and Fermi level shift, exhibit a systematic trend of change following certain physical laws. During training, the physical branch network captures these common features determined by the essential properties of the functional groups, causing the physical embedding features corresponding to the same type of functional group to form cohesive clustered regions, i.e., distribution mapping intervals, in the high-dimensional feature space. Different functional group categories (such as hydroxyl, carbonyl, and ester groups) occupy distinct regions in the feature space.

[0041] When the target gas is an unknown homologue not included in the model training set (e.g., the training set only contains data for methanol and propanol, while ethanol appears in the testing phase), the system first obtains the corresponding microscopic physical descriptor (adsorption energy, charge transfer, etc. of ethanol) from the physical descriptor database. Since ethanol shares the hydroxyl functional group with methanol and propanol in the training set, after mapping by the physical branch network, the embedding features of ethanol naturally fall within the distribution mapping range of alcohol homologues in the high-dimensional feature space. Based on the distribution mapping range into which the physical embedding features of the unknown homologue fall, the system outputs the generalized category result of the unknown homologue through logical transfer inference, classifying it as "alcohol". The essence of this process is that the physical branch network learns the electronic interaction rules between functional groups and active sites, rather than the mechanical memorization of sample labels. Therefore, even if a specific compound has not appeared in the training data, as long as its functional groups exhibit a characteristic distribution trend consistent with known homologues in the physical descriptor space, the system can make reliable inferences based on physical laws.

[0042] To verify the generalization recognition performance, this embodiment conducted a leave-one-out cross-validation experiment. In the experiment, one compound (such as ethanol) was sequentially removed from the training set of alcohols, and the model was trained using only the remaining homologues. The model's ability to correctly identify the removed compound was then tested during the testing phase. Experimental results show that the generalization recognition accuracy for unknown homologues using the physics-data dual-drive method reached over 95%, while the control scheme using only a data branch network achieved a generalization recognition accuracy of less than 60% under the same conditions. This result demonstrates that the introduction of microscopic physical descriptors can effectively endow the model with the ability to extrapolate and predict chemicals outside the training set.

[0043] Example 3 This embodiment, based on Embodiment 1 or Embodiment 2, further introduces an attribution analysis method based on Shapley additive interpretation (SHAP) to verify the reliability of the model decision after outputting the category identification results of the target gas.

[0044] In practical applications, especially in fields with extremely high reliability requirements such as assisted medical diagnosis or industrial safety early warning, simply obtaining classification results is insufficient. It is also necessary to confirm that the model's decisions are indeed based on real physicochemical laws rather than interference from environmental noise or signal artifacts. Therefore, this embodiment, after obtaining the category recognition results, introduces the SHAP attribution analysis algorithm to quantify the contribution weights corresponding to each input microscopic physical descriptor.

[0045] Specifically, the SHAP algorithm is based on the Shapley value theory in cooperative game theory, determining the importance of each input feature by calculating its marginal contribution to the model output across all possible combinations of feature subsets. For the input of the physical branch network... A microscopic physical descriptor Its SHAP value The calculation formula is: in It is the set of all microscopic physical descriptors. This represents the total number of features in the set (i.e., the dimension of the physical descriptor). For features not included Any feature subset, For subset The number of features contained in it To represent factorial operation, To use only a subset of features The model's predicted output value is used as input. For in subset Add features based on The difference between the predicted output value of the later model and the predicted output value. That is, characteristics In subset The marginal contribution of the model output under given conditions. The coefficients in the formula. This is a weighting factor used to perform a weighted average over all possible subset combinations to ensure fairness and consistency in the distribution of contributions for each feature. The SHAP value for each descriptor feature. This reflects the direction and magnitude of the feature's contribution to the final classification decision. A positive value indicates that the feature has a positive promoting effect on the current category determination, while a negative value indicates that it has an inhibitory effect.

[0046] After obtaining the SHAP values ​​of each microscopic physical descriptor, the contribution weights are ranked and analyzed. For example, when classifying a certain ester compound, the SHAP analysis results show that the contribution weight of the Fermi level shift ranks first, followed by charge transfer and adsorption energy. This result is highly consistent with the physical mechanism revealed by density functional theory calculations, namely, the electronic interaction between the ester functional group and the palladium single atom is indeed mainly manifested through the modulation of the Fermi level. The ranking analysis results based on the contribution weights can verify whether the classification decision logic of the joint feature representation is dominated by the real microscopic electronic interaction law. If the SHAP analysis shows that parameters directly related to chemisorption (such as adsorption energy and charge transfer) in the physical descriptor have a dominant contribution, it proves that the model's decision logic conforms to physical expectations and eliminates the interference of environmental noise and signal artifacts; if the analysis results are abnormal (such as descriptors unrelated to the chemical mechanism having a dominant contribution), it suggests that there may be data quality problems or abnormal model training, which requires further investigation. This decision reliability verification mechanism enables a semi-quantitative opening of the "black box" of deep learning models, making every judgment of the model verifiable and reasonable.

[0047] It should be noted that SHAP attribution analysis is an optional verification method and not a necessary step for implementing the basic identification method. Even without performing attribution analysis, the physical-data dual-driven gas identification method described in Example 1 can still run independently and completely and output category identification results. In addition, besides the SHAP method, attribution analysis can also be replaced by other interpretability analysis methods such as LIME (Locally Interpretable Model-Independent Interpretation), Gradient Weighted Class Activation Mapping (Grad-CAM), or integral gradient method.

[0048] To further verify the effectiveness and interpretability of the aforementioned physical-data dual-driven framework, Figure 3 Experimental validation results of the system are presented from five dimensions: feature relevance, feature-class mapping relationship, feature importance ranking, SHAP attribution distribution, and classification performance.

[0049] Figure 3 Subplot (a) shows the correlation matrix of all input features. The size and color intensity of the circular markers represent the Pearson correlation coefficient between any two features. From the graph, it can be observed that microscopic physical descriptors (such as adsorption energy)... Charge transfer amount Fermi level d-zone center The physical descriptors and macroscopic statistical features (such as mean, standard deviation, peak-to-peak value, skewness, and kurtosis) exhibit significant differential correlation patterns. Specifically, the correlation between similar descriptors is high, while the correlation between microscopic physical descriptors and macroscopic statistical features is generally low. This indicates that the information extracted from the physical branch and the data branch is highly complementary, and their fusion can provide the classifier with richer discriminative criteria, rather than redundant and repetitive information.

[0050] Figure 3 Subgraph (b) in the diagram is a feature-category Sankey diagram, which visually illustrates the mapping relationship and information flow intensity between each input feature and the five gas categories (alkanes, alcohols, ketones, aldehydes, and esters). The diagram shows that different gas categories exhibit significantly different response patterns at the feature level. For example, the identification of alcohols and aldehydes is highly dependent on specific microscopic physical descriptors (such as charge transfer and Fermi levels), while alkanes are more driven by macroscopic statistical features. This differentiated feature dependency pattern between categories verifies the necessity of the collaborative work between the physical branch and the data branch in the dual-branch architecture—relying solely on any one feature branch cannot simultaneously cover the effective identification of all categories.

[0051] Figure 3 Subplot (c) in the figure is a bar chart ranking the global feature importance of the model. This chart measures the overall contribution of each feature to the classification decision by calculating the mean absolute SHAP value of each feature across all test samples. From the ranking results, the mean, charge transfer, and other features are ranked higher. The root mean square (RMS) value ranked in the top three, indicating that the model relies on both the amplitude characteristics of the macroscopic response signal and the microscopic electron transfer parameters when making classification decisions. Notably, among the top ten most important features, microscopic physical descriptors and macroscopic statistical features each account for a considerable proportion, further confirming the rationality of the cross-modal fusion strategy. If only a purely data-driven single-branch network is used, the model will completely lose the chemical cognitive information carried by the microscopic physical descriptors, resulting in a significant decrease in its ability to identify homologues with similar physicochemical properties.

[0052] Figure 3Subplot (d) is a Beeswarm Plot, showing the distribution of SHAP values ​​for each feature across all test samples. Each point in the plot represents a test sample, and the color of the point indicates the strength of the corresponding feature (red for high values, blue for low values). The position of the point on the horizontal axis represents the SHAP contribution of that feature to the classification result of the current sample. It is clear from the plot that the distribution of SHAP values ​​for each feature exhibits a distinct directionality. For example, high-value samples (red points) of the Mean feature are mainly distributed in the positive SHAP value region, indicating that a higher Mean response positively influences the model's determination of a specific gas category; while the charge transfer (… The SHAP distribution of the model exhibits a symmetrical positive-negative differentiation trend across different categories, indicating that this microscopic parameter plays a crucial "watershed" role in distinguishing different functional group types. This distribution pattern is highly consistent with the electron interaction mechanism revealed by density functional theory, verifying that the model has indeed learned physicochemical laws, rather than simply memorizing training data.

[0053] Figure 3 Subplot (e) shows the Receiver Operating Characteristic (ROC) curve. This curve, with the false positive rate on the horizontal axis and the true positive rate on the vertical axis, comprehensively evaluates the overall performance of the classifier under different decision thresholds. As can be seen from the figure, the ROC curve of the model using the physics-data dual-driven framework (MLP+CNN) is close to the upper left corner, and the area under the curve (AUC) reaches 1.00, indicating that the model achieves perfect classification of the five gas categories on all test samples, with a true positive rate approaching 100% and a false positive rate approaching 0. This result statistically validates the superior classification performance of the physics-data dual-driven method in gas identification tasks.

[0054] comprehensive Figure 3 The experimental results presented across five dimensions lead to the following conclusions: the microscopic electronic structure features extracted by the physics branch and the macroscopic dynamic statistical features extracted by the data branch exhibit good information complementarity, and their cross-modal fusion significantly enhances the model's discriminative ability; SHAP attribution analysis confirms that the model's decision-making logic is dominated by real physicochemical laws, demonstrating good interpretability; and the ROC curve and AUC index verify the classification accuracy and reliability of the method from an overall performance perspective.

[0055] Example 4 This embodiment provides a gas identification system based on a physical-data dual-driven framework, used to implement the gas identification method described in any of Embodiments 1 to 3. The technical features of Embodiments 1 to 3 can be used in this embodiment. Figure 2 As shown, the system includes a sensor signal acquisition module 201, a micro-feature retrieval module 202, a dual-drive deep learning computing module 203, and a classification output module 204.

[0056] The sensor signal acquisition module is used to acquire macroscopic dynamic sensing signals triggered by the target gas on the sensor array. This module includes the sensor array and its associated signal conditioning and acquisition circuitry. The sensor array contains eight sensing units, each with a sensing interface constructed by coating a composite material with catalytically active sites supported on a conductive carrier onto a flexible polyimide substrate with gold interdigitated electrodes. The signal conditioning circuitry includes a constant current source drive circuit and an operational amplifier conditioning circuit. The constant current source provides a stable bias current to each sensing channel, and the operational amplifier amplifies the weak resistance change signal from the sensor to the effective range of the analog-to-digital converter (ADC). The ADC converts the analog signal into a digital signal. The acquired digital signal is transmitted to the dual-drive deep learning computing module via a serial peripheral interface.

[0057] The microscopic feature retrieval module is used to obtain microscopic physical descriptors corresponding to a preset set of candidate gases. This module is implemented as a structured database, which stores the microscopic physical descriptor vectors of each candidate gas molecule obtained through theoretical calculations. The database uses a key-value pair index structure, with the gas molecule name or molecular identifier encoded as the key and the corresponding physical descriptor vector as the value. During the detection process, the microscopic feature retrieval module reads the corresponding physical descriptor data in batches according to the index list of the candidate gas set and transmits it to the input end of the physical branch network of the dual-drive deep learning computing module. The database can be deployed in local storage to reduce retrieval latency, or it can be obtained from a remote server via a network interface.

[0058] The dual-drive deep learning computing module is connected to both the sensor signal acquisition module and the microscopic feature retrieval module, serving as the core computing unit of the entire system. Internally, this module comprises three sub-modules: a data branch network 2031, a physical branch network 2032, and a feature fusion sub-module 2033. The data branch network receives macroscopic dynamic sensor signals from the sensor signal acquisition module and extracts macroscopic sensor features using a one-dimensional convolutional neural network. The physical branch network receives microscopic physical descriptors from the microscopic feature retrieval module and obtains microscopic physical embedded features through multilayer perceptron mapping. The feature fusion sub-module performs cross-modal feature fusion between the macroscopic sensor features output from the data branch network and the microscopic physical embedded features output from the physical branch network to obtain a joint feature representation. In terms of hardware implementation, the dual-drive deep learning computing module can be deployed on a server equipped with a GPU accelerator card to support high-performance parallel inference, or it can be deployed on an embedded edge computing platform (such as an ARM processor with a neural network accelerator) after model quantization and compression to achieve real-time inference on the edge.

[0059] The classification output module is connected to the output of the dual-drive deep learning computation module. It is used to perform classification calculations based on joint feature representations and output the category identification results of the target gas. The classification output module internally contains a fully connected layer and a Softmax probability mapping layer. Its output includes the probability value corresponding to each candidate category and the final determined gas category name. The classification results can be displayed locally on a screen or uploaded to a higher-level monitoring system or cloud platform for further processing via a communication interface (such as Ethernet or a wireless communication module).

[0060] One embodiment of this application also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps in the above-described method embodiments. The computer-readable storage medium may include any type of volatile or non-volatile memory, or any combination thereof, storing computer program code. Specifically, the computer-readable storage medium may include, but is not limited to: magnetic storage devices (e.g., magnetic tape, hard disk, floppy disk), optical storage devices (e.g., optical disc (CD), digital versatile optical disc (DVD)), magneto-optical storage devices (e.g., magneto-optical disc), semiconductor memory (e.g., read-only memory (ROM), random access memory (RAM), electrically erasable programmable read-only memory (EEPROM), flash memory, etc.) or other storage technologies. In this application, the computer-readable storage medium refers to a non-transitory readable storage medium, excluding the transient propagation signal itself (e.g., modulated data signals, carrier waves, etc.).

[0061] Furthermore, one embodiment of this application provides a gas identification system based on a physical-data dual-drive framework, comprising: a memory and a processor. The memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the steps in the above-described method embodiments of this application. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0062] Furthermore, one embodiment of this application also provides a computer program product, including computer-executable instructions that, when executed by a processor, implement the steps in the above-described method embodiments.

[0063] It should be noted that the terms "first," "second," etc., used in this specification are only used to distinguish similar objects, and the elements defined by them do not necessarily require or imply any actual relationship or order between these elements, nor are they used to describe a specific order or sequence.

[0064] The terms “comprising,” “including,” or any other variations thereof, as used herein, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The phrase "execute according to a certain element" as used in this specification means at least according to that element, including both "execute only according to that element" and "execute according to that element and other elements".

[0066] Those skilled in the art will understand that variations of the above-described method steps can be obtained based on the embodiments disclosed in this application without any inventive effort. For example, the sequence numbers of the steps described in the method embodiments of this specification do not themselves constitute a limitation on the execution order of these steps. Unless there is an explicit specific limitation in the context, the steps may be executed in a different order than in the embodiments (e.g., the steps with larger sequence numbers are executed first, followed by the steps with smaller sequence numbers), or they may be executed in parallel. Furthermore, other steps may be inserted between multiple steps with consecutively numbered sequence numbers.

[0067] The above description is merely a specific embodiment of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application. Contents not described in detail in this specification belong to prior art known to those skilled in the art.

Claims

1. A gas identification method based on a physical-data dual-driven framework, characterized in that, include: Acquire macroscopic dynamic sensing signals triggered by the target gas on a sensor array, wherein the sensing interface of the sensor array includes a conductive carrier and catalytic active sites dispersed on the conductive carrier at the atomic or sub-nanometer level. Obtain a microphysical descriptor corresponding to a preset set of candidate gases. The microphysical descriptor is used to characterize the atomic-level electronic interaction between gas molecules in the set of candidate gases and the catalytic active site. The macroscopic dynamic sensing signal is input into a preset data branch network for feature extraction to obtain macroscopic sensing features; The microscopic physical descriptor is input into a preset physical branch network for feature mapping to obtain microscopic physical embedding features; The macroscopic sensing features and the microscopic physical embedded features are fused across modes by the feature fusion submodule to obtain a joint feature representation constrained by the atomic-level electron interaction law; Based on the joint feature representation, classification calculations are performed, and the category identification result of the target gas is output.

2. The gas identification method according to claim 1, characterized in that, The microscopic physical descriptor is a quantized parameter obtained through first-principles calculations of quantum mechanics, molecular dynamics simulations, or semi-empirical methods of quantum chemistry. The microscopic physical descriptor characterizing the atomic-level electronic interaction law includes at least one of the following microscopic parameters: adsorption energy, charge transfer amount, Fermi level shift, system band gap change, d-band center eigenvalue, energy level gap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital, dipole moment, molecular polarizability, and ionization energy.

3. The gas identification method according to claim 1, characterized in that, The data branch network includes at least one of a one-dimensional convolutional neural network, a long short-term memory network, a gated recurrent unit, an attention mechanism network, or a temporal convolutional network. The physical branch network includes a multilayer perceptron network; The cross-modal feature fusion of the macroscopic sensing features and the microscopic physical embedding features through the feature fusion submodule includes: fusing the macroscopic sensing features and the microscopic physical embedding features through splicing operators, element-level multiplication operators, weighted summation mechanisms, or attention-based cross-alignment operations; wherein, when the dimensions of the macroscopic sensing features and the microscopic physical embedding features are inconsistent, they are mapped to a unified feature space dimension through linear transformation layers before fusion.

4. The gas identification method according to claim 1, characterized in that, The method supports generalized identification of unknown homologues to be tested, and the method further includes: During the model training or initialization phase, the physical branch network is used to learn and extract the common physical laws and distribution mapping intervals of known homologs belonging to the same homologous series on the corresponding microscopic physical descriptors. Members of the same homologous series share the same core chemical functional groups. The output of the target gas category identification result specifically includes: when the target gas is an unknown homologue not included in the model training set, based on the structural characteristics that the unknown homologue belongs to the same homologue series and has the same core chemical functional group as the known homologue, the distribution mapping interval in the feature space where the microscopic physical embedding feature corresponding to the unknown homologue falls, and the output of the generalized category result of the unknown homologue through logical transfer inference.

5. The gas identification method according to claim 1, characterized in that, The conductive carrier is a material with a high specific surface area; The catalytic active sites are composed of transition metals or noble metals and are discretely distributed in the form of single atoms or ultrafine nanoclusters with a particle size of less than 2 nanometers. They are used to provide a uniform atomic-level coordination environment for the adsorption of chemical gases and to convert the microscopic differences of the functional groups of the target gas into identifiable electrical signal characteristics.

6. The gas identification method according to claim 5, characterized in that, The conductive carrier is reduced graphene oxide, and the catalytic active site is a single palladium atom anchored in a defect site on the surface of the reduced graphene oxide. The palladium single atom generates a specific electronic fingerprint signal by undergoing differentiated orbital hybridization with functional groups of different types of target gases through its unsaturated d orbitals.

7. The gas identification method according to any one of claims 1 to 6, characterized in that, After outputting the category identification result of the target gas, the following steps are also included: An attribution analysis algorithm based on Shapley additive interpretation is introduced to calculate the contribution weight of each input microphysical descriptor for the final category identification result. Based on the ranking analysis of the contribution weights, it is verified whether the classification decision logic of the joint feature representation is dominated by the real microscopic electronic interaction law, so as to eliminate the interference of environmental noise and signal artifacts.

8. The gas identification method according to claim 1, characterized in that, After acquiring the macroscopic dynamic sensing signal triggered by the target gas on the sensor array, and before inputting the macroscopic dynamic sensing signal into the preset data branch network, a signal preprocessing step is also included: The time series data of the acquired macroscopic dynamic sensing signals are extracted using the sliding window technique; Extract multidimensional statistical features from each captured signal window, wherein the multidimensional statistical features include at least one of the signal sequence mean, standard deviation, peak-to-peak value, skewness, or kurtosis; The multidimensional statistical features are input into the data branch network as a representation of the macroscopic dynamic sensing signal.

9. A gas identification system based on a physical-data dual-drive framework, characterized in that, include: The sensing signal acquisition module is used to acquire the macroscopic dynamic sensing signal triggered by the target gas on the sensor array. The sensing interface of the sensor array includes a conductive carrier and catalytic active sites dispersed on the conductive carrier at the atomic or sub-nanometer level. The microscopic feature retrieval module is used to obtain microscopic physical descriptors corresponding to a preset set of candidate gases. The microscopic physical descriptors are used to characterize the atomic-level electronic interaction rules between gas molecules in the set of candidate gases and the catalytic active sites. A dual-drive deep learning computing module is connected to the sensing signal acquisition module and the microscopic feature retrieval module, respectively. Internally, it includes a data branch network, a physical branch network, and a feature fusion submodule. The data branch network extracts the macroscopic dynamic sensing signals to obtain macroscopic sensing features; the physical branch network maps the microscopic physical descriptors to obtain microscopic physical embedded features; and the feature fusion submodule performs cross-modal feature fusion of the macroscopic sensing features and the microscopic physical embedded features to obtain a joint feature representation constrained by the atomic-level electron interaction laws. The classification output module is used to perform classification calculations based on the joint feature representation and output the category identification result of the target gas.

10. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the gas identification method as described in any one of claims 1 to 8.

Images