Intelligent electronic nose gas recognition method, system, medium and equipment

Abstract

The invention discloses an intelligent electronic nose gas recognition method and system, a medium and equipment. The method comprises the following steps: acquiring basic feature data required by electronic nose recognition; according to the basic feature data, information fusion is carried out based on a multilayer context structure mapping network, and context enhancement representation is obtained; according to the context enhancement representation, dynamic fusion of feature data and compensation features is achieved through a self-consistent gating strategy, and compensated input features are obtained; and according to the input features, generating a prediction probability of a final gas type based on a feature interaction vector, and performing model training to obtain a final gas recognition result. Through the joint effect of all the modules, the system has relatively high stability and adaptability in the aspects of known gas recognition, environmental interference suppression, unknown category detection and the like, so that an electronic nose intelligent recognition platform with strong interpretability and good expandability is constructed.

Description

Technical Field

[0001] This invention relates to a method, system, medium, and device for intelligent electronic nose gas recognition, belonging to the fields of electronic nose signal processing, artificial intelligence perception and reasoning, and environmental gas analysis technology. Background Technology

[0002] Electronic nose (E-nose) systems acquire gas response signals through multi-channel gas sensor arrays and use algorithms to identify gas types, concentrations, and source regions, representing a significant direction in the field of artificial intelligence perception in recent years. However, the biggest challenge facing electronic noses in long-term practical applications is that sensor drift, device heterogeneity, and environmental changes cause continuous shifts in signal distribution, resulting in inconsistencies between data from the model training phase and the operational phase, thus significantly reducing recognition accuracy.

[0003] Traditional methods rely on feature engineering, statistical compensation, or shallow classifiers, which are ill-suited to the complexity of drift. In recent years, various drift compensation schemes based on machine learning, transfer learning, and deep neural networks have been proposed, gradually forming two main technological development paths:

[0004] Path 1: Drift compensation technology based on ensemble learning and cross-domain transfer.

[0005] For example, patent CN111103325A (An electronic nose signal drift compensation method based on integrated neural network learning) trains shallow neural networks with data from different time periods, then constructs a weighted integrated classifier, and uses historical drift patterns to infer future drift trends, achieving lightweight drift compensation without complex deep structures. It has the characteristics of low computational cost, suitability for resource-constrained devices, and construction of compensation models through historical offsets.

[0006] Path 2: Feature alignment method based on deep learning and domain adaptation.

[0007] For example, patent CN113837085A (drift compensation method based on domain adaptive convolutional neural network) extracts high-dimensional features through multi-path convolutional network and introduces MMD for cross-domain alignment. It can achieve drift compensation without using target domain data. Its value lies in its ability to automatically learn cross-domain feature projection, flexible architecture, customizability, and applicability to complex and nonlinear gas response modes.

[0008] Despite the positive progress made by the above technologies in electronic nose signal processing, drift compensation, and cross-domain recognition, key shortcomings still exist:

[0009] 1. Traditional drift compensation models lack multi-layer semantic context modeling capabilities: Existing technologies such as CN111103325A typically handle sensor drift through shallow neural networks or transfer mapping of extreme learning machines. The feature processing and compensation logic of these methods mainly focus on numerical-level distribution alignment, achieving only local alignment in the numerical space and lacking semantic-level structural modeling. However, real drift is often associated with multiple factors such as device usage time, environmental evolution, usage scenarios, and gas source changes, exhibiting a multi-layered contextual change with semantic logic.

[0010] 2. Transfer learning methods are highly dependent on the target scene and lack open adaptability: For example, CN113837085A uses transfer learning or domain adaptation to align the source and target domains. Although it can improve cross-domain recognition performance, it has two significant limitations: (1) the model needs to obtain target domain data (with or without labels) to perform transfer or adaptation; (2) the structure of the transfer mapping is fixed and it does not have the ability to generalize to unknown scenes, unknown gases or combination patterns.

[0011] In real-world applications, the composition of ambient gases is highly uncertain, and the emergence of new gases and mixtures is often unpredictable, resulting in existing migration methods lacking open-world processing capabilities.

[0012] Electronic nose recognition systems suffer from decreased recognition accuracy, insufficient compensation for environmental and temporal drift, difficulty adapting models to contextual changes, and inability to effectively handle unknown odors in open environments under complex or dynamic conditions. Therefore, there is an urgent need to develop an electronic nose recognition system that can comprehensively address these challenges and maintain stable and accurate gas recognition capabilities in open environments. Summary of the Invention

[0013] The purpose of this invention is to provide an intelligent electronic nose gas recognition method, system, medium, and device. It proposes an open electronic nose intelligent recognition system based on multi-layer context structure mapping and a self-consistent compensation mechanism. This system achieves highly robust modeling of sensor array features, dynamic compensation for environmental disturbances, and open recognition of unknown odor samples through the coordinated operation of a multi-layer structure mapping network and a context-driven compensation model.

[0014] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0015] In a first aspect, the present invention provides an intelligent electronic nose gas recognition method, comprising:

[0016] Obtain the basic feature data required for electronic nose recognition;

[0017] Based on the aforementioned basic feature data, information fusion is performed using a multi-layer context structure mapping network to obtain a context-enhanced representation.

[0018] Based on the context-enhanced representation, a self-consistent gating strategy is used to dynamically fuse feature data and compensated features to obtain compensated input features.

[0019] Based on the input features, the predicted probability of the final gas type is generated based on the feature interaction vector, and the model is trained to obtain the final gas identification result.

[0020] Furthermore, the basic feature data required for electronic nose recognition is collected and the model is initialized. The collected basic feature data includes the original response matrix of the sensor array. Basic context variables and array structure diagram ;

[0021] The original response matrix of the sensor array includes the transient response, steady-state characteristics and frequency domain parameters of each sensor in the array under different gases, concentrations and environmental conditions;

[0022] The basic context variables include temperature, humidity, airflow velocity, background gas concentration, and power conditions;

[0023] The array structure graph is an undirected graph or a weighted graph, used to describe the physical layout and coupling relationship between sensors. Nodes represent sensors, and edges represent compensation, association, or cooperative relationships.

[0024] Furthermore, learnable embedding vectors are generated for entities such as sensors, gas categories, and environmental contexts in the system to characterize the inherent characteristics of sensors (sensitivity, drift characteristics, etc.), the internal characteristics of gases (molecular structure, response modes, etc.), and the potential influence modes of the environmental context.

[0025] In the method of this invention, each embedding is randomly initialized with a Gaussian distribution and used as a trainable parameter in subsequent modeling.

[0026] Furthermore, a multi-level structure mapping network is constructed to map the original sensor response to a deep feature space containing array structure relationships and contextual semantics, achieving cross-sensor collaborative representation, including:

[0027] First, the multi-layer context structure mapping network is built on the above array structure diagram. In this architecture, each sensor node is treated as a computing node, used to extract and fuse information from neighboring nodes; the edge structure is used to define the feature propagation path and weights.

[0028] In the method of this invention, each layer of the network corresponds to one "cross-sensor collaborative information propagation", thereby enabling the sensors to absorb structural information from the entire array layer by layer.

[0029] Secondly, after the multi-layer context structure mapping network is constructed, iterative propagation of structural dependencies is performed, setting the initial sensor vector as... ,conduct Layer iteration, the expression is:

[0030] ;

[0031] in, For sensors The neighborhood group, For the first Sensor characteristics after layer propagation For the propagation weights between sensors, Learnable gating coefficients to control the update ratio. For activation function, For bias vectors, Neighbor Sensors In the The feature vector of the layer;

[0032] go through After layer propagation, each sensor obtains a structured vector containing array structure context information. ;

[0033] Finally, the sensor vectors after structure propagation are combined with the environmental context. and gas category embedding Information is fused to construct a "context-enhanced representation," which includes enhanced gas vector representation and enhanced sensor vector representation.

[0034] The enhanced gas vector representation By fusing structured vectors The expression is:

[0035] ;

[0036] in, It can be any one of average pooling, weighted pooling, attention pooling, or learnable graph attention module; For sensors;

[0037] The enhanced sensor vector representation By fusing structured vectors Embedded with Context The expression is:

[0038] ;

[0039] in, This is the weight matrix. It is the bias vector;

[0040] In the method of this invention, the enhancement mechanism used simultaneously injects physical structure, environmental dependence and scene semantics, enabling the model to obtain stronger contextual understanding ability; through the structured modeling method of layer-by-layer propagation, the system can simultaneously capture the cooperative relationship between sensors, the coupling effect brought about by array layout and the influence of environmental context, and form more discriminative deep feature representations in both spatial and temporal dimensions. Compared with traditional modeling methods that rely on single-layer convolution or basic graph structure, it has higher expressive power and universality.

[0041] Furthermore, to compensate for sensor response shifts caused by factors such as time drift and environmental disturbances, an open recognition capability for unknown odor categories is established. The compensation mechanism constructs a consistent reasoning process based on multimodal contextual information and achieves dynamic fusion of original and compensated features through a self-consistent gating strategy. Specifically, it includes:

[0042] First, construct the context vector. It is obtained by combining the sensor enhancement vector, the gas enhancement vector, and the environmental context vector, and the expression is: The vector represents the comprehensive context conditions under the current "gas-environment-array state";

[0043] Secondly, the compensation features are obtained through latent feature compensation mechanism reasoning, and an attention mechanism module is designed to... For query vectors, use a structured sensor vector matrix. For key / value pairs, the expression is:

[0044] ;

[0045] ;

[0046] in, For the dimension of the query vector, For potential compensation vectors;

[0047] Finally, to improve the adaptability of the compensation mechanism under different contextual conditions, the present invention introduces a dynamic gating strategy based on contextual scenario vectors to flexibly adjust between the original features and potential compensation features.

[0048] The gating strategy extracts dependency information related to environmental disturbances, sensor consistency, and structural stability from the contextual vector, and generates dynamic gating coefficients through a learnable nonlinear mapping function. The nonlinear mapping function is not limited to a specific form and can be any monotonically differentiable normalized function, including but not limited to sigmoid, softmax or other normalized weighted mappings; the gating strategy can be further combined with historical residual information, uncertainty estimation or confidence features to improve the reliability of context judgment.

[0049] Based on the gating coefficient, the original feature set With potential compensation vector Perform weighted fusion and generate the compensated optimal feature representation. The fusion process employs weighted mixing, attention fusion, or other fusion methods with differentiable structures, provided that the compensation weights can be adaptively adjusted based on the context.

[0050] The method of this invention, through its gating compensation strategy, can maintain the stability and consistency of feature representation under different environmental conditions, drift levels, and array structure changes, providing reliable input features for subsequent recognition modules. Furthermore, this mechanism does not rely on frequent manual calibration, and can maintain the robustness of the system output, overcoming the technical limitations of existing methods that generally employ static calibration or manual modeling, which lack adaptability.

[0051] Furthermore, for gas identification and reasoning, an interaction vector is first constructed for each sampled data point, expressed as:

[0052] ;

[0053] The matching degree is obtained through a multilayer perceptron (MLP), expressed as:

[0054] ;

[0055] The specific structure of MLP includes the number of layers, activation functions, and regularization, which are configured according to the application scenario and computing power limitations.

[0056] Secondly, the matching degree obtained above The predicted probability of the final gas type is generated using Softmax or other normalized probability mapping functions, expressed as follows:

[0057] ;

[0058] in, This is the weight matrix. It is the bias vector;

[0059] The system supports OOD detection for known and unknown gas types, as well as dynamic incremental learning for new gas types.

[0060] Finally, the model is trained end-to-end, where the cross-entropy loss expression is:

[0061] ;

[0062] The Adam optimization algorithm is used to update the model parameters, which include the embedding vector, structure propagation network parameters, attention parameters, compensation module parameters, and classification head parameters.

[0063] The method of this invention introduces an open space prediction mechanism and a confidence interval inference method into the model structure. When the system identifies a feature pattern that does not conform to the existing category, it can actively determine that the sample is an unknown category or an out-of-distribution sample (OOD), thus realizing a truly open electronic nose recognition.

[0064] Furthermore, the obtained optimal feature representation after compensation is... Compared with the final gas identification results Data storage and structured report generation are performed, and the data storage includes the sensor's raw response matrix. Enhanced feature vectors Gas vector representation Self-consistent compensation characteristics and gas identification results ;

[0065] The structured report includes gas identification results, reasoning explanations, risk levels, and recommended actions.

[0066] In the method of this invention, data is stored for subsequent historical analysis, drift pattern backtracking, and incremental learning, ensuring the long-term traceability and performance optimization of the system.

[0067] In a second aspect, the present invention provides an intelligent electronic nose gas recognition system, comprising:

[0068] The data acquisition module is used to acquire the basic feature data required for electronic nose recognition;

[0069] The multi-layer structure mapping module is used to perform information fusion based on the basic feature data and a multi-layer context structure mapping network to obtain a context-enhanced representation.

[0070] The compensation and fusion module is used to dynamically fuse feature data and compensated features based on the context-enhanced representation through a self-consistent gating strategy to obtain compensated input features.

[0071] The training and recognition module is used to generate the predicted probability of the final gas type based on the feature interaction vector according to the input features and to train the model to obtain the final gas recognition result.

[0072] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the intelligent electronic nose gas recognition method described in any of the first aspects.

[0073] Fourthly, the present invention provides a computer device, comprising:

[0074] Memory, used to store computer programs / instructions;

[0075] A processor for executing the computer program / instructions to implement the steps of the intelligent electronic nose gas recognition method as described in any one of the first aspects.

[0076] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0077] 1. The intelligent electronic nose gas recognition method provided by this invention includes: acquiring basic feature data required for electronic nose recognition; performing information fusion based on a multi-layer context structure mapping network to obtain context-enhanced information. This invention solves the problems of insufficient feature expression ability and weak cross-scene generalization of existing electronic noses in complex environments, and improves the structured representation ability; realizing the dynamic fusion of feature data and compensation features through a self-consistent gating strategy to obtain compensated input features. The open recognition method adopted by this invention supports unknown gases and abnormal states, overcomes the limitation of traditional electronic noses that "can only classify, not exclude," and improves the reliability and risk prevention and control capabilities in safety-sensitive scenarios; generating the predicted probability of the final gas type based on feature interaction vectors and training the model to obtain the final gas recognition result; storing the data and generating a structured report. This invention generates a structured report with causal chains, which improves the problem of opaque decision-making in traditional electronic noses and facilitates manual review and quick understanding by on-site operators.

[0078] 2. The intelligent electronic nose gas recognition system provided by this invention integrates data acquisition, multi-layer structure mapping, compensation and recognition, prediction and training, and storage and visualization modules into a unified end-to-end system. Through shared feature representation and joint optimization, multiple functional modules can improve performance simultaneously. The method of this invention overcomes the problem of fragmented "feature processing-compensation correction-classification recognition" links commonly found in traditional electronic nose systems, significantly reducing information loss and improving overall reasoning efficiency. In addition, the system of this invention has good modularity, and can access additional environmental perception information (such as images, acoustic signals or external environmental sensor data) while keeping the core recognition process unchanged. This enables the system to have stronger environmental understanding and auxiliary reasoning capabilities in complex scenarios, and also provides convenient expansion interfaces for subsequent functional enhancement and cross-modal research.

[0079] 3. The computer-readable storage medium and computer device provided by the present invention can execute the steps of the intelligent electronic nose gas recognition method provided by the present invention. Attached Figure Description

[0080] Figure 1 This is an overall flowchart of the intelligent electronic nose gas recognition method provided according to an embodiment of the present invention;

[0081] Figure 2 This is a schematic diagram of the system structure of the intelligent electronic nose gas recognition method provided in an embodiment of the present invention. Detailed Implementation

[0082] It should be noted that:

[0083] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0084] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0085] Example 1

[0086] In certain industrial waste gas emission monitoring scenarios, the gas composition is complex and environmental disturbances are significant, such as large fluctuations in temperature and humidity, changes in wind speed, and long-term drift caused by sensor aging. The open electronic nose intelligent recognition system provided by this invention can operate stably in the above-mentioned open environment, and its specific implementation is as follows:

[0087] like Figure 1 As shown in the figure, this embodiment introduces a smart electronic nose gas recognition method, including:

[0088] Obtain the basic feature data required for electronic nose recognition;

[0089] Based on the aforementioned basic feature data, information fusion is performed using a multi-layer context structure mapping network to obtain a context-enhanced representation.

[0090] Based on the context-enhanced representation, a self-consistent gating strategy is used to dynamically fuse feature data and compensated features to obtain compensated input features.

[0091] Based on the input features, the predicted probability of the final gas type is generated based on the feature interaction vector, and the model is trained to obtain the final gas identification result.

[0092] Furthermore, in this embodiment, the response signals of the sensor array are continuously acquired, and environmental parameters such as temperature and humidity and equipment status information are recorded simultaneously, while historical case data is retrieved; the system initialization module constructs the initial structure of the multi-layer context structure mapping network according to the type of sensor array used and the application scenario, and loads the basic category set and its structured representation.

[0093] The basic feature data collected includes the raw response matrix of the sensor array. Basic context variables and array structure diagram ;

[0094] The original response matrix of the sensor array includes the transient response, steady-state characteristics and frequency domain parameters of each sensor in the array under different gases, concentrations and environmental conditions;

[0095] The basic context variables include temperature, humidity, airflow velocity, background gas concentration, and power conditions;

[0096] The array structure graph is an undirected graph or a weighted graph, used to describe the physical layout and coupling relationship between sensors. Nodes represent sensors, and edges represent compensation, association, or cooperative relationships.

[0097] Furthermore, a multi-layer structure mapping module is constructed. When the array generates a response waveform at a certain moment, a multi-layer parsing process that strictly corresponds to the core technology process is executed sequentially:

[0098] The first layer preprocesses the data, extracting the instantaneous change features of the local response curves to prepare basic features for subsequent structured modeling.

[0099] The second layer is based on the array structure diagram. By using iterative propagation of a graph neural network, the response correlation structure and synergistic effect between different sensor channels are modeled, with the initial sensor vector set as... ,conduct Layer iteration, the expression is:

[0100] ;

[0101] in, For sensors The neighborhood group, For the first Sensor characteristics after layer propagation For the propagation weights between sensors, Learnable gating coefficients to control the update ratio. For activation function, For bias vectors, Neighbor Sensors In the The feature vector of the layer;

[0102] go through After layer propagation, each sensor obtains a structured vector containing array structure context information. ;

[0103] To make the array structure diagram The physical meaning is clearer in this embodiment. The construction method will be further explained in this embodiment: The graph structure Used to characterize the hierarchical relationships, correlations, and cooperative response features within a gas sensor array, depending on the specific device, array layout, or application requirements. Any of the following methods of construction shall fall within the scope of protection of this invention, including:

[0104] (1) Nearest neighbor association graph construction method based on physical layout: calculate the Euclidean distance based on the geometric coordinates of the sensors on the array, and when the distance between two sensors is less than a preset threshold, the Euclidean distance is calculated. In this case, an undirected edge is established between the two to reflect their natural neighborhood relationship;

[0105] (2) Construction method of statistical correlation diagram based on historical response: Based on a large amount of collected gas sample data, calculate the Pearson correlation coefficient between any two sensor response sequences, and use the correlation coefficient as the basis for the construction of the statistical correlation diagram. As the weight of the edge, it reflects the collaborative response characteristics in terms of functionality;

[0106] (3) Fully connected graph structure based on full connectivity: When the sensor correlation cannot be determined in advance, a fully connected graph can be used to allow the model to automatically learn the optimal edge weights during training;

[0107] (4) Dynamic graph structure based on task or environment: The edge weights or adjacency relationships are dynamically adjusted according to different gas categories, different environmental scenarios or context information at different times, so that the structure mapping network has stronger dynamic adaptability.

[0108] The third layer combines the sensor vectors after structural propagation with the environmental context. and gas category embedding Information is fused to construct a "context-enhanced representation," which includes enhanced gas vector representation and enhanced sensor vector representation.

[0109] The enhanced gas vector representation By fusing structured vectors The expression is:

[0110] ;

[0111] in, It can be any one of average pooling, weighted pooling, attention pooling, or learnable graph attention module; For sensors;

[0112] The enhanced sensor vector representation By fusing structured vectors Embedded with Context The expression is:

[0113] ;

[0114] in, This is the weight matrix. This is the bias vector.

[0115] Furthermore, to compensate for sensor response shifts caused by factors such as time drift and environmental disturbances, an open recognition capability for unknown odor categories is established. The compensation mechanism constructs a consistent reasoning process based on multimodal contextual information and achieves dynamic fusion of original and compensated features through a self-consistent gating strategy. Specifically, it includes:

[0116] First, construct the context vector. It is obtained by combining the sensor enhancement vector, the gas enhancement vector, and the environmental context vector, and the expression is: The vector represents the comprehensive context conditions under the current "gas-environment-array state";

[0117] Secondly, the compensation features are obtained through latent feature compensation mechanism reasoning, and an attention mechanism module is designed to... For query vectors, use a structured sensor vector matrix. For key / value pairs, the expression is:

[0118] ;

[0119] ;

[0120] in, For the dimension of the query vector, For potential compensation vectors;

[0121] Finally, to improve the adaptability of the compensation mechanism under different contextual conditions, the present invention introduces a dynamic gating strategy based on contextual scenario vectors to flexibly adjust between the original features and potential compensation features.

[0122] The gating strategy extracts dependency information related to environmental disturbances, sensor consistency, and structural stability from the contextual vector, and generates dynamic gating coefficients through a learnable nonlinear mapping function. The nonlinear mapping function is not limited to a specific form and can be any monotonically differentiable normalized function, including but not limited to sigmoid, softmax or other normalized weighted mappings; the gating strategy can be further combined with historical residual information, uncertainty estimation or confidence features to improve the reliability of context judgment.

[0123] Based on the gating coefficient, the original feature set With potential compensation vector Perform weighted fusion and generate the compensated optimal feature representation. The fusion process employs weighted mixing, attention fusion, or other fusion methods with differentiable structures, provided that the compensation weights can be adaptively adjusted based on the context.

[0124] In this embodiment, a consistency determination mechanism is used to evaluate the degree of matching between the current feature and the existing category representation space:

[0125] When it is found that the response of some channels deviates from the normal range, the self-consistent compensation mechanism is triggered, and the drift deviation is automatically corrected according to the historical mapping relationship to obtain a stable signal after compensation.

[0126] When the compensation signal still cannot match any known category, the system activates the open identification unit to perform boundary confidence inference and then determine whether it belongs to an unknown gas, a mixed gas, or other abnormal mode.

[0127] Furthermore, for gas identification and reasoning, an interaction vector is first constructed for each sampled data point, expressed as:

[0128] ;

[0129] The matching degree is obtained through a multilayer perceptron (MLP), expressed as:

[0130] ;

[0131] The specific structure of MLP includes the number of layers, activation functions, and regularization, which are configured according to the application scenario and computing power limitations.

[0132] Secondly, the matching degree obtained above The predicted probability of the final gas type is generated using Softmax or other normalized probability mapping functions, expressed as follows:

[0133] ;

[0134] in, This is the weight matrix. This is the bias vector.

[0135] The system supports OOD detection for known and unknown gas types, as well as dynamic incremental learning for new gas types.

[0136] Finally, the model is trained end-to-end, where the cross-entropy loss expression is:

[0137] ;

[0138] The Adam optimization algorithm is used to update the model parameters, which include the embedding vector, structure propagation network parameters, attention parameters, compensation module parameters, and classification head parameters.

[0139] In this embodiment, when the system is in continuous learning mode, the model can be incrementally updated based on manual confirmation or feedback from the host computer to adapt to long-term environmental changes and the effects of sensor aging.

[0140] Furthermore, the obtained optimal feature representation after compensation is... Compared with the final gas identification results Data storage and structured report generation are performed, and the data storage includes the sensor's raw response matrix. Enhanced feature vectors Gas vector representation Self-consistent compensation characteristics and gas identification results ;

[0141] In this embodiment, the gas identification results, such as detecting sulfide gases or unknown gases, and the inference explanation, such as an abnormal response in a certain channel, are compensated to improve the similarity with historical patterns, the risk level, and suggested measures, such as improving ventilation or issuing an early warning command. The event records are archived in the data storage module for subsequent incremental training and model maintenance.

[0142] Example 2

[0143] like Figure 2 As shown, based on the intelligent electronic nose gas recognition method described in Embodiment 1, this embodiment introduces an intelligent electronic nose gas recognition system, including:

[0144] The data acquisition module is used to acquire the basic feature data required for electronic nose recognition;

[0145] The multi-layer structure mapping module is used to perform information fusion based on the basic feature data and a multi-layer context structure mapping network to obtain a context-enhanced representation.

[0146] The compensation and fusion module is used to dynamically fuse feature data and compensated features based on the context-enhanced representation through a self-consistent gating strategy to obtain compensated input features.

[0147] The training and recognition module is used to generate the predicted probability of the final gas type based on the feature interaction vector according to the input features and to train the model to obtain the final gas recognition result.

[0148] Example 3

[0149] Based on the intelligent electronic nose gas recognition method described in Embodiment 1, this embodiment also has some alternative technical solutions, including:

[0150] 1. Extended solutions for multi-source heterogeneous gas scenarios:

[0151] Option 1 (Identification Method Based on Cross-Modal Gas Information Fusion): In scenarios with complex environmental interference or multi-source gas emissions (such as industrial leak monitoring and urban compound pollution monitoring), in addition to the electronic nose array, heterogeneous sensors such as infrared spectroscopy, ultrasonic disturbance detection, and temperature and humidity arrays may be used. A cross-modal fusion module can be introduced into the data acquisition and initialization process and the construction of the multi-layer context structure mapping network in this embodiment to achieve alignment and collaborative encoding of the electronic nose signal and multi-modal information such as infrared absorption spectra. In the open inference structure for unknown gases in the electronic nose, multi-modal features are used as extended context inputs, enabling the model to identify complex gas combinations that are difficult to distinguish using traditional electronic noses.

[0152] This embodiment can significantly improve the system's ability to distinguish gas types and their mixtures in complex scenarios.

[0153] Option 2 (Weak Signal Enhancement and Recognition for Extremely Low Concentration Gases): For target gases at ppb level or lower, this invention can introduce a weak signal reconstruction mechanism based on variational autoencoders in the signal structuring stage of constructing a multi-layer context structure mapping network. By reconstructing the weak response with denoising and structure preservation, it is more suitable for the processing flow of the multi-layer structure mapping network.

[0154] This embodiment enables the system to maintain stable recognition performance even in extremely low concentration scenarios through the "weak signal reconstruction → multi-layer structure mapping" link.

[0155] 2. Alternative solutions for multi-layer structure mapping modules:

[0156] Option 1 (Hybrid Structure Mapping Network Based on Graph Convolution (GCN) and Graph Attention (GAT): In the construction of the multi-layer structure mapping network in this embodiment, a hybrid feature propagation method of GCN and GAT can be adopted: GCN is used to model the diffusion relationship between local sensors, and GAT is used to dynamically learn the importance of key sensor nodes in the global scope; by combining the two, this embodiment can simultaneously maintain local sensitivity and global reasoning ability, and improve the ability to characterize complex structural odor patterns.

[0157] Option 2 (Gas Structure Relationship Modeling Based on Diffusion Generation Model): In the process of structure-dependent iterative propagation, the structural relationships between sensors are modeled. The diffusion generation model can be used as an alternative to generate potential gas diffusion relationships through a "stepwise denoising" modeling method, enabling the system to automatically infer the potential connection structures of unknown or rare gases. This embodiment allows the potential structural relationships between open set gases to be automatically generated by the model, which is beneficial to improving the representation capability in open set environments.

[0158] 3. Alternatives to the self-consistent compensation mechanism:

[0159] During the training phase of the self-consistent compensation module, learnable noise can be injected into the input features, and an adversarial training strategy can be adopted to make the compensation mechanism more robust when facing actual working conditions such as gas-sensitive drift, temperature and humidity disturbances, and component aging. This embodiment further enhances the long-term stability of the system.

[0160] 4. Alternatives for system-level deployment and service-oriented architecture:

[0161] When large-scale industrial deployments or online scenarios are required, all modules in the system of this invention can be split as follows. In this embodiment, they are split into: edge side (front end): responsible for raw data acquisition, basic preprocessing and some low-level feature extraction; cloud platform: responsible for multi-layer structure mapping, self-consistent compensation, open reasoning and continuous learning. This embodiment is suitable for large-scale industrial monitoring systems that require centralized statistics, unified model updates or cross-regional iterative optimization, which facilitates the realization of service-oriented and cloud-based collaborative operation.

[0162] Example 4

[0163] Based on the intelligent electronic nose gas recognition method described in Embodiment 1, this embodiment introduces a computer-readable storage medium storing a computer program / instruction thereon. When the computer program / instruction is executed by a processor, it implements the steps of the intelligent electronic nose gas recognition method as described in any of Embodiment 1.

[0164] Example 5

[0165] Based on the intelligent electronic nose gas recognition method described in Embodiment 1, this embodiment provides a computer device, including:

[0166] Memory, used to store computer programs / instructions;

[0167] A processor is configured to execute the computer program / instructions to implement the steps of the intelligent electronic nose gas recognition method as described in any one of Embodiment 1.

[0168] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for intelligent electronic nose gas recognition, characterized in that, include: Obtain the basic feature data required for electronic nose recognition; Based on the aforementioned basic feature data, information fusion is performed using a multi-layer context structure mapping network to obtain a context-enhanced representation. Based on the context-enhanced representation, a self-consistent gating strategy is used to dynamically fuse feature data and compensated features to obtain compensated input features. Based on the input features, the predicted probability of the final gas type is generated based on the feature interaction vector, and the model is trained to obtain the final gas identification result.

2. The intelligent electronic nose gas recognition method according to claim 1, characterized in that, The basic feature data includes the raw response matrix of the sensor array. Basic context variables and array structure diagram ; The original response matrix of the sensor array includes the transient response, steady-state characteristics and frequency domain parameters of each sensor in the array under different gases, concentrations and environmental conditions; The basic context variables include temperature, humidity, airflow velocity, background gas concentration, and power conditions; The array structure graph is an undirected graph or a weighted graph, used to describe the physical layout and coupling relationship between sensors. Nodes represent sensors, and edges represent compensation, association, or cooperative relationships.

3. The intelligent electronic nose gas recognition method according to claim 1, characterized in that, The multi-layer context structure mapping network is built on an array structure graph, where each sensor node is regarded as a computing node, used to extract and fuse information from neighboring nodes; Edge structures are used to define feature propagation paths and weights.

4. The intelligent electronic nose gas recognition method according to claim 3, characterized in that, After the multi-layer context structure mapping network is constructed, iterative propagation of structural dependencies is performed. The propagation process is as follows: Set the initial sensor vector as follows: ,conduct Layer iteration, the expression is: ； in, For sensors The neighborhood group, For the first Sensor characteristics after layer propagation For the propagation weights between sensors, Learnable gating coefficients to control the update ratio. For activation function, For bias vectors, Neighbor Sensors In the The feature vector of the layer; go through After layer propagation, each sensor obtains a structured vector containing array structure context information. .

5. The intelligent electronic nose gas recognition method according to claim 1, characterized in that, The context-enhanced representation includes enhanced gas vector representation and enhanced sensor vector representation; The enhanced gas vector representation By fusing structured vectors The expression is: ； in, It can be any one of average pooling, weighted pooling, attention pooling, or learnable graph attention module; For sensors; The enhanced sensor vector representation By fusing structured vectors Embedded with Context The expression is: ； in, This is the weight matrix. It is the bias vector; Contextual Vector The expression is: .

6. The intelligent electronic nose gas recognition method according to claim 1, characterized in that, The compensation features are obtained through an attention score-based compensation mechanism, where the attention score is calculated by the attention module using the context vector. For query vectors, use a structured sensor vector matrix. For key / value pairs, the expression is: ； ； in, For the dimension of the query vector, This is the potential compensation vector.

7. The intelligent electronic nose gas recognition method according to claim 1, characterized in that, Based on the gas identification results, data is stored and a structured report is generated; the data storage includes the sensor's original response matrix. Enhanced feature vectors Gas vector representation Self-consistent compensation characteristics and gas identification results ; The structured report includes gas identification results, reasoning explanations, risk levels, and recommended actions.

8. An intelligent electronic nose gas recognition system, characterized in that, include: The data acquisition module is used to acquire the basic feature data required for electronic nose recognition; The multi-layer structure mapping module is used to perform information fusion based on the basic feature data and a multi-layer context structure mapping network to obtain a context-enhanced representation. The compensation and fusion module is used to dynamically fuse feature data and compensated features based on the context-enhanced representation through a self-consistent gating strategy to obtain compensated input features. The training and recognition module is used to generate the predicted probability of the final gas type based on the feature interaction vector according to the input features and to train the model to obtain the final gas recognition result.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the intelligent electronic nose gas recognition method according to any one of claims 1 to 7.

10. A computer device / equipment / system, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the intelligent electronic nose gas recognition method according to any one of claims 1 to 7.

Images