Detection method and device based on multi-mode gas sensor

Abstract

The invention relates to the technical field of sensor detection, and discloses a detection method and device based on a multi-mode gas sensor, and the method comprises the following steps: carrying out blending processing on a metal oxide semiconductor and a conductive material to obtain a bimodal nano composite material, and creating a multi-mode gas detection module, the preprocessed multi-mode signal data is acquired and obtained; acquiring and analyzing data of the temperature sensor and the humidity sensor to obtain environmental factor compensation parameters; performing adaptive calibration processing on the preprocessed multi-modal signal data according to the environmental factor compensation parameters to obtain calibrated multi-modal signal data; constructing a multi-input neural network model based on the calibrated multi-modal signal data, and training to obtain a gas recognition model; multi-mode signals collected and calibrated in real time are analyzed and judged based on the gas recognition model, gas type and concentration information is obtained, and high-precision and high-reliability gas detection is achieved.

Description

Technical Field

[0001] The present invention relates to the field of sensor detection technology, and in particular to a detection method and device based on a multi-modal gas sensor. Background Art

[0002] With the growing demand for industrial production and environmental monitoring, gas detection technology plays an increasingly important role in many fields. Traditional single-mode gas sensors have obvious limitations in detection accuracy, anti-interference ability, and identification of complex gas mixtures. This has prompted researchers to continuously explore new multimodal gas sensing technologies to improve detection accuracy and reliability. However, current multimodal gas sensors still face many challenges in material design, signal processing, and pattern recognition.

[0003] Existing multimodal gas sensors often use simple material composites, which makes it difficult to fully utilize the synergistic effects of the components, resulting in limited improvement in sensing performance. At the same time, traditional signal processing methods are difficult to effectively handle complex noise and interference in multimodal data, affecting the stability of detection results. In addition, existing gas identification algorithms often perform poorly when processing high-dimensional, strongly correlated multimodal data, and it is difficult to accurately distinguish similar gases or low-concentration gases. Summary of the invention

[0004] The present invention provides a detection method and device based on a multi-modal gas sensor, and the present invention realizes high-precision and high-reliability gas detection.

[0005] In a first aspect, the present invention provides a detection method based on a multimodal gas sensor, the detection method based on the multimodal gas sensor comprising: A metal oxide semiconductor and a conductive material are co-mixed and synthesized to obtain a bimodal nanocomposite material, and a sensor array having a three-dimensional microchannel network is created based on the bimodal nanocomposite material; The sensor array is packaged and integrated to obtain a multi-modal gas detection module, and based on the multi-modal gas detection module, resistance signals and capacitance signals are synchronously collected and digitally filtered to obtain pre-processed multi-modal signal data; Collecting and analyzing the temperature sensor and humidity sensor data in the multi-modal gas detection module to obtain environmental factor compensation parameters; Performing adaptive calibration processing on the preprocessed multimodal signal data according to the environmental factor compensation parameter to obtain calibrated multimodal signal data; Building a multi-input neural network model based on the calibrated multimodal signal data and training it to obtain a gas recognition model; Based on the gas identification model, the multimodal signals collected and calibrated in real time are analyzed and judged to obtain gas type and concentration information.

[0006] In a second aspect, the present invention provides a detection device based on a multimodal gas sensor, the detection device based on the multimodal gas sensor comprising: A creation module is used to perform a co-compounding process on a metal oxide semiconductor and a conductive material to obtain a bimodal nanocomposite material, and to create a sensor array having a three-dimensional microchannel network based on the bimodal nanocomposite material; An acquisition module is used to package and integrate the sensor array to obtain a multi-modal gas detection module, and to synchronously acquire and digitally filter the resistance signal and the capacitance signal based on the multi-modal gas detection module to obtain pre-processed multi-modal signal data; An analysis module, used to collect and analyze data from the temperature sensor and humidity sensor in the multi-modal gas detection module to obtain environmental factor compensation parameters; A calibration module, configured to perform adaptive calibration processing on the pre-processed multimodal signal data according to the environmental factor compensation parameter to obtain calibrated multimodal signal data; A training module, used to construct a multi-input neural network model based on the calibrated multimodal signal data and perform training to obtain a gas recognition model; The processing module is used to analyze and judge the multi-modal signals collected and calibrated in real time based on the gas recognition model to obtain gas type and concentration information.

[0007] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned detection method based on the multimodal gas sensor.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned detection method based on a multi-modal gas sensor.

[0009] In the technical solution provided by the present invention, the design and preparation technology of the dual-mode nanocomposite material is to dope SnO with Fe 2The synergistic effect of nanoparticles and graphene significantly improves the gas-sensitive performance and selectivity of the sensing material. The design and manufacture of the three-dimensional microchannel network structure increases the contact area between the gas and the sensitive material, improves the response speed and sensitivity of the sensor, and is conducive to the generation and transmission of multimodal signals. The application of temperature and humidity compensation technology effectively reduces the impact of environmental changes on the detection results through the dynamic environmental factor compensation model, and improves the stability and reliability of the sensor in complex environments. The adaptive calibration processing technology, combined with methods such as wavelet packet transform and empirical mode decomposition, realizes the precise calibration of multimodal signals, effectively improving the signal-to-noise ratio and feature expression ability of the signal. The design of the multi-input neural network model integrates advanced algorithms such as multi-scale residual networks, bidirectional gated recurrent unit networks, and multi-head self-attention mechanisms, significantly improving the recognition ability of complex gases and the accuracy of concentration estimation. The adoption of multi-task learning strategies optimizes gas classification and concentration regression tasks at the same time, improves the generalization ability and robustness of the model, and enables it to better cope with various complex situations in practical applications. The present invention realizes high-precision and high-reliability gas detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 Schematic diagram of the steps of a detection method based on a multi-modal gas sensor in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a detection device based on a multi-modal gas sensor in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The embodiments of the present invention provide a detection method and device based on a multimodal gas sensor. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of a detection method based on a multi-modal gas sensor in an embodiment of the present invention includes: Step S1, performing a co-compounding process on a metal oxide semiconductor and a conductive material to obtain a bimodal nanocomposite material, and creating a sensor array with a three-dimensional microchannel network based on the bimodal nanocomposite material; It is understandable that the execution subject of the present invention may be a detection device based on a multi-modal gas sensor, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0014] Specifically, the tin source and the iron source are mixed in a preset molar ratio to obtain a metal precursor solution, and the solution is placed in a specific reaction container, and subjected to a hydrothermal reaction treatment so that the metal ions react under high temperature and high pressure conditions to form Fe-doped SnO 2 Nanoparticles. Tin oxide nanoparticles doped with iron have excellent gas-sensitive properties and are one of the basic materials for forming dual-modal nanocomposites. Graphene is subjected to ultrasonic dispersion treatment to obtain a uniform graphene suspension. The Fe-doped SnO 2 The nanoparticles and graphene suspension are fully mixed in a preset mass ratio to form a composite nanomaterial suspension with good electrical conductivity and gas sensing properties. The composite nanomaterial suspension is treated by spray drying technology to rapidly evaporate the solvent under the action of hot air to form a uniform granular substance, which is then calcined to remove possible residual organic matter and enhance the crystallinity of the material to obtain a dual-modal nanocomposite material with stable performance. The silicon substrate is oxidized to form a uniform layer of SiO on its surface. 2 Insulation layer to improve the electrical stability and isolation effect of the sensor. 2The substrate of the insulating layer is subjected to photolithography and metal deposition processing. By this method, a designed comb electrode pattern is formed on the surface of the substrate. The electrode structure helps to enhance the electrical response of the sensor. After the electrode is prepared, the substrate with the comb electrode is processed using laser direct writing technology to form a complex three-dimensional microchannel network structure on the substrate surface. The three-dimensional structure can significantly improve the diffusion efficiency of the gas and the sensitivity of the sensor. The prepared bimodal nanocomposite material is dispersed again to obtain a nanomaterial suspension. The suspension is accurately deposited on the substrate structure with a three-dimensional microchannel network by inkjet printing technology to achieve precise positioning and uniform distribution of the material. The substrate structure deposited with the nanomaterial is heat treated to remove possible residual solvents and enhance the adhesion and stability of the material to form a sensor array with a three-dimensional microchannel network.

[0015] Step S2, packaging and integrating the sensor array to obtain a multimodal gas detection module, and synchronously collecting and digitally filtering the resistance signal and the capacitance signal based on the multimodal gas detection module to obtain pre-processed multimodal signal data; Specifically, the low-temperature co-fired ceramic material is precisely processed to manufacture a sensor housing with preset air inlets and outlets to ensure that the gas can flow evenly on the sensor array and improve the accuracy of detection. The sensor array is fixed to the sensor housing, and a sensor assembly that is closely integrated with the sensor housing is formed through precise positioning and bonding processes to ensure its stability in complex environments. For the sensor assembly, ultrasonic bonding technology is used to connect the electrodes of the sensor array with external leads to form a fully functional sensor unit. At the same time, the temperature sensor and humidity sensor are integrated and installed to obtain a multi-modal gas detection module. Based on the multi-modal gas detection module, a micro air pump and gas distribution system are designed to ensure the uniform distribution of gas inside the sensor and improve the stability and accuracy of the sensor response. At the same time, in response to the electrical signal processing requirements of the sensor system, the signal conditioning circuit is designed and installed to construct a signal processing unit with preamplification, filtering and analog-to-digital conversion functions, which can effectively enhance weak signals and filter out interference. By programming and configuring the microcontroller, a data acquisition and communication control system is formed. The system realizes the synchronous acquisition of resistance signals and capacitance signals through reasonable software and hardware design, and ensures the synchronous acquisition and processing of multimodal signals through parallel data acquisition settings. Based on the synchronous acquisition scheme, the collected original signal is processed by Butterworth low-pass filtering to remove high-frequency noise, reduce interference components in the signal, and retain effective information. For the filtered signal, the wavelet transform technology is used to remove the baseline drift, effectively eliminating the baseline drift in the signal and obtaining more stable and accurate signal data. The pre-processed multimodal signal data is obtained.

[0016] Step S3, collecting and analyzing the temperature sensor and humidity sensor data in the multi-modal gas detection module to obtain environmental factor compensation parameters; Specifically, the temperature sensor and humidity sensor are sampled to obtain the original temperature and humidity time series data, which reflects the dynamic changes of the ambient temperature and humidity over time. The original temperature and humidity time series data are processed by median filtering to remove the outliers. The filtering process can effectively eliminate the data deviation caused by occasional interference or sensor noise, and obtain smoother and more accurate temperature and humidity data. Fast Fourier transform is performed on the temperature and humidity data without outliers to obtain the spectrum characteristics of the temperature and humidity data. By analyzing the spectrum characteristics, the main periodic components and characteristic frequencies in the temperature and humidity changes are identified, and the target mode of the temperature and humidity changes is extracted. The target mode reflects the main law of the change of temperature and humidity over time. Based on the target mode, an autoregressive moving average model is constructed to predict the future trend of temperature and humidity. Kalman filtering is performed on the constructed temperature and humidity change trend prediction model to obtain the optimized temperature and humidity estimation value. Kalman filtering is a linear optimal estimation method that can effectively reduce the noise influence in the prediction and improve the reliability of the estimation result. Based on the optimized temperature and humidity estimation value, the influence coefficient of temperature and humidity on the sensitivity of the gas sensor is calculated to obtain the initial compensation factor, thereby correcting the influence of the ambient temperature and humidity changes on the detection accuracy of the gas sensor. The initial compensation factor is fitted with nonlinear least squares to obtain a more accurate temperature and humidity compensation function. Based on the temperature and humidity compensation function, an adaptive fuzzy neural network is constructed to form a dynamic environmental factor compensation model. The compensation model can automatically adjust the compensation parameters according to the real-time temperature and humidity changes to ensure the accuracy of gas detection. The dynamic environmental factor compensation model is optimized and the optimal environmental factor compensation parameters are obtained through multiple iterations and adjustments.

[0017] Step S4, performing adaptive calibration processing on the pre-processed multimodal signal data according to the environmental factor compensation parameter to obtain calibrated multimodal signal data; Specifically, the preprocessed multimodal signal data is segmented into time windows and divided into multiple fixed-length signal segments to ensure that the signal maintains temporal consistency during the analysis process. Based on the environmental factor compensation parameters, the specific effects of temperature and humidity on each signal segment are calculated to obtain an environmental impact matrix that reflects the degree of influence of environmental changes on the signal. The environmental impact matrix is ​​decomposed using the singular value decomposition technique to extract the target influence pattern of environmental factors. The target influence pattern can identify the main influence direction and characteristics of environmental factors on the signal. Based on the target influence pattern, an environmental noise estimation model is constructed, which filters out the interference of environmental noise on the signal so that the effective components of the signal are retained. In order to extract useful information from the signal, the signal segments and the environmental noise estimation model are transformed by wavelet packets to obtain multi-scale time-frequency features. The time-frequency features can not only describe the time and frequency changes of the signal, but also reveal the correlation between the signal and the environmental noise at different scales. By calculating the mutual information between the multi-scale time-frequency features, the correlation index between the signal and the environmental noise is obtained to quantify how much information in the signal is interfered by the environmental noise. According to the correlation index, the signal fragments are subjected to adaptive threshold denoising to remove the parts that are more affected by environmental noise, and the preliminary calibrated signal data is obtained. The preliminary calibrated signal data is subjected to empirical mode decomposition to separate the intrinsic mode functions of different frequency components. Each intrinsic mode function represents the main change trend of the signal in a certain frequency band, which can help identify the essential characteristics of the signal. The Hilbert-Huang transform spectrum is constructed based on the intrinsic mode function, and the instantaneous frequency and amplitude of the signal are extracted from it to reflect the dynamic changes of the signal. The instantaneous frequency and amplitude are nonlinearly mapped and reconstructed to correct the signal distortion caused by environmental factors and obtain the calibrated multimodal signal data.

[0018] Step S5: constructing a multi-input neural network model based on the calibrated multimodal signal data and training it to obtain a gas recognition model; Specifically, the time-frequency domain feature extraction is performed on the calibrated multimodal signal data to obtain the feature sets of the resistance signal and the capacitance signal. Based on the extracted resistance signal feature set, a first multiscale residual network branch is constructed. This branch consists of three parallel residual blocks, each of which contains convolutional layers and pooling layers of different scales. Through multi-scale processing, the feature information of different frequency bands and time domains in the resistance signal can be captured to obtain a trained multi-scale resistance feature representation. Similarly, for the capacitance signal feature set, a second multiscale residual network branch is constructed. The structure of this branch is the same as that of the resistance signal branch, which can effectively extract the multi-scale features of the capacitance signal and obtain a trained multi-scale capacitance feature representation. In order to optimize and enhance the feature expression ability, the trained multi-scale resistance feature representation and multi-scale capacitance feature representation are adaptively recalibrated. By introducing the channel attention mechanism and the spatial attention mechanism, the features are weighted, the importance of different features is automatically adjusted, and the training recalibration features are obtained, which effectively highlight the key features for gas identification, while suppressing redundant information and improving the overall expressiveness of the model. A bidirectional gated recurrent unit network is constructed based on the training recalibration features to mine the timing information in the signal. The network consists of two layers, each with 128 units. It uses the tanh activation function and can capture the contextual information of the signal in the time series through bidirectional processing to obtain the training time series context features. A multi-head self-attention mechanism is applied to the training time series context features. The mechanism contains 8 attention heads, each with a dimension of 64. In this way, the model can understand the dependencies between features from a global perspective and obtain the training global dependency features. A pyramid pooling module is constructed based on the training global dependency features. Through pooling operations of three different scales, 1x1, 2x2, and 4x4, multi-scale global information is extracted to obtain the training multi-scale global features. The global features are fused, and the multi-scale features are further integrated through element-by-element addition and 1x1 convolution operations to obtain the training fusion feature vector. A multi-task learning head is constructed based on the training fusion feature vector to simultaneously process gas classification and concentration regression tasks. For the gas classification task, the Softmax activation function is used to obtain the probability distribution of gas types; for the concentration regression task, the linear activation function is used to output the estimated value of the gas concentration. Through multi-task learning, the model can handle classification and regression problems at the same time, improving the accuracy of gas type and concentration recognition. The entire multi-input neural network model is trained end-to-end, using the weighted cross entropy loss function and Huber loss function as optimization targets. The loss function can balance the importance of classification and regression tasks, and the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to update parameters, and finally a trained gas recognition model is obtained.

[0019] Step S6: Analyze and judge the multimodal signals collected and calibrated in real time based on the gas recognition model to obtain gas type and concentration information.

[0020] Specifically, the resistance signal and capacitance signal collected in real time are synchronously sampled and digitized to obtain the original multimodal signal data, which reflects the real-time response of the sensor to the ambient gas. The original multimodal signal data is adaptively calibrated to eliminate the influence of environmental factors and obtain the calibrated real-time multimodal signal. The time-frequency domain feature extraction is performed on the calibrated real-time multimodal signal, and the real-time resistance feature set and the real-time capacitance feature set are extracted therefrom respectively. The real-time resistance feature set is input into the first multiscale residual network branch of the gas recognition model, and the real-time multiscale resistance feature representation is obtained by calculating the three parallel residual blocks in the branch. Similarly, the real-time capacitance feature set is input into the second multiscale residual network branch of the gas recognition model, and the real-time multiscale capacitance feature representation is obtained by calculating the three parallel residual blocks of the same structure. The real-time multiscale resistance feature representation and the real-time multiscale capacitance feature representation are adaptively recalibrated. Through the weighted processing of the channel attention mechanism and the spatial attention mechanism, the recalibrated feature representation is obtained, so that important features are highlighted, irrelevant features are suppressed, and the sensitivity of the model to key information is improved. The recalibrated features are input into a bidirectional gated recurrent unit (GRU) network. Through two layers, each containing 128 units, the contextual relationship of the signal in the time series is captured to obtain the real-time temporal contextual features. A multi-head self-attention mechanism is applied to the real-time temporal contextual features. Through the parallel calculation of 8 attention heads, the model can identify the complex dependencies between signals and obtain the real-time global dependency features. The global dependency features are input into the pyramid pooling module. Through pooling operations of three different scales, 1x1, 2x2 and 4x4, the multi-scale global information of the signal is extracted to obtain the real-time multi-scale global features. The real-time multi-scale global features are subjected to feature fusion operations. Through element-wise addition and 1x1 convolution operations, the information of different scales is integrated to obtain the real-time fused feature vector. The real-time fused feature vector is input into the multi-task learning head for parallel gas classification and concentration regression calculations. For the gas classification task, the Softmax activation function is used to obtain the probability distribution of gas types; for the concentration regression task, the linear activation function is used to output the estimated value of gas concentration. The model can provide gas type identification and concentration estimation at the same time, meeting the needs of multi-modal gas detection. Threshold judgment and time series smoothing are performed on the real-time gas type probability distribution and concentration estimation values. Through threshold judgment, low probability or unreliable results are filtered out, and through time series smoothing, outliers caused by instantaneous fluctuations are eliminated, and finally stable and reliable gas type and concentration information is obtained.

[0021] In the embodiment of the present invention, the design and preparation technology of the dual-mode nanocomposite material is to dope SnO with Fe 2 The synergistic effect of nanoparticles and graphene significantly improves the gas-sensitive performance and selectivity of the sensing material. The design and manufacture of the three-dimensional microchannel network structure increases the contact area between the gas and the sensitive material, improves the response speed and sensitivity of the sensor, and is conducive to the generation and transmission of multimodal signals. The application of temperature and humidity compensation technology effectively reduces the impact of environmental changes on the detection results through the dynamic environmental factor compensation model, and improves the stability and reliability of the sensor in complex environments. The adaptive calibration processing technology, combined with methods such as wavelet packet transform and empirical mode decomposition, realizes the precise calibration of multimodal signals, effectively improving the signal-to-noise ratio and feature expression ability of the signal. The design of the multi-input neural network model integrates advanced algorithms such as multi-scale residual networks, bidirectional gated recurrent unit networks, and multi-head self-attention mechanisms, significantly improving the recognition ability of complex gases and the accuracy of concentration estimation. The adoption of multi-task learning strategies optimizes gas classification and concentration regression tasks at the same time, improves the generalization ability and robustness of the model, and enables it to better cope with various complex situations in practical applications. The present invention realizes high-precision and high-reliability gas detection.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: The tin source and the iron source are mixed according to a preset molar ratio to obtain a metal precursor solution, and the metal precursor solution is subjected to a hydrothermal reaction to obtain Fe-doped SnO 2 Nanoparticles; The graphene was subjected to ultrasonic dispersion treatment to obtain a graphene suspension, and Fe-doped SnO 2 The nanoparticles and the graphene suspension are mixed according to a preset mass ratio to obtain a composite nanomaterial suspension; The composite nanomaterial suspension is spray dried and calcined to obtain a dual-modal nanocomposite material, and the silicon substrate is oxidized to obtain a SiO 2 The base of the insulating layer; Based on SiO 2 The substrate of the insulating layer is subjected to photolithography and metal deposition processing to obtain a substrate with comb-shaped electrodes, and the substrate with the comb-shaped electrodes is subjected to laser direct writing processing to obtain a substrate structure with a three-dimensional microchannel network; The bimodal nanocomposite material is dispersed to obtain a nanomaterial suspension, and the nanomaterial suspension is deposited on a base structure with a three-dimensional microchannel network by inkjet printing technology. The base structure with the deposited nanomaterial is heat treated to obtain a sensor array with a three-dimensional microchannel network.

[0023] Specifically, the tin source (Sn) and the iron source (Fe) are accurately proportioned. Usually, stannous oxide (SnO) and ferric chloride (FeCl 3 ) as the tin source and iron source. In actual operation, SnO and FeCl 3 Dissolve them separately in deionized water to obtain their own solutions. By mixing the two solutions, a metal precursor solution is formed, which contains Sn²⁺ and Fe³⁺ ions. In chemical reactions, the precise control of the molar ratio directly affects the Fe doping concentration in the final product. Assuming the molar ratio used is x:1, then x represents the number of moles of SnO, and 1 represents the number of moles of FeCl 3 Controlling the appropriate molar ratio can ensure that Fe 2 The metal precursor solution is placed in an autoclave for hydrothermal reaction. The hydrothermal reaction is usually carried out at a temperature of 180°C to 200°C. Under this condition, Sn²⁺ and Fe³⁺ ions are hydrolyzed and precipitated in the solution to form Fe-doped SnO 2 Nanoparticles. The process of hydrothermal reaction can be expressed as the following chemical reaction equation: ; Among them, Sn²⁺ and Fe³⁺ are the metal ions in the reaction, H 2 O is the reaction medium, and the final product is Fe-SnO 2 Fe-doped SnO 2 Nanoparticles. Due to the doping of Fe, the gas-sensing performance of the material can be significantly improved, especially in the detection of specific gases such as nitrogen dioxide (NO 2 ) showed higher sensitivity. 2 After the nanoparticles, the conductivity and mechanical strength of the material need to be further improved, and graphene is used as another part of the composite material. Graphene is known for its excellent conductivity and flexibility. In actual operation, the graphene powder is evenly dispersed in the solution through ultrasonic dispersion treatment to form a graphene suspension. 2 Nanoparticles and graphene suspension are mixed in a preset mass ratio to form a composite nanomaterial suspension. The setting of the mass ratio determines the ratio of graphene to Fe-SnO in the final composite material. 2By optimizing this ratio, the conductivity and mechanical properties of the material are adjusted, so that it maintains good gas sensing performance while also having excellent conductivity and structural stability. The mixed suspension is treated by spray drying technology to evaporate the solution quickly, leaving a uniform granular substance, which is then calcined to remove residual organic matter and enhance the crystallinity and stability of the material to form a bimodal nanocomposite material with good physical and chemical properties. In order to construct the base structure of the gas sensor, the silicon substrate is oxidized. This process generates a uniform layer of SiO on the surface of the silicon substrate through an oxidation reaction at high temperature. 2 Insulation layer. This insulation layer can not only effectively isolate the current and prevent current leakage, but also provide good substrate conditions in subsequent photolithography and metal deposition processes. 2 Photolithography and metal deposition are performed on the substrate of the insulating layer. Through this process, a precise comb-shaped electrode structure is formed on the surface of the substrate. The comb-shaped electrode can significantly increase the contact area between the electrode and the gas, and improve the response speed and sensitivity of the sensor. After the electrode structure is formed, a three-dimensional microchannel network is constructed on the substrate using laser direct writing technology. This network structure allows the gas to contact the sensor surface more fully, improving the efficiency of gas diffusion and the response performance of the sensor. The bimodal nanocomposite prepared in the early stage is dispersed again to form a nanomaterial suspension. The suspension is precisely deposited on the substrate structure with a three-dimensional microchannel network through inkjet printing technology. The advantage of inkjet printing technology is its high precision and low material consumption, which enables nanomaterials to be uniformly deposited on complex three-dimensional structures to form a continuous and uniform active layer. The substrate structure deposited with nanomaterials is heat treated to enhance the adhesion and stability of the material and remove any possible residual solvents and impurities. The heat-treated structure forms a sensor array with a three-dimensional microchannel network. This sensor array has extremely high gas sensitivity and selectivity, can accurately detect specific gas components in the environment, and provide timely feedback based on changes in gas concentration.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: The low temperature co-fired ceramic material is precisely processed to obtain a sensor housing with a preset air inlet and air outlet, and the sensor array is fixed to obtain a sensor assembly combined with the sensor housing; Ultrasonic bonding is performed based on the sensor assembly to obtain a sensor unit in which electrodes are connected to external leads, and a temperature sensor and a humidity sensor are integrated and installed to obtain a multi-modal gas detection module; Based on the multi-modal gas detection module, a micro air pump and gas distribution system are designed to obtain a sensor system with uniform gas distribution. The signal conditioning circuit of the sensor system is designed and installed to obtain a signal processing unit with preamplification, filtering and analog-to-digital conversion functions. Programming and configuring the microcontroller based on the signal processing unit to obtain a data acquisition and communication control system, and setting the data acquisition and communication control system for parallel data acquisition to obtain a synchronous acquisition scheme for resistance signals and capacitance signals; Based on the synchronous acquisition scheme, the collected original signal is processed by Butterworth low-pass filtering to obtain a signal with high-frequency noise removed, and the signal with high-frequency noise removed is processed by wavelet transform to remove baseline drift to obtain pre-processed multimodal signal data.

[0025] Specifically, the low-temperature co-fired ceramic material is precisely processed. Through high-precision machining equipment, a sensor housing with a specific geometric shape is cut out on the low-temperature co-fired ceramic material according to preset design parameters, and precise air inlet and air outlet are designed on the housing. The pre-prepared sensor array is fixed to the inside of the housing. The sensor array usually includes multiple sensing units, each of which can respond to a specific gas component. In order to ensure the close combination of the sensor array and the housing, a precision fixing process such as dispensing or welding is usually used to firmly mount the array at a specified position of the housing. Ultrasonic bonding is performed based on the sensor assembly. Ultrasonic bonding is an efficient welding technology that uses high-frequency ultrasonic vibration energy to locally heat and melt the contact surface under pressure to achieve a firm metal connection. In this process, the electrodes on the sensor array are precisely connected to the external leads through ultrasonic bonding technology to form a complete sensor unit. The unit not only includes a gas sensor, but also integrates a temperature sensor and a humidity sensor to achieve multimodal detection functions. A micro air pump and gas distribution system are designed to ensure uniform distribution of gas on the sensor surface. The design of the micro air pump needs to consider parameters such as gas flow rate, pressure and stability, while the gas distribution system needs to ensure that the gas can flow evenly through each sensor unit through reasonable pipeline design and flow control. Through this system, the sensor can obtain stable and consistent gas samples to avoid detection errors caused by uneven airflow. Design and install the signal conditioning circuit for the sensor system. The signal conditioning circuit is the core part of the entire detection system, and its main functions include signal preamplification, filtering and analog-to-digital conversion. In the sensor array, gas sensors usually output weak electrical signals, which need to be amplified by a preamplifier. The formula is as follows: ; in, represents the output voltage after amplification, Represents the original voltage signal output by the sensor, is the gain of the amplifier, which is usually a constant greater than 1. The amplified signal will pass through the filter, and a Butterworth low-pass filter will be used to remove high-frequency noise and retain the useful components in the signal. The Butterworth filter can effectively remove high-frequency interference without introducing additional phase distortion. The filtered signal is converted into a digital signal through an analog-to-digital converter. After completing the design and installation of the hardware system, the microcontroller is programmed and configured based on the signal processing unit to build a data acquisition and communication control system. The microcontroller is the control core of the system, responsible for coordinating the data acquisition, signal processing and data transmission of the sensor array. When programming, a parallel data acquisition scheme needs to be set to ensure that the resistance signal and the capacitance signal can be collected and processed synchronously. Through reasonable programming and scheduling, the microcontroller can efficiently manage the acquisition process of multimodal signals to avoid data loss and delay. The collected original signal is processed by a Butterworth low-pass filter, and the formula is as follows: ; in, is the transfer function of the filter, is the cut-off frequency, is the order of the filter. The function of this filter is to remove high-frequency noise in the signal, retain the effective information of the low-frequency part, and obtain a smoother signal. The signal with high-frequency noise removed is subjected to wavelet transform to remove baseline drift. Wavelet transform is a multi-scale analysis tool that can effectively separate the useful components and baseline drift in the signal to obtain a more accurate signal representation. After wavelet transform processing, the pre-processed multimodal signal data is obtained.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: The temperature sensor and humidity sensor are sampled to obtain the original temperature and humidity time series data, and the original temperature and humidity time series data are processed by median filtering to obtain the temperature and humidity data with outliers removed; Perform fast Fourier transform on the temperature and humidity data with outliers removed to obtain the frequency spectrum characteristics of the temperature and humidity data, and analyze the frequency spectrum characteristics of the temperature and humidity data to obtain the target mode of temperature and humidity changes; Based on the target pattern of temperature and humidity changes, an autoregressive moving average model is constructed to obtain a temperature and humidity change trend prediction model, and the temperature and humidity change trend prediction model is processed by Kalman filtering to obtain optimized temperature and humidity estimation values; Based on the optimized estimated values ​​of temperature and humidity, the influence coefficient of temperature and humidity on the sensitivity of the gas sensor is calculated to obtain the initial compensation factor, and the initial compensation factor is fitted by nonlinear least squares to obtain the temperature and humidity compensation function; An adaptive fuzzy neural network is constructed based on the temperature and humidity compensation function to obtain a dynamic environmental factor compensation model. The parameters of the dynamic environmental factor compensation model are optimized to obtain environmental factor compensation parameters.

[0027] Specifically, the temperature sensor and humidity sensor are sampled to obtain the original temperature and humidity time series data. The original temperature and humidity time series data is processed by median filtering. Median filtering is a commonly used denoising method. Its principle is to sort the data in a window by size, and then take the middle value as the output value of the window. This process can effectively remove sharp noise and outliers while retaining the main characteristics of the signal to obtain temperature and humidity data without outliers. Fast Fourier transform is performed based on the temperature and humidity data without outliers to obtain the spectral characteristics of the data. Fourier transform is a mathematical tool that converts time domain signals into frequency domain signals. Fast Fourier transform can be used to analyze the energy distribution of temperature and humidity data at different frequencies to reveal the periodic characteristics of temperature and humidity changes. The spectral characteristics of temperature and humidity data can be expressed by the following formula: ; in, Indicates temperature and humidity data at frequency The amplitude of Indicates Temperature and humidity data at a time point, is the total length of the data. By analyzing the spectral characteristics of the temperature and humidity data, the main period and amplitude of the temperature and humidity changes are identified, and the target pattern of temperature and humidity changes is obtained, reflecting the changing laws of temperature and humidity on different time scales. Based on the target pattern of temperature and humidity changes, an autoregressive moving average (ARMA) model is constructed to predict the future trend of temperature and humidity changes. The ARMA model is a classic statistical model for time series prediction. It can model the linear relationship in the time series by combining autoregressive (AR) and moving average (MA) components. The expression of the ARMA model is: ; in, Indicates time The temperature and humidity values ​​at is a constant term, and are the autoregressive coefficient and the moving average coefficient, is the white noise error term, and are the orders of autoregression and moving average, respectively. By training the ARMA model, a prediction model for temperature and humidity changes is obtained. The output of the ARMA model is processed by Kalman filtering. Kalman filtering is a recursive estimation algorithm that dynamically adjusts the estimation error by weighting the model prediction value and the actual observation value to obtain the optimized temperature and humidity estimation value. The core formula of Kalman filtering is: ; in, is the estimated value of the state at the current moment, is the predicted value at the previous moment, is the Kalman gain, is the current observation value, is the observation matrix. The calculation of the Kalman gain ensures the balance of the estimated value to the actual observed value and the model predicted value, so as to obtain a more accurate temperature and humidity estimate in a noisy environment. Based on the optimized temperature and humidity estimates, the influence coefficient of temperature and humidity on the sensitivity of the gas sensor is calculated to obtain the initial compensation factor. The relationship between the temperature and humidity data and the output of the gas sensor is analyzed, and the influence of temperature and humidity changes on the response of the gas sensor is fitted through experimental data. The initial compensation factor is used to adjust the output of the gas sensor so that it can maintain a high detection accuracy under different environmental conditions. The initial compensation factor is fitted by nonlinear least squares to obtain the temperature and humidity compensation function. The nonlinear least squares method is a numerical optimization method that finds the best fitting parameters by minimizing the sum of squares of the error between the actual value and the fitted value. The temperature and humidity compensation function can be expressed as: ; in, is the gas sensor sensitivity after temperature and humidity compensation, is the temperature, For humidity, are fitting parameters. By optimizing these parameters, an accurate compensation model is obtained, which makes the response of the gas sensor more consistent under different temperature and humidity conditions. Based on the temperature and humidity compensation function, an adaptive fuzzy neural network is constructed to establish a dynamic environmental factor compensation model. The adaptive fuzzy neural network combines the advantages of fuzzy logic and neural networks. It can handle nonlinear problems through fuzzy rules and adjust parameters by using the learning ability of neural networks. By training the adaptive fuzzy neural network model, real-time compensation of the gas sensor response is achieved, so that it can maintain high-precision gas detection performance under different environmental conditions. The parameters of the dynamic environmental factor compensation model are optimized to obtain the optimal environmental factor compensation parameters. Through iterative optimization and cross-validation, it is ensured that the compensation model can operate stably under various environmental conditions, and finally achieve effective compensation for the influence of temperature and humidity, and improve the detection accuracy and reliability of the gas sensor.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: The preprocessed multimodal signal data is segmented into time windows to obtain signal segments of fixed length, and the influence coefficients of temperature and humidity on each signal segment are calculated based on the environmental factor compensation parameters to obtain the environmental influence matrix; Performing singular value decomposition on the environmental impact matrix to obtain the target impact pattern of the environmental factors, and constructing an adaptive filter based on the target impact pattern of the environmental factors to obtain an environmental noise estimation model; Perform wavelet packet transform on the signal fragments and the environmental noise estimation model to obtain multi-scale time-frequency features, and calculate the mutual information based on the multi-scale time-frequency features to obtain the correlation index between the signal and the environmental noise; Adaptively threshold denoising is performed on the signal segments according to the correlation index to obtain preliminary calibrated signal data, and empirical mode decomposition is performed on the preliminary calibrated signal data to obtain an eigenmode function; The Hilbert-Huang transform spectrum is constructed based on the eigenmode function to obtain the instantaneous frequency and amplitude of the signal, and the instantaneous frequency and amplitude are nonlinearly mapped and reconstructed to obtain the calibrated multimodal signal data.

[0029] Specifically, the preprocessed multimodal signal data is segmented into time windows. The continuous signal data is divided into signal segments of fixed length to maintain the time consistency of the signal and the comparability of the features during the analysis process. For example, assuming that the sampling frequency of the multimodal signal data is 100 Hz, and the time window length of each signal segment is 1 second, then each signal segment will contain 100 data points. Based on the environmental factor compensation parameters, the influence coefficients of temperature and humidity on each signal segment are calculated, the impact of environmental conditions on the signal is quantified, and an environmental impact matrix is ​​constructed. The elements of the environmental impact matrix represent the degree of correction of the signal characteristics by temperature and humidity within a specific time window. Assuming that the environmental impact matrix is ​​a two-dimensional matrix, in which rows represent different time windows and columns represent different signal features, each element in the matrix can be expressed by the following formula: ; in, Indicates The signal characteristics are The environmental impact value within a time window, and are the average temperature and humidity in the time window, and are the corresponding temperature and humidity influence coefficients. By calculating the signal fragments of all time windows, a complete environmental influence matrix is ​​constructed. Singular value decomposition is performed on the environmental influence matrix to extract the main influence mode of environmental factors. Singular value decomposition is a matrix decomposition technique that can decompose a matrix into three parts, namely a left singular matrix, a diagonal matrix, and a right singular matrix. Through singular value decomposition, the main components of the matrix are extracted, that is, the main influence mode of environmental factors on signal characteristics. Assume that the environmental influence matrix is , then the singular value decomposition can be performed as follows: ; in, and are the left singular matrix and the right singular matrix respectively, is a diagonal matrix, and the elements on the diagonal are singular values, which represent the importance of different modes. By analyzing the size of the singular values, we can determine which modes are the main environmental impact modes, and use these modes for the subsequent adaptive filter construction. Based on the target impact mode of environmental factors, an adaptive filter is constructed to generate an environmental noise estimation model. The adaptive filter can adjust its parameters in real time according to the changes in the input signal, thereby effectively removing or suppressing the impact of environmental noise on the signal. The environmental noise estimation model is continuously updated through the adaptive filter to reflect the changing characteristics of noise under different environmental conditions. The signal fragments and the environmental noise estimation model are transformed by wavelet packet to extract multi-scale time-frequency features. Wavelet packet transform is an enhanced wavelet transform method that can provide finer frequency band division, thereby better capturing the time-frequency characteristics of the signal. Through wavelet packet transform, the signal can be decomposed into time-frequency representations at different scales, which not only reflect the local characteristics of the signal, but also reveal the correlation between the signal and the environmental noise. In order to quantify this correlation, mutual information is calculated based on multi-scale time-frequency features. Mutual information is a measure of the degree of dependence between two random variables. By calculating the mutual information between the signal and the environmental noise, their correlation index is obtained. According to the correlation index, the signal fragments are subjected to adaptive threshold denoising. The adaptive threshold denoising method can automatically select the optimal threshold, remove the part of the signal that is highly correlated with the environmental noise, retain the useful information components, and obtain the preliminary calibrated signal data. Perform empirical mode decomposition (EMD) on the preliminary calibrated signal data. EMD is an adaptive decomposition method that can decompose complex signals into several intrinsic mode functions (IMFs), which represent different frequency components of the signal. Each IMF corresponds to an independent oscillation mode and can reveal the intrinsic structure and dynamic changes of the signal. Construct the Hilbert-Huang transform spectrum based on IMF. The Hilbert-Huang transform is a time-frequency analysis method that obtains the instantaneous frequency and amplitude of the signal by performing a Hilbert transform on the IMF. Assume that an IMF is , then its Hilbert transform is: ; in, is the signal after Hilbert transformation, and PV represents principal value integral. By calculating the instantaneous frequency and amplitude, the Hilbert spectrum of the signal is generated to describe the dynamic change characteristics of the signal. Nonlinear mapping and reconstruction are performed on the instantaneous frequency and amplitude. Nonlinear mapping can transform the complex patterns of the signal into a form that is easier to analyze, while reconstruction integrates these patterns into calibrated multimodal signal data.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Performing time-frequency domain feature extraction on the calibrated multimodal signal data to obtain a resistance signal feature set and a capacitance signal feature set; A first multi-scale residual network branch is constructed based on the resistance signal feature set, wherein the first multi-scale residual network branch includes three parallel residual blocks, each residual block includes convolutional layers and pooling layers of different scales, and a training multi-scale resistance feature representation is obtained; A second multi-scale residual network branch is constructed based on the capacitance signal feature set, wherein the structure of the second multi-scale residual network branch is the same as that of the first multi-scale residual network branch, and a training multi-scale capacitance feature representation is obtained; Adaptively recalibrate the training multi-scale resistance feature representation and multi-scale capacitance feature representation, and weight the features through the channel attention mechanism and the spatial attention mechanism to obtain the training recalibrated features; A bidirectional gated recurrent unit network is constructed based on the training recalibration features. The bidirectional gated recurrent unit network consists of 2 layers, each with 128 units. The tanh activation function is used to obtain the training temporal context features. A multi-head self-attention mechanism is applied to the training temporal context features. The multi-head self-attention mechanism contains 8 attention heads, each with a dimension of 64, to obtain the training global dependency features; A pyramid pooling module is constructed based on the training global dependency features. Multi-scale global information is extracted through pooling operations of three different scales: 1x1, 2x2, and 4x4, to obtain the training multi-scale global features. Perform feature fusion on the training multi-scale global features, and obtain the training fusion feature vector through element-by-element addition and 1x1 convolution operation; A multi-task learning head is constructed based on the training fusion feature vector, including gas classification task and concentration regression task. The gas classification task uses the Softmax activation function, and the concentration regression task uses the linear activation function to obtain the probability distribution of the training gas types and the concentration estimation value. The multi-input neural network model is trained end-to-end, and the weighted cross entropy loss function and Huber loss function are used as optimization objectives. The AdamW optimizer and cosine annealing learning rate scheduling strategy are used to update parameters to obtain the gas recognition model.

[0031] Specifically, the time-frequency domain feature extraction is performed on the calibrated multimodal signal data, the useful information in the original signal is separated, and the signal is characterized in two dimensions of time and frequency. For multimodal signal data, resistance signal and capacitance signal are two main input sources. Through time-frequency domain feature extraction, resistance signal feature sets are extracted from resistance signals, and capacitance signal feature sets are extracted from capacitance signals. These feature sets contain the change information of the signal at different time points and frequency ranges. The first multiscale residual network branch is constructed based on the resistance signal feature set. The multiscale residual network is a deep learning model that can capture information of different scales in the signal and is suitable for processing signal data with multi-level features. In the first multiscale residual network branch, three parallel residual blocks are designed, each residual block contains convolution layers and pooling layers of different scales to ensure that multiple features in the resistance signal can be captured. The function of the convolution layer is to extract local features of the signal through convolution operations, while the pooling layer compresses the features through downsampling operations, reduces the feature dimensions and retains important information. Through multi-scale processing, the network can generate a training multi-scale resistor feature representation, which represents the key information of the resistor signal at different scales. Similarly, a second multi-scale residual network branch is constructed based on the capacitor signal feature set. The structure of this branch is the same as the first multi-scale residual network branch, which contains three parallel residual blocks. The convolutional layers and pooling layers in each residual block are set to different scales to capture the multi-level features of the capacitor signal, and finally the training multi-scale capacitor feature representation is obtained. Adaptive feature recalibration is performed on the training multi-scale resistor feature representation and the multi-scale capacitor feature representation. Adaptive feature recalibration uses channel attention mechanism and spatial attention mechanism to weight the features, highlight important features, and suppress irrelevant or redundant features. The channel attention mechanism mainly targets the channel dimension of the feature and adjusts the importance of the feature by learning the weights of different channels; while the spatial attention mechanism weights the spatial dimension of the feature map so that the network can pay more attention to the key spatial positions. Through these operations, the training recalibrated features are obtained. Based on the training recalibrated features, a bidirectional gated recurrent unit (GRU) network is constructed. The bidirectional GRU network is a recurrent neural network that can capture contextual information in time series data. The network contains two layers, each layer consists of 128 units, and uses the tanh activation function. The output range of the tanh activation function is [-1, 1], which can effectively handle nonlinear relationships in the signal. Through bidirectional processing, the network can simultaneously consider the forward and backward information of the signal to obtain the training time series context features, which represent the global dependency of the signal on the time series. Based on the training time series context features, the multi-head self-attention mechanism is applied. The multi-head self-attention mechanism contains 8 attention heads, each with a dimension of 64. The attention mechanism automatically selects the most important information for the current task by calculating the correlation between different features.Through the design of multiple heads, the network can analyze and process features from multiple angles to obtain training global dependency features. A pyramid pooling module is constructed based on the training global dependency features. Pyramid pooling is a multi-scale pooling technology that extracts multi-scale global information of the signal through pooling operations of three different scales: 1x1, 2x2, and 4x4. The pooling operation can further compress the feature dimension and reduce the computational complexity while maintaining the spatial structure of the feature map. Through the pyramid pooling module, training multi-scale global features are obtained, which have stronger global information expression capabilities. Multi-scale global features are fused through element-by-element addition and 1x1 convolution operations. Element-by-element addition can add elements of different feature layers one by one, while the 1x1 convolution operation further compresses and integrates the features to obtain a training fusion feature vector. Based on the training fusion feature vector, a multi-task learning head is constructed. The multi-task learning head processes the gas classification task and the concentration regression task at the same time, where the gas classification task uses the Softmax activation function. The Softmax function can map the input features to a probability distribution to determine the type of gas; while the concentration regression task uses a linear activation function to estimate the concentration of the gas. The multi-task learning method not only improves the utilization efficiency of the model, but also enables the model to share features between different tasks and improve the overall performance. The entire multi-input neural network model is trained end-to-end. In order to optimize the model, the weighted cross entropy loss function and the Huber loss function are used as optimization targets. The weighted cross entropy loss function is used for classification tasks, and different weights are assigned to different categories to deal with the category imbalance problem; the Huber loss function is used for regression tasks, and its smooth error curve can show better robustness when dealing with outliers. The parameter update uses the AdamW optimizer, which combines the adaptive learning rate and weight decay mechanism of the Adam optimizer to effectively prevent overfitting. In order to improve the training effect of the model, the cosine annealing learning rate scheduling strategy is adopted. By gradually reducing the learning rate, the convergence process of the model is stabilized, and finally a high-precision gas recognition model is obtained.

[0032] In a specific embodiment, the process of executing step S6 may specifically include the following steps: The resistance signal and the capacitance signal collected in real time are synchronously sampled and digitally processed to obtain original multimodal signal data, and adaptive calibration processing is performed based on the original multimodal signal data to obtain a calibrated real-time multimodal signal; Performing time-frequency domain feature extraction on the calibrated real-time multimodal signal to obtain a real-time resistance feature set and a real-time capacitance feature set; The real-time resistance feature set is input into the first multi-scale residual network branch of the gas recognition model, and the real-time multi-scale resistance feature representation is obtained through three parallel residual blocks. The real-time capacitance feature set is input into the second multi-scale residual network branch of the gas recognition model, and is calculated through three parallel residual blocks to obtain the real-time multi-scale capacitance feature representation; Adaptively recalibrate the real-time multi-scale resistance feature representation and the real-time multi-scale capacitance feature representation to obtain a real-time recalibrated feature; The real-time recalibration features are input into a bidirectional gated recurrent unit network, and the real-time temporal context features are obtained through sequential calculation of 2 layers of 128 units; A multi-head self-attention mechanism is applied to the real-time temporal context features, and real-time global dependency features are obtained through parallel calculation of 8 attention heads; The real-time global dependency features are input into the pyramid pooling module, and the real-time multi-scale global features are obtained through pooling operations at three different scales. Perform feature fusion operation on real-time multi-scale global features to obtain real-time fused feature vectors; The real-time fused feature vector is input into the multi-task learning head, and the real-time gas type probability distribution and concentration estimation are obtained through parallel classification and regression calculations. The real-time gas type probability distribution and concentration estimation values ​​are subjected to threshold judgment and time series smoothing to obtain gas type and concentration information.

[0033] Specifically, the resistance signal and capacitance signal collected in real time are synchronously sampled and digitized. Synchronous sampling can avoid errors caused by time differences. The analog signal is converted into a digital signal through a high-precision analog-to-digital converter to obtain the original multimodal signal data. The original multimodal signal data is adaptively calibrated to eliminate the influence of environmental factors such as temperature and humidity on the signal and improve the accuracy of the signal. Through the pre-trained environmental compensation model, the deviation in the signal is adjusted in real time so that the calibrated real-time multimodal signal can more accurately reflect the response of the sensor to the target gas. The time-frequency domain features of the calibrated real-time multimodal signal are extracted. Through time-frequency analysis methods such as short-time Fourier transform or wavelet transform, the signal is decomposed into different time and frequency intervals to obtain more detailed feature information. A real-time resistance feature set is extracted from the resistance signal, and a real-time capacitance feature set is extracted from the capacitance signal. The real-time resistance feature set is input into the first multi-scale residual network branch of the gas recognition model. This branch consists of three parallel residual blocks, each of which contains convolutional layers and pooling layers of different scales, aiming to capture multi-level features in the resistance signal. The residual network effectively solves the problem of gradient vanishing in deep networks through skip connections. At the same time, convolution operations can extract local features, and pooling operations are used to compress feature maps to reduce computational complexity. Through multi-scale processing, the network can generate a real-time multi-scale resistance feature representation, which effectively describes the important patterns in the resistance signal. Similarly, the real-time capacitance feature set is input into the second multi-scale residual network branch of the gas recognition model. The structure of this branch is the same as that of the resistance signal branch. Through convolution and pooling operations of different scales, a real-time multi-scale capacitance feature representation is generated. Adaptive feature recalibration is performed on the real-time multi-scale resistance feature representation and the real-time multi-scale capacitance feature representation. Through the channel attention mechanism and the spatial attention mechanism, the features are weighted to highlight the most important features and suppress irrelevant or redundant features to obtain real-time recalibrated features. The channel attention mechanism dynamically adjusts the importance of each channel feature by learning the weights of different channels, while the spatial attention mechanism highlights the key spatial positions according to the response strength of different positions in the feature map. The real-time recalibrated features are input into the bidirectional gated recurrent unit (GRU) network for processing. The bidirectional GRU network can consider the forward and backward information of the signal at the same time, so that the model can capture the global dependency of time series data. The network contains two layers, each with 128 units, and uses the tanh activation function to handle nonlinear relationships. The GRU network effectively solves the gradient vanishing problem in long time series data through the gating mechanism and obtains real-time time series context features. The multi-head self-attention mechanism is applied to the real-time time series context features. The multi-head self-attention mechanism calculates different dependencies in the signal through 8 parallel attention heads and integrates this information. The dimension of each attention head is 64, which can capture the complex interrelationships between features and generate real-time global dependency features.The real-time global dependency features are input into the pyramid pooling module. Pyramid pooling extracts the multi-scale global information of the signal through pooling operations of three different scales: 1x1, 2x2, and 4x4. The multi-scale pooling method can retain the important information in the feature map while compressing the feature dimension and reducing the computational complexity. Through the pyramid pooling module, the generated real-time multi-scale global features have stronger global information expression capabilities. The real-time multi-scale global features are subjected to feature fusion operations. Feature fusion integrates information from different scales into a unified feature vector through element-by-element addition and 1x1 convolution operations. Element-by-element addition can maintain the relative proportion of features at each scale, while 1x1 convolution further compresses the channel dimension of the feature map, reduces redundant information, and obtains a real-time fused feature vector. The real-time fused feature vector is input into the multi-task learning head for gas type classification and concentration regression calculation. In the classification task, the Softmax activation function is used to output the probability distribution of the gas type; in the regression task, the linear activation function is used to estimate the gas concentration value. Through parallel calculation, the model can simultaneously complete gas recognition and concentration estimation, providing comprehensive detection results. Threshold judgment and time series smoothing are performed on the real-time gas type probability distribution and concentration estimation. Threshold judgment is used to filter low-probability or uncertain results to ensure the stability of the output; time series smoothing eliminates instantaneous fluctuations through methods such as weighted averaging, making the final output of gas type and concentration information more stable and reliable.

[0034] The above describes the detection method based on the multi-modal gas sensor in the embodiment of the present invention. The following describes the detection device based on the multi-modal gas sensor in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a detection device based on a multi-modal gas sensor includes: A creation module is used for performing a co-compounding process on a metal oxide semiconductor and a conductive material to obtain a bimodal nanocomposite material, and creating a sensor array having a three-dimensional microchannel network based on the bimodal nanocomposite material; An acquisition module is used to package and integrate the sensor array to obtain a multi-modal gas detection module, and to synchronously acquire and digitally filter the resistance signal and the capacitance signal based on the multi-modal gas detection module to obtain pre-processed multi-modal signal data; An analysis module is used to collect and analyze the data of the temperature sensor and humidity sensor in the multi-modal gas detection module to obtain environmental factor compensation parameters; A calibration module, used for performing adaptive calibration processing on the pre-processed multi-modal signal data according to the environmental factor compensation parameter to obtain calibrated multi-modal signal data; A training module is used to construct a multi-input neural network model based on the calibrated multi-modal signal data and perform training to obtain a gas recognition model; The processing module is used to analyze and judge the multi-modal signals collected and calibrated in real time based on the gas recognition model to obtain gas type and concentration information.

[0035] Through the synergistic cooperation of the above components, the design and preparation technology of bimodal nanocomposites are developed. 2 The synergistic effect of nanoparticles and graphene significantly improves the gas-sensitive performance and selectivity of the sensing material. The design and manufacture of the three-dimensional microchannel network structure increases the contact area between the gas and the sensitive material, improves the response speed and sensitivity of the sensor, and is conducive to the generation and transmission of multimodal signals. The application of temperature and humidity compensation technology effectively reduces the impact of environmental changes on the detection results through the dynamic environmental factor compensation model, and improves the stability and reliability of the sensor in complex environments. The adaptive calibration processing technology, combined with methods such as wavelet packet transform and empirical mode decomposition, realizes the precise calibration of multimodal signals, effectively improving the signal-to-noise ratio and feature expression ability of the signal. The design of the multi-input neural network model integrates advanced algorithms such as multi-scale residual networks, bidirectional gated recurrent unit networks, and multi-head self-attention mechanisms, significantly improving the recognition ability of complex gases and the accuracy of concentration estimation. The adoption of multi-task learning strategies optimizes gas classification and concentration regression tasks at the same time, improves the generalization ability and robustness of the model, and enables it to better cope with various complex situations in practical applications. The present invention realizes high-precision and high-reliability gas detection.

[0036] The present invention also provides a computer device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the detection method based on the multi-modal gas sensor in the above-mentioned embodiments.

[0037] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the detection method based on a multi-modal gas sensor.

[0038] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0039] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0040] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A detection method based on a multimodal gas sensor, characterized in that: The method comprises: A metal oxide semiconductor and a conductive material are co-mixed and synthesized to obtain a bimodal nanocomposite material, and a sensor array having a three-dimensional microchannel network is created based on the bimodal nanocomposite material; The sensor array is packaged and integrated to obtain a multi-modal gas detection module, and based on the multi-modal gas detection module, resistance signals and capacitance signals are synchronously collected and digitally filtered to obtain pre-processed multi-modal signal data; Collecting and analyzing the temperature sensor and humidity sensor data in the multi-modal gas detection module to obtain environmental factor compensation parameters; Performing adaptive calibration processing on the preprocessed multimodal signal data according to the environmental factor compensation parameter to obtain calibrated multimodal signal data; Building a multi-input neural network model based on the calibrated multimodal signal data and training it to obtain a gas recognition model; Based on the gas identification model, the multimodal signals collected and calibrated in real time are analyzed and judged to obtain gas type and concentration information.

2. The detection method based on a multimodal gas sensor according to claim 1, characterized in that: The metal oxide semiconductor and the conductive material are subjected to a co-mixing synthesis process to obtain a bimodal nanocomposite material, and a sensor array having a three-dimensional microchannel network is created based on the bimodal nanocomposite material, comprising: The tin source and the iron source are mixed according to a preset molar ratio to obtain a metal precursor solution, and the metal precursor solution is subjected to a hydrothermal reaction treatment to obtain Fe-doped SnO2 nanoparticles; Performing ultrasonic dispersion treatment on graphene to obtain a graphene suspension, and mixing the Fe-doped SnO2 nanoparticles and the graphene suspension according to a preset mass ratio to obtain a composite nanomaterial suspension; The composite nanomaterial suspension is spray-dried and calcined to obtain a bimodal nanocomposite material, and the silicon substrate is oxidized to obtain a substrate with a SiO2 insulating layer; Performing photolithography and metal deposition on the substrate with the SiO2 insulating layer to obtain a substrate with comb-shaped electrodes, and performing laser direct writing on the substrate with the comb-shaped electrodes to obtain a substrate structure with a three-dimensional microchannel network; The bimodal nanocomposite material is dispersed to obtain a nanomaterial suspension, and the nanomaterial suspension is deposited on the substrate structure with the three-dimensional microchannel network by inkjet printing technology. The substrate structure with the deposited nanomaterial is heat treated to obtain a sensor array with a three-dimensional microchannel network.

3. The detection method based on a multimodal gas sensor according to claim 1, characterized in that: The sensor array is packaged and integrated to obtain a multi-modal gas detection module, and based on the multi-modal gas detection module, resistance signals and capacitance signals are synchronously collected and digitally filtered to obtain pre-processed multi-modal signal data, including: Precision-processing the low-temperature co-fired ceramic material to obtain a sensor housing with a preset air inlet and air outlet, and fixing the sensor array to obtain a sensor assembly combined with the sensor housing; Ultrasonic bonding is performed based on the sensor assembly to obtain a sensor unit in which electrodes are connected to external leads, and a temperature sensor and a humidity sensor are integrated and installed to obtain a multi-modal gas detection module; Based on the multimodal gas detection module, a micro air pump and a gas distribution system are designed to obtain a sensor system with uniform gas distribution, and a signal conditioning circuit is designed and installed for the sensor system to obtain a signal processing unit with preamplification, filtering and analog-to-digital conversion functions; Based on the signal processing unit, a microcontroller is programmed and configured to obtain a data acquisition and communication control system, and a parallel data acquisition setting is performed on the data acquisition and communication control system to obtain a synchronous acquisition scheme for resistance signals and capacitance signals; Based on the synchronous acquisition scheme, the collected original signal is processed by Butterworth low-pass filtering to obtain a signal with high-frequency noise removed, and the signal with high-frequency noise removed is processed by wavelet transformation to remove baseline drift to obtain pre-processed multimodal signal data.

4. The detection method based on a multimodal gas sensor according to claim 1, characterized in that: The collecting and analyzing the temperature sensor and humidity sensor data in the multi-modal gas detection module to obtain the environmental factor compensation parameters includes: Sampling data from the temperature sensor and the humidity sensor to obtain original temperature and humidity time series data, and performing median filtering on the original temperature and humidity time series data to obtain temperature and humidity data with outliers removed; Performing a fast Fourier transform on the temperature and humidity data with outliers removed to obtain frequency spectrum characteristics of the temperature and humidity data, and analyzing the frequency spectrum characteristics of the temperature and humidity data to obtain a target mode of temperature and humidity changes; Based on the target pattern of temperature and humidity changes, an autoregressive moving average model is constructed to obtain a temperature and humidity change trend prediction model, and the temperature and humidity change trend prediction model is subjected to Kalman filtering to obtain optimized temperature and humidity estimation values; Calculating the influence coefficient of temperature and humidity on the sensitivity of the gas sensor based on the optimized temperature and humidity estimation value to obtain an initial compensation factor, and performing nonlinear least squares fitting on the initial compensation factor to obtain a temperature and humidity compensation function; An adaptive fuzzy neural network is constructed based on the temperature and humidity compensation function to obtain a dynamic environmental factor compensation model, and the parameters of the dynamic environmental factor compensation model are optimized to obtain environmental factor compensation parameters.

5. The detection method based on a multimodal gas sensor according to claim 1, characterized in that: The step of performing adaptive calibration processing on the pre-processed multi-modal signal data according to the environmental factor compensation parameter to obtain calibrated multi-modal signal data includes: Performing time window segmentation on the preprocessed multimodal signal data to obtain signal segments of fixed length, and calculating the influence coefficient of temperature and humidity on each signal segment based on the environmental factor compensation parameter to obtain an environmental influence matrix; Performing singular value decomposition on the environmental impact matrix to obtain a target impact pattern of environmental factors, and constructing an adaptive filter based on the target impact pattern of environmental factors to obtain an environmental noise estimation model; Performing wavelet packet transform on the signal segment and the environmental noise estimation model to obtain multi-scale time-frequency features, and calculating mutual information based on the multi-scale time-frequency features to obtain a correlation index between the signal and the environmental noise; Performing adaptive threshold denoising on the signal segments according to the correlation index to obtain preliminary calibrated signal data, and performing empirical mode decomposition on the preliminary calibrated signal data to obtain an eigenmode function; A Hilbert-Huang transform spectrum is constructed based on the eigenmode function to obtain the instantaneous frequency and amplitude of the signal, and the instantaneous frequency and amplitude are nonlinearly mapped and reconstructed to obtain calibrated multimodal signal data.

6. The detection method based on a multimodal gas sensor according to claim 1, characterized in that: The multi-input neural network model is constructed based on the calibrated multi-modal signal data and trained to obtain a gas recognition model, including: Performing time-frequency domain feature extraction on the calibrated multimodal signal data to obtain a resistance signal feature set and a capacitance signal feature set; Constructing a first multi-scale residual network branch based on the resistance signal feature set, wherein the first multi-scale residual network branch includes three parallel residual blocks, each residual block includes convolutional layers and pooling layers of different scales, and obtains a training multi-scale resistance feature representation; constructing a second multi-scale residual network branch based on the capacitance signal feature set, wherein the structure of the second multi-scale residual network branch is the same as that of the first multi-scale residual network branch, and obtaining a training multi-scale capacitance feature representation; Adaptively recalibrating the training multi-scale resistance feature representation and the multi-scale capacitance feature representation, and weighting the features through a channel attention mechanism and a spatial attention mechanism to obtain training recalibrated features; Based on the training recalibration features, a bidirectional gated recurrent unit network is constructed, wherein the bidirectional gated recurrent unit network comprises 2 layers, each layer has 128 units, and a tanh activation function is used to obtain training temporal context features; Applying a multi-head self-attention mechanism to the training temporal context feature, the multi-head self-attention mechanism comprising 8 attention heads, each with a dimension of 64, to obtain a training global dependency feature; A pyramid pooling module is constructed based on the training global dependency features, and multi-scale global information is extracted through pooling operations of three different scales: 1x1, 2x2, and 4x4, to obtain training multi-scale global features; Performing feature fusion on the training multi-scale global features, and obtaining a training fusion feature vector through element-by-element addition and 1x1 convolution operation; Based on the training fusion feature vector, a multi-task learning head is constructed, including a gas classification task and a concentration regression task, wherein the gas classification task uses a Softmax activation function, and the concentration regression task uses a linear activation function, to obtain a probability distribution of training gas types and a concentration estimation value; The multi-input neural network model is trained end-to-end, a weighted cross entropy loss function and a Huber loss function are used as optimization targets, and an AdamW optimizer and a cosine annealing learning rate scheduling strategy are used to update parameters to obtain a gas recognition model.

7. The detection method based on a multimodal gas sensor according to claim 1, characterized in that: The multimodal signals collected and calibrated in real time are analyzed and judged based on the gas identification model to obtain gas type and concentration information, including: The resistance signal and the capacitance signal collected in real time are synchronously sampled and digitally processed to obtain original multimodal signal data, and adaptively calibrated based on the original multimodal signal data to obtain a calibrated real-time multimodal signal; Performing time-frequency domain feature extraction on the calibrated real-time multimodal signal to obtain a real-time resistance feature set and a real-time capacitance feature set; Inputting the real-time resistance feature set into the first multi-scale residual network branch of the gas identification model, and calculating through three parallel residual blocks to obtain a real-time multi-scale resistance feature representation; Inputting the real-time capacitance feature set into the second multi-scale residual network branch of the gas recognition model, and calculating through three parallel residual blocks to obtain a real-time multi-scale capacitance feature representation; Performing adaptive feature recalibration on the real-time multi-scale resistance feature representation and the real-time multi-scale capacitance feature representation to obtain real-time recalibrated features; Input the real-time recalibration feature into a bidirectional gated recurrent unit network, and obtain the real-time temporal context feature through sequential calculation of 2 layers of 128 units; Applying a multi-head self-attention mechanism to the real-time temporal context features, and obtaining real-time global dependency features through parallel calculation of 8 attention heads; The real-time global dependency features are input into a pyramid pooling module, and real-time multi-scale global features are obtained through pooling operations at three different scales; Performing a feature fusion operation on the real-time multi-scale global feature to obtain a real-time fused feature vector; The real-time fused feature vector is input into a multi-task learning head, and real-time gas type probability distribution and concentration estimation value are obtained through parallel classification and regression calculation; Threshold determination and time series smoothing are performed on the real-time gas type probability distribution and concentration estimation value to obtain gas type and concentration information.

8. A detection device based on a multimodal gas sensor, characterized in that: Used to perform the detection method based on a multimodal gas sensor as described in any one of claims 1 to 7, the detection device based on the multimodal gas sensor comprises: A creation module is used to perform a co-compounding process on a metal oxide semiconductor and a conductive material to obtain a bimodal nanocomposite material, and to create a sensor array having a three-dimensional microchannel network based on the bimodal nanocomposite material; An acquisition module is used to package and integrate the sensor array to obtain a multi-modal gas detection module, and to synchronously acquire and digitally filter the resistance signal and the capacitance signal based on the multi-modal gas detection module to obtain pre-processed multi-modal signal data; An analysis module, used to collect and analyze data from the temperature sensor and humidity sensor in the multi-modal gas detection module to obtain environmental factor compensation parameters; A calibration module, configured to perform adaptive calibration processing on the pre-processed multimodal signal data according to the environmental factor compensation parameter to obtain calibrated multimodal signal data; A training module, used to construct a multi-input neural network model based on the calibrated multimodal signal data and perform training to obtain a gas recognition model; The processing module is used to analyze and judge the multi-modal signals collected and calibrated in real time based on the gas recognition model to obtain gas type and concentration information.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the detection method based on a multimodal gas sensor described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the detection method based on a multi-modal gas sensor according to any one of claims 1 to 7.

Images