Air quality monitoring method based on olfaction chip

Abstract

The invention relates to the technical field of gas monitoring, and discloses an air quality monitoring method based on an olfactory chip, which comprises the following steps: step 1, preparing a multi-modal sensor system which comprises an olfactory chip module, a surface enhanced Raman scattering module and a microfluidic time division multiplexing module, the sensor surface of the olfaction chip module is modified with a metal organic framework material, and the substrate of the surface enhanced Raman scattering module is made of a core-shell structure nano material; and 2, synchronously collecting resistance response data of the olfaction chip module, spectral data of the surface enhanced Raman scattering module and temperature, humidity and wind speed data of the environment sensor. According to the invention, a multi-modal sensor cooperative detection technical scheme is adopted, the high-sensitivity detection effect on low-concentration harmful gas in air is achieved, and compared with a technical scheme of a single sensor and a fixed fusion algorithm in the prior art, the defect that accurate recognition of trace gas cannot be realized under ppb-level concentration is overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of gas monitoring, and specifically to an air quality monitoring method based on an olfactory chip. Background Art

[0002] With the acceleration of the industrialization and urbanization processes, the monitoring demand for trace harmful gases in the air has become increasingly urgent. Traditional air quality monitoring technologies mainly rely on single-modal detection means such as electrochemical sensors and optical spectrometers, combined with fixed-threshold alarms or simple data fusion algorithms. However, there are significant technical bottlenecks in the detection of low-concentration gases, anti-interference ability, long-term stability, and pollution source tracing in complex environments. The specific deficiencies are as follows:

[0003] Most existing technologies use single-type sensors, and the detection lower limit is usually at the ppm level, which cannot meet the monitoring requirements for trace gases at the ppb level. Due to the inherent defects of the hardware, the false alarm rate of such technologies is as high as 30% in low-concentration scenarios, and the sensitivity cannot be fundamentally improved through algorithm optimization.

[0004] Traditional methods use linear temperature compensation models or static Kalman filters to suppress environmental interference, but they perform poorly under complex conditions such as sudden changes in temperature and humidity and turbulent airflows. Field measurements show that when the environmental temperature fluctuates by more than ±5°C, the monitoring error of existing technologies is amplified to ±15%, seriously affecting the accuracy of the assessment.

[0005] Therefore, the present invention proposes an air quality monitoring method based on an olfactory chip to solve the above-mentioned problems. Summary of the Invention

[0006] In view of the deficiencies of the prior art, the present invention provides an air quality monitoring method based on an olfactory chip to solve the problems raised in the above background art.

[0007] To achieve the above objectives, the present invention is realized through the following technical solutions: An air quality monitoring method based on an olfactory chip, comprising:

[0008] Step 1, preparing a multi-modal sensor system, the system comprising an olfactory chip module, a surface-enhanced Raman scattering module, and a microfluidic time-division multiplexing module. The surface of the sensor of the olfactory chip module is modified with a metal-organic framework material, and the substrate of the surface-enhanced Raman scattering module is a core-shell structured nanomaterial;

[0009] Step 2, synchronously collecting the resistance response data of the olfactory chip module, the spectral data of the surface-enhanced Raman scattering module, and the temperature, humidity, and wind speed data of the environmental sensor, and performing noise reduction and time-domain alignment on the data;

[0010] Step 3, generating a synthetic data set based on physical constraint adversarial training, and training a long short-term memory network to establish a mapping relationship between environmental parameters and sensor errors;

[0011] Step 4, dynamically allocate the data weights of the olfactory chip module and the surface-enhanced Raman scattering module through a dual-channel attention mechanism, and fuse multi-source data by combining double-layer Kalman filtering;

[0012] Step 5, trigger a closed-loop self-calibration process, correct the sensor parameters according to the standard gas concentration, and update the lightweight model;

[0013] Step 6, deploy an edge-side real-time monitoring and cloud pollution tracing module, generate a pollution heat map, and analyze the diffusion path.

[0014] Preferably, in the said Step 1, the preparation of the multimodal sensor system further includes:

[0015] Sub-step 1.1, prepare the olfactory chip module: use the microelectromechanical system process to deposit a metal oxide semiconductor thin film on a silicon substrate to form a sensor array, and modify the surface of the sensor with ZIF-8 type metal-organic framework material with a pore size of 0.3 to 0.5 nanometers by chemical vapor deposition. The thickness d of the metal-organic framework material satisfies the formula:

[0016] d = k·T·t,

[0017] where k is the deposition rate coefficient, T is the deposition temperature, and t is the deposition time;

[0018] Sub-step 1.2, prepare the surface-enhanced Raman scattering module: prepare a pyramid-shaped gold-silver core-shell nanostructure substrate on a silicon substrate by the seed growth method. The height h of the substrate, the concentration C of the reactant, and the reduction time τ satisfy the formula:

[0019] h = α·C 0.5 ·ln(τ + 1),

[0020] where α is the morphology adjustment coefficient, C is the concentration of the chloroauric acid solution, and τ is the reduction reaction time;

[0021] Sub-step 1.3, integrate the microfluidic time-division multiplexing module: set a three-way piezoelectric thin film valve between the olfactory chip module and the surface-enhanced Raman scattering module. The gas flow rate switching threshold θ is determined by the response signal change rate ΔR / Δt of the olfactory chip module and satisfies the formula:

[0022]

[0023] where β is the sensitivity coefficient, ΔR is the change in the resistance of the olfactory chip, and Δt is the time interval.

[0024] Preferably, in the step 2, the resistance response data of the olfactory chip module, the spectral data of the surface-enhanced Raman scattering module, and the temperature, humidity, and wind speed data of the environmental sensor are synchronously collected, and the data denoising and time-domain alignment further include:

[0025] Sub-step 2.1, multi-source data synchronous acquisition: Synchronize the data acquisition timings of the olfactory chip module, the surface-enhanced Raman scattering module, and the environmental sensor through a hardware trigger signal, and the timestamp alignment error Δt satisfies the formula:

[0026]

[0027] where f min is the lowest sampling frequency, the sampling frequency of the resistance response data of the olfactory chip module is 10 Hz, and the sampling frequency of the spectral data of the surface-enhanced Raman scattering module is 1 Hz;

[0028] Sub-step 2.2, olfactory chip data denoising: Perform moving window mean filtering on the resistance response data R(t) of the olfactory chip module, and the filtered data R filtered (t) satisfies the formula:

[0029]

[0030] where N is the window size, w i is the weight coefficient, t is the current timestamp, and R filtered (t) is the filtered resistance response value;

[0031] Sub-step 2.3, surface-enhanced Raman scattering spectrum denoising: Perform wavelet transform denoising on the spectral data S(λ) of the surface-enhanced Raman scattering module, and the decomposed signal S denoised (λ) satisfies the formula:

[0032]

[0033] where L is the number of wavelet decomposition layers, D j (λ) is the detail coefficient of the j-th layer, τ is the threshold, and the SoftThreshold function is defined as:

[0034] SoftThreshold(x,τ) = sign(x)·max(|x| - τ, 0),

[0035] where x is the wavelet decomposition detail coefficient and max is the maximum value.

[0036] Preferably, in the step 3, generating a synthetic data set based on physical constraint adversarial training and training a long short-term memory network to establish a mapping relationship between environmental parameters and sensor errors further includes:

[0037] Sub-step 3.1, generate physically constrained synthetic data: Based on computational fluid dynamics to simulate the gas diffusion process, solve the Navier-Stokes equations to generate the velocity field u(x,t) and the concentration field C(x,t), and the simulation data satisfies the formula:

[0038]

[0039] where D is the gas diffusion coefficient, is the Laplace operator, x is the spatial coordinate, and t is the time;

[0040] Sub-step 3.2, construct a physically constrained generative adversarial network: Define the generator G(z,θ) and the discriminator The generator takes the latent vector z and the environmental parameter θ as inputs and outputs the synthetic concentration field The loss function L G includes the adversarial loss L adv and the physical constraint term L phy , and satisfies the formula:

[0041] L G =λ1L adv +λ2L phy ,

[0042]

[0043] where λ1 and λ2 are weight coefficients, is the square of the L2 norm, L adv is the adversarial loss, L phy is the physical constraint loss, is the time derivative of the generated concentration field, is the convection term, is the diffusion term;

[0044] Sub-step 3.3, train the long short-term memory network mapping error: Input the synthetic data set into the long short-term memory network. The network takes the environmental parameter θ and the original sensor data S(t) as inputs and outputs the sensor error ∈(t). The loss function L LSTM satisfies the formula:

[0045]

[0046] where L LSTM is the loss function, T is the time step, ∈ true (t) is the calibrated sensor error true value, and ∈(t) is the sensor error predicted by the long short-term memory network.

[0047] Preferably, in step 4, the data weights of the olfactory chip module and the surface-enhanced Raman scattering module are dynamically allocated through a dual-channel attention mechanism. The multi-source data fusion by combining double-layer Kalman filtering further includes:

[0048] Sub-step 4.1, calculating the environmental attention weight: Based on the mapping relationship between the environmental parameters and the sensor error output by the long short-term memory network trained in step 3, calculate the weight w MOS of the olfactory chip module and the weight w SERS of the surface-enhanced Raman scattering module, satisfying the formula:

[0049]

[0050] w SERS = 1 - w MOS ,

[0051] where T is the temperature, H is the humidity, V is the wind speed, and α, β, γ are environmental compensation coefficients;

[0052] Sub-step 4.2, calculating the pollution event attention weight: If the surface-enhanced Raman scattering module detects the characteristic peak of the target gas and the peak intensity I peak exceeds the threshold I th , dynamically increase the weight of the surface-enhanced Raman scattering module to w′ SERS , satisfying the formula:

[0053]

[0054] where η is the sensitivity coefficient, and I th is the trace gas detection threshold;

[0055] Sub-step 4.3, performing double-layer Kalman filtering fusion: The first-layer filtering performs environmental compensation on the original data of the olfactory chip module and the surface-enhanced Raman scattering module. The state equation and the observation equation are:

[0056] x k = Ax k-1 + Bu k + w k ,

[0057] z k = Hx k + v k ,

[0058] where x k is the state vector, x k-1 is the state vector of the previous moment, A and B are the state transition matrices, u k is the environmental parameter input, w k , v k are the process noise and the observation noise, and zk is the observation vector, and H is the observation matrix;

[0059] The second - layer filtering dynamically adjusts the covariance matrix P based on the weights of sub - steps 4.1 and 4.2 k , satisfying the formula:

[0060] P k =P k|k-1 -K k HP k|k-1 ,

[0061]

[0062] where K k is the Kalman gain, R k is the observation noise covariance matrix, P k|k-1 is the prior prediction covariance matrix, and P k is the covariance matrix of the state estimation error at the current moment.

[0063] Preferably, in step 5, triggering the closed - loop self - calibration process, correcting the sensor parameters according to the calibration gas concentration and updating the lightweight model further includes:

[0064] Sub - step 5.1, triggering the calibration condition determination: When the variance σ 2 of the output data of the olfactory chip module or the surface - enhanced Raman scattering module continuously exceeds the set threshold σ th , the calibration process is triggered, and the variance calculation satisfies the formula:

[0065]

[0066] where M is the number of data points within the time window, x i is the sensor output value, and μ is the mean value within the window;

[0067] Sub - step 5.2, performing sensor parameter correction: Releasing the standard gas with a concentration of C ref in the calibration gas cylinder, calculating the sensor gain coefficient error ΔK, satisfying the formula:

[0068]

[0069] where C meas is the measured concentration value of the sensor, and the corrected gain coefficient K new is:

[0070] K new =K old ·(1 + λΔK),

[0071] where λ is the error correction proportionality coefficient, and K old is the original gain coefficient;

[0072] Sub-step 5.3, update the weights of the lightweight model: When ΔK > 10%, use the calibration data to update the weights W of the fully connected layer of the lightweight model, and the update amount ΔW satisfies the formula:

[0073]

[0074] where η is the learning rate, is the gradient of the model output with respect to the weights.

[0075] Preferably, in step 6, deploying the edge-side real-time monitoring and cloud pollution tracing module, generating a pollution heat map and analyzing the diffusion path further includes:

[0076] Sub-step 6.1, edge-side real-time data compression and transmission: Perform principal component analysis dimensionality reduction on the fused pollutant concentration data C(t) at the edge side, retain the first k groups of principal components, and the data compression rate ρ satisfies the formula:

[0077]

[0078] where d is the dimension of the original data, k is the number of retained principal components, and the compressed data is transmitted to the cloud through the LoRa wireless protocol with a transmission interval of 5 seconds;

[0079] Sub-step 6.2, cloud pollution heat map generation: Based on the pollutant concentration C i (x, y) transmitted by multiple nodes, use the Kriging interpolation algorithm to generate a spatial distribution heat map, and the interpolation weight λ j satisfies the formula:

[0080]

[0081] where γ(h ij ) is the variogram, h ij is the spatial distance between nodes i and j, λ j is the Kriging interpolation weight coefficient, and μ is the Lagrange multiplier;

[0082] Sub-step 6.3, graph neural network pollution tracing: Construct a graph neural network model, and the message passing function m ij satisfies the formula:

[0083] m ij = ReLU(W m ·[C i , T i , V i , C j , T j , V j ),

[0084] where W mis a trainable weight matrix, C i , T i , V i , C j , T j , V j are all node features, m ij is a message passing function.

[0085] Preferably, in the step 5, the closed-loop self-calibration process is triggered by the sensor output variance continuously exceeding the set threshold. The standard gas concentration is 1 ppm toluene, and when the calibration error exceeds 10%, the weights of the fully connected layer of the lightweight model are updated.

[0086] A terminal device includes a memory, a processor, and a computer program stored on the memory. When the processor executes the computer program, the air quality monitoring method based on the olfactory chip is implemented.

[0087] A storage medium stores a computer program, and when the computer program is executed by a processor, the air quality monitoring method based on the olfactory chip is implemented.

[0088] The present invention provides an air quality monitoring method based on an olfactory chip. It has the following beneficial effects:

[0089] 1. The present invention adopts a multi-modal sensor collaborative detection technical solution to achieve a high-sensitivity detection effect for low-concentration harmful gases in the air. Compared with the technical solutions of single sensors and fixed fusion algorithms in the prior art, it solves the problem of inability to accurately identify trace gases at the ppb level concentration.

[0090] 2. The present invention adopts a physical constraint adversarial training and dynamic weight fusion technical solution to achieve a high-precision monitoring effect against environmental interference. Compared with the technical solutions of relying on a single environmental compensation model and static filtering algorithms in the prior art, it solves the problem of error amplification caused by sudden changes in temperature, humidity, and wind speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0092] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0093] The present invention will be described in detail below with reference to the accompanying drawings:

[0094] Example:

[0095] Please refer to the attached Figure 1 , this embodiment of the present invention provides an air quality monitoring method based on an olfactory chip, including:

[0096] Step 1, prepare a multimodal sensor system, the system includes an olfactory chip module, a surface-enhanced Raman scattering module and a microfluidic time-division multiplexing module. The sensor surface of the olfactory chip module is modified with a metal-organic framework material, and the substrate of the surface-enhanced Raman scattering module is a core-shell structure nanomaterial;

[0097] Sub-step 1.1, prepare the olfactory chip module: use microelectromechanical system technology to deposit a metal oxide semiconductor thin film on a silicon substrate to form a sensor array, and modify the sensor surface with a ZIF-8 type metal-organic framework material with a pore size of 0.3 to 0.5 nanometers by chemical vapor deposition. The thickness d of the metal-organic framework material satisfies the formula:

[0098] d = k·T·t,

[0099] where k is the deposition rate coefficient, T is the deposition temperature, and t is the deposition time;

[0100] Sub-step 1.2, prepare the surface-enhanced Raman scattering module: prepare a pyramid-shaped gold-silver core-shell nanostructure substrate on a silicon substrate by the seed growth method. The height h of the substrate, the reactant concentration C, and the reduction time τ satisfy the formula:

[0101] h = α·C 0.5 ·ln(τ + 1),

[0102] where α is the morphology adjustment coefficient, C is the concentration of chloroauric acid solution, and τ is the reduction reaction time;

[0103] Sub-step 1.3, integrate the microfluidic time-division multiplexing module: set a three-way piezoelectric thin film valve between the olfactory chip module and the surface-enhanced Raman scattering module. The gas flow rate switching threshold θ is determined by the response signal change rate ΔR / Δt of the olfactory chip module and satisfies the formula:

[0104]

[0105] where β is the sensitivity coefficient, ΔR is the change in the resistance of the olfactory chip, and Δt is the time interval;

[0106] Step 2, synchronously collect the resistance response data of the olfactory chip module, the spectral data of the surface-enhanced Raman scattering module, and the temperature, humidity, and wind speed data of the environmental sensor, and perform noise reduction and time-domain alignment on the data;

[0107] Sub-step 2.1, Multi-source data synchronous acquisition: Synchronize the data acquisition timings of the olfactory chip module, surface-enhanced Raman scattering module, and environmental sensor through a hardware trigger signal. The timestamp alignment error Δt satisfies the formula:

[0108]

[0109] where, f min is the lowest sampling frequency. The sampling frequency of the resistance response data of the olfactory chip module is 10 Hz, and the sampling frequency of the spectral data of the surface-enhanced Raman scattering module is 1 Hz;

[0110] Sub-step 2.2, Olfactory chip data denoising: Perform sliding window mean filtering on the resistance response data R(t) of the olfactory chip module. The filtered data R filtered (t) satisfies the formula:

[0111]

[0112] where, N is the window size, w i is the weight coefficient, t is the current timestamp, and R filtered (t) is the filtered resistance response value;

[0113] Sub-step 2.3, Surface-enhanced Raman scattering spectrum denoising: Perform wavelet transform denoising on the spectral data S(λ) of the surface-enhanced Raman scattering module. The decomposed signal S denoised (λ) satisfies the formula:

[0114]

[0115] where, L is the number of wavelet decomposition layers, D j (λ) is the detail coefficient of the j-th layer, τ is the threshold, and the SoftThreshold function is defined as:

[0116] SoftThreshold(x,τ) = sign(x)·max(|x| - τ, 0),

[0117] where, x is the wavelet decomposition detail coefficient, and max is the maximum value;

[0118] Step 3, Generate a synthetic dataset based on physical constraint adversarial training, and train a long short-term memory network to establish a mapping relationship between environmental parameters and sensor errors;

[0119] Sub-step 3.1, Generate physical constraint synthetic data: Based on computational fluid dynamics to simulate the gas diffusion process, solve the Navier-Stokes equation to generate the velocity field u(x,t) and concentration field C(x,t). The simulation data satisfies the formula:

[0120]

[0121] Among them, D is the gas diffusion coefficient, is the Laplace operator, x is the spatial coordinate, and t is the time;

[0122] Sub-step 3.2, construct a physical constraint adversarial generation network: Define the generator G(z,θ) and the discriminator The generator takes the latent vector z and the environmental parameter θ as inputs and outputs the synthetic concentration field The loss function L G includes the adversarial loss L adv and the physical constraint term L phy , and satisfies the formula:

[0123] L G = λ1L adv + λ2L phy ,

[0124]

[0125] Among them, λ1 and λ2 are weight coefficients, is the square of the L2 norm, L adv is the adversarial loss, L phy is the physical constraint loss, is the time derivative of the generated concentration field, is the convection term, is the diffusion term;

[0126] Sub-step 3.3, train the long short-term memory network mapping error: Input the synthetic data set into the long short-term memory network. The network input is the environmental parameter θ and the original sensor data S(t), and the output is the sensor error ∈(t). The loss function L LSTM satisfies the formula:

[0127]

[0128] Among them, L LSTM is the loss function, T is the time step, ∈ true (t) is the calibrated sensor error true value, and ∈(t) is the sensor error predicted by the long short-term memory network;

[0129] Step 4, dynamically allocate the data weights of the olfactory chip module and the surface-enhanced Raman scattering module through a dual-channel attention mechanism, and fuse multi-source data by combining a double-layer Kalman filter;

[0130] Sub-step 4.1, calculate the environmental attention weight: Based on the mapping relationship between the environmental parameter and the sensor error output by the long short-term memory network trained in Step 3, calculate the weight w of the olfactory chip module MOS and the weight w of the surface-enhanced Raman scattering module SERS, satisfying the formula:

[0131]

[0132] w SERS = 1 - w MOS ,

[0133] where T is the temperature, H is the humidity, V is the wind speed, and α, β, γ are environmental compensation coefficients;

[0134] Sub-step 4.2, calculating the attention weight of the pollution event: If the surface-enhanced Raman scattering module detects the characteristic peak of the target gas and the peak intensity I peak exceeds the threshold I th at this time, dynamically increase the weight of the surface-enhanced Raman scattering module to w' SERS , satisfying the formula:

[0135]

[0136] where η is the sensitivity coefficient and I th is the trace gas detection threshold;

[0137] Sub-step 4.3, performing double-layer Kalman filter fusion: The first-layer filter performs environmental compensation on the original data of the olfactory chip module and the surface-enhanced Raman scattering module. The state equation and the observation equation are:

[0138] x k = Ax k-1 + Bu k + w k ,

[0139] z k = Hx k + v k ,

[0140] where x k is the state vector, x k-1 is the state vector of the previous moment, A and B are the state transition matrices, u k is the environmental parameter input, w k , v k are the process noise and the observation noise, z k is the observation vector, and H is the observation matrix;

[0141] The second-layer filter dynamically adjusts the covariance matrix P k based on the weights in Sub-steps 4.1 and 4.2, satisfying the formula:

[0142] P k = P k|k-1 - K k HP k|k-1 ,

[0143]

[0144] Among them, K k is the Kalman gain, R k is the observation noise covariance matrix, P k|k-1 is the prior prediction covariance matrix, P k is the covariance matrix of the state estimation error at the current moment;

[0145] Step 5: Trigger the closed-loop self-calibration process, correct the sensor parameters according to the calibration gas concentration, and update the lightweight model;

[0146] Sub-step 5.1: Trigger the calibration condition determination: The variance σ 2 of the output data of the olfactory chip module or the surface-enhanced Raman scattering module continuously exceeds the set threshold σ th to trigger the calibration process. The variance calculation satisfies the formula:

[0147]

[0148] Among them, M is the number of data points within the time window, x i is the sensor output value, and μ is the mean value within the window;

[0149] Sub-step 5.2: Perform sensor parameter correction: Release the standard gas with a concentration of C ref in the calibration gas cylinder, and calculate the sensor gain coefficient error ΔK, which satisfies the formula:

[0150]

[0151] Among them, C meas is the measured concentration value of the sensor, and the corrected gain coefficient K new is:

[0152] K new = K old ·(1 + λΔK),

[0153] Among them, λ is the error correction proportionality coefficient, and K old is the original gain coefficient;

[0154] Sub-step 5.3: Update the lightweight model weights: If ΔK > 10%, use the calibration data to update the weights W of the fully connected layer of the lightweight model. The update amount ΔW satisfies the formula:

[0155]

[0156] Among them, η is the learning, is the gradient of the model output with respect to the weights;

[0157] Step 6: Deploy the edge - side real - time monitoring and cloud - side pollution traceability module to generate a pollution heat map and analyze the diffusion path;

[0158] Sub - step 6.1: Edge - side real - time data compression and transmission: Perform principal component analysis dimensionality reduction on the fused pollutant concentration data C(t) at the edge, retain the first k groups of principal components, and the data compression rate ρ satisfies the formula:

[0159]

[0160] where d is the dimension of the original data, k is the number of retained principal components, and the compressed data is transmitted to the cloud through the LoRa wireless protocol with a transmission interval of 5 seconds;

[0161] Sub - step 6.2: Cloud - side pollution heat map generation: Based on the pollutant concentration C i (x,y) transmitted by multiple nodes, use the Kriging interpolation algorithm to generate a spatial distribution heat map, and the interpolation weight λ j satisfies the formula:

[0162]

[0163] where γ(h ij ) is the variogram, h ij is the spatial distance between nodes i and j, λ j is the Kriging interpolation weight coefficient, and μ is the Lagrange multiplier;

[0164] Sub - step 6.3: Pollution traceability using graph neural network: Construct a graph neural network model, and the message - passing function m ij satisfies the formula:

[0165] m ij =ReLU(W m ·[C i ,T i ,V i ,C j ,T j ,V j ),

[0166] where W m is the trainable weight matrix, C i 、T i 、V i 、C j 、T j 、V j are all node features, and m ij is the message - passing function.

[0167] Benefits of Step 1: By integrating the olfactory chip module, surface-enhanced Raman scattering module, and microfluidic time-division multiplexing module, the collaborative complementarity of multimodal gas detection capabilities is achieved. The ZIF-8 type MOFs material of the olfactory chip selectively adsorbs target gases through size sieving effect, enhancing the detection specificity; the gold-silver core-shell pyramid structure substrate of the SERS module enhances the Raman signal by more than 10 6 times through the local surface plasmon resonance effect, breaking through the sensitivity bottleneck of traditional spectral detection; the microfluidic time-division multiplexing module dynamically switches the gas flow direction through a piezoelectric valve to avoid cross-contamination of sensors. This solution solves the problems of narrow detection range and insufficient sensitivity of a single sensor at the hardware level, providing a high-precision data source for trace gas identification.

[0168] Benefits of Step 2: Synchronize the acquisition timings of multi-source data based on the hardware trigger signal to ensure the time alignment of resistance response, spectral data, and environmental parameters; the data of the olfactory chip suppresses high-frequency noise through sliding window mean filtering, and the SERS spectrum separates the effective peaks and background interference through wavelet soft threshold denoising. This step ensures the consistency of multimodal information from the data source, reduces the interference of environmental noise on subsequent analysis, and provides high-quality input for physical constraint modeling and dynamic fusion.

[0169] Benefits of Step 3: Simulate gas diffusion through the Navier-Stokes equation to generate a physical constraint data set, combine with a generative adversarial network to synthesize concentration field data containing environmental interference, and force the generated data to conform to the laws of fluid mechanics; the long short-term memory network uses the synthesized data to establish a mapping relationship between environmental parameters and sensor errors, quantifying the impact of environmental mutations on sensor output. This step solves the problem of poor generalization of traditional data-driven models, providing an environmental error compensation benchmark for dynamic weight allocation.

[0170] Benefits of Step 4: The dual-channel attention mechanism dynamically allocates weights to the olfactory chip and SERS module according to environmental parameters, and at the same time triggers the pollution event channel when the characteristic peak of the target gas is detected, increasing the SERS weight to more than 0.7; the first layer of the double-layer Kalman filter suppresses environmental noise, and the second layer optimizes the signal-to-noise ratio of multi-source data fusion based on weight adjustment of the covariance matrix. This solution reduces the monitoring error in complex environments from ±15% to ±3%, breaking through the bottleneck of insufficient adaptability of static fusion algorithms in dynamic scenarios.

[0171] Benefits of Step 5: Real-time monitor the sensor stability through variance threshold and trigger the standard gas calibration process; dynamically correct the sensor gain coefficient and lightweight model weight according to the standard gas concentration to achieve adaptive compensation for sensor drift and model failure. This step enables the cumulative error of the system to be stabilized within 5% after continuous operation for 6 months, reducing the traditional manual calibration maintenance frequency by 80%, and significantly improving the long-term reliability.

[0172] Benefits of Step 6: The edge device compresses data to 37.5% of its original volume through principal component analysis and transmits it with low power consumption using the LoRa protocol. The cloud generates a pollution heat map based on Kriging interpolation and combines it with a graph neural network to predict the coordinates of pollution sources in real time. This solution reduces the pollution source localization error from 200 meters to 50 meters and shortens the response time from 2 hours to 5 minutes, achieving a closed-loop monitoring-tracing system.

[0173] In Step 5, the closed-loop self-calibration process is triggered when the variance of the sensor output continuously exceeds the set threshold. The calibration gas concentration is 1 ppm toluene, and when the calibration error exceeds 10%, the weights of the fully connected layer of the lightweight model are updated.

[0174] A terminal device includes a memory, a processor, and a computer program stored on the memory. When the processor executes the computer program, it implements an air quality monitoring method based on an olfactory chip.

[0175] A storage medium stores a computer program. When the computer program is executed by a processor, it implements an air quality monitoring method based on an olfactory chip.

[0176] Benefits of the terminal device: The processor adopts a heterogeneous architecture of ARM Cortex-M7 and NPU (Neural Network Processing Unit). The NPU is dedicated to accelerating the physical constraint adversarial training model in Step 3 and the graph neural network traceability calculation in Step 6, making the real-time processing delay of LSTM error mapping and Kalman filtering lower than 50 ms, meeting the millisecond-level response requirements of industrial scenarios.

[0177] Through dynamic voltage and frequency scaling technology, the main frequency of the processor drops to 50 MHz during pollution-free events, and the power consumption is lower than 100 mW, supporting the edge device to work continuously for more than 30 days without an external power supply, adapting to the long-term monitoring requirements of Step 6.

[0178] An internal hardware-level timestamp synchronization module ensures that the data acquisition timing alignment error of the olfactory chip, SERS, and environmental sensors in Step 2 is ≤ 1 ms; the memory uses the eMMC 5.1 protocol, supporting the cyclic storage of calibration data in Step 5 for 90 days to avoid data loss.

[0179] Benefits of the storage medium: The storage medium supports multi-format file systems such as FAT32 / exFAT / NTFS, ensuring seamless migration of the synthetic dataset in Step 3, Kalman filter parameters in Step 4, and lightweight model weights in Step 5 between Windows / Linux / embedded RTOS systems, adapting to heterogeneous hardware such as industrial PCs and edge gateways.

[0180] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An air quality monitoring method based on an olfactory chip, characterized in that, Including: Step 1: Prepare a multimodal sensor system, which includes an olfactory chip module, a surface-enhanced Raman scattering module, and a microfluidic time-division multiplexing module. The sensor surface of the olfactory chip module is modified with a metal-organic framework material, and the substrate of the surface-enhanced Raman scattering module is a core-shell structured nanomaterial; Step 2: Synchronously collect the resistance response data of the olfactory chip module, the spectral data of the surface-enhanced Raman scattering module, and the temperature, humidity, and wind speed data of the environmental sensor, and perform noise reduction and time-domain alignment on the data; Step 3: Generate a synthetic dataset based on physically constrained adversarial training, and train a long short-term memory network to establish a mapping relationship between environmental parameters and sensor errors; Step 4: Dynamically allocate the data weights of the olfactory chip module and the surface-enhanced Raman scattering module through a dual-channel attention mechanism, and fuse multi-source data with a double-layer Kalman filter; Step 5: Trigger a closed-loop self-calibration process, correct the sensor parameters according to the standard gas concentration, and update the lightweight model; Step 6: Deploy an edge-side real-time monitoring and cloud pollution tracing module to generate a pollution heat map and analyze the diffusion path.

2. The air quality monitoring method based on an olfactory chip according to claim 1, wherein In the said Step 1, the preparation of the multimodal sensor system further includes: Sub-step 1.1: Prepare the olfactory chip module: Use microelectromechanical system technology to deposit a metal oxide semiconductor thin film on a silicon substrate to form a sensor array, and modify the sensor surface with a ZIF-8 type metal-organic framework material with a pore size of 0.3 to 0.5 nanometers through chemical vapor deposition. The thickness d of the metal-organic framework material satisfies the formula: d = k·T·t, where k is the deposition rate coefficient, T is the deposition temperature, and t is the deposition time; Sub-step 1.2: Prepare the surface-enhanced Raman scattering module: Prepare a pyramid-shaped gold-silver core-shell nanostructure substrate on a silicon substrate by the seed growth method. The height h of the substrate, the reactant concentration C, and the reduction time τ satisfy the formula: h = α·C 0.5 ·ln(τ + 1), where α is the morphology adjustment coefficient, C is the concentration of chloroauric acid solution, and τ is the reduction reaction time; Sub-step 1.3: Integrate the microfluidic time-division multiplexing module: Set a three-way piezoelectric thin film valve between the olfactory chip module and the surface-enhanced Raman scattering module. The gas flow rate switching threshold θ is determined by the response signal change rate ΔR / Δt of the olfactory chip module and satisfies the formula: where β is the sensitivity coefficient, ΔR is the change in the resistance of the olfactory chip, and Δt is the time interval.

3. The air quality monitoring method based on an olfactory chip according to claim 1, wherein In the said Step 2, the synchronous collection of the resistance response data of the olfactory chip module, the spectral data of the surface-enhanced Raman scattering module, and the temperature, humidity, and wind speed data of the environmental sensor, and the noise reduction and time-domain alignment of the data further include: Sub-step 2.1: Synchronous collection of multi-source data: Synchronize the data acquisition time sequences of the olfactory chip module, the surface-enhanced Raman scattering module, and the environmental sensor through a hardware trigger signal. The time stamp alignment error Δt satisfies the formula: where f min is the lowest sampling frequency, the sampling frequency of the resistance response data of the olfactory chip module is 10 Hz, and the sampling frequency of the spectral data of the surface-enhanced Raman scattering module is 1 Hz; Sub-step 2.2, olfactory chip data denoising: Perform moving window mean filtering on the resistance response data R(t) of the olfactory chip module. The filtered data R filtered (t) satisfies the formula: where N is the window size, w i is the weight coefficient, t is the current timestamp, and R filtered (t) is the filtered resistance response value; Sub-step 2.3, denoising of surface-enhanced Raman scattering spectrum: Perform wavelet transform denoising on the spectral data S(λ) of the surface-enhanced Raman scattering module. After decomposition, the signal S denoised (λ) satisfies the formula: where L is the number of wavelet decomposition levels, D j (λ) is the detail coefficient of the j-th layer, τ is the threshold, and the SoftThreshold function is defined as: SoftThreshold(x,τ) = sign(x)·max(|x| - τ, 0), where x is the wavelet decomposition detail coefficient and max is the maximum value.

4. The air quality monitoring method based on an olfactory chip according to claim 1, wherein In the said Step 3, the generation of a synthetic dataset based on physically constrained adversarial training and the training of a long short-term memory network to establish a mapping relationship between environmental parameters and sensor errors further include: Sub-step 3.1, generating physically constrained synthetic data: Based on computational fluid dynamics to simulate the gas diffusion process, solve the Navier-Stokes equations to generate the velocity field u(x,t) and the concentration field C(x,t), and the simulation data satisfies the formula: where D is the gas diffusion coefficient, is the Laplace operator, x is the spatial coordinate, and t is the time; Sub-step 3.2, construct a physical constraint adversarial generation network: Define the generator G(z,θ) and the discriminator The generator takes the latent vector z and the environmental parameter θ as inputs and outputs a synthetic concentration field The loss function L G includes the adversarial loss L adv and the physical constraint term L phy , satisfying the formula: L G = λ1L adv + λ2L phy , where λ1 and λ2 are weight coefficients, is the square of the L2 norm, L adv is the adversarial loss, L phy is the physical constraint loss, is the time derivative of the generated concentration field, is the convection term, is the diffusion term; Sub-step 3.3, training the long short-term memory network mapping error: Input the synthetic data set into the long short-term memory network. The network inputs are the environmental parameter θ and the original sensor data S(t), and the output is the sensor error ∈(t). The loss function L LSTM satisfies the formula: Among them, L LSTM is the loss function, T is the time step, and ∈ true (t) is the true value of the calibrated sensor error, and ∈(t) is the sensor error predicted by the long short-term memory network.

5. A method for monitoring air quality based on an olfactory chip according to claim 1, characterized in that, In the said step 4, dynamically allocating the data weights of the olfactory chip module and the surface-enhanced Raman scattering module through a dual-channel attention mechanism, and further including fusing multi-source data with a double-layer Kalman filter: Sub-step 4.1, calculate the environmental attention weight: Based on the mapping relationship between the environmental parameters and the sensor error output by the long short-term memory network trained in step 3, calculate the weight w of the olfactory chip module MOS and the weight w of the surface-enhanced Raman scattering module SERS , satisfying the formula: w SERS = 1 - w MOS , where, T is the temperature, H is the humidity, V is the wind speed, and α, β, γ are environmental compensation coefficients; Sub-step 4.2, calculate the attention weight of the pollution event: If the surface-enhanced Raman scattering module detects the characteristic peak of the target gas and the peak intensity I peak exceeds the threshold I th at this time, dynamically increase the weight of the surface-enhanced Raman scattering module to w′ SERS , satisfying the formula: where η is the sensitivity coefficient and I th is the trace gas detection threshold; Sub-step 4.3, performing double-layer Kalman filter fusion: The first-layer filter performs environmental compensation on the original data of the olfactory chip module and the surface-enhanced Raman scattering module, and the state equation and the observation equation are: x k = Ax k-1 + Bu k + w k , z k = Hx k + v k , where, x k is the state vector, x k-1 is the state vector at the previous moment, A and B are state transition matrices, u k is the environmental parameter input, w k , v k are the process noise and the observation noise, z k is the observation vector, and H is the observation matrix; The second layer of filtering dynamically adjusts the covariance matrix P based on the weights of sub-steps 4.1 and 4.2 k , satisfying the formula: P k = P k|k-1 - K k HP k|k-1 , K k = P k|k-1 H T (HP k|k-1 H T + R k ) -1 , Among them, K k is the Kalman gain, R k is the observation noise covariance matrix, P k|k-1 is the prior prediction covariance matrix, and P k is the covariance matrix of the state estimation error at the current moment.

6. The air quality monitoring method based on an olfactory chip according to claim 1, wherein In the said step 5, triggering a closed-loop self-calibration process, correcting the sensor parameters according to the standard gas concentration and updating the lightweight model further includes: Sub-step 5.1, trigger calibration condition determination: the variance σ of the output data of the olfactory chip module or the surface-enhanced Raman scattering module 2 continuously exceeds the set threshold σ th to trigger the calibration process, and the variance calculation satisfies the formula: where M is the number of data points within the time window, x i is the sensor output value, and μ is the mean value within the window; Sub-step 5.2, perform sensor parameter correction: Release the standard gas with a concentration of C ref in the calibration gas cylinder, calculate the sensor gain coefficient error ΔK, which satisfies the formula: Among them, C meas is the measured concentration value of the sensor, and the corrected gain coefficient K new is as follows: K new = K old ·(1 + λΔK), Among them, λ is the error correction proportionality coefficient, and K old is the original gain coefficient; Sub-step 5.3, updating the weights of the lightweight model: When ΔK > 10%, use the calibration data to update the weights W of the fully connected layer of the lightweight model, and the update amount ΔW satisfies the formula: Among them, η is for learning, which is the gradient of the model output with respect to the weights.

7. The air quality monitoring method based on an olfactory chip according to claim 1, characterized in that, In the said step 6, deploying an edge-side real-time monitoring and cloud-side pollution source tracing module, generating a pollution heat map and analyzing the diffusion path further includes: Sub-step 6.1, edge-side real-time data compression and transmission: Perform principal component analysis dimensionality reduction on the fused pollutant concentration data C(t) at the edge side, retain the first k groups of principal components, and the data compression rate ρ satisfies the formula: where, d is the dimension of the original data, k is the number of retained principal components, and the compressed data is transmitted to the cloud through the LoRa wireless protocol, and the transmission interval is 5 seconds; Sub-step 6.2, Cloud Pollution Heat Map Generation: Based on the pollutant concentration C i (x, y) transmitted by multiple nodes, a spatial distribution heat map is generated using the Kriging interpolation algorithm, and the interpolation weight λ j satisfies the formula: Among them, γ(h ij ) is the variogram, h ij is the spatial distance between nodes i and j, λ j is the Kriging interpolation weight coefficient, and μ is the Lagrange multiplier; Sub-step 6.3, Graph Neural Network Pollution Tracing: Construct a graph neural network model, and the message passing function m ij Satisfies the formula: m ij = ReLU(W m · [C i , T i , V i , C j , T j , V j ) Among them, W m is a trainable weight matrix, C i , T i , V i , C j , T j , V j are all node features, and m ij is a message passing function.

8. The air quality monitoring method based on an olfactory chip according to claim 1, characterized in that In the said step 5, the closed-loop self-calibration process is triggered by the sensor output variance continuously exceeding the set threshold. When the standard gas concentration is 1 ppm toluene and the calibration error exceeds 10%, update the weights of the fully connected layer of the lightweight model.

9. A terminal device, characterized in that, Comprising a memory, a processor, and a computer program stored on the memory, when the processor executes the computer program, it implements the olfactory-chip-based air quality monitoring method according to any one of claims 1 to 8.

10. A storage medium, characterized in that, Stored with a computer program, when the computer program is executed by a processor, it implements the olfactory-chip-based air quality monitoring method according to any one of claims 1 to 8.

Images