Multi-identification and concentration quantitative detection system based on sensor array

Abstract

The application discloses a multi-identification and concentration quantitative detection system based on a sensor array, belongs to the technical field of gas detection, and specifically comprises the following steps: a curve acquisition module continuously acquires the resistance value of each sensor element to form a resistance response curve in the process that a gas sensor array is exposed to a mixed gas to be detected; a feature extraction module extracts a feature parameter set reflecting the gas adsorption and desorption process by segmenting the resistance response curve; a vector fusion module combines the feature parameter set and environmental temperature and humidity into a multi-dimensional feature vector; a decoupling reasoning module inputs the multi-dimensional feature vector into an adsorption competition decoupling neural network to output the initial concentration estimation value of each gas and a competition influence matrix among the gases; a concentration inversion module obtains the actual concentration value of each gas by solving a linear equation set; and a storage output module stores and outputs the actual concentration value in association. The application realizes the synchronous quantitative detection of the concentration of multiple components in a mixed gas.

Description

Technical Field

[0001] This invention relates to the field of gas detection technology, and more specifically to various identification and concentration quantitative detection systems based on sensor arrays. Background Technology

[0002] Gas sensor array detection technology is an important development direction in the field of volatile organic compound (VOC) monitoring. It utilizes multiple gas-sensitive elements with cross-sensitivity characteristics to form a sensor array, combined with pattern recognition algorithms to achieve qualitative and quantitative analysis of various gases. With the development of the Internet of Things (IoT) and embedded technologies, sensor array-based gas detection devices are widely used in industrial safety, environmental monitoring, and indoor air quality assessment, placing higher technical demands on the simultaneous identification and concentration inversion of multiple components in mixed gases.

[0003] Existing gas sensor array detection methods typically collect steady-state response values ​​after the sensors reach a stable state. These values ​​from multiple sensors are combined into a feature vector, which is then input into a pattern recognition model such as a neural network or support vector machine for training. The trained model is then used to identify the type and predict the concentration of unknown gas samples. Some improved methods incorporate environmental temperature and humidity sensors, using temperature and humidity measurements as auxiliary input features to model the gas along with the sensor response values, thus reducing the impact of environmental changes on detection accuracy. Other research attempts to extract dynamic features such as the rise rate or response time of the sensor response curve to supplement steady-state response information and improve the model's generalization ability.

[0004] However, the single steady-state response value of existing methods only reflects the final state after gas adsorption equilibrium and cannot fully describe the dynamic process of gas-sensitive material interaction. When multiple gases are present at the same time, different gas molecules compete for adsorption on the sensor surface, causing mutual interference in the adsorption rate, equilibrium state and desorption characteristics of each gas. It is difficult to effectively separate the signal contribution of each component in the mixed gas based solely on the steady-state response value. The method of modeling temperature and humidity as independent input features is a passive correction and fails to solve the nonlinear interference caused by competitive adsorption at the mechanistic level. Moreover, the extraction method of dynamic features is relatively simple, usually using only a single dynamic index, which fails to make full use of the rich information of the entire process of adsorption rise, equilibrium stability and desorption fall, thus limiting the accuracy and reliability of multi-component concentration inversion in mixed gas. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-sensor array-based identification and concentration quantification system to solve the following technical problems: Existing methods rely on steady-state response values, neglecting dynamic process information of gas adsorption and desorption. Temperature and humidity corrections are passive compensations and cannot solve the nonlinear interference caused by competitive adsorption at the mechanistic level. The dynamic feature extraction method is simple and does not make full use of the rich information of the entire process of adsorption rise, equilibrium stability, and desorption fall, which limits the accuracy of multi-component concentration inversion in mixed gases.

[0006] The objective of this invention can be achieved through the following technical solutions: Various identification and concentration quantification detection systems based on sensor arrays include: The curve acquisition module is used to continuously acquire the resistance value of each sensor element at fixed time intervals during the process of the gas sensor array being exposed to the gas mixture to be tested, and form the resistance response curve of each sensor element over time. The feature extraction module is used to extract segmented features from the resistance response curve of each sensor element to obtain a set of feature parameters that reflect the dynamic process of gas adsorption and desorption. The vector fusion module is used to simultaneously collect the ambient temperature and relative humidity values ​​of the environment where the gas sensor array is located, and combine the feature parameter set with the ambient temperature and relative humidity values ​​into a multi-dimensional feature vector. The decoupling inference module is used to input multi-dimensional feature vectors into a pre-trained adsorption competition decoupling neural network, which outputs an initial concentration estimate for each gas to be detected and a gas competition influence matrix. The concentration inversion module is used to obtain the actual concentration value of each gas to be detected in the gas mixture by solving a system of linear equations based on the initial concentration estimate of each gas to be detected and the inter-gas competition influence matrix. The storage output module is used to associate and store the calculated actual concentration value of each gas to be detected with the feature parameter set, and then output it to the display device after arranging it in the preset gas type order.

[0007] As a further aspect of the present invention: the specific process of segmenting feature extraction for the resistance response curve of each sensor element in the feature extraction module is as follows: Starting from the moment when the gas sensor array is exposed to the gas mixture to be tested, the first and second derivatives of the resistance value of each sensor element with respect to time are continuously calculated. The moment when the first derivative continuously decreases from a positive value to close to zero and the second derivative turns from negative to positive is marked as the end of the adsorption rising phase. The period when the first derivative fluctuates near zero and the absolute value of the second derivative is less than a preset fluctuation threshold is marked as the equilibrium and stability phase. Starting from the moment when the gas mixture to be tested is stopped being introduced into the gas sensor array, the moment when the first derivative is negative and gradually rises back to close to zero and the second derivative turns from positive to negative is marked as the end of the desorption falling phase. The adsorption rate integral, adsorption time constant, and adsorption capacity characteristic value are extracted from the adsorption rise phase; the equilibrium resistance value and equilibrium fluctuation variance value are extracted from the equilibrium stability phase; and the desorption rate integral, desorption time constant, and desorption residual ratio characteristic value are extracted from the desorption fall phase. The above eight characteristic values ​​are used as the characteristic parameter set of the sensor element.

[0008] As a further aspect of the present invention, the specific extraction process of the adsorption rate integral value, adsorption time constant, and adsorption capacity characteristic value is as follows: During the adsorption ascent phase, the first derivative of the resistance value with respect to time is integrated, and the result is used as the adsorption rate integral value. The resistance response curve during the adsorption ascent phase is fitted using a first-order exponential ascent model with nonlinear least squares, and the fitted time constant parameter is used as the adsorption time constant. The difference between the resistance value at the end of the adsorption ascent phase and the resistance value at the beginning of the adsorption ascent phase is calculated, and the difference is divided by the difference between the average resistance value during the equilibrium and stable phase and the resistance value at the beginning of the adsorption ascent phase. The result of the division operation is used as the adsorption capacity characteristic value.

[0009] As a further aspect of the present invention, the specific extraction process of the desorption rate integral value, desorption time constant, and desorption residue ratio characteristic value is as follows: During the desorption descent phase, the absolute value of the first derivative of the resistance value with respect to time is integrated, and the result is used as the desorption rate integral value. The resistance response curve during the desorption descent phase is fitted using a first-order exponential decay model with nonlinear least squares, and the fitted time constant parameter is used as the desorption time constant. The first difference between the resistance value at the end of the desorption descent phase and the resistance value at the beginning of the desorption descent phase is calculated, and the second difference between the resistance value at the beginning of the desorption descent phase and the average resistance value during the equilibrium and stable phase is calculated. The ratio of the first difference to the second difference is obtained, and the desorption residue ratio characteristic value is obtained by subtracting the ratio from 1.

[0010] As a further aspect of the present invention: the specific process of combining the feature parameter set with the ambient temperature value and the ambient relative humidity value into a multidimensional feature vector in the vector fusion module is as follows: The sensor elements are sorted in descending order according to the peak response reached by each sensor element during the adsorption and ascent phase. The eight characteristic parameters of each sensor element are arranged in the sorted order to form a characteristic parameter sequence. The ambient temperature value and the ambient relative humidity value are inserted into the middle position of the characteristic parameter sequence to form a multidimensional feature vector with a dimension equal to the product of the number of sensor elements and 8 plus 2. Before inputting the multidimensional feature vector into the adsorption competition decoupling neural network, the values ​​of each dimension are dynamically normalized based on the reference gas response. The dynamic normalization process uses the average resistance value of all sensor elements in the equilibrium and stable phase during the current measurement cycle as the reference value. The values ​​of each dimension are divided by the corresponding reference value and mapped to a closed interval of 0 to 2.

[0011] As a further aspect of the present invention: the specific structure of the adsorption competition decoupling neural network in the decoupling inference module is as follows: The adsorption competition decoupling neural network consists of an input layer, three competitive feature extraction layers, two competitive interaction layers, and an output layer. The number of nodes in the input layer is set to the dimension of the multidimensional feature vector. The three competitive feature extraction layers sequentially use a two-dimensional convolutional structure with progressively larger kernel size to perform spatial transformation on the input features. Each competitive feature extraction layer is followed by a batch normalization layer and a leakage correction linear unit activation function. The two competitive interaction layers adopt a graph attention network structure, using the features output by the competitive feature extraction layer as graph node features and the prior competitive relationship between the gas types to be detected as the initial value of the graph edge weights. The graph node features and edge weights are updated through a multi-head attention mechanism.

[0012] As a further aspect of the present invention: the specific training process of the adsorption competition decoupling neural network is as follows: Samples of single and multiple gases with different concentration combinations were prepared. For each sample, the curve acquisition module and the vector fusion module were repeatedly executed to obtain multi-dimensional feature vectors as training input samples. A two-stage training strategy was adopted. In the first stage, single gas samples were used for pre-training. In the initial concentration estimate expected by the output layer, the gas was set to a known concentration value, other gases were set to zero, the gas competition influence matrix was set to an identity matrix, and the loss function was a weighted sum of mean square error and relative error of predicted concentration. The second stage uses a multi-gas mixed sample for fine-tuning. The actual concentration values ​​of each gas in the mixed sample are determined by gas chromatography as a benchmark. The expected initial concentration estimate of the output layer is set as the known concentration value of each gas. The matrix is ​​solved by iterative optimization to minimize the error between the solution result of the linear equation system in the concentration inversion module and the value determined by gas chromatography. The matrix obtained by the solution is used as the expected inter-gas competition influence matrix. At the same time, matrix symmetry and diagonal dominance constraints are introduced as regularization terms into the loss function.

[0013] As a further aspect of the present invention: the output layer of the adsorption competition decoupling neural network is specifically configured as follows: The number of output layer nodes is set to the number of gas types to be detected plus the square of the number of gas types to be detected. The first few nodes of the output layer output the initial concentration estimate of each gas. The remaining nodes of the output layer output each element in the inter-gas competition influence matrix in the order of row first and column second. The inter-gas competition influence matrix is ​​a square matrix with the number of rows and columns equal to the number of gas types to be detected. The diagonal elements in the matrix are set to a fixed value of 1, and the off-diagonal elements are normalized to the interval of 0 to 1 by a flexible maximum value function and then used as the mutual inhibition coefficient between the corresponding two gases.

[0014] As a further aspect of the present invention: the specific process by which the concentration inversion module obtains the actual concentration value of each gas to be detected in the gas mixture by solving a system of linear equations is as follows: Let the number of gases to be detected be n. Construct a column vector of length n from the initial concentration estimates. Treat the inter-gas competition influence matrix as an n-order square matrix. Construct a system of linear equations in which the product of the square matrix and the unknown concentration column vector equals the initial concentration estimate column vector. Decompose the inter-gas competition influence matrix into the sum of a diagonal matrix, a strictly upper triangular matrix, and a strictly lower triangular matrix. Calculate the matrix using the Gauss-Seidel iterative format. After each iteration, apply non-negativity constraints and upper concentration constraints to the unknown concentration column vector. Stop when the Euclidean distance between two adjacent iterations is less than the convergence threshold. Use the converged column vector elements as the corresponding actual gas concentration values. Store the characteristic parameter set, ambient temperature value, ambient relative humidity value, initial concentration estimate, inter-gas competition influence matrix, iterative intermediate variable sequence, and actual concentration value obtained within the same measurement period as a single historical measurement record.

[0015] The beneficial effects of this invention are: This invention addresses the limitation in the accuracy of multi-component concentration inversion in mixed gases by introducing a dynamic response segmentation and competitive decoupling mechanism, thus resolving the issue at the mechanistic level. First, the sensor's resistance response curve is segmented into its entirety—adsorption rise, equilibrium stabilization, and desorption fall—to extract eight dynamic feature parameters, including the adsorption rate integral, adsorption time constant, and desorption residue ratio. This fully preserves the adsorption and desorption information at each stage of the gas-sensitive material interaction, overcoming the insufficient information provided by a single steady-state response value. Second, the dynamic feature parameters, along with ambient temperature and humidity, are used to construct a multi-dimensional feature vector, which is then input into an adsorption competition decoupling neural network. This network employs a competitive feature extraction layer and a graph attention interaction layer structure. Through a two-stage training strategy, it learns the response characteristics of individual gases and the competitive relationship between the mixed gases, outputting an initial concentration estimate and a gas competition influence matrix. This mathematically decouples the mutual inhibition levels of different gases on the sensor surface. Finally, a linear equation system is constructed by combining the initial concentration estimate with the competition influence matrix. This system is then solved using a physically constrained Gaussian-Seidel iterative approach, with non-negativity and upper concentration limits applied. The actual concentrations of each component in the gas mixture are then retrieved through inversion. The dynamic characteristics, environmental parameters, competition matrix, and concentration values ​​from the same measurement period are correlated and stored as a complete historical record. This invention forms a complete technical chain from dynamic response capture and competition relationship decoupling to concentration inversion, significantly improving the accuracy and reliability of multi-component concentration measurements in gas mixtures. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a multi-sensor array-based identification and concentration quantification detection system, comprising: The curve acquisition module performs data acquisition while the gas sensor array is exposed to the test gas mixture. The gas sensor array consists of 25 metal oxide semiconductor gas-sensitive elements, which have cross-sensitivity to volatile organic compounds. The curve acquisition module continuously acquires the resistance value of each sensor element at a fixed time interval of 10 times per second, and records the acquisition time in relation to the corresponding resistance value to form a resistance response curve of each sensor element over time. When the test gas mixture is a single gas or a mixture of two gases selected from formaldehyde, ethanol, ethyl acetate, benzene, and pentamidine, the curve acquisition module continuously acquires all resistance data from the moment the gas is introduced until 120 seconds after the gas is stopped.

[0020] The feature extraction module performs segmented feature extraction on the resistance response curve of each sensor element. Starting from the initial moment when the gas sensor array is exposed to the test gas mixture, the feature extraction module continuously calculates the first and second derivatives of the resistance value of each sensor element with respect to time. The feature extraction module marks the moment when the first derivative continuously decreases from a positive value to near zero and the second derivative turns from negative to positive as the end of the adsorption rising phase, and marks the period when the first derivative fluctuates near zero and the absolute value of the second derivative is less than a preset fluctuation threshold of 0.01 as the equilibrium and stability phase. Starting from the moment when the test gas mixture is stopped being introduced into the gas sensor array, the feature extraction module marks the moment when the first derivative is negative and gradually rises back to near zero while the second derivative turns from positive to negative as the end of the desorption falling phase. The feature extraction module extracts the adsorption rate integral value, adsorption time constant, and adsorption capacity feature value from the adsorption rise phase, the equilibrium resistance value and equilibrium fluctuation variance value from the equilibrium stability phase, and the desorption rate integral value, desorption time constant, and desorption residue ratio feature value from the desorption fall phase. The above eight feature values ​​are used as the feature parameter set of the sensor element.

[0021] The vector fusion module synchronously acquires the ambient temperature and relative humidity values ​​of the environment where the gas sensor array is located. The module sorts the sensor elements in descending order based on the peak response value reached by each element during the adsorption rise phase. Then, it arranges the eight characteristic parameters of each sensor element in sequence according to this sorted order, forming a characteristic parameter sequence. The module inserts the ambient temperature and relative humidity values ​​into the middle of this sequence, creating a 202-dimensional feature vector with dimensions equal to the product of the number of sensor elements (25 + 8) plus 2. The module then performs dynamic normalization on each dimension of the feature vector based on the reference gas response. Using the average resistance value of all sensor elements during the equilibrium and stable phase within the current measurement period as the baseline value, each dimension's value is divided by the corresponding baseline value and mapped to a closed interval between 0 and 2.

[0022] The decoupling inference module inputs the normalized 202-dimensional multidimensional feature vector into a pre-trained adsorption competition decoupling neural network. The adsorption competition decoupling neural network consists of one input layer, three competitive feature extraction layers, two competitive interaction layers, and one output layer. The input layer has 202 nodes. The three competitive feature extraction layers sequentially use two-dimensional convolutional structures with kernel sizes of 3, 5, and 7 to spatially transform the input features. Each competitive feature extraction layer is followed by a batch normalization layer and a leakage correction linear unit activation function. The two competitive interaction layers employ a graph attention network structure, using the features output by the competitive feature extraction layers as graph node features and the prior competition relationships among the five gases (formaldehyde, ethanol, ethyl acetate, benzene, and pentanone) as initial values ​​for graph edge weights. A multi-head attention mechanism updates the graph node features and edge weights. The output layer has 5 + 5 squared = 30 nodes. The first 5 nodes of the output layer output the initial concentration estimates for each gas to be detected, and the remaining 25 nodes output the elements of the gas competition influence matrix in row-first, column-last order. The gas competition influence matrix is ​​a 5x5 square matrix. The diagonal elements in the matrix are set to a fixed value of 1, and the off-diagonal elements are normalized to the range of 0 to 1 using a flexible maximum value function and then used as the mutual inhibition coefficients between the two corresponding gases.

[0023] The adsorption competition decoupling neural network employs a two-stage training strategy. Single-gas samples and two-gas mixtures with varying concentrations are prepared. For each sample, the curve acquisition, feature extraction, and vector fusion modules are repeatedly executed to obtain multi-dimensional feature vectors, which serve as training input samples. In the first stage, pre-training is performed using single-gas samples. The initial concentration estimate of the desired gas in the output layer is set to a known concentration value, while other gases are set to 0. The inter-gas competition influence matrix is ​​set to an identity matrix, and the loss function is a weighted sum of mean square error and relative error of predicted concentration. In the second stage, fine-tuning is performed using two-gas mixtures. The actual concentration values ​​of each gas in the mixture are determined by gas chromatography as a baseline. The initial concentration estimate of the desired gas in the output layer is set to the known concentration values ​​of each gas. The matrix is ​​iteratively optimized to minimize the error between the linear equations solved by the subsequent concentration inversion module and the gas chromatography measurement values. The resulting matrix is ​​used as the desired inter-gas competition influence matrix. Matrix symmetry and diagonal dominance constraints are introduced as regularization terms in the loss function.

[0024] The concentration inversion module calculates the actual concentration value based on the initial concentration estimate and the inter-gas competition influence matrix. The module denotes the number of gases to be detected (5 as n), constructs a column vector of length 5 from the initial concentration estimates, and treats the inter-gas competition influence matrix as a 5th-order square matrix. It then constructs a system of linear equations where the product of this square matrix and the unknown concentration column vector equals the initial concentration estimate column vector. The module decomposes the inter-gas competition influence matrix into the sum of a diagonal matrix, a strictly upper triangular matrix, and a strictly lower triangular matrix, and calculates the values ​​using a Gauss-Seidel iterative scheme. After each iteration, a non-negativity constraint and a concentration upper limit constraint are applied to the unknown concentration column vector. The upper limit for formaldehyde and benzene is set to 10 ppm, while the upper limit for ethanol, ethyl acetate, and pentanone is set to 100 ppm. Iteration stops when the Euclidean distance between two consecutive iterations is less than the convergence threshold of 0.001, and the elements of the converged column vector are taken as the actual concentration values ​​of the corresponding gases.

[0025] The storage output module associates and stores the characteristic parameter sets of 25 sensor elements obtained within the same measurement period, along with ambient temperature, ambient relative humidity, initial concentration estimates, inter-gas competition matrix, iterative intermediate variable sequences, and actual concentration values, into a single historical measurement record. The storage output module then arranges the actual concentration values ​​in a preset gas type order (formaldehyde, ethanol, ethyl acetate, benzene, and pentanone) and outputs them to the display device for display.

[0026] In a preferred embodiment of the present invention, the specific process of segmenting feature extraction for the resistance response curve of each sensor element in the feature extraction module is as follows: The feature extraction module first calculates the first and second derivatives of the resistance value of each sensor element with respect to time, starting from the initial moment when the gas sensor array is exposed to the gas mixture to be measured. The first derivative reflects the rate of change of the resistance value with time, and the second derivative reflects the trend of this rate of change. The feature extraction module monitors the first and second derivative values ​​of each sensor element in real time to delineate different stages of the resistance response curve.

[0027] The feature extraction module marks the end of the adsorption rise phase as the moment when the first derivative continuously decreases from a positive value to near zero and the second derivative turns from negative to positive. During the adsorption rise phase, gas molecules begin to adsorb onto the surface of the sensor's sensitive material, causing the resistance to rise rapidly, and the first derivative is positive. As adsorption gradually approaches saturation, the rate of increase in resistance slows down, and the first derivative begins to decrease continuously from its maximum value. When the first derivative decreases to near zero, it indicates that the resistance has essentially stopped increasing, and at this point, the second derivative turns from negative to positive, marking the transition of the adsorption process from a rapid rise to a stable state. The feature extraction module defines the time interval from the start of gas exposure to this marked moment as the adsorption rise phase.

[0028] The feature extraction module marks periods where the first derivative fluctuates around zero and the absolute value of the second derivative is less than a preset fluctuation threshold as a stable equilibrium phase. After the adsorption rise phase ends, the resistance value enters a relatively stable equilibrium state, with the first derivative fluctuating slightly around zero and the absolute value of the second derivative also being relatively small. The feature extraction module presets a fluctuation threshold of 0.01. When the absolute values ​​of the first and second derivatives at multiple consecutive sampling points are all less than this threshold, the current period is determined to be in a stable equilibrium phase. The feature extraction module records the start time of this period as the end time of the adsorption rise phase, and the end time is triggered by subsequent desorption operations.

[0029] Starting from the moment the gas mixture to be tested stops flowing into the gas sensor array, the feature extraction module marks the moment when the first derivative is negative and gradually rises to near zero, while the second derivative turns from positive to negative, as the end of the desorption descent phase. After the gas flow stops, gas molecules adsorbed on the sensor surface begin to desorb, the resistance gradually decreases, and the first derivative becomes negative. As the desorption process continues, the rate of resistance decrease slows down, and the first derivative gradually rises from its minimum negative value to near zero. When the second derivative turns from positive to negative, it indicates that the desorption process has transitioned from a rapid descent to a slow recovery. The feature extraction module defines the time interval from the moment the gas flow stops to this marked moment as the desorption descent phase.

[0030] Through the above process, the feature extraction module divides the resistance response curve of each sensor element into three consecutive time intervals: the adsorption rising stage, the equilibrium stabilization stage, and the desorption falling stage, laying the foundation for subsequent feature extraction.

[0031] The feature extraction module extracts three parameters from the adsorption ascent phase: the adsorption rate integral value, the adsorption time constant, and the adsorption capacity feature value.

[0032] The adsorption rate integral value characterizes the total adsorption amount during the adsorption ascent phase. Within the time interval of the adsorption ascent phase, the feature extraction module integrates the resistance value at each sampling moment with respect to the first derivative with respect to time. Since the first derivative represents the instantaneous adsorption rate, integrating with time yields the total change in resistance value during that phase, reflecting the total adsorption capacity of the sensitive material for the gas. Specifically, the feature extraction module uses a numerical integration method, accumulating the product of the first derivative values ​​of each sampling interval within the adsorption ascent phase and the time interval; the accumulated result is used as the adsorption rate integral value.

[0033] The adsorption time constant characterizes the rate of adsorption. The feature extraction module uses a first-order exponential rise model to perform nonlinear least squares fitting on the resistance response curve during the adsorption ascent phase. The mathematical form of the first-order exponential rise model is Rt equal to R0 plus A multiplied by the negative t of 1 minus e (in parentheses) divided by τ, where Rt is the resistance value at time t, R0 is the resistance value at the start of adsorption, A is the adsorption amplitude, and τ is the adsorption time constant. The feature extraction module aims to minimize the sum of squared errors between the actual resistance values ​​and the model predictions at each sampling point during the adsorption ascent phase. It solves for the model parameters through an iterative optimization algorithm and uses the fitted time constant τ as the adsorption time constant. The smaller the value of this parameter, the faster the adsorption rate, and vice versa.

[0034] The adsorption capacity characteristic value represents the proportion of the adsorbed amount at the end of the adsorption ascent phase to the theoretical maximum adsorption amount. The feature extraction module first calculates the difference between the resistance value at the end of the adsorption ascent phase and the resistance value at the beginning of the adsorption ascent phase, denoted as ΔRrise. Then, it calculates the difference between the average resistance value of the equilibrium steady-state phase and the resistance value at the beginning of the adsorption ascent phase, denoted as ΔReq. The average resistance value of the equilibrium steady-state phase is obtained by taking the arithmetic mean of the resistance values ​​at all sampling points within this phase. The feature extraction module divides ΔRrise by ΔReq, and the result of the division operation is used as the adsorption capacity characteristic value. A value closer to 1 indicates more complete adsorption, while a value closer to 0 indicates less complete adsorption.

[0035] The feature extraction module extracts two parameters from the equilibrium and stable phase: the equilibrium resistance value and the equilibrium fluctuation variance value.

[0036] The equilibrium resistance value characterizes the stable response level of the sensor under adsorption equilibrium conditions. The feature extraction module calculates the arithmetic mean of the resistance values ​​at all sampling points during the equilibrium steady-state phase, and uses this average value as the equilibrium resistance value. This parameter reflects the steady-state response strength of the sensor to the target gas and is a commonly used single feature in traditional methods, but in this invention, it is only used as a supplement to the dynamic features.

[0037] The equilibrium fluctuation variance value characterizes the signal stability during the equilibrium and stable phase. The feature extraction module calculates the squared difference between the resistance value at each sampling point and the equilibrium resistance value during the equilibrium and stable phase. The sum of all squared differences is then divided by the number of sampling points to obtain the variance value for that phase. This variance value is used as the equilibrium fluctuation variance value. A smaller value indicates a more stable signal during the equilibrium phase, while a larger value indicates more severe signal fluctuations, reflecting environmental interference or sensor noise levels.

[0038] The feature extraction module extracts three parameters from the desorption descent phase: the desorption rate integral value, the desorption time constant, and the desorption residual ratio feature value.

[0039] The desorption rate integral value represents the total amount of desorption during the desorption descent phase. The feature extraction module integrates the absolute value of the first derivative of the resistance value with respect to time within the time interval of the desorption descent phase. Since the first derivative is negative during the desorption phase, integrating the absolute value yields the total decrease in resistance during this phase, reflecting the desorption capacity of the adsorbed gas on the sensitive material. Specifically, the feature extraction module accumulates the products of the absolute values ​​of the first derivatives of each sampling interval within the desorption descent phase and the time interval; the accumulated result is used as the desorption rate integral value.

[0040] The desorption time constant characterizes the speed of the desorption process. The feature extraction module uses a first-order exponential decay model to perform nonlinear least squares fitting on the resistance response curve during the desorption descent phase. The mathematical form of the first-order exponential decay model is Rt equal to Rend plus B multiplied by e^(-t) divided by τd, where Rt is the resistance value at time t, Rend is the resistance value at the end of desorption, B is the desorption amplitude, and τd is the desorption time constant. The feature extraction module aims to minimize the sum of squared errors between the actual resistance values ​​and the model predictions at each sampling point during the desorption descent phase. It solves for the model parameters through an iterative optimization algorithm and uses the fitted time constant τd as the desorption time constant. The smaller the value of this parameter, the faster the desorption rate, and vice versa.

[0041] The desorption-residue ratio characteristic value characterizes the proportion of gas remaining on the sensor surface after the desorption stage ends. The feature extraction module first calculates the first difference between the resistance value at the end of the desorption descent stage and the resistance value at the beginning of the desorption descent stage; this difference represents the actual resistance decrease during the desorption stage. Then, it calculates the second difference between the resistance value at the beginning of the desorption descent stage and the average resistance value during the equilibrium and stable stage; this difference represents the total resistance decrease theoretically that should occur from equilibrium to complete desorption. The feature extraction module obtains the ratio of the first difference to the second difference; this ratio represents the proportion of actual desorption to the total theoretical desorption. Subtracting this ratio from 1 yields the desorption-residue ratio characteristic value. For example, if the average resistance during the equilibrium steady-state phase is 5000 ohms, the resistance at the start of desorption is 5000 ohms, and the resistance at the end of desorption is 1200 ohms, then the second difference is 5000 minus 1000 equals 4000 ohms, the first difference is 5000 minus 1200 equals 3800 ohms, the ratio is 3800 divided by 4000 equals 0.95, and the characteristic value of the desorption residue ratio is 1 minus 0.95 equals 0.05, indicating that 5% of the response residue was not completely desorbed. The closer this characteristic value is to 0, the more complete the desorption; the closer it is to 1, the more severe the residue.

[0042] The feature extraction module uses the above eight feature values ​​as the feature parameter set of the sensor element: adsorption rate integral value, adsorption time constant, adsorption capacity feature value, equilibrium resistance value, equilibrium fluctuation variance value, desorption rate integral value, desorption time constant, and desorption residue ratio feature value.

[0043] For each sensor element in the gas sensor array, the feature extraction module performs the above segmentation and feature extraction operations. When the sensor array contains 25 sensor elements, the feature extraction module obtains a total of 25 sets of feature parameters, each containing 8 feature values, for a total of 200 feature values. These feature values ​​fully describe the dynamic response characteristics of each sensor element throughout the entire process of adsorption rise, equilibrium stabilization, and desorption fall, providing a rich information foundation for subsequent mixture composition inversion.

[0044] In another preferred embodiment of the present invention, the specific process of combining the feature parameter set with the ambient temperature value and the ambient relative humidity value into a multidimensional feature vector in the vector fusion module is as follows: The vector fusion module first acquires the feature parameter set of each sensor element output by the feature extraction module. Each sensor element's feature parameter set includes eight feature values: adsorption rate integral, adsorption time constant, adsorption capacity feature value, equilibrium resistance value, equilibrium fluctuation variance value, desorption rate integral, desorption time constant, and desorption residue ratio feature value. When the gas sensor array contains 25 sensor elements, the vector fusion module obtains 25 sets of feature parameters, totaling 200 feature values.

[0045] The vector fusion module sorts the sensor elements. It obtains the peak response value of each sensor element during the adsorption ascent phase, where the peak response value is the maximum resistance value of that sensor element during this phase. The module then sorts the 25 sensor elements in descending order of their peak response values. Sensor elements with larger peak response values ​​are more sensitive to the current gas mixture being measured. Placing these sensor elements at the top of the sequence helps the subsequent neural network prioritize the response information of sensitive elements. For example, in a measurement, if sensor element number 7 has a peak response value of 5200 ohms, sensor element number 12 has a peak response value of 4800 ohms, and sensor element number 3 has a peak response value of 4500 ohms, then after sorting, sensor element number 7 is ranked first, sensor element number 12 is ranked second, sensor element number 3 is ranked third, and the remaining sensor elements are arranged sequentially.

[0046] The vector fusion module arranges the eight characteristic parameters of each sensor element in the sorted order, forming a sequence of characteristic parameters. Taking the first sensor element as an example, the vector fusion module first arranges the adsorption rate integral value, then its adsorption time constant, and then the adsorption capacity characteristic value, equilibrium resistance value, equilibrium fluctuation variance value, desorption rate integral value, desorption time constant, and desorption residue ratio characteristic value. Then it arranges the eight characteristic values ​​of the second sensor element, and so on, until all 200 characteristic values ​​of the 25 sensor elements are arranged.

[0047] The vector fusion module simultaneously acquires the ambient temperature and relative humidity values ​​of the environment surrounding the gas sensor array. The ambient temperature value is read from a digital temperature sensor, in degrees Celsius. The ambient relative humidity value is read from a digital humidity sensor, in percentage. The vector fusion module inserts the ambient temperature and relative humidity values ​​into the middle of the feature parameter sequence. The middle position refers to the area near the midpoint between the two halves of the feature parameter sequence. Specifically, the ambient temperature value can be inserted after the 100th feature value, and the ambient relative humidity value before the 101st feature value, or the two environmental parameters can be inserted consecutively in the exact middle of the sequence. By inserting the environmental parameters into the middle of the sequence, the neural network can simultaneously consider the interaction between the sensor features and environmental parameters in both halves when processing the feature vector. After inserting the environmental parameters, the vector fusion module forms a 202-dimensional multidimensional feature vector with dimensions equal to the product of 25 and 8 (the number of sensor elements) plus 2. The first to the 200th dimensions of this multidimensional feature vector are the 200 sensor feature values ​​after sorting, and the two dimensions near the 101st or 102nd dimension are the ambient temperature value and the ambient relative humidity value.

[0048] Before the multi-dimensional feature vectors are input into the adsorption competition decoupling neural network, the vector fusion module performs dynamic normalization processing on the values ​​of each dimension based on the reference gas response. The dynamic normalization process uses the average resistance value of all sensor elements during the equilibrium stabilization phase within the current measurement cycle as the reference value. Specifically, the vector fusion module obtains the equilibrium resistance value of each sensor element calculated by the feature extraction module during the equilibrium stabilization phase; this equilibrium resistance value is the arithmetic mean of the resistance values ​​at all sampling points during that phase. The vector fusion module calculates the arithmetic mean of the equilibrium resistance values ​​of 25 sensor elements and uses this average value as the reference reference value for the current measurement cycle. For example, if the equilibrium resistance values ​​of the 25 sensor elements are 5120 ohms, 4980 ohms, 5230 ohms up to 4850 ohms, and the calculated average is 5050 ohms, then the reference reference value for the current measurement cycle is 5050 ohms.

[0049] The vector fusion module performs dynamic normalization on the values ​​of each dimension of the multidimensional feature vector. For dimensions derived from sensor feature parameters, the feature value of that dimension is divided by a reference value. For example, if the adsorption rate integral value of a sensor element is 3800 and the reference value is 5050, the normalized value is 3800 divided by 5050, which is approximately 0.752. For the two dimensions of ambient temperature and relative humidity, their values ​​are also divided by the reference value. Ambient temperature values ​​are typically between 20 and 30 degrees Celsius, and relative humidity values ​​are between 20 and 95%. After dividing by the reference value, these values ​​are mapped to a closed interval of 0 to 2. For example, if the ambient temperature is 25 degrees Celsius and the reference value is 5050, the normalized value is 25 divided by 5050, which is approximately 0.005. Although this value is extremely small, it preserves the relative proportionality. After dynamic normalization, the values ​​of all dimensions are distributed within a closed interval of 0 to 2. The vector fusion module outputs the normalized 202-dimensional multidimensional feature vector to the adsorption competition decoupling neural network.

[0050] In another preferred embodiment of the present invention, the specific structure of the adsorption competition decoupling neural network in the decoupling inference module is as follows: The adsorption-competitive decoupling neural network consists of one input layer, three competitive feature extraction layers, two competitive interaction layers, and one output layer. The number of nodes in the input layer is set to the number of dimensions of the multidimensional feature vector, i.e., 202 nodes. Each node receives a value from one dimension of the normalized 202-dimensional multidimensional feature vector.

[0051] Three competing feature extraction layers sequentially employ a two-dimensional convolutional structure with progressively larger kernel sizes to spatially transform the input features. The first competing feature extraction layer uses a 3x3 two-dimensional convolution with 128 kernels, a stride of 1, and padding to maintain the feature map size. This layer rearranges the 202-dimensional input features, converting the one-dimensional feature vector into a two-dimensional feature map (e.g., rearranging the 202-dimensional features into a 14x15 two-dimensional feature map, padding with zeros if necessary). After the first convolutional operation, 128 feature maps are output. The second competing feature extraction layer uses a 5x5 two-dimensional convolution with 64 kernels, a stride of 1, and padding to maintain the feature map size. It performs the convolution operation on the 128 feature maps output from the first layer, outputting 64 feature maps. The third competitive feature extraction layer employs a 7x7 2D convolution with 32 kernels and a stride of 1. Padding is set to maintain the feature map size. It convolves the 64 feature maps output from the second layer, outputting 32 feature maps. By progressively increasing the kernel size, the three competitive feature extraction layers gradually expand the receptive field, capturing a wider range of interactive information between sensor features. Each competitive feature extraction layer is followed by a batch normalization layer and a leakage correction linear unit activation function. The batch normalization layer normalizes each batch of input data, stabilizing the data distribution. The leakage correction linear unit activation function assigns a small positive slope (0.01) to negative inputs, preventing neuron death.

[0052] The two competitive interaction layers employ a graph attention network structure. The first competitive interaction layer flattens the 32 feature maps output by the third competitive feature extraction layer into a one-dimensional feature vector. The dimension of this feature vector is 32 multiplied by the height and width of the feature map. Each element in this feature vector is considered a graph node, and the number of graph nodes equals the dimension of the feature vector. The first competitive interaction layer uses the prior competition relationships between the gas types to be detected as the initial values ​​for the graph edge weights. Five gas types are selected: formaldehyde, ethanol, ethyl acetate, benzene, and pentamphenone. The prior competition relationships are pre-set based on the similarity of the chemical structures of the gas molecules. For example, the initial competition coefficient between formaldehyde and ethanol is set to 0.3, between formaldehyde and ethyl acetate to 0.4, between benzene and pentamphenone to 0.6, and the competition coefficient for the same gas itself is set to 0. The prior competition relationships between the five gases are extended to the dimension of the number of graph nodes. Each graph node corresponds to a competition weight vector with a length of 5. The first competitive interaction layer updates the graph node features and edge weights through a multi-head attention mechanism. The multi-head attention mechanism uses eight attention heads, each independently calculating the attention coefficients between nodes. The outputs of the eight attention heads are concatenated to obtain the updated node features. The attention coefficients are calculated by taking the dot product of the current node's features and those of its neighboring nodes, then normalizing using a flexible maximum function. The edge weights are updated by taking a weighted average of the attention coefficients and the initial values ​​of the original edge weights. The second competitive interaction layer uses the same graph attention network structure as the first competitive interaction layer, taking the updated node features and edge weights from the first layer as input, and then updating again using the multi-head attention mechanism to further strengthen the representation of the competitive relationships between nodes. Through the processing of the two competitive interaction layers, the network fully learns the mutual inhibition relationships between different gas types in the sensor response.

[0053] The specific training process of the adsorption competition decoupling neural network is as follows. First, single gas samples and two-gas mixtures with different concentration combinations are prepared. Single gas samples include formaldehyde, ethanol, ethyl acetate, benzene, and pentamidine at different concentrations, ranging from 0 to 100 ppm. At least 10 different concentration points are prepared for each gas, and each concentration point is sampled 5 times. Two-gas mixtures include 10 combinations: formaldehyde and ethanol, formaldehyde and ethyl acetate, formaldehyde and benzene, formaldehyde and pentamidine, ethanol and ethyl acetate, ethanol and benzene, ethanol and pentamidine, ethyl acetate and benzene, ethyl acetate and pentamidine, and benzene and pentamidine. At least 5 different concentration ratios are prepared for each combination, and each ratio is sampled 5 times. For each sample, the curve acquisition module, feature extraction module, and vector fusion module are repeatedly executed to obtain a multi-dimensional feature vector as the training input sample.

[0054] A two-stage training strategy is employed for network training. The first stage uses a single gas sample for pre-training. The training input sample is a multi-dimensional feature vector of a single gas sample. In the output layer's expected initial concentration estimate, the gas is set to a known concentration value, while other gases are set to zero. For example, for a formaldehyde sample with a concentration of 50 ppm, the expected output initial concentration estimate for formaldehyde is 50, while the initial concentration estimates for ethanol, ethyl acetate, benzene, and pentanone are all 0. The expected inter-gas competition influence matrix for the output layer is set as an identity matrix, i.e., diagonal elements are 1s and off-diagonal elements are 0s. The loss function is a weighted sum of mean squared error and relative error of predicted concentration. The mean squared error is calculated as the average of the squared differences between the predicted and expected values ​​for all nodes in the output layer. The relative error of predicted concentration is calculated only for the gas type corresponding to the current sample, by dividing the absolute value of the difference between the predicted and expected concentration values ​​by the expected concentration value. The weight of the mean squared error in the weighted sum is set to 0.7, and the weight of the relative error of predicted concentration is set to 0.3. The network parameters are updated using the backpropagation algorithm, with the number of training iterations set to 100 and the learning rate set to 0.001.

[0055] The second stage uses a mixture of multiple gases for fine-tuning. The training input sample is a multidimensional feature vector of a mixture of two gases, and the initial concentration estimate expected by the output layer is set to the known concentration values ​​of each gas. For example, for a mixture of formaldehyde at 30 ppm and ethanol at 20 ppm, the expected initial concentration estimate for formaldehyde in the output is 30, for ethanol it is 20, and for ethyl acetate, benzene, and pentanone it is 0. The actual concentration values ​​of each gas in the mixture are determined by gas chromatography as the baseline concentration values. The gas competition influence matrix expected by the output layer is not pre-set to a specific value, but is solved by an iterative optimization algorithm to minimize the error between the solution of the linear equation system of the subsequent concentration inversion module and the gas chromatography measurement value. Specifically, during training, the gas competition influence matrix is ​​treated as a learnable parameter matrix, and each element in the matrix participates in backpropagation as a network parameter. For each training sample, the initial concentration estimate and the gas competition influence matrix output by the network are input into the concentration inversion module, which solves the linear equation system according to the Gauss-Seidel iterative method to obtain the calculated actual concentration values ​​of each gas. The mean squared error (MSE) between the calculated value and the gas chromatography-measured value is calculated and used as part of the loss function. Matrix symmetry and diagonal dominance constraints are also introduced as regularization terms into the loss function. The matrix symmetry constraint requires that the element in the i-th row and j-th column of the gas competition influence matrix be as equal as possible to the element in the j-th row and i-th column; the loss term is the sum of the squares of the differences between all elements i ≠ j. The diagonal dominance constraint requires that the diagonal element 1 is greater than the sum of all off-diagonal elements in that row; the loss term is the sum of the squares of the off-diagonal elements in all rows whose sum exceeds 1. The final loss function is a weighted sum of the mean squared error, the relative error of the predicted concentration, the matrix symmetry constraint, and the diagonal dominance constraint. The network parameters and the gas competition influence matrix parameters are updated simultaneously using a backpropagation algorithm. The number of training iterations is set to 200, and the learning rate is set to 0.0005. After training, the adsorption competition decoupling neural network can accurately output the initial concentration estimates of each gas and the gas competition influence matrix based on the input multidimensional feature vector.

[0056] The output layer of the adsorption competition decoupling neural network is specifically configured as follows. The number of output layer nodes is set to the number of gas species to be detected plus the square of the number of gas species to be detected. Since there are 5 gas species to be detected, the number of output layer nodes is 5 plus 5 squared, which is 30 nodes. The first 5 nodes of the output layer output the initial concentration estimate for each gas species, arranged in the order of formaldehyde, ethanol, ethyl acetate, benzene, and pentamidine. The first node outputs the initial concentration estimate for formaldehyde, the second for ethanol, the third for ethyl acetate, the fourth for benzene, and the fifth for pentamidine. The remaining 25 nodes of the output layer output the elements of the gas competition influence matrix in a row-first, column-later order. The gas competition influence matrix is ​​a 5x5 square matrix, with rows and columns corresponding to the order of formaldehyde, ethanol, ethyl acetate, benzene, and pentamidine. The 6th node corresponds to the element in the 1st row and 1st column of the matrix, the 7th node to the element in the 1st row and 2nd column, the 8th node to the element in the 1st row and 3rd column, the 9th node to the element in the 1st row and 4th column, the 10th node to the element in the 1st row and 5th column, the 11th node to the element in the 2nd row and 1st column, and so on until the 30th node corresponds to the element in the 5th row and 5th column. The output value of the diagonal elements (i.e., nodes in the 1st row and 1st column, 2nd row and 2nd column, 3rd row and 3rd column, 4th row and 4th column, and 5th row and 5th column) is set to a fixed value of 1, indicating that the competition coefficient of each gas against itself is 1. The off-diagonal elements are normalized to the interval between 0 and 1 using a flexible maximum function and then used as the mutual inhibition coefficient between the corresponding two gases. The flexible maximum function exponentially normalizes the output values ​​of all off-diagonal elements in the same row, making the sum of the off-diagonal elements in that row less than 1, thus satisfying the diagonal dominance property. For example, the element in the first row and second column of the matrix represents the inhibition coefficient of formaldehyde on ethanol. The closer the value is to 1, the stronger the inhibition of formaldehyde on ethanol; the closer the value is to 0, the weaker the inhibition. Through the above output layer settings, the adsorption competition decoupling neural network simultaneously outputs 5 initial concentration estimates and 25 competition coefficients in one forward calculation, fully describing the concentrations of each component in the mixed gas and their mutual inhibition relationships.

[0057] In another preferred embodiment of the present invention, the specific process of obtaining the actual concentration value of each gas to be detected in the gas mixture by solving a system of linear equations in the concentration inversion module is as follows: The number of gases to be detected was set to 5, including formaldehyde, ethanol, ethyl acetate, benzene, and pentanone.

[0058] The concentration inversion module first obtains the initial concentration estimates and the inter-gas competition influence matrix from the decoupled reasoning module. The initial concentration estimates are five values, arranged in the order of formaldehyde, ethanol, ethyl acetate, benzene, and pentamidine, forming a column vector of length 5, denoted as b. For example, in a certain measurement, the initial concentration estimates are formaldehyde 12.5, ethanol 8.3, ethyl acetate 0.2, benzene 0.1, and pentamidine 0.4, then b is equal to the column vector with elements of 12.5, 8.3, 0.2, 0.1, and 0.4. The inter-gas competition influence matrix is ​​a 5x5 square matrix, denoted as A. The element in the i-th row and j-th column of the matrix represents the competition inhibition coefficient of the j-th gas on the i-th gas. The diagonal elements are all 1, and the off-diagonal elements are values ​​between 0 and 1.

[0059] The concentration inversion module constructs a linear equation system Ax equal to b, where x is a column vector of unknown concentrations, representing the actual concentration values ​​of the five gases to be determined. The physical meaning of this equation system is that, after considering inter-gas competition suppression, the initial concentration estimate measured by the sensor array is equal to the product of the competition matrix and the actual concentration.

[0060] The concentration inversion module uses the Gauss-Seidel iterative method to solve the linear equation system. First, matrix A is decomposed into the sum of three matrices: A equals D plus L plus U. D is a diagonal matrix, consisting of the diagonal elements of A; since all diagonal elements are 1, D is the identity matrix. L is a strictly lower triangular matrix, containing all elements of A below the diagonal, i.e., the elements in the 2nd row, 1st column, 3rd row, 1st column, 3rd row, 2nd column up to the 5th row, 4th column. U is a strictly upper triangular matrix, containing all elements of A above the diagonal, i.e., the elements in the 1st row, 2nd column, 1st row, 3rd column up to the 4th row, 5th column.

[0061] The concentration inversion module sets the initial iteration vector x. 0 This is a column vector of 5 elements, all of which are 0. Then, the calculation is performed using the Gauss-Seidel iterative formula, where the iterative formula is xᵢ. ( ᵏ +1) It equals bᵢ inside the parentheses minus Aᵢⱼ multiplied by xⱼ from j=1 to i-1. ( ᵏ +1) The sum minus Aᵢⱼ multiplied by xⱼ from j equal to i plus 1 to n ( ᵏ ) The sum of the parentheses is divided by Aᵢᵢ. Since Aᵢᵢ is always 1, the formula simplifies to xᵢ. ( ᵏ +1) It equals bᵢ minus Aᵢⱼ multiplied by xⱼ from j=1 to i-1. ( ᵏ +1) The sum minus Aᵢⱼ multiplied by xⱼ from j equal to i plus 1 to n ( ᵏ )The concentration inversion module calculates the new value of each unknown concentration in the order of i from 1 to 5, using the first i minus 1 components updated in the current iteration step and the last n minus i components from the previous iteration step.

[0062] After each iteration of calculations from i equal to 1 to 5, the concentration inversion module applies non-negativity constraints and concentration upper limit constraints to the x vector obtained in the current iteration. The non-negativity constraint requires that every element in the x vector must be greater than or equal to 0; if an element's calculated value is negative, it is forcibly set to 0. The concentration upper limit constraint is set according to the gas type: 10 ppm for formaldehyde and benzene, and 100 ppm for ethanol, ethyl acetate, and pentanone. If an element's calculated value exceeds the corresponding upper limit, it is forcibly set to the upper limit value. For example, if the calculated value of formaldehyde in a certain iteration is 13.2 ppm, exceeding 10 ppm, it is changed to 10 ppm.

[0063] The concentration inversion module calculates the Euclidean distance between two adjacent iterations after each iteration. The Euclidean distance is calculated by subtracting the square of the previous iteration value from the current iteration value for each component, summing the squares of all components, and then taking the square root. The preset convergence threshold is set to 0.001. When the Euclidean distance is less than 0.001, the concentration inversion module stops iterating and uses the x-vector obtained from the current iteration as the final solution. The five elements in the vector correspond to the actual concentration values ​​of formaldehyde, ethanol, ethyl acetate, benzene, and pentanone in the gas mixture being tested.

[0064] If the convergence condition is not met after 100 iterations, the concentration inversion module will also stop iterating, output the result of the last iteration as the actual concentration value, and mark the measurement as an abnormal convergence.

[0065] The concentration inversion module stores all data obtained within the same measurement period into a complete historical measurement record. This data includes the characteristic parameter set of 25 sensor elements (200 characteristic values), ambient temperature, ambient relative humidity, initial concentration estimate vector, inter-gas competition matrix, intermediate variable sequence for each iteration, and the final calculated actual concentration value. The iterative intermediate variable sequence records five values ​​of the x-vector after each iteration, used for subsequent analysis of the iteration convergence process or anomaly detection. Historical measurement records are stored in local storage or a cloud database in chronological order of measurement, with each record containing a complete measurement timestamp and all the aforementioned data fields. The storage output module outputs the actual concentration values ​​in the order of formaldehyde, ethanol, ethyl acetate, benzene, and pentanone to a display device for display. The display interface simultaneously shows the concentration values ​​of the five gases, all in ppm.

[0066] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A multi-sensor array-based identification and concentration quantification system, characterized in that, include: The curve acquisition module is used to continuously acquire the resistance value of each sensor element at fixed time intervals during the process of the gas sensor array being exposed to the gas mixture to be tested, and form the resistance response curve of each sensor element over time. The feature extraction module is used to extract segmented features from the resistance response curve of each sensor element to obtain a set of feature parameters that reflect the dynamic process of gas adsorption and desorption. The vector fusion module is used to simultaneously collect the ambient temperature and relative humidity values ​​of the environment where the gas sensor array is located, and combine the feature parameter set with the ambient temperature and relative humidity values ​​into a multi-dimensional feature vector. The decoupling inference module is used to input multi-dimensional feature vectors into a pre-trained adsorption competition decoupling neural network, which outputs an initial concentration estimate for each gas to be detected and a gas competition influence matrix. The concentration inversion module is used to obtain the actual concentration value of each gas to be detected in the gas mixture by solving a system of linear equations based on the initial concentration estimate of each gas to be detected and the inter-gas competition influence matrix. The storage output module is used to associate and store the calculated actual concentration value of each gas to be detected with the feature parameter set, and then output it to the display device after arranging it in the preset gas type order.

2. The multi-identification and concentration quantitative detection system based on sensor array according to claim 1, characterized in that, The specific process of segmenting and extracting features from the resistance response curve of each sensor element in the feature extraction module is as follows: Starting from the moment when the gas sensor array is exposed to the gas mixture to be tested, the first and second derivatives of the resistance value of each sensor element with respect to time are continuously calculated. The moment when the first derivative continuously decreases from a positive value to close to zero and the second derivative turns from negative to positive is marked as the end of the adsorption rising phase. The period when the first derivative fluctuates near zero and the absolute value of the second derivative is less than a preset fluctuation threshold is marked as the equilibrium and stability phase. Starting from the moment when the gas mixture to be tested is stopped being introduced into the gas sensor array, the moment when the first derivative is negative and gradually rises back to close to zero and the second derivative turns from positive to negative is marked as the end of the desorption falling phase. The adsorption rate integral, adsorption time constant, and adsorption capacity characteristic value are extracted from the adsorption rise phase; the equilibrium resistance value and equilibrium fluctuation variance value are extracted from the equilibrium stability phase; and the desorption rate integral, desorption time constant, and desorption residual ratio characteristic value are extracted from the desorption fall phase. The above eight characteristic values ​​are used as the characteristic parameter set of the sensor element.

3. The multi-identification and concentration quantitative detection system based on sensor array according to claim 2, characterized in that, The specific extraction process for the adsorption rate integral value, adsorption time constant, and adsorption capacity characteristic value is as follows: During the adsorption ascent phase, the first derivative of the resistance value with respect to time is integrated, and the result is used as the adsorption rate integral value. The resistance response curve during the adsorption ascent phase is fitted using a first-order exponential ascent model with nonlinear least squares, and the fitted time constant parameter is used as the adsorption time constant. The difference between the resistance value at the end of the adsorption ascent phase and the resistance value at the beginning of the adsorption ascent phase is calculated, and the difference is divided by the difference between the average resistance value during the equilibrium and stable phase and the resistance value at the beginning of the adsorption ascent phase. The result of the division operation is used as the adsorption capacity characteristic value.

4. The multi-identification and concentration quantitative detection system based on sensor array according to claim 2, characterized in that, The specific extraction process for the desorption rate integral value, desorption time constant, and desorption residue ratio characteristic value is as follows: During the desorption descent phase, the absolute value of the first derivative of the resistance value with respect to time is integrated, and the result is used as the desorption rate integral value. The resistance response curve during the desorption descent phase is fitted using a first-order exponential decay model with nonlinear least squares, and the time constant parameter obtained from the fitting is used as the desorption time constant. Calculate the first difference between the resistance value at the end of the desorption descent phase and the resistance value at the beginning of the desorption descent phase, calculate the second difference between the resistance value at the beginning of the desorption descent phase and the average resistance value during the equilibrium and stable phase, obtain the ratio of the first difference to the second difference, and subtract the ratio from 1 to obtain the characteristic value of the desorption residual ratio.

5. The multi-identification and concentration quantitative detection system based on a sensor array according to claim 2, characterized in that, In the vector fusion module, the specific process of combining the feature parameter set with the ambient temperature value and the ambient relative humidity value into a multidimensional feature vector is as follows: The sensor elements are sorted in descending order according to the peak response reached by each sensor element during the adsorption and ascent phase. The eight characteristic parameters of each sensor element are arranged in the sorted order to form a characteristic parameter sequence. The ambient temperature value and the ambient relative humidity value are inserted into the middle position of the characteristic parameter sequence to form a multidimensional feature vector with a dimension equal to the product of the number of sensor elements and 8 plus 2. Before inputting the multidimensional feature vector into the adsorption competition decoupling neural network, the values ​​of each dimension are dynamically normalized based on the reference gas response. The dynamic normalization process uses the average resistance value of all sensor elements in the equilibrium and stable phase during the current measurement cycle as the reference value. The values ​​of each dimension are divided by the corresponding reference value and mapped to a closed interval of 0 to 2.

6. The multi-identification and concentration quantitative detection system based on sensor array according to claim 1, characterized in that, The specific structure of the adsorption competition decoupling neural network in the decoupling inference module is as follows: The adsorption competition decoupling neural network consists of an input layer, three competitive feature extraction layers, two competitive interaction layers, and an output layer. The number of nodes in the input layer is set to the dimension of the multidimensional feature vector. The three competitive feature extraction layers sequentially use a two-dimensional convolutional structure with progressively larger kernel size to perform spatial transformation on the input features. Each competitive feature extraction layer is followed by a batch normalization layer and a leakage correction linear unit activation function. The two competitive interaction layers adopt a graph attention network structure, using the features output by the competitive feature extraction layer as graph node features and the prior competitive relationship between the gas types to be detected as the initial value of the graph edge weights. The graph node features and edge weights are updated through a multi-head attention mechanism.

7. The multi-identification and concentration quantitative detection system based on a sensor array according to claim 6, characterized in that, The specific training process of the adsorption competition decoupling neural network is as follows: Samples of single and multiple gases with different concentration combinations were prepared. For each sample, the curve acquisition module and the vector fusion module were repeatedly executed to obtain multi-dimensional feature vectors as training input samples. A two-stage training strategy was adopted. In the first stage, single gas samples were used for pre-training. In the initial concentration estimate expected by the output layer, the gas was set to a known concentration value, other gases were set to zero, the gas competition influence matrix was set to an identity matrix, and the loss function was a weighted sum of mean square error and relative error of predicted concentration. The second stage uses a multi-gas mixed sample for fine-tuning. The actual concentration values ​​of each gas in the mixed sample are determined by gas chromatography as a benchmark. The expected initial concentration estimate of the output layer is set as the known concentration value of each gas. The matrix is ​​solved by iterative optimization to minimize the error between the solution result of the linear equation system in the concentration inversion module and the value determined by gas chromatography. The matrix obtained by the solution is used as the expected inter-gas competition influence matrix. At the same time, matrix symmetry and diagonal dominance constraints are introduced as regularization terms into the loss function.

8. The multi-identification and concentration quantitative detection system based on a sensor array according to claim 7, characterized in that, The output layer of the adsorption competition decoupling neural network is specifically configured as follows: The number of output layer nodes is set to the number of gas types to be detected plus the square of the number of gas types to be detected. The first few nodes of the output layer output the initial concentration estimate of each gas. The remaining nodes of the output layer output each element in the inter-gas competition influence matrix in the order of row first and column second. The inter-gas competition influence matrix is ​​a square matrix with the number of rows and columns equal to the number of gas types to be detected. The diagonal elements in the matrix are set to a fixed value of 1, and the off-diagonal elements are normalized to the interval of 0 to 1 by a flexible maximum value function and then used as the mutual inhibition coefficient between the corresponding two gases.

9. The multi-identification and concentration quantitative detection system based on sensor array according to claim 1, characterized in that, In the concentration inversion module, the specific process of obtaining the actual concentration value of each gas to be detected in the gas mixture by solving a system of linear equations is as follows: Let the number of gases to be detected be n. Construct a column vector of length n from the initial concentration estimates. Treat the inter-gas competition influence matrix as an n-order square matrix. Construct a system of linear equations in which the product of the square matrix and the unknown concentration column vector equals the initial concentration estimate column vector. Decompose the inter-gas competition influence matrix into the sum of a diagonal matrix, a strictly upper triangular matrix, and a strictly lower triangular matrix. Calculate the matrix using the Gauss-Seidel iterative format. After each iteration, apply non-negativity constraints and upper concentration constraints to the unknown concentration column vector. Stop when the Euclidean distance between two adjacent iterations is less than the convergence threshold. Use the converged column vector elements as the corresponding actual gas concentration values. Store the characteristic parameter set, ambient temperature value, ambient relative humidity value, initial concentration estimate, inter-gas competition influence matrix, iterative intermediate variable sequence, and actual concentration value obtained within the same measurement period as a single historical measurement record.

Images