Signal acquisition method and system of MEMS gas sensor array

Abstract

The invention relates to the technical field of signal acquisition, and discloses a signal acquisition method and system for an MEMS gas sensor array, and the method comprises the steps: carrying out the analog-to-digital conversion and temperature and humidity compensation of resistance value signals synchronously acquired by a plurality of sensing units in the MEMS gas sensor array, and obtaining a compensated response matrix; performing fast Fourier transform and dynamic feature extraction to obtain a frequency domain feature matrix of each sensing unit; performing sensing unit crosstalk coefficient calculation and response component decoupling to obtain a decoupling feature vector; carrying out Hilbert transform and adsorption-desorption characteristic extraction to obtain a gas sensitive characteristic sequence; inputting a self-adaptive power spectrum analysis model to carry out cross sensitivity identification processing to obtain a target gas response component; according to the method, the accuracy of target gas detection and concentration quantitative analysis is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal acquisition, and in particular to a signal acquisition method and system for a MEMS gas sensor array. Background Art

[0002] MEMS gas sensor arrays are widely used due to their high integration, low cost and high portability. However, due to the high integration of multiple sensor units on the same chip, there is serious crosstalk between the sensor units, and changes in ambient temperature and humidity can cause sensor baseline drift, affecting detection accuracy.

[0003] There are three main problems with the existing MEMS gas sensor array signal acquisition methods: traditional temperature and humidity compensation methods often use simple linear models, which cannot effectively deal with the nonlinear temperature and humidity response characteristics of the sensor; the crosstalk effect and cross-sensitivity between sensing units make it difficult to extract the target gas characteristics, reducing the accuracy of gas identification; conventional signal processing methods fail to fully consider the adsorption-desorption kinetics between gas molecules and sensitive materials, resulting in inaccurate quantitative analysis results of gas concentration. Summary of the invention

[0004] The present invention provides a signal acquisition method and system for a MEMS gas sensor array, which improves the accuracy of target gas detection and concentration quantitative analysis.

[0005] In a first aspect, the present invention provides a signal acquisition method for a MEMS gas sensor array, the signal acquisition method for a MEMS gas sensor array comprising: The resistance value signals synchronously collected by multiple sensing units in the MEMS gas sensor array are converted into analog-to-digital and compensated for temperature and humidity to obtain a compensated response matrix; Performing fast Fourier transform and dynamic feature extraction on the compensated response matrix to obtain a frequency domain feature matrix of each sensor unit; Calculate the crosstalk coefficient of the sensor unit and decouple the response component according to the frequency domain characteristic matrix to obtain a decoupling characteristic vector; Performing Hilbert transformation and adsorption-desorption characteristic extraction on the decoupling feature vector to obtain a gas-sensitive feature sequence; Inputting the gas-sensitive characteristic sequence into an adaptive power spectrum analysis model for cross-sensitivity identification processing to obtain a target gas response component; Autoregressive modeling and gas concentration quantitative mapping are performed on the target gas response component to obtain target gas concentration parameters.

[0006] In a second aspect, the present invention provides a signal acquisition device for a MEMS gas sensor array, the signal acquisition device for a MEMS gas sensor array comprising: A compensation module is used to perform analog-to-digital conversion and temperature and humidity compensation on the resistance value signals synchronously collected by multiple sensing units in the MEMS gas sensor array to obtain a compensated response matrix; An extraction module, used for performing fast Fourier transform and dynamic feature extraction on the compensated response matrix to obtain a frequency domain feature matrix of each sensor unit; A decoupling module, used to calculate the crosstalk coefficient of the sensor unit and decouple the response component according to the frequency domain characteristic matrix to obtain a decoupling characteristic vector; A transformation module, used for performing Hilbert transformation and adsorption-desorption characteristic extraction on the decoupling feature vector to obtain a gas-sensitive feature sequence; A processing module, used for inputting the gas-sensitive feature sequence into an adaptive power spectrum analysis model for cross-sensitivity identification processing to obtain a target gas response component; The mapping module is used to perform autoregressive modeling and gas concentration quantitative mapping on the target gas response component to obtain the target gas concentration parameter.

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

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the signal acquisition method of the MEMS gas sensor array described above.

[0009] In the technical solution provided by the present invention, by adopting high-precision analog-to-digital conversion and nonlinear temperature and humidity compensation models, the baseline drift caused by environmental factors is effectively eliminated, and high-fidelity acquisition of sensor unit response signals is achieved; based on the transfer function model and orthogonalization processing, the crosstalk relationship between the sensor units is accurately identified and separated, and the independence of the response of each sensor unit is ensured through independent component analysis; the gas response characteristics are comprehensively extracted by combining frequency domain analysis and Hilbert transform, the adsorption-desorption kinetics are accurately captured, and an accurate mapping relationship between gas-sensitive response and concentration is established; the fast algorithm and segmented processing strategy are adopted to reduce the computational complexity and improve the data processing efficiency, and the adaptive algorithm ensures the real-time nature of the processing results; the selectivity of target gas detection is improved through cross-sensitivity identification, the autoregressive model improves the accuracy of quantitative concentration analysis, and the multi-stage filtering optimization ensures the stability of the detection results, so that the present invention can meet the high-precision requirements of gas detection in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0011] Figure 1 A schematic diagram of the steps of a signal acquisition method of a MEMS gas sensor array in an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a signal acquisition device of a MEMS gas sensor array in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

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

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the signal acquisition method of the MEMS gas sensor array in the embodiment of the present invention includes: Step S1, performing analog-to-digital conversion and temperature and humidity compensation on the resistance value signals synchronously collected by multiple sensing units in the MEMS gas sensor array to obtain a compensated response matrix; It is understandable that the execution subject of the present invention may be a signal acquisition device of a MEMS gas sensor array, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0014] Specifically, multiple sensing units in the MEMS gas sensor array are sampled and triggered in parallel, and data is collected at a sampling frequency of 1000Hz to obtain the original resistance value sequence of each sensing unit. The original resistance value sequence is input into a 16-bit high-precision analog-to-digital converter (ADC) for digital-to-analog conversion to convert the analog signal into a digital signal. The 16-bit high-precision analog-to-digital converter can provide a high resolution, so that weak signal changes can be accurately captured. The digital signal is filtered by a Butterworth low-pass filter with a cutoff frequency of 100Hz to remove high-frequency noise interference in the collected signal. The Butterworth filter has a good flat frequency response characteristic, which can eliminate high-frequency noise while retaining the main components of the original signal to obtain a relatively pure digital filter signal. The sampling point data of each sensing unit in the digital filter signal are arranged in chronological order, and an M×N original data matrix is ​​constructed, where M represents the number of sensing units and N represents the number of sampling points. The real-time data collected by the temperature and humidity sensor is subjected to a polynomial fitting operation with the calibration curve to obtain the temperature and humidity correction parameters. Through the correction parameters, the corresponding compensation model is established to calculate the temperature and humidity response coefficient of each sensor unit and obtain the compensation coefficient matrix. The matrix can correct the response of each sensor unit under different temperature and humidity environments to eliminate the errors caused by environmental factors. The original data matrix is ​​processed by zero mean and variance normalization. By adjusting the data mean to zero and the variance to 1, the difference between data of different dimensions is reduced and the comparability of the data is enhanced. Subsequently, the data is preprocessed by singular value decomposition to extract the main components in the data and remove redundant noise to obtain a standardized data matrix. The standardized data matrix and the compensation coefficient matrix are Hadamard product operated to eliminate the influence of temperature and humidity drift. Hadamard product is a matrix multiplication operation method. Through this operation, the data of each sensor unit is corrected point by point to obtain the initial compensation result. The initial compensation result is nonlinearly corrected and adaptive weight iteratively optimized according to the sensitivity curve of each sensor unit. Nonlinear correction is used to process the nonlinear response characteristics of the sensor unit so that the compensation result is more in line with the actual sensor characteristics. The adaptive weight iterative optimization minimizes the error and improves the compensation accuracy by continuously adjusting the weights, and finally obtains the compensated response matrix.

[0015] Step S2, performing fast Fourier transform and dynamic feature extraction on the compensated response matrix to obtain a frequency domain feature matrix of each sensor unit; Specifically, each sensor unit signal of the compensated response matrix is ​​segmented into 1024 points in turn, and the longer acquisition signal is divided into multiple equal-length segments according to a fixed length to facilitate subsequent frequency domain analysis. The Hanning window function product operation is performed on each signal segment, and weighting is applied to the edge of the signal to reduce the spectrum leakage effect caused by signal segmentation. The Hanning window function effectively reduces the impact of signal discontinuity on spectrum analysis, making the spectrum result smoother and better reflecting the true characteristics of the original signal. After weighted processing by the Hanning window, the preprocessed signal segment is obtained. The preprocessed signal segment is subjected to a fast Fourier transform to convert the signal from the time domain to the frequency domain, and the amplitude and phase information of the signal at different frequencies are extracted. In the result of the Fourier transform, the complex spectrum of the first N / 2 frequency points is extracted to capture the main frequency components, thereby reducing the amount of calculation while retaining key information to obtain the initial spectrum data. The amplitude-frequency characteristic analysis of the initial spectrum data is performed to identify the main frequency components contained in the signal, and a characteristic frequency set is constructed. The response characteristics of the signal at different frequencies are clarified through the characteristic frequency set. The phase information corresponding to each frequency point in the characteristic frequency set is phase-unwrapped, and the discontinuity of the phase jump is eliminated by phase unwrapping, making the phase curve smoother. The phase change curve is fitted on the unfolded phase information to obtain the phase eigenvector. The phase eigenvector can reflect the phase change of the signal at different frequencies. On this basis, the phase eigenvector and the characteristic frequency set are reconstructed by complex numbers, and the instantaneous frequency and instantaneous amplitude of the signal are calculated to generate a dynamic parameter matrix. The calculation of instantaneous frequency and amplitude can effectively describe the instantaneous change characteristics of the signal in time and frequency. The dynamic parameter matrix is ​​sorted according to the size of the frequency component to clarify the importance of each frequency component, and the feature contribution is calculated by weighted superposition to obtain the feature weight sequence. The feature weight sequence represents the weight of each frequency component in the entire signal feature. Through the feature weight sequence, the frequency components that have a greater impact on the sensor response at different frequencies are identified. Matrix multiplication is performed on the feature weight sequence and the dynamic parameter matrix to obtain the feature projection matrix. By weighting the different feature components in the dynamic parameter matrix, the final frequency domain feature can highlight the contribution of the key frequency components. In the feature projection matrix, the feature vector combinations whose projection values ​​are greater than the set threshold are screened out, and these feature vector combinations are orthogonalized to ensure that the final frequency domain features are independent of each other and avoid the correlation between different features affecting the discriminability of the features. The frequency domain feature matrix of each sensor unit is obtained.

[0016] Step S3, calculating the crosstalk coefficient of the sensor unit and decoupling the response component according to the frequency domain characteristic matrix to obtain a decoupling characteristic vector; Specifically, the response signal of each sensor unit in the frequency domain feature matrix is ​​inversely transformed by Fourier transform to restore the signal from the frequency domain to the time domain. According to the spatial distribution of the sensor units, the crosstalk coefficient of the time domain response signal is analyzed to obtain the time domain crosstalk component, which reflects the coupling phenomenon caused by the physical proximity between different sensor units. The cross-correlation function of the time domain crosstalk component is calculated, and the similarity of the signals between the sensor units is evaluated by cross-correlation analysis. In order to determine which sensor unit pairs have significant crosstalk relationships, the correlation coefficient threshold is set to 0.85, and the sensor unit pairs exceeding the threshold are considered to have a strong crosstalk effect. A crosstalk relationship is established for the sensor unit pairs exceeding the threshold, and the paired sensor units constitute a crosstalk unit pair set. The crosstalk unit pair set is modeled by transfer function to quantify the degree of crosstalk between each unit pair. Through transfer function modeling, the crosstalk coefficient between each unit pair is obtained, and the crosstalk coefficient matrix is ​​constructed. The matrix can quantify the crosstalk relationship between different sensor units, specifically reflecting the coupling strength between each pair of sensors. The crosstalk coefficient matrix is ​​input into the singular value decomposition model, and the crosstalk coefficient matrix is ​​decomposed by singular value decomposition. The singular values ​​and eigenvectors greater than 0.1 are selected as the main crosstalk components to construct the main crosstalk feature space. The establishment of the main crosstalk feature space helps to retain the main crosstalk components in the signal, thereby focusing on the significant coupling relationship between the sensor units. On this basis, the main crosstalk feature space is transformed into an orthogonal basis, and the decoupling basis vector is constructed by the Gram-Schmidt orthogonalization method. Gram-Schmidt orthogonalization is a method for constructing an orthogonal basis. Features with strong correlation are orthogonalized to obtain orthogonal decoupling basis vectors to achieve effective decoupling of the main crosstalk components. The orthogonal decoupling basis vectors are projected with the response signals of each sensor unit to obtain the initial decoupling result. The initial decoupling result reflects the projection of the response signals of each sensor unit under the orthogonal basis, which can effectively remove the crosstalk effects between the sensor units. In order to improve the decoupling effect, an independent component analysis is performed on the initial decoupling result to separate the independent sources in the signal, so that the response signals of each sensor unit are more independent. Based on the independent component analysis, the optimal separation matrix is ​​calculated to ensure that the independence between the separated signals reaches the optimal state and obtain independent response components. The independent response components are sorted by variance contribution, and the response component combination with a cumulative contribution rate of 95% is selected to obtain the decoupling eigenvector. The sorting of variance contribution and the selection of cumulative contribution rate are aimed at retaining the main information components in the signal while reducing redundancy and noise, ensuring that the final decoupling eigenvector has high representativeness and accuracy.

[0017] Step S4, performing Hilbert transform and adsorption-desorption characteristic extraction on the decoupling feature vector to obtain a gas-sensitive feature sequence; Specifically, the decoupled eigenvector is subjected to Hilbert transform to obtain an analytical signal sequence. Hilbert transform is a processing method for generating complex signals, and its result is used to extract the instantaneous characteristics of the signal. The analytical signal sequence obtained by Hilbert transform contains the amplitude and phase information of the original signal, so that the temporal variation characteristics of the signal can be characterized. Complex number operations are performed on the analytical signal sequence to obtain instantaneous amplitude and instantaneous phase sequences. These two sequences represent the intensity and phase changes of the signal at a specific time point, respectively. The instantaneous amplitude reflects the envelope of the signal, while the instantaneous phase represents the phase change of the signal in time. The instantaneous amplitude and instantaneous phase sequences are subjected to sliding window processing to better describe the time-varying characteristics of the signal. The sliding window processing obtains a time-varying envelope matrix by selecting continuous time periods on the time axis for analysis. The time-varying envelope matrix can reflect the variation trend and local characteristics of the signal in each time period by calculating the local changes of the instantaneous amplitude and instantaneous phase. The adsorption process and desorption process in the time-varying envelope matrix are extracted respectively to construct a kinetic parameter matrix. The adsorption process and desorption process reflect the response behavior of the sensor when it contacts the gas and when the gas leaves, respectively. Extracting these two parts of characteristics is helpful to analyze the interaction between the gas and the sensor. The constructed kinetic parameter matrix is ​​subjected to eigendecomposition to reveal the main change pattern in the matrix and identify different adsorption and desorption characteristics. Through eigendecomposition, the eigenvalues ​​corresponding to each eigencomponent are obtained, and the adsorption rate and desorption rate are sorted according to the size of these eigenvalues ​​to obtain the characteristic coefficient sequence. The characteristic coefficient sequence describes the importance of the adsorption and desorption processes in the kinetic parameters. In order to improve the discrimination of the characteristic coefficient sequence, the S-shaped mapping function is used to perform nonlinear transformation on it. The S-shaped mapping function can enhance the difference between the characteristic coefficients, so that different types of gases can be better distinguished in subsequent analysis. At the same time, in the process of nonlinear transformation, the response curvature before and after the transformation is calculated to quantify the change trend of the characteristic coefficient and obtain the response eigenvector. The response eigenvector is decomposed to obtain the gas-sensitive response component. The gas-sensitive response components are combined and reconstructed according to the energy contribution, retaining the part of the signal that contributes more to the total energy, while reducing the influence of noise and minor components. Through combined reconstruction, a signal containing the main response information is generated, and it is normalized and scaled to a uniform scale range to obtain a gas-sensitive feature sequence.

[0018] Step S5, inputting the gas-sensitive feature sequence into an adaptive power spectrum analysis model for cross-sensitivity identification processing to obtain a target gas response component; Specifically, the gas-sensitive characteristic sequence is segmented and the entire characteristic sequence is divided into several segments with fixed lengths to obtain a segmented characteristic matrix. Segmented processing can improve the local analysis accuracy of the data and help to better capture the changing characteristics of the signal in different time periods. The segmented characteristic matrix is ​​input into the adaptive power spectrum analysis model to analyze the frequency domain characteristics of the signal. In the process of power spectrum estimation, the Welch method is used for calculation. By segmenting, windowing and averaging the data, a smooth power spectrum estimation result is obtained. The Welch method calculates the self-power spectrum and cross-power spectrum with a frequency resolution of 0.1Hz to obtain a multi-channel power spectrum matrix. The multi-channel power spectrum matrix effectively reflects the response characteristics of different sensing units at each frequency and their relationship with each other. The multi-channel power spectrum matrix is ​​input into the adaptive filter bank for processing. The adaptive filter bank consists of a 64-order finite impulse response (FIR) filter and a recursive least squares (RLS) algorithm with a forgetting factor of 0.98. The FIR filter is used to smooth the power spectrum, while the RLS algorithm achieves adaptive spectrum smoothing by continuously updating the coefficients of the filter, reducing the noise components and transient interference in the power spectrum, and obtaining a smooth power spectrum sequence. The power spectra between channels in the smoothed power spectrum sequence are cross-correlated. The degree of mutual influence between each sensor unit is evaluated. Through cross-correlation analysis, the mutual sensitivity between sensor unit pairs is identified, and a cross-sensitivity matrix is ​​constructed. This matrix can quantify the cross-influence between different channels. The cross-sensitivity matrix is ​​input into the feature separation unit, and the eigenvector is extracted by the subspace decomposition method to obtain the target response subspace. Subspace decomposition is an effective method for separating signals into independent sub-signals, which helps to extract components related to the target gas from complex sensor responses. The target response subspace is adaptively weighted to determine the relative importance of each eigenvector in the target response. The adaptive weight update is iteratively calculated and optimized so that the weight can effectively reflect the contribution of each eigenvector in the target gas identification, and a response weight vector is obtained. The weight vector indicates which features in the target response subspace play a key role in the response of the target gas. The response weight vector is projected onto the smoothed power spectrum sequence to extract the response component of the target gas. Through the projection operation, the multi-channel power spectrum information is concentrated on the characteristic response of the target gas to obtain the response component reflecting the characteristics of the target gas.

[0019] Step S6: performing autoregressive modeling and gas concentration quantitative mapping on the target gas response component to obtain the target gas concentration parameter.

[0020] Specifically, based on the target gas response component, the order of the autoregressive model is determined to be 8th order through partial correlation function analysis. Partial correlation function analysis helps to identify the key time correlation in the signal. The 8th order model can capture the dynamic characteristics of the target gas response component and obtain a more accurate autoregressive model parameter sequence. The autoregressive model parameter sequence is input into the Levinson-Durbin recursive algorithm, and the forward and reverse prediction errors are calculated by the algorithm to obtain the autoregressive coefficients. The Levinson-Durbin algorithm is known for its high efficiency and can quickly calculate the coefficients of the autoregressive model under limited computing resources. The obtained autoregressive coefficients are used to construct a characteristic coefficient matrix to describe the characteristic relationship of the response component in the time series. The characteristic coefficient matrix is ​​optimized by Gaussian-Newton iteration, and the model error is minimized by continuously adjusting the parameters to obtain the optimized model parameters. The Gaussian-Newton iteration method is a commonly used algorithm in nonlinear least squares optimization. Through multiple iterations, the model is more accurately fitted to the actual response data. The optimized model parameters are polynomially fitted with the known gas concentration calibration curve to establish a mapping relationship between gas concentration and sensor response. Through fitting, the mapping function matrix is ​​obtained, and the root mean square of the fitting error is calculated to evaluate the accuracy of the fitting. The smaller the root mean square error, the more accurately the fitting model can reflect the response relationship of the sensor to the gas concentration. The mapping function matrix is ​​processed by piecewise linearization. The mapping function is segmented according to the concentration range, and the corresponding slope and intercept are calculated for each segment to obtain the piecewise linear parameters. The piecewise linearization approximates the complex nonlinear mapping relationship into several linear segments, improves the calculation efficiency of the model, and maintains reasonable fitting accuracy. The piecewise linear parameters are input into the concentration mapping model, and nonlinear transformation and feature space mapping are performed to obtain the concentration mapping vector. The concentration mapping vector describes the mapping relationship between the sensor response and the target gas concentration, and can convert the original response characteristics into the corresponding concentration value. The concentration mapping vector is processed by Kalman filtering to obtain the filtered concentration sequence. Kalman filtering is a recursive algorithm for signal processing. It can perform state estimation in a noisy environment based on the observed value of the sensor and the predicted state of the system, effectively reduce the influence of random noise on the concentration estimation, and obtain a smoother concentration change curve. After Kalman filtering, in order to reduce short-term random fluctuations, the filtered concentration sequence is subjected to moving average processing. By taking the average of multiple adjacent data points, the fluctuation in concentration estimation is reduced to obtain the final target gas concentration parameter.

[0021] In the embodiment of the present invention, by adopting high-precision analog-to-digital conversion and nonlinear temperature and humidity compensation models, the baseline drift caused by environmental factors is effectively eliminated, and high-fidelity acquisition of the response signal of the sensor unit is achieved; based on the transfer function model and orthogonalization processing, the crosstalk relationship between the sensor units is accurately identified and separated, and the independence of the response of each sensor unit is ensured through independent component analysis; the frequency domain analysis and Hilbert transform are combined to comprehensively extract the gas response characteristics, accurately capture the adsorption-desorption kinetics, and establish a precise mapping relationship between the gas-sensitive response and the concentration; the fast algorithm and segmented processing strategy are adopted to reduce the computational complexity and improve the data processing efficiency, and the adaptive algorithm ensures the real-time processing results; the selectivity of target gas detection is improved through cross-sensitivity identification, the autoregressive model improves the accuracy of quantitative concentration analysis, and the multi-stage filtering optimization ensures the stability of the detection results, so that the present invention can meet the high-precision requirements of gas detection in complex environments.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Parallel sampling and triggering are performed on multiple sensing units in the MEMS gas sensor array, and data is collected at a sampling frequency of 1000 Hz to obtain a sequence of original resistance values; The original resistance value sequence is input into a 16-bit high-precision analog-to-digital converter for digital-to-analog conversion, and the noise is filtered out through a Butterworth low-pass filter with a cutoff frequency of 100 Hz to obtain a digital filtered signal; The sampling point data of each sensor unit in the digital filter signal are arranged in time order, and an M×N matrix is ​​constructed, where M is the number of sensor units and N is the number of sampling points, to obtain the original data matrix; The real-time data collected by the temperature and humidity sensor is subjected to polynomial fitting operation with the calibration curve to obtain the temperature and humidity correction parameters, and a compensation model is established based on the temperature and humidity correction parameters to calculate the temperature and humidity response coefficient of each sensor unit to obtain the compensation coefficient matrix; The original data matrix was processed with zero mean and variance normalization, and data preprocessing was performed through singular value decomposition to obtain a standardized data matrix; The Hadamard product operation is performed on the standardized data matrix and the compensation coefficient matrix, and the influence of temperature and humidity drift is eliminated through matrix decomposition to obtain the initial compensation result. The initial compensation result is subjected to nonlinear correction and adaptive weight iterative optimization according to the sensitivity curve of each sensor unit to obtain the compensated response matrix.

[0023] Specifically, multiple sensor units in the array are sampled and triggered in parallel, and data is collected at a sampling frequency of 1000 Hz to obtain an original resistance value sequence. In order to ensure the synchronization of multiple sensor units, a synchronous clock signal is used for triggering, and each sensor unit collects resistance values ​​at the same time interval. This process is implemented using a high-speed data acquisition system. For example, the resistance value of each sensor unit is continuously recorded at a sampling frequency of 1000 Hz to obtain an original resistance value sequence. Assume that the collected original resistance value sequence is ,in It is a time series The resistance value under is in ohms (Ω), and the data length is . The original resistance value sequence is input into a 16-bit high-precision analog-to-digital converter (ADC) for digital-to-analog conversion, converting the analog signal into a digital signal. The 16-bit ADC can provide a higher resolution, ensure the accuracy of the signal, and avoid the subsequent processing affected by quantization errors. At the same time, in order to filter out the high-frequency noise in the signal, a Butterworth low-pass filter with a cutoff frequency of 100Hz is used for noise filtering. The Butterworth low-pass filter has a flat frequency response and can effectively remove high-frequency noise components. The filtered digital signal is expressed as The sampling point data in the digital filter signal of each sensor unit are arranged in time sequence and a The matrix of represents the number of sensing units, represents the number of sampling points of each sensor unit. Each row of the matrix represents the sampling sequence of a sensor unit, and each column corresponds to a specific time point. Assume that sensor units, each sensor unit Each time point has a corresponding resistance value, and the constructed original data matrix is ​​expressed as: ; Each element in the matrix Representative The sensor unit is The resistance value at a time point. In order to perform temperature and humidity compensation, the real-time data collected by the temperature and humidity sensor is combined. The temperature and humidity correction parameters are obtained by performing polynomial fitting calculations with the calibration curve. The polynomial fitting of the calibration curve uses the least squares method to solve the best fitting function, as shown below: ; in represents temperature, Represents humidity, and The relationship between temperature, humidity and correction parameters is obtained through polynomial fitting. The temperature and humidity correction parameters are obtained through polynomial fitting. , and establish a compensation model based on the correction parameters to calculate the temperature and humidity response coefficient of each sensor unit. The temperature and humidity response coefficient of each sensor unit is , the compensation coefficient matrix is ​​expressed as: ; The original data matrix is ​​subjected to zero mean and variance normalization. Zero mean is to subtract the mean of each data point to ensure that the mean of the data is zero; variance normalization is to divide each data point by its standard deviation to ensure that the variance of the data is 1. This process is achieved through the following formula: ; in, and They are The mean and standard deviation of the sensor units, is the normalized resistance value. The matrix after zero mean and variance normalization is When performing data preprocessing, singular value decomposition is used to reduce the dimension of the standardized data matrix. Singular value decomposition extracts the main features of the data by decomposing the matrix into the product of three matrices. The specific formula is: ; in, and are the left singular matrix and the right singular matrix respectively, is a singular value matrix. Through singular value decomposition, we can extract important information from the data and ignore unimportant noise to obtain a more concise feature matrix. and compensation coefficient matrix Perform Hadamard product operation. Hadamard product is an element-level matrix multiplication operation, that is, corresponding elements of the matrix are multiplied to obtain a new matrix. This operation eliminates the influence of temperature and humidity drift and obtains the initial compensation result: ; in, represents the Hadamard product operation. The initial compensation result is Represents the response signal of each sensor unit after temperature and humidity compensation. The initial compensation result is subjected to nonlinear correction and adaptive weight iterative optimization. Through nonlinear correction, appropriate adjustments are made according to the sensitivity curve of each sensor unit. Assuming that the sensor unit The sensitivity curve is , the optimization process is completed through an iterative algorithm as follows: ; in, is the learning rate, It is the sensing unit Through multiple iterations, the response matrix gradually converges to the optimal solution, and finally the compensated response matrix is ​​obtained. .

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Each sensor unit signal of the compensated response matrix is ​​segmented into 1024 points in turn, and each segment signal is multiplied by the Hanning window function to obtain a preprocessed signal segment; Perform fast Fourier transform on the preprocessed signal segment, extract the complex spectrum of the first N / 2 frequency points, obtain the initial spectrum data, and perform amplitude-frequency characteristic analysis on the initial spectrum data to obtain the characteristic frequency set; The phase information corresponding to the characteristic frequency set is expanded and the phase change curve is fitted to obtain the phase characteristic vector, and the phase characteristic vector and the characteristic frequency set are reconstructed by complex numbers, and the instantaneous frequency and instantaneous amplitude are calculated to obtain the dynamic parameter matrix; The dynamic parameter matrix is ​​sorted according to the size of the frequency component, and the feature contribution is calculated by weighted superposition to obtain the feature weight sequence; The feature weight sequence and the dynamic parameter matrix are multiplied to obtain the feature projection matrix, and the feature vector combinations with projection values ​​greater than the set threshold in the feature projection matrix are orthogonalized to obtain the frequency domain feature matrix of each sensor unit.

[0025] Specifically, the signal of each sensor unit is processed in 1024-point segments, and each segment of the signal is multiplied by the Hanning window function to obtain a preprocessed signal segment. The signal is divided into time windows of fixed length to facilitate the analysis of the characteristics of the signal in the frequency domain. Assume that the response matrix after compensation is ,in yes The signals of the sensor units, each signal contains sampling points. In the signal of each sensor unit, the signal length is The data is divided into multiple segments of size 1024. Each segment of data is in the form of: ; in Indicates the index of the signal segment, It is The first sensor unit For each segment of signal , apply the Hanning window function for windowing. Hanning window function The form is: ; The function of this window function is to reduce the spectrum leakage of the signal and improve the spectrum resolution of Fourier transform. The signal after windowing by the Hanning window function is: ; The windowing results of each signal segment are obtained. Fast Fourier transform is performed on each windowed signal segment to convert the time domain signal to the frequency domain. The formula of fast Fourier transform is: ; in, is a time domain signal, is the corresponding frequency domain signal, is the frequency, is the number of data points. After fast Fourier transform, the complex spectrum data is obtained ,in Indicates Sensor unit Frequency domain representation of segment data. In order to extract the main frequency components of the signal, only the frequency points are retained. ,in is the Nyquist frequency, is the sampling frequency. The extracted spectrum data is expressed as: ; On this basis, the amplitude-frequency characteristics of the initial spectrum data are analyzed to calculate the amplitude of each frequency point. and Phase The purpose of amplitude-frequency characteristic analysis is to identify the energy distribution of frequency components in order to extract the main characteristic frequencies. Through amplitude-frequency characteristic analysis, the characteristic frequency set is obtained. , these frequency points are the main spectral components of the signal. The phase information corresponding to the characteristic frequency set is expanded and the phase change curve is fitted. Phase information It represents the phase angle of the spectrum data, reflecting the phase change of the signal in the frequency domain. In order to extract effective phase information from the frequency points, a curve fitting method, such as polynomial fitting or spline interpolation, is used to fit the discrete phase points into a continuous phase curve. Assume that the fitted phase curve is , through the fitting curve, we get the phase characteristic vector : ; The phase eigenvector is reconstructed by complex reconstruction. The instantaneous frequency of the signal is reconstructed based on the characteristic frequency set and the phase eigenvector obtained by fitting. and the instantaneous amplitude , and its calculation formula is: ; Among them, the instantaneous frequency Represents the signal at time Frequency distribution at the moment, instantaneous amplitude Represents the signal at time By calculating these parameters, we can get the dynamic parameter matrix , which contains the time-varying information of the instantaneous frequency and instantaneous amplitude of the signal. The dynamic parameter matrix is ​​sorted and weighted superposition is performed according to the size of the frequency components. Through weighted superposition, the characteristic contribution of each frequency component is calculated : ; Feature weight sequence calculated by weighted superposition Optimize the frequency domain feature analysis of the signal. Perform matrix multiplication on the feature weight sequence and the dynamic parameter matrix to obtain the feature projection matrix : ; For each eigenvector in the projection matrix, the most important frequency domain features are screened out by setting a threshold. Assume that the threshold is set to , when the eigenprojection value is greater than When , it indicates that the eigenvector is an important frequency domain feature. The important eigenvectors are combined and orthogonalized, and the orthogonal decoupling basis vectors are obtained by the Gram-Schmidt orthogonalization method, and finally the frequency domain feature matrix of each sensor unit is obtained.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Perform inverse Fourier transform on the response signal of each sensor unit of the frequency domain feature matrix, and perform crosstalk coefficient analysis according to the spatial distribution of the sensor units to obtain the time domain crosstalk component; The time domain crosstalk component is cross-correlated and the correlation coefficient threshold is set to 0.85, and a crosstalk relationship is established for the sensor unit pairs exceeding the threshold, thereby obtaining a crosstalk unit pair set. The transfer function modeling is performed on the set of crosstalk unit pairs, and the crosstalk coefficients between each unit pair are calculated to obtain a crosstalk coefficient matrix; The crosstalk coefficient matrix is ​​input into the singular value decomposition model, singular values ​​and eigenvectors greater than 0.1 are selected as the main crosstalk components to obtain the main crosstalk feature space, and the main crosstalk feature space is transformed into an orthogonal basis, and the decoupling basis vectors are constructed by the Gram-Schmidt orthogonalization method to obtain the orthogonal decoupling basis vectors; Projecting the orthogonal decoupling basis vectors with the response signals of each sensor unit to obtain an initial decoupling result, and performing independent component analysis and optimal separation matrix calculation on the initial decoupling result to obtain independent response components; The independent response components are sorted by variance contribution, and the response component combination with a cumulative contribution rate of 95% is selected to obtain the decoupled eigenvector.

[0027] Specifically, the response signal of each sensor unit in the frequency domain feature matrix is ​​subjected to inverse Fourier transform to convert it from the frequency domain back to the time domain. The frequency domain feature matrix is ​​set to , where each Indicates The frequency domain characteristic signal of each sensor unit. The calculation formula of the inverse Fourier transform is: ; in, Indicates The time domain response signal of each sensor unit, is the corresponding frequency domain signal, is the number of sampling points, is the frequency point, is time. Through inverse Fourier transform, the frequency domain characteristic signal of each sensor unit is converted into a time domain response signal. The crosstalk coefficient is analyzed based on the time domain response signal. Since multiple sensor units in the sensor array will produce crosstalk, especially between sensor units close to each other in space, crosstalk will cause mutual interference of signals. These signals are analyzed. For each pair of sensor units, the cross-correlation function between their time domain response signals is calculated. Assume that and The time domain signals of each sensor unit are and , their cross-correlation function is defined as: ; in, and They are and By calculating the cross-correlation function of each pair of sensor units, the correlation between them is obtained. If the correlation coefficient is greater than the set threshold , then it is considered that there is a crosstalk relationship between the two sensor units. All sensor unit pairs exceeding this threshold are included in the crosstalk unit pair set , which contains all the sensor unit pairs that have a large mutual influence. For the crosstalk unit pair set , a transfer function model is established to describe the crosstalk relationship between each pair of sensor units. Assuming that the sensor unit and The transfer function between , expressed as: ; To calculate the transfer function, the least squares method or maximum likelihood estimation is used to fit the data to obtain the crosstalk coefficient matrix between the sensing units. , where each element Indicates Sensor units to The crosstalk coefficient of each sensor unit is calculated. The degree of crosstalk between different sensor units is quantified by calculating the crosstalk coefficient. Input the singular value decomposition model to perform matrix decomposition. The formula for singular value decomposition is: ; in, and is the unit orthogonal matrix, is a diagonal matrix containing the matrix The singular values ​​of . Through singular value decomposition, the principal components of the crosstalk coefficient matrix are obtained. If some singular values ​​are small (less than 0.1), the eigenvectors they represent contribute little to the crosstalk analysis and can be ignored. In this process, eigenvectors greater than 0.1 are selected as the main crosstalk components to form the main crosstalk feature space . For the main crosstalk feature space Perform orthogonal basis transformation. Use the Gram-Schmidt orthogonalization method to construct a decoupling basis vector to eliminate the correlation between the sensor units, making the response signal of each sensor unit more independent. The basic steps of Gram-Schmidt orthogonalization are: select a basis vector , and then calculate the orthogonal basis vectors by the following formula : ; in, yes exist The projection on , the projection calculation formula is: ; In this way, the orthogonal decoupled basis vectors are constructed. , these basis vectors can effectively remove the crosstalk between sensor units. Perform projection calculation with the response signals of each sensor unit to obtain the initial decoupling result : ; Perform independent component analysis and optimal separation matrix calculation on the initial decoupling results. The goal of independent component analysis is to extract independent source signals from multiple mixed signals, which helps to improve the decoupling effect. The independent component analysis method uses the principle of maximizing non-Gaussianity to separate signals and obtains the optimal separation matrix through optimization algorithms (such as gradient descent method). , and apply it to the decoupled result to obtain the independent response components : ; In order to ensure the stability and effectiveness of the decoupling effect, the variance contribution of the independent response components is sorted, and the response component combination with a cumulative contribution rate of 95% is selected. The most representative decoupling features are screened out, and the variance contribution is calculated using the following formula: ; Select the response components with larger variance contribution, reconstruct and normalize them, and obtain the final decoupled feature vector , these vectors represent the independent response characteristics of each sensor unit after crosstalk suppression.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Performing Hilbert transformation on the decoupled eigenvector to obtain an analytical signal sequence, and performing complex operations on the analytical signal sequence to obtain an instantaneous amplitude and instantaneous phase sequence; The instantaneous amplitude and instantaneous phase sequences are processed by sliding windows to obtain a time-varying envelope matrix, and the adsorption process and desorption process in the time-varying envelope matrix are extracted respectively to construct a kinetic parameter matrix. The kinetic parameter matrix is ​​characteristically decomposed, and the eigenvalues ​​of the adsorption rate and the desorption rate are sorted to obtain a characteristic coefficient sequence, which is then nonlinearly transformed using an S-shaped mapping function. At the same time, the response curvature before and after the transformation is calculated to obtain a response characteristic vector. The response characteristic vector is decomposed to obtain gas-sensitive response components, and the gas-sensitive response components are combined, reconstructed and normalized according to energy contribution to obtain a gas-sensitive characteristic sequence.

[0029] Specifically, assume that the decoupled eigenvector is expressed as ,in Indicates The time domain response signal of each sensor unit. Apply the Hilbert transform to obtain its analytical signal , the analytical signal consists of the real part and the imaginary part The imaginary part is the Hilbert transform of the signal. The formula of Hilbert transform is: ; in, yes The Hilbert transform of is an imaginary unit. Through Hilbert transform, we get It is a complex signal, which contains the amplitude information and phase information of the signal. The instantaneous amplitude and instantaneous phase are extracted from the analytical signal through complex number operation. and instantaneous phase The calculation formula is: ; ; in, Indicates The sensor unit at the time The instantaneous amplitude of is the corresponding instantaneous phase. Through these operations, the instantaneous amplitude and instantaneous phase sequence of each sensor unit at different time points are obtained. Sliding window processing is performed on the instantaneous amplitude and instantaneous phase sequence. Assume that the window size is , the instantaneous amplitude and instantaneous phase sequences are locally analyzed by sliding window technology. and the instantaneous phase sequence , defines the time-varying envelope after sliding window processing and time-varying phase for: ; ; These two items represent the mean of the amplitude and phase within the sliding window, respectively. In this way, a time-varying envelope matrix is ​​obtained, in which each column corresponds to the smoothed results of the instantaneous amplitude and phase within a time period. The adsorption process and desorption process are extracted from the time-varying envelope matrix. The adsorption process refers to the reaction process between gas molecules and the sensor surface, while the desorption process is the process of gas molecules being released from the sensor surface. In the time-varying envelope matrix, the adsorption process is manifested as an increase in amplitude with time, while the desorption process is manifested as a decrease in amplitude with time. A threshold is set to separate the adsorption process and the desorption process, and the kinetic parameters of the two processes are extracted respectively. Assume that is the time-varying envelope matrix The instantaneous amplitude of each sensor unit is used to construct the dynamic parameter matrix by the following method: ; In this way, the kinetic parameter matrix Contains the amplitude information of the adsorption and desorption processes. Perform eigendecomposition on the kinetic parameter matrix to extract the eigenvalues ​​of the adsorption rate and desorption rate. By analyzing the time evolution of the adsorption and desorption processes, important dynamic features are extracted. Assumptions is the constructed kinetic parameter matrix, which contains the adsorption and desorption information of all sensing units. Perform eigenvalue decomposition: ; in, is the eigenvalue matrix, is the corresponding eigenvector matrix. By sorting these eigenvalues, the characteristic coefficient sequence is obtained according to the size of the adsorption rate and desorption rate. For example, a larger eigenvalue indicates a higher adsorption rate, while a smaller eigenvalue indicates a higher desorption rate. In order to analyze the change of the characteristic coefficient, the S-shaped mapping function is used to perform a nonlinear transformation on the characteristic coefficient sequence. The S-shaped mapping function is expressed by the following formula: ; This function can compress the characteristic coefficient sequence into the interval [0,1], thereby performing standardization. Response curvature before and after transformation Calculated by the following formula: ; in, is the change of the characteristic coefficient after transformation, is the change of the input value. By calculating the response curvature and judging the change rate of the characteristic coefficient, a more accurate gas-sensitive response characteristic is obtained. The response characteristic vector is decomposed and reconstructed to extract the gas-sensitive response component. The decomposition process is performed through principal component analysis. Or independent component analysis (ICA) is used to extract the main components representing the gas-sensitive response from numerous sensor signals. It is the decoupled feature matrix, which is decomposed by PCA to obtain the main components: ; in, is the eigenvector matrix, is the eigenvalue matrix. The most important gas-sensitive response components are selected according to the energy contribution, and they are combined and normalized to obtain the final gas-sensitive characteristic sequence These characteristic sequences can effectively characterize the gas-sensitive response characteristics of the sensor and be used for the identification and quantitative analysis of gas concentration.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: The gas-sensitive characteristic sequence is segmented to obtain a segmented characteristic matrix, and the segmented characteristic matrix is ​​input into the adaptive power spectrum analysis model for power spectrum estimation. The auto-power spectrum and cross-power spectrum with a frequency resolution of 0.1 Hz are calculated by the Welch method to obtain a multi-channel power spectrum matrix. The multi-channel power spectrum matrix is ​​input into the adaptive filter bank, and the spectrum is smoothed by a 64-order FIR filter and a recursive least squares algorithm with a forgetting factor of 0.98 to obtain a smoothed power spectrum sequence; The power spectra between channels in the smoothed power spectrum sequence are cross-correlated and analyzed to obtain a cross-sensitivity matrix, which is then input into a feature separation unit to extract feature vectors using a subspace decomposition method to obtain a target response subspace. The target response subspace is adaptively weighted to obtain a response weight vector, and the response weight vector is projected onto the smoothed power spectrum sequence to obtain the target gas response component.

[0031] Specifically, the gas-sensitive feature sequence is segmented. Assume that the gas-sensitive feature sequence ,in Indicates The gas-sensitive characteristics of each sensor unit are divided into multiple sub-segments according to a certain time window. The window length is set to , and the step length is , and obtain the segmented feature matrix , where each column represents a segmented gas-sensitive feature sequence. Segments It is expressed as: ; in, is the total number of segments, and represents the number of sensing units. The length of each segment is ,Include The response data of each sensor unit. The segmented feature matrix is ​​input into the adaptive power spectrum analysis model to perform power spectrum estimation. The Welch method is used for power spectrum estimation. The Welch method is a commonly used technique for estimating power spectrum density. It reduces variance by overlapping the signal and calculating the Fourier transform of each segment. The frequency resolution is set to 0.1 Hz, and the Fourier transform is performed on each segment feature matrix to obtain the frequency domain features of each segment. Assume that The spectrum of each segment is , then its autopower spectrum and cross power spectrum They are calculated as: ; ; in, represents the Fourier transform operation, is the frequency, and * represents the complex conjugate operation. After Fourier transforming all the segments, the multi-channel power spectrum matrix is ​​obtained. , contains the power spectrum information of all sensor units at different frequencies. The multi-channel power spectrum matrix Input adaptive filter bank for spectrum smoothing. In order to improve the estimation accuracy of power spectrum, 64-order FIR filter and recursive least squares algorithm (RLS) are used to smooth the power spectrum. RLS algorithm adaptively updates the filter coefficients by minimizing the weighted square error, thereby improving the stability of estimation. Assume that the weight vector of the filter is , then at every moment , the smoothed value of the power spectrum Updated by the following recursive formula: ; in, is the error at the current moment, are the adaptive weights of the filter, is the smoothed power spectrum sequence. The cross-correlation analysis of the smoothed power spectrum sequence is performed to analyze the correlation between different channels and reveal the mutual influence between the sensor units. Assume and They are and The smoothed power spectrum sequence of the sensor units, the cross-correlation function Defined as: ; By calculating the cross-correlation function, the cross-sensitivity matrix is ​​obtained , describes the mutual influence between different sensing units in the frequency domain. The cross-sensitivity matrix Input feature separation unit for feature extraction. Decompose the multi-channel signal into several independent signal components, which represent different physical characteristics. Use subspace decomposition methods, such as singular value decomposition or principal component analysis, to decompose the cross sensitivity matrix and obtain the target response subspace. , which contains the characteristic direction that is most sensitive to the target gas response. Performing feature decomposition, we get: ; in, is the left singular vector matrix, is the singular value matrix, is a right singular vector matrix. Sort and select the most important eigenvector as the target response subspace. Update the target response subspace with adaptive weights. According to the response characteristics of the target gas, assign appropriate weights to each eigenvector. The adaptive weight update is performed by calculating the weight vector To achieve this, the formula is as follows: ; in, is the weight vector updated according to the target response subspace. By replacing the response weight vector And the smoothed power spectrum series Perform projection calculation to obtain the response component of the target gas , which represents the characteristics related to the target gas concentration in the power spectrum domain. Target gas response component Represents the signal component that best reflects the characteristics of the target gas among multiple sensing units.

[0032] Among them, the power spectra between channels in the smooth power spectrum sequence are cross-correlated and analyzed to obtain a cross-sensitivity matrix, which is then input into a feature separation unit. The feature vector is extracted by a subspace decomposition method to obtain a target response subspace. The specific steps include: performing power normalization processing on each sensor channel data in the smooth power spectrum sequence to eliminate the influence of power differences between channels and obtain a normalized power spectrum; pairing the normalized power spectra in pairs according to sensor channels, calculating the correlation between each pair of channels and obtaining a channel correlation matrix; setting a correlation threshold of 0.8 for the channel correlation matrix, marking channel pairs with correlations exceeding the threshold and obtaining a high-correlation channel group; constructing a cross-sensitivity calculation model based on the high-correlation channel group to analyze the response of each channel to the target gas. response intensity to obtain sensitivity parameters; arrange the sensitivity parameters into a matrix form according to the sensor channels, and perform row and column normalization processing to obtain a cross-sensitivity matrix; assign weights to each element in the cross-sensitivity matrix, and the weight value is proportional to the sensitivity to obtain a weighted sensitivity matrix; perform feature decomposition on the weighted sensitivity matrix, extract the eigenvectors corresponding to the main response characteristics, and obtain a set of eigenvectors; perform feature space projection on the eigenvector set, map the high-dimensional features to the low-dimensional space, and obtain a reduced-dimensional feature space; construct a target gas response subspace based on the reduced-dimensional feature space, and set the projection coefficient threshold to 0.75 to obtain an initial response subspace; perform orthogonal processing on the initial response subspace to eliminate the linear correlation between features and obtain a target response subspace.

[0033] In a specific embodiment, the process of executing step S6 may specifically include the following steps: Based on the target gas response component, the order of the autoregressive model is determined to be 8th order through partial correlation function analysis, and the model parameter sequence is obtained; The model parameter sequence is input into the Levinson-Durbin recursive algorithm, the autoregressive coefficient is calculated through forward and backward prediction error operation, the characteristic coefficient matrix is ​​obtained, and the characteristic coefficient matrix is ​​optimized by Gaussian-Newton iteration to obtain the optimized model parameters; The optimized model parameters and the gas concentration calibration curve are subjected to polynomial fitting and the root mean square calculation of the fitting error to obtain a mapping function matrix, and the mapping function matrix is ​​subjected to piecewise linearization processing, and the slope and intercept of each segment are calculated to obtain piecewise linear parameters; The piecewise linear parameters are input into the concentration mapping model for nonlinear transformation and feature space mapping to obtain a concentration mapping vector; The concentration mapping vector is processed by Kalman filtering to obtain a filtered concentration sequence, and the filtered concentration sequence is processed by moving average to obtain the target gas concentration parameter.

[0034] Specifically, the order of the autoregressive model is determined by partial correlation function analysis. The autoregressive (AR) model is a time series modeling method that predicts the current value by the past value of the independent variable. In this task, the response component of the target gas As time series data input for modeling. The partial correlation function is an important tool for determining the order of the AR model. Indicates that before a given Under the condition of data points, The linear relationship between the data point and the current data point. Perform partial correlation analysis to determine the optimal autoregressive order , assuming that the order is determined to be 8th through analysis, the model is constructed as an 8th-order autoregressive model. In the autoregressive model, the target gas response component It is represented by the values ​​of the first 8 moments, namely: ; in, is the coefficient of the AR model, is the white noise error term. In this way, 8 past time points are used to predict the current response value. According to the form of the model, the order of the autoregressive model is determined by the partial correlation function, and the corresponding model parameters are calculated The calculated model parameter sequence is input into the Levinson-Durbin recursive algorithm for parameter optimization. The Levinson-Durbin algorithm is an efficient recursive algorithm for solving the coefficients of the autoregressive model. In this process, the autoregressive coefficients are calculated by forward and backward prediction errors. Assume that a characteristic coefficient matrix is ​​obtained by this algorithm , which contains the estimated values ​​of each model parameter. This matrix is ​​obtained by optimization method and can accurately fit the dynamic characteristics of the target gas. The Gaussian-Newton iterative optimization algorithm is used to optimize the characteristic coefficient matrix. The Gaussian-Newton algorithm is an iterative optimization method suitable for nonlinear least squares problems. During the optimization process, the Jacobian matrix of the characteristic coefficient matrix is ​​calculated and used to update the characteristic coefficients to minimize the prediction error. The optimization process is carried out by the following iterative formula: ; in, It is The characteristic coefficient of the iteration, is the Jacobian matrix, is the residual vector, which represents the difference between the model prediction value and the actual observation value. Through iterative optimization, more accurate model parameters are obtained. The model parameters are polynomially fitted with the gas concentration calibration curve. The calibration curve is obtained through experiments and describes the relationship between the sensor response and the target gas concentration. Assume that the calibration curve is ,in is the gas concentration, is the response component of the target gas. Using the polynomial fitting method, the optimized AR model parameters Estimate gas concentration. Assume that a quadratic polynomial is selected for fitting, and the fitting function is: ; in, are the fitting parameters of the polynomial. These fitting parameters are calculated by the least squares method to minimize the fitting error. The fitting error is measured by the root mean square error (RMSE), and the calculation formula is: ; in, is the number of samples, It is the concentration prediction value obtained by polynomial fitting. In this way, a mapping function matrix is ​​obtained, which represents the nonlinear relationship between the target gas response component and the gas concentration. The mapping function matrix is ​​subjected to piecewise linearization. Since the mapping function obtained by polynomial fitting has a large nonlinearity in the entire concentration range, the mapping function matrix is ​​processed in segments. Assume that the gas concentration range is divided into several segments, and a linear function is used to approximate each segment. For each segment, the slope and intercept of the segment are calculated to obtain the piecewise linear parameters. ,in and They are The slope and intercept of the segment. In this way, the concentration value is more accurately mapped to the target gas response component. The piecewise linear parameters are input into the concentration mapping model for nonlinear transformation and feature space mapping. Nonlinear transformation aims to improve mapping accuracy, and feature space mapping helps the system represent gas concentration in multidimensional space. In this way, the concentration mapping vector is obtained. , which represents an accurate estimate of the gas concentration. The concentration mapping vector is processed by Kalman filtering. Kalman filtering is a recursive filtering algorithm used for state estimation of dynamic systems. Kalman filtering is used to smooth the concentration sequence to remove noise and improve prediction accuracy. The basic formula of Kalman filtering is: ; in, is the estimated concentration at the current moment, is the Kalman gain, is the actual observed value, is the observation matrix. Through Kalman filtering, we get the filtered concentration sequence In the filtered concentration sequence, the concentration sequence is processed by moving average. By performing sliding window averaging on the filtered concentration sequence, short-term fluctuations are effectively removed, the concentration change curve is smoothed, and the final target gas concentration parameter is obtained. Assume that the sliding window size is , the moving average formula is: ; Get the final gas concentration estimate , that is, the target gas concentration parameter.

[0035] Optionally, the piecewise linear parameters are input into the concentration mapping model for nonlinear transformation and feature space mapping to obtain a concentration mapping vector; the concentration mapping vector is subjected to Kalman filtering to obtain a filtered concentration sequence, and the filtered concentration sequence is subjected to moving average processing to obtain the target gas concentration parameter. The specific steps include: performing a first data segmentation on the piecewise linear parameters, dividing the parameters into two main intervals, the sudden change area and the steady-state area, according to the inflection point characteristics of the concentration response curve, to obtain a first partition parameter group; performing a second data segmentation on the first partition parameter group, subdividing the sudden change area and the steady-state area into multiple sub-intervals according to the slope changes in each interval, to obtain a second partition parameter group; inputting the second partition parameter group into a hyperbolic tangent function for a first nonlinear transformation, focusing on enhancing the sudden change area to obtain an enhanced mapping function; performing a second nonlinear transformation on the enhanced mapping function, compressing the steady-state area using a Sigmoid function to obtain an optimized mapping curve; The curve is multiplied by the dynamic weighting coefficient, and different weights are assigned to different concentration intervals through an adaptive weight adjustment algorithm to obtain a first concentration mapping vector; the first concentration mapping vector is corrected for the first time, and a polynomial compensation model is established to eliminate the nonlinear error of the system to obtain a second concentration mapping vector; the second concentration mapping vector is input into the Kalman filter, and an adaptive covariance matrix update algorithm is used for state estimation to obtain a first filter sequence; the first filter sequence is filtered for the second time, and noise interference of different frequencies is eliminated by constructing a multi-scale filter group to obtain a second filter sequence; the second filter sequence is input into an adaptive sliding window processing unit, and the window length is dynamically adjusted according to the signal change rate to obtain a first smooth sequence; multiple quantization mapping is performed on the first smooth sequence, and a segmented mapping function is established in combination with the gas concentration calibration curve to obtain a second smooth sequence; the second smooth sequence is finally calibrated through an error compensation network to achieve high-precision concentration quantitative mapping to obtain the target gas concentration parameter.

[0036] The signal acquisition method of the MEMS gas sensor array in the embodiment of the present invention is described above. The signal acquisition device of the MEMS gas sensor array in the embodiment of the present invention is described below. Figure 2 , an embodiment of the signal acquisition device of the MEMS gas sensor array in the embodiment of the present invention includes: A compensation module is used to perform analog-to-digital conversion and temperature and humidity compensation on the resistance value signals synchronously collected by multiple sensing units in the MEMS gas sensor array to obtain a compensated response matrix; An extraction module is used to perform fast Fourier transform and dynamic feature extraction on the compensated response matrix to obtain a frequency domain feature matrix of each sensor unit; A decoupling module is used to calculate the crosstalk coefficient of the sensor unit and decouple the response component according to the frequency domain characteristic matrix to obtain a decoupling characteristic vector; A transformation module is used to perform Hilbert transformation on the decoupled feature vector and extract the adsorption-desorption characteristics to obtain a gas-sensitive feature sequence; A processing module, used for inputting the gas-sensitive feature sequence into an adaptive power spectrum analysis model for cross-sensitivity identification processing to obtain a target gas response component; The mapping module is used to perform autoregressive modeling and gas concentration quantitative mapping on the target gas response component to obtain the target gas concentration parameter.

[0037] Through the coordinated cooperation of the above-mentioned components, by adopting high-precision analog-to-digital conversion and nonlinear temperature and humidity compensation models, the baseline drift caused by environmental factors is effectively eliminated, and high-fidelity acquisition of the response signal of the sensor unit is achieved; based on the transfer function model and orthogonalization processing, the crosstalk relationship between the sensor units is accurately identified and separated, and the independence of the response of each sensor unit is ensured through independent component analysis; combining frequency domain analysis and Hilbert transform to comprehensively extract gas response characteristics, accurately capture adsorption-desorption kinetics, and establish a precise mapping relationship between gas-sensitive response and concentration; adopting fast algorithms and segmented processing strategies to reduce computational complexity and improve data processing efficiency, and adaptive algorithms ensure the real-time nature of processing results; through cross-sensitivity identification, the selectivity of target gas detection is improved, the autoregressive model improves the accuracy of quantitative concentration analysis, and multi-stage filtering optimization ensures the stability of detection results, so that the present invention can meet the high-precision requirements of gas detection in complex environments.

[0038] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0039] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0040] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0041] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0042] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0043] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0044] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A signal acquisition method for a MEMS gas sensor array, characterized in that: The method comprises: The resistance value signals synchronously collected by multiple sensing units in the MEMS gas sensor array are converted into analog-to-digital and compensated for temperature and humidity to obtain a compensated response matrix; Performing fast Fourier transform and dynamic feature extraction on the compensated response matrix to obtain a frequency domain feature matrix of each sensor unit; Calculate the crosstalk coefficient of the sensor unit and decouple the response component according to the frequency domain characteristic matrix to obtain a decoupling characteristic vector; Performing Hilbert transformation and adsorption-desorption characteristic extraction on the decoupling feature vector to obtain a gas-sensitive feature sequence; Inputting the gas-sensitive characteristic sequence into an adaptive power spectrum analysis model for cross-sensitivity identification processing to obtain a target gas response component; Autoregressive modeling and gas concentration quantitative mapping are performed on the target gas response component to obtain target gas concentration parameters.

2. The signal acquisition method of the MEMS gas sensor array according to claim 1, characterized in that: The step of performing analog-to-digital conversion and temperature and humidity compensation on the resistance value signals synchronously collected by multiple sensing units in the MEMS gas sensor array to obtain a compensated response matrix includes: Parallel sampling and triggering are performed on multiple sensing units in the MEMS gas sensor array, and data is collected at a sampling frequency of 1000 Hz to obtain a sequence of original resistance values; The original resistance value sequence is input into a 16-bit high-precision analog-to-digital converter for digital-to-analog conversion, and noise is filtered out through a Butterworth low-pass filter with a cut-off frequency of 100 Hz to obtain a digital filtered signal; Arranging the sampling point data of each sensor unit in the digital filtering signal in time sequence and constructing an M×N matrix, where M is the number of sensor units and N is the number of sampling points, to obtain an original data matrix; The real-time data collected by the temperature and humidity sensor is subjected to a polynomial fitting operation with the calibration curve to obtain temperature and humidity correction parameters, and a compensation model is established based on the temperature and humidity correction parameters to calculate the temperature and humidity response coefficient of each sensing unit to obtain a compensation coefficient matrix; The original data matrix is ​​subjected to zero mean conversion and variance normalization processing, and data preprocessing is performed by singular value decomposition to obtain a standardized data matrix; The standardized data matrix and the compensation coefficient matrix are subjected to a Hadamard product operation, and the influence of temperature and humidity drift is eliminated by matrix decomposition to obtain an initial compensation result. The initial compensation result is subjected to nonlinear correction and adaptive weight iterative optimization according to the sensitivity curve of each sensor unit to obtain a compensated response matrix.

3. The signal acquisition method of the MEMS gas sensor array according to claim 2, characterized in that: The method of performing fast Fourier transform and dynamic feature extraction on the compensated response matrix to obtain a frequency domain feature matrix of each sensor unit includes: Sequentially segmenting each sensor unit signal of the compensated response matrix into 1024 points, and performing a Hanning window function product operation on each segment signal to obtain a preprocessed signal segment; Performing a fast Fourier transform on the preprocessed signal segment, extracting the complex spectrum of the first N / 2 frequency points to obtain initial spectrum data, and performing amplitude-frequency characteristic analysis on the initial spectrum data to obtain a characteristic frequency set; The phase information corresponding to the characteristic frequency set is expanded and the phase change curve is fitted to obtain a phase characteristic vector, and the phase characteristic vector and the characteristic frequency set are reconstructed into complex numbers, and the instantaneous frequency and instantaneous amplitude are calculated to obtain a dynamic parameter matrix; The dynamic parameter matrix is ​​sorted according to the magnitude of the frequency components, and the feature contribution is calculated by weighted superposition to obtain a feature weight sequence; A matrix multiplication operation is performed on the feature weight sequence and the dynamic parameter matrix to obtain a feature projection matrix, and a feature vector combination whose projection value in the feature projection matrix is ​​greater than a set threshold is orthogonalized to obtain a frequency domain feature matrix of each sensor unit.

4. The signal acquisition method of the MEMS gas sensor array according to claim 3, characterized in that: The calculating of the crosstalk coefficient of the sensing unit and the decoupling of the response component according to the frequency domain characteristic matrix to obtain the decoupling characteristic vector includes: Performing inverse Fourier transform on the response signal of each sensor unit of the frequency domain characteristic matrix, and performing crosstalk coefficient analysis according to the spatial distribution of the sensor units to obtain a time domain crosstalk component; The time domain crosstalk component is subjected to cross-correlation function calculation, and a correlation coefficient threshold is set to 0.85, and a crosstalk relationship is established for the sensor unit pairs exceeding the threshold, thereby obtaining a crosstalk unit pair set; Performing transfer function modeling on the set of crosstalk unit pairs, and calculating the crosstalk coefficients between each unit pair to obtain a crosstalk coefficient matrix; Input the crosstalk coefficient matrix into a singular value decomposition model, select singular values ​​and eigenvectors greater than 0.1 as main crosstalk components, obtain a main crosstalk feature space, perform an orthogonal basis transformation on the main crosstalk feature space, construct a decoupling basis vector by a Gram-Schmidt orthogonalization method, and obtain an orthogonal decoupling basis vector; Performing a projection operation on the orthogonal decoupling basis vectors and the response signals of each sensor unit to obtain an initial decoupling result, and performing independent component analysis and optimal separation matrix calculation on the initial decoupling result to obtain independent response components; The independent response components are sorted by variance contribution, and a response component combination with a cumulative contribution rate of 95% is selected to obtain a decoupled eigenvector.

5. The signal acquisition method of the MEMS gas sensor array according to claim 4, characterized in that: The step of performing Hilbert transformation and adsorption-desorption characteristic extraction on the decoupling feature vector to obtain a gas-sensitive feature sequence includes: Performing Hilbert transformation on the decoupled eigenvector to obtain an analytical signal sequence, and performing complex operations on the analytical signal sequence to obtain an instantaneous amplitude and an instantaneous phase sequence; Performing sliding window processing on the instantaneous amplitude and the instantaneous phase sequence to obtain a time-varying envelope matrix, and extracting the adsorption process and the desorption process in the time-varying envelope matrix respectively to construct a kinetic parameter matrix; Performing characteristic decomposition on the kinetic parameter matrix, sorting according to the eigenvalues ​​of adsorption rate and desorption rate to obtain a characteristic coefficient sequence, and performing nonlinear transformation on the characteristic coefficient sequence through an S-shaped mapping function, and at the same time, calculating the response curvature before and after the transformation to obtain a response characteristic vector; The response characteristic vector is decomposed to obtain gas-sensitive response components, and the gas-sensitive response components are combined, reconstructed and normalized according to energy contribution to obtain a gas-sensitive characteristic sequence.

6. The signal acquisition method of the MEMS gas sensor array according to claim 5, characterized in that: The step of inputting the gas-sensitive feature sequence into an adaptive power spectrum analysis model for cross-sensitivity identification processing to obtain a target gas response component includes: The gas-sensitive characteristic sequence is segmented to obtain a segmented characteristic matrix, and the segmented characteristic matrix is ​​input into an adaptive power spectrum analysis model for power spectrum estimation, and the auto-power spectrum and cross-power spectrum with a frequency resolution of 0.1 Hz are calculated by the Welch method to obtain a multi-channel power spectrum matrix; The multi-channel power spectrum matrix is ​​input into an adaptive filter bank, and spectrum smoothing is performed by a 64-order FIR filter and a recursive least squares algorithm with a forgetting factor of 0.98 to obtain a smoothed power spectrum sequence; Performing cross-correlation analysis on the power spectra between the channels in the smoothed power spectrum sequence to obtain a cross-sensitivity matrix, and inputting the cross-sensitivity matrix into a feature separation unit, extracting feature vectors by a subspace decomposition method, and obtaining a target response subspace; The target response subspace is adaptively weighted to obtain a response weight vector, and the response weight vector is projected onto the smoothed power spectrum sequence to obtain a target gas response component.

7. The signal acquisition method of the MEMS gas sensor array according to claim 6, characterized in that: The step of performing autoregressive modeling and gas concentration quantitative mapping on the target gas response component to obtain a target gas concentration parameter includes: Based on the target gas response component, the order of the autoregressive model is determined to be 8th order through partial correlation function analysis, and a model parameter sequence is obtained; Inputting the model parameter sequence into the Levinson-Durbin recursive algorithm, calculating the autoregressive coefficients through forward and backward prediction error operations to obtain a characteristic coefficient matrix, and performing Gaussian-Newton iterative optimization on the characteristic coefficient matrix to obtain optimized model parameters; Performing polynomial fitting and fitting error root mean square calculation on the optimization model parameters and the gas concentration calibration curve to obtain a mapping function matrix, and performing piecewise linearization processing on the mapping function matrix to calculate the slope and intercept of each segment to obtain piecewise linear parameters; Inputting the piecewise linear parameters into a concentration mapping model for nonlinear transformation and feature space mapping to obtain a concentration mapping vector; The concentration mapping vector is subjected to Kalman filtering to obtain a filtered concentration sequence, and the filtered concentration sequence is subjected to moving average processing to obtain a target gas concentration parameter.

8. A signal acquisition device for a MEMS gas sensor array, characterized in that: The device is used to perform a signal acquisition method for a MEMS gas sensor array according to any one of claims 1 to 7, the device comprising: A compensation module is used to perform analog-to-digital conversion and temperature and humidity compensation on the resistance value signals synchronously collected by multiple sensing units in the MEMS gas sensor array to obtain a compensated response matrix; An extraction module, used for performing fast Fourier transform and dynamic feature extraction on the compensated response matrix to obtain a frequency domain feature matrix of each sensor unit; A decoupling module, used to calculate the crosstalk coefficient of the sensor unit and decouple the response component according to the frequency domain characteristic matrix to obtain a decoupling characteristic vector; A transformation module, used for performing Hilbert transformation and adsorption-desorption characteristic extraction on the decoupling feature vector to obtain a gas-sensitive feature sequence; A processing module, used for inputting the gas-sensitive feature sequence into an adaptive power spectrum analysis model for cross-sensitivity identification processing to obtain a target gas response component; The mapping module is used to perform autoregressive modeling and gas concentration quantitative mapping on the target gas response component to obtain the target gas concentration parameter.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the signal acquisition method of the MEMS gas sensor array described in any one of claims 1 to 7 is implemented. 10 . A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the signal acquisition method for a MEMS gas sensor array according to any one of claims 1 to 7.

Images