Gas sensor array based flow sensing feedback regulation method and system

Abstract

The application relates to the technical field of industrial automation control, and discloses a flow sensing feedback regulation method and system based on a gas sensor array. The method comprises the following steps: acquiring an array voltage signal; processing the array voltage signal to extract node peak intensity; performing concentration field reconstruction processing according to the node peak intensity and a preset spatial physical coordinate to obtain a spatial concentration distribution graph; acquiring a real-time flow parameter; extracting a concentration time sequence feature from the spatial concentration distribution graph, inputting the concentration time sequence feature into a pre-trained neural network to perform time sequence deduction to obtain a predicted concentration trend; performing calculation according to the predicted concentration trend and the real-time flow parameter to obtain a flow prediction sequence; acquiring a temperature and humidity parameter; processing the temperature and humidity parameter and the flow prediction sequence to obtain a regulation instruction set; performing closed-loop error correction operation according to the regulation instruction set and a valve feedback flow to generate an opening degree driving signal. The method can realize high-precision dynamic regulation of gas flow in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of industrial control and environmental monitoring technology, and in particular to a flow sensing feedback regulation method and system based on a gas sensor array. Background Technology

[0002] Currently, in modern industrial production and environmental monitoring applications, precise control of gas flow in complex multi-gas mixed environments is a crucial core element, directly related to the safety, stability, and resource utilization efficiency of the production process.

[0003] In existing technologies, single-channel gas sensors or conventional PID controllers are typically used for static monitoring of overall concentration and passive valve regulation. However, real industrial environments often experience drastic fluctuations in temperature and humidity, as well as the cross-mixing of various interfering components. The aforementioned existing technologies rely solely on static threshold determination and single-dimensional concentration calculation, lacking a specific identification mechanism for target gases at complex frequency domain levels. They cannot effectively isolate the coupling effects of environmental interference and cross-components. Furthermore, this isolated sensing and monitoring exhibits significant response lag compared to backend flow control, making it impossible to perform time-series feedforward extrapolation and proactive compensation based on the dynamic evolution of spatial concentration. This logical disconnect between the sensing and execution layers makes the control system highly susceptible to steady-state deviations in flow matching and long-term operational oscillations when facing weak signals or frequency domain interference.

[0004] Existing technologies suffer from poor specific identification and anti-interference capabilities for target gases in complex mixed environments, as well as severe lag in flow feedback adjustment and low matching accuracy. Summary of the Invention

[0005] This invention provides a flow sensing feedback regulation method and system based on a gas sensor array to solve the problems of poor specific identification and anti-interference ability of target gas in complex mixed environments, as well as serious lag and low matching accuracy of flow feedback regulation in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a flow sensing feedback regulation method based on a gas sensor array, comprising:

[0007] The array voltage signal in a multi-gas environment is acquired, and the array voltage signal is subjected to channel-specific feature extraction processing to obtain a node feature sequence.

[0008] The node feature sequence is subjected to frequency domain transformation to obtain the node feature spectrum. The matching similarity between the node feature spectrum and the pre-constructed gas template library is calculated. The node peak intensity of the target gas is extracted based on the matching similarity.

[0009] The concentration field is reconstructed based on the node peak intensity and the preset spatial physical coordinates to obtain a spatial concentration distribution map.

[0010] The real-time flow parameters of the monitoring area are obtained, the concentration time-series features are extracted from the spatial concentration distribution map, the concentration time-series features are input into a pre-trained neural network for time-series extrapolation, the predicted concentration trend is obtained, and the predicted concentration trend is matched and calculated with the real-time flow parameters to obtain the flow prediction sequence.

[0011] The temperature and humidity parameters of the environment fluctuation are obtained, and environmental compensation mapping processing is performed on the temperature and humidity parameters and the flow prediction sequence to obtain the control instruction set;

[0012] Obtain the valve feedback flow at the pipeline regulation end, perform closed-loop error correction calculation based on the regulation instruction set and the valve feedback flow, and generate an opening drive signal.

[0013] In a second aspect, the present invention provides a flow sensing feedback regulation system based on a gas sensor array, comprising:

[0014] The data acquisition module is used to acquire array voltage signals in a multi-gas environment, and to perform channel-specific feature extraction processing on the array voltage signals to obtain node feature sequences.

[0015] The spectrum recognition module is used to perform frequency domain transformation processing on the node feature sequence to obtain the node feature spectrum, calculate the matching similarity between the node feature spectrum and the pre-built gas template library, and extract the node peak intensity of the target gas based on the matching similarity.

[0016] The spatial imaging module is used to reconstruct the concentration field based on the peak intensity of the nodes and the preset spatial physical coordinates to obtain a spatial concentration distribution map.

[0017] The spatiotemporal prediction module is used to acquire real-time flow parameters of the monitoring area, extract concentration time-series features from the spatial concentration distribution map, input the concentration time-series features into a pre-trained neural network for time-series extrapolation, obtain the predicted concentration trend, and perform matching calculations based on the predicted concentration trend and the real-time flow parameters to obtain the flow prediction sequence.

[0018] The feedforward compensation module is used to acquire temperature and humidity parameters of environmental fluctuations, and perform environmental compensation mapping processing based on the temperature and humidity parameters and the flow prediction sequence to obtain a set of control instructions.

[0019] The closed-loop execution module is used to obtain the valve feedback flow at the pipeline regulation end, perform closed-loop error correction calculation based on the regulation instruction set and the valve feedback flow, and generate an opening drive signal.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) The present invention obtains node feature sequences through channel feature extraction and further performs frequency domain transformation and gas template library matching similarity calculation, thereby realizing accurate stripping and specific identification of target signals in multi-gas mixed environment; this extraction mechanism based on frequency domain feature peaks can effectively eliminate cross interference caused by non-target gases and high-frequency noise, improve the selective identification accuracy of the system in complex gas composition environment, and solve the technical problem of low identification accuracy in multi-component mixed scenario of the prior art.

[0022] (2) This invention reconstructs the spatial concentration distribution map through spatial interpolation and uses a neural network to perform time-series deduction on the concentration time-series characteristics of continuous sampling periods to obtain the predicted concentration trend, realizing the dimensional leap from single-point static monitoring to spatiotemporal dynamic prediction; the evolution law extraction based on the neural network can perceive the trend of concentration fluctuation in advance, and combine with real-time flow parameters to complete feedforward matching calculation, effectively compensating for the inherent physical lag of the pipeline regulation system, making the flow regulation leap from passive response to active compensation, and shortening the system regulation response time.

[0023] (3) This invention obtains temperature and humidity parameters to perform environmental compensation mapping processing on the flow prediction sequence, and combines the valve feedback flow at the pipeline regulation end to perform closed-loop error correction calculation, thus constructing a composite control architecture that combines multi-dimensional feedforward and negative feedback; through environmental factors to compensate for phase shift and gain calculation on steady-state deviation, the measurement deviation and mechanical error at the execution end caused by environmental fluctuations are effectively eliminated, the steady-state error of flow matching is reduced, and the high stability and high reliability of gas flow output under variable operating conditions are guaranteed. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of the flow sensing feedback regulation method based on a gas sensor array provided in the first embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the flow sensing feedback regulation system based on a gas sensor array provided in the second embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1 The first embodiment of the present invention provides a flow sensing feedback regulation method based on a gas sensor array, comprising the following steps:

[0028] S1, acquire the array voltage signal of the multi-gas environment, and perform channel-by-channel feature extraction processing on the array voltage signal to obtain the node feature sequence;

[0029] S2, perform frequency domain transformation on the node feature sequence to obtain the node feature spectrum, calculate the matching similarity between the node feature spectrum and the pre-built gas template library, and extract the node peak intensity of the target gas based on the matching similarity;

[0030] S3, perform concentration field reconstruction processing based on the node peak intensity and preset spatial physical coordinates to obtain a spatial concentration distribution map;

[0031] S4, obtain real-time flow parameters of the monitoring area, extract concentration time-series features from the spatial concentration distribution map, input the concentration time-series features into a pre-trained neural network for time-series extrapolation, obtain the predicted concentration trend, and perform matching calculations based on the predicted concentration trend and the real-time flow parameters to obtain the flow prediction sequence;

[0032] S5, acquire the temperature and humidity parameters of the environmental fluctuations, and perform environmental compensation mapping processing based on the temperature and humidity parameters and the flow prediction sequence to obtain the control instruction set;

[0033] S6, obtain the valve feedback flow at the pipeline regulation end, perform closed-loop error correction calculation based on the regulation instruction set and the valve feedback flow, and generate an opening drive signal.

[0034] In step S1, the array voltage signal of the multi-gas environment is acquired, and the array voltage signal is subjected to channel feature extraction processing to obtain the node feature sequence.

[0035] Specifically, the array voltage signal undergoes channel-specific feature extraction processing to obtain a node feature sequence, including:

[0036] The array voltage signal is windowed to obtain a local voltage waveform;

[0037] The local voltage waveform is subjected to wavelet transform denoising to obtain a low-frequency effective signal.

[0038] Principal component features are extracted based on the low-frequency effective signal to obtain the node feature sequence.

[0039] It should be noted that the array voltage signal is a multidimensional analog electrical signal sequence that reflects the response intensity of various gases in the monitoring area. In this embodiment, the local voltage waveform is obtained by dividing the time domain, and high-frequency noise is removed by multi-scale wavelet decomposition. Finally, the spatial dimension features are aggregated by principal component analysis to output a node feature sequence that reflects the environmental characteristics.

[0040] In one implementation, this embodiment uses a multi-channel gas sensor array deployed in a multi-gas environment to synchronously acquire data, obtaining continuous raw voltage waveforms from multiple channels, which are then identified as the array voltage signal for the multi-gas environment. This embodiment performs a sliding step truncation operation along the time axis on the array voltage signal, i.e., it performs windowing processing on the array voltage signal to obtain a local voltage waveform.

[0041] It should be noted that the window time length of the windowing process is determined by acquiring the average adsorption equilibrium time constant of the target gas from the gas sensor offline, and the sliding step size is set to half of the window time length. The average adsorption equilibrium time constant is multiplied by the sensor's sampling frequency and rounded down to obtain the total number of discrete sampling points contained within the corresponding sliding window.

[0042] In one implementation, this embodiment performs wavelet transform denoising on the local voltage waveform to obtain a low-frequency effective signal. Specifically, this embodiment uses the Daubechies wavelet basis function to perform discrete wavelet decomposition on the local voltage waveform, decomposing it into approximate coefficients representing the overall trend of signal change and detail coefficients representing local abrupt changes.

[0043] It is worth noting that the number of decomposition levels in the discrete wavelet decomposition operation is determined by calculating the logarithmic ratio of the sensor's Nyquist frequency to a preset high-frequency noise lower cutoff frequency and then rounding it down. The preset high-frequency noise lower cutoff frequency is determined by acquiring continuous monitoring signals from the sensor in a background environment with no gas leakage, performing power spectral density analysis on the continuous monitoring signals, extracting the inflection point frequency value corresponding to when the cumulative energy distribution accounts for 90% of the total energy, and determining the inflection point frequency value as the preset high-frequency noise lower cutoff frequency.

[0044] In this embodiment, all detail coefficients within the decomposition layers are set to zero and discarded, while the approximation coefficients of the deepest layer are retained. The approximation coefficients of the deepest layer are then reconstructed in the time domain using the inverse wavelet transform algorithm to generate a smooth time series after filtering out high-frequency interference components, which is the low-frequency effective signal.

[0045] In one implementation, this embodiment extracts principal component features based on the low-frequency effective signals to obtain a node feature sequence. Specifically, this embodiment calculates the covariance matrix between the low-frequency effective signals of each channel and performs eigenvalue decomposition on the covariance matrix to obtain an eigenvalue set and an eigenvector matrix. This embodiment sorts the eigenvalue set in descending order of value, sequentially accumulates the eigenvalues ​​corresponding to each principal component, and divides by the sum of the eigenvalues ​​to calculate the cumulative variance contribution rate. The number of principal components corresponding to the first time the cumulative variance contribution rate reaches a preset contribution threshold is extracted and determined as the target dimensionality reduction dimension.

[0046] It should be noted that the preset contribution threshold is set at 95%, which is an objective baseline determined based on the principles of mathematical statistics to preserve most of the feature information of the original data space.

[0047] In this embodiment, column vectors corresponding to the target dimensionality reduction dimension are extracted from the feature vector matrix to construct a dimensionality reduction projection matrix; matrix multiplication mapping calculation is performed between the low-frequency effective signal and the dimensionality reduction projection matrix to obtain a data vector containing low-dimensional principal component feature values; the data vector is then concatenated in one dimension to obtain a node feature sequence.

[0048] For example, the system acquires data at a frequency of 10 Hz using an eight-channel gas sensor array to obtain the array voltage signal. Based on the sensor's adsorption equilibrium time constant, the system extracts the local voltage waveform using a sliding window with 512 sampling points. Subsequently, the system calls the Daubechies wavelet basis function, performs wavelet decomposition based on the decomposition level determined by the background noise spectrum analysis (e.g., 5 levels), discards high-frequency detail coefficients, and reconstructs the low-frequency effective signal. The system calculates the covariance matrix of the low-frequency effective signal and performs eigenvalue decomposition. When the first two principal component eigenvalues ​​are accumulated, the calculated cumulative variance contribution rate reaches 95.3%, exceeding the preset contribution threshold of 95%, thus determining the target dimensionality reduction dimension to be 2. Based on this, the system constructs a two-dimensional dimensionality reduction projection matrix, performs matrix multiplication mapping calculations on the low-frequency effective signal, and outputs a sequence containing the eigenvalues ​​of the two principal components, which is then combined to generate a node feature sequence.

[0049] In step S2, the node feature sequence is subjected to frequency domain transformation to obtain the node feature spectrum. The matching similarity between the node feature spectrum and the pre-constructed gas template library is calculated. Based on the matching similarity, the node peak intensity of the target gas is extracted.

[0050] The process of performing frequency domain transformation on the node feature sequence to obtain the node feature spectrum includes:

[0051] The frequency domain sequence is obtained by performing a Fast Fourier Transform on the node feature sequence.

[0052] Extract power spectral density features and characteristic frequency points from the frequency domain sequence;

[0053] The node feature spectrum is obtained by splicing and assembling the power spectral density characteristics and the characteristic frequency points.

[0054] It should be noted that the node feature sequence is time-domain data that reflects the principal component characteristics of the sensor node response waveform; in this embodiment, the energy distribution characteristics are extracted by performing zero-padding expansion and frequency domain mapping on the node feature sequence to obtain the node feature spectrum.

[0055] In one implementation, this embodiment performs a Fast Fourier Transform (FFT) on the node feature sequence to obtain a frequency domain sequence. To meet the radix-2 butterfly operation requirement of the FFT algorithm and improve frequency resolution, this embodiment performs zero-padding at the end of the node feature sequence, adding a zero-value sequence of the same length as the original sequence, extending the sequence length to 1024 data points. The extended sequence is then input into the FFT algorithm for calculation, outputting a sequence containing complex amplitudes, thus obtaining the frequency domain sequence.

[0056] It should be noted that the number of sampling points for the Fast Fourier Transform (FFT) process is set to 1024. This value is determined based on the original sampling frequency of the sensor array and the preset frequency resolution requirement; the original sampling frequency is divided by the preset frequency resolution to calculate the minimum number of sampling points, and the nearest power of 2 greater than the minimum number of sampling points is selected as the number of sampling points.

[0057] In one implementation, this embodiment extracts power spectral density features and feature frequency points from the frequency domain sequence. Specifically, the square of the modulus is calculated for the complex value corresponding to each frequency point in the frequency domain sequence to obtain the power value of each frequency point; the power value is divided by the total number of sampling points to obtain the power spectral density value reflecting the distribution of signal energy with frequency, and this value is determined as the power spectral density feature. The power spectral density features across the entire frequency band are traversed to search for local maxima of the power spectral density features, and the frequency values ​​corresponding to the local maxima are determined as feature frequency points.

[0058] It is worth noting that the specific logic for finding local maxima is as follows: a sliding frequency window is set. When the power spectral density value corresponding to the center frequency point of the sliding frequency window is greater than the value corresponding to other frequency points within the sliding frequency window, and is also greater than the preset average background noise energy, the center frequency point is determined to be a characteristic frequency point.

[0059] It should be noted that the method for determining the preset average background noise energy is as follows: during the system initialization phase, the sensor background signal under a clean air environment is collected, a spectrum transformation is performed, the arithmetic mean of the power spectral density values ​​of all frequency points in the full frequency band is extracted, and the arithmetic mean is solidified as the preset average background noise energy.

[0060] In one implementation, this embodiment assembles the power spectral density features and the feature frequency points to obtain the node feature spectrum. Specifically, the extracted feature frequency points and their corresponding power spectral density features are mapped using two-dimensional coordinates in ascending order of frequency, and a feature vector reflecting the frequency domain fingerprint characteristics of the signal is output to obtain the node feature spectrum.

[0061] The calculation of the matching similarity between the node feature spectrum and a pre-built gas template library, and the extraction of the node peak intensity of the target gas based on the matching similarity, includes:

[0062] The Euclidean distance between the node feature spectrum and the standard spectrum template in the pre-constructed gas template library is calculated to obtain the matching similarity.

[0063] Extract feature frequency components that satisfy a preset similarity threshold from the node feature spectrum;

[0064] The signal amplitude is read based on the characteristic frequency components to obtain the node peak intensity.

[0065] Matching similarity is a geometric index that measures the degree of similarity between the frequency domain features of the signal under test and the features of a known gas. In this embodiment, the Euclidean distance between the node feature spectrum and the standard spectrum template is calculated, and the signal amplitude of the target gas is screened by combining a preset similarity threshold to obtain the node peak intensity.

[0066] In one implementation, this embodiment calculates the Euclidean distance between the node feature spectrum and the standard spectrum template in a pre-built gas template library to obtain the matching similarity. Specifically, the numerical values ​​of each dimension of the feature vector in the node feature spectrum are extracted, and component subtraction, summation of squares, and square root operations are performed with the standard spectrum template in the pre-built gas template library to calculate the geometric distance between the two, thus obtaining the matching similarity.

[0067] It should be noted that the construction process of the pre-constructed gas template library is as follows: in a controlled laboratory environment, standard gases of typical industrial components such as methane, ammonia, carbon monoxide and carbon dioxide are sequentially introduced; the response signals of the sensor array to each standard gas are collected and frequency domain conversion processing is performed to obtain the standard spectrum template of each standard gas; the standard spectrum template is associated with its corresponding chemical component label and stored to generate the pre-constructed gas template library.

[0068] In one implementation, this embodiment extracts feature frequency components that meet a preset similarity threshold condition from the node feature spectrum. Specifically, the matching similarity is compared with the preset similarity threshold; if the matching similarity is less than or equal to the preset similarity threshold, the current signal is determined to be successfully matched, and the frequency range corresponding to the matched standard spectrum template is extracted from the node feature spectrum to obtain the feature frequency components.

[0069] It is worth noting that the method for determining the preset similarity threshold is as follows: obtain a verification dataset containing known samples, calculate the similarity distribution between each group of samples in the verification dataset and its corresponding standard template; use the receiver operating characteristic curve analysis method to find the similarity boundary value that maximizes the sum of recall and precision for target gas recognition, and label the similarity boundary value as the preset similarity threshold.

[0070] In one implementation, this embodiment performs a signal amplitude reading operation based on the characteristic frequency component to obtain the node peak intensity. Specifically, within the frequency range corresponding to the characteristic frequency component, the frequency point with the maximum power spectral density value is retrieved, and the original analog voltage amplitude corresponding to the frequency point is read to obtain the node peak intensity.

[0071] For example, the system acquires a node feature sequence with 512 sampling points, fills the end of the sequence with 512 zero values, expands it to 1024 points, and then inputs it into a Fast Fourier Transform algorithm. After calculation, the system determines that the power spectral density value is greater than the preset average background noise energy, and extracts the power spectral density feature at a frequency of 2.5 Hz, which is then assembled into a node feature spectrum. The system extracts the feature vector of the node feature spectrum and performs component subtraction, summation of squares, and square root operations with a carbon dioxide template in a pre-built gas template library, calculating a matching similarity of 0.02. The system retrieves a preset similarity threshold of 0.05, compares it, and determines that 0.02 is less than 0.05, thereby identifying the target gas and extracting the corresponding feature frequency component. The system retrieves the frequency point with the maximum power spectral density among the feature frequency components, reads the original signal amplitude corresponding to the frequency point, and obtains a node peak intensity of 1.38 volts that conforms to the analog-to-digital converter range.

[0072] In step S3, a concentration field reconstruction process is performed based on the node peak intensity and preset spatial physical coordinates to obtain a spatial concentration distribution map, including:

[0073] The node concentration value is obtained by mapping and converting the node peak intensity to a pre-built concentration calibration model;

[0074] The kriging interpolation algorithm is used to calculate the weight assignment of the preset grid nodes to be tested, and the grid weight matrix is ​​obtained.

[0075] A spatial concentration distribution map is obtained by performing spatial mapping processing based on the grid weight matrix and the node concentration values.

[0076] It should be noted that, during the system installation phase, the preset spatial physical coordinates are determined in advance using a laser rangefinder or ultrasonic positioning device to measure the relative three-dimensional positions of each gas sensor node within the pipeline cross-section or reaction chamber, and these coordinates are stored as spatial physical coordinates in the system's non-volatile memory. The node concentration value is a numerical characteristic that quantifies the target gas mass concentration at each sampling point. In this embodiment, the node peak intensity is converted into the node concentration value using a pre-constructed concentration calibration model. The grid weight matrix is ​​a combination of coefficients that quantifies the contribution of each known node to the concentration of unknown points in space. In this embodiment, the grid weight matrix is ​​calculated using the Kriging interpolation algorithm and linearly weighted by combining it with the node concentration values ​​to generate a continuous map covering the monitoring area, i.e., a spatial concentration distribution map.

[0077] In one implementation, this embodiment maps the node peak intensities to a pre-built concentration calibration model to obtain node concentration values. Specifically, the extracted peak intensities of each node are input as independent variables into the concentration calibration model. The concentration calibration model is then used to perform a nonlinear function mapping operation to calculate the corresponding concentration dimension value, which is then determined as the node concentration value.

[0078] It should be noted that the process of constructing the concentration calibration model is as follows: under standard operating conditions, the sensor array is placed in a standard gas with known concentration gradients, and the peak voltage output of the sensor array is recorded; the least squares method is used to fit the nonlinear regression curve between the peak voltage and the standard gas concentration, the regression coefficients are extracted, and the concentration calibration model is constructed.

[0079] In one implementation, this embodiment utilizes the Kriging interpolation algorithm to calculate the weight allocation for preset grid nodes to be measured, obtaining a grid weight matrix. Specifically, the monitoring area is divided into a two-dimensional spatial grid with a fixed resolution. Each coordinate point to be calculated in the two-dimensional spatial grid is extracted and identified as a grid node to be measured. The fixed resolution is determined by half of the minimum spatial distance between adjacent nodes in the gas sensor array, ensuring that the Kriging interpolation algorithm can effectively capture the spatial variation characteristics of the concentration field at this resolution. The spatial straight-line distance between the grid node to be measured and the corresponding spatial physical coordinates of all known sensor nodes is calculated. The spatial autocorrelation characteristics between nodes are analyzed using the exponential semi-variogram model in the ordinary Kriging interpolation algorithm. The Kriging equations are solved to obtain the contribution coefficients of each known sensor node relative to the grid node to be measured. All the contribution coefficients are then matrix-arranged to obtain the grid weight matrix.

[0080] It should be noted that the calculation formula for the exponential semivariogram model is as follows:

[0081]

[0082] in, This represents the semivariogram values ​​between spatial grid nodes; It represents the straight-line distance between nodes, and its value is derived by extracting the three-dimensional coordinate values ​​of the corresponding spatial physical coordinates of the two nodes and performing Euclidean distance calculation; It represents the nugget value, characterizing local random fluctuations caused by measurement errors or micro-variations; This represents the partial sill value, which characterizes the structural variance caused by the spatial autocorrelation component; This indicates the range of influence of spatial autocorrelation.

[0083] It is worth noting that the value of the gold nuggets... The off-base slab value and the range The determination method is as follows: historical monitoring data of the sensor array under the condition of no gas leakage is collected, the variance distribution of known observations under different spatial straight-line distances is calculated, and an experimental semivariogram scatter plot is drawn; the experimental semivariogram scatter plot is fitted with a nonlinear regression algorithm, and the mathematical parameters corresponding to the fitted curve are extracted and solidified as the nugget value, the sill value and the range, respectively.

[0084] In one implementation, this embodiment performs spatial mapping processing based on the grid weight matrix and the node concentration values ​​to obtain a spatial concentration distribution map. Specifically, it extracts the weight coefficients corresponding to each of the grid nodes to be tested from the grid weight matrix, and performs a linear weighted summation operation on each weight coefficient and the corresponding node concentration value to obtain the estimated concentration of the grid node to be tested. The estimated concentrations of all the grid nodes to be tested are calculated iteratively, and then rendered using color gradation according to the spatial physical coordinates to generate a continuous spatial concentration distribution map.

[0085] For example, the system reads a peak intensity of 1.38 volts at a certain sensor node. Using a pre-built concentration calibration model, it performs mapping calculations and obtains a node concentration of 45.6 ppm at that location. The system sets the grid resolution to 0.5 meters and calculates the spatial straight-line distance between the node to be measured and known nodes. For example, 2.0 meters. The system retrieves parameters calibrated by fitting historical observation data, nugget value. 0.05, off-center sill value For 0.85 and range The distance is 10.0 meters. Substituting this value into an exponential semivariogram model, the semivariogram value is calculated. Then, the ordinary kriging equations are solved to obtain the grid weight matrix. After linear weighted summation, the system generates a spatial concentration distribution map, showing that the peak concentration at the center of the monitoring area is 48.2 ppm, decreasing to 32.7 ppm towards the edge.

[0086] In step S4, real-time flow parameters of the monitoring area are obtained, concentration time-series features are extracted from the spatial concentration distribution map, the concentration time-series features are input into a pre-trained neural network for time-series extrapolation to obtain a predicted concentration trend, and the predicted concentration trend is matched and calculated with the real-time flow parameters to obtain a flow prediction sequence.

[0087] The process includes extracting concentration time-series features from the spatial concentration distribution map, inputting these features into a pre-trained neural network for time-series extrapolation, and obtaining a predicted concentration trend, including:

[0088] Within a preset sliding time window, the center concentration value is extracted from the spatial concentration distribution map;

[0089] Sequence combination operations are performed based on the central concentration value to obtain the concentration time-series characteristics;

[0090] The concentration time-series features are input into a pre-trained long short-term memory network model to deduce the evolution law and obtain the predicted concentration trend.

[0091] It should be noted that the concentration time series feature is a one-dimensional feature vector that quantifies the dynamic evolution trend of the target gas concentration in the monitoring area over time. In this embodiment, the concentration time series feature reflecting the core fluctuation is obtained by performing truncation and extreme value extraction operations on the spatial concentration distribution map in the time dimension. The predicted concentration trend represents the system's prediction parameters for the environmental state within a preset time period. By inputting the concentration time series feature into a neural network model with time memory characteristics, the nonlinear evolution law is extracted to obtain the predicted concentration trend referenced by subsequent flow compensation.

[0092] In one implementation, this embodiment extracts the center concentration value from the spatial concentration distribution map within a preset sliding time window. Specifically, this embodiment acquires the continuously generated spatial concentration distribution maps at a preset sampling frequency and loads them into a memory queue; iterates through the memory queue, locks the grid node corresponding to the center region of the geographical location in each frame of the spatial concentration distribution map, reads the mass concentration value of the grid node, and obtains the center concentration value.

[0093] It should be noted that the time span of the preset sliding time window is determined by acquiring the physical geometric scale of the monitoring area and the average diffusion rate of the target gas under standard operating conditions. The characteristic length of the monitoring area is divided by the average diffusion rate to calculate the characteristic time for the airflow to pass through the area. This characteristic time is then multiplied by a preset safety factor, such as a constant of 2.0, and the result is used to determine the time span of the preset sliding time window. The preset safety factor is an objective redundancy coefficient derived based on the 3-sigma principle, using the variance offset of the pipeline gas flow rate under maximum load fluctuations from historical monitoring data. The sampling frequency is determined by reading the minimum response time of the sensor and performing a reciprocal operation, ensuring that the extracted dynamic data fully covers the transient fluctuation characteristics of the concentration.

[0094] In one implementation, this embodiment performs a sequence combination operation based on the central concentration values ​​to obtain concentration time-series features. Specifically, this embodiment arranges multiple central concentration values ​​extracted within the same preset sliding time window into an array according to the sampling time sequence, generating a one-dimensional vector reflecting the concentration fluctuation over time, thus obtaining the concentration time-series features.

[0095] In one implementation, this embodiment inputs the concentration time-series features into a pre-trained Long Short-Term Memory (LSTM) network model to deduce evolutionary patterns and obtain a predicted concentration trend. Specifically, this embodiment inputs the concentration time-series features as an input tensor into the LTM network model. The input gate, forget gate, and output gate structures within the LTM network model are used to weight the long-term dependencies and short-term fluctuations in the time-series data, outputting a numerical sequence reflecting the concentration trend within a future prediction period, thus obtaining the predicted concentration trend.

[0096] It is worth noting that the construction process of the pre-trained Long Short-Term Memory (LSTM) network model is as follows: In this embodiment, an LSM network architecture with two hidden layers is constructed, with each hidden layer having 64 neurons. In a standard industrial pipeline environment that has been running continuously for 30 days, the concentration response sequence of the sensor array and the feedback sequence of the high-precision flow meter at the back end are collected at a frequency of 1 Hz. A sequence segment with a length of 600 seconds is extracted, and rolling sampling is performed with a sliding step size of 172 seconds. The first 500 seconds of the sequence are used as input feature parameters, and the real concentration trend of the last 100 seconds is used as the label parameter for supervised learning, thus constructing a training sample set of 15,000 sets. The network weights are iteratively updated using an adaptive moment estimation optimization algorithm. When the root mean square error of the model on the validation set decreases to the preset convergence index and the number of iterations reaches 200, the network parameters are fixed, and the pre-trained LSM network model is obtained.

[0097] It should be noted that the method for determining the number of neurons, 64, is as follows: using a grid search method, the search range for the number of hidden layer neurons is set to 16 to 128, and the step size is set to 16. On the validation set, the root mean square error is minimized as the objective function to perform traversal optimization, and the number of neurons when the root mean square error curve reaches the convergence inflection point is extracted, i.e., 64 neurons, which is then solidified as a model hyperparameter. It is worth noting that, considering that sliding with a step size of 172 seconds may cause overlap and autocorrelation of time series data, leading to data leakage during training, as an optional embodiment, the sliding step size can be set to 10 seconds.

[0098] The flow prediction sequence is obtained by matching the predicted concentration trend with the real-time flow parameters, including:

[0099] The dynamic coupling coefficient is obtained by regression calculation based on the predicted concentration trend and the real-time flow parameters;

[0100] The adjustment cost is calculated based on the dynamic coupling coefficient to obtain the compensation flow value;

[0101] The flow prediction sequence is obtained by superimposing the compensated flow value with the preset baseline flow.

[0102] It should be noted that the dynamic coupling coefficient is a dimensionless parameter that quantifies the physical mapping relationship between the concentration change and the flow regulation. In this embodiment, the sensitivity of the environmental concentration to the flow feedback is determined by regression analysis to obtain the dynamic coupling coefficient. The compensation flow value represents the flow correction amount to smooth out future concentration fluctuations. The compensation amount is optimized by performing normalized weighted cost calculation, and the flow prediction sequence for driving the actuator is generated by combining the basic operating parameters of the system.

[0103] In one implementation, this embodiment performs regression calculations based on the predicted concentration trend and the real-time flow parameter to obtain a dynamic coupling coefficient. Specifically, this embodiment obtains the total flow value of the pipeline regulation system in real-time operation and determines it as the real-time flow parameter; it constructs a linear regression equation using the predicted concentration trend as the dependent variable and the real-time flow parameter as the independent variable; it then uses the least squares method to fit the historical observation data of the predicted concentration trend and the real-time flow parameter, solves for the slope term of the equation, and outputs it as the dynamic coupling coefficient. The dynamic coupling coefficient, in a physical sense, represents the corresponding response of the predicted concentration trend when the flow rate changes by one unit span.

[0104] In one implementation, this embodiment calculates the adjustment cost based on the dynamic coupling coefficient to obtain the compensation flow value. Specifically, this embodiment constructs a cost evaluation function. The calculation logic of the cost evaluation function is as follows: it calculates the deviation of the predicted concentration trend from a preset safety threshold, and performs maximum-minimum normalization on the deviation to obtain a first dimensionless parameter; the preset safety threshold is determined by retrieving the national industrial health standard concentration limit of the corresponding target gas, multiplying it by a preset engineering warning coefficient, which can be determined by fitting the limit distribution of historical leakage accident data of the system, and solidifying it as the preset safety threshold. The manufacturer's specifications for the valve actuator are retrieved, and the rated regulating life and mean time between failures (MTBF) of the valve actuator are extracted. The ratio of the two is calculated and normalized to obtain a second dimensionless parameter, i.e., a preset mechanical loss factor. A preset safety weight coefficient, such as 0.7, is assigned to the first dimensionless parameter, and a preset economic weight coefficient, such as 0.3, is assigned to the second dimensionless parameter. The preset safety weight coefficient and the preset economic weight coefficient are determined by using a multi-objective optimization algorithm, such as the NSGA-II algorithm, to perform Pareto front optimization on the concentration exceeding penalty cost and valve mechanical wear reset cost in the historical operation log, and extracting and fixing the weight allocation ratio corresponding to the inflection point in the Pareto optimal solution set. The weighted first dimensionless parameter and the second dimensionless parameter are then algebraically summed to obtain the comprehensive adjustment cost value. In this embodiment, multiple schemes are compared in conjunction with the dynamic coupling coefficient, and the flow rate change that minimizes the comprehensive adjustment cost value is selected as the compensation flow rate value.

[0105] Specifically, the second dimensionless parameter can be obtained by acquiring the number of adjustments and the cumulative running time of the valve actuator in real time; dividing the number of adjustments by the rated adjustment life to obtain the mechanical wear rate, and dividing the cumulative running time by the mean time between failures to obtain the equipment aging rate; adding the mechanical wear rate and the equipment aging rate, and performing maximum and minimum value normalization processing on the sum, thereby obtaining the second dimensionless parameter that dynamically changes with the equipment status, so as to accurately reflect the mechanical wear risk of the valve.

[0106] In one implementation, this embodiment combines the compensated flow rate value with a preset baseline flow rate to obtain a flow prediction sequence. Specifically, this embodiment obtains an initial flow rate setpoint to maintain steady-state system operation and determines it as the preset baseline flow rate. Specifically, the initial flow rate setpoint is obtained by arithmetic averaging the rated operating flow rate during the pipeline system's process design phase with the average flow rate under steady-state leak-free conditions from historical operating data. The compensated flow rate value is then added to the preset baseline flow rate at time points to generate an instruction sequence containing target flow rate setpoints for future time periods, thus obtaining the flow prediction sequence.

[0107] For example, the system extracts the central concentration value within a 10-minute sliding time window, for instance, within the range of 46.5 parts per million (ppm) to 49.1 ppm. This central concentration value is input into a Long Short-Term Memory (LSTM) network model with 64 hidden layer units for extrapolation, calculating the predicted concentration trend for the next hour. The predicted peak concentration of this trend is 50.3 ppm. Subsequently, the system acquires real-time flow parameters, for example, 5.0 cubic meters per hour, and obtains a dynamic coupling coefficient of 0.7 through regression analysis. In the cost assessment phase, the system normalizes the concentration deviation and multiplies it by a safety weight of 0.7, normalizes the valve loss and multiplies it by an economic weight of 0.3, and calculates the minimum value by summing and determining a compensation flow rate of 0.2 cubic meters per hour that minimizes the overall adjustment cost. This compensation flow rate is then superimposed on the baseline flow rate to generate a flow prediction sequence, for example, with the target flow rate set at 5.2 cubic meters per hour.

[0108] In step S5, temperature and humidity parameters of environmental fluctuations are obtained, and environmental compensation mapping processing is performed on the temperature and humidity parameters and the flow prediction sequence to obtain a control instruction set.

[0109] The control instruction set is obtained by performing environmental compensation mapping processing based on the temperature and humidity parameters and the flow prediction sequence, including:

[0110] The signal response delay is obtained by evaluating and calculating based on the temperature and humidity parameters.

[0111] Optimized prediction parameters are obtained by performing a leading phase shift calculation based on the signal response delay and the traffic prediction sequence.

[0112] The optimized prediction parameters are searched and matched in a preset control strategy mapping table to obtain a set of control instructions.

[0113] It should be noted that the control instruction set is the set of underlying machine parameters that drive the physical actuator to adjust the gas flow, such as pipeline regulating valves. In this embodiment, by extracting the characteristics of environmental temperature and humidity fluctuations, the response delay caused by environmental temperature and humidity fluctuations to the physicochemical adsorption process of gas sensors is evaluated. Then, the flow prediction sequence is shifted backward on the time axis, i.e., advanced phase shift. Finally, the compensated flow demand is converted into specific hardware execution instructions through table lookup mapping, thus eliminating the feedforward control lag caused by environmental temperature and humidity fluctuations at the data processing link level.

[0114] In one implementation, this embodiment evaluates and calculates the signal response delay based on the temperature and humidity parameters. Specifically, this embodiment acquires real-time temperature and relative humidity values ​​from environmental sensors deployed in the monitoring area, combines the temperature and relative humidity values ​​to determine the temperature and humidity parameters representing environmental fluctuations. This embodiment inputs the temperature and humidity parameters into a pre-built support vector machine regression model to perform nonlinear mapping operations, outputting a scalar value representing the time dimension, which is then determined as the signal response delay.

[0115] It should be noted that the construction process of the pre-built support vector machine regression model is as follows: In a controllable environmental simulation chamber, multiple temperature and humidity cross-gradient conditions are set; under each cross-gradient condition, a target gas of standard concentration is transiently injected into the environmental simulation chamber, and the time required for the gas sensor to reach the true steady-state concentration value is recorded; the time and the standard environment are arithmetically subtracted from the baseline response time at 20 degrees Celsius and 50% relative humidity, and the difference is extracted as training label data; a radial basis function is selected as the kernel function, and a grid search algorithm is used, setting the search range of the penalty parameter to 1 to 100 and the search range of the kernel width to 0.01 to 1.0; combined with k-fold cross-validation, five-fold cross-validation is used to find the parameter combination with the minimum mean square error of the validation set, determining the optimal penalty parameter and the optimal kernel width; the optimal penalty parameter and the optimal kernel width are substituted into the model, and the regression model is fitted and the parameters are solidified by minimizing the structural risk loss function, resulting in the pre-built support vector machine regression model.

[0116] In one implementation, this embodiment performs a leading phase shift calculation based on the signal response delay and the flow prediction sequence to obtain optimized prediction parameters. Specifically, this embodiment extracts the timestamp nodes corresponding to each predicted flow value in the flow prediction sequence; subtracts the signal response delay value from the value of each timestamp node to obtain a corrected timestamp sequence. Mathematically, this operation is equivalent to shifting the flow prediction curve along the time axis in a direction earlier than the current time; using a linear interpolation algorithm, the shifted discrete curve is resampled and aligned to the standard clock control cycle time point of the underlying control system; the aligned target flow sequence is then extracted and determined as the optimized prediction parameters.

[0117] In one implementation, this embodiment retrieves and matches the optimized prediction parameters in a preset control strategy mapping table to obtain a control instruction set. This embodiment iterates through the target flow values ​​for each time period in the optimized prediction parameters, calculates the absolute difference of the values ​​in the retrieval index column of the preset control strategy mapping table, and locks the matching row with the smallest absolute difference. The hardware control reference parameters corresponding to the matching row are extracted, and the hardware control reference parameters are encapsulated into machine-recognizable data packets according to the underlying communication protocol of the actuator to generate the control instruction set.

[0118] It is worth noting that the method for constructing the preset control strategy mapping table is as follows: before the pipeline system is officially put into operation, an offline physical calibration experiment is conducted, and the various hardware control reference parameters of the pipeline actuator are gradually adjusted in fixed small steps, such as the basic valve opening percentage; after the gas hydrodynamic state in the pipeline reaches a steady state, the actual flow rate value fed back by the high-precision standard flow meter is recorded; the actual flow rate value is used as the primary key index, and the corresponding hardware control reference parameters are used as associated data, and entered into a relational database to generate the preset control strategy mapping table.

[0119] For example, the application scenario of this embodiment is an industrial reaction system or ventilation network system with fluid pipelines; the sensor array is uniformly arranged at the cross-sectional nodes of the pipeline or the inner wall of the mixing reaction chamber, and the spatial concentration distribution map is a two-dimensional or three-dimensional concentration grid map of the pipeline cross-section or the interior of the chamber; the real-time flow parameters and valve feedback flow are both fluid flow rates within the pipeline; further, in this embodiment, the signal response delay is replaced by a sensitivity attenuation compensation coefficient, specifically, the acquired temperature and relative humidity values ​​are input into a pre-built support vector machine regression model, and a sensitivity attenuation compensation coefficient ranging from 0.8 to 1.2 is output; each target flow value in the flow prediction sequence is multiplied by the sensitivity attenuation compensation coefficient to obtain optimized prediction parameters; this can effectively eliminate sensor baseline drift caused by temperature and humidity fluctuations and improve the accuracy of flow prediction.

[0120] In step S6, the valve feedback flow at the pipeline regulating end is obtained, and a closed-loop error correction calculation is performed based on the regulation instruction set and the valve feedback flow to generate an opening drive signal.

[0121] Specifically, the process of performing closed-loop error correction calculations based on the control instruction set and the valve feedback flow to generate an opening drive signal includes:

[0122] Extract the target flow value from the set of control instructions;

[0123] The steady-state deviation is obtained by calculating the deviation between the target flow rate and the valve feedback flow rate.

[0124] The corrected gain is obtained by performing proportional-integral-differential gain calculation based on the steady-state deviation.

[0125] The opening drive signal is obtained by superimposing the corrected gain with the preset opening reference.

[0126] It should be noted that steady-state deviation is an objective physical quantity characterizing the numerical difference between the expected flow rate issued by the command and the actual flow rate executed in the physical pipeline; correction gain is a dynamic compensation amount calculated using proportional-integral-derivative (PID) control logic to address the aforementioned physical execution error; and the opening degree drive signal is the electrical drive parameter ultimately output to the actuator's underlying layer. This embodiment combines the feedforward command set with the negative feedback data from the execution end through closed-loop error correction calculations to generate an opening degree drive signal containing dynamic compensation, thus eliminating mechanical execution errors and pipeline pressure disturbances at the control architecture level.

[0127] In one implementation, this embodiment extracts the target flow rate value from the control instruction set. Specifically, this embodiment parses the data packets of the control instruction set, extracts the numerical parameters representing the desired output flow rate, and obtains the target flow rate value. Simultaneously, the instantaneous volumetric flow rate of the actual fluid passing through is obtained by a flow sensor deployed at the downstream end of the gas pipeline regulating valve, thus obtaining the valve feedback flow rate.

[0128] In one implementation, this embodiment calculates the steady-state deviation based on the deviation between the target flow rate and the valve feedback flow rate. Specifically, this embodiment uses the target flow rate as the minuend and the valve feedback flow rate as the subtrahend, performs an arithmetic subtraction operation, calculates the difference representing the current physical execution error, and obtains the steady-state deviation.

[0129] In one implementation, this embodiment performs proportional-integral-differential gain calculations based on the steady-state deviation to obtain a correction gain. Specifically, this embodiment multiplies the steady-state deviation by a preset proportionality coefficient to obtain the proportional gain; performs cumulative integration on the steady-state deviation in the time domain to obtain the integral deviation, and multiplies the integral deviation by a preset integration coefficient to obtain the integral gain; calculates the first derivative of the steady-state deviation over time to obtain the rate of change of the deviation, and multiplies the rate of change of the deviation by a preset differential coefficient to obtain the differential gain. This embodiment performs an algebraic summation operation on the proportional gain, the integral gain, and the differential gain to obtain the correction gain.

[0130] It should be noted that the preset proportional coefficient, the preset integral coefficient, and the preset derivative coefficient were determined through physical experimental calibration using the critical proportional gain method, i.e., the Ziegler-Nichols method, before the pipeline system was officially put into operation. The specific calibration and derivation logic is as follows: First, the integral and derivative control actions of the system are disconnected. Under pure proportional control, the proportional gain of the controller is gradually increased until the flow output of the system pipeline exhibits continuous constant-amplitude oscillations. The critical proportional gain value corresponding to this point, and the critical oscillation period value corresponding to the constant-amplitude oscillation, are recorded. After obtaining the critical proportional gain value and the critical oscillation period value, an algebraic formula transformation is performed according to the Ziegler-Nichols empirical parameter mapping rule.

[0131] The formula for calculating the preset proportional coefficient is as follows:

[0132]

[0133] The formula for calculating the preset integral coefficient is as follows:

[0134]

[0135] The preset formula for calculating the differential coefficients is as follows:

[0136]

[0137] in, This represents the preset proportionality coefficient. This represents the preset integral coefficient. This represents the preset differential coefficient; This represents the critical proportional gain value objectively observed in a closed-loop constant-amplitude oscillation physical experiment. This represents the critical oscillation period value objectively observed in the physical experiment.

[0138] In one implementation, this embodiment superimposes the corrected gain with a preset opening reference to obtain the opening drive signal. Specifically, this embodiment extracts the base valve opening value output from the preset control strategy mapping table in the preceding steps and determines it as the preset opening reference. This embodiment performs arithmetic addition on the corrected gain and the preset opening reference to generate a final valve opening target value after closed-loop compensation. The final valve opening target value is then converted into a pulse width modulation duty cycle parameter corresponding to the underlying hardware communication protocol to obtain the opening drive signal directly acting on the hardware execution unit.

[0139] For example, the system extracts the target flow rate of 5.2 cubic meters per hour from the control command set, and collects the current valve feedback flow rate of 5.18 cubic meters per hour in real time through the sensor network. The system performs arithmetic subtraction to calculate the steady-state deviation as 0.02 cubic meters per hour. Subsequently, the system calls the proportional-integral-derivative (PID) control algorithm, retrieving objective parameters measured by the system through closed-loop constant-amplitude oscillation experiments before commissioning, such as a critical proportional gain of 4.17 and a critical oscillation period of 1.2 minutes. The system substitutes the empirical parameter mapping formula to calculate the preset proportional coefficient as 2.5 (obtained by multiplying the critical proportional gain of 4.17 by a constant of 0.6), the preset integral coefficient as 4.17 (obtained by multiplying the critical proportional gain of 4.17 by a constant of 1.2 and then dividing by the critical oscillation period of 1.2), and the preset derivative coefficient as 0.38 (obtained by multiplying the critical proportional gain of 4.17 by a constant of 0.075 and then multiplying by the critical oscillation period of 1.2). The system combines the aforementioned coefficients and steady-state deviation to calculate the proportional, integral, and differential terms, and then performs algebraic summation to obtain the correction gain characterizing the valve opening adjustment, for example, a 0.2% opening increment. The system extracts the current preset opening reference of 45.0%, adds the preset opening reference to the 0.2% correction gain, generates the parameter for the final valve opening of 45.2%, and converts this parameter into an opening drive signal, which is then sent to the valve actuator for physical adjustment.

[0140] In summary, this invention constructs a deep control logic from multi-dimensional perception to proactive prediction by integrating channel-specific feature extraction, frequency domain feature specificity identification, spatial concentration field reconstruction, and concentration evolution time-series extrapolation based on long short-term memory networks. Furthermore, by combining phase-advance compensation for environmental fluctuations and closed-loop error correction calculations based on pipeline regulation feedback, it achieves high-precision, adaptive control of target gas flow under multi-gas mixing interference and dynamic operating conditions. This invention not only significantly enhances the system's selective identification capability and anti-interference performance for specific gases, but also completely overcomes the lag of traditional feedback regulation through spatiotemporal feedforward extrapolation, significantly reducing the steady-state error of flow matching, and providing a highly reliable intelligent solution for precise gas flow management in industrial and environmental monitoring fields.

[0141] Reference Figure 2 The second embodiment of the present invention provides a flow sensing feedback regulation system based on a gas sensor array, comprising:

[0142] The data acquisition module is used to acquire array voltage signals in a multi-gas environment, and to perform channel-specific feature extraction processing on the array voltage signals to obtain node feature sequences.

[0143] The spectrum recognition module is used to perform frequency domain transformation processing on the node feature sequence to obtain the node feature spectrum, calculate the matching similarity between the node feature spectrum and the pre-built gas template library, and extract the node peak intensity of the target gas based on the matching similarity.

[0144] The spatial imaging module is used to reconstruct the concentration field based on the peak intensity of the nodes and the preset spatial physical coordinates to obtain a spatial concentration distribution map.

[0145] The spatiotemporal prediction module is used to acquire real-time flow parameters of the monitoring area, extract concentration time-series features from the spatial concentration distribution map, input the concentration time-series features into a pre-trained neural network for time-series extrapolation, obtain the predicted concentration trend, and perform matching calculations based on the predicted concentration trend and the real-time flow parameters to obtain the flow prediction sequence.

[0146] The feedforward compensation module is used to acquire temperature and humidity parameters of environmental fluctuations, and perform environmental compensation mapping processing based on the temperature and humidity parameters and the flow prediction sequence to obtain a set of control instructions.

[0147] The closed-loop execution module is used to obtain the valve feedback flow at the pipeline regulation end, perform closed-loop error correction calculation based on the regulation instruction set and the valve feedback flow, and generate an opening drive signal.

[0148] It should be noted that the flow sensing feedback regulation system based on a gas sensor array provided in this embodiment of the invention is used to execute all the process steps of the flow sensing feedback regulation method based on a gas sensor array in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0149] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0150] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A flow sensing feedback regulation method based on a gas sensor array, characterized in that, include: The array voltage signal in a multi-gas environment is acquired, and the array voltage signal is subjected to channel-specific feature extraction processing to obtain a node feature sequence. The node feature sequence is subjected to frequency domain transformation to obtain the node feature spectrum. The matching similarity between the node feature spectrum and the pre-constructed gas template library is calculated. The node peak intensity of the target gas is extracted based on the matching similarity. The concentration field is reconstructed based on the node peak intensity and the preset spatial physical coordinates to obtain a spatial concentration distribution map. The real-time flow parameters of the monitoring area are obtained, the concentration time-series features are extracted from the spatial concentration distribution map, the concentration time-series features are input into a pre-trained neural network for time-series extrapolation, the predicted concentration trend is obtained, and the predicted concentration trend is matched and calculated with the real-time flow parameters to obtain the flow prediction sequence. The temperature and humidity parameters of the environment fluctuation are obtained, and environmental compensation mapping processing is performed on the temperature and humidity parameters and the flow prediction sequence to obtain the control instruction set; Obtain the valve feedback flow at the pipeline regulation end, perform closed-loop error correction calculation based on the regulation instruction set and the valve feedback flow, and generate an opening drive signal.

2. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The step of performing channel-specific feature extraction processing on the array voltage signal to obtain a node feature sequence includes: The array voltage signal is windowed to obtain a local voltage waveform; The local voltage waveform is subjected to wavelet transform denoising to obtain a low-frequency effective signal. Principal component features are extracted based on the low-frequency effective signal to obtain the node feature sequence.

3. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The step of performing frequency domain transformation on the node feature sequence to obtain the node feature spectrum includes: The frequency domain sequence is obtained by performing a Fast Fourier Transform on the node feature sequence. Extract power spectral density features and characteristic frequency points from the frequency domain sequence; The node feature spectrum is obtained by splicing and assembling the power spectral density characteristics and the characteristic frequency points.

4. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The calculation of the matching similarity between the node feature spectrum and the pre-built gas template library, and the extraction of the node peak intensity of the target gas based on the matching similarity, includes: The Euclidean distance between the node feature spectrum and the standard spectrum template in the pre-constructed gas template library is calculated to obtain the matching similarity. Extract feature frequency components that satisfy a preset similarity threshold from the node feature spectrum; The signal amplitude is read based on the characteristic frequency components to obtain the node peak intensity.

5. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The process of reconstructing the concentration field based on the node peak intensity and preset spatial physical coordinates to obtain a spatial concentration distribution map includes: The node concentration value is obtained by mapping and converting the node peak intensity to a pre-built concentration calibration model; The kriging interpolation algorithm is used to calculate the weight assignment of the preset grid nodes to be tested, and the grid weight matrix is ​​obtained. A spatial concentration distribution map is obtained by performing spatial mapping processing based on the grid weight matrix and the node concentration values.

6. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The step of extracting concentration time-series features from the spatial concentration distribution map and inputting these features into a pre-trained neural network for time-series extrapolation to obtain a predicted concentration trend includes: Within a preset sliding time window, the center concentration value is extracted from the spatial concentration distribution map; Sequence combination operations are performed based on the central concentration value to obtain the concentration time-series characteristics; The concentration time-series features are input into a pre-trained long short-term memory network model to deduce the evolution law and obtain the predicted concentration trend.

7. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The step of matching and calculating the predicted concentration trend with the real-time flow parameters to obtain the flow prediction sequence includes: The dynamic coupling coefficient is obtained by regression calculation based on the predicted concentration trend and the real-time flow parameters; The adjustment cost is calculated based on the dynamic coupling coefficient to obtain the compensation flow value; The flow prediction sequence is obtained by superimposing the compensated flow value with the preset baseline flow.

8. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The step of performing environmental compensation mapping processing based on the temperature and humidity parameters and the flow prediction sequence to obtain a control instruction set includes: The signal response delay is obtained by evaluating and calculating based on the temperature and humidity parameters. Optimized prediction parameters are obtained by performing a leading phase shift calculation based on the signal response delay and the traffic prediction sequence. The optimized prediction parameters are searched and matched in a preset control strategy mapping table to obtain a set of control instructions.

9. The flow sensing feedback regulation method based on a gas sensor array according to claim 1, characterized in that, The step of performing closed-loop error correction calculation based on the control command set and the valve feedback flow to generate an opening drive signal includes: Extract the target flow value from the set of control instructions; The steady-state deviation is obtained by calculating the deviation between the target flow rate and the valve feedback flow rate. The corrected gain is obtained by performing proportional-integral-differential gain calculation based on the steady-state deviation. The opening drive signal is obtained by superimposing the corrected gain with the preset opening reference.

10. A flow sensing feedback control system based on a gas sensor array, characterized in that, include: The data acquisition module is used to acquire array voltage signals in a multi-gas environment, and to perform channel-specific feature extraction processing on the array voltage signals to obtain node feature sequences. The spectrum recognition module is used to perform frequency domain transformation processing on the node feature sequence to obtain the node feature spectrum, calculate the matching similarity between the node feature spectrum and the pre-built gas template library, and extract the node peak intensity of the target gas based on the matching similarity. The spatial imaging module is used to reconstruct the concentration field based on the peak intensity of the nodes and the preset spatial physical coordinates to obtain a spatial concentration distribution map. The spatiotemporal prediction module is used to acquire real-time flow parameters of the monitoring area, extract concentration time-series features from the spatial concentration distribution map, input the concentration time-series features into a pre-trained neural network for time-series extrapolation, obtain the predicted concentration trend, and perform matching calculations based on the predicted concentration trend and the real-time flow parameters to obtain the flow prediction sequence. The feedforward compensation module is used to acquire temperature and humidity parameters of environmental fluctuations, and perform environmental compensation mapping processing based on the temperature and humidity parameters and the flow prediction sequence to obtain a set of control instructions. The closed-loop execution module is used to obtain the valve feedback flow at the pipeline regulation end, perform closed-loop error correction calculation based on the regulation instruction set and the valve feedback flow, and generate an opening drive signal.

Images