A gas analysis method and system for gas source purification processes

Abstract

This application discloses a gas analysis method and system for gas source purification processes, relating to the field of gas detection and analysis technology. The method includes: simultaneously acquiring optical, electrochemical, and conductivity detection signals in the gas source purification pipeline, and combining these with environmental parameters such as temperature, humidity, and flow rate; filtering, anomaly removal, and normalization of the detection signals to construct a multi-dimensional feature vector and reduce its dimensionality to obtain principal component features; dynamically allocating weights through an adaptive weighted fusion model to calculate gas purification feature values; establishing a drift prediction model based on historical data, inputting environmental parameters and runtime to calculate the predicted drift amount, and correcting the gas purification feature values; when the drift amount exceeds a threshold or reaches the calibration period, introducing a standard gas to detect its average response value, calculating a calibration coefficient based on the standard gas concentration, and feeding it back to update the weighted fusion model and drift prediction model, thereby achieving self-calibration and high-precision analysis of the gas purification process.

Description

Technical Field

[0001] This application relates to the field of gas detection and analysis technology, specifically a gas analysis method and system for gas source purification processes. Background Technology

[0002] With the increasing reliance on high-purity gases in semiconductor manufacturing, precision chemicals, and the medical industry, gas purification systems have become a crucial link in ensuring process quality and safety. During gas purification, real-time monitoring of gas composition, purity, and impurity content is necessary to adjust purification parameters and replace adsorption materials promptly, ensuring the output gas reaches the target purity. Existing gas analysis devices typically use a single sensor (such as an infrared sensor, electrochemical sensor, or conductivity sensor) to monitor specific gas components. However, single sensors are susceptible to changes in temperature and humidity, flow fluctuations, and aging, leading to data drift. To ensure detection accuracy, existing systems often require frequent manual calibration or sensor replacement, resulting in high maintenance costs and insufficient continuous monitoring capabilities.

[0003] On the other hand, to reduce the calibration frequency, existing technologies have introduced compensation models with fixed parameters or dual-sensor comparisons. However, these methods generally rely on static models established under laboratory conditions, which are difficult to cope with the dynamic characteristics of environmental factors and gas composition changes during gas purification, resulting in limited compensation accuracy. Summary of the Invention

[0004] The purpose of this application is to provide a gas analysis method and system for gas source purification processes, so as to solve the problems mentioned in the background art.

[0005] In a first aspect, one embodiment of this application provides a gas analysis method for a gas source purification process. The method includes: simultaneously acquiring optical detection signals, electrochemical detection signals, and conductivity detection signals as gas detection signals in a gas source purification pipeline, and simultaneously acquiring temperature, humidity, and flow rate as environmental parameters; filtering, outlier removal, and normalization of the gas detection signals to construct a multidimensional feature vector containing the normalized values ​​of each detection signal and the environmental parameters; reducing the dimensionality of the multidimensional feature vector to obtain principal component features, and calculating gas purification feature values ​​through an adaptive weighted fusion model, wherein the adaptive weighted fusion model dynamically... The system assigns weight coefficients to each detection signal; it establishes a drift prediction model for each detection signal based on historical operating data, substitutes the current temperature, humidity, and operating time into the drift prediction model to calculate the predicted drift amount, and obtains the corrected gas purification characteristic value based on the gas purification characteristic value and the predicted drift amount; when the predicted drift amount exceeds the preset calibration trigger threshold or reaches the preset calibration cycle, it switches the gas path to introduce standard gas for detection and collects the average response value of the standard gas; it calculates the calibration coefficient based on the average response value and the known concentration of the standard gas, and feeds back the calibration coefficient to update the parameters of the adaptive weighted fusion model and the drift prediction model.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the multidimensional feature vector is reduced in dimensionality to obtain principal component features, and the gas purification feature value is calculated through an adaptive weighted fusion model. This includes: calculating the covariance matrix of the multidimensional feature vector and solving for the eigenvalues ​​and eigenvectors of the covariance matrix; sorting the eigenvalues ​​by size and extracting the top K principal components with a cumulative contribution rate not lower than a preset threshold to form the dimensionality-reduced principal component features; and inputting the principal component features into the adaptive weighted fusion model to obtain the gas purification feature value through weighted summation.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, an adaptive weighted fusion model is used to dynamically assign weight coefficients to each detection signal, including: calculating the corresponding signal-to-noise ratio (SNR) index based on the stability of each detection signal output; extracting the deviation statistics of each detection signal from the historical operating data and using the deviation statistics as historical deviation indicators; determining the influence coefficient of environmental parameters on various types of detection signals based on the degree of deviation of the current environmental parameters from the standard operating range; and assigning weight coefficients to each detection signal by comprehensively considering the SNR index, historical deviation indicators, and the influence coefficients of environmental parameters on various types of detection signals according to a predefined weight allocation function.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, a drift prediction model is established for each detection signal based on historical operating data, including: establishing a historical database for each detection signal, storing the output value of each detection signal, synchronously acquired environmental parameters, and reference values ​​during calibration in a time series; for detection signals exhibiting linear drift characteristics, a recursive least squares algorithm is used to establish a drift prediction model, representing the drift amount as a multiple linear regression function of time, temperature, and humidity, and updating the regression coefficients recursively; for detection signals exhibiting nonlinear drift characteristics, an adaptive Kalman filter algorithm is used to establish a drift prediction model, constructing a state-space model containing the true output value, drift error, and drift rate, and iteratively estimating the drift error through prediction and update steps; periodically performing a drift characteristic analysis task, extracting data from the historical database to identify linear drift rates, periodic fluctuation components, and abrupt change points for optimizing the drift prediction model.

[0009] In conjunction with the first aspect, in certain implementations of the first aspect, the current temperature, humidity, and runtime are substituted into the drift prediction model to calculate the predicted drift amount, and a corrected gas purification feature value is obtained based on the gas purification feature value and the predicted drift amount. This includes: inputting the current temperature, humidity, and runtime into the drift prediction model; calculating the predicted drift amount using a multiple linear regression function for a prediction model established using a recursive least squares algorithm; extracting the drift error component from the state estimate as the predicted drift amount for a prediction model established using an adaptive Kalman filter algorithm; weighting and summing the predicted drift amounts of each detection signal according to the weight coefficients of each detection signal in the adaptive weighted fusion model to obtain a comprehensive predicted drift amount; and performing drift compensation calculations on the gas purification feature value and the comprehensive predicted drift amount to obtain the corrected gas purification feature value.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, when the predicted drift exceeds a preset calibration trigger threshold or reaches a preset calibration period, the gas path is switched to introduce standard gas for detection, and the average response value of the standard gas is collected. This includes: determining whether the calibration trigger condition is met, whereby the calibration trigger condition includes time period triggering or drift anomaly triggering; when the calibration trigger condition is met, the normal detection channel is cut off and the reference gas channel is opened through the gas path switching device, allowing the standard gas to flow through each sensor; waiting for each detection signal to reach a stable response state, where the stability criterion is that the continuous change amplitude of the detection signal output value is less than a preset threshold; continuously collecting the output value of each detection signal in the standard gas environment, and statistically processing the collected data to obtain the average response value of each detection signal to the standard gas.

[0011] In conjunction with the first aspect, in certain implementations of the first aspect, calibration coefficients are calculated based on the average response value and the known concentration of the standard gas, and the calibration coefficients are fed back to update the parameters of the adaptive weighted fusion model and the drift prediction model. This includes: calculating calibration coefficients using a multiplicative calibration mode or an additive calibration mode based on the average response value of each detection signal and the known concentration value of the standard gas; feeding the calibration coefficients back to the adaptive weighted fusion model to update the baseline parameters of each detection signal; using the difference in the output of the detection signal before and after calibration as a new drift observation sample to update the regression coefficients or state estimates of the drift prediction model; and dynamically adjusting the next calibration cycle based on the change in the calibration coefficients, extending the calibration cycle when the change in the calibration coefficients is less than a preset stability threshold, and shortening the calibration cycle when the change in the calibration coefficients is greater than a preset change threshold.

[0012] In conjunction with the first aspect, in certain implementations of the first aspect, the gas detection signal is filtered, outlier removed, and normalized to construct a multidimensional feature vector containing the normalized values ​​of each detection signal and environmental parameters. This includes: filtering the gas detection signal using a moving average filter or median filter to obtain a filtered gas detection signal; identifying and removing outlier data points in the filtered gas detection signal that do not conform to the normal range based on statistical criteria; normalizing the gas detection signals and environmental parameters of different dimensions to a unified numerical range; and combining the normalized detection signal values ​​and environmental parameter values ​​to form a multidimensional feature vector.

[0013] Secondly, one embodiment of this application provides a gas analysis system for a gas source purification process, comprising: a sensing acquisition module, installed in the gas source purification pipeline, for simultaneously acquiring optical detection signals, electrochemical detection signals, and conductivity detection signals as gas detection signals, and simultaneously acquiring temperature, humidity, and flow rate as environmental parameters; a data processing module, connected to the sensing acquisition module, for filtering, outlier removal, and normalization of the gas detection signals, constructing a multidimensional feature vector containing the normalized values ​​of each detection signal and environmental parameters; reducing the dimensionality of the multidimensional feature vector to obtain principal component features, and calculating gas purification feature values ​​through an adaptive weighted fusion model, wherein the adaptive weighted fusion model dynamically assigns values ​​to each detection signal. The system includes a weighting coefficient and a drift correction module connected to the data processing module. The drift correction module is used to establish a drift prediction model for each detection signal based on historical operating data. It substitutes the current temperature, humidity, and operating time into the drift prediction model to calculate the predicted drift amount, and obtains the corrected gas purification characteristic value based on the gas purification characteristic value and the predicted drift amount. The calibration and control module, connected to the sensing acquisition module and the drift correction module, is used to switch the gas path to introduce standard gas for detection when the predicted drift amount exceeds the preset calibration trigger threshold or reaches the preset calibration cycle. It collects the average response value of the standard gas, calculates the calibration coefficient based on the average response value and the known concentration of the standard gas, and feeds back the calibration coefficient to update the parameters of the adaptive weighted fusion model and the drift prediction model.

[0014] In conjunction with the second aspect, in some implementations of the second aspect, the data processing module includes: a covariance calculation subunit, used to calculate the covariance matrix of the multidimensional feature vectors and solve for the eigenvalues ​​and eigenvectors of the covariance matrix; a principal component extraction subunit, connected to the covariance calculation subunit, used to sort the feature values ​​by size and extract the top K principal components with a cumulative contribution rate not lower than a preset threshold to form the dimensionality-reduced principal component features; and a weighted fusion subunit, connected to the principal component extraction subunit, used to input the principal component features into an adaptive weighted fusion model and obtain the gas purification feature values ​​through weighted summation.

[0015] Compared with the prior art, the beneficial effects of this application are:

[0016] 1. This application enables online analysis of gas composition and purity during gas purification through the synergy of sensor fusion and correction algorithms. By utilizing the complementary characteristics of different types of sensors and combining dynamic correction based on environmental parameters and historical drift characteristics, detection errors caused by temperature and humidity fluctuations and sensor aging are effectively suppressed, ensuring that the detection results remain stable and reliable over the long term.

[0017] 2. Through online calibration and adaptive model updates, the system can maintain stable accuracy during long-term operation, reduce the frequency of manual calibration, and improve the automation and intelligence level of the gas source purification process. Attached Figure Description

[0018] Figure 1 A schematic diagram of the structure of a gas analysis system for gas source purification process provided in an embodiment of this application;

[0019] Figure 2 A schematic diagram of a gas analysis system for a gas source purification process provided in another embodiment of this application;

[0020] Figure 3 A schematic flowchart of a gas analysis method for gas source purification provided in an embodiment of this application;

[0021] Figure 4 This is a schematic diagram of a process for reducing the dimensionality of a multidimensional feature vector to obtain principal component features and calculating gas purification feature values ​​through an adaptive weighted fusion model, as provided in an embodiment of this application.

[0022] Figure 5 This is a flowchart illustrating how an adaptive weighted fusion model dynamically assigns weight coefficients to each detection signal, as provided in an embodiment of this application.

[0023] Figure 6 This is a flowchart illustrating the process of establishing a drift prediction model for each detection signal based on historical operating data, as provided in an embodiment of this application.

[0024] Figure 7 This is a schematic diagram of a process provided in an embodiment of the present application, in which the current temperature, humidity and running time are substituted into the drift prediction model to calculate the predicted drift amount, and the corrected gas purification characteristic value is obtained based on the gas purification characteristic value and the predicted drift amount.

[0025] Figure 8 This is a schematic diagram of a process provided in an embodiment of the present application, which shows that when the predicted drift exceeds a preset calibration trigger threshold or reaches a preset calibration cycle, the gas path is switched to introduce standard gas for detection and the average response value of the standard gas is collected.

[0026] Figure 9 This is a schematic diagram illustrating the process of calculating calibration coefficients based on the average response value and the known concentration of the standard gas, and feeding back the calibration coefficients to update the parameters of the adaptive weighted fusion model and the drift prediction model, according to an embodiment of this application.

[0027] Figure 10 This is a flowchart illustrating the process of filtering, outlier removal, and normalization of gas detection signals to construct a multidimensional feature vector containing normalized values ​​of each detection signal and environmental parameters, according to an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Figure 1 This is a schematic diagram of a gas analysis system for a gas source purification process provided in one embodiment of this application. Figure 1 As shown, one embodiment of this application provides a gas analysis system for a gas source purification process. The system includes a sensing acquisition module 900, installed in the gas source purification pipeline, used to simultaneously acquire gas detection signals and environmental parameters based on two different detection principles. Each detection data point is marked with a unified timestamp. A data processing module 901, connected to the sensing acquisition module 900, is used to preprocess the gas detection signals to construct a multi-dimensional feature vector. A dimensionality reduction algorithm is used to process the multi-dimensional feature vector to obtain principal component features. An adaptive weighted fusion model is then used to calculate gas purification feature values. The adaptive weighted fusion model dynamically allocates weight coefficients based on the real-time performance indicators of each detection signal. The drift correction module 902, connected to the data processing module 901, is used to establish a drift prediction model for each detection signal based on historical operating data. It substitutes current environmental and operating parameters into the drift prediction model to calculate the predicted drift amount and corrects the gas purification characteristic values ​​based on the predicted drift amount. The calibration and control module 903, connected to the sensor acquisition module 900 and the drift correction module 902, is used to switch the gas path to introduce reference gas for detection and collect the average response value when the preset calibration trigger conditions are met. It calculates the calibration coefficient based on the detection response value and the known characteristics of the reference gas and feeds the calibration coefficient back to the data processing module 901 and the drift correction module 902. By combining multimodal signal acquisition, principal component dimensionality reduction, adaptive weighted fusion, and drift correction, high-precision, real-time analysis of the gas source purification process is achieved, effectively improving the stability and reliability of gas detection results.

[0030] Figure 2 This is a schematic diagram of a gas analysis system for a gas source purification process, provided as another embodiment of this application. (See diagram below.) Figure 2As shown, the data processing module of the system also includes a covariance calculation subunit 904, which receives multi-dimensional feature vector input from the data processing module 901, calculates the covariance matrix of the multi-dimensional feature vector, and solves for the eigenvalues ​​and eigenvectors of the covariance matrix; a principal component extraction subunit 905, connected to the covariance calculation subunit 904, is used to sort the features according to their size and extract the top few principal components whose cumulative contribution rate reaches a preset threshold, forming the dimensionality-reduced principal component features; and a weighted fusion subunit 906, connected to the principal component extraction subunit 905, is used to input the principal component features into an adaptive weighted fusion model and obtain the gas purification feature values ​​through weighted summation.

[0031] The following is combined Figures 3 to 10 The gas analysis method for gas source purification process provided in this application is described in detail.

[0032] Figure 3 This is a schematic flowchart of a gas analysis method for gas source purification provided in an embodiment of this application. Figure 3 As shown in the embodiments of this application, the gas analysis method for gas source purification process includes the following steps:

[0033] Step 100: Simultaneously collect optical detection signals, electrochemical detection signals, and conductivity detection signals as gas detection signals in the gas source purification pipeline, and simultaneously collect temperature, humidity, and flow rate as environmental parameters.

[0034] Step 101: Filter, remove outliers and normalize the gas detection signals to construct a multidimensional feature vector containing the normalized values ​​of each detection signal and environmental parameters.

[0035] It should be understood that combining multiple preprocessed detection signal values ​​and environmental parameter values ​​into an ordered array, namely a multidimensional feature vector, is used to comprehensively characterize the gas state at a certain moment.

[0036] Step 102: Dimensionality reduction of the multidimensional feature vector is performed to obtain principal component features, and gas purification feature values ​​are calculated through an adaptive weighted fusion model. Weight coefficients are dynamically assigned to each detection signal through the adaptive weighted fusion model.

[0037] Step 103: Establish a drift prediction model for each detection signal based on historical operating data, substitute the current temperature, humidity and running time into the drift prediction model to calculate the predicted drift amount, and obtain the corrected gas purification characteristic value based on the gas purification characteristic value and the predicted drift amount.

[0038] Step 104: When the predicted drift exceeds the preset calibration trigger threshold or reaches the preset calibration cycle, switch the gas path to introduce standard gas for detection and collect the average response value of the standard gas.

[0039] Step 105: Calculate the calibration coefficients based on the average response value and the known concentration of the standard gas, and feed the calibration coefficients back to update the parameters of the adaptive weighted fusion model and the drift prediction model.

[0040] It should be understood that principal component features refer to a few new, independent features extracted from the original multidimensional features using dimensionality reduction algorithms. An adaptive weighted fusion model is a mathematical model that dynamically adjusts the weights of each signal based on signal quality or importance, used to calculate a comprehensive evaluation value. Gas purification feature values ​​are numerical indicators used to comprehensively quantify and evaluate the current gas purity, calculated by fusing multiple detection signals. A drift prediction model is a mathematical model based on historical data used to predict the magnitude of deviation in sensor measurements over time or due to environmental factors. The predicted drift amount is the estimated deviation between the current sensor signal and the true value, calculated by the drift prediction model. The calibration trigger threshold is a preset critical value; when the predicted sensor drift amount exceeds this value, the system automatically initiates the calibration process. A standard gas is a gas with known precise concentrations, used as a reference to calibrate the measurement sensor. The calibration coefficient is a correction parameter calculated by comparing the sensor's response to the standard gas with a known concentration, used to compensate for sensor drift errors.

[0041] The gas analysis method for gas source purification provided in this application achieves long-term stable monitoring of gas purity during the gas source purification process through the synergy of multi-sensor fusion, adaptive weighting, drift prediction compensation, and automatic calibration. It effectively suppresses detection errors caused by environmental fluctuations and sensor aging, significantly reduces the frequency of manual calibration, and improves the system's automation level.

[0042] Figure 4 This is a schematic diagram illustrating a process for reducing the dimensionality of a multidimensional feature vector to obtain principal component features, and then calculating gas purification feature values ​​using an adaptive weighted fusion model, as provided in one embodiment of this application. Figure 4 As shown in the embodiments of this application, the gas analysis method for gas source purification process reduces the dimensionality of a multidimensional feature vector to obtain principal component features, and calculates gas purification feature values ​​through an adaptive weighted fusion model, including the following steps:

[0043] Step 200: Calculate the covariance matrix of the multidimensional eigenvectors, and solve for the eigenvalues ​​and eigenvectors of the covariance matrix.

[0044] Specifically, the sample mean of each feature dimension is calculated, followed by the calculation of each element of the covariance matrix. After the covariance matrix calculation is completed, eigenvalue decomposition is performed. This process simultaneously yields eigenvectors corresponding to each eigenvalue. These eigenvectors are column vectors with the same dimensions as the original eigenvectors, representing the direction of the principal components. The covariance calculation subunit then transmits all the solved eigenvalues, corresponding eigenvectors, and the covariance matrix to the principal component extraction subunit via an internal data interface, completing the calculation task for this stage.

[0045] Step 201: Sort the feature values ​​by size and extract the top K principal components whose cumulative contribution rate is not lower than a preset threshold to form the dimensionality-reduced principal component features.

[0046] Specifically, all eigenvalues ​​are sorted in descending order of numerical value to form an eigenvalue sequence, where the first element is the largest eigenvalue and the last element is the smallest eigenvalue. The variance contribution rate of each eigenvalue is calculated, which is the percentage of a single eigenvalue relative to the sum of all eigenvalues. Starting with the first eigenvalue, the variance contribution rates of each eigenvalue are accumulated sequentially to calculate the cumulative variance contribution rate. The cumulative variance contribution rate is equal to the sum of the variance contribution rates of the first few eigenvalues. A preset threshold for the cumulative variance contribution rate is set; in a preferred embodiment, the preset threshold is set to 95% to ensure that more than 95% of the effective information of the original data is retained after dimensionality reduction. The variance contribution rates are accumulated one by one according to the eigenvalue sorting order. When the cumulative variance contribution rate first reaches or exceeds the preset threshold, the number of eigenvalues ​​accumulated at this time, K, is recorded. The first K eigenvalues ​​and their corresponding K eigenvectors are extracted, and these K eigenvectors constitute the principal component transformation matrix. The original multidimensional eigenvectors are multiplied by the principal component transformation matrix to obtain the dimensionality-reduced principal component eigenvectors. The original multidimensional feature vectors have a dimension of 8 to 12, containing three gas detection signals, three environmental parameters and their derived features. After dimensionality reduction, the principal component feature vectors have a dimension of 3 to 5.

[0047] Furthermore, the value of K ranges from 3 to 6, preferably 4 or 5. The specific value of K is determined based on the inflection point of the cumulative variance contribution rate curve or a preset threshold, in order to achieve a balance between information retention rate and computational complexity.

[0048] Step 202: Input the principal component features into the adaptive weighted fusion model, and obtain the gas purification feature values ​​by weighted summation.

[0049] Each component of the principal component feature vector is multiplied by the weight coefficient of the corresponding detection signal, and then all weighted components are summed. In a specific embodiment, if the principal component feature vector contains three components, each corresponding to a principal component of the optical detection signal... Electrochemical detection signal principal component and the principal component of the conductivity detection signal The gas purification characteristic value is calculated by weighted summation. :

[0050] ;

[0051] Among them, the weighting coefficient , , The calculated gas purification feature values ​​are scalar values, 0.5, 0.3, and 0.2 respectively. These values ​​comprehensively reflect the current purity level of the gas; a higher value indicates higher purity or lower impurity content. The gas purification feature values ​​are output to the drift correction module as a monitoring indicator of the system. The weighted fusion subunit receives the dimensionality-reduced principal component feature vectors from the principal component extraction subunit and calls the adaptive weighted fusion model to perform the weighted fusion calculation task.

[0052] Figure 5 This is a flowchart illustrating how an adaptive weighted fusion model dynamically assigns weight coefficients to each detected signal, as provided in one embodiment of this application. Figure 5 As shown in the embodiments of this application, the gas analysis method for gas source purification process dynamically assigns weight coefficients to each detection signal through an adaptive weighted fusion model, including the following steps:

[0053] Step 300: Calculate the corresponding signal-to-noise ratio based on the stability of each detection signal output.

[0054] Specifically, the system extracts the signal-to-noise ratio (SNR) of the continuous output sequence of optical detection signals within the most recent evaluation time window from the data buffer. for:

[0055] ;

[0056] ;

[0057] in, For signal strength, For noise intensity, This represents the signal-to-noise ratio (SNR) in dB. In a typical operating scenario, an SNR of 30 to 50 dB for optical detection signals is considered normal, while a SNR below 20 dB indicates poor signal quality. The SNR of electrochemical and conductivity detection signals are calculated using the same method.

[0058] Step 301: Extract the deviation statistics of each detection signal from the standard reference value from the historical operation data, and use the deviation statistics as the historical deviation index.

[0059] Specifically, the system accesses the historical database and locates the calibration record table corresponding to each detection signal. The main fields of the calibration record table include calibration time, sensor identifier, standard reference concentration value, actual sensor output value, status flags before and after calibration, and environmental parameter snapshots. The query condition is set to extract the most recent 3 to 5 successfully completed calibration records. In a preferred embodiment, the most recent 5 calibration records are extracted, covering a time span of one to three months of operation. For each calibration record of the optical detection signal, the absolute deviation between the actual sensor output value and the standard reference value during that calibration is calculated. The absolute deviation equals the sensor output value minus the standard reference value; a positive value indicates that the sensor output is too high, and a negative value indicates that the sensor output is too low. To eliminate the influence of dimensions, the absolute deviation is divided by the standard reference value to obtain the relative deviation percentage. The relative deviation percentages of the extracted 5 calibration records form a deviation sample sequence, and the root mean square value of this sequence is calculated as the historical deviation index of the optical detection signal. Historical deviation indices are calculated separately for the electrochemical detection signal and the conductivity detection signal.

[0060] Step 302: Determine the influence coefficients of environmental parameters on various detection signals based on the degree of deviation between the current environmental parameters and the standard working range.

[0061] Specifically, the system acquires current environmental parameter values ​​in real time, including current temperature and humidity. Simultaneously, it reads preset standard temperature and humidity ranges from the system configuration parameter table. The absolute value of the deviation between the current temperature value and the center value of the standard temperature is calculated. The humidity deviation is calculated using the same method, with the absolute value of the humidity deviation divided by 10% of the standard humidity half-width.

[0062] Retrieve the temperature and humidity sensitivity coefficients for various detection signals from the sensor characteristic parameter database. For optical detection signals, calculate their environmental impact coefficients.

[0063] ;

[0064] ;

[0065] ;

[0066] in, This is the environmental impact factor. Temperature deviation, This is the temperature sensitivity coefficient. Humidity deviation, Humidity sensitivity coefficient This is the current temperature value. This is the center value of the standard temperature range. Half the width of the standard temperature range This is the current humidity value. This is the center value of the standard humidity range. This represents half the standard humidity range. The environmental impact coefficients for both the electrochemical and conductivity detection signals are calculated using the same formula.

[0067] Step 303: Based on the predefined weight allocation function, the influence coefficients of the signal-to-noise ratio index, historical deviation index, and environmental parameters on various detection signals are combined to assign weight coefficients to each detection signal.

[0068] Specifically, the signal-to-noise ratio (SNR) index is first forward normalized to extract the SNR values ​​of all detection signals involved in the fusion, thus determining the maximum and minimum values. For the SNR of the optical detection signals, a linear normalization formula is used for mapping:

[0069] ;

[0070] in, This represents the normalized signal-to-noise ratio. For the first The signal-to-noise ratio of each detected signal. The minimum signal-to-noise ratio among all detected signals. This represents the maximum signal-to-noise ratio (SNR) among all detected signals. The normalized result is in the range of 0 to 1; the closer the value is to 1, the better the SNR of the signal. Next, the historical deviation index is inversely normalized. Since a smaller historical deviation index value indicates better sensor performance, inverse mapping is necessary. The formula for calculating the inversely normalized historical deviation is:

[0071] ;

[0072] in, This represents the historical bias after reverse normalization. For the first Historical deviation of each detection signal The minimum historical deviation among all detected signals. This represents the maximum historical deviation among all detected signals. The converted value also falls within the range of 0 to 1, with the value closer to 1 indicating the smallest relative historical deviation for that signal. For example, the historical deviation for optical signals is 2%, for electrochemical signals it is 4%, and for conductivity signals it is 3%, with a minimum deviation of 2% and a maximum deviation of 4%. The inverse normalized historical deviation for optical signals is 1.0, for electrochemical signals it is 0.0, and for conductivity signals it is 0.5.

[0073] The environmental impact coefficients are reverse normalized. The calculation of the reverse normalized environmental impact coefficient is similar to that of the historical deviation index, using the method of subtracting the normalization ratio from 1. If the environmental impact coefficient of the optical signal is 0.064, that of the electrochemical signal is 0.15, and that of the conductivity signal is 0.22, after reverse normalization, the sensor with the highest environmental adaptability score obtains a value close to 1.

[0074] The weight coefficients of each inspection signal are calculated and immediately updated in the weight parameter table of the adaptive weighted fusion model. The weight update triggers a parameter refresh operation of the fusion model, and subsequent weighted summation calculations will be performed based on the latest weight coefficients.

[0075] Figure 6 This is a schematic diagram illustrating a process for establishing a drift prediction model for each detection signal based on historical operational data, as provided in one embodiment of this application. Figure 6 As shown in the embodiments of this application, the gas analysis method for gas source purification process establishes a drift prediction model for each detection signal based on historical operating data, including the following steps:

[0076] Step 400: Establish a historical database for each detection signal, storing the output value of each detection signal, synchronously acquired environmental parameters, and reference values ​​during calibration in a time series.

[0077] Specifically, an independent historical database table structure is established for each detection signal, and the database is implemented using a relational database management system. The data tables are designed with a time-series structure, with the timestamp as the primary key field, ensuring that data records are strictly arranged in chronological order, facilitating time-series querying and analysis.

[0078] The sensor output value field stores the raw output value of the detected signal at the corresponding moment. The ambient temperature field stores the synchronously acquired ambient temperature value. The ambient humidity field stores the relative humidity percentage value, ranging from 0 to 100%. The gas flow field stores the instantaneous flow rate value measured by the mass flow meter. A calibration flag field is set in the data table to identify whether the current record is data collected during the calibration process. When the calibration flag is true, the data record additionally includes a standard reference value field, which stores the known concentration value or known characteristic parameters of the standard gas used during calibration. The standard reference value serves as the benchmark for drift calculation, and its numerical accuracy and reliability directly affect the accuracy of the drift prediction model. The system ensures that the standard reference value comes from standard gases verified by national metrology institutions or authoritative third-party testing institutions, and its concentration uncertainty is controlled within 1%.

[0079] Step 401: For the detection signal exhibiting linear drift characteristics, a drift prediction model is established using the recursive least squares algorithm. The drift amount is expressed as a multiple linear regression function of time, temperature, and humidity, and the regression coefficients are updated recursively.

[0080] Specifically, the system first identifies the drift characteristics of each detected signal to determine whether its drift behavior conforms to linear or quasi-linear characteristics. For detected signals identified as having linear drift characteristics, the system establishes a multiple linear regression model to describe the drift behavior. The mathematical form of the multiple linear regression function is:

[0081] ;

[0082] in, This is the sensor drift amount. This is the time drift coefficient. This is a time variable, representing the cumulative runtime since the last calibration. For temperature sensitivity coefficient, Temperature deviation represents the deviation of the current ambient temperature from the standard calibration temperature. Humidity sensitivity coefficient Humidity deviation represents the deviation of the current ambient humidity from the standard calibrated humidity. This is a constant term. The system extracts operational data for 7 to 14 consecutive days after the initial calibration. In a preferred embodiment, 10 days of data are selected as the initial training set. The initial training set contains all sampling records within this time period. Each record provides a set of independent variable observations and one dependent variable observation. The dependent variable observation, i.e., the drift, is calculated by subtracting the baseline value determined during the initial calibration from the current sensor output value. The independent variable observations include the time interval since the initial calibration, temperature deviation, and humidity deviation. The system constructs the initial training set in the form of an observation matrix and an observation vector. Each row of the observation matrix corresponds to the independent variable value of one record, and each element of the observation vector corresponds to the dependent variable value of one record.

[0083] The initial regression coefficients are solved using batch least squares to minimize the sum of squared residuals between the model-predicted drift and the actual observed drift. The specific calculation is as follows:

[0084] ;

[0085] in, The initial regression coefficient vector contains , , , Four components; This is an observation matrix, where each row corresponds to the value of one record of the independent variable. This is an observation vector, where each element corresponds to a record of the dependent variable value. The information matrix is ​​used. After initialization, the system enters the recursive update phase. Each time a new calibration operation is completed, the system extracts the actual output value of the sensor before calibration, the standard reference value during calibration, the time variable at the calibration time, and the temperature and humidity deviations, forming a new observation sample. The dependent variable of the new observation sample is the difference between the output value before calibration and the standard reference value, representing the cumulative drift at the instant before calibration. The independent variables are the time, temperature, and humidity values ​​corresponding to the calibration time. The recursive update calculation process is based on a recursive formula, which expresses the regression coefficient at the current time as the sum of the regression coefficient and the correction term at the previous time. The correction term is determined by the gain matrix, the independent variable vector of the new observation sample, and the prediction residual. The prediction residual equals the actual dependent variable value of the new observation sample minus the dependent variable value predicted using the old regression coefficients, reflecting the model prediction error. The calculation of the gain matrix involves the recursive update of the inverse of the information matrix. The matrix inversion lemma is used to avoid direct matrix inversion operations, reducing computational complexity. The system calculates the gain matrix, prediction residuals, and regression coefficient update amount sequentially according to the recursive formula, and adds the update amount to the old regression coefficients to obtain the new regression coefficients.

[0086] Step 402: For the detection signal exhibiting nonlinear drift characteristics, an adaptive Kalman filter algorithm is used to establish a drift prediction model. A state-space model containing the true output value, drift error, and drift rate is constructed. The drift error is estimated iteratively through prediction and update steps.

[0087] Specifically, a Kalman filter algorithm is constructed based on a state-space model framework. The system defines a three-dimensional state vector. The first state variable is the sensor's true output value, representing the output the sensor should produce under ideal drift-free conditions. The second state variable is the drift error, representing the deviation of the sensor's actual output from the true output. The third state variable is the drift rate, representing the rate of change of the drift error over time. The state-space model consists of two parts: a state transition equation and an observation equation. The state transition equation describes the evolution of the state vector from the previous moment to the current moment. The true output value is assumed to be a slowly changing process, and its state transition relationship is expressed as the current true output value equals the previous true output value plus process noise. The process noise follows a Gaussian distribution with a mean of 0, and the standard deviation is set to be several times the sensor resolution. The observation equation describes the relationship between the sensor's actual measured value and the state vector. The sensor's actual measured value equals the true output value plus the drift error, plus the observation noise. The observation noise represents the random error in the measurement process, following a Gaussian distribution with a mean of 0, and the standard deviation is determined based on the sensor's repeatability index. The matrix form of the observation equation is:

[0088] ;

[0089] in, , for The actual measured value of the time sensor; The observation matrix is ​​a row vector in the form [1,1,0][1, 1, 0][1,1,0]. State vector = , This is the actual output value of the sensor. For drift error, For drift rate, To observe noise.

[0090] The update step uses the actual measurements at the current moment to correct the prior state estimate. The system first calculates the Kalman gain matrix, which quantifies the confidence weight of the measurement information relative to the model prediction. The formula for calculating the Kalman gain is:

[0091] ;

[0092] in, for The Kalman gain matrix at time t. The prior values ​​of the error covariance matrix are... For the observation matrix, To observe the noise variance.

[0093] The measurement residuals were then calculated:

[0094] ;

[0095] in, To measure the residual, These are prior state estimates. The measurement residuals are used to predict the measured values. They reflect the difference between actual observations and model predictions, and contain new information for revising the state estimates.

[0096] Step 403: Periodically perform drift characteristic analysis tasks to extract data from historical databases to identify linear drift rates, periodic fluctuation components, and abrupt change points, which are used to optimize the drift prediction model.

[0097] Specifically, the system sets the execution interval for the periodic drift characteristic analysis task to every hour to every four hours. When the task starts, it extracts data records for a specified time window from the historical database. The length of the time window is set according to the analysis objective. For linear drift rate identification, data records from the most recent one to two weeks are extracted. For periodic fluctuation component identification, data records from the most recent one to three months are extracted to capture possible weekly or monthly cycle patterns. For abrupt change point detection, high-density sampling data from the most recent three to seven days are extracted.

[0098] The system filters all records with true calibration flags from the historical database, extracting the calibration time, standard reference value, and sensor output value before calibration. It calculates the deviation between the sensor output value and the standard reference value for each calibration, constructing a deviation-time data pair sequence. A univariate linear regression analysis is performed on this sequence, expressing the deviation as a linear function of time, with the regression slope representing the average linear drift rate. The results of the drift characteristic analysis are integrated into the model optimization decision. When the analysis shows an accelerated linear drift rate for a sensor, the system shortens the calibration period for that sensor from thirty days to ten days, ensuring maintained detection accuracy even at faster drift rates. When the analysis identifies significant periodic fluctuations, the system adds a periodic compensation term to the drift prediction model. The compensation term parameters include period, amplitude, and initial phase, determined by fitting a periodic curve. When the analysis does not detect any abrupt changes and the linear drift rate is stable, the system evaluates the current model performance as good and maintains the existing model structure and parameter update strategy.

[0099] Figure 7 This is a schematic diagram illustrating a process in one embodiment of the present application, which calculates the predicted drift by substituting the current temperature, humidity, and runtime into a drift prediction model, and obtains a corrected gas purification characteristic value based on the gas purification characteristic value and the predicted drift. Figure 7 As shown in the embodiments of this application, the gas analysis method for gas source purification process includes the following steps: substituting the current temperature, humidity, and running time into the drift prediction model to calculate the predicted drift amount, and obtaining the corrected gas purification characteristic value based on the gas purification characteristic value and the predicted drift amount.

[0100] Step 500: Input the current temperature, humidity and running time into the drift prediction model. For the prediction model established by the recursive least squares algorithm, calculate the predicted drift amount through the multiple linear regression function. For the prediction model established by the adaptive Kalman filter algorithm, extract the drift error component from the state estimate as the predicted drift amount.

[0101] Specifically, the system receives real-time environmental parameters and operational status information. It collects ambient temperature values ​​and simultaneously reads the latest collected ambient humidity values. A recursive least squares algorithm is used to establish the detection signal for the prediction model. The system reads the currently valid regression coefficient vector of this signal from the drift prediction model parameter table. The regression coefficient vector contains four components: the first component is the time drift coefficient, representing the drift rate of the sensor output as operating time increases; the second component is the temperature sensitivity coefficient, representing the response strength of the sensor output to ambient temperature deviations; the third component is the humidity sensitivity coefficient, representing the response strength of the sensor output to ambient humidity deviations; and the fourth component is a constant term, representing the model's baseline offset. After the model has been sufficiently trained and calibrated, the value of the constant term is close to 0.

[0102] For a detection signal whose prediction model is built using an adaptive Kalman filter algorithm, the first component of the state vector is the estimate of the true output value, the second component is the estimate of the drift error, and the third component is the estimate of the drift rate. The system directly extracts the second component of the state vector, i.e., the drift error estimate, as the predicted drift amount of the detection signal. The drift error estimate is obtained by the Kalman filter algorithm through the fusion of historical model predictions and real-time measurement information, and its value reflects the cumulative drift amplitude of the sensor at the current moment. In an electrochemical detection signal, the state vector output by the Kalman filter has the following components: the first component 4.52 is the estimate of the true output value, the second component 0.36 is the extracted predicted drift amount, and the third component 0.008 is the estimate of the drift rate.

[0103] Step 501: Based on the weight coefficients of each detection signal in the adaptive weighted fusion model, the predicted drift of each detection signal is weighted and summed to obtain the comprehensive predicted drift. The gas purification feature value is then compared with the comprehensive predicted drift to perform drift compensation calculation, resulting in the corrected gas purification feature value.

[0104] Specifically, the system reads the currently valid weight coefficients for each detected signal from the weight parameter table of the adaptive weighted fusion model. The system performs a weighted summation operation to calculate the comprehensive predicted drift. For each detected signal, its predicted drift is multiplied by the corresponding weight coefficient to obtain the weighted drift contribution of that signal. The comprehensive predicted drift reflects the overall drift trend of multiple sensors under the current operating conditions, and the uncertainty of single-sensor drift prediction is reduced through weighted fusion.

[0105] The system receives gas purification feature values ​​from the adaptive weighted fusion model output by the data processing module. These feature values ​​are obtained through principal component analysis and weighted fusion calculations, comprehensively reflecting the current gas purity level or impurity content. The gas purification feature values ​​are scalar values, with dimensions consistent with the sensor output units. Their range is calibrated according to the application scenario, mapped to a standardized interval of 0 to 10 or a percentage interval of 0 to 100. In a specific example, the original gas purification feature value without drift compensation is 6.85, obtained based on the weighted fusion of principal component features from three detection signals. Drift compensation is performed to correct the gas purification feature value. Drift compensation uses a subtraction operation mode, subtracting the comprehensive predicted drift amount from the original gas purification feature value to obtain the corrected gas purification feature value. In the aforementioned example, the corrected gas purification feature value is equal to 6.407. The difference of 0.443 between the original and corrected feature values ​​is the drift compensation amount, which accounts for 6.5% of the original feature value, indicating that drift correction has a significant effect on improving detection accuracy. The corrected gas purification feature value is the final output of the system and is transmitted to multiple downstream modules.

[0106] Figure 8This is a schematic diagram illustrating a process in one embodiment of the present application where, when the predicted drift exceeds a preset calibration trigger threshold or reaches a preset calibration period, a standard gas is introduced for detection by switching the gas path, and the average response value of the standard gas is collected. Figure 8 As shown in the embodiments of this application, the gas analysis method for gas source purification process, when the predicted drift exceeds a preset calibration trigger threshold or reaches a preset calibration cycle, switches the gas path to introduce standard gas for detection and collects the average response value of the standard gas, including the following steps:

[0107] Step 600: Determine whether the calibration trigger conditions are met. Calibration trigger conditions include time period triggering or drift anomaly triggering.

[0108] Specifically, the calibration and control module continuously runs a calibration trigger condition monitoring task, which polls the system's operating status and drift prediction results at fixed time intervals. The polling interval is set to once every minute to once every five minutes.

[0109] The determination of drift anomaly trigger conditions is based on the predicted drift value output by the drift correction module. The system reads the latest calculated predicted drift value for each detection signal from the drift compensation data table and the corresponding calibration trigger threshold from the configuration parameter table. The calibration trigger threshold for optical detection signals is set to 5% to 10% of full scale. The system calculates the absolute value of the predicted drift value and compares it with the calibration trigger threshold. When the absolute value of the predicted drift value is greater than or equal to the calibration trigger threshold, the detection signal is determined to meet the drift anomaly trigger condition.

[0110] Step 601: When the calibration trigger condition is met, the normal detection channel is cut off and the reference gas channel is opened through the gas path switching device, so that the standard gas flows through each sensor.

[0111] Specifically, when the calibration and control module detects that any detection signal meets the calibration trigger condition, the system initiates the gas path switching preparation process. The system first sends a calibration status notification message to the data processing module. Upon receiving the notification, the data processing module switches the adaptive weighted fusion model to calibration mode. In this mode, the real-time output of the corrected gas purification characteristic values ​​is paused, maintaining the output of the last valid characteristic value before calibration initiation to avoid transient data interfering with the process control system's judgment logic during gas path switching. Simultaneously, the drift correction module pauses the parameter update operation of the drift prediction model, freezing the current model state until calibration is complete. The system sends valve control commands to the gas path switching device via the digital output interface. After valve switching is completed, the standard gas begins to flow through the sensors of the sensing and acquisition module. The gas flow path sequentially passes through the gas chamber of the optical sensor, the diffusion cavity of the electrochemical sensor, and the measurement electrode gap of the conductivity sensor.

[0112] Step 602: Wait for each detection signal to reach a stable response state. The stability criterion is that the continuous change amplitude of the detection signal output value is less than a preset threshold.

[0113] Specifically, after the gas path switching is completed and the standard gas begins to flow through each sensor, the system enters a stabilization waiting phase. The purpose of this phase is to ensure that the sensors are fully in contact with the standard gas, that the gas composition within the gas chamber or measurement cavity is completely replaced by the standard gas, that the physical or chemical response processes of the sensors reach equilibrium, and that the output signal stabilizes at the response level corresponding to the standard gas. The system sets up an independent stabilization monitoring task for each detection signal, with each task running in parallel, continuously acquiring sensor output values ​​and determining whether the stability criteria are met. The stability criteria are based on the continuous variation amplitude of the detection signal output value. The system sets a time window as the observation period for stability assessment, with a window length of 5 to 20 seconds. Within this time window, the system continuously acquires sensor output values ​​at a normal sampling frequency. If the sampling frequency is once per second, 10 data points are acquired within a 10-second window; if the sampling frequency is twice per second, 20 data points are acquired. The system calculates the range of the output value sequence within the observation window, which is the maximum value minus the minimum value. The range reflects the fluctuation amplitude of the signal within that time period; a smaller range indicates a more stable signal.

[0114] The system compares the calculated range within the observation window with a stability criterion threshold. When the range is less than the threshold, the system determines that the detected signal meets the stability requirements within the current observation window. The system continues to monitor the next observation window, using a sliding window mechanism. After each sampling cycle, the window slides forward one data point, deleting the oldest data point and adding the newest data point, recalculating the range, and re-evaluating stability. When N consecutive observation windows meet the stability requirements, the system finally determines that the detected signal has reached a stable response state. The number of consecutive windows N is set to 2 to 5, and in a preferred embodiment, it is set to 3, meaning that the range of three consecutive 10-second windows must be less than the threshold, corresponding to a stable output for 30 consecutive seconds. This ensures the reliability of the stability judgment while avoiding excessively long waiting times that could extend the calibration process.

[0115] Step 603: Continuously collect the output values ​​of each detection signal under a standard gas environment, and perform statistical processing on the collected data to obtain the average response value of each detection signal to the standard gas.

[0116] Specifically, once all calibrated or normal detection signals have reached a stable response state, the system initiates the data acquisition phase. The acquired data sequences are arranged in chronological order, with each data point containing a timestamp and an output value. The system performs preprocessing on the data sequences before statistical processing, identifying and removing potential outlier data points. The system calculates the arithmetic mean of the preprocessed valid sample sequences as the average response value of the detection signal to the standard gas.

[0117] Figure 9 This is a schematic diagram illustrating a process for calculating calibration coefficients based on the average response value and the known concentration of a standard gas, and feeding these calibration coefficients back to update the parameters of the adaptive weighted fusion model and the drift prediction model, according to an embodiment of this application. Figure 9 As shown in the embodiments of this application, the gas analysis method for gas source purification process calculates calibration coefficients based on the average response value and the known concentration of the standard gas, and feeds back the calibration coefficients to update the parameters of the adaptive weighted fusion model and the drift prediction model, including the following steps:

[0118] Step 700: Calculate the calibration coefficient using either multiplicative or additive calibration mode based on the average response value of each detection signal and the known concentration value of the standard gas.

[0119] Specifically, the system reads the average response value of each detection signal measured under a standard gas environment from the calibration data table, and simultaneously reads the known concentration value of the standard gas used in this calibration from the standard gas information table. The system determines the applicable calibration mode based on the physical measurement principle and output characteristics of each detection signal. Optical detection signals use a multiplicative calibration mode. Conductivity detection signals measure the conductivity of a gas; conductivity is proportional to the concentration of certain impurities, and its output characteristics are similar to those of optical sensors, making it suitable for a multiplicative calibration mode as well. Electrochemical detection signals are based on the current or voltage generated by the electrochemical reaction at the electrodes, and use an additive calibration mode or a dual-parameter calibration mode that includes both multiplicative and additive corrections.

[0120] Step 701: Feed the calibration coefficients back to the adaptive weighted fusion model to update the baseline parameters of each detection signal.

[0121] Specifically, the system transmits the calibration coefficients calculated from each detection signal to the adaptive weighted fusion model module via a data interface. For detection signals using multiplicative calibration mode, the system writes the newly calculated multiplicative calibration coefficients into the calibration coefficient field of the reference parameter table, replacing the old calibration coefficients of the sensor. The update of the multiplicative calibration coefficients directly affects the subsequent concentration conversion calculation. In the data preprocessing stage of the fusion algorithm, the system reads the sensor's original output value, multiplies it by the currently valid multiplicative calibration coefficient, and obtains the corrected concentration value. For detection signals using additive calibration mode, the system writes the newly calculated additive calibration coefficients into the zero-point offset field of the reference parameter table. In the data preprocessing stage of the fusion algorithm, the system reads the sensor's original output value, adds the currently valid additive calibration coefficient, and obtains the corrected concentration value. For detection signals using dual-parameter calibration mode, the system updates both the multiplicative coefficient and additive coefficient fields simultaneously. In the data preprocessing stage, the system first multiplies the original output value by the multiplicative coefficient, then adds the additive coefficient, completing two levels of correction. The correction formula is:

[0122] ;

[0123] in, This is the corrected concentration value. The original output value of the sensor. For multiplicative calibration coefficients, These are additive calibration coefficients. Dual-parameter calibration simultaneously corrects for both the sensor's gain error and zero-point error, providing higher accuracy across the entire measurement range.

[0124] Step 702: Use the difference in the output of the detection signal before and after calibration as a new drift observation sample to update the regression coefficients or state estimates of the drift prediction model.

[0125] Specifically, before the calibration process begins, the system records the pre-calibration output values ​​of each detection signal. The pre-calibration output value is defined as the sensor's output value within the last effective sampling period before the gas path switches to the reference gas channel. The system extracts the sensor output value sequence from the data buffer within one minute prior to the calibration start time and calculates the arithmetic mean of this sequence as a stable pre-calibration output value, avoiding the introduction of random errors using a single instantaneous value. The system calculates the difference in the detection signal output before and after calibration, defined as the pre-calibration output value minus the calibration reference value multiplied by the effective calibration coefficient from the previous calibration. The system uses the calculated output difference as a new drift observation sample, along with relevant input variables. For the drift prediction model established using the recursive least squares algorithm, the system substitutes the new drift observation sample into the recursive update formula to update the regression coefficients. For the drift prediction model established using the adaptive Kalman filter algorithm, the system uses the new drift observation sample to update the state estimate and error covariance matrix.

[0126] Step 703: Dynamically adjust the next calibration cycle based on the change in the calibration coefficient. When the change in the calibration coefficient is less than the preset stable threshold, extend the calibration cycle; when the change in the calibration coefficient is greater than the preset change threshold, shorten the calibration cycle.

[0127] Specifically, the system extracts the calibration coefficient values ​​for the current and previous calibrations from the calibration coefficient history table and calculates the change in the calibration coefficients. The system reads the preset stability threshold and preset variation threshold for the detection signal from the configuration parameter table. The system compares the calculated relative change rate of the calibration coefficients with the preset stability threshold and preset variation threshold. When the relative change rate is less than the preset stability threshold, the system determines that the sensor's drift characteristics are stable, the drift rate is slow, and the current calibration cycle setting is relatively conservative, which can be appropriately extended to reduce calibration frequency and maintenance costs. When the relative change rate is greater than the preset variation threshold, the system determines that the sensor's drift characteristics are unstable, the drift rate is accelerated or abnormal changes occur, the current calibration cycle setting is too long, and it needs to be shortened to increase the calibration frequency and ensure detection accuracy. The system calculates a calibration cycle shortening coefficient, which is set to 0.5 to 0.8, preferably 0.7. When the relative change rate is between the stability threshold and the variation threshold, the system determines that the sensor's drift characteristics are normal, the current calibration cycle setting is reasonable, and the system maintains the calibration cycle unchanged.

[0128] Figure 10 This is a flowchart illustrating a process for filtering, outlier removal, and normalization of gas detection signals, constructing a multidimensional feature vector containing normalized values ​​of each detection signal and environmental parameters, according to an embodiment of this application. Figure 10 As shown in the embodiments of this application, the gas analysis method for gas source purification process filters, removes outliers, and normalizes the gas detection signal to construct a multi-dimensional feature vector containing the normalized values ​​of each detection signal and environmental parameters, including the following steps:

[0129] Step 800: The gas detection signal is filtered using a moving average filter or median filter to obtain the filtered gas detection signal.

[0130] Specifically, the optical detection signal is filtered using a moving average method. Moving average filtering is applied to the optical detection signal. Median filtering is applied to the electrochemical and conductivity detection signals.

[0131] Step 801: Based on statistical criteria, identify and remove abnormal data points in the filtered gas detection signal that do not conform to the normal range.

[0132] Specifically, the system performs outlier detection and removal on the filtered gas detection signal. Outliers are defined as data points that significantly deviate from the normal signal distribution range. The system employs a statistically based outlier detection method, implementing 3D filtering on the optical detection signal. Outlier detection criteria.

[0133] Step 802: Normalize the gas detection signals and environmental parameters of different dimensions to a unified numerical range.

[0134] Specifically, the system uses a minimum-maximum normalization method to map each parameter to a standard range of 0 to 1.

[0135] Step 803: Combine the normalized detection signal values ​​and environmental parameter values ​​to form a multidimensional feature vector.

[0136] Specifically, the system reads the normalized values ​​of each detection signal and environmental parameter from the normalized data buffer and combines them in a predefined order to form a multi-dimensional feature vector. The dimension of the feature vector is equal to the sum of the number of detection signals and the number of environmental parameters involved in the fusion. The arrangement order of the feature vectors is as follows: first component - normalized value of optical detection signal; second component - normalized value of electrochemical detection signal; third component - normalized value of conductivity detection signal; fourth component - normalized value of ambient temperature; fifth component - normalized value of ambient humidity; and sixth component - normalized value of gas flow rate. This arrangement groups similar parameters, with detection signals occupying the first three components and environmental parameters occupying the last three components, facilitating the differentiation between sensor measurement information and environmental state information during processing.

[0137] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method of gas analysis for a gas source purification process, characterized by, The method comprises the following steps: Synchronously collecting optical detection signals, electrochemical detection signals and conductivity detection signals in a gas source purification pipeline as gas detection signals, and simultaneously collecting temperature, humidity and flow as environmental parameters; Filtering, outlier rejection and normalization processing are performed on the gas detection signals to construct a multi-dimensional feature vector containing the normalized values of each detection signal and the environmental parameters; Dimensionality reduction is performed on the multi-dimensional feature vector to obtain principal component features, and a gas purification characteristic value is calculated through an adaptive weighted fusion model, wherein the adaptive weighted fusion model dynamically assigns weight coefficients to each detection signal; A drift prediction model is established for each detection signal based on historical operation data, the temperature, humidity and operation time at the current time are substituted into the drift prediction model to calculate a predicted drift amount, and a corrected gas purification characteristic value is obtained based on the gas purification characteristic value and the predicted drift amount; When the predicted drift amount exceeds a preset calibration trigger threshold or reaches a preset calibration period, the gas circuit is switched to introduce standard gas for detection, and the average response value of the standard gas is collected; A calibration coefficient is calculated according to the average response value and the known concentration of the standard gas, and the calibration coefficient is fed back and updated to the parameters of the adaptive weighted fusion model and the drift prediction model.

2. A method for gas analysis for a gas source purification process according to claim 1, characterized in that, The dimensionality reduction of the multi-dimensional feature vector to obtain principal component features and the calculation of the gas purification characteristic value through the adaptive weighted fusion model comprise: The covariance matrix of the multi-dimensional feature vector is calculated, and the eigenvalues and eigenvectors of the covariance matrix are solved; The eigenvalues are sorted by size, the first K principal components with a cumulative contribution rate not less than a preset threshold are extracted to form the principal component features after dimensionality reduction; The principal component features are input into the adaptive weighted fusion model to obtain the gas purification characteristic value through weighted summation.

3. A method of gas analysis for a gas source purification process according to claim 2, wherein, The adaptive weighted fusion model dynamically assigns weight coefficients to each detection signal, which comprises: According to the stability of the output of each detection signal, the signal-to-noise ratio index is calculated; The deviation statistics of each detection signal and the standard reference value are extracted from the historical operation data, and the deviation statistics are used as historical deviation indexes; According to the deviation degree of the current environmental parameters from the standard working range, the influence coefficients of the environmental parameters on each type of detection signal are determined; According to a predefined weight allocation function, the signal-to-noise ratio index, the historical deviation index and the influence coefficients of the environmental parameters on each type of detection signal are integrated to assign weight coefficients to each detection signal.

4. A method for gas analysis for a gas source purification process according to any one of claims 1 to 3, characterized in that, The drift prediction model is established for each detection signal based on historical operation data, which comprises: A historical database is established for each detection signal, and the output values of each detection signal, the synchronously collected environmental parameters and the reference values at the calibration time are stored in time sequence; For detection signals exhibiting linear drift characteristics, a recursive least squares algorithm is used to establish a drift prediction model, the drift amount is represented as a multiple linear regression function of time, temperature and humidity, and the regression coefficients are updated in a recursive manner; For the detection signal presenting nonlinear drift characteristics, an adaptive Kalman filter algorithm is used to establish a drift prediction model, a state space model containing a true output value, a drift error and a drift rate is constructed, and the drift error is iteratively estimated through a prediction step and an update step; Periodically performing a drift characteristic analysis task, extracting data from the historical database to identify a linear drift rate, a periodic fluctuation component and a mutation point, for optimizing the drift prediction model.

5. A method for gas analysis for a gas source purification process according to any one of claims 1 to 3, characterized in that, The current temperature, humidity and running time are substituted into the drift prediction model to calculate a predicted drift amount, and based on the gas purification characteristic value and the predicted drift amount, a corrected gas purification characteristic value is obtained, including: The current temperature, humidity and running time are input into the drift prediction model, and for the prediction model established using a recursive least squares algorithm, the predicted drift amount is calculated by a multiple linear regression function, and for the prediction model established using an adaptive Kalman filter algorithm, the drift error component is extracted from the state estimate value as the predicted drift amount; According to the weight coefficients of the detection signals in the adaptive weighted fusion model, the predicted drift amounts of the detection signals are weighted and summed to obtain a comprehensive predicted drift amount, and the gas purification characteristic value and the comprehensive predicted drift amount are subjected to drift compensation operation to obtain the corrected gas purification characteristic value.

6. A method for gas analysis for a gas source purification process according to any one of claims 1 to 3, characterized in that, When the predicted drift amount exceeds a preset calibration trigger threshold or reaches a preset calibration period, the gas path is switched to introduce standard gas for detection, and the average response value of the standard gas is collected, including: Determine whether the calibration trigger condition is met, the calibration trigger condition includes a time period trigger or a drift anomaly trigger; When it is detected that the calibration trigger condition is met, the normal detection channel is cut off and the reference gas channel is opened by the gas path switching device, so that the standard gas flows through each sensor; Wait for the detection signals to reach a stable response state, and the stability criterion is that the continuous change amplitude of the detection signal output value is less than a preset threshold; Collect the output values of each detection signal in the standard gas environment continuously, and obtain the average response value of each detection signal to the standard gas by statistical processing of the collected data.

7. A method for gas analysis for a gas source purification process according to any one of claims 1 to 3, characterized in that, The calibration coefficient is calculated according to the average response value and the known concentration of the standard gas, and the calibration coefficient is fed back to update the parameters of the adaptive weighted fusion model and the drift prediction model, including: According to the average response value of each detection signal and the known concentration of the standard gas, the calibration coefficient is calculated using a multiplicative calibration mode or an additive calibration mode; The calibration coefficient is fed back to the adaptive weighted fusion model to update the reference parameters of the detection signals; The output difference of the detection signals before and after calibration is used as a new drift observation sample to update the regression coefficients or state estimates of the drift prediction model; According to the change amount of the calibration coefficient, the next calibration period is dynamically adjusted, when the change amount of the calibration coefficient is less than a preset stability threshold, the calibration period is lengthened, and when the change amount of the calibration coefficient is greater than a preset change threshold, the calibration period is shortened.

8. A method for gas analysis for a gas source purification process according to any one of claims 1 to 3, characterized in that, The gas detection signal is filtered, outlier removed and normalized, a multi-dimensional feature vector containing each detection signal normalized value and environmental parameter is constructed, including: The gas detection signal is filtered by using a sliding average filter or a median filter method to obtain a filtered gas detection signal; Based on statistical criteria, abnormal data points in the filtered gas detection signal that do not conform to the normal range are identified and removed; The gas detection signal and the environmental parameter of different dimensions are normalized to a unified numerical interval; The normalized detection signal value and the environmental parameter value are combined to form a multi-dimensional feature vector.

9. A gas analysis system for a gas source purification process, characterized by, It includes: The sensing acquisition module is arranged in the gas source purification pipeline and is used for synchronously collecting optical detection signals, electrochemical detection signals and conductivity detection signals as gas detection signals, and collecting temperature, humidity and flow as environmental parameters; The data processing module is connected with the sensing acquisition module and is used for filtering, outlier removing and normalizing the gas detection signal, and constructing a multi-dimensional feature vector containing each detection signal normalized value and environmental parameter; The multi-dimensional feature vector is reduced to obtain principal component features, and a gas purification feature value is calculated through an adaptive weighted fusion model, wherein the adaptive weighted fusion model dynamically assigns weight coefficients to the detection signals; The drift correction module is connected with the data processing module and is used for establishing a drift prediction model for each detection signal based on historical operation data, substituting temperature, humidity and operation time at the current moment into the drift prediction model to calculate a predicted drift amount, and obtaining a corrected gas purification feature value based on the gas purification feature value and the predicted drift amount; The calibration and control module is connected with the sensing acquisition module and the drift correction module, and is used for switching the gas circuit to introduce standard gas for detection when the predicted drift amount exceeds a preset calibration trigger threshold or reaches a preset calibration period, collecting the average response value of the standard gas, calculating a calibration coefficient according to the average response value and the known concentration of the standard gas, and feeding back the calibration coefficient to update the parameters of the adaptive weighted fusion model and the drift prediction model.

10. A gas analysis system for a gas source purification process according to claim 9, wherein, The data processing module includes: The covariance calculation subunit is used for calculating the covariance matrix of the multi-dimensional feature vector and solving the eigenvalue and eigenvector of the covariance matrix; The principal component extraction subunit is connected with the covariance calculation subunit and is used for sorting the eigenvalues, extracting the first K principal components with a cumulative contribution rate not lower than a preset threshold, and constituting the reduced principal component features; The weighted fusion subunit is connected with the principal component extraction subunit and is used for inputting the principal component features into the adaptive weighted fusion model to obtain the gas purification feature value through weighted summation.

Images