Method and system for detecting residual gas pollutants based on multi-source data

Abstract

The application relates to the technical field of gas detection and analysis, in particular to a gas pollutant residual detection method and system based on multi-source data. The method comprises the following steps: acquiring multi-source sensor data in a target limited space, including an environmental humidity value and an original response value sequence; decomposing and reconstructing the original response value sequence through multi-scale wavelet transformation to obtain an enhanced original response value sequence; performing humidity interference compensation to obtain a humidity-corrected response value sequence; performing adaptive cross-sensitivity decoupling processing on the humidity-corrected response value sequence; taking the latest value in a residual concentration contribution value sequence as a detection result at a current moment; calculating a variation coefficient of the residual concentration contribution value sequence; calculating the confidence of the residual concentration contribution value of each target gas pollutant according to the variation coefficient; and outputting the detection result and the confidence. The method effectively improves the detection accuracy and reliability of residual concentrations of multiple gas pollutants in a limited space.

Description

Technical Field

[0001] This invention relates to the field of gas detection and analysis technology, specifically to a method and system for detecting residual gaseous pollutants based on multi-source data. Background Technology

[0002] During maintenance work in confined spaces at wastewater treatment plants, accurate detection of residual gaseous pollutants is crucial for assessing operational safety risks. Because these confined spaces are typically in a closed or semi-closed state, various toxic and harmful gases, such as hydrogen sulfide, methane, and ammonia, produced by the decomposition of organic matter, can easily accumulate and remain within them. If workers fail to accurately determine the residual concentrations of these gaseous pollutants before entering, it can easily lead to poisoning, explosions, and other safety accidents. Therefore, accurate and reliable detection of residual gaseous pollutant concentrations in confined spaces is a critical prerequisite for ensuring operational safety.

[0003] However, the actual monitoring environment in confined spaces of wastewater treatment plants is extremely complex, posing numerous challenges to the accurate detection of gaseous pollutants. These environments are consistently characterized by high humidity, with high water vapor content that fluctuates significantly with changing operating conditions. Humidity variations cause systematic deviations in sensor output values, making it difficult to extract the true gas concentration information from the raw response signal. Furthermore, multiple gaseous pollutants often coexist within the confined spaces of wastewater treatment plants, resulting in aliased sensor output signals representing the combined effects of various gas responses. Therefore, a method is needed to accurately and reliably detect the residual concentrations of multiple gaseous pollutants in confined spaces under complex environments with multiple sources of interference. Summary of the Invention

[0004] (1) Technical problems to be solved The purpose of this invention is to provide a method and system for detecting residual gaseous pollutants based on multi-source data, which solves the problem that the concentration of residual gaseous pollutants cannot be accurately and reliably detected by sensors due to cross-sensitivity in confined spaces during maintenance work in wastewater treatment plants, where environmental humidity interference and signal noise superposition occur.

[0005] (2) Technical solution To achieve the above objectives, in one aspect, the present invention provides a method for detecting residual gaseous pollutants based on multi-source data, the method comprising: S1. Acquire multi-source sensor data within the target confined space. The multi-source sensor data includes the ambient humidity value collected in real time by a humidity sensor and the raw response value sequence collected in real time by a gas sensor array. The gas sensor array contains multiple gas-sensitive units that have cross-sensitivity to gaseous pollutants.

[0006] S2. The original response value sequence is decomposed into multiple wavelet coefficient components of different frequency scales through multi-scale wavelet transform. The scale components that match the response time constant of typical gaseous pollutants are identified and retained. The retained scale components are reconstructed by wavelet transform to obtain the enhanced original response value sequence.

[0007] S3. Perform humidity interference compensation on the enhanced original response value sequence based on the ambient humidity value to obtain the humidity-corrected response value sequence.

[0008] S4. The humidity-corrected response value sequence is subjected to adaptive cross-sensitivity decoupling processing to decompose it into multiple independent source signal time components, and the residual concentration contribution value sequence corresponding to each target gaseous pollutant is extracted from the source signal time components.

[0009] S5. Take the latest value in the residual concentration contribution value sequence as the detection result at the current moment; calculate the coefficient of variation of the residual concentration contribution value sequence within the current sliding time window, calculate the confidence level of the residual concentration contribution value of each target gaseous pollutant based on the coefficient of variation, and output the detection result and confidence level.

[0010] Furthermore, the method for identifying and retaining scale components that match the response time constant of typical gaseous pollutants includes: After calculating the actual response time constant of each wavelet coefficient component, the waveform asymmetry coefficient of each wavelet coefficient component is also calculated. The typical gaseous pollutant response time constant of each target gaseous pollutant and the typical waveform asymmetry coefficient of each target gaseous pollutant are combined to form a two-dimensional feature template.

[0011] The two-dimensional measured features, consisting of the actual response time constant and waveform asymmetry coefficient of each wavelet coefficient component, are compared with the two-dimensional feature template of each target gaseous pollutant to calculate the Mahalanobis distance. When the Mahalanobis distance is less than the preset Mahalanobis distance threshold, the wavelet coefficient component is determined to be a scale component that matches the corresponding target gaseous pollutant and is retained.

[0012] Furthermore, before performing Mahalanobis distance calculation, the ratio of the total energy of each wavelet coefficient component to the sum of the total energy of all wavelet coefficient components is calculated to obtain the energy proportion of each wavelet coefficient component; when the energy proportion of a wavelet coefficient component is lower than the preset energy proportion threshold, the wavelet coefficient component is directly identified as a noise component and removed.

[0013] Further, the method for compensating for humidity interference in the enhanced original response value sequence based on the ambient humidity value to obtain a humidity-corrected response value sequence includes: Obtain the ambient humidity value at the sampling time corresponding to each response value in the enhanced original response value sequence, and obtain the historical ambient humidity values ​​of N consecutive sampling times before the current sampling time, arranged in chronological order to form a historical humidity memory vector; multiply each historical ambient humidity value in the historical humidity memory vector by the response value of each gas-sensitive unit in the enhanced original response value sequence at the current sampling time to generate a set of humidity-response cross-feature terms; concatenate the humidity-response cross-feature terms with the historical humidity memory vector, the ambient humidity value at the current sampling time, and the response values ​​of each gas-sensitive unit in the enhanced original response value sequence at the current sampling time to form an expanded input feature vector.

[0014] The extended input feature vector is input into the humidity interference compensation model. After forward propagation calculation by the feedforward neural network, the humidity-compensated response value corresponding to each gas-sensitive unit is obtained from the output layer node. The humidity-compensated response values ​​of all gas-sensitive units are combined in the same order as the enhanced original response value sequence to obtain the humidity-corrected response value sequence.

[0015] Furthermore, a feedforward neural network is constructed as a humidity interference compensation model, wherein the number of input layer nodes of the feedforward neural network is set to be equal to the dimension of the expanded input feature vector.

[0016] The number of hidden layer nodes in the feedforward neural network is set to twice the number of input layer nodes plus one, and a nonlinear activation function is used to perform nonlinear mapping on the expanded input feature vector. The number of output layer nodes in the feedforward neural network is equal to the number of gas-sensitive units in the gas sensor array. Each output layer node corresponds to the response value of a gas-sensitive unit after humidity interference compensation. The loss function used by the humidity interference compensation model during the training phase is the weighted mean square error loss function.

[0017] Furthermore, the method for performing adaptive cross-sensitivity decoupling processing on the humidity-corrected response value sequence to decompose it into multiple independent source signal time components includes: The humidity-corrected response value sequence is organized into an observation matrix, where the rows of the observation matrix correspond to each gas-sensitive unit in the gas sensor array, and the columns of the observation matrix correspond to the humidity-corrected response value at each sampling time.

[0018] The observation matrix is ​​subjected to blind source separation using the independent component analysis algorithm. An unmixing matrix is ​​obtained by iterative optimization. The source signal matrix is ​​obtained by multiplying the unmixing matrix with the observation matrix. Each row of the source signal matrix represents an independent source signal time component.

[0019] During the iterative optimization process, the independent component analysis algorithm adaptively adjusts the update direction of the unmixing matrix according to the actual data distribution of the humidity-corrected response value sequence until the mutual information between the time components of each source signal in the source signal matrix converges to the minimum value, thus completing the adaptive cross-sensitive decoupling process and obtaining multiple independent source signal time components.

[0020] Furthermore, the method of performing blind source separation processing on the observation matrix using the independent component analysis algorithm and iteratively optimizing to solve for an unmixing matrix includes: A composite optimization objective is formed by adding the statistical independence measure among the response values ​​of each gas-sensitive unit in the observation matrix with a sparsity penalty term. The sparsity penalty term is the sum of the L1 norms of all source signal time components in the source signal matrix obtained by multiplying the current unmixed matrix with the observation matrix.

[0021] A sparsity regularization intensity coefficient is set, which is determined based on the maximum number of target gaseous pollutant types appearing simultaneously in historical monitoring data before maintenance operations in the confined space of the wastewater treatment plant; in each iteration of iterative optimization to solve the unmixing matrix, the natural gradient of the composite optimization objective relative to the current unmixing matrix is ​​calculated, and the unmixing matrix is ​​updated according to the direction of the natural gradient.

[0022] Furthermore, the method for extracting the residual concentration contribution value sequence corresponding to each target gaseous pollutant from the time component of the source signal includes: A reference response mode vector is pre-constructed for each target gaseous pollutant. The reference response mode vector is obtained by introducing a single target gaseous pollutant of known concentration into the gas sensor array in a laboratory environment and recording the steady-state response value of each gas-sensitive unit in the gas sensor array. The dimension of the reference response mode vector is equal to the number of gas-sensitive units in the gas sensor array.

[0023] Each source signal time component is considered as a source signal vector to be matched, and the similarity between each source signal vector to be matched and the reference response mode vector of each target gaseous pollutant is calculated.

[0024] Select the source signal vector with the highest similarity to the reference response mode vector of the target gaseous pollutant from all the source signal vectors to be matched as the initial source signal component of the target gaseous pollutant.

[0025] Using the initial source signal component as the independent variable and the actual humidity-corrected response value of each gas sensing unit in the gas sensor array as the dependent variable, the contribution coefficient of each initial source signal component to the response value of each gas sensing unit is solved by the least squares method. The amplitude of each sampling moment in the initial source signal component is multiplied by the corresponding contribution coefficient to obtain the residual concentration contribution value of the target gaseous pollutant. The residual concentration contribution values ​​of all sampling moments are arranged in chronological order to obtain the residual concentration contribution value sequence.

[0026] Furthermore, the step of performing online adaptive updates to the reference response mode vector includes: At each sampling moment, the fitting residual vector between the initially selected source signal component and the actual humidity-corrected response value of each gas sensing unit in the gas sensor array is calculated, and the root mean square error of the fitting residual vector is calculated within a sliding time window; when the root mean square error is continuously less than a preset stability threshold for a preset trigger duration, the actual humidity-corrected response values ​​of each gas sensing unit in the gas sensor array at the current sampling moment are arranged into a measured response mode vector according to the order of each gas sensing unit; the vector obtained by weighted averaging the measured response mode vector and the currently stored reference response mode vector replaces the original reference response mode vector.

[0027] On the other hand, based on the same inventive concept, this invention also provides a gaseous pollutant residue detection system based on multi-source data, the system comprising: A multi-source sensor data acquisition module is used to acquire multi-source sensor data within a limited space of the target. The multi-source sensor data includes the ambient humidity value acquired in real time by a humidity sensor and the raw response value sequence acquired in real time by a gas sensor array. The gas sensor array contains multiple gas-sensitive units that have cross-sensitivity to gaseous pollutants.

[0028] The signal enhancement module is used to decompose the original response value sequence into multiple wavelet coefficient components of different frequency scales through multi-scale wavelet transform, identify and retain the scale components that match the response time constant of typical gaseous pollutants, and perform wavelet reconstruction on the retained scale components to obtain the enhanced original response value sequence.

[0029] The humidity interference compensation module is used to perform humidity interference compensation on the enhanced original response value sequence based on the ambient humidity value, so as to obtain a humidity-corrected response value sequence.

[0030] The cross-sensitive adaptive decoupling module is used to perform adaptive cross-sensitive decoupling processing on the humidity-corrected response value sequence, decompose it into multiple independent source signal time components, and extract the residual concentration contribution value sequence corresponding to each target gaseous pollutant from the source signal time components.

[0031] The detection result output module is used to take the latest value in the residual concentration contribution value sequence as the detection result at the current moment; calculate the coefficient of variation of the residual concentration contribution value sequence within the current sliding time window; calculate the confidence level of the residual concentration contribution value of each target gaseous pollutant based on the coefficient of variation; and output the detection result and the confidence level.

[0032] (3) Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: 1. Noise reduction and feature enhancement of sensor signals are achieved through multi-scale wavelet transform. Combined with a neural network compensation model based on historical humidity memory vector, dynamic interference from environmental humidity is eliminated. Then, adaptive independent component analysis is used to decouple the signals of the multi-sensor array through cross-sensitivity. This enables the residual concentration of each target gaseous pollutant to be accurately separated and quantified in a limited space environment where noise, humidity interference and cross-sensitivity coexist. This solves the problem of detection result distortion caused by multi-source interference coupling.

[0033] 2. By calculating the coefficient of variation of the residual concentration contribution value sequence in real time within the sliding time window, a quantitative confidence assessment is added to each detection result. At the same time, the reference response mode vector is updated online adaptively using the fitting residual triggering mechanism. This enables the detection system to automatically correct its own model as long-term changes such as sensor aging and environmental condition drift occur. Thus, while ensuring detection accuracy, the reliability of the detection results and the long-term stability of the system are further improved, providing more valuable data support for safety decision-making in confined space maintenance operations at wastewater treatment plants. Attached Figure Description

[0034] Figure 1 This is a flowchart of the gaseous pollutant residue detection method based on multi-source data according to the present invention.

[0035] Figure 2 This is a schematic diagram of the module composition of the gaseous pollutant residue detection system based on multi-source data of the present invention. Detailed Implementation

[0036] 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.

[0037] Example 1: As Figure 1 As shown, this embodiment provides a method for detecting residual gaseous pollutants based on multi-source data. The method includes: S1. Acquire multi-source sensor data within the target confined space. The multi-source sensor data includes the ambient humidity value collected in real time by a humidity sensor and the raw response value sequence collected in real time by a gas sensor array. The gas sensor array contains multiple gas-sensitive units that have cross-sensitivity to gaseous pollutants.

[0038] For example, multi-source sensor data includes two types: first, the ambient humidity value collected in real time by a humidity sensor, continuously recorded in percentage relative humidity (%RH); and second, the raw response value sequence collected in real time by a gas sensor array. The gas sensor array consists of multiple gas-sensitive units, each exhibiting cross-sensitivity to gaseous pollutants. This means a single gas-sensitive unit may respond to multiple gaseous pollutants, and different units may respond to the same pollutant to varying degrees. For instance, a gas sensor array may consist of six gas-sensitive units, each using a metal oxide semiconductor material, responding to target gaseous pollutants such as hydrogen sulfide, methane, ammonia, and carbon monoxide to varying degrees. The raw response value sequence refers to the time sequence formed by arranging the output signals of each gas-sensitive unit over time at fixed sampling intervals. The sampling interval can be set according to on-site requirements, such as once per second.

[0039] S2. The original response value sequence is decomposed into multiple wavelet coefficient components at different frequency scales using multi-scale wavelet transform. The purpose of multi-scale wavelet transform processing on the original response value sequence is to extract the effective signal components related to the actual gaseous pollutant response from the original signal containing noise interference, while simultaneously removing irrelevant components such as electrical interference and vibration noise. Multi-scale wavelet transform decomposes the original response value sequence at different frequency scales, obtaining multiple wavelet coefficient components, each corresponding to a specific frequency scale range. The wavelet coefficient components at different frequency scales reflect the changing characteristics of the original signal at different time scales; low-frequency scale components correspond to slowly changing signal components, while high-frequency scale components correspond to rapidly changing signal components. The number of wavelet transform scales can be determined based on the sampling rate and signal bandwidth. For example, for a sensor signal with a 1Hz sampling rate, a five-level wavelet decomposition can be performed to obtain five wavelet coefficient components at different scales.

[0040] Identifying and retaining scale components that match the response time constant of typical gaseous pollutants; the method for identifying and retaining scale components that match the response time constant of typical gaseous pollutants includes: After calculating the actual response time constant of each wavelet coefficient component, the waveform asymmetry coefficient of each wavelet coefficient component is also calculated. The typical gaseous pollutant response time constant of each target gaseous pollutant and the typical waveform asymmetry coefficient of each target gaseous pollutant are combined to form a two-dimensional feature template.

[0041] For example, the waveform asymmetry coefficient is defined as the ratio of the rising edge duration to the falling edge duration of the wavelet coefficient component, taken as the result of a logarithmic transformation with the natural logarithm as the base: Waveform asymmetry coefficient = ln(Rising edge duration / Falling edge duration); where the rising edge duration refers to the time it takes for the amplitude of the wavelet coefficient component to rise from 10% to 90% of the baseline, and the falling edge duration refers to the time it takes for the amplitude to fall from 90% to 10%. Through logarithmic transformation, the waveform asymmetry coefficient can be made numerically symmetrical about zero. The coefficient is zero when the rising and falling edges are of equal length, positive when the rising edge is longer, and negative when the falling edge is longer. This corresponds to the differences in adsorption-desorption kinetics of different gaseous pollutants on sensors. For example, hydrogen sulfide typically exhibits a faster rising edge and a slower falling edge response on metal oxide sensors, resulting in a negative waveform asymmetry coefficient; while some organic gases may exhibit the opposite waveform asymmetry characteristics.

[0042] The pre-stored typical response time constant and typical waveform asymmetry coefficient for each target gaseous pollutant were obtained by introducing a single target gaseous pollutant of known concentration into the gas sensor array during laboratory calibration experiments. The time constant and waveform asymmetry coefficient of the response signal for each sensor channel were recorded in each experiment, and the average value of multiple repeated experiments was taken. These two parameters together constitute a two-dimensional feature template for that target gaseous pollutant. The two-dimensional feature template also includes a covariance matrix obtained statistically from multiple calibration experiments, used for calculating Mahalanobis distance.

[0043] Before performing Mahalanobis distance calculation, the ratio of the total energy of each wavelet coefficient component to the sum of the total energy of all wavelet coefficient components is calculated to obtain the energy proportion of each wavelet coefficient component. When the energy proportion of a wavelet coefficient component is lower than the preset energy proportion threshold, the wavelet coefficient component is directly identified as a noise component and removed.

[0044] For example, before calculating Mahalanobis distance, it is necessary to screen the energy proportions of each wavelet coefficient component to remove noise components with excessively low energy. The total energy of each wavelet coefficient component is defined as the sum of the squares of the amplitudes of all sampling points in the time series of that component. After calculating the sum of the total energy of all wavelet coefficient components, the energy proportion of each wavelet coefficient component is obtained, which is the ratio of the total energy of a single component to the sum of the total energy of all components. The energy proportion of each wavelet coefficient component is compared with a preset energy proportion threshold. When the energy proportion is lower than the energy proportion threshold, the wavelet coefficient component is directly identified as a noise component and removed, without further Mahalanobis distance matching. The energy proportion threshold is not a fixed constant but is dynamically adjusted based on the signal-to-noise ratio estimate of the original response value sequence. The signal-to-noise ratio (SNR) estimate is obtained by calculating the ratio of the total energy of the high-frequency wavelet coefficients to the total energy of the low-frequency wavelet coefficients in the original response value sequence. Its physical meaning is as follows: the higher the proportion of high-frequency energy in the signal, the stronger the noise; in this case, the energy proportion threshold should be appropriately increased to more rigorously filter out noise components. Conversely, if the high-frequency energy proportion is low, the signal quality is good, and the energy proportion threshold can be appropriately decreased to retain more effective information. Specifically, the energy proportion threshold can be dynamically set as follows: a baseline energy proportion threshold (e.g., 0.02, or 2%) is predetermined. When the SNR estimate is lower than the preset SNR reference value, the baseline energy proportion threshold is linearly scaled according to the reciprocal of the SNR estimate, so that the energy proportion threshold increases as the SNR decreases; when the SNR estimate is higher than the SNR reference value, the energy proportion threshold remains unchanged at the baseline energy proportion threshold.

[0045] The two-dimensional measured features, consisting of the actual response time constant and waveform asymmetry coefficient of each wavelet coefficient component, are compared with the two-dimensional feature template of each target gaseous pollutant to calculate the Mahalanobis distance. When the Mahalanobis distance is less than the preset Mahalanobis distance threshold, the wavelet coefficient component is determined to be a scale component that matches the corresponding target gaseous pollutant and is retained.

[0046] For example, after screening by energy proportion, the Mahalanobis distance between the remaining wavelet coefficient components and their two-dimensional measured features (i.e., the two-dimensional vector composed of the actual response time constant and the waveform asymmetry coefficient) and the two-dimensional feature template of each target gaseous pollutant is calculated. The formula for calculating the Mahalanobis distance is: ;in This represents the two-dimensional measured eigenvectors of the wavelet coefficient components. Let be the mean vector of a two-dimensional feature template for a certain target gaseous pollutant. is the inverse of the covariance matrix. Compared to Euclidean distance, Mahalanobis distance considers the correlation between the two feature dimensions and the dimensional differences between each dimension, and can more accurately reflect the statistical similarity between two-dimensional measured features and two-dimensional template features.

[0047] When the Mahalanobis distance is less than a preset Mahalanobis distance threshold, the wavelet coefficient component is determined to match the corresponding target gaseous pollutant response characteristics and is retained. The Mahalanobis distance threshold can be determined based on the distribution range of laboratory calibration data. For example, the 95th percentile of the Mahalanobis distance distribution in various target gaseous pollutant calibration experiments can be used as the threshold to ensure that the vast majority of true response components can be retained under normal calibration conditions.

[0048] For wavelet coefficient components whose Mahalanobis distance to two-dimensional feature templates of multiple target gaseous pollutants is less than a preset Mahalanobis distance threshold, the target gaseous pollutant corresponding to the minimum Mahalanobis distance is marked, and the wavelet coefficient component is retained as a scale component that matches each of the multiple target gaseous pollutants, so as to avoid the loss of signal components in mixed gas scenarios due to excessive removal.

[0049] The enhanced original response value sequence is obtained by wavelet reconstruction of the retained scale components. All retained effective scale components are then reconstructed using wavelet transform, i.e., the retained wavelet coefficients are superimposed and synthesized to obtain the enhanced original response value sequence. The enhanced original response value sequence, after eliminating noise components, retains the time-scale components that match the response characteristics of the target gaseous pollutant, resulting in signal quality superior to the original acquired data.

[0050] S3. Perform humidity interference compensation on the enhanced original response value sequence based on the ambient humidity value to obtain a humidity-corrected response value sequence; the method for performing humidity interference compensation on the enhanced original response value sequence based on the ambient humidity value to obtain a humidity-corrected response value sequence includes: Obtain the ambient humidity value at the sampling time corresponding to each response value in the enhanced original response value sequence, and obtain the historical ambient humidity values ​​of N consecutive sampling times before the current sampling time, arranged in chronological order to form a historical humidity memory vector; multiply each historical ambient humidity value in the historical humidity memory vector by the response value of each gas-sensitive unit in the enhanced original response value sequence at the current sampling time to generate a set of humidity-response cross-feature terms; concatenate the humidity-response cross-feature terms with the historical humidity memory vector, the ambient humidity value at the current sampling time, and the response values ​​of each gas-sensitive unit in the enhanced original response value sequence at the current sampling time to form an expanded input feature vector.

[0051] For example, high humidity in a confined space that fluctuates over time can systematically interfere with the response values ​​of gas-sensitive units such as those containing metal oxides. Without correction, this can lead to inaccurate gas concentration estimations. For each sampling moment in the enhanced original response value sequence, the current ambient humidity value corresponding to that sampling moment is obtained, along with the ambient humidity values ​​of the N consecutive sampling moments preceding the current sampling moment. These are arranged chronologically to form a historical humidity memory vector with dimension N. The value of N is determined based on the dynamic response characteristics of the gas sensor array to humidity changes. Generally, it is selected as the number of sampling points corresponding to 3 to 5 times the sensor's humidity response time constant. For example, if the sensor's humidity response time constant is 30 seconds and the sampling interval is 1 second, then N can be 90 to 150. The reason for introducing the historical humidity memory vector is that the gas sensor's response to humidity is not instantaneous; the humidity values ​​at historical moments still have a residual influence on the current sensor response value. Simply using the current humidity value for compensation would introduce systematic errors.

[0052] If the historical humidity memory vector has a dimension of N and the number of gas-sensitive units is M, then there are N×M humidity-response cross-feature terms. Each cross-feature term represents the coupling relationship between the humidity level at a certain historical moment and the current response value of a certain sensor, enabling the compensation model to capture the nonlinear characteristics of humidity-response coupling. The humidity-response cross-feature terms (N×M dimensions), the historical humidity memory vector (N dimensions), the ambient humidity value at the current sampling moment (1 dimension), and the enhanced original response values ​​of each gas-sensitive unit at the current sampling moment (M dimensions) are concatenated to form the expanded input feature vector.

[0053] A feedforward neural network is constructed as a humidity interference compensation model, and the number of input layer nodes of the feedforward neural network is set to be equal to the dimension of the expanded input feature vector.

[0054] The number of hidden layer nodes in the feedforward neural network is set to twice the number of input layer nodes plus one, and a nonlinear activation function is used to perform nonlinear mapping on the expanded input feature vector. The number of output layer nodes in the feedforward neural network is equal to the number of gas-sensitive units in the gas sensor array. Each output layer node corresponds to the response value of a gas-sensitive unit after humidity interference compensation. The loss function used by the humidity interference compensation model during the training phase is the weighted mean square error loss function.

[0055] For example, the humidity interference compensation model employs a three-layer feedforward neural network structure. The number of nodes in the input layer equals the total dimension of the expanded input feature vector. The number of nodes in the hidden layer is set to twice the number of nodes in the input layer plus one; that is, if the number of nodes in the input layer is D, then the number of nodes in the hidden layer is 2D+1. The hidden layer uses a nonlinear activation function (such as the hyperbolic tangent function tanh or the modified linear unit ReLU) to perform a nonlinear mapping on the expanded input feature vector to capture the nonlinear components in humidity interference. The number of nodes in the output layer is equal to the number M of gas-sensitive units in the gas sensor array, and each output layer node corresponds to the response value of a gas-sensitive unit after humidity interference compensation.

[0056] The neural network was trained offline. Training samples were collected as follows: In a laboratory environment, at different controlled humidity levels (e.g., 30%RH, 50%RH, 70%RH, 90%RH, etc.), target gaseous pollutants of known concentrations were introduced into the gas sensor array, and the response values ​​of each gas-sensitive unit and the corresponding ambient humidity values ​​were recorded simultaneously. The response values ​​recorded at a reference humidity level (e.g., 50%RH) were used as the target output without humidity interference, and the response values ​​recorded at other humidity levels, along with the corresponding humidity data, were used as inputs to construct training sample pairs. A weighted mean square error loss function was used during the training phase, where the weight of each output node was proportional to the humidity sensitivity coefficient of the corresponding gas-sensitive unit in historical calibration experiments. The humidity sensitivity coefficient was obtained by statistically analyzing the ratio of humidity change to response value change in the calibration experiments. Gas-sensitive units with higher humidity sensitivity had a larger weight in the loss function, making the model focus more on the compensation accuracy of channels with greater humidity interference.

[0057] The extended input feature vector is input into the humidity interference compensation model. After forward propagation calculation by the feedforward neural network, the humidity-compensated response value corresponding to each gas-sensitive unit is obtained from the output layer node. The humidity-compensated response values ​​of all gas-sensitive units are combined in the same order as the enhanced original response value sequence to obtain the humidity-corrected response value sequence.

[0058] S4. The humidity-corrected response value sequence is subjected to adaptive cross-sensitivity decoupling processing to decompose it into multiple independent source signal time components; the method for performing adaptive cross-sensitivity decoupling processing on the humidity-corrected response value sequence to decompose it into multiple independent source signal time components includes: The humidity-corrected response value sequence is organized into an observation matrix, where the rows of the observation matrix correspond to each gas-sensitive unit in the gas sensor array, and the columns of the observation matrix correspond to the humidity-corrected response value at each sampling time.

[0059] For example, adaptive cross-sensitivity decoupling processing is performed on the humidity-corrected response value sequence to separate the components contributed by different gaseous pollutants in the response values ​​of each gas-sensitive unit. Since each gas-sensitive unit responds to multiple gaseous pollutants, its actual output is the superposition of the effects of multiple gaseous pollutants. Directly estimating the concentration of each pollutant from the response value will produce serious cross-sensitivity errors. Decoupling processing is the core step to eliminate cross-sensitivity interference.

[0060] The humidity-corrected response value sequence is organized into an observation matrix. The rows of the observation matrix correspond to each gas-sensitive unit in the gas sensor array (M rows in total), and the columns correspond to the humidity-corrected response values ​​at each sampling time (T columns in total, where T is the length of the time series). The dimension of the observation matrix is ​​M×T.

[0061] Independent Component Analysis (ICA) is used to perform blind source separation on the observation matrix. An unmixing matrix is ​​obtained through iterative optimization. Multiplying the unmixing matrix by the observation matrix yields the source signal matrix. Each row in the source signal matrix represents an independent time component of the source signal. ICA is based on the premise that the observed signal is a linear superposition of multiple statistically independent source signals. By solving the unmixing matrix, the observed signal is restored to several independent source signal components. In this method, the temporal changes in the concentration of various gaseous pollutants are considered as independent source signals, and the mixed response of the gas sensor array is considered as a linear mixture of these source signals.

[0062] The method of performing blind source separation processing on the observation matrix using independent component analysis algorithm and iteratively optimizing to solve an unmixing matrix includes: A composite optimization objective is formed by adding the statistical independence measure among the response values ​​of each gas-sensitive unit in the observation matrix with a sparsity penalty term. The sparsity penalty term is the sum of the L1 norms of all source signal time components in the source signal matrix obtained by multiplying the current unmixed matrix with the observation matrix.

[0063] A sparsity regularization intensity coefficient is set, which is determined based on the maximum number of target gaseous pollutant types appearing simultaneously in historical monitoring data before maintenance operations in the confined space of the wastewater treatment plant; in each iteration of iterative optimization to solve the unmixing matrix, the natural gradient of the composite optimization objective relative to the current unmixing matrix is ​​calculated, and the unmixing matrix is ​​updated according to the direction of the natural gradient.

[0064] In the optimization objective of solving the unmixing matrix, a composite optimization objective is formed by adding the statistical independence measure and the sparsity penalty term to the response values ​​of each gas-sensitive unit. The statistical independence measure typically adopts the criteria of maximizing negative entropy or minimizing mutual information. The sparsity penalty term is the sum of the L1 norms of all time components of the source signals in the source signal matrix obtained by multiplying the current unmixing matrix and the observation matrix. The sparsity penalty is introduced because the number of target gaseous pollutants that actually exceed the standard at any time in the finite space is usually much smaller than the total number of gas-sensitive units. That is, most source signals have amplitudes close to zero at most times, exhibiting sparsity characteristics. L1 regularization helps the algorithm converge to a physically more reasonable sparse solution.

[0065] The sparsity regularization intensity coefficient is determined based on the maximum number of target gaseous pollutant types simultaneously occurring in historical monitoring data prior to maintenance work in the confined space of the wastewater treatment plant; the two are inversely proportional. If a maximum of K target gaseous pollutants were simultaneously detected exceeding the standard in the historical data, the sparsity regularization intensity coefficient can be set according to λ=λ0 / K, where λ0 is the baseline regularization intensity coefficient (set based on experimental experience, for example, 0.1). When K is small (e.g., historically only two gases have exceeded the standard simultaneously), the sparsity is stronger, the regularization intensity coefficient is larger, and the algorithm tends to generate a sparse source signal matrix; when K is large, the regularization intensity coefficient is reduced to avoid excessive sparsity that would suppress effective source signal components.

[0066] In each iteration, the natural gradient of the composite optimization objective relative to the current unmixed matrix is ​​calculated, and the unmixed matrix is ​​updated according to the direction of the natural gradient. Compared with the ordinary gradient, the natural gradient is a better update direction in the Riemannian manifold sense, which can accelerate the convergence speed and improve the stability of the solution.

[0067] During the iterative optimization process, the Independent Component Analysis (ICA) algorithm adaptively adjusts the update direction of the unmixing matrix based on the actual data distribution of the humidity-corrected response value sequence until the mutual information between the time components of each source signal in the source signal matrix converges to a minimum, completing the adaptive cross-sensitivity decoupling process and obtaining multiple independent source signal time components. The iterative process continues until the mutual information between the time components of each source signal in the source signal matrix converges to a minimum, meaning the statistical independence of each source signal time component reaches its optimal value, completing the adaptive cross-sensitivity decoupling process. "Adaptive" means that the ICA algorithm automatically adjusts the selection of the nonlinear function and the update step size based on the actual data distribution characteristics of the humidity-corrected response value sequence during the iteration process, without requiring manual pre-specification of the data distribution type. The source signal matrix is ​​obtained by multiplying the unmixing matrix and the observation matrix. Each row in the source signal matrix represents an independent source signal time component, and the number of source signal time components is equal to the number M of gas sensing units.

[0068] The method for extracting the residual concentration contribution value sequence corresponding to each target gaseous pollutant from the time component of the source signal includes: A reference response pattern vector is pre-constructed for each target gaseous pollutant. This vector is obtained by introducing a single target gaseous pollutant of known concentration into a gas sensor array under laboratory conditions and recording the steady-state response values ​​of each gas-sensitive unit in the array. The dimension of the reference response pattern vector is equal to the number of gas-sensitive units in the gas sensor array. The vector is also obtained by introducing a single target gaseous pollutant of known concentration into the gas sensor array under laboratory conditions and recording the steady-state response values ​​of each gas-sensitive unit. The vector dimension is equal to the number of gas-sensitive units M, and each element in the vector corresponds to the steady-state response value of a gas-sensitive unit under the influence of a single target gaseous pollutant. The reference response pattern vector reflects the "fingerprint" characteristics of different target gaseous pollutants on the sensor array.

[0069] The steps for online adaptive updating of the reference response mode vector include: At each sampling moment, the fitting residual vector between the initial source signal component and the actual humidity-corrected response value of each gas sensing unit in the gas sensor array is calculated, and the root mean square error of the fitting residual vector is calculated within a sliding time window. When the root mean square error is continuously less than a preset stability threshold for a preset trigger duration, the actual humidity-corrected response values ​​of each gas sensing unit in the gas sensor array at the current sampling moment are arranged into a measured response pattern vector according to the order of each gas sensing unit. The measured response pattern vector is then weighted and averaged with the currently stored reference response pattern vector to replace the original reference response pattern vector. The reference response pattern vector supports online adaptive updates to cope with changes in response characteristics caused by aging, drift, and other factors during long-term sensor use. At each sampling moment, the fitting residual vector of the current initial source signal component to the response value of each gas sensing unit is calculated. The fitting residual vector is a vector composed of the difference between the actual humidity-corrected response value of each gas sensing unit and the fitted value restored by the initial source signal component according to the contribution coefficient, arranged in the order of the sensors. The root mean square error (RMSE) of the fitted residual vector is calculated within a sliding time window (e.g., 60 sampling points, corresponding to 60 seconds). When the RMSE is continuously less than a preset stability threshold (the stability threshold is determined based on the typical fluctuation range of the sensor response value in the calibration experiment, e.g., 1% of the sensor's full scale) for a period of time that reaches a preset trigger duration (e.g., continuously stable for 120 seconds), the sensor response is considered to have reached a stable state. The actual humidity-corrected response values ​​of each gas-sensitive unit at the current sampling time are arranged in sensor order to form a measured response pattern vector. The measured response pattern vector is then weighted and averaged with the currently stored reference response pattern vector. The weight coefficient of the measured response pattern vector is inversely proportional to the RMSE (i.e., the better the fitting quality, the higher the weight of the measured vector, which can be calculated as weight coefficient = 1 / (1 + RMSE / stability threshold)). The weight coefficient of the currently stored reference response pattern vector is equal to 1 minus the weight coefficient of the measured response pattern vector. The result of the weighted average replaces the original reference response pattern vector, completing one online adaptive update.

[0070] Each source signal time component is considered as a source signal vector to be matched, and the similarity between each source signal vector to be matched and the reference response mode vector of each target gaseous pollutant is calculated.

[0071] From all the source signal vectors to be matched, the one with the highest similarity to the reference response pattern vector of the target gaseous pollutant is selected as the initial source signal component of the target gaseous pollutant. When extracting the residual concentration contribution value sequence, the source signal time component of each row in the source signal matrix is ​​regarded as a source signal vector to be matched, and the similarity between each source signal vector to be matched and the reference response pattern vector of each target gaseous pollutant is calculated (e.g., using normalized cross-correlation coefficient or cosine similarity). From all the source signal vectors to be matched, the one with the highest similarity to the reference response pattern vector of a certain target gaseous pollutant is selected as the initial source signal component of that target gaseous pollutant.

[0072] Using the initially selected source signal component as the independent variable and the actual humidity-corrected response value of each gas sensing unit in the gas sensor array as the dependent variable, the contribution coefficient of each initially selected source signal component to the response value of each gas sensing unit is solved using the least squares method. The residual concentration contribution value of the target gaseous pollutant is obtained by multiplying the amplitude of each sampling moment in the initially selected source signal component by the corresponding contribution coefficient. The residual concentration contribution values ​​at all sampling moments are arranged in chronological order to obtain a sequence of residual concentration contribution values. The contribution coefficient refers to the change in the response value of the corresponding gas sensing unit caused by a unit amplitude change in the initially selected source signal component; its value comprehensively reflects the concentration level of the target gaseous pollutant and the sensor sensitivity.

[0073] S5. Take the latest value in the residual concentration contribution value sequence as the detection result at the current moment; calculate the coefficient of variation of the residual concentration contribution value sequence within the current sliding time window, calculate the confidence level of the residual concentration contribution value of each target gaseous pollutant based on the coefficient of variation, and output the detection result and confidence level.

[0074] For example, the latest value in the residual concentration contribution value sequence (i.e., the residual concentration contribution value at the current sampling time) is output as the detection result at the current time. To provide an assessment of the reliability of the detection results, the confidence level of each target gaseous pollutant detection result also needs to be calculated. The coefficient of variation is defined as the ratio of the standard deviation to the mean of the residual concentration contribution value within the sliding time window (i.e., the relative standard deviation). The smaller the coefficient of variation, the smaller the fluctuation of the residual concentration contribution value within the window, the more stable the signal, and the more reliable the detection result; the larger the coefficient of variation, the larger the fluctuation of the concentration contribution value, which may be affected by unresolved interference, and the reliability of the detection result is correspondingly reduced. When calculating the confidence level based on the coefficient of variation, a monotonically decreasing mapping function can be used, for example, confidence level = 1 / (1 + coefficient of variation), so that the confidence level value falls between 0 and 1. When the coefficient of variation is zero, the confidence level is 1, and as the coefficient of variation increases, the confidence level monotonically decreases. The length of the sliding time window can be set according to the on-site detection needs, for example, taking the time period corresponding to the most recent 30 sampling points. The current detection results (residual concentration contribution value) of each target gaseous pollutant and the corresponding confidence level are used as the system output to provide on-site safety management personnel with a basis for judging the residual status of gaseous pollutants in the confined space and the reliability of the detection results, and to provide a basis for safety decisions before entering the work.

[0075] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a gaseous pollutant residue detection system based on multi-source data, the system comprising: A multi-source sensor data acquisition module is used to acquire multi-source sensor data within a limited space of the target. The multi-source sensor data includes the ambient humidity value acquired in real time by a humidity sensor and the raw response value sequence acquired in real time by a gas sensor array. The gas sensor array contains multiple gas-sensitive units that have cross-sensitivity to gaseous pollutants.

[0076] The signal enhancement module is used to decompose the original response value sequence into multiple wavelet coefficient components of different frequency scales through multi-scale wavelet transform, identify and retain the scale components that match the response time constant of typical gaseous pollutants, and perform wavelet reconstruction on the retained scale components to obtain the enhanced original response value sequence.

[0077] The humidity interference compensation module is used to perform humidity interference compensation on the enhanced original response value sequence based on the ambient humidity value, so as to obtain a humidity-corrected response value sequence.

[0078] The cross-sensitive adaptive decoupling module is used to perform adaptive cross-sensitive decoupling processing on the humidity-corrected response value sequence, decompose it into multiple independent source signal time components, and extract the residual concentration contribution value sequence corresponding to each target gaseous pollutant from the source signal time components.

[0079] The detection result output module is used to take the latest value in the residual concentration contribution value sequence as the detection result at the current moment; calculate the coefficient of variation of the residual concentration contribution value sequence within the current sliding time window; calculate the confidence level of the residual concentration contribution value of each target gaseous pollutant based on the coefficient of variation; and output the detection result and the confidence level.

[0080] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0081] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting residual gaseous pollutants based on multi-source data, characterized in that, The method includes: Acquire multi-source sensor data within a limited target space. The multi-source sensor data includes ambient humidity values ​​collected in real time by a humidity sensor and raw response value sequences collected in real time by a gas sensor array. The gas sensor array contains multiple gas-sensitive units that have cross-sensitivity to gaseous pollutants. The original response value sequence is decomposed into multiple wavelet coefficient components of different frequency scales through multi-scale wavelet transform. The scale components that match the response time constant of typical gaseous pollutants are identified and retained. The retained scale components are then reconstructed by wavelet transform to obtain the enhanced original response value sequence. Based on the ambient humidity value, humidity interference compensation is performed on the enhanced original response value sequence to obtain a humidity-corrected response value sequence. The humidity-corrected response value sequence is subjected to adaptive cross-sensitivity decoupling processing, decomposed into multiple independent source signal time components, and the residual concentration contribution value sequence corresponding to each target gaseous pollutant is extracted from the source signal time components. The latest value in the residual concentration contribution value sequence is taken as the detection result at the current moment; the coefficient of variation of the residual concentration contribution value sequence within the current sliding time window is calculated, the confidence level of the residual concentration contribution value of each target gaseous pollutant is calculated based on the coefficient of variation, and the detection result and confidence level are output.

2. The method for detecting residual gaseous pollutants based on multi-source data according to claim 1, characterized in that, The method for identifying and retaining scale components that match the response time constant of typical gaseous pollutants includes: After calculating the actual response time constant of each wavelet coefficient component, the waveform asymmetry coefficient of each wavelet coefficient component is also calculated. The typical gaseous pollutant response time constant of each target gaseous pollutant and the typical waveform asymmetry coefficient of each target gaseous pollutant stored in the pre-stored data are used together to form a two-dimensional feature template. The two-dimensional measured features, consisting of the actual response time constant and waveform asymmetry coefficient of each wavelet coefficient component, are compared with the two-dimensional feature template of each target gaseous pollutant to calculate the Mahalanobis distance. When the Mahalanobis distance is less than the preset Mahalanobis distance threshold, the wavelet coefficient component is determined to be a scale component that matches the corresponding target gaseous pollutant and is retained.

3. The gaseous pollutant residue detection method based on multi-source data according to claim 2, characterized in that, Before performing Mahalanobis distance calculation, the ratio of the total energy of each wavelet coefficient component to the sum of the total energy of all wavelet coefficient components is calculated to obtain the energy proportion of each wavelet coefficient component. When the energy proportion of a wavelet coefficient component is lower than the preset energy proportion threshold, the wavelet coefficient component is directly identified as a noise component and removed.

4. The method for detecting residual gaseous pollutants based on multi-source data according to claim 1, characterized in that, The method for compensating the enhanced original response value sequence for humidity interference based on the ambient humidity value to obtain a humidity-corrected response value sequence includes: Obtain the ambient humidity value at the sampling time corresponding to each response value in the enhanced original response value sequence, and obtain the historical ambient humidity values ​​of N consecutive sampling times before the current sampling time, arranged in chronological order to form a historical humidity memory vector; multiply each historical ambient humidity value in the historical humidity memory vector by the response value of each gas-sensitive unit in the enhanced original response value sequence at the current sampling time to generate a set of humidity-response cross-feature terms; concatenate the humidity-response cross-feature terms with the historical humidity memory vector, the ambient humidity value at the current sampling time, and the response values ​​of each gas-sensitive unit in the enhanced original response value sequence at the current sampling time to form an expanded input feature vector; The extended input feature vector is input into the humidity interference compensation model. After forward propagation calculation by the feedforward neural network, the humidity-compensated response value corresponding to each gas-sensitive unit is obtained from the output layer node. The humidity-compensated response values ​​of all gas-sensitive units are combined in the same order as the enhanced original response value sequence to obtain the humidity-corrected response value sequence.

5. The method for detecting residual gaseous pollutants based on multi-source data according to claim 4, characterized in that, A feedforward neural network is constructed as a humidity interference compensation model, wherein the number of nodes in the input layer of the feedforward neural network is set to be equal to the dimension of the expanded input feature vector; The number of hidden layer nodes in the feedforward neural network is set to twice the number of input layer nodes plus one, and a nonlinear activation function is used to perform nonlinear mapping on the expanded input feature vector. The number of output layer nodes in the feedforward neural network is equal to the number of gas-sensitive units in the gas sensor array. Each output layer node corresponds to the response value of a gas-sensitive unit after humidity interference compensation. The loss function used by the humidity interference compensation model during the training phase is the weighted mean square error loss function.

6. The method for detecting residual gaseous pollutants based on multi-source data according to claim 1, characterized in that, The method for performing adaptive cross-sensitivity decoupling processing on the humidity-corrected response value sequence to decompose it into multiple independent source signal time components includes: The humidity-corrected response value sequence is organized into an observation matrix, where the rows of the observation matrix correspond to each gas-sensitive unit in the gas sensor array, and the columns of the observation matrix correspond to the humidity-corrected response value at each sampling time. The observation matrix is ​​subjected to blind source separation using an independent component analysis algorithm. An unmixing matrix is ​​obtained through iterative optimization. The source signal matrix is ​​obtained by multiplying the unmixing matrix with the observation matrix. Each row of the source signal matrix represents an independent time component of the source signal. During the iterative optimization process, the independent component analysis algorithm adaptively adjusts the update direction of the unmixing matrix according to the actual data distribution of the humidity-corrected response value sequence until the mutual information between the time components of each source signal in the source signal matrix converges to the minimum value, thus completing the adaptive cross-sensitive decoupling process and obtaining multiple independent source signal time components.

7. The method for detecting residual gaseous pollutants based on multi-source data according to claim 6, characterized in that, The method of performing blind source separation processing on the observation matrix using independent component analysis algorithm and iteratively optimizing to solve an unmixing matrix includes: A composite optimization objective is formed by adding the statistical independence measure between the response values ​​of each gas sensitive unit in the observation matrix and the sparsity penalty term. The sparsity penalty term is the sum of the L1 norms of all source signal time components in the source signal matrix obtained by multiplying the current unmixed matrix and the observation matrix. A sparsity regularization intensity coefficient is set, which is determined based on the maximum number of target gaseous pollutant types appearing simultaneously in historical monitoring data before maintenance operations in the confined space of the wastewater treatment plant; in each iteration of iterative optimization to solve the unmixing matrix, the natural gradient of the composite optimization objective relative to the current unmixing matrix is ​​calculated, and the unmixing matrix is ​​updated according to the direction of the natural gradient.

8. The method for detecting residual gaseous pollutants based on multi-source data according to claim 1, characterized in that, The method for extracting the residual concentration contribution value sequence of each target gaseous pollutant from the time component of the source signal includes: A reference response mode vector is pre-constructed for each target gaseous pollutant. The reference response mode vector is obtained by introducing a single target gaseous pollutant of known concentration into the gas sensor array in a laboratory environment and recording the steady-state response value of each gas sensitive unit in the gas sensor array. The dimension of the reference response mode vector is equal to the number of gas sensitive units in the gas sensor array. Each source signal time component in the source signal time component is regarded as a source signal vector to be matched, and the similarity between each source signal vector to be matched and the reference response mode vector of each target gaseous pollutant is calculated. Select the source signal vector with the highest similarity to the reference response pattern vector of the target gaseous pollutant from all the source signal vectors to be matched as the initial source signal component of the target gaseous pollutant. Using the initial source signal component as the independent variable and the actual humidity-corrected response value of each gas sensing unit in the gas sensor array as the dependent variable, the contribution coefficient of each initial source signal component to the response value of each gas sensing unit is solved by the least squares method. The amplitude of each sampling moment in the initial source signal component is multiplied by the corresponding contribution coefficient to obtain the residual concentration contribution value of the target gaseous pollutant. The residual concentration contribution values ​​of all sampling moments are arranged in chronological order to obtain the residual concentration contribution value sequence.

9. The method for detecting residual gaseous pollutants based on multi-source data according to claim 8, characterized in that, The steps for online adaptive updating of the reference response mode vector include: At each sampling moment, the fitting residual vector between the initially selected source signal component and the actual humidity-corrected response value of each gas sensing unit in the gas sensor array is calculated, and the root mean square error of the fitting residual vector is calculated within a sliding time window; when the root mean square error is continuously less than a preset stability threshold for a preset trigger duration, the actual humidity-corrected response values ​​of each gas sensing unit in the gas sensor array at the current sampling moment are arranged into a measured response mode vector according to the order of each gas sensing unit; the vector obtained by weighted averaging the measured response mode vector and the currently stored reference response mode vector replaces the original reference response mode vector.

10. A gaseous pollutant residue detection system based on multi-source data, characterized in that, The system includes: A multi-source sensor data acquisition module is used to acquire multi-source sensor data within a limited space of the target. The multi-source sensor data includes the ambient humidity value acquired in real time by a humidity sensor and the raw response value sequence acquired in real time by a gas sensor array. The gas sensor array contains multiple gas-sensitive units that have cross-sensitivity to gaseous pollutants. The signal enhancement module is used to decompose the original response value sequence into multiple wavelet coefficient components of different frequency scales through multi-scale wavelet transform, identify and retain the scale components that match the response time constant of typical gaseous pollutants, and perform wavelet reconstruction on the retained scale components to obtain the enhanced original response value sequence. A humidity interference compensation module is used to perform humidity interference compensation on the enhanced original response value sequence based on the ambient humidity value, so as to obtain a humidity-corrected response value sequence. The cross-sensitive adaptive decoupling module is used to perform adaptive cross-sensitive decoupling processing on the humidity-corrected response value sequence, decompose it into multiple independent source signal time components, and extract the residual concentration contribution value sequence corresponding to each target gaseous pollutant from the source signal time components. The detection result output module is used to take the latest value in the residual concentration contribution value sequence as the detection result at the current moment; calculate the coefficient of variation of the residual concentration contribution value sequence within the current sliding time window; calculate the confidence level of the residual concentration contribution value of each target gaseous pollutant based on the coefficient of variation; and output the detection result and the confidence level.

Images