A method for high-sensitivity detection of gas decomposition products by fusing multi-dimensional sensing data

Abstract

The application discloses a kind of gas decomposition product high sensitivity detection methods of fusion multidimensional sensing data, the present application relates to gas detection technical field, construct heterogeneous three modal sensing array, with mutual information maximization dynamic optimization sensing parameter, combine bayesian credibility fusion and KL divergence anomaly separation, improve trace detection precision, simultaneously realize sensor self-diagnosis and fault-tolerant operation, the present application is in that Advantages lie in: by constructing chemical resistance type, electrochemical amperometric type, ultraviolet-visible absorption spectrum type three heterogeneous physical modal sensing array, combined with multidimensional response signal synergistic fusion analysis, break through the information dimension limitation of single sensing modal, can effectively weaken the cross interference influence between multiple-component gas, improve the component identification ability and quantitative detection stability of gas decomposition product under trace concentration, provide more reliable detection support for early fault warning of power equipment insulating gas.

Description

Technical Field

[0001] This invention relates to the field of gas detection technology, specifically to a highly sensitive detection method for gas decomposition products that integrates multidimensional sensor data. Background Technology

[0002] In power transmission and distribution systems, high-voltage electrical equipment such as gas-insulated switchgear and gas-insulated transformers are widely used due to their excellent insulation performance and arc-extinguishing capabilities. As the core medium for equipment insulation and arc extinguishing, the operating status of insulating gas is directly related to the safe and stable operation of the entire power grid. When there are latent faults such as partial discharge and overheating inside the equipment, the insulating gas will decompose and generate a variety of characteristic decomposition products. The composition and concentration of these products are directly related to the fault type and severity. Therefore, accurate detection of gas decomposition products is the core technical means to realize equipment status perception, early fault warning and condition maintenance. Existing gas decomposition product detection schemes mainly rely on single-physical-mode sensor arrays combined with backend intelligent algorithms to identify and detect the concentration of insulating gas decomposition components. However, these schemes have several drawbacks. First, the information dimensions of a single sensor mode are inherently limited, cross-interference between multiple gas components is difficult to eliminate effectively, and component identification capabilities at trace concentrations are insufficient, failing to provide reliable support for early fault warnings of power equipment. Second, the fixed operating parameters of the sensor units prevent proactive optimization of the complementarity of multi-dimensional information and lack an adaptive compensation mechanism for sensor performance degradation, resulting in insufficient long-term operational stability and requiring frequent manual calibration and maintenance. Third, they cannot effectively distinguish between real gas concentration anomalies and sensor malfunctions, easily leading to false alarms and missed alarms due to sensor drift or failure, making them unsuitable for long-term online monitoring needs in unattended scenarios. Therefore, we propose a highly sensitive gas decomposition product detection method that integrates multi-dimensional sensor data. Summary of the Invention

[0003] The purpose of this invention is to provide a highly sensitive detection method for gas decomposition products that integrates multidimensional sensor data.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a highly sensitive detection method for gas decomposition products that integrates multi-dimensional sensor data, the method comprising the following steps: Step 1: Construct a heterogeneous three-modal sensing array that includes a chemical resistive MOS sensor array, an electrochemical amperometric sensor array, and a UV-Vis absorption spectroscopy micro-sensor. Complete the initial zero point and sensitivity calibration, and establish a benchmark library of multimodal joint response probability distribution under normal operating conditions. Step 2: Simultaneously acquire the raw response data of the three sensing modes, and obtain the dimension-matched response vectors of each mode after baseline correction, denoising and standardization preprocessing. Step 3: Calculate the mutual information between each pair of the three modes using the k-nearest neighbor probability density estimation method, and quantify the information redundancy and complementarity of different modes; Step 4: Using the maximization of mutual information between the three-modal joint response and the concentration of the gas component to be measured as the objective function, the core operating parameters of each mode are dynamically optimized through an online gradient descent algorithm; Step 5: Based on independent quantitative analysis models, each mode outputs the estimated concentration of the target gas component and the corresponding measurement uncertainty. Step 6: Dynamically update the Bayesian confidence weights of each mode through cross-modal concentration consistency verification, and fuse and output the comprehensive concentration estimate and the corresponding confidence interval; Step 7: Calculate the KL divergence between the current joint response and the baseline distribution, and, in conjunction with the consistency of the three-mode response, distinguish between two types of events: abnormal gas concentration and sensor failure. Step 8: Output the final detection results and perform adaptive rolling updates on the normal operating benchmark library.

[0005] As a further aspect of the present invention: In step one, three types of sensing units—a chemical resistance MOS sensor array, an electrochemical amperometric sensor array, and a UV-Vis absorption spectroscopy micro-sensor—are sealed and integrated in the same detection chamber to ensure that the gas to be measured can act synchronously and uniformly on all sensing units; the decomposition products of the target gas at known standard concentrations are used to perform initial zero-point calibration and sensitivity calibration on each sensing mode in sequence, and the initial operating temperature of the MOS sensor, the initial working electrode bias voltage of the electrochemical sensor, the initial integration time of the spectral sensor, and the light source intensity reference parameters are set respectively. At the same time, multiple sets of normal operating state response data are collected to establish a multi-modal joint response probability distribution reference library for the normal operating state of the system.

[0006] As a further aspect of the present invention: In step two, the gas to be tested is controlled to be introduced into the detection chamber at a set stable flow rate. After the pressure and temperature parameters in the chamber stabilize, the raw response data of three sensing modes are collected synchronously at a uniform sampling frequency. The chemical resistance mode collects the resistance change time series data, the electrochemical mode collects the limiting diffusion current time series data, and the spectral mode collects the absorbance spectrum data of the entire ultraviolet-visible band. The raw data of each mode are sequentially processed by baseline drift correction, sliding window smoothing and noise reduction, and outlier removal to obtain the dimension-matched and dimensionally standardized response vectors of each mode.

[0007] As a further aspect of the present invention: In step three, for the three preprocessed modal response vectors, the k-nearest neighbor probability density estimation method is used to calculate the pairwise mutual information values ​​between the chemical resistance mode and the electrochemical mode, the chemical resistance mode and the spectral mode, and the electrochemical mode and the spectral mode, respectively; the mutual information values ​​are used to quantify the degree of information redundancy and independent complementarity between different modes. The lower the mutual information value, the lower the overlap of the detection information of the two modes and the stronger the independent complementarity, and vice versa, the higher the information redundancy of the two modes.

[0008] As a further aspect of the present invention: In step four, the global optimization objective is to maximize the mutual information between the joint response of the three modes and the concentration of the gas component to be measured. The following core objective function is constructed, and an online gradient descent algorithm is used to iteratively adjust the core operating parameters of each mode. The adjusted parameters include the operating temperature of the MOS sensor, the bias voltage of the working electrode of the electrochemical sensor, the integration time of the spectral sensor, and the light source intensity. After each set number of continuous measurements, a parameter optimization iteration is performed to maintain the information complementarity of the three modes and the overall detection information in a globally optimal state. ; In the formula, Indicates maximization operation; subscript This indicates that the set of working parameters is used as the optimization variable; The set of operating parameters for the three sensing modes, where For the operating parameters of the chemical resistance mode, For the operating parameters of the electrochemical mode, These are the operating parameters for the spectral modes; This represents the response vector of the chemical resistance mode under the corresponding operating parameters. This represents the response vector of the electrochemical mode under the corresponding operating parameters. This represents the response vector of the spectral mode under the corresponding operating parameters; This represents the component concentration vector of the gas to be measured; The operator is for mutual information calculation, with the two sets of random variables before and after the semicolon representing the two sets of random variables participating in the mutual information calculation.

[0009] As a further aspect of the present invention: In step five, based on the optimized working parameters and the standardized response vectors of each mode, the quantitative analysis models pre-trained for the corresponding modes are called respectively, and the concentration estimates and corresponding measurement uncertainties of the decomposition products of each target gas are output independently; wherein the chemical resistance mode adopts an extremely random tree regression model, the electrochemical mode adopts the limiting current concentration calibration curve, and the spectral mode adopts the Lambert-Beer law combined with a partial least squares regression model, ensuring that the quantitative results of the three modes are generated based on completely independent physical measurement principles.

[0010] As a further aspect of the present invention: In step six, the relative deviations of the gas component concentration estimates output independently by the three modes are calculated pairwise, and the Bayesian confidence weights of each mode are dynamically updated based on the magnitude of the deviations and a preset deviation threshold. When the deviation between the output result of a certain mode and the other two modes exceeds the set threshold, it is determined that the mode has signal drift, surface poisoning, and / or performance degradation, and its confidence weight is automatically reduced proportionally. Finally, based on the Bayesian evidence fusion framework, the comprehensive concentration estimate of the gas component and the corresponding confidence interval are calculated by combining the real-time confidence weights and measurement uncertainties of each mode.

[0011] As a further aspect of the present invention: In step seven, the KL divergence between the current three-modal joint response distribution and the normal operation distribution in the benchmark library is calculated in real time, and the abnormal event is classified and determined in combination with the response consistency of the three modes; when the KL divergence exceeds the set abnormal threshold and the response change trend of the three modes is consistent, it is determined to be a real gas concentration abnormal event, corresponding to the power equipment insulation fault warning; when the KL divergence exceeds the set abnormal threshold but only a single mode response is abnormal and the responses of the other two modes are within the normal fluctuation range, it is determined to be a sensor fault event of the corresponding mode, and the corresponding sensor operation and maintenance warning signal is generated.

[0012] As a further aspect of the present invention: In step eight, the final gas component comprehensive concentration result, event type determination result and corresponding confidence level are output, and the working parameters, response data and confidence weight of each mode are recorded simultaneously; when it is determined to be in normal operation and multiple sets of detection results are stable, the current multimodal joint response data is rolled into the benchmark library, and the normal operation distribution benchmark is adaptively updated to adapt to the slow aging of the sensor and the slow change of environmental parameters.

[0013] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows: 1. This invention constructs sensor arrays with three heterogeneous physical modes: chemical resistance, electrochemical amperometric, and ultraviolet-visible absorption spectroscopy. By combining the synergistic fusion analysis of multi-dimensional response signals, it breaks through the information dimension limitations of a single sensor mode. It can effectively reduce the cross-interference between multi-component gases, improve the component identification ability and quantitative detection stability of gas decomposition products at trace concentrations, and provide more reliable detection support for early fault warning of insulating gases in power equipment. 2. This invention dynamically adjusts the operating parameters of each sensing mode with the goal of maximizing cross-modal mutual information, actively optimizes the information complementarity between multiple modes, and, in conjunction with cross-modal consistency verification and dynamic weighting of reliability, can automatically adjust the fusion weight to maintain detection accuracy when a single sensing unit experiences performance degradation or signal drift. This helps to improve the long-term operational stability and environmental adaptability of the detection system and reduce the frequency of manual operation and calibration. 3. By combining the difference analysis of multimodal joint response distribution with modal consistency judgment logic, this invention can classify and identify abnormal events, distinguish between two types of situations: changes in real gas concentration and sensor malfunctions. This reduces false alarms or missed alarms caused by sensor failures, provides more accurate decision-making references for equipment operation and maintenance, and is suitable for long-term online monitoring needs in unattended scenarios. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method steps in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the heterogeneous three-modal sensing and detection system in an embodiment of the present invention; Figure 3 This is a simplified schematic diagram of the high-sensitivity detection method for gas decomposition products in an embodiment of the present invention; Figure 4 This is a schematic diagram of Bayesian credibility dynamic weighted fusion in an embodiment of the present invention; Figure 5 This is a bar chart showing the performance comparison between Example 1 and the comparative example in this invention. Detailed Implementation

[0015] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0016] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] Please see the appendix Figure 1 -Appendix Figure 5 This invention discloses a highly sensitive detection method for gas decomposition products that integrates multi-dimensional sensor data. The method includes the following steps: Step 1: Construct a heterogeneous three-modal sensing array that includes a chemical resistive MOS sensor array, an electrochemical amperometric sensor array, and a UV-Vis absorption spectroscopy micro-sensor. Complete the initial zero point and sensitivity calibration, and establish a benchmark library of multimodal joint response probability distribution under normal operating conditions. Step 2: Simultaneously acquire the raw response data of the three sensing modes, and obtain the dimension-matched response vectors of each mode after baseline correction, denoising and standardization preprocessing. Step 3: Calculate the mutual information between each pair of the three modes using the k-nearest neighbor probability density estimation method, and quantify the information redundancy and complementarity of different modes; Step 4: Using the maximization of mutual information between the three-modal joint response and the concentration of the gas component to be measured as the objective function, the core operating parameters of each mode are dynamically optimized through an online gradient descent algorithm; Step 5: Based on independent quantitative analysis models, each mode outputs the estimated concentration of the target gas component and the corresponding measurement uncertainty. Step 6: Dynamically update the Bayesian confidence weights of each mode through cross-modal concentration consistency verification, and fuse and output the comprehensive concentration estimate and the corresponding confidence interval; Step 7: Calculate the KL divergence between the current joint response and the baseline distribution, and, in conjunction with the consistency of the three-mode response, distinguish between two types of events: abnormal gas concentration and sensor failure. Step 8: Output the final detection results and perform adaptive rolling updates on the normal operating benchmark library.

[0018] Example 1 This embodiment is an implementation plan for the detection of SF6 insulating gas decomposition products. The detection targets are four typical SF6 discharge decomposition products: SO2, H2S, CO, and SOF2. The detection concentration range covers the trace range of 50ppb-1000ppb. The specific implementation process is as follows: The first step involved the construction and system initialization calibration of a heterogeneous three-modal sensor array. A 4-channel noble metal-doped MOS sensor array, a 3-channel electrochemical amperometric sensor array, and a miniature UV-Vis spectroscopy sensor were integrated and sealed within a single 50mL stainless steel detection chamber. The gas inlet and outlet were located on opposite sides of the chamber to ensure a uniform flow field of the analyte gas simultaneously acting on all sensing units. The initial operating parameters for each mode were set as follows: the initial operating temperature of the MOS sensor array was 300℃, with a temperature adjustment range of 200℃-400℃; the initial bias voltage of the electrochemical sensor's working electrode was 0.6V, with an adjustment range of 0.2V-1.2V; the initial integration time of the UV-Vis spectroscopy sensor was 100ms, the wavelength detection range was 200nm-400nm, and the initial light source intensity was 70% of the rated power, with an adjustment range of 30%-100%. Initial calibration was performed using standard gases prepared by the National Center for Standard Materials. Standard gas mixtures of SO2, H2S, CO, and SOF2 with concentrations of 0 ppb, 50 ppb, 200 ppb, 500 ppb, and 1000 ppb were introduced, with each concentration stabilized for 10 minutes. Response data for each mode were collected to complete zero-point calibration and sensitivity calibration. Twenty sets of multimodal response data under pure SF6 gas conditions were collected to construct a benchmark library of the joint response probability distribution under normal operation in the form of a multivariate Gaussian distribution. The benchmark library contains the mean vector and covariance matrix of 10-dimensional features.

[0019] The second step involves simultaneous acquisition and preprocessing of multimodal response data. The SF6 gas to be tested is introduced into the detection chamber at a stable flow rate of 100 mL / min. After the pressure inside the chamber stabilizes at 0.1 MPa and the temperature at 25℃, raw response data for three modes are simultaneously acquired at a uniform sampling frequency of 1 Hz. Specifically, the chemical resistance mode acquires resistance time-series data from four MOS sensors, the electrochemical mode acquires limiting diffusion current time-series data from three sensors, and the spectral mode acquires continuous absorbance spectral data in the 200 nm-400 nm band with a spectral resolution of 1 nm. The raw data are preprocessed sequentially: baseline drift correction is performed using the first-order difference method; smoothing and noise reduction are performed using a moving average filter with a window length of 5; and outlier removal is performed using the 3σ criterion. The data for each mode are then normalized using min-max standardization, mapped to the [0,1] interval, resulting in standardized response vectors of dimension 200, denoted as follows: (Chemical resistance mode) (Electrochemical modes) (Spectral modes).

[0020] The third step is cross-modal mutual information quantization calculation. The k-nearest neighbor probability density estimation method is used to calculate the pairwise mutual information between the three modes, with the nearest neighbor number k set to 6 and the sliding window length set to 200 sampling points. The calculations are performed separately. The calculated pairwise mutual information values ​​for the three modes under the initial parameters are 0.42nat, 0.31nat, and 0.27nat, respectively. Lower mutual information values ​​indicate stronger independence and better complementarity between the two modes. The electrochemical mode and the spectral mode exhibit the lowest information redundancy and the strongest complementarity.

[0021] The fourth step is dynamic parameter optimization based on maximizing joint mutual information. Taking the maximization of mutual information between the three-modal joint response and the gas component concentration vector Y as the global optimization objective, the following core objective function is constructed: ; In the formula, Indicates maximization operation; subscript This indicates that the set of working parameters is used as the optimization variable; The set of operating parameters for the three sensing modes, where For the operating parameters of the chemical resistance mode, For the operating parameters of the electrochemical mode, These are the operating parameters for the spectral modes; This represents the response vector of the chemical resistance mode under the corresponding operating parameters. This represents the response vector of the electrochemical mode under the corresponding operating parameters. This represents the response vector of the spectral mode under the corresponding operating parameters; This represents the component concentration vector of the gas to be measured; The operator for mutual information calculation is denoted by a semicolon, with the two sets of random variables before and after the semicolon representing the two sets of random variables involved in the mutual information calculation. A stochastic gradient descent algorithm is used for iterative parameter optimization, with a learning rate set to 0.02. One optimization iteration is performed every 10 consecutive measurements, with 20 iterations per iteration. The convergence threshold is set to a single-round increase in mutual information of less than 0.01 nat. After 5 rounds of optimization iterations, the operating temperature of the MOS sensor is optimized to 320℃, the bias voltage of the electrochemical sensor to 0.75V, the integration time of the spectral sensor to 150ms, and the light source intensity to 80% of the rated power. After optimization, the joint mutual information of the three modes and the concentration vector increases from the initial 1.18 nat to 1.57 nat, significantly improving information complementarity.

[0022] Step 5: Independent quantitative output of single-mode concentrations. Concentration inversion is performed using the quantitative models corresponding to each mode: For the chemical resistance mode, four channels of MOS response data are input, and a pre-trained extremely random tree regression model outputs the estimated concentration value. The model has 100 trees with a maximum depth of 12. For the electrochemical mode, the concentration value is calculated using linear interpolation based on the limiting current-concentration calibration curve. For the spectral mode, the concentration of each component is inverted using the Lambert-Beer law combined with a partial least squares regression model, with 6 latent variables. Measurement uncertainties are output simultaneously for each mode: 8% for the chemical resistance mode, 5% for the electrochemical mode, and 6% for the spectral mode. Taking 100 ppb SO2 standard gas as an example, the independent output concentrations for the three modes are 97.2 ppb, 101.3 ppb, and 99.5 ppb, with corresponding uncertainties of 7.8 ppb, 5.1 ppb, and 6.0 ppb, respectively.

[0023] Step 6: Cross-modal Bayesian confidence weighted fusion. The pairwise relative deviations of the three-modal concentration outputs are calculated. For the SO2 component, the deviation between the chemical resistance and electrochemical modes is 4.0%, the deviation between the chemical resistance and spectral modes is 2.3%, and the deviation between the electrochemical and spectral modes is 1.8%, all below the preset deviation threshold of 10%. Therefore, the initial confidence weight for each of the three modes is 1 / 3. The comprehensive concentration estimate is calculated based on the Bayesian evidence fusion framework, using the following formula: ; In the formula, This is a comprehensive concentration estimate. For the first The credibility weight of each modality For the first Concentration estimates for each mode, For the first Measurement uncertainty of each mode.

[0024] This embodiment uses specific parameters: .

[0025] Calculate the numerator: ; Calculate the denominator: ; Finally, the estimated comprehensive SO2 concentration was obtained. The corresponding 95% confidence interval is [96.2ppb, 102.6ppb].

[0026] Step 7: Classification and determination of abnormal events based on KL divergence. Calculate the KL divergence between the current multimodal joint response and the baseline distribution. In this embodiment, both the normal baseline distribution and the current response distribution follow a multivariate Gaussian distribution. The calculation formula is as follows: ; In the formula, Let be the probability density function of the current joint response, corresponding to the mean vector of the 10-dimensional features. Covariance matrix ; Let be the probability density function of the baseline distribution, corresponding to the mean vector of the 10-dimensional features. Covariance matrix .

[0027] Substituting the mean vector and covariance matrix parameters into the example provided, the result is obtained using the analytical formula for the KL divergence of a multivariable Gaussian distribution. =2.73.

[0028] An anomaly threshold of 2.0 is set. If the KL divergence is less than 2.0, it is considered a normal operating state. If the KL divergence is greater than or equal to 2.0 and the three modal responses show a consistent trend, it is considered a gas concentration anomaly, triggering a fault warning. If the KL divergence is greater than or equal to 2.0 but only a single modal response deviates, it is considered a sensor fault, triggering an operation and maintenance warning. In this embodiment, when 100ppbSO2 standard gas is introduced, the KL divergence is 2.73, and all three modal responses show an upward trend, which is determined to be a real gas concentration anomaly event.

[0029] Step 8: Result Output and Benchmark Library Update. The final detection result is output: SO2 concentration 99.4 ppb, confidence interval [96.2, 102.6] ppb, event type: gas concentration anomaly, confidence level 95%. The working parameters, response data, and confidence weights for this detection are recorded simultaneously. When five consecutive detections are determined to be in normal operating condition, the corresponding data is rolled into the benchmark library to achieve adaptive updates of the benchmark distribution. The sliding window is updated to the 100 most recent sets of normal data.

[0030] Example 2 This embodiment is a detection implementation scheme for the decomposition products of C4F7N / CO2 environmentally friendly insulating gas. The detection targets are four typical decomposition products: CF3CN, CO, CF4, and C2F6. The detection method architecture of this embodiment is completely consistent with that of Embodiment 1. It can be adapted to the new gas system only by adjusting the working parameters of each mode and the quantitative model, as detailed below: The hardware architecture of the heterogeneous trimodal sensor array remained unchanged, but the operating parameter ranges of each mode were adjusted: the operating temperature range of the MOS sensor was optimized to 220℃-380℃, with an initial value of 280℃; the bias voltage range of the electrochemical sensor was adjusted to 0.3V-1.1V, with an initial value of 0.5V; and the wavelength detection range of the spectral sensor was extended to 200nm-450nm, with an initial integration time of 120ms. Standard decomposition products of the C4F7N / CO2 mixed gas matrix were used for calibration and model training to establish a baseline library of normal operating distribution for the corresponding gas system. The dimensions of the baseline library were consistent with those in Example 1. Subsequent detection procedures were completely consistent with Example 1, including data acquisition and preprocessing, mutual information calculation, dynamic parameter optimization, single-modal quantification, Bayesian fusion, and anomaly detection steps. The learning rate for mutual information optimization was adjusted to 0.025, the deviation threshold was set to 12%, and the anomaly threshold was set to 2.2.

[0031] This embodiment verifies the portability of the method of the present invention. Without changing the core method architecture, it can be applied to the detection of decomposition products of different types of insulating gases simply by adapting the parameters, and has a wide range of engineering applications.

[0032] Comparative Example This comparative example uses the exact same experimental conditions, detection targets, and testing environment as Example 1, simulating the detection scheme of a single-mode MOS sensor array in the prior art. Specifically, it uses only a 4-channel noble metal-doped MOS sensor array, combined with a lightweight convolutional neural network for gas type identification and concentration quantification. It does not include electrochemical and spectral modes, nor does it include mutual information parameter optimization and cross-modal verification mechanisms. The specific experimental setup and comparative results are as follows: Experimental environmental conditions: The ambient temperature was kept constant at 25℃, the relative humidity was controlled at 40%±5%, the gas chamber pressure was kept stable at 0.1MPa, and the standard gas used was the same SF6 matrix mixed standard gas as in Example 1. Each concentration test point was measured repeatedly 10 times, and the average value was taken as the final result. The statistical indicators included six core indicators: component identification accuracy, quantitative detection relative error, detection limit, response time, anti-drift performance, and false alarm rate.

[0033] The test was conducted with four concentration gradients: 50 ppb, 100 ppb, 500 ppb, and 1000 ppb for a mixed gas of SO2, H2S, CO, and SOF2. The performance comparison data are as follows: Component identification accuracy: The average component identification accuracy of Example 1 was 95.2% across the entire concentration range, with an accuracy of 92.5% at a trace concentration of 50 ppb. The average identification accuracy of the comparative example across the entire concentration range was 71.8%, while the accuracy at a trace concentration of 50 ppb was only 63.3%, showing a significant difference in identification capability across the trace concentration range.

[0034] Quantitative detection accuracy: In Example 1, the average relative detection error of the four components was 3.2%, of which the relative error of SO2 was 2.9%, H2S was 3.5%, CO was 3.1%, and SOF2 was 3.3%; in the comparative example, the average relative detection error of the four components was 11.7%, with the SOF2 component having the highest error of 15.2%.

[0035] Detection limits: Example 1 shows a detection limit of 12 ppb for SO2 and 18 ppb for H2S; the comparative example shows a detection limit of 45 ppb for SO2 and 62 ppb for H2S. The detection sensitivity of this scheme is significantly improved.

[0036] Response recovery characteristics: The 90% response time of Example 1 is 42s and the recovery time is 58s; the 90% response time of the comparative example is 38s and the recovery time is 55s. The response speeds of the two are comparable. This solution improves detection performance without sacrificing response efficiency.

[0037] Anti-drift fault tolerance performance: By controlling the heating temperature deviation of the MOS sensor to simulate the 20% signal drift scenario caused by surface poisoning, Example 1 automatically identifies MOS mode anomalies through cross-modal consistency verification and reduces its confidence weight from 1 / 3 to 0.1, with an overall detection relative error of 3.5%, which can maintain stable detection accuracy; the comparative example has no verification mechanism, and the detection relative error rises to 28.7%, resulting in serious detection deviation, and the system cannot autonomously identify this drift fault.

[0038] Anomaly differentiation and false alarm rate: During the 72-hour continuous operation test, three sensor drift faults and three real gas concentration anomalies were artificially injected. Example 1 can accurately distinguish between the two types of events, with an overall false alarm rate of 1.2% and a false alarm rate of 0.8%. The comparative example cannot distinguish between sensor drift and real gas changes, with an overall false alarm rate of 14.6% and a false alarm rate of 8.3%.

[0039] Summarize In summary, this invention, through heterogeneous three-physical-modal sensing fusion and information-theory-driven parameter optimization, achieves a more than 20 percentage point improvement in component identification accuracy at trace concentrations and a reduction of approximately 60% in the detection limit compared to traditional single-modal detection schemes, significantly enhancing quantitative accuracy. It also possesses cross-modal self-diagnosis and fault tolerance capabilities, effectively suppressing detection biases caused by sensor drift and significantly reducing false alarm and false negative rates without sacrificing response speed. The method's architecture is general and transferable, adaptable to various insulating gas systems without modifying the core process, achieving a dual improvement in detection performance and operational reliability. It provides reliable technical support for online monitoring and early fault warning of insulating gases in power equipment, demonstrating significant engineering application value.

[0040] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. A highly sensitive method for detecting gas decomposition products by integrating multidimensional sensor data, characterized in that, The method includes the following steps: Step 1: Construct a heterogeneous three-modal sensing array that includes a chemical resistive MOS sensor array, an electrochemical amperometric sensor array, and a UV-Vis absorption spectroscopy micro-sensor. Complete the initial zero point and sensitivity calibration, and establish a benchmark library of multimodal joint response probability distribution under normal operating conditions. Step 2: Simultaneously acquire the raw response data of the three sensing modes, and obtain the dimension-matched response vectors of each mode after baseline correction, denoising and standardization preprocessing. Step 3: Calculate the mutual information between each pair of the three modes using the k-nearest neighbor probability density estimation method, and quantify the information redundancy and complementarity of different modes; Step 4: Using the maximization of mutual information between the three-modal joint response and the concentration of the gas component to be measured as the objective function, the core operating parameters of each mode are dynamically optimized through an online gradient descent algorithm; Step 5: Based on independent quantitative analysis models, each mode outputs the estimated concentration of the target gas component and the corresponding measurement uncertainty. Step 6: Dynamically update the Bayesian confidence weights of each mode through cross-modal concentration consistency verification, and fuse and output the comprehensive concentration estimate and the corresponding confidence interval; Step 7: Calculate the KL divergence between the current joint response and the baseline distribution, and, in conjunction with the consistency of the three-mode response, distinguish between two types of events: abnormal gas concentration and sensor failure. Step 8: Output the final detection results and perform adaptive rolling updates on the normal operating benchmark library.

2. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step one, three types of sensing units—a chemical resistance MOS sensor array, an electrochemical amperometric sensor array, and a UV-Vis absorption spectroscopy micro-sensor—are sealed and integrated into the same detection chamber. Initial zero-point calibration and sensitivity calibration of each sensing mode are performed sequentially using target gas decomposition products of known standard concentrations. Initial operating temperature of the MOS sensor, initial working electrode bias voltage of the electrochemical sensor, initial integration time of the spectral sensor, and light source intensity reference parameters are set respectively. Simultaneously, multiple sets of normal operating state response data are collected to establish a multi-modal joint response probability distribution reference library for the system's normal operating state.

3. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step two, the gas to be tested is controlled to be introduced into the detection chamber at a set stable flow rate. After the pressure and temperature parameters in the chamber stabilize, the raw response data of three sensing modes are collected synchronously at a uniform sampling frequency. The chemical resistance mode collects the time series data of resistance change, the electrochemical mode collects the time series data of limiting diffusion current, and the spectral mode collects the full-band absorbance spectrum data of the ultraviolet-visible band. The raw data of each mode are then subjected to baseline drift correction, sliding window smoothing and noise reduction, and outlier removal to obtain the dimension-matched and dimensionally standardized response vectors of each mode.

4. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step three, for the three preprocessed modal response vectors, the k-nearest neighbor probability density estimation method is used to calculate the pairwise mutual information values ​​between the chemical resistance mode and the electrochemical mode, the chemical resistance mode and the spectral mode, and the electrochemical mode and the spectral mode, respectively. The mutual information value is used to quantify the degree of information redundancy and independence complementarity between different modes. The lower the mutual information value, the lower the overlap of the detection information of the two modes and the stronger the independence complementarity. Conversely, the higher the mutual information value, the higher the information redundancy of the two modes.

5. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step four, the global optimization objective is to maximize the mutual information between the joint response of the three modes and the concentration of the gas component to be measured. The following core objective function is constructed, and the core operating parameters of each mode are iteratively adjusted using an online gradient descent algorithm. The adjusted parameters include the operating temperature of the MOS sensor, the bias voltage of the working electrode of the electrochemical sensor, the integration time of the spectral sensor, and the light source intensity. After each set number of continuous measurements, a parameter optimization iteration is performed to maintain the information complementarity of the three modes and the overall detection information in a globally optimal state. ； In the formula, Indicates maximization operation; subscript This indicates that the set of working parameters is used as the optimization variable; The set of operating parameters for three sensing modes, where For the operating parameters of the chemical resistance mode, For the operating parameters of the electrochemical mode, These are the operating parameters for the spectral modes; This represents the response vector of the chemical resistance mode under the corresponding operating parameters. This represents the response vector of the electrochemical mode under the corresponding operating parameters. This represents the response vector of the spectral mode under the corresponding operating parameters; This represents the component concentration vector of the gas to be measured; The operator is for mutual information calculation, with the two sets of random variables before and after the semicolon representing the two sets of random variables participating in the mutual information calculation.

6. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step five, based on the optimized working parameters and the standardized response vectors of each mode, the quantitative analysis models pre-trained for the corresponding modes are called to independently output the concentration estimates and corresponding measurement uncertainties of each target gas decomposition product; among them, the chemical resistance mode adopts an extremely random tree regression model, the electrochemical mode adopts the limiting current concentration calibration curve, and the spectral mode adopts the Lambert-Beer law combined with a partial least squares regression model.

7. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step six, the relative deviations of the concentration estimates of the same gas component output independently by the three modes are calculated pairwise. The Bayesian confidence weights of each mode are dynamically updated based on the magnitude of the deviations and a preset deviation threshold. When the deviation between the output of a certain mode and the other two modes exceeds the set threshold, it is determined that the mode has signal drift, surface poisoning, and / or performance degradation, and its confidence weight is automatically reduced proportionally. Finally, based on the Bayesian evidence fusion framework, the comprehensive concentration estimate of the gas component and the corresponding confidence interval are calculated by combining the real-time confidence weights of each mode and the measurement uncertainty.

8. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step seven, the KL divergence between the current three-modal joint response distribution and the normal operation distribution in the benchmark library is calculated in real time, and the abnormal event classification is determined in combination with the response consistency of the three modes. When the KL divergence exceeds the set abnormal threshold and the response change trends of the three modes are consistent, it is determined to be a real gas concentration abnormal event, corresponding to an early warning of insulation fault in power equipment. When the KL divergence exceeds the set abnormal threshold but only a single modal response is abnormal while the responses of the other two modalities are within the normal fluctuation range, it is determined to be a sensor fault event of the corresponding modality, and an operation and maintenance early warning signal for the corresponding sensor is generated.

9. The highly sensitive detection method for gas decomposition products by fusing multidimensional sensor data according to claim 1, characterized in that, In step eight, the final comprehensive concentration result of gas components, the event type determination result and the corresponding confidence level are output, and the working parameters, response data and confidence weight of each mode are recorded simultaneously. When the system is determined to be in normal operation and multiple sets of test results are stable, the current multimodal joint response data is rolled into the benchmark library, and the normal operation distribution benchmark is adaptively updated.

Images