Methane gas leakage detection method based on multi-sensor data fusion

Abstract

The invention relates to the technical field of gas detection, and discloses a methane gas leakage detection method based on multi-sensor data fusion. According to the method, an absorption intensity sequence and a wavelength drift distance are obtained through a plurality of tunable semiconductor laser absorption spectrum sensors. The method comprises the following steps: constructing a dynamic characteristic association network taking sensor positions as nodes, taking sensor data as node initial states, comparing adjacent node state vectors in a cross manner along a preset path, and calculating a similarity matrix of absorption spectral line characteristics; a sensor subset with a cooperative change pattern is identified, and a background absorption component and a potential leakage signal component are separated from the sensor subset. Performing space-time alignment and integration on the leakage signal to generate a regional fusion concentration distribution diagram, further calculating the methane concentration gradient and change rate, and judging leakage when the methane concentration gradient and change rate exceed a preset threshold value. According to the method, through dynamic network analysis and cooperative signal extraction, the accuracy and anti-interference capability of leakage detection in a complex environment are improved.

Description

Technical Field

[0001] This invention relates to the field of gas detection technology, specifically a method for detecting methane gas leaks based on multi-sensor data fusion. Background Technology

[0002] Existing methane leak detection technologies widely employ tunable semiconductor laser absorption spectroscopy sensors. These technologies primarily rely on single or a group of independently operating sensors for concentration measurement. In multi-sensor deployment scenarios, conventional data fusion methods typically involve directly averaging or weighting the concentration values ​​output by each sensor to obtain a comprehensive monitoring result. This approach treats each sensor as an independent data source, failing to fully utilize the correlation information inherent in the spatial distribution of the sensor array.

[0003] This approach of independent processing or simple fusion has its drawbacks. Because it fails to consider the spatial topological relationships between sensors and the inherent correlations between data, the system struggles to effectively distinguish between spatially transitive and cooperative signals triggered by real leakage events, and spatially discrete or non-cooperative signals generated by environmental background fluctuations or random interference. This results in insufficient sensitivity for detecting weak leakage signals in complex environments and susceptibility to false alarms due to common-mode interference, hindering accurate identification of leakage sources and reliable early warning. Therefore, a data fusion processing method is needed that can mine and utilize the dynamic correlation characteristics between sensors to improve signal recognition and anti-interference capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting methane gas leaks based on multi-sensor data fusion, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for detecting methane gas leaks based on multi-sensor data fusion, the method comprising:

[0006] The raw light absorption signals collected by multiple tunable semiconductor laser absorption spectroscopy sensors are obtained, and the absorption intensity sequence and corresponding wavelength shift of each sensor are obtained.

[0007] A dynamic feature association network is constructed with sensor locations as nodes, and the absorption intensity sequence and wavelength drift of each sensor are used as the initial state vector of the corresponding node.

[0008] In the dynamic feature association network, the state vectors of adjacent nodes are cross-compared along a preset topological path to calculate the similarity matrix of the absorption spectral features of each pair of adjacent sensors.

[0009] Based on the similarity matrix, a subset of sensors with cooperative change patterns is identified, and the background absorption component and potential leakage signal component are separated from the absorption intensity sequence corresponding to the sensor subset.

[0010] Spatiotemporal alignment and integration are performed on the potential leakage signal components to generate a fusion concentration distribution map for the region where the sensor subset is located.

[0011] Based on the fused concentration distribution map, the methane concentration gradient and concentration change rate within the coverage space of the sensor subset are calculated;

[0012] When the methane concentration gradient exceeds a preset morphological threshold, or the concentration change rate exceeds a preset dynamic threshold, a gas leak event is determined to exist within the space covered by the sensor subset.

[0013] Preferably, the step of acquiring the raw light absorption signals collected by multiple tunable semiconductor laser absorption spectral sensors, and obtaining the absorption intensity sequence and corresponding wavelength shift of each sensor, includes:

[0014] Each sensor is positioned at a different location in the monitoring space and records the corresponding laser path length information;

[0015] Temporal denoising and spectral feature extraction are performed on the raw optical absorption signal of each sensor to obtain the absorption intensity sequence and corresponding wavelength shift of each sensor.

[0016] The process of performing temporal denoising and spectral feature extraction on the raw optical absorption signal of each sensor includes:

[0017] The original optical absorption signal of each sensor is decomposed using adaptive wavelet transform to obtain signal components at different scales;

[0018] A background noise model is established based on the historical monitoring data of each sensor, and the signal components are then subjected to threshold filtering using the background noise model.

[0019] The filtered signal components are reconstructed to obtain the denoised optical absorption signal;

[0020] In the denoised optical absorption signal, the signal segment corresponding to the characteristic absorption spectrum of methane is identified by the peak detection algorithm;

[0021] The center wavelength, full width at half maximum (FWHM), and integrated absorption intensity of each identified signal segment are extracted to form the absorption intensity sequence and wavelength shift of the sensor.

[0022] Preferably, the generation of the absorption intensity sequence and wavelength shift further includes:

[0023] Based on the historical calibration data of laser operating temperature and drive current of each sensor, the extracted center wavelength is nonlinearly corrected.

[0024] The corrected center wavelength is compared with the standard absorption wavelength, and the difference obtained from the comparison is taken as the wavelength drift of this measurement.

[0025] Laser power fluctuation compensation and optical path attenuation compensation are applied to the integral absorption intensity to obtain the calibrated absorption intensity value;

[0026] The calibrated absorption intensity values ​​are arranged in chronological order to form the absorption intensity sequence of the sensor.

[0027] Preferably, in the dynamic feature association network, the step of cross-comparing the state vectors of adjacent nodes along a preset topological path includes:

[0028] Define the connection rules between nodes in the dynamic feature association network, which are determined based on the actual physical distance between sensors and the quality of the communication link.

[0029] Each connection is assigned an initial weight, which is inversely proportional to the physical distance between the corresponding sensor pairs;

[0030] Within each fusion calculation cycle, the statistical characteristics of the absorption intensity sequence within the current time window are synchronously exchanged along connected node pairs.

[0031] Calculate the cross-correlation function of the absorption intensity sequences between node pairs, and use the peak value of the cross-correlation function as the temporal similarity;

[0032] Calculate the Euclidean distance between node pairs based on wavelength drift, and use the reciprocal of the Euclidean distance as the frequency domain similarity.

[0033] The time-domain similarity and frequency-domain similarity are combined and multiplied by the initial weight to generate the element value at the corresponding position in the similarity matrix.

[0034] Preferably, identifying the subset of sensors with cooperative change patterns based on the similarity matrix includes:

[0035] Spectral clustering analysis is performed on the similarity matrix to divide the sensor nodes into several candidate subsets;

[0036] Calculate the cohesion of each candidate subset, whereby cohesion is defined as the average similarity value between all node pairs within the subset;

[0037] Candidate subsets with cohesion higher than a preset subset threshold are marked as high-cohesion subsets;

[0038] For each highly cohesive subset, analyze the consistency of the direction of change of the absorption intensity sequence of all sensors within it at the same time, and calculate the direction consistency ratio;

[0039] A subset of sensors with a high cohesion rate exceeding a preset consistency threshold is selected and identified as a subset of sensors with a cooperative change pattern.

[0040] Preferably, separating the background absorption component and the potential leakage signal component from the absorption intensity sequence corresponding to the subset of sensors includes:

[0041] For each sensor in the aforementioned sensor subset, its absorption intensity sequence is input into a fully trained variational mode decomposition model.

[0042] The variational mode decomposition model automatically decomposes the input sequence into multiple intrinsic mode functions;

[0043] Based on the center frequency and energy distribution of each intrinsic mode function, they are classified into a background mode set and a signal mode set;

[0044] The background absorption component of the sensor is reconstructed by superimposing all intrinsic mode functions in the background mode set.

[0045] The potential leakage signal component of the sensor is obtained by subtracting the background absorption component from the original absorption intensity sequence of the sensor.

[0046] The potential leakage signal components of all sensors in the sensor subset are summarized to form a subset of potential leakage signals.

[0047] Preferably, the spatiotemporal alignment and integration of the potential leakage signal components includes:

[0048] A local spatial coordinate system is established with the geometric center of the sensor subset as the reference origin;

[0049] Based on the position of each sensor in the local spatial coordinate system and its laser path length, calculate the spatial effective weight of the potential leakage signal component of each sensor;

[0050] Based on a high-precision time synchronization signal, all potential leakage signal components of the sensors are aligned to a unified time scale;

[0051] In the local spatial coordinate system, the position of each sensor is taken as a discrete point, and the value of the potential leakage signal component after spatiotemporal alignment is multiplied by the effective spatial weight is taken as the concentration contribution value of the discrete point.

[0052] The concentration contribution value in the local spatial coordinate system is interpolated using the Kriging space interpolation algorithm to generate a continuous concentration field distribution.

[0053] The concentration field distribution is integrated over a preset time interval to obtain the fusion concentration distribution map.

[0054] Preferably, the step of calculating the methane concentration gradient and concentration change rate within the coverage space of the sensor subset based on the fused concentration distribution map includes:

[0055] On the fusion concentration distribution map, calculate the partial derivative of the concentration at each spatial point along the spatial coordinate axis;

[0056] The magnitude of the partial derivatives in each direction is taken as the methane concentration gradient at the spatial point.

[0057] The time series of the fusion concentration distribution map is obtained at consecutive timestamps;

[0058] Calculate the difference between adjacent time stamps in the fusion concentration distribution map;

[0059] The average absolute value of the difference map within the spatial range covered by the sensor subset is taken as the methane concentration change rate within the time period corresponding to the adjacent timestamps.

[0060] Preferably, the step of establishing a background noise model based on historical monitoring data of each sensor and using the background noise model to perform threshold filtering on the signal components is achieved through the following steps:

[0061] Collect the historical light absorption signal of each sensor under a leak-free state to form the background signal sample set of the sensor;

[0062] The background signal sample set is subjected to the same adaptive wavelet transform as the original optical absorption signal to decompose the background signal components at different scales.

[0063] For each scale, calculate the amplitude probability distribution of all background signal components at that scale, and set the upper limit of the amplitude in the preset high probability interval as the noise threshold of that scale.

[0064] When performing threshold filtering on the signal components of the corresponding scale obtained by decomposing the original optical absorption signal, the portion of the signal component with an amplitude lower than the noise threshold of the scale is set to zero, while the portion with an amplitude higher than or equal to the noise threshold of the scale is retained.

[0065] Preferably, assigning an initial weight to each connection, wherein the initial weight is inversely proportional to the physical distance between the corresponding sensor pairs, is achieved through the following steps:

[0066] Obtain the actual physical distance between each pair of connected sensor nodes in the dynamic feature association network;

[0067] For each pair of sensor nodes, their physical distance is input into a preset weight allocation function, which is a monotonically decreasing function, such that the output initial weight value is inversely proportional to the input actual physical distance value.

[0068] The calculated initial weight values ​​are assigned to the connections between the corresponding sensor nodes;

[0069] In subsequent calculation cycles of the dynamic feature association network, the initial weight value remains unchanged, or the initial weight value is dynamically adjusted based on the real-time monitoring results of the communication link quality between sensor nodes.

[0070] Compared with the prior art, the beneficial effects of the present invention are:

[0071] A dynamic feature association network is constructed with sensor locations as nodes. The absorption intensity sequence and wavelength drift of each sensor are used as the initial state vector of each node, and cross-comparisons are performed along a pre-defined topological path. This method organizes discrete sensor data into a spatially related graph model, allowing the data fusion process to explicitly utilize the physical layout and connectivity between sensors. Based on the network structure, cross-comparisons can capture consistent patterns of leak signal propagation in a spatial dimension, thereby more accurately delineating the area affected by the leak, improving the ability to identify leak events from spatial distribution features, and enhancing the spatial directionality of leak source location.

[0072] Based on the similarity matrix generated by comparing sensor nodes, a subset of sensors with cooperative variation patterns is identified. Furthermore, background absorption components and potential leakage signal components are separated from the absorption intensity sequence of this subset. This method identifies true leakage signals by mining the correlation features between sensor groups, narrowing the analytical focus from all sensors to a highly correlated subset. The signal separation technique effectively removes common-mode background interference caused by environmental factors, highlighting the specific signals generated by leakage events. This makes the extraction of weak leakage signals more sensitive and reliable in complex environmental noise backgrounds, reducing misjudgments caused by background fluctuations or isolated interference. Attached Figure Description

[0073] Figure 1 This is a schematic diagram illustrating the working principle of the methane gas leak detection method based on multi-sensor data fusion described in this invention.

[0074] Figure 2 A flowchart for the post-processing of the absorption intensity sequence and wavelength shift;

[0075] Figure 3 A flowchart illustrating the construction of a dynamic feature association network and the calculation of the similarity matrix;

[0076] Figure 4 This represents a subset of the sensor region representing the spatial distribution of methane concentration gradients.

[0077] Figure 5 This is a time series graph showing the real-time monitoring of methane concentration gradient and rate of change. Detailed Implementation

[0078] 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.

[0079] Please see Figure 1 This invention provides a method for detecting methane gas leaks based on multi-sensor data fusion. The method includes: acquiring raw optical absorption signals collected by multiple tunable semiconductor laser absorption spectroscopy sensors deployed at different locations in the monitoring space; processing the signals to obtain an independent absorption intensity sequence and corresponding wavelength drift for each sensor; constructing a dynamic feature association network based on the physical layout and communication relationships between the sensors, using the location of each sensor as a node; and using the absorption intensity sequence and wavelength drift of each sensor as the initial state vector of its corresponding node; cross-comparing the state vectors of adjacent nodes along a preset topological path in the network; calculating the similarity of each pair of adjacent sensors in absorption spectral features; and forming a similarity matrix. Based on this matrix, one or more sensor subsets with highly coordinated absorption signal change trends can be identified. From the absorption intensity sequences corresponding to these identified sensor subsets, a signal decomposition technique is used to separate the prevalent background absorption component and the potential leakage signal component that may be caused by the leak. These potential leakage signal components from different sensors are time-synchronized and spatially aligned, and an integral operation is used to generate a fused concentration distribution map specifically for the area covered by the sensor subset. Based on this distribution map, the spatial gradient of methane concentration and the rate of change of concentration over time within the region are calculated. When the calculated concentration gradient exceeds a preset morphological threshold, or the rate of change of concentration exceeds a preset dynamic threshold, a methane gas leak event is determined to have occurred within the spatial area covered by this subset of sensors.

[0080] In one embodiment of the present invention, see [reference] Figure 2Multiple tunable semiconductor laser absorption spectroscopy sensors are configured at different locations in the monitoring space, and the laser path length information between the laser emitter and receiver of each sensor is accurately recorded. For the raw optical absorption signal acquired by each sensor, time-domain denoising and spectral feature extraction are performed to obtain usable absorption intensity sequences and wavelength shifts. Time-domain denoising and spectral feature extraction are achieved through the following steps: The raw optical absorption signal is decomposed using adaptive wavelet transform to obtain a series of signal components at different scales. A dedicated background noise model is established based on historical data obtained from long-term monitoring of each sensor under leak-free conditions. This model is then used to perform threshold filtering on the signal components at each scale to suppress noise. The retained signal components after filtering are reconstructed to obtain the denoised optical absorption signal. In the denoised signal, a peak detection algorithm is used to accurately locate the signal segment corresponding to the characteristic absorption spectral line of methane molecules. For each identified signal segment, key features such as its center wavelength, full width at half maximum (FWHM), and integrated absorption intensity are extracted. The generation of the absorption intensity sequence and wavelength shift further includes a calibration step. Based on the historical calibration data of the laser's operating temperature and drive current for each sensor, the extracted center wavelength is nonlinearly corrected to compensate for the effects of factors such as device temperature drift. The corrected center wavelength is compared with the center wavelength of the known methane standard absorption spectrum; the difference between the two is the wavelength drift corresponding to this measurement. The extracted integrated absorption intensity needs to be compensated for laser source output power fluctuations and optical path attenuation based on the known laser path length, thus obtaining calibrated and comparable absorption intensity values. The absorption intensity values ​​acquired and calibrated in chronological order are arranged sequentially to form the absorption intensity sequence of the sensor.

[0081] In practice, data acquisition and preliminary processing are achieved through the following process: Multiple tunable semiconductor laser absorption spectrometers are deployed in the area to be monitored, such as near different sections of a natural gas pipeline. Each tunable semiconductor laser absorption spectrometer's laser emitter and photodetector form an independent measurement optical path, and the laser path length is pre-measured and recorded in the system. Each tunable semiconductor laser absorption spectrometer continuously acquires the raw light absorption signal, which is a time-varying voltage or current sequence reflecting the intensity attenuation of the laser after passing through the measured gas area.

[0082] In some embodiments, temporal denoising and spectral feature extraction of the original optical absorption signal are the core preprocessing steps. The system performs adaptive wavelet transform on the original optical absorption signal of each tunable semiconductor laser absorption spectroscopy sensor. The adaptive wavelet transform automatically selects the optimal wavelet basis and decomposition level based on the local features of the original optical absorption signal, decomposing the original optical absorption signal into wavelet coefficient components of multiple different scales. Based on the historical optical absorption signal data accumulated by each tunable semiconductor laser absorption spectroscopy sensor under known leakage-free conditions over a long period of time, the system independently establishes a background noise model for each tunable semiconductor laser absorption spectroscopy sensor. The background noise model statistically analyzes the amplitude distribution characteristics of wavelet coefficients at each scale under leakage-free conditions. Using the background noise model, threshold filtering is performed on the wavelet coefficient components of each scale obtained in real time decomposition, setting coefficients with amplitudes below the corresponding scale's preset noise threshold to zero, while retaining coefficients above the threshold. All scale wavelet coefficient components after threshold filtering are reconstructed to obtain the denoised optical absorption signal. The components related to random noise in the denoised optical absorption signal are significantly suppressed. In the denoised light absorption signal, the signal waveform is scanned by a preset peak detection algorithm to identify signal segments with obvious signal amplitude protrusions that match the standard absorption curve of methane molecules. Each identified signal segment corresponds to a complete gas absorption scan process.

[0083] In practice, the center wavelength, full width at half maximum (FWHM), and integrated absorption intensity characteristics of each identified signal segment are extracted. The center wavelength is determined by fitting the peak position of the signal segment, the FWHM is determined by calculating the width corresponding to half the peak height, and the integrated absorption intensity is obtained by numerically integrating the area under the curve of the entire signal segment. These extracted raw features need to be calibrated to eliminate systematic errors. Based on the historical calibration data of operating temperature and drive current obtained during the factory and periodic calibration of each tunable semiconductor laser absorption spectroscopy sensor, a nonlinear correction function for the center wavelength is established to compensate for the extracted center wavelength. The nonlinear correction function can be in the form of a polynomial or a lookup table, used to describe the combined effect of temperature and current fluctuations on the laser output center wavelength. The corrected center wavelength is compared with the standard absorption wavelength of methane at a specific pressure and temperature stored in the database, and the difference between the two is recorded as the wavelength drift of this measurement. The wavelength drift reflects the frequency stability of the laser itself and the slight influence of environmental pressure on the absorption line position. For the extracted integrated absorption intensity, laser power fluctuation compensation and optical path attenuation compensation are required. Laser power fluctuation compensation is achieved by monitoring the real-time output power of the laser and normalizing the absorption intensity according to the Beer-Lambert law. Optical path attenuation compensation, based on pre-recorded laser path length information, corrects the absorption intensity inversely to the path length, ensuring comparability of absorption intensity data from tunable semiconductor laser absorption spectrometers at different locations and with different optical paths. The fully calibrated absorption intensity values ​​are arranged chronologically according to acquisition time, forming the absorption intensity sequence of the tunable semiconductor laser absorption spectrometer within the current monitoring period. Simultaneously, the wavelength drift corresponding to each measurement is also recorded for subsequent frequency domain characteristic analysis.

[0084] In one embodiment of the present invention, see [reference] Figure 3In dynamic feature association networks, the connection rules between nodes are determined based on the actual physical distance between sensors and the quality of the communication link. Each connection in the network is assigned an initial weight, which is inversely proportional to the actual physical distance between the pair of sensor nodes it connects to. The specific assignment of the initial weight is achieved through the following steps: Obtain the actual physical distance data between each pair of connected sensor nodes in the network. For each pair of sensor nodes, input its physical distance value into a preset weight allocation function, which is designed to be a monotonically decreasing function, ensuring that the output initial weight value is inversely proportional to the input distance value. The calculated initial weight value is assigned to the corresponding connection. During subsequent network operation, this weight value can remain unchanged or be dynamically fine-tuned based on real-time monitoring results of the communication link quality between sensor nodes. Within each set fusion calculation cycle, along all connected node pairs in the network, synchronously exchange the statistical characteristics of the absorption intensity sequence of the sensors represented by the nodes within the current time window. Calculate the cross-correlation function of the absorption intensity sequences between each pair of nodes, and use the peak value of this cross-correlation function as a measure of the temporal similarity between the two pairs to measure the consistency of their temporal changes. Simultaneously, the Euclidean distance of wavelength drift between each pair of nodes is calculated, and the reciprocal of this Euclidean distance is used as the frequency domain similarity to measure the closeness of their frequency domain states. The calculated time domain similarity and frequency domain similarity are combined and multiplied by the initial weights of the connections corresponding to the node pair, and the result is the element value at the corresponding position in the similarity matrix.

[0085] In practical implementation, the extraction of inter-sensor correlation features is achieved by constructing and calculating a dynamic feature correlation network. The construction of the dynamic feature correlation network uses each tunable semiconductor laser absorption spectroscopy (TSLS) sensor deployed within the monitoring area as a node. The connection rules between nodes are determined jointly based on the actual physical distance between the TSLS sensors and the quality of the communication link used for data exchange. In a specific natural gas station monitoring example, twelve TSLS sensors are installed around different process units, and the system records the three-dimensional coordinates of each sensor. The communication link quality is assessed through periodic packet transmission and reception tests, using packet loss rate and latency as quantitative indicators. In practical implementation, each established connection in the dynamic feature correlation network is assigned an initial weight, which is inversely proportional to the actual physical distance between the pair of TSLS sensor nodes it connects to. The specific assignment of the initial weight is achieved through the following steps: the system obtains the actual physical distance between each pair of connected TSLS sensor nodes in the dynamic feature correlation network. For each pair of tunable semiconductor laser absorption spectroscopy sensor nodes, their physical distance data is input into a preset weighting function, which is designed to be a monotonically decreasing function. One possible form of the weighting function is:

[0086]

[0087] in: This represents the initial weight of the edge connecting node i and node j. This represents the actual physical distance between node i and node j. and A preset constant greater than zero is used to adjust the curve shape of the weight decay with distance. The weight assignment function ensures the initial weight values ​​of the output. Physical distance value from input The inverse relationship is maintained. The calculated initial weight values ​​are assigned to the connections between the corresponding tunable semiconductor laser absorption spectroscopy sensor nodes. In some embodiments, the initial weight values ​​remain unchanged in subsequent calculation cycles of the dynamic feature association network. In other embodiments, the initial weight values ​​are dynamically adjusted based on real-time monitoring results of the communication link quality between the tunable semiconductor laser absorption spectroscopy sensor nodes; when the communication quality deteriorates, the weight value of the corresponding connection is appropriately reduced.

[0088] In practical implementation, the dynamic feature association network performs state comparison and similarity calculation within each preset fusion calculation cycle. Within each fusion calculation cycle, along all connected node pairs in the network, the system synchronously exchanges the statistical characteristics of the absorption intensity sequence of the tunable semiconductor laser absorption spectroscopy sensor represented by each node within the current time window. These statistical characteristics include mean, variance, and extreme values. The cross-correlation function of the absorption intensity sequences between each pair of nodes is calculated. This cross-correlation function measures the similarity between two time series at different time offsets, and the maximum peak value of the cross-correlation function across all time offsets is taken as the temporal similarity of the node pair. The Euclidean distance between the wavelength drift sequences recorded between each pair of nodes within the same time window is calculated. The Euclidean distance directly calculates the straight-line distance between two vectors in space, and the reciprocal of the calculated Euclidean distance is taken as the frequency domain similarity of the node pair. It can be understood that temporal similarity and frequency domain similarity reflect the correlation between sensor data from two dimensions: the trend of signal amplitude variation and the stability of light source frequency, respectively. Finally, the time-domain similarity and frequency-domain similarity are combined, and the combined result is multiplied by the initial weights of the connections corresponding to the node pair. The resulting value is the element value at the corresponding position in the similarity matrix. The similarity matrix is ​​a symmetric matrix whose dimension is equal to the number of tunable semiconductor laser absorption spectroscopy sensors in the network. Each element value in the matrix quantitatively describes the overall state similarity of a corresponding pair of sensor nodes in the latest monitoring period.

[0089] In one embodiment of the present invention, based on the calculated similarity matrix, all sensor nodes are divided into several candidate subsets through spectral clustering analysis. The cohesion of each candidate subset is calculated, defined as the average value of the corresponding values ​​in the similarity matrix among all possible node pairs within the subset. Candidate subsets with cohesion higher than a preset subset threshold are marked as high-cohesion subsets. For each high-cohesion subset, it is further analyzed whether the change direction of the absorption intensity sequences of all sensors within it is consistent at the same time, and the proportion of data points with consistent directions is calculated. High-cohesion subsets with a consistent direction ratio exceeding a preset consistency threshold are selected and finally confirmed as sensor subsets with cooperative change patterns. For each sensor in the confirmed sensor subset, its absorption intensity sequence is input into a pre-trained variational mode decomposition model. This variational mode decomposition model can automatically decompose the input absorption intensity sequence into several intrinsic mode functions with different center frequencies. Based on the center frequency and energy distribution characteristics of each intrinsic mode function, it is classified into a background mode set or a signal mode set. By superimposing all intrinsic mode functions categorized into the background mode set, the background absorption component of the sensor in the current environment can be reconstructed. Subtracting the reconstructed background absorption component from the original absorption intensity sequence of the sensor yields the potential leakage signal component. Summarizing the potential leakage signal components calculated from all sensors in this subset forms a subset-level potential leakage signal set for subsequent fusion analysis.

[0090] In practical implementation, collaborative sensor groups are identified and signal separation is performed from a dynamic feature association network. The dynamic feature association network identifies sensor subsets based on a similarity matrix, where the dimension of the similarity matrix corresponds to the number of tunable semiconductor laser absorption spectroscopy (TSLS) sensors in the network. In a specific example of monitoring a pump area in a natural gas processing plant, eight TSLS sensors are deployed around the pump group. The system calculates an 8×8 similarity matrix, where each element represents the overall similarity between a pair of TSLS sensor nodes in the time and frequency domains. Spectral clustering analysis is performed on the similarity matrix. This analysis treats the similarity matrix as an adjacency matrix of a graph structure. By calculating the Laplacian matrix of the graph and clustering its eigenvectors, all eight TSLS sensor nodes are divided into several candidate subsets, such as three.

[0091] In practice, the cohesion of each candidate subset is calculated. Cohesion is defined as the arithmetic mean of the corresponding element values ​​in the similarity matrix among all possible pairs of nodes within the subset. The formula for calculating cohesion can be expressed as:

[0092]

[0093] in: Represents candidate subsets cohesion, Represents candidate subsets The number of tunable semiconductor laser absorption spectroscopy sensor nodes included. This indicates the nodes in the similarity matrix. With nodes The similarity value is used to sum the results and iterate through the candidate subsets. The system identifies all unique node pairs within a subset. The cohesion of all candidate subsets is compared to a preset subset threshold; subsets with cohesion exceeding this threshold are designated as high-cohesion subsets. For each high-cohesion subset, the consistency of the absorption intensity sequences of all tunable semiconductor laser absorption spectroscopy sensors within it at the same time point is further analyzed. The direction of change is determined by calculating the difference in absorption intensity values ​​between adjacent sampling points. The system counts the number of data points within a high-cohesion subset whose absorption intensity sequences of all tunable semiconductor laser absorption spectroscopy sensors exhibit identical directions of change within an analysis time window. This number is then divided by the total number of sampling points within the time window to obtain the direction consistency ratio. High-cohesion subsets with direction consistency ratios exceeding a preset consistency threshold are selected and ultimately identified as sensor subsets with a cooperative change pattern. This cooperative change pattern indicates that these tunable semiconductor laser absorption spectroscopy sensors may have been collectively affected by the diffusion of the same gas source.

[0094] In some embodiments, for each tunable semiconductor laser absorption spectroscopy sensor in the confirmed subset of sensors with cooperative variation modes, its absorption intensity sequence is input into a pre-trained variational mode decomposition (VMD) model. VMD is an adaptive signal decomposition method that decomposes the input absorption intensity sequence into several narrowband intrinsic mode functions (EMFs) with specific center frequencies. The trained VMD model has optimized its hyperparameters, such as the number of decomposition layers and the penalty factor, using historical data. The VMD model automatically decomposes the input absorption intensity sequence of the tunable semiconductor laser absorption spectroscopy sensor into multiple EMFs. Based on the center frequency and energy distribution characteristics of each EMF obtained, the system classifies it into either the background mode set or the signal mode set. The classification rules are based on prior knowledge; for example, EMFs with extremely low center frequencies and dominant energy are classified into the background mode set, while EMFs with center frequencies within a certain range and exhibiting transient or periodic energy fluctuations are classified into the signal mode set. By linearly superimposing all intrinsic mode functions classified into the background mode set, the background absorption component of the tunable semiconductor laser absorption spectrometer within the current monitoring period is reconstructed. The reconstructed background absorption component is subtracted from the original absorption intensity sequence of the tunable semiconductor laser absorption spectrometer, and the resulting difference sequence is the potential leakage signal component of the tunable semiconductor laser absorption spectrometer. The potential leakage signal components calculated from all tunable semiconductor laser absorption spectrometers in the subset of sensors with cooperative variation modes are summarized to form a subset of potential leakage signal sets. This subset of potential leakage signal sets contains multiple temporally synchronized and spatially correlated potential leakage signal sequences, which are used for subsequent spatiotemporal fusion and concentration field reconstruction.

[0095] In one embodiment of the present invention, a local spatial coordinate system is established with the geometric center point of the identified subset of sensors as the reference origin. Based on the precise position coordinates of each sensor in this local spatial coordinate system and its corresponding laser path length, the effective weight of the potential leakage signal component of each sensor in spatial fusion is calculated. Based on a high-precision time synchronization signal, the timestamps of the potential leakage signal components of all sensors are aligned to a unified time scale. In this local spatial coordinate system, the position of each sensor is considered a discrete spatial point. The value of the potential leakage signal component at this point, after spatiotemporal alignment, is multiplied by its spatial effective weight to obtain the contribution value of this discrete point to the spatial concentration field. A Kriging spatial interpolation algorithm is used to interpolate the concentration contribution values ​​at these discrete spatial points, generating a continuous spatial concentration field distribution covering the local area. This spatial concentration field distribution is integrated over a preset time interval to obtain a fused concentration distribution map reflecting the cumulative concentration during that time period. Based on the generated fused concentration distribution map, the partial derivative of the concentration at each spatial point on the map is calculated along the spatial coordinate axis. The magnitude of the partial derivative in each direction is used as the methane concentration gradient value at that spatial point. At consecutive time stamps, sequentially generated fused concentration distribution maps are acquired to form a time series. The difference map between two fused concentration distribution maps corresponding to adjacent time stamps is calculated. The absolute value of this difference map is averaged over the entire spatial range covered by the sensor subset; this average value is used as the rate of change of methane concentration within the time period corresponding to these two adjacent time stamps.

[0096] In practical implementation, potential leakage signals undergo spatial reconstruction and dynamic parameter calculation. After identifying a subset of sensors with cooperative variation modes, the system establishes a two-dimensional or three-dimensional local spatial coordinate system with the geometric center point of the sensor subset as the origin. The axis of the local spatial coordinate system is aligned with the actual orientation of the monitoring area. Based on the precise position coordinates of each tunable semiconductor laser absorption spectroscopy sensor in the local spatial coordinate system and its pre-recorded laser path length information, the spatial effective weight of the potential leakage signal component of each tunable semiconductor laser absorption spectroscopy sensor is calculated. The spatial effective weight reflects the proportion of the sensor measurement value in the spatial interpolation. Based on a high-precision time synchronization protocol in the network, the timestamps of the potential leakage signal components of all tunable semiconductor laser absorption spectroscopy sensors are aligned to a unified time scale to ensure data temporal consistency.

[0097] In the local spatial coordinate system, the position of each tunable semiconductor laser absorption spectroscopy sensor is considered as a discrete spatial point. The value of the potential leakage signal component corresponding to each discrete spatial point, after spatiotemporal alignment, is multiplied by its spatial effective weight. The resulting product is the concentration contribution value of that discrete point to the final concentration field. Using the Kriging spatial interpolation algorithm, with these discrete spatial points and their concentration contribution values ​​as known conditions, the concentration values ​​at unknown locations within the local spatial coordinate system are optimally estimated without bias, generating a continuous spatial concentration field distribution covering the region where the sensor subset is located. This spatial concentration field distribution is integrated over a preset time interval, such as the past 60 seconds. The integration operation reflects the concentration accumulation effect over time, ultimately generating a fused concentration distribution map, which is a concentration value matrix on a two-dimensional spatial grid. Refer to Table 1, which shows the coordinates of a subset containing four sensors in the local spatial coordinate system, the calculated spatial effective weights, and the concentration contribution value at a single sampling time.

[0098] Table 1: Sensor Spatial Coordinates, Effective Weights, and Concentration Contribution Values

[0099] Sensor number X coordinate (meters) Y coordinate (meters) Laser path length (meters) Spatial effective weights Potential leakage signal component (AU) Concentration contribution (AU·m) S05 -2.1 1.5 10.0 0.28 0.015 0.0042 S08 0.0 0.0 8.5 0.24 0.022 0.0053 S11 1.8 -1.0 9.0 0.25 0.018 0.0045 S12 -0.5 2.0 9.5 0.23 0.012 0.0028

[0100] Note: AU is an arbitrary absorption unit. The effective spatial weight is calculated by normalization based on the sensor's laser path length and relative position.

[0101] In practice, based on the generated fused concentration distribution map, the methane concentration gradient within the sensor subset's coverage area is calculated. On the two-dimensional concentration field grid represented by the fused concentration distribution map, the partial derivatives of the concentration value at each grid point are calculated along the X and Y axes of the spatial coordinate system. and The partial derivatives are obtained by central difference method or spline interpolation. The magnitudes of the partial derivatives in both directions at each grid point are taken as the methane concentration gradient at that point. The calculation relationship is as follows:

[0102]

[0103] in: This represents the methane concentration gradient value. This represents the concentration value at a point on the fusion concentration distribution map. This represents the partial derivative of concentration in the X direction. This represents the partial derivative of the concentration in the Y direction. It can be understood as the concentration gradient value. It characterizes the degree and direction of concentration changes at various points in space.

[0104] In some embodiments, the system runs continuously and periodically generates fused concentration distribution maps, forming a time series of fused concentration distribution maps arranged chronologically. A difference map is calculated between two fused concentration distribution maps corresponding to adjacent time stamps. This difference map is obtained by subtracting the concentration value of the corresponding grid at the previous time stamp from the concentration value at the next time stamp point by point. The average absolute value of the resulting difference map over the entire spatial range covered by the sensor subset is taken as the rate of change of methane concentration within the time period corresponding to these two adjacent time stamps. (Methane concentration change rate) The discrete calculation formula is:

[0105]

[0106] in: Indicates the rate of change in methane concentration. This represents the total number of grid points in the fusion concentration distribution map. Indicates time The fusion concentration distribution map on the first grid points Concentration value at that location, Indicates time The fusion concentration distribution map on the first grid points The concentration value at that location. Optionally, when the fusion concentration distribution map is generated via kriging interpolation, This represents the number of regular sampling points used to assess the rate of concentration change. Concentration gradient. With the rate of change of concentration The calculation results provide core dynamic morphological parameters for subsequent determination of gas leak events.

[0107] See Figure 4 The presented implementation results correspond to the output of the "concentration gradient calculation" stage in the methane leak detection process based on multi-sensor data fusion. The figure uses a two-dimensional spatial grid to show the methane concentration gradient distribution within the sensor subset's coverage area. The color scale corresponds to the gradient value (unit: AU·m⁻¹), and red triangles mark the locations of the sensors involved in the fusion. Specifically, the generation of this figure is based on the following process: a local coordinate system is established with the geometric center of the identified sensor subset as the origin; the concentration contribution values ​​of the sensors are spatially reconstructed using Kriging interpolation to form a continuous concentration field; then, the partial derivatives of the concentration at each grid point are calculated along the X and Y axes, and the magnitude of the partial derivative is used as the concentration gradient value at that point (i.e., the quantization result corresponding to the colors in the figure). The yellow area in the figure corresponds to a region with a high gradient value (approximately 0.0045 AU·m⁻¹), reflecting a more drastic spatial change in concentration at that location; the purple area has a gradient value close to 0, representing a more gradual concentration distribution. The spatial correlation between the sensor location and the high gradient region can help locate the relative position of potential leak sources.

[0108] In one embodiment of the present invention, the background noise model is established through the following steps: Historical light absorption signals obtained from long-term monitoring of each sensor under conditions confirming no methane leakage are collected, forming a background signal sample set for the sensor. This background signal sample set is decomposed using the same adaptive wavelet transform process as the original light absorption signal, yielding background signal components at different scales. For each decomposition scale of the wavelet transform, the amplitude probability distribution of all background signal components at that scale is calculated, and the upper limit of the amplitude falling within a preset high probability interval is set as the noise threshold corresponding to the signal component at that scale. When denoising the original light absorption signal acquired by the sensor in real time, the original signal, after being decomposed by the adaptive wavelet transform, yields signal components at each scale. For each signal component at each scale, its amplitude is compared with the noise threshold preset according to the background noise model for that scale. Amplitudes below the noise threshold are set to zero, while amplitudes above or equal to the noise threshold are retained, thus completing threshold filtering based on the background noise model.

[0109] In practical implementation, each tunable semiconductor laser absorption spectroscopy sensor establishes its own background noise model and applies it to real-time signal filtering. The background noise model is established based on historical light absorption signals collected under conditions where no methane leakage is confirmed. In a specific example of natural gas compressor room monitoring, six tunable semiconductor laser absorption spectroscopy sensors are installed at different heights and orientations in the room. Under leak-free conditions with the compressor normally shut down and good ventilation, the system continuously collects historical light absorption signals from each tunable semiconductor laser absorption spectroscopy sensor for hundreds of hours, forming a background signal sample set for each sensor. The background signal sample set covers various background conditions such as ambient temperature fluctuations, equipment vibration, interference from other non-methane gases, and slight drift of the laser itself.

[0110] In some embodiments, historical optical absorption signals in the background signal sample set are processed to establish a background noise model. The system decomposes each historical optical absorption signal in the background signal sample set using the same adaptive wavelet transform process as when processing the real-time raw optical absorption signal. The adaptive wavelet transform decomposes each historical optical absorption signal into background signal components of the same series of scales. For each decomposition scale of the adaptive wavelet transform, the background signal components corresponding to all signals in the background signal sample set at that scale are collected, and the amplitude probability distribution of these background signal components is calculated. It can be understood that the amplitude probability distribution describes the statistical characteristics of the signal component amplitude at that scale under pure background conditions. The upper limit of the amplitude falling within a preset high probability interval is set as the noise threshold corresponding to that scale. The preset high probability interval can be, for example, set as the amplitude range corresponding to the cumulative probability distribution from 0 to 95%. A formulaic expression for setting the noise threshold is as follows:

[0111]

[0112] in: Indicates the first Noise threshold at each decomposition scale Indicates the first The cumulative amplitude distribution function of background signal components at each scale. It is the inverse function of the cumulative distribution function. This is a preset high probability value (e.g., 0.95). This formula indicates the noise threshold. Set to make the amplitude less than or equal to The background signal component accounts for the proportion of the total. .

[0113] In practice, a pre-established background noise model is used to perform threshold filtering on the real-time acquired raw optical absorption signal. After the system acquires the raw optical absorption signal from the tunable semiconductor laser absorption spectroscopy sensor, it first performs the same adaptive wavelet transform on the raw optical absorption signal, decomposing it into real-time signal components of the same multiple scales. For each scale of the real-time signal component, its amplitude is compared with the noise threshold preset for that scale in the background noise model. Compare the components. Components with amplitudes below the corresponding scale noise threshold in the real-time signal are included. Partially zeroing out the noise, while retaining amplitudes higher than or equal to the corresponding scale noise threshold. This process, known as threshold filtering based on a background noise model, involves reconstructing the retained signal components at each scale after threshold filtering. The resulting denoised optical absorption signal is then used for subsequent spectral feature extraction. Optionally, the background noise model is not static; the system can periodically collect background signal samples and update the model when the monitoring environment changes significantly, ensuring the noise threshold remains adaptable to the current background. This threshold filtering based on historical background statistics effectively suppresses steady-state or slowly varying noise similar to historical background characteristics, thereby enhancing the detection sensitivity for sudden leaks.

[0114] See Figure 5 In the corresponding methane leak detection process, the core is to perform real-time time-series analysis of the methane concentration gradient and rate of change within the sensor subset coverage area based on the fused concentration distribution map. The figure uses time (s) as the horizontal axis, simultaneously displaying two core monitoring indicators: the left vertical axis represents the concentration gradient (unit: ppm / m³), and the right vertical axis represents the concentration change rate (unit: ppm / s); the gradient threshold (5.0 ppm / m³) and the change rate threshold (2.0 ppm / s) are also marked, and the leak event period is indicated by a pink area. Specifically, during the leak event period (approximately 40-70 seconds), the concentration gradient (blue curve) rapidly increases and repeatedly exceeds the gradient threshold, while the concentration change rate (red curve) surges synchronously and exceeds the change rate threshold. Both exhibit coordinated fluctuations, consistent with the abrupt changes in methane concentration in both spatial and temporal dimensions during a leak. During non-leakage periods, both the concentration gradient and the change rate remain at low levels and do not reach the thresholds, reflecting the stable characteristics under background conditions. This figure, through the time-series correlation of multiple indicators, intuitively presents the dynamic response pattern of concentration gradient and rate of change in a leakage event, providing a visualized quantitative basis for the determination of leakage events.

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting methane gas leaks based on multi-sensor data fusion, characterized in that, Includes the following steps: The raw light absorption signals collected by multiple tunable semiconductor laser absorption spectroscopy sensors are obtained, and the absorption intensity sequence and corresponding wavelength shift of each sensor are obtained. A dynamic feature association network is constructed with sensor locations as nodes, and the absorption intensity sequence and wavelength drift of each sensor are used as the initial state vector of the corresponding node. In the dynamic feature association network, the state vectors of adjacent nodes are cross-compared along a preset topological path to calculate the similarity matrix of the absorption spectral features of each pair of adjacent sensors. Based on the similarity matrix, a subset of sensors with cooperative change patterns is identified, and the background absorption component and potential leakage signal component are separated from the absorption intensity sequence corresponding to the sensor subset. Spatiotemporal alignment and integration are performed on the potential leakage signal components to generate a fusion concentration distribution map for the region where the sensor subset is located. Based on the fused concentration distribution map, the methane concentration gradient and concentration change rate within the coverage space of the sensor subset are calculated; When the methane concentration gradient exceeds a preset morphological threshold, or the concentration change rate exceeds a preset dynamic threshold, a gas leak event is determined to exist within the space covered by the sensor subset.

2. The methane gas leak detection method based on multi-sensor data fusion according to claim 1, characterized in that, The process of acquiring the raw light absorption signals collected by multiple tunable semiconductor laser absorption spectral sensors, and obtaining the absorption intensity sequence and corresponding wavelength shift of each sensor, includes: Each sensor is configured at a different location in the monitoring space and records the corresponding laser path length information; Temporal denoising and spectral feature extraction are performed on the raw optical absorption signal of each sensor to obtain the absorption intensity sequence and corresponding wavelength shift of each sensor. The process of performing temporal denoising and spectral feature extraction on the raw optical absorption signal of each sensor includes: The original optical absorption signal of each sensor is decomposed using adaptive wavelet transform to obtain signal components at different scales; A background noise model is established based on the historical monitoring data of each sensor, and the signal components are then subjected to threshold filtering using the background noise model. The filtered signal components are reconstructed to obtain the denoised optical absorption signal; In the denoised optical absorption signal, the signal segment corresponding to the characteristic absorption spectrum of methane is identified by the peak detection algorithm; The center wavelength, full width at half maximum (FWHM), and integrated absorption intensity of each identified signal segment are extracted to form the absorption intensity sequence and wavelength shift of the sensor.

3. The methane gas leak detection method based on multi-sensor data fusion according to claim 2, characterized in that, The generation of the absorption intensity sequence and wavelength shift further includes: Based on the historical calibration data of laser operating temperature and drive current of each sensor, the extracted center wavelength is nonlinearly corrected. The corrected center wavelength is compared with the standard absorption wavelength, and the difference obtained from the comparison is taken as the wavelength drift of this measurement. Laser power fluctuation compensation and optical path attenuation compensation are applied to the integral absorption intensity to obtain the calibrated absorption intensity value; The calibrated absorption intensity values ​​are arranged in chronological order to form the absorption intensity sequence of the sensor.

4. The methane gas leak detection method based on multi-sensor data fusion according to claim 1, characterized in that, In the dynamic feature association network, the cross-comparison of the state vectors of adjacent nodes along a preset topological path includes: Define the connection rules between nodes in the dynamic feature association network, which are determined based on the actual physical distance between sensors and the quality of the communication link. Each connection is assigned an initial weight, which is inversely proportional to the physical distance between the corresponding sensor pairs; Within each fusion calculation cycle, the statistical characteristics of the absorption intensity sequence within the current time window are synchronously exchanged along connected node pairs. Calculate the cross-correlation function of the absorption intensity sequences between node pairs, and use the peak value of the cross-correlation function as the temporal similarity; Calculate the Euclidean distance between node pairs based on wavelength drift, and use the reciprocal of the Euclidean distance as the frequency domain similarity. The time-domain similarity and frequency-domain similarity are combined and multiplied by the initial weight to generate the element value at the corresponding position in the similarity matrix.

5. The methane gas leak detection method based on multi-sensor data fusion according to claim 1, characterized in that, The step of identifying a subset of sensors with cooperative change patterns based on the similarity matrix includes: Spectral clustering analysis is performed on the similarity matrix to divide the sensor nodes into several candidate subsets; Calculate the cohesion of each candidate subset, whereby cohesion is defined as the average similarity value between all node pairs within the subset; Candidate subsets with cohesion higher than a preset subset threshold are marked as high-cohesion subsets; For each highly cohesive subset, analyze the consistency of the direction of change of the absorption intensity sequence of all sensors within it at the same time, and calculate the direction consistency ratio; A subset of sensors with a high cohesion rate exceeding a preset consistency threshold is selected and identified as a subset of sensors with a cooperative change pattern.

6. The methane gas leak detection method based on multi-sensor data fusion according to claim 5, characterized in that, The step of separating the background absorption component and the potential leakage signal component from the absorption intensity sequence corresponding to the subset of sensors includes: For each sensor in the aforementioned sensor subset, its absorption intensity sequence is input into a fully trained variational mode decomposition model. The variational mode decomposition model automatically decomposes the input sequence into multiple intrinsic mode functions; Based on the center frequency and energy distribution of each intrinsic mode function, they are classified into a background mode set and a signal mode set; The background absorption component of the sensor is reconstructed by superimposing all intrinsic mode functions in the background mode set. The potential leakage signal component of the sensor is obtained by subtracting the background absorption component from the original absorption intensity sequence of the sensor. The potential leakage signal components of all sensors in the sensor subset are summarized to form a subset of potential leakage signals.

7. The methane gas leak detection method based on multi-sensor data fusion according to claim 6, characterized in that, The spatiotemporal alignment and integration of the potential leak signal components include: A local spatial coordinate system is established with the geometric center of the sensor subset as the reference origin; Based on the position of each sensor in the local spatial coordinate system and its laser path length, calculate the spatial effective weight of the potential leakage signal component of each sensor; Based on a high-precision time synchronization signal, all potential leakage signal components of the sensors are aligned to a unified time scale; In the local spatial coordinate system, the position of each sensor is taken as a discrete point, and the value of the potential leakage signal component after spatiotemporal alignment is multiplied by the effective spatial weight is taken as the concentration contribution value of the discrete point. The concentration contribution value in the local spatial coordinate system is interpolated using the Kriging space interpolation algorithm to generate a continuous concentration field distribution. The concentration field distribution is integrated over a preset time interval to obtain the fusion concentration distribution map.

8. The methane gas leak detection method based on multi-sensor data fusion according to claim 7, characterized in that, The calculation of the methane concentration gradient and concentration change rate within the coverage space of the sensor subset based on the fused concentration distribution map includes: On the fusion concentration distribution map, calculate the partial derivative of the concentration at each spatial point along the spatial coordinate axis; The magnitude of the partial derivatives in each direction is taken as the methane concentration gradient at the spatial point. The time series of the fusion concentration distribution map is obtained at consecutive timestamps; Calculate the difference between adjacent time stamps in the fusion concentration distribution map; The average absolute value of the difference map within the spatial range covered by the sensor subset is taken as the methane concentration change rate within the time period corresponding to the adjacent timestamps.

9. The methane gas leak detection method based on multi-sensor data fusion according to claim 2, characterized in that, The process of establishing a background noise model based on historical monitoring data from each sensor, and then using this model to perform threshold filtering on the signal components, is achieved through the following steps: Collect the historical light absorption signal of each sensor under a leak-free state to form the background signal sample set of the sensor; The background signal sample set is subjected to the same adaptive wavelet transform as the original optical absorption signal to decompose the background signal components at different scales. For each scale, calculate the amplitude probability distribution of all background signal components at that scale, and set the upper limit of the amplitude in the preset high probability interval as the noise threshold of that scale. When performing threshold filtering on the signal components of the corresponding scale obtained by decomposing the original optical absorption signal, the portion of the signal component with an amplitude lower than the noise threshold of the scale is set to zero, while the portion with an amplitude higher than or equal to the noise threshold of the scale is retained.

10. The methane gas leak detection method based on multi-sensor data fusion according to claim 4, characterized in that, The process of assigning an initial weight to each connection, wherein the initial weight is inversely proportional to the physical distance between the corresponding sensor pairs, is achieved through the following steps: Obtain the actual physical distance between each pair of connected sensor nodes in the dynamic feature association network; For each pair of sensor nodes, their physical distance is input into a preset weight allocation function, which is a monotonically decreasing function, such that the output initial weight value is inversely proportional to the input actual physical distance value. The calculated initial weight values ​​are assigned to the connections between the corresponding sensor nodes; In subsequent calculation cycles of the dynamic feature association network, the initial weight value remains unchanged, or the initial weight value is dynamically adjusted based on the real-time monitoring results of the communication link quality between sensor nodes.

Images