Methane leakage anomaly detection system and method based on dual-light fusion

Abstract

The invention provides a methane leakage anomaly detection system and method based on dual-light fusion, and belongs to the field of industrial safety detection. According to the system, a double-light-path holder type laser detector is arranged in a monitoring area, and methane concentration and image data are synchronously acquired through a 1653nm near-infrared laser module and a visible light image module. Time features (mean values, trends and the like) are extracted from the concentration data through time sequence analysis, and spatial features (edges, textures and the like) are extracted from the image data through deep learning; the system fuses time and space features to generate a comprehensive feature vector, reduces the calculation complexity through a dimension reduction technology, and retains core features at the same time. And constructing an anomaly detection model based on an unsupervised learning algorithm, identifying anomaly features in real time, and triggering an alarm. The system realizes real-time and accurate detection of methane leakage, and enhances the reliability and efficiency of industrial safety.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial safety detection, and in particular to a methane leakage anomaly detection system and method based on dual-light fusion. Background Art

[0002] Currently, methane (CH 4 ) As an important greenhouse gas and industrial raw material, its leakage not only has a serious impact on the environment, but may also cause safety hazards. Traditional methane leak detection methods, such as handheld gas detectors and fixed gas concentration monitors, can detect methane concentrations to a certain extent, but their efficiency and accuracy are limited by human operation and environmental factors. With the increasingly stringent environmental regulations and the improvement of safety production requirements, modern detection technologies have been gradually introduced, mainly including laser spectroscopy analysis technology, infrared gas imaging technology and multi-sensor network technology. These modern technologies can quickly and accurately identify and locate methane leaks through high sensitivity and real-time monitoring capabilities, thereby effectively reducing its impact on the environment and ensuring the safe operation of production equipment. However, the existing leak detection technology still has some shortcomings, mainly manifested in the lack of comprehensive consideration of environmental factors, reliance on a single sensor, and slow response speed.

[0003] The prior art has the following deficiencies:

[0004] Most of the existing methane leak detection systems rely on a single sensor for monitoring, and this design has obvious limitations in complex and changing environments. A single sensor cannot fully capture environmental changes and gas concentration fluctuations, which affects the reliability and accuracy of detection; traditional methane leak detection methods are usually based on gas concentration monitoring, which has low detection sensitivity for low-concentration methane leaks. In practical applications, low-concentration methane leaks are often difficult to detect in a timely manner, increasing safety hazards and environmental risks; existing detection technologies have obvious shortcomings in response speed, especially when gas concentrations change dramatically, and cannot respond quickly. This delay may lead to the expansion of leakage incidents, increasing the difficulty and risk of handling; existing methane leak detection technologies often lack the ability to comprehensively analyze multiple factors and cannot comprehensively assess leakage risks. This single-dimensional analysis method is difficult to cope with the needs of leak detection in complex environments, increasing the possibility of leak incidents being ignored.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The object of the present invention is to provide a methane leak anomaly detection system and method based on dual-light fusion to solve the problems raised in the above background technology.

[0007] A methane leak anomaly detection system based on dual-light fusion, comprising:

[0008] The data acquisition module is used to mark the transformer surface, the flange where the pipeline is connected to the transformer, and the inlet and outlet positions of all connected pipelines as detection points in sequence, and to arrange thermometers, mid-infrared thermal imagers, and non-dispersive infrared gas analyzers at the detection points. The surface temperature of the equipment is collected by the mid-infrared thermal imager, and the gas concentration data of the environment is collected by the non-dispersive infrared gas analyzer.

[0009] The data acquisition module is used to deploy dual-optical path pan-tilt laser methane leak detectors at key locations inside and around the monitoring area, and synchronously collect methane concentration data and image data in the environment through a 1653nm near-infrared laser transmitting and receiving module and a visible light image acquisition module.

[0010] The multi-dimensional feature extraction module is used to perform time series analysis on the collected infrared laser concentration data and extract time series features, such as mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes. At the same time, spatial feature extraction is performed on image data, and deep learning technology is used to extract spatial features in the image, such as edges, textures, shapes, etc., to capture the spatial characteristics of the leakage source.

[0011] The multispectral feature fusion module is used to deeply fuse the extracted temporal features and spatial features to generate a comprehensive feature vector, thereby improving the comprehensiveness and accuracy of the features.

[0012] The anomaly detection module is used to analyze time series features to predict future leakage trends and distinguish different types of leakage; it is used to analyze spatial features to locate the leakage source and identify the leakage diffusion pattern; based on the feature vector after dimensionality reduction, an anomaly detection model is constructed, and an unsupervised learning algorithm is used to identify abnormal points in the feature vector to achieve real-time anomaly detection of methane leakage.

[0013] Furthermore, the specific steps for obtaining the current methane concentration data and image data are as follows:

[0014] The near-infrared laser reflected light intensity in the current environment is obtained through the infrared laser transmitting and receiving device. Specifically, this system uses the near-infrared 1653nm spectrum to measure the methane concentration according to the Beer-Lambert law. The formula of the Beer-Lambert law is the relationship between the reflected light intensity and the methane concentration in the current environment:

[0015] I=I 0 ·e -α·C·d

[0016] Where I is the intensity of reflected light received by the infrared laser transmitting and receiving device from the outside world, I 0 is the light intensity emitted by the infrared laser transmitting and receiving device, α is the absorption coefficient of methane gas, C is the methane concentration, and d is the optical path length.

[0017] By inversely solving the formula, the methane concentration at the detection point is obtained. The formula for obtaining the methane concentration data in the current environment is:

[0018]

[0019] Among them, C i represents the methane concentration at the i-th detection point.

[0020] Furthermore, the specific steps for obtaining the current image data are:

[0021] The visible light image acquisition module outputs an image frame based on the visible light image data collected in the monitoring area. The value of each pixel is the color intensity within a certain color space range. The relationship between color intensity and environmental status is expressed by the formula:

[0022] P ij =ADC(I ij ·S+D)

[0023] Among them, P ij Represents the pixel value of the pixel in the i-th row and j-th column of the image, I ij represents the incident light intensity of the pixel, S represents the sensitivity of the camera (i.e., the gain coefficient), D represents the dark current of the camera (i.e., the offset), and ADC represents the analog-to-digital converter function, which is used to convert the analog signal into a digital signal.

[0024] Through the above steps, the visible light image data in the current environment can be acquired in real time and accurately.

[0025] Furthermore, the specific logic for obtaining the time feature is as follows:

[0026] The collected infrared laser concentration data are analyzed in time series to extract time series features: mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes.

[0027] First, the collected infrared laser concentration data is preprocessed, including data cleaning, missing value processing and outlier detection, to ensure the quality of the data. Secondly, the preprocessed infrared laser concentration data is arranged in chronological order to construct time series data. Finally, the features in the time series are extracted through time series analysis technology.

[0028] Mean: Calculate the average value of infrared laser concentration in the time series. The formula is:

[0029]

[0030] Among them, x i represents the i-th observation in the time series, and N represents the total number of observations.

[0031] Variance: Calculates the fluctuation degree of infrared laser concentration in the time series. The formula is:

[0032]

[0033] Trend: To analyze the long-term trend of changes in time series, a linear regression model can be used to fit the trend. The formula is:

[0034] Trend=α+βt

[0035] Among them, α and β are parameters obtained by least squares fitting, and t represents time.

[0036] Periodicity: To identify periodic changes in time series, Fourier transform can be used to analyze periodicity. The formula is:

[0037]

[0038] Where F(k) represents the Fourier coefficient at frequency k, x n Represents the nth observation in the time series.

[0039] Furthermore, the logic for screening out all warning points is as follows:

[0040] The comprehensive monitoring index of each detection point is compared with the preset index threshold. If the comprehensive detection index of the i-th detection point is higher than the index threshold, the i-th detection point is set as a warning point and the position is marked as i 0 , if the comprehensive monitoring index of the i-th detection point is lower than the index threshold, the i-th detection point is regarded as a safe point.

[0041] Furthermore, the specific logic for obtaining spatial features is as follows:

[0042] First, the collected image data is preprocessed, including image denoising, enhancement, normalization, etc., to ensure the quality and consistency of the image data. The preprocessed image data is used to train the deep learning model to automatically extract spatial features in the image.

[0043] The preprocessed image data is divided into training set, validation set and test set, and the VGG convolutional neural network model architecture is selected. The training set data is then used to train the deep learning model, and the validation set data is used for model tuning and hyperparameter adjustment. The cross entropy loss function is used, and the formula is:

[0044]

[0045] Among them, L(θ) is the loss function, y i is the true label, is the predicted label and N is the number of samples.

[0046] Through the trained deep learning model, the input image is forward propagated to extract the spatial features in the image.

[0047] The spatial features in the image are extracted through convolution operation. The formula of convolution operation is:

[0048]

[0049] Among them, Z i,j,k is the output of the convolutional layer, W m,n,c is the convolution kernel, X i+m,j+n,c is the input image, b k is the bias term, i,j are the spatial positions, k is the feature map channel, and m,n are the sizes of the convolution kernels.

[0050] The texture information in the image is extracted through pooling operation. Commonly used pooling methods include maximum pooling and average pooling, and the formula is:

[0051]

[0052]

[0053] Among them, P max (i,j) and P avg (i, j) are the results of maximum pooling and average pooling respectively, and R(i, j) is the pooling window.

[0054] The shape information in the image is extracted through the fully connected layer. The output of the fully connected layer is expressed as:

[0055] F c =W·X+b

[0056] Among them, F c is the output of the fully connected layer, W is the weight matrix, X is the input feature vector, and b is the bias vector.

[0057] Furthermore, the specific logic for obtaining the comprehensive feature vector is:

[0058] First, feature alignment is performed to align features from different sources to the same dimension to ensure the synchronization of each feature in time and space. The time feature is based on a continuous time series, and the spatial and spectral features also need to be aligned in time.

[0059] Then, each feature is normalized to eliminate the dimensional difference and make the features comparable on the same scale.

[0060] Secondly, different feature fusion methods are used to fuse the temporal features, spatial features and spectral features to generate a comprehensive feature vector.

[0061] Simple concatenation: directly concatenate the feature vectors together in order. The formula is:

[0062] F concat =[F time ,f spatial ,F spectral ]

[0063] Among them, F concat is the concatenated feature vector, F time is the time eigenvector, F spatial is the spatial eigenvector, F sqpectral is the spectral eigenvector.

[0064] Weighted fusion: By assigning different weights to each feature, weighted fusion is performed. The formula is:

[0065] F weighted =αF time +βF spatial +γF spectral

[0066] Among them, F weighted It is the feature vector after weighted fusion, α, β, γ are the weights of each feature, satisfying α+β+γ=1.

[0067] Multimodal neural network fusion: Feature fusion is performed through a multimodal fusion network. The formula is:

[0068]

[0069] Among them, F fusion is the fused feature vector, FFN represents the feedforward neural network, Represents a fusion operation (dot product, concatenation, etc.).

[0070] Finally, the fused feature vector is subjected to dimensionality reduction processing to reduce the feature dimension and avoid overfitting, including principal component analysis (PCA) and linear discriminant analysis (LDA).

[0071] PCA:

[0072] F PCA =W T F concat

[0073] Among them, F PCA is the eigenvector after dimensionality reduction, W is the principal component matrix of PCA, and F concat is the concatenated feature vector.

[0074] LDA:

[0075]

[0076] Among them, F PCA is the eigenvector after dimensionality reduction, is the projection matrix of LDA, F concat is the concatenated feature vector.

[0077] Finally, a comprehensive feature vector is generated.

[0078] Furthermore, the specific logic for predicting future leakage trends and distinguishing different types of leakage is as follows:

[0079] The ARIMA (Auto Regressive Integrated Moving Average) model is used to predict leakage trends. The formula is:

[0080]

[0081] in, is the predicted methane concentration, φ and θ j are model parameters and ∈ is the error term.

[0082] The time series classification method Dynamic Time Warping (DTW) is used for pattern classification. The formula is:

[0083]

[0084] Among them, X and Y are two sets of time series, π is the path mapping, ||x i -y j || is the distance between two points.

[0085] Furthermore, the specific logic for locating the leak source and identifying the leak diffusion pattern is as follows:

[0086] Use the deep learning-based object detection algorithm YOLO to locate the leakage source. The formula is:

[0087]

[0088] Among them, YOLO(X) is the detected object bounding box, and P(b|X) is the confidence of the bounding box b.

[0089] Using image processing methods such as edge detection to analyze the diffusion pattern, the formula is:

[0090]

[0091] Among them, E(x,y) is the edge strength, G x and G y is the gradient component.

[0092] Furthermore, the specific logic for judging abnormal methane leakage is as follows:

[0093] Select an anomaly detection model based on machine learning, use the comprehensive feature vector of the normal state to train the anomaly detection model, and divide the comprehensive feature vector of the normal state into a training set, a validation set, and a test set. Use the training set data to train the anomaly detection model, and use the validation set data to perform model tuning and hyperparameter adjustment.

[0094] Isolated forest:

[0095]

[0096] Among them, score(x) is the anomaly score, h(x) is the length of the path from the root node to the leaf node, E(h(x)) is the expected path length, and c(n) is the average path length. When score(x) is less than a certain threshold, it is judged as an anomaly.

[0097] The present invention further provides a methane leakage anomaly detection method based on dual-light fusion, which is performed by the above-mentioned methane leakage anomaly detection system based on dual-light fusion, and the specific steps include:

[0098] Step 1: Deploy dual-optical path pan-tilt laser methane leak detectors at key locations inside and around the monitoring area, and synchronously collect methane concentration data and image data in the environment through a 1653nm near-infrared laser transmitting and receiving module and a visible light image acquisition module;

[0099] Step 2: Perform time series analysis on the collected infrared laser concentration data and extract time series features, such as mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes. At the same time, perform spatial feature extraction on the image data and use deep learning technology to extract spatial features in the image, such as edges, textures, shapes, etc., to capture the spatial characteristics of the leakage source;

[0100] Step 3: Deeply fuse the extracted temporal features and spatial features to generate a comprehensive feature vector to improve the comprehensiveness and accuracy of the features;

[0101] Step 4: Analyze time series features to predict future leakage trends and distinguish different types of leakage; analyze spatial features to locate the source of leakage and identify leakage diffusion patterns; build an anomaly detection model based on the feature vector after dimensionality reduction, and use an unsupervised learning algorithm to identify abnormal points in the feature vector to achieve real-time anomaly detection of methane leakage.

[0102] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0103] The present invention deploys a dual-optical path pan-tilt laser methane leak detector, uses a 1653nm near-infrared laser transmitting and receiving module and a visible light image acquisition module to synchronously collect methane concentration data and image data in the environment, and extracts the temporal characteristics of methane concentration changes and the spatial characteristics of the leakage source through time series analysis and deep learning technology, successfully capturing key information of the leakage event. Furthermore, by deeply fusing the comprehensive feature vector generated by the temporal and spatial features, the comprehensiveness and accuracy of the detection are improved, and high-precision, comprehensive and accurate detection of methane leaks in the monitoring area is achieved, effectively reducing the false alarm rate and missed detection rate, and providing comprehensive data support for leakage analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0105] Figure 2 It is a schematic diagram of the overall method flow of the present invention. DETAILED DESCRIPTION

[0106] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0107] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0108] Example:

[0109] See also Figure 1 The present invention provides a methane leakage anomaly detection system based on dual-light fusion, comprising:

[0110] The data acquisition module is used to deploy dual-optical path pan-tilt laser methane leak detectors at key locations inside and around the monitoring area, and synchronously collect methane concentration data and image data in the environment through a 1653nm near-infrared laser transmitting and receiving module and a visible light image acquisition module;

[0111] In this embodiment, the specific steps for obtaining the methane concentration data in the current environment are:

[0112] The near-infrared laser reflected light intensity in the current environment is obtained through the infrared laser transmitting and receiving device. Specifically, this system uses the near-infrared 1653nm spectrum to measure the methane concentration according to the Beer-Lambert law. The formula of the Beer-Lambert law is the relationship between the reflected light intensity and the methane concentration in the current environment:

[0113] I=I 0 ·e -α·C·d

[0114] Where I is the intensity of reflected light received by the infrared laser transmitting and receiving device from the outside world, I 0 is the light intensity emitted by the infrared laser transmitting and receiving device, α is the absorption coefficient of methane gas, C is the methane concentration, and d is the optical path length.

[0115] By inversely solving the formula, the methane concentration at the detection point is obtained. The formula for obtaining the methane concentration data in the current environment is:

[0116]

[0117] Among them, C i represents the methane concentration at the i-th detection point.

[0118] Near-infrared absorption spectroscopy detection technology uses the high penetration and high sensitivity of laser beams to analyze the concentration of methane in the environment in real time in a non-contact manner, which is crucial for complex environments such as gas storage tanks and industrial pipelines. This non-contact detection not only avoids potential interference with equipment, but also improves the accuracy and safety of detection. Near-infrared absorption spectroscopy can continuously monitor changes in methane concentration, quickly locate methane leaks by comparing set thresholds, and ensure that timely measures are taken to prevent the spread of leaks. In addition, the technology can accurately identify small changes in concentration, help detect possible leakage risks or other operational abnormalities in equipment, and reduce methane leaks caused by equipment failures.

[0119] In this embodiment, the specific steps for obtaining image data in the current environment are:

[0120] The visible light image acquisition module outputs an image frame based on the visible light image data collected in the monitoring area. The value of each pixel is the color intensity within a certain color space range. The relationship between color intensity and environmental status is expressed by the formula:

[0121] P ij =ADC(I ij ·S+D)

[0122] Among them, P ij Represents the pixel value of the pixel in the i-th row and j-th column of the image, I ij represents the incident light intensity of the pixel, S represents the sensitivity of the camera (i.e., the gain coefficient), D represents the dark current of the camera (i.e., the offset), and ADC represents the analog-to-digital converter function, which is used to convert the analog signal into a digital signal.

[0123] Visible light cameras record scene images when methane leaks occur, which is particularly important for accurately analyzing leak points and assessing the severity of leaks. The captured scene images can provide intuitive environmental data, and combined with the methane concentration information detected by infrared lasers, they form a comprehensive analysis basis for multi-dimensional data. This not only improves the comprehensiveness and accuracy of detection, but also makes subsequent leak source analysis more precise, greatly reduces the false alarm rate, and enhances the monitoring capability and application value of the overall system.

[0124] The multi-dimensional feature extraction module is used to perform time series analysis on the collected infrared laser concentration data and extract time series features, such as mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes. At the same time, spatial feature extraction is performed on image data, and deep learning technology is used to extract spatial features in the image, such as edges, textures, shapes, etc., to capture the spatial characteristics of the leakage source;

[0125] In this embodiment, the specific logic for obtaining the time feature and the space feature is as follows:

[0126] The collected infrared laser concentration data are analyzed in time series to extract time series features: mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes.

[0127] First, the collected infrared laser concentration data is preprocessed, including data cleaning, missing value processing and outlier detection, to ensure the quality of the data. Secondly, the preprocessed infrared laser concentration data is arranged in chronological order to construct time series data. Finally, the features in the time series are extracted through time series analysis technology.

[0128] Mean: Calculate the average value of infrared laser concentration in the time series. The formula is:

[0129]

[0130] Among them, x i represents the i-th observation in the time series, and N represents the total number of observations.

[0131] Variance: Calculates the fluctuation degree of infrared laser concentration in the time series. The formula is:

[0132]

[0133] Trend: To analyze the long-term trend of changes in time series, a linear regression model can be used to fit the trend. The formula is:

[0134] Trend=α+βt

[0135] Among them, α and β are parameters obtained by least squares fitting, and t represents time.

[0136] Periodicity: To identify periodic changes in time series, Fourier transform can be used to analyze periodicity. The formula is:

[0137]

[0138] Where F(k) represents the Fourier coefficient at frequency k, x n Represents the nth observation in the time series.

[0139] First, the collected image data is preprocessed, including image denoising, enhancement, normalization, etc., to ensure the quality and consistency of the image data. The preprocessed image data is used to train the deep learning model to automatically extract spatial features in the image.

[0140] The preprocessed image data is divided into training set, validation set and test set, and the VGG convolutional neural network model architecture is selected. The training set data is then used to train the deep learning model, and the validation set data is used for model tuning and hyperparameter adjustment. The cross entropy loss function is used, and the formula is:

[0141]

[0142] Among them, L(θ) is the loss function, y i is the true label, is the predicted label and N is the number of samples.

[0143] Through the trained deep learning model, the input image is forward propagated to extract the spatial features in the image.

[0144] The spatial features in the image are extracted through convolution operation. The formula of convolution operation is:

[0145]

[0146] Among them, Z i,j,k is the output of the convolutional layer, W m,n,c is the convolution kernel, X i+m,j+n,c is the input image, b k is the bias term, i,j are the spatial positions, k is the feature map channel, and m,n are the sizes of the convolution kernels.

[0147] The texture information in the image is extracted through pooling operation. Commonly used pooling methods include maximum pooling and average pooling, and the formula is:

[0148]

[0149]

[0150] Among them, P max (i,j) and P avg (i, j) are the results of maximum pooling and average pooling respectively, and R(i, j) is the pooling window.

[0151] The shape information in the image is extracted through the fully connected layer. The output of the fully connected layer is expressed as:

[0152] F c =W·X+b

[0153] Among them, F c is the output of the fully connected layer, W is the weight matrix, X is the input feature vector, and b is the bias vector.

[0154] The time feature and the space feature complement each other in the present invention and together constitute the core analysis method for high-precision methane leak detection. The time feature captures the dynamic change pattern of the detection data by analyzing the time series of the change of methane concentration in the leakage event, and can reveal important information such as the speed, duration and periodicity of the leakage. This feature helps the system to quickly identify abnormal concentration fluctuations and effectively filter false alarms. At the same time, the spatial feature analyzes the image data through a deep learning algorithm to extract the specific location and spatial distribution characteristics of the leakage source in the environment, and further optimize the positioning accuracy of the leakage point. By combining the concentration change law of the time feature and the leakage point identification of the spatial feature, the system can generate a comprehensive and accurate leakage source feature vector, realize all-round monitoring and precise positioning of methane leakage, improve the efficiency and reliability of detection, and provide strong technical support for the timely response and effective handling of leakage accidents. The warning point screening module is used to compare the comprehensive monitoring indicators with the preset indicators, mark the location of the abnormal comprehensive monitoring indicators as a warning point, and screen out all the warning points;

[0155] The multispectral feature fusion module is used to deeply fuse the extracted temporal features and spatial features to generate a comprehensive feature vector, thereby improving the comprehensiveness and accuracy of the features.

[0156] In this embodiment, the logic for obtaining the comprehensive feature vector is:

[0157] First, feature alignment is performed to align features from different sources to the same dimension to ensure the synchronization of each feature in time and space. The time feature is based on a continuous time series, and the spatial and spectral features also need to be aligned in time.

[0158] Then, each feature is normalized to eliminate the dimensional difference and make the features comparable on the same scale.

[0159] Secondly, different feature fusion methods are used to fuse the temporal features, spatial features and spectral features to generate a comprehensive feature vector.

[0160] Simple concatenation: directly concatenate the feature vectors together in order. The formula is:

[0161] F concat =[F time ,F spatial ,F spectral ]

[0162] Among them, F concat is the concatenated feature vector, F time is the time eigenvector, F spatial is the spatial eigenvector, F sqpectral is the spectral eigenvector.

[0163] Weighted fusion: By assigning different weights to each feature, weighted fusion is performed. The formula is:

[0164] F weighted =αF time +βF spatial +γF spectral

[0165] Among them, F weighted It is the feature vector after weighted fusion, α, β, γ are the weights of each feature, satisfying α+β+γ=1.

[0166] Multimodal neural network fusion: Feature fusion is performed through a multimodal fusion network. The formula is:

[0167]

[0168] Among them, F fusion is the fused feature vector, FFN represents the feedforward neural network, Represents a fusion operation (dot product, concatenation, etc.).

[0169] Finally, the fused feature vector is subjected to dimensionality reduction processing to reduce the feature dimension and avoid overfitting, including principal component analysis (PCA) and linear discriminant analysis (LDA).

[0170] PCA:

[0171] F PCA =WTF concat

[0172] Among them, F PCA is the eigenvector after dimensionality reduction, W is the principal component matrix of PCA, and F concat is the concatenated feature vector.

[0173] LDA:

[0174]

[0175] Among them, F PCA is the eigenvector after dimensionality reduction, is the projection matrix of LDA, F concat is the concatenated feature vector.

[0176] The comprehensive feature vector is the basis of the high-precision methane leak detection system in the present invention. It integrates the dynamic change pattern of methane concentration with time characteristics and the leakage source location information with spatial characteristics. Through the extraction of time characteristics, the system can capture the changing trend of methane concentration over time, identify key information such as the rapid occurrence and duration of leakage, and thus accurately judge the severity and urgency of the leakage. The introduction of spatial features analyzes the scene image through a deep learning model to determine the specific location of the leakage source and its distribution in space, thereby optimizing the specific location of the leakage point. The comprehensive feature vector organically combines these two dimensions to form a feature set that comprehensively describes the leakage event. This comprehensive feature set enables the system to more accurately identify, locate and evaluate methane leakage events, which not only improves the accuracy and reliability of detection, but also provides a scientific basis for formulating efficient emergency response strategies, ensuring that leakage accidents can be controlled and handled in a timely and effective manner.

[0177] The anomaly detection module is used to analyze time series features to predict future leakage trends and distinguish different types of leakage; it is used to analyze spatial features to locate the leakage source and identify the leakage diffusion pattern; based on the feature vector after dimensionality reduction, an anomaly detection model is constructed, and an unsupervised learning algorithm is used to identify abnormal points in the feature vector to achieve real-time anomaly detection of methane leakage.

[0178] In this embodiment, the specific logic for predicting future leakage trends, distinguishing different types of leakage, locating the leakage source, identifying the leakage diffusion pattern, and judging the abnormality of methane leakage is as follows:

[0179] The ARIMA (Auto Regressive Integrated Moving Average) model is used to predict leakage trends. The formula is:

[0180]

[0181] in, is the predicted methane concentration, φ i and θ j are model parameters and ∈ is the error term.

[0182] The time series classification method Dynamic Time Warping (DTW) is used for pattern classification. The formula is:

[0183]

[0184] Among them, X and Y are two sets of time series, π is the path mapping, ||x i -y j || is the distance between two points.

[0185] Use the deep learning-based object detection algorithm YOLO to locate the leakage source. The formula is:

[0186]

[0187] Among them, YOLO(X) is the detected object bounding box, and P(b|X) is the confidence of the bounding box b.

[0188] Using image processing methods such as edge detection to analyze the diffusion pattern, the formula is:

[0189]

[0190] Among them, E(x,y) is the edge strength, G x and G y is the gradient component.

[0191] Select an anomaly detection model based on machine learning, use the comprehensive feature vector of the normal state to train the anomaly detection model, and divide the comprehensive feature vector of the normal state into a training set, a validation set, and a test set. Use the training set data to train the anomaly detection model, and use the validation set data to perform model tuning and hyperparameter adjustment.

[0192] Isolated forest:

[0193]

[0194] Among them, score(x) is the anomaly score, h(x) is the length of the path from the root node to the leaf node, E(h(x)) is the expected path length, and c(n) is the average path length. When score(x) is less than a certain threshold, it is judged as an anomaly.

[0195] Predicting future leakage trends can identify potential risks in advance, support preventive maintenance strategies, and ensure the continued stable operation of the system. Differentiating between different types of leaks, through pattern recognition technology, provides specific directions for targeted repairs and maintenance, optimizes maintenance processes, and reduces unnecessary inspections and repairs. The accuracy of locating the source of the leak greatly improves the response efficiency, helps to take prompt measures at the early stage of the leak, and reduces environmental pollution and safety risks. Identifying the leakage diffusion pattern provides a scientific basis for emergency plans and regional control, ensuring effective evacuation and protection of on-site personnel in crisis situations. The function of judging abnormal methane leaks can detect problems at an early stage, avoid the expansion of accidents, and ensure the safety of personnel and equipment. These combined effects make methane leak detection and management more intelligent and efficient, and comprehensively improve industrial safety standards.

[0196] See also Figure 2The present invention also provides a methane leakage anomaly detection method based on dual-light fusion, which is performed by the above-mentioned methane leakage anomaly detection system based on dual-light fusion, and the specific steps include:

[0197] Step 1: Deploy dual-optical path pan-tilt laser methane leak detectors at key locations inside and around the monitoring area, and synchronously collect methane concentration data and image data in the environment through a 1653nm near-infrared laser transmitting and receiving module and a visible light image acquisition module;

[0198] Step 2: Perform time series analysis on the collected infrared laser concentration data and extract time series features, such as mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes. At the same time, perform spatial feature extraction on the image data and use deep learning technology to extract spatial features in the image, such as edges, textures, shapes, etc., to capture the spatial characteristics of the leakage source;

[0199] Step 3: Deeply fuse the extracted temporal features and spatial features to generate a comprehensive feature vector to improve the comprehensiveness and accuracy of the features;

[0200] Step 4: Analyze time series features to predict future leakage trends and distinguish different types of leakage; analyze spatial features to locate the source of leakage and identify leakage diffusion patterns; build an anomaly detection model based on the feature vector after dimensionality reduction, and use an unsupervised learning algorithm to identify abnormal points in the feature vector to achieve real-time anomaly detection of methane leakage.

[0201] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0202] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0203] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0204] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A methane leak anomaly detection system based on dual-light fusion, characterized in that: include: The data acquisition module is used to deploy dual-optical path pan-tilt laser methane leak detectors at key locations inside and around the monitoring area, and synchronously collect methane concentration data and image data in the environment through a 1653nm near-infrared laser transmitting and receiving module and a visible light image acquisition module. The multi-dimensional feature extraction module is used to perform time series analysis on the collected infrared laser concentration data and extract time series features, such as mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes. At the same time, spatial feature extraction is performed on image data, and deep learning technology is used to extract spatial features in the image, such as edges, textures, shapes, etc., to capture the spatial characteristics of the leakage source. The multispectral feature fusion module is used to deeply fuse the extracted temporal features and spatial features to generate a comprehensive feature vector, thereby improving the comprehensiveness and accuracy of the features. The anomaly detection module is used to analyze time series features to predict future leakage trends and distinguish different types of leakage; it is used to analyze spatial features to locate the leakage source and identify the leakage diffusion pattern; based on the feature vector after dimensionality reduction, an anomaly detection model is constructed, and an unsupervised learning algorithm is used to identify abnormal points in the feature vector to achieve real-time anomaly detection of methane leakage.

2. The system according to claim 1, characterized in that The dual-optical path pan-tilt laser methane leak detector also has the following features: The measuring distance is greater than 200 meters, the gas concentration detection range is 0-100000ppm﹒ m, and the response time is ≤0.02s; it can realize continuous scanning monitoring of a large area of ​​360° horizontally and 180° vertically; it has fast response speed and high detection accuracy; it has unique selectivity for the target gas and good stability; 220VAC working voltage, RS485 data signal output, optical fiber video signal output; multiple preset position settings, and the cruise route can be freely set.

3. According to claim 1, a methane leak anomaly detection system based on dual light fusion is characterized in that: The specific steps for obtaining the methane concentration data in the current environment are: The near-infrared laser reflected light intensity in the current environment is obtained through the infrared laser transmitting and receiving device. Specifically, this system uses the near-infrared 1653nm spectrum to measure the methane concentration according to the Beer-Lambert law. The formula of the Beer-Lambert law is the relationship between the reflected light intensity and the methane concentration in the current environment: I=I0·e -α·C·d Among them, I is the intensity of reflected light received by the infrared laser transmitting and receiving device from the outside, I0 is the intensity of light emitted by the infrared laser transmitting and receiving device, α is the absorption coefficient of methane gas, C is the methane concentration, and d is the optical path length. By inversely solving the formula, the methane concentration at the detection point is obtained. The formula for obtaining the methane concentration data in the current environment is: Among them, C i represents the methane concentration at the i-th detection point.

4. According to claim 1, a methane leak anomaly detection system based on dual light fusion is characterized in that: The specific steps for obtaining image data in the current environment are: The visible light image acquisition module outputs an image frame based on the visible light image data collected in the monitoring area. The value of each pixel is the color intensity within a certain color space range. The relationship between color intensity and environmental status is expressed by the formula: P ij =ADC(I ij ·S+D) Among them, P ij Represents the pixel value of the pixel in the i-th row and j-th column of the image, I ij represents the incident light intensity of the pixel, S represents the sensitivity of the camera (i.e., the gain coefficient), D represents the dark current of the camera (i.e., the offset), and ADC represents the analog-to-digital converter function, which is used to convert the analog signal into a digital signal. Through the above steps, the visible light image data in the current environment can be acquired in real time and accurately.

5. The methane leak anomaly detection system based on dual-light fusion according to claim 2 is characterized in that: The specific logic for obtaining time features is: The collected infrared laser concentration data are analyzed in time series to extract time series features: mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes. First, the collected infrared laser concentration data is preprocessed, including data cleaning, missing value processing and outlier detection, to ensure the quality of the data. Secondly, the preprocessed infrared laser concentration data is arranged in chronological order to construct time series data. Finally, the features in the time series are extracted through time series analysis technology. Mean: Calculate the average value of infrared laser concentration in the time series. The formula is: Among them, x i represents the i-th observation in the time series, and N represents the total number of observations. Variance: Calculates the fluctuation degree of infrared laser concentration in the time series. The formula is: Trend: To analyze the long-term trend of changes in time series, a linear regression model can be used to fit the trend. The formula is: Trend=α+βt Among them, α and β are parameters obtained by least squares fitting, and t represents time. Periodicity: To identify periodic changes in time series, Fourier transform can be used to analyze periodicity. The formula is: Where F(k) represents the Fourier coefficient at frequency k, x n Represents the nth observation in the time series.

6. The methane leak anomaly detection system based on dual-light fusion according to claim 2 is characterized in that: The specific logic for obtaining spatial features is: First, the collected image data is preprocessed, including image denoising, enhancement, normalization, etc., to ensure the quality and consistency of the image data. The preprocessed image data is used to train the deep learning model to automatically extract spatial features in the image. The preprocessed image data is divided into training set, validation set and test set, and the VGG convolutional neural network model architecture is selected. The training set data is then used to train the deep learning model, and the validation set data is used for model tuning and hyperparameter adjustment. The cross entropy loss function is used, and the formula is: Among them, L(θ) is the loss function, y i is the true label, is the predicted label and N is the number of samples. Through the trained deep learning model, the input image is forward propagated to extract the spatial features in the image. The spatial features in the image are extracted through convolution operation. The formula of convolution operation is: Among them, Z i,j,k is the output of the convolutional layer, W m,n,c is the convolution kernel, X i+m,j+n,c is the input image, b k is the bias term, i,j are the spatial positions, k is the feature map channel, and m,n are the sizes of the convolution kernels. The texture information in the image is extracted through pooling operation. Commonly used pooling methods include maximum pooling and average pooling, and the formula is: Among them, P max (i,j) and P avg (i, j) are the results of maximum pooling and average pooling respectively, and R(i, j) is the pooling window. The shape information in the image is extracted through the fully connected layer. The output of the fully connected layer is expressed as: F c =W·X+b Among them, F c is the output of the fully connected layer, W is the weight matrix, X is the input feature vector, and b is the bias vector.

7. The methane leak anomaly detection system based on dual-light fusion according to claim 3 is characterized in that: The specific logic for obtaining the comprehensive feature vector is: First, feature alignment is performed to align features from different sources to the same dimension to ensure the synchronization of each feature in time and space. The time feature is based on a continuous time series, and the spatial and spectral features also need to be aligned in time. Then, each feature is normalized to eliminate the dimensional difference and make the features comparable on the same scale. Secondly, different feature fusion methods are used to fuse the temporal features, spatial features and spectral features to generate a comprehensive feature vector. Simple concatenation: directly concatenate the feature vectors together in order. The formula is: F concat =[F time ,F spatial ,F spectral ] Among them, F concat is the concatenated feature vector, F time is the time eigenvector, F spatial is the spatial eigenvector, F spectral is the spectral eigenvector. Weighted fusion: By assigning different weights to each feature, weighted fusion is performed. The formula is: F weighted =αF time +βF spatial +γF spectral Among them, F weighted It is the feature vector after weighted fusion, α, β, γ are the weights of each feature, satisfying α+β+γ=1. Multimodal neural network fusion: Feature fusion is performed through a multimodal fusion network. The formula is: Among them, F fusion is the fused feature vector, FFN represents the feedforward neural network, Represents a fusion operation (dot product, concatenation, etc.). Finally, the fused feature vector is subjected to dimensionality reduction processing to reduce the feature dimension and avoid overfitting, including principal component analysis (PCA) and linear discriminant analysis (LDA). PCA: F PCA =W T F concat Among them, F PCA is the eigenvector after dimensionality reduction, W is the principal component matrix of PCA, and F concat is the concatenated feature vector. LDA: Among them, F PCA is the eigenvector after dimensionality reduction, is the projection matrix of LDA, F concat is the concatenated feature vector. Finally, a comprehensive feature vector is generated.

8. The methane leak anomaly detection system based on dual-light fusion according to claim 4 is characterized in that: The specific logic for predicting future leakage trends and distinguishing different types of leakage is as follows: The ARIMA (Auto Regressive Integrated Moving Average) model is used to predict leakage trends. The formula is: in, is the predicted methane concentration, φ i and θ j are model parameters and ∈ is the error term. The time series classification method Dynamic Time Warping (DTW) is used for pattern classification. The formula is: Among them, X and Y are two sets of time series, π is the path mapping, ||x i -y j || is the distance between two points.

9. The methane leak anomaly detection system based on dual-light fusion according to claim 4 is characterized in that: The specific logic for locating the leak source and identifying the leak diffusion pattern is: Use the deep learning-based object detection algorithm YOLO to locate the leakage source. The formula is: Among them, YOLO(X) is the detected object bounding box, and P(b|X) is the confidence of the bounding box b. Using image processing methods such as edge detection to analyze the diffusion pattern, the formula is: Among them, E(x,y) is the edge strength, G x and G y is the gradient component.

10. The methane leak anomaly detection system based on dual-light fusion according to claim 4, characterized in that: The specific logic for judging abnormal methane leakage is: Select an anomaly detection model based on machine learning, use the comprehensive feature vector of the normal state to train the anomaly detection model, and divide the comprehensive feature vector of the normal state into a training set, a validation set, and a test set. Use the training set data to train the anomaly detection model, and use the validation set data to perform model tuning and hyperparameter adjustment. Isolated forest: Among them, score(x) is the anomaly score, h(x) is the length of the path from the root node to the leaf node, E(h(x)) is the expected path length, and c(n) is the average path length. When score(x) is less than a certain threshold, it is judged as an anomaly.

11. A methane leak anomaly detection method based on dual-light fusion, characterized in that: The methane leakage anomaly detection method is performed by the methane leakage anomaly detection system based on dual-light fusion according to any one of claims 1 to 9, and the specific steps include: Step 1: Deploy dual-optical path pan-tilt laser methane leak detectors at key locations inside and around the monitoring area, and synchronously collect methane concentration data and image data in the environment through a 1653nm near-infrared laser transmitting and receiving module and a visible light image acquisition module; Step 2: Perform time series analysis on the collected infrared laser concentration data and extract time series features, such as mean, variance, trend, periodicity, etc., to capture the temporal characteristics of concentration changes. At the same time, perform spatial feature extraction on the image data and use deep learning technology to extract spatial features in the image, such as edges, textures, shapes, etc., to capture the spatial characteristics of the leakage source; Step 3: Deeply fuse the extracted temporal features and spatial features to generate a comprehensive feature vector to improve the comprehensiveness and accuracy of the features; Step 4: Analyze time series features to predict future leakage trends and distinguish different types of leakage; analyze spatial features to locate the source of leakage and identify leakage diffusion patterns; build an anomaly detection model based on the feature vector after dimensionality reduction, and use an unsupervised learning algorithm to identify abnormal points in the feature vector to achieve real-time anomaly detection of methane leakage.

Images