An LNG leakage cloud detection and emergency response method based on AI vision

Abstract

The application discloses an LNG leakage cloud cluster detection and emergency response method based on AI vision, and comprises the following steps: collecting a continuous infrared video frame sequence to form an original infrared video frame sequence; pre-processing the original infrared video frame sequence to obtain a pre-processed infrared video frame sequence; establishing a detection partition according to the process partition of an emergency shut-off valve, and establishing a background statistical model for each process partition and fixed grid sub-area; constructing a feature vector according to a pixel block to obtain a feature vector set; performing abnormal score calculation on the feature vector by using Reed-Xiaoli abnormal detection and combining the background statistical model to obtain a suspected LNG leakage cloud cluster segmentation result; obtaining a leakage position and a diffusion trend parameter; determining an alarm level in combination with the leakage position and the diffusion trend, starting an audible and visual alarm when the linkage level is reached, and sending a linkage signal. The application improves the real-time performance of LNG leakage alarm grading and linkage disposal.

Description

Technical Field

[0001] This invention relates to the field of industrial safety monitoring technology, and in particular to a method for detecting and responding to LNG leak clouds based on AI vision. Background Technology

[0002] There is a risk of leakage during loading, unloading, storage and transportation of liquefied natural gas (LNG) in the plant area. After leakage, the gas is prone to form clouds with indistinct visual characteristics under the influence of low temperature and spread along the wind direction. If stable identification and rapid response cannot be achieved in the early stage, it can easily cause personal injury and plant-level accidents. Existing leakage monitoring in industrial sites mainly relies on contact gas detectors, point alarms and manual inspections. Contact sensors are usually limited by the density and installation location of the points, and it is difficult to cover complex areas such as valves, flanges and pipe corridors. Moreover, they are prone to missed or delayed detection when the concentration is low in the early stage of leakage or when the diffusion path deviates from the detection area. Manual inspections are affected by visibility, lighting conditions, personnel experience and response time, and it is difficult to meet the requirements of continuous monitoring and real-time alarm. In addition, there is a risk of entering in hazardous environments.

[0003] With the development of infrared imaging technology, non-contact infrared imaging cameras are used for visual monitoring of hazardous gas leaks. Some solutions extract suspected gas areas by performing threshold segmentation, background subtraction, or optical flow analysis on infrared video frame sequences, and trigger alarms by combining simple rules. At the same time, anomaly detection ideas are introduced to detect abnormal areas by constructing background models. Although this improves the monitoring range and visualization capabilities to some extent, there are still significant shortcomings in LNG scenarios. Industrial plant areas are often accompanied by dynamic background changes caused by steam, vehicle exhaust, heat source fluctuations, and equipment vibration. Conventional threshold or background subtraction can easily misjudge steam and exhaust as leak clouds, leading to frequent false alarms.

[0004] Therefore, how to provide an AI vision-based method for detecting and responding to LNG leak clouds is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI vision-based method for LNG leak cloud detection and emergency response. This invention achieves early and stable identification of weak cold plume clouds by statistically modeling the background of process zones and fixed grid sub-regions, combining pixel block feature vectors with Reed-Xiaoli anomaly detection to generate anomaly score maps. It also utilizes cross-frame correlation to obtain the leak location and diffusion trend, triggers audible and visual alarms according to alarm levels, closes the emergency shut-off valves of the corresponding process zones, and simultaneously sends linkage signals, thereby improving monitoring accuracy and real-time response.

[0006] An AI-based vision-based method for LNG leak cloud detection and emergency response according to an embodiment of the present invention includes the following steps: Connect to an explosion-proof infrared imaging camera that is sensitive to the methane band, collect continuous infrared video frame sequences, and form the original infrared video frame sequence; The original infrared video frame sequence is preprocessed to obtain the preprocessed infrared video frame sequence. Based on the process zoning of the emergency shut-off valve, a detection zone is established. The process zoning is completed using the pre-processed infrared video frame sequence. Fixed grid sub-regions are divided within each process zone. A background statistical model is established for each process zone and fixed grid sub-region. Based on the preprocessed infrared video frame sequence and combined with the background statistical model, feature vectors are constructed according to pixel blocks to obtain the feature vector set; For pixel blocks that meet the cold plume threshold condition, Reed-Xiaoli anomaly detection is used and combined with the background statistical model to calculate the anomaly score of the feature vector, generate an anomaly score map, determine the candidate cloud region based on the anomaly score map and process it to obtain the segmentation result of the suspected LNG leak cloud. Perform cross-frame correlation and tracking on the segmentation results of suspected LNG leak cloud to obtain parameters of leak location and spread trend; The alarm level is determined by combining the location of the leak and the spread trend. When the linkage level is reached, the audible and visual alarm is activated, the emergency shut-off valve of the corresponding process zone is closed, and a linkage signal is sent to the emergency command center and the industrial safety instrumented system.

[0007] Optionally, the formation of the original infrared video frame sequence specifically includes: Configure the explosion-proof infrared imaging camera to operate in a preset band sensitive to methane, configure the image resolution, frame rate, exposure time and gain parameters of the explosion-proof infrared imaging camera, and start the camera's continuous acquisition mode. In continuous acquisition mode, the monitoring scene is imaged and acquired. The infrared image frames output by the camera are received frame by frame according to the configured frame rate. Each infrared image frame is assigned an incremental frame sequence number according to the acquisition order, and each infrared image frame is written into the buffer in the order of the frame sequence number to form a continuous infrared video frame sequence. For each infrared image frame entering the buffer, a time stamp is written and bound to the frame sequence number. Based on the time stamp, the continuous infrared video frame sequence is time-ordered and duplicate and invalid frames are removed. The verified continuous infrared video frame sequence is output as the original infrared video frame sequence.

[0008] Optionally, obtaining the preprocessed infrared video frame sequence specifically includes: Non-uniformity correction is performed frame by frame on the original infrared video frame sequence. For each frame of infrared image, preset gain correction parameters and offset correction parameters are obtained at the pixel level. Gain correction and offset compensation are performed on each pixel value of the current frame of infrared image to obtain the non-uniformity corrected infrared video frame sequence. Denoising is performed frame by frame on the non-uniformity-corrected infrared video frame sequence to obtain the denoised infrared video frame sequence. Brightness temperature normalization is performed frame by frame on the denoised infrared video frame sequence. For each frame of infrared image, the minimum and maximum values ​​of pixel brightness temperature are counted within the current frame, and linearly mapped to the preset normalization interval according to the minimum and maximum values. At the same time, the frame number and time order remain unchanged, and the preprocessed infrared video frame sequence is output.

[0009] Optionally, the establishment of the background statistical model specifically includes: Receive the preprocessed infrared video frame sequence and obtain the process partition configuration information corresponding to the emergency shut-off valve. Map the partition boundary in the process partition configuration information to the pixel coordinate system of the monitoring screen, complete the process partition division according to the partition boundary, and write each pixel block into the process partition identifier. Within each process zone, rows and columns are divided according to a preset grid size to generate fixed grid sub-regions and write grid sub-region identifiers. The process zone identifiers are then bound to the grid sub-region identifiers to form a zone index table. Based on the preprocessed infrared video frame sequence, a background statistical model is established for each fixed grid sub-region in the partition index table. For each fixed grid sub-region, a preset number of background modeling frames are selected and the brightness temperature data of the pixel blocks in the fixed grid sub-region is extracted. The average level of the brightness temperature data in the background modeling frames is statistically analyzed to form a background mean vector. The joint fluctuation relationship of the brightness temperature data in the background modeling frames is statistically analyzed to form a background covariance matrix. The background mean vector and the background covariance matrix are associated and stored with the corresponding process partition identifier and grid sub-region identifier to form a background statistical model.

[0010] Optionally, obtaining the feature vector set specifically includes: Each frame of the preprocessed infrared video frame sequence is divided into blocks according to a preset block width and a preset block height to obtain multiple pixel blocks covering the monitoring screen. A pixel block identifier, its corresponding process partition identifier, and a fixed grid sub-region identifier are written to each pixel block. For each pixel block, the representative value of the brightness temperature within the block in the current frame is statistically analyzed and used as the brightness temperature feature. For the same pixel block, the brightness temperature change between adjacent frames is statistically analyzed and used as the temporal difference feature. For the same pixel block, the representative value of the spatial gradient within the block is statistically analyzed and used as the boundary feature. For the same pixel block, the brightness temperature fluctuation within a sliding time window of a preset length is statistically analyzed and used as the temporal fluctuation feature. For each pixel block, infrared image blocks are extracted and stacked in order of frame number within a sliding time window to form a time window image block sequence. The time window image block sequence is then input into a convolutional neural network, which outputs a fixed-dimensional depth feature vector. Based on the process partition identifier and fixed grid sub-region identifier of the pixel block, the background mean vector and background covariance matrix in the corresponding background statistical model are read. The brightness temperature feature, temporal difference feature, boundary feature, temporal fluctuation feature, and depth feature vector of the pixel block are processed by removing the mean and normalized according to the background covariance. The normalized features of each dimension are concatenated into the feature vector of the pixel block in a preset order and stored in association with the pixel block identifier to form a feature vector set.

[0011] Optionally, obtaining the segmentation results of the suspected LNG leak cloud specifically includes: In the preprocessed infrared video frame sequence, the pixel block set is located according to the process partition identifier and the fixed grid sub-region identifier. The feature vector of each pixel block in the pixel block set is read and the cold plume threshold condition is determined to filter and obtain the cold plume pixel block set. Within the range corresponding to the process zone identifier and the fixed grid sub-zone identifier, a set of homogeneous background pixel blocks is selected based on the background statistical model; Based on the background homogeneous pixel block set, the homogeneous pixel transformation is estimated, and the homogeneous pixel transformation is applied to the feature vector of the cold plume pixel block set to obtain the predicted background feature vector. The feature vector of the cold plume pixel block set and the predicted background feature vector are then subtracted to obtain the residual feature vector. The residual feature vectors are written into the residual feature vector set according to the spatial location of the pixel block, and a local background window is defined for each pixel block of the residual feature vector set within a fixed grid sub-region. The residual feature vectors corresponding to the background pixel blocks are selected from the local background window and aggregated to form a background residual sample set. Reed-Xiaoli anomaly detection computation is performed on the residual feature vector set. The background mean vector is calculated on the background residual sample set of the local background window, and robust covariance estimation is performed. The robust covariance estimation uses Tyler robust shape estimation. The covariance matrix is ​​initialized and updated cyclically according to the upper limit of the iteration count. In each iteration, the weights of each background residual sample set are calculated and updated weighted. Trace normalization is performed on the updated covariance matrix. The change in the covariance matrix is ​​monitored during iteration, and iteration stops when the change falls below the convergence threshold, resulting in a robust background covariance matrix. The background covariance matrix is ​​regularized. The corresponding robust background covariance matrix and background mean vector are called according to the pixel block process partition identifier and fixed grid sub-region identifier. The residual feature vector is mean-removed and whitening transformation is performed based on the robust background covariance matrix. Anomaly scores are calculated based on the whitening transformation results and written according to the pixel block spatial location to obtain anomaly score map. The background distribution of anomaly scores is statistically analyzed in the fixed grid sub-region and the partition threshold is determined. The anomaly score map is segmented by partition threshold to obtain cloud candidate regions. Connectivity extraction and morphological processing are performed on the cloud candidate regions to output the segmentation results of suspected LNG leak cloud.

[0012] Optionally, obtaining the leakage location and diffusion trend parameters specifically includes: For each frame of suspected LNG leak cloud segmentation results, perform connected region marking, obtain the boundary of the connected region and the set of pixel block coordinates contained in the connected region, calculate the area of ​​the connected region and the centroid coordinates of the connected region based on the set of pixel block coordinates; Cross-frame association is performed on connected regions of adjacent frames in the order of frame number. The centroid distance is calculated based on the centroid coordinates of the connected regions of adjacent frames, and the region overlap is calculated based on the boundary of the connected regions. Centroid distance threshold and overlap threshold are set. Connected regions that meet the centroid distance threshold and overlap threshold are determined to be the same cloud target and written into the same tracking trajectory sequence. Otherwise, they are written into different tracking trajectory sequences. By performing trajectory analysis on the effective tracking trajectory sequence, the leakage location and diffusion trend parameters are determined based on the centroid coordinates of the connected region in the starting frame of the tracking trajectory sequence, and then bound to the acquisition time stamp of the starting frame.

[0013] Optionally, the process of determining the alarm level by combining the leak location and diffusion trend, activating the audible and visual alarm when the linkage level is reached, closing the emergency shut-off valve of the corresponding process zone, and sending a linkage signal to the emergency command center and the industrial safety instrumented system specifically includes: Receive parameters of leak location and spread trend, match process zone identifiers based on leak location and generate alarm assessment objects; The alarm level is determined for the alarm assessment object. If the comparison result meets the preset linkage determination conditions, the alarm level is determined to be the linkage level. When the alarm level is determined to be the linkage level, an audible and visual alarm command is generated and output to the audible and visual alarm device. At the same time, an emergency shut-off interlock command is generated and output to the emergency shut-off valve of the process zone matching the leak location to perform the shut-off action. Simultaneously, a linkage signal is generated and written and sent to the emergency command center and the industrial safety instrumented system.

[0014] The beneficial effects of this invention are: This invention acquires a continuous infrared video frame sequence using an explosion-proof infrared imaging camera in the methane-sensitive band, and performs preprocessing such as non-uniformity correction, noise reduction, and brightness temperature normalization on the original infrared video frame sequence. This improves the infrared data in terms of noise suppression and scale consistency, providing a stable input for subsequent visual inspection, thereby reducing false detections and missed detections caused by imaging drift and random noise.

[0015] This invention establishes detection zones based on the process zones of emergency shut-off valves. Within each process zone, fixed grid sub-regions are further divided, and background statistical models are established separately. This effectively characterizes the background differences in different regions. By combining pixel block feature vector construction with Reed-Xiaoli anomaly detection calculation, an anomaly score map is obtained. This can highlight suspected cloud areas that deviate significantly from the background statistics under complex industrial background interference such as steam and exhaust gas, thereby improving the early segmentation stability of weak, semi-transparent LNG cold plume clouds.

[0016] This invention performs cross-frame correlation and tracking on the segmentation results of suspected LNG leak cloud clusters to obtain leak location and diffusion trend parameters. Based on the leak location and diffusion trend, it determines the alarm level. When the linkage level is reached, it activates audible and visual alarms, closes the emergency shut-off valves of the corresponding process zone, and sends linkage signals to the emergency command center and industrial safety instrument system. This achieves closed-loop linkage between detection results and handling actions, thereby improving the real-time performance of alarm classification and linkage handling and reducing the risk of misoperation caused by information inconsistency. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an AI vision-based LNG leak cloud detection and emergency response method proposed in this invention; Figure 2 This is a schematic diagram of anomaly score map generation and cloud segmentation in an AI vision-based LNG leak cloud detection and emergency response method proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-2 A method for detecting and responding to LNG leak clouds based on AI vision, comprising the following steps: Connect to an explosion-proof infrared imaging camera that is sensitive to the methane band, collect continuous infrared video frame sequences, and form the original infrared video frame sequence; The original infrared video frame sequence is preprocessed to obtain the preprocessed infrared video frame sequence. Based on the process zoning of the emergency shut-off valve, a detection zone is established. The process zoning is completed using the pre-processed infrared video frame sequence. Fixed grid sub-regions are divided within each process zone. A background statistical model is established for each process zone and fixed grid sub-region. Based on the preprocessed infrared video frame sequence and combined with the background statistical model, feature vectors are constructed according to pixel blocks to obtain the feature vector set; For pixel blocks that meet the cold plume threshold condition, Reed-Xiaoli anomaly detection is used and combined with the background statistical model to calculate the anomaly score of the feature vector, generate an anomaly score map, determine the candidate cloud region based on the anomaly score map and process it to obtain the segmentation result of the suspected LNG leak cloud. Perform cross-frame correlation and tracking on the segmentation results of suspected LNG leak cloud to obtain parameters of leak location and spread trend; The alarm level is determined by combining the location of the leak and the spread trend. When the linkage level is reached, the audible and visual alarm is activated, the emergency shut-off valve of the corresponding process zone is closed, and a linkage signal is sent to the emergency command center and the industrial safety instrumented system.

[0020] In this embodiment, the formation of the original infrared video frame sequence specifically includes: Configure the explosion-proof infrared imaging camera to operate in a preset band sensitive to methane, configure the image resolution, frame rate, exposure time and gain parameters of the explosion-proof infrared imaging camera, and start the camera's continuous acquisition mode. In continuous acquisition mode, the monitoring scene is imaged and acquired. The infrared image frames output by the camera are received frame by frame according to the configured frame rate. Each infrared image frame is assigned an incremental frame sequence number according to the acquisition order, and each infrared image frame is written into the buffer in the order of the frame sequence number to form a continuous infrared video frame sequence. For each infrared image frame entering the buffer, a time stamp is written and bound to the frame sequence number. Based on the time stamp, the continuous infrared video frame sequence is time-ordered and duplicate and invalid frames are removed. The verified continuous infrared video frame sequence is output as the original infrared video frame sequence.

[0021] In this embodiment, obtaining the preprocessed infrared video frame sequence specifically includes: Non-uniformity correction is performed frame by frame on the original infrared video frame sequence. For each frame of infrared image, preset gain correction parameters and offset correction parameters are obtained at the pixel level. Gain correction and offset compensation are performed on each pixel value of the current frame of infrared image so that the pixel output corresponding to the same uniform radiation input in the whole image tends to be consistent, and the non-uniformity corrected infrared video frame sequence is obtained. Denoising is performed frame by frame on the infrared video frame sequence after non-uniformity correction. The denoising process uses a preset two-dimensional smoothing filter kernel to perform neighborhood weighting operation on each frame of infrared image and outputs the denoised pixel value to suppress high-frequency random noise and isolated noise points, thus obtaining the denoised infrared video frame sequence. Brightness temperature normalization is performed frame by frame on the denoised infrared video frame sequence. For each frame of infrared image, the minimum and maximum values ​​of pixel brightness temperature are counted within the current frame, and linearly mapped to the preset normalization interval according to the minimum and maximum values. At the same time, the frame number and time order remain unchanged, and the preprocessed infrared video frame sequence is output.

[0022] In this embodiment, the establishment of the background statistical model specifically includes: Receive the preprocessed infrared video frame sequence and obtain the process partition configuration information corresponding to the emergency shut-off valve. Map the partition boundary in the process partition configuration information to the pixel coordinate system of the monitoring screen, complete the process partition division according to the partition boundary, and write each pixel block into the process partition identifier. Within each process zone, rows and columns are divided according to a preset grid size to generate fixed grid sub-regions and write grid sub-region identifiers. The process zone identifiers are then bound to the grid sub-region identifiers to form a zone index table. Based on the preprocessed infrared video frame sequence, a background statistical model is established for each fixed grid sub-region in the partition index table. For each fixed grid sub-region, a preset number of background modeling frames are selected and the brightness temperature data of the pixel blocks in the fixed grid sub-region is extracted. The average level of the brightness temperature data in the background modeling frames is statistically analyzed to form a background mean vector. The joint fluctuation relationship of the brightness temperature data in the background modeling frames is statistically analyzed to form a background covariance matrix. The background mean vector and the background covariance matrix are associated and stored with the corresponding process partition identifier and grid sub-region identifier to form a background statistical model.

[0023] In this embodiment, obtaining the feature vector set specifically includes: Each frame of the preprocessed infrared video frame sequence is divided into blocks according to a preset block width and a preset block height to obtain multiple pixel blocks covering the monitoring screen. A pixel block identifier, its corresponding process partition identifier, and a fixed grid sub-region identifier are written to each pixel block. For each pixel block, the representative value of the brightness temperature within the block in the current frame is statistically analyzed and used as the brightness temperature feature. For the same pixel block, the brightness temperature change between adjacent frames is statistically analyzed and used as the temporal difference feature. For the same pixel block, the representative value of the spatial gradient within the block is statistically analyzed and used as the boundary feature. For the same pixel block, the brightness temperature fluctuation within a sliding time window of a preset length is statistically analyzed and used as the temporal fluctuation feature. For each pixel block, infrared image blocks are extracted and stacked in order of frame number within a sliding time window to form a time window image block sequence. The time window image block sequence is then input into a convolutional neural network, which outputs a fixed-dimensional depth feature vector. Based on the process partition identifier and fixed grid sub-region identifier of the pixel block, the background mean vector and background covariance matrix in the corresponding background statistical model are read. The brightness temperature feature, temporal difference feature, boundary feature, temporal fluctuation feature, and depth feature vector of the pixel block are processed by removing the mean and normalized according to the background covariance. The normalized features of each dimension are concatenated into the feature vector of the pixel block in a preset order and stored in association with the pixel block identifier to form a feature vector set.

[0024] This invention integrates the brightness temperature features, temporal difference features, boundary features, and temporal fluctuation features of infrared video with the deep feature vector output by a convolutional neural network by constructing pixel-level feature vectors. It also combines background statistical models corresponding to process partitions and fixed grid sub-regions to perform mean removal and normalization, so that the feature expression has both physical interpretability and data-driven discrimination capability. This improves the separability of weak LNG leakage clouds and steam exhaust interference in complex industrial backgrounds, and enhances the stability and consistency of anomaly detection and segmentation.

[0025] In this embodiment, obtaining the segmentation results of the suspected LNG leak cloud specifically includes: In the preprocessed infrared video frame sequence, the pixel block set is located according to the process partition identifier and the fixed grid sub-region identifier. The feature vector of each pixel block in the pixel block set is read and the cold plume threshold condition is determined. The cold plume threshold condition determination includes obtaining the average brightness temperature of the pixel block in the current frame and the average brightness temperature of the same pixel block in the previous frame and calculating the brightness temperature difference value. It is determined that the brightness temperature difference value meets the cooling direction and reaches the cold plume cooling amplitude threshold. The average brightness temperature sequence of the pixel block in the sliding time window is obtained and the brightness temperature fluctuation meets the cold plume stability threshold. The average brightness temperature of the neighboring ring area of ​​the pixel block is obtained and the brightness temperature comparison between the pixel block and the neighboring ring area meets the cold plume comparison threshold. The continuous frame gate count is obtained and the number of consecutive frames that meet the threshold reaches the continuity threshold. The cold plume pixel block set is obtained by filtering. Within the range corresponding to the process partition identifier and the fixed grid sub-region identifier, a set of homogeneous background pixel blocks is selected based on the background statistical model. The set of homogeneous background pixel blocks satisfies the condition that the abnormal score is lower than the homogeneity screening threshold and meets the allowable range of the background statistical model. Homogeneous pixel transformation is estimated based on a set of homogeneous background pixel blocks. The homogeneous pixel transformation represents the homogeneous change relationship of the background in adjacent frames. The homogeneous pixel transformation is applied to the feature vector of the cold plume pixel block set to obtain the predicted background feature vector. The feature vector of the cold plume pixel block set and the predicted background feature vector are then subtracted to obtain the residual feature vector. The residual feature vectors are written into the residual feature vector set according to the spatial location of the pixel block. A local background window is defined for each pixel block of the residual feature vector set within a fixed grid sub-region. The local background window contains background pixel blocks that are spatially adjacent to the pixel block and are located in the same fixed grid sub-region. The residual feature vectors corresponding to the background pixel blocks are selected from the local background window and aggregated to form a background residual sample set. Reed-Xiaoli anomaly detection computation is performed on the residual feature vector set. The background mean vector is calculated on the background residual sample set of the local background window, and robust covariance estimation is performed using Tyler robust shape estimation. The covariance matrix is ​​initialized and updated iteratively according to the upper limit of the iteration count. In each iteration, the weights of each background residual sample set are calculated and updated weighted. Trace normalization is performed on the updated covariance matrix. The change in the covariance matrix is ​​monitored during iteration, and iteration stops when the change falls below the convergence threshold, resulting in a robust background covariance matrix. The background covariance matrix is ​​regularized. The corresponding robust background covariance matrix and background mean vector are called according to the pixel block process partition identifier and fixed grid sub-region identifier. The residual feature vector is subjected to mean removal processing and whitening transformation is performed based on the robust background covariance matrix. Anomaly scores are calculated based on the whitening transformation results and written according to the spatial position of the pixel block to obtain an anomaly score map. The background distribution of anomaly scores is statistically analyzed in the fixed grid sub-region and the partition threshold is determined. The anomaly score map is segmented by partition threshold to obtain cloud candidate regions. Connectivity extraction and morphological processing are performed on the cloud candidate regions to output the segmentation results of suspected LNG leak cloud clusters.

[0026] This invention utilizes cold plume threshold gating and process partition gridded background statistical modeling, combined with homogeneous pixel transformation to generate residual features, and introduces Reed-Xiaoli anomaly detection using local background windows and Tyler robust shape estimation. This yields stable anomaly score maps and partition threshold segmentation results, thereby improving the early detection rate and segmentation robustness of weak, semi-transparent LNG leak clouds under dynamic interference backgrounds such as steam and exhaust gas, reducing false alarms and false negatives caused by background drift, and improving the detection consistency of different process regions.

[0027] In this embodiment, obtaining the parameters of leakage location and diffusion trend specifically includes: For each frame of suspected LNG leak cloud segmentation results, perform connected region marking, obtain the boundary of the connected region and the set of pixel block coordinates contained in the connected region, calculate the area of ​​the connected region and the centroid coordinates of the connected region based on the set of pixel block coordinates, the area is taken as the number of elements in the set of pixel block coordinates, and the centroid coordinates are taken as the average of the horizontal coordinate and the average of the vertical coordinate in the set of pixel block coordinates. Cross-frame association is performed on connected regions of adjacent frames in the order of frame number. The centroid distance is calculated based on the centroid coordinates of the connected regions of adjacent frames, and the region overlap is calculated based on the boundary of the connected regions. Centroid distance threshold and overlap threshold are set. Connected regions that meet the centroid distance threshold and overlap threshold are determined to be the same cloud target and written into the same tracking trajectory sequence. Otherwise, they are written into different tracking trajectory sequences. By performing trajectory analysis on the effective tracking trajectory sequence, the leakage location and diffusion trend parameters are determined based on the centroid coordinates of the connected regions in the starting frame of the tracking trajectory sequence. These parameters are then bound to the acquisition time stamp of the starting frame. The diffusion trend parameters include the trajectory point sequence, the main dispersion direction, the dispersion velocity, the area change trend, and the diffusion front advance. The trajectory point sequence is obtained by sorting the centroid coordinates of the connected regions in each frame of the tracking trajectory sequence according to the frame number. The main dispersion direction is the direction from the starting point to the ending point of the trajectory point sequence. The dispersion velocity is calculated by combining the displacement of adjacent points in the trajectory point sequence with the acquisition time stamp. The area change trend is obtained by the change of the area of ​​the connected regions in the tracking trajectory sequence with the frame number. The diffusion front advance is obtained by the change of the farthest boundary point of the connected region in the main dispersion direction with the frame number.

[0028] This invention uses the segmentation results of suspected LNG leak cloud clusters to mark connected regions and perform cross-frame correlation tracking. By analyzing the centroid coordinates, regional overlap, and trajectory, the leak location is obtained and diffusion trend parameters are extracted simultaneously. This makes the location of the leak source and the assessment of the diffusion situation continuous and quantifiable, thereby improving the tracking stability of changes in the direction and speed of cloud cluster dispersion, reducing the impact of false detections in a single frame on the decision, and improving the consistency and real-time nature of the basis for alarm classification and linkage response.

[0029] In this embodiment, the process of determining the alarm level based on the leak location and diffusion trend, activating an audible and visual alarm when the linkage level is reached, closing the emergency shut-off valve of the corresponding process zone, and sending a linkage signal to the emergency command center and industrial safety instrumented system specifically includes: Receive parameters of leak location and spread trend, match process zone identifiers based on leak location and generate alarm assessment objects; The alarm level is determined for the alarm assessment object. The alarm level determination includes calculating the propulsion of the diffusion front based on the diffusion trend parameter and comparing it with the propulsion threshold, calculating the area change trend based on the diffusion trend parameter and comparing it with the growth threshold, calculating the drift velocity based on the diffusion trend parameter and comparing it with the velocity threshold, calculating the stability of the main drift direction based on the diffusion trend parameter and comparing it with the stability threshold. When the comparison result meets the preset linkage determination conditions, the alarm level is determined to be the linkage level. When the alarm level is determined to be the linkage level, an audible and visual alarm command is generated and output to the audible and visual alarm device. At the same time, an emergency shut-off interlock command is generated and output to the emergency shut-off valve of the process zone matching the leak location to execute the shut-off action. Simultaneously, a linkage signal is generated and written and sent to the emergency command center and the industrial safety instrumented system. This is used to initiate command and dispatch and safety interlock execution in parallel under the same triggering conditions, so as to shorten the emergency response delay and reduce the risk of misoperation caused by inconsistent information.

[0030] Example 1: To verify the feasibility of the present invention in practice, it was applied to a typical process area for LNG loading and unloading and storage tank transportation. This area includes multiple process zones with emergency shut-off valves. Common interference sources are present in the image, including intermittent steam emissions, vehicle exhaust, and changes in surface radiation from hot equipment. Furthermore, the equipment and pipe racks create numerous obstructions and highly reflective edges, causing the cold plume cloud formed by low-concentration, early leaks to exhibit weak contrast, semi-transparency, and discontinuous boundaries. Conventional threshold segmentation and simple background subtraction are prone to misjudging steam as leak clouds in such scenarios, triggering frequent false alarms. At the same time, there are detection lags and missed detections in the early stages of the actual cold plume. This example aims to solve the engineering pain points of the difficulty in stably identifying and quickly responding to early LNG leak clouds in complex industrial dynamics.

[0031] In this scenario, explosion-proof infrared imaging cameras sensitive to methane wavelengths are deployed to continuously acquire monitoring images. Following this invention, the video stream undergoes non-uniformity correction, noise reduction, and brightness temperature normalization to obtain a stable pre-processed infrared video frame sequence. Based on this, the monitoring images are regionalized according to the process zones of the emergency shut-off valve. Within each zone, fixed grid sub-regions are further divided, and background statistical models are maintained separately for each sub-region. This ensures that each grid sub-region has independent background mean and background fluctuation characterization, avoiding the mixing of background characteristics from high-steam areas and road exhaust areas into the same global background model, which could lead to threshold mismatch. Subsequently, each frame is divided into pixel blocks, and feature vectors are constructed. The feature vector simultaneously includes brightness temperature level, brightness temperature change between adjacent frames, spatial gradient representation, and brightness temperature fluctuation representation within a sliding time window. The infrared image blocks of pixels are stacked in frame order within the sliding time window as AI visual input. A convolutional neural network is used to perform visual feature extraction and inference on the spatiotemporal image blocks, outputting a fixed-dimensional depth feature vector. The depth feature vector is fused with physical statistical features and normalized according to the background statistical model to form a unified representation for subsequent anomaly detection. This allows the feature expression to retain the physical laws of the cooling trend and stable diffusion of the cold plume, while also reflecting the AI ​​vision's ability to discriminate complex background textures and interference patterns.

[0032] In the cloud detection stage, cold plume gating is used to screen pixel blocks. The gating simultaneously constrains the cooling direction, cooling amplitude, time window stability, neighborhood ring band comparison, and continuous frame persistence to reduce the probability of transient high responses caused by steam bursts and exhaust gas drift entering the main detection link. For the set of pixel blocks that pass through the gating, a homogeneous pixel transformation is estimated by combining it with a set of homogeneous background pixel blocks. This transformation is used to characterize the homogeneous change relationship of the background in adjacent frames. This transformation is applied to the feature vector of the gated pixel block to obtain the predicted background feature vector, and the difference between this and the actual feature vector yields the residual feature vector, allowing overall background drift and local non-leakage disturbances to be preferentially canceled. Subsequently, in a fixed grid sub-region... For each pixel block, a local background window is defined. Background residual samples are collected from the local background window and the background mean is calculated. At the same time, Tyler robust shape estimation is used to perform robust covariance estimation and regularization to ensure reversibility. The residual features are whitened under robust background statistics and Reed-Xiaoli anomaly scores are calculated to generate an anomaly score map. The anomaly score threshold is determined by a partitioned threshold method within a fixed grid sub-region to avoid the global threshold being raised in dense steam areas or lowered in low-noise areas, which could lead to regional false alarms or missed alarms. After threshold segmentation to obtain candidate cloud regions, connected component screening and morphological processing are performed to output the segmentation results of suspected LNG leak cloud regions.

[0033] To verify the beneficial effects of this invention, multiple sets of "weak leakage - moderate leakage - strong leakage" scenarios were constructed using a controlled release method, covering different background interference intensities and wind speed conditions. Long-term infrared video clips with and without leakage were collected for evaluation. Two common baseline schemes in industrial settings were selected for comparison: one is the traditional method of superimposing brightness temperature thresholds with background subtraction and performing morphological processing; the other is the standard Reed-Xiaoli anomaly detection method using a global background statistical model but excluding homogeneous pixel transformation residuals, local background windows, and robust covariance estimation. The three methods were subjected to offline playback and online simulation statistics under the same video and the same partition mapping conditions. The specific comparison data is shown in Table 1. Table 1 Comparison of LNG Leakage Cloud Identification and Response Performance

[0034] Table 1 shows that the present invention achieves a weak leakage detection rate of 0.90, which is higher than the traditional threshold + background difference method (0.62) and the standard RX method (0.74). It also reduces the false alarm rate without leakage to 0.6 times / hour, a significant decrease compared to 8.4 times / hour and 3.1 times / hour, and shortens the median detection delay to 4.2 seconds, which is better than 13.6 seconds and 8.9 seconds. Regarding segmentation consistency, the present invention has an average overlap of 0.68, higher than 0.41 and 0.53, and the median positioning error is reduced to 2.9 meters. The diffusion trend stability is improved to an average trajectory jitter of 1.6 meters, all of which are better than the comparative methods. Finally, the linkage success rate reaches 0.99, demonstrating more stable detection output and more reliable linkage closed-loop capability.

[0035] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting and responding to LNG leak clouds based on AI vision, characterized in that, Includes the following steps: Connect to an explosion-proof infrared imaging camera that is sensitive to the methane band, collect continuous infrared video frame sequences, and form the original infrared video frame sequence; The original infrared video frame sequence is preprocessed to obtain the preprocessed infrared video frame sequence. Based on the process zoning of the emergency shut-off valve, a detection zone is established. The process zoning is completed using the pre-processed infrared video frame sequence. Fixed grid sub-regions are divided within each process zone. A background statistical model is established for each process zone and fixed grid sub-region. Based on the preprocessed infrared video frame sequence and combined with the background statistical model, feature vectors are constructed according to pixel blocks to obtain the feature vector set; For pixel blocks that meet the cold plume threshold condition, Reed-Xiaoli anomaly detection is used and combined with the background statistical model to calculate the anomaly score of the feature vector, generate an anomaly score map, determine the candidate cloud region based on the anomaly score map and process it to obtain the segmentation result of the suspected LNG leak cloud. Perform cross-frame correlation and tracking on the segmentation results of suspected LNG leak cloud to obtain parameters of leak location and spread trend; The alarm level is determined by combining the location of the leak and the spread trend. When the linkage level is reached, the audible and visual alarm is activated, the emergency shut-off valve of the corresponding process zone is closed, and a linkage signal is sent to the emergency command center and the industrial safety instrumented system.

2. The method for detecting and responding to LNG leak clouds based on AI vision according to claim 1, characterized in that, The formation of the original infrared video frame sequence specifically includes: Configure the explosion-proof infrared imaging camera to operate in a preset band sensitive to methane, configure the image resolution, frame rate, exposure time and gain parameters of the explosion-proof infrared imaging camera, and start the camera's continuous acquisition mode. In continuous acquisition mode, the monitoring scene is imaged and acquired. The infrared image frames output by the camera are received frame by frame according to the configured frame rate. Each infrared image frame is assigned an incremental frame sequence number according to the acquisition order, and each infrared image frame is written into the buffer in the order of the frame sequence number to form a continuous infrared video frame sequence. For each infrared image frame entering the buffer, a time stamp is written and bound to the frame sequence number. Based on the time stamp, the continuous infrared video frame sequence is time-ordered and duplicate and invalid frames are removed. The verified continuous infrared video frame sequence is output as the original infrared video frame sequence.

3. The method for detecting and responding to LNG leak clouds based on AI vision according to claim 1, characterized in that, The process of obtaining the preprocessed infrared video frame sequence specifically includes: Non-uniformity correction is performed frame by frame on the original infrared video frame sequence. For each frame of infrared image, preset gain correction parameters and offset correction parameters are obtained at the pixel level. Gain correction and offset compensation are performed on each pixel value of the current frame of infrared image to obtain the non-uniformity corrected infrared video frame sequence. Denoising is performed frame by frame on the non-uniformity-corrected infrared video frame sequence to obtain the denoised infrared video frame sequence. Brightness temperature normalization is performed frame by frame on the denoised infrared video frame sequence. For each frame of infrared image, the minimum and maximum values ​​of pixel brightness temperature are counted within the current frame, and linearly mapped to the preset normalization interval according to the minimum and maximum values. At the same time, the frame number and time order remain unchanged, and the preprocessed infrared video frame sequence is output.

4. The method for detecting and responding to LNG leak clouds based on AI vision according to claim 1, characterized in that, The establishment of the background statistical model specifically includes: Receive the preprocessed infrared video frame sequence and obtain the process partition configuration information corresponding to the emergency shut-off valve. Map the partition boundary in the process partition configuration information to the pixel coordinate system of the monitoring screen, complete the process partition division according to the partition boundary, and write each pixel block into the process partition identifier. Within each process zone, rows and columns are divided according to a preset grid size to generate fixed grid sub-regions and write grid sub-region identifiers. The process zone identifiers are then bound to the grid sub-region identifiers to form a zone index table. Based on the preprocessed infrared video frame sequence, a background statistical model is established for each fixed grid sub-region in the partition index table. For each fixed grid sub-region, a preset number of background modeling frames are selected and the brightness temperature data of the pixel blocks in the fixed grid sub-region is extracted. The average level of the brightness temperature data in the background modeling frames is statistically analyzed to form a background mean vector. The joint fluctuation relationship of the brightness temperature data in the background modeling frames is statistically analyzed to form a background covariance matrix. The background mean vector and the background covariance matrix are associated and stored with the corresponding process partition identifier and grid sub-region identifier to form a background statistical model.

5. The method for detecting and responding to LNG leak clouds based on AI vision according to claim 1, characterized in that, The specific steps involved in obtaining the feature vector set are as follows: Each frame of the preprocessed infrared video frame sequence is divided into blocks according to a preset block width and a preset block height to obtain multiple pixel blocks covering the monitoring screen. A pixel block identifier, its corresponding process partition identifier, and a fixed grid sub-region identifier are written to each pixel block. For each pixel block, the representative value of the brightness temperature within the block in the current frame is statistically analyzed and used as the brightness temperature feature. For the same pixel block, the brightness temperature change between adjacent frames is statistically analyzed and used as the temporal difference feature. For the same pixel block, the representative value of the spatial gradient within the block is statistically analyzed and used as the boundary feature. For the same pixel block, the brightness temperature fluctuation within a sliding time window of a preset length is statistically analyzed and used as the temporal fluctuation feature. For each pixel block, infrared image blocks are extracted and stacked in order of frame number within a sliding time window to form a time window image block sequence. The time window image block sequence is then input into a convolutional neural network, which outputs a fixed-dimensional depth feature vector. Based on the process partition identifier and fixed grid sub-region identifier of the pixel block, the background mean vector and background covariance matrix in the corresponding background statistical model are read. The brightness temperature feature, temporal difference feature, boundary feature, temporal fluctuation feature, and depth feature vector of the pixel block are processed by removing the mean and normalized according to the background covariance. The normalized features of each dimension are concatenated into the feature vector of the pixel block in a preset order and stored in association with the pixel block identifier to form a feature vector set.

6. The method for detecting and responding to LNG leak clouds based on AI vision according to claim 1, characterized in that, The specific steps involved in obtaining the segmentation results of the suspected LNG leak cloud include: In the preprocessed infrared video frame sequence, the pixel block set is located according to the process partition identifier and the fixed grid sub-region identifier. The feature vector of each pixel block in the pixel block set is read and the cold plume threshold condition is determined to filter and obtain the cold plume pixel block set. Within the range corresponding to the process zone identifier and the fixed grid sub-zone identifier, a set of homogeneous background pixel blocks is selected based on the background statistical model; Based on the background homogeneous pixel block set, the homogeneous pixel transformation is estimated, and the homogeneous pixel transformation is applied to the feature vector of the cold plume pixel block set to obtain the predicted background feature vector. The feature vector of the cold plume pixel block set and the predicted background feature vector are then subtracted to obtain the residual feature vector. The residual feature vectors are written into the residual feature vector set according to the spatial location of the pixel block, and a local background window is defined for each pixel block of the residual feature vector set within a fixed grid sub-region. The residual feature vectors corresponding to the background pixel blocks are selected from the local background window and aggregated to form a background residual sample set. Reed-Xiaoli anomaly detection computation is performed on the residual feature vector set. The background mean vector is calculated on the background residual sample set of the local background window, and robust covariance estimation is performed. The robust covariance estimation uses Tyler robust shape estimation. The covariance matrix is ​​initialized and updated cyclically according to the upper limit of the iteration count. In each iteration, the weights of each background residual sample set are calculated and updated weighted. Trace normalization is performed on the updated covariance matrix. The change in the covariance matrix is ​​monitored during iteration, and iteration stops when the change falls below the convergence threshold, resulting in a robust background covariance matrix. The background covariance matrix is ​​regularized. The corresponding robust background covariance matrix and background mean vector are called according to the pixel block process partition identifier and fixed grid sub-region identifier. The residual feature vector is mean-removed and whitening transformation is performed based on the robust background covariance matrix. Anomaly scores are calculated based on the whitening transformation results and written according to the pixel block spatial location to obtain anomaly score map. The background distribution of anomaly scores is statistically analyzed in the fixed grid sub-region and the partition threshold is determined. The anomaly score map is segmented by partition threshold to obtain cloud candidate regions. Connectivity extraction and morphological processing are performed on the cloud candidate regions to output the segmentation results of suspected LNG leak cloud.

7. The method for detecting and responding to LNG leak clouds based on AI vision according to claim 1, characterized in that, The acquisition of the leakage location and diffusion trend parameters specifically includes: For each frame of suspected LNG leak cloud segmentation results, perform connected region marking, obtain the boundary of the connected region and the set of pixel block coordinates contained in the connected region, calculate the area of ​​the connected region and the centroid coordinates of the connected region based on the set of pixel block coordinates; Cross-frame association is performed on connected regions of adjacent frames in the order of frame number. The centroid distance is calculated based on the centroid coordinates of the connected regions of adjacent frames, and the region overlap is calculated based on the boundary of the connected regions. Centroid distance threshold and overlap threshold are set. Connected regions that meet the centroid distance threshold and overlap threshold are determined to be the same cloud target and written into the same tracking trajectory sequence. Otherwise, they are written into different tracking trajectory sequences. By performing trajectory analysis on the effective tracking trajectory sequence, the leakage location and diffusion trend parameters are determined based on the centroid coordinates of the connected region in the starting frame of the tracking trajectory sequence, and then bound to the acquisition time stamp of the starting frame.

8. The method for detecting and responding to LNG leak clouds based on AI vision according to claim 1, characterized in that, The process of determining the alarm level by combining the leak location and diffusion trend, activating the audible and visual alarm when the linkage level is reached, closing the emergency shut-off valve of the corresponding process zone, and sending a linkage signal to the emergency command center and the industrial safety instrumented system specifically includes: Receive parameters of leak location and spread trend, match process zone identifiers based on leak location and generate alarm assessment objects; The alarm level is determined for the alarm assessment object. If the comparison result meets the preset linkage determination conditions, the alarm level is determined to be the linkage level. When the alarm level is determined to be the linkage level, an audible and visual alarm command is generated and output to the audible and visual alarm device. At the same time, an emergency shut-off interlock command is generated and output to the emergency shut-off valve of the process zone matching the leak location to perform the shut-off action. Simultaneously, a linkage signal is generated and written and sent to the emergency command center and the industrial safety instrumented system.

Images