Dangerous chemical storage tank leakage detection device, medium and electronic equipment

Abstract

The invention relates to the technical field of dangerous chemical transportation safety monitoring, in particular to a dangerous chemical storage tank leakage detection device, a medium and electronic equipment, which comprises physical monitoring equipment arranged at a storage tank part and a control device in communication connection with the physical monitoring equipment, the physical monitoring equipment comprises an infrared thermal imaging high-definition video acquisition unit and a photoacoustic spectrometry gas sensing array. The control device integrates a data preprocessing module, a visual feature extraction module, a gas concentration feature extraction module, a spatio-temporal context sensing module, a self-adaptive decision fusion module, a digital twin simulation module and a linkage control execution module, and through multi-source data spatio-temporal alignment and fusion analysis, physical model simulation verification is combined. And intelligent identification and grading response of leakage characteristics are realized. The defect of false alarm caused by isolation of video and gas data in an existing detection method is effectively overcome, and the accuracy and reliability of leakage detection in a complex transportation environment are remarkably improved.

Description

A leak detection device, medium, and electronic equipment for hazardous chemical storage tanks. Technical Field

[0001] This invention relates to the field of safety monitoring technology for the transportation of hazardous chemicals, and in particular to a leak detection device, medium, and electronic equipment for hazardous chemical storage tanks. Background Technology

[0002] Real-time monitoring of tank leaks during the operation of hazardous chemical transport vehicles primarily relies on sensing technologies based on various physical or chemical principles. When a hazardous chemical leaks from inside the tank, certain properties of the surrounding medium change accordingly, and the monitoring system detects these changes to provide early warnings. For example, this can be achieved by continuously detecting the concentration of specific gases using an array of sensors arranged along the tank, or by immersing a monitoring probe into the space between the tank and its outer shell to detect liquid leaks.

[0003] Existing technologies for detecting leaks in hazardous chemical storage tanks suffer from the following technical challenges: In real-time safety monitoring scenarios involving hazardous chemical transport vehicles, methods relying on a single physical quantity for leak detection lack reliability. For example, relying solely on gas sensors may trigger false alarms due to interference from other non-leakage sources in the environment or the vehicle passing through a contaminated area. Conversely, relying solely on video image analysis is susceptible to interference from changes in daylight, weather factors such as rain and snow, or normal condensation on the tank surface, potentially misclassifying harmless scenarios as leaks. Furthermore, because video monitoring data and gas concentration detection data are independent, an effective cross-validation and intelligent linkage decision-making mechanism has not been established. The system cannot fuse and analyze multi-source information to distinguish between actual leaks and environmental interference, resulting in frequent false alarms and reducing the reliability and practicality of the monitoring system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a hazardous chemical storage tank leak detection device, medium, and electronic equipment, solving the technical problem of frequent false alarms caused by the lack of effective linkage between video monitoring and gas sensing data in existing leak detection methods in complex transportation environments.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: Firstly, the present invention provides a leak detection device for a hazardous chemical storage tank, comprising a physical monitoring device arranged at the location of the storage tank and a control device communicatively connected to the physical monitoring device; the physical monitoring device includes a high-definition video acquisition unit with infrared thermal imaging function and a gas sensing array employing photoacoustic spectroscopy technology; the control device includes: a data preprocessing module, used to perform timestamp alignment and spatial registration on the video stream acquired by the high-definition video acquisition unit and the gas concentration readings acquired by the gas sensing array, and to perform noise reduction processing to output standardized data; a visual feature extraction module, used to receive video data from the standardized data, apply a target detection model to identify leak-related visual features, and output visual detection results with confidence levels; a gas concentration feature extraction module, used to receive... The system receives standardized gas concentration data, analyzes the time-series concentration data, identifies abnormal fluctuations, and outputs gas concentration characteristics. A spatiotemporal context awareness module receives visual detection results and gas concentration characteristics, constructs a spatial field model to analyze the correlation between visual and gas signals in the spatiotemporal dimension, and outputs a correlation assessment result. An adaptive decision fusion module receives visual detection results, gas concentration characteristics, and correlation assessment results, performs uncertainty reasoning, dynamically adjusts the weights of different evidence sources, and outputs a joint decision result. A digital twin simulation module receives the joint decision result to perform leak scenario simulation, predicts diffusion patterns through simulation, and compares and verifies them with real-time monitoring data, outputting verification results. A linkage control execution module integrates the joint decision result and verification results to execute a tiered response strategy to trigger alarms or handling procedures.

[0006] Furthermore, in the hazardous chemical storage tank leakage detection device of the present invention, the data preprocessing module is configured to: perform microsecond-level synchronization processing on the video frames acquired by the high-definition video acquisition unit and the gas concentration sampling data acquired by the gas sensor array using a precision clock protocol to obtain synchronization data; based on the synchronization data, establish the transformation relationship between the camera coordinate system and the sensor array coordinate system through a calibration board, and calculate the intrinsic and extrinsic parameters using the Zhang Zhengyou calibration method to complete spatial registration and obtain registration data; perform multi-scale decomposition of the video data in the registration data using Gaussian pyramids, perform frequency band division of the gas data in the registration data using wavelet packet transform, reconstruct the signal after processing the transform coefficients through a threshold function, complete the noise reduction processing, and output a standardized data packet with a spatiotemporal index.

[0007] Furthermore, in the hazardous chemical storage tank leakage detection device of the present invention, the visual feature extraction module is configured to: receive video data in a standardized data packet; process the video data using an object detection model, wherein the object detection model is a neural network model based on YOLOv7 that introduces a cross-stage local attention mechanism, and adds an adaptive spatial feature pyramid structure to the neck network; extract multi-scale feature maps from the video data using the object detection model; fuse deep and shallow features in the multi-scale feature maps using a path aggregation network; generate bounding box coordinates and class confidence based on the fused features; apply a non-maximum suppression algorithm to process the bounding box coordinates and class confidence, remove overlapping detections, and output a visual detection result with confidence; simultaneously generate feature importance evaluation results during the extraction of multi-scale feature maps; feed back the feature importance evaluation results to the data preprocessing module; and adaptively adjust the threshold function in the noise reduction processing according to the feature importance evaluation results.

[0008] Furthermore, in the hazardous chemical storage tank leakage detection device of the present invention, the gas concentration feature extraction module is configured to: receive gas concentration data from a standardized data packet; model and analyze the gas concentration time series data using a gated cyclic unit long short-term memory network, and dynamically adjust the input sequence length through a sliding time window mechanism; during the analysis of the concentration time series data, dynamically update the abnormal fluctuation detection threshold using a particle swarm optimization algorithm with an adaptive adjustment strategy incorporating inertial weights to identify concentration anomalies; simultaneously analyze the attenuation characteristics of the sensor response curves in the gas sensor array to monitor the sensor health status; mark and compensate sensor data identified as abnormal based on the results of sensor health status monitoring; and generate gas concentration features based on the marked and compensated gas concentration data, the analysis results of the concentration time series data, and the results of sensor health status monitoring. The gas concentration features include quantitative indicators of concentration change trends and sensor reliability assessment parameters.

[0009] Furthermore, in the hazardous chemical storage tank leakage detection device of the present invention, the spatiotemporal context perception module is configured to: receive visual detection results and gas concentration characteristics; construct a three-dimensional spatial field model using an octree data structure for discretization based on the spatial location information and monitoring values ​​included in the visual detection results and gas concentration characteristics; fill in the blank areas of the spatial monitoring points using the Kriging interpolation algorithm during the construction of the three-dimensional spatial field model; analyze the spatial diffusion law and temporal evolution trend of leakage characteristics in the three-dimensional spatial field model using a spatiotemporal graph attention network, wherein the node features of the spatiotemporal graph attention network include physical coordinates, monitoring values ​​from the visual detection results and gas concentration characteristics, and timestamps, and the edge weights are calculated based on the distance between nodes and preset fluid dynamic parameters; calculate the similarity between the visual sequence and the gas concentration sequence under spatiotemporal transformation using an improved dynamic time warping algorithm; and output the correlation evaluation result based on the analysis results and similarity calculation results of the spatiotemporal graph attention network.

[0010] Furthermore, in the hazardous chemical storage tank leakage detection device of the present invention, the adaptive decision fusion module is configured to: receive visual detection results, gas concentration characteristics, and correlation assessment results; perform uncertainty reasoning on the evidence sources using DS evidence theory, the basic probability allocation function of DS evidence theory being designed based on fuzzy membership functions; use confidence data in the visual detection results, fluctuation indicators in the gas concentration characteristics, and correlation assessment results as evidence sources; use the correlation assessment results as evidence weight adjustment factors in the calculation of the basic probability allocation function; apply evidence synthesis rules to synthesize evidence, the evidence synthesis rules introducing conflict evidence redistribution algorithms; during the evidence synthesis process, use Jousselme distance to measure the degree of conflict between evidence; and generate a joint decision probability distribution based on the results of evidence synthesis using the evidence synthesis rules.

[0011] Furthermore, in the hazardous chemical storage tank leakage detection device of the present invention, the digital twin simulation module is configured to: receive joint decision results; establish a physical model of the storage tank and medium based on the parametric finite element method; determine the boundary condition parameters of the physical model through experimental data inversion; perform computational fluid dynamics simulation on the established and parametric physical model using the large eddy simulation method, with the computational fluid dynamics simulation coupling component transport equations and energy equations; generate multiple sets of simulation scenarios based on the joint decision results using the Monte Carlo algorithm to perform leakage scenario deduction; compare the simulation data generated from the leakage scenario deduction with the real-time monitoring data curves collected by the physical monitoring equipment using the dynamic time warping algorithm; correct the simulation parameters of the physical model in real time based on the comparison results of the simulation data and the real-time monitoring data curves using extended Kalman filtering to obtain the corrected simulation parameters; and perform simulation verification based on the corrected simulation parameters to generate verification results.

[0012] Furthermore, in the hazardous chemical storage tank leakage detection device of the present invention, the linkage control execution module is configured to: receive joint decision results and verification results; determine graded response levels based on the joint decision results and verification results, with each graded response level corresponding to a specific combination of disposal procedures; perform decision consistency judgment on the joint decision results and verification results using a voting mechanism combined with confidence weighting, with the decision consistency judgment using a dynamic consistency threshold set based on historical data; manage the state transition of the disposal process using a finite state machine model, with each state of the finite state machine model having a timeout rollback mechanism; select and trigger the corresponding disposal procedure from the disposal procedure library based on the determined graded response level and decision consistency judgment results, the disposal procedure library includes ventilation system control commands, leakage sealing device activation commands, and emergency communication trigger commands; after triggering the disposal procedure, collect disposal effect data and feed the disposal effect data back to the visual feature extraction module, gas concentration feature extraction module, spatiotemporal context perception module, adaptive decision fusion module, and digital twin simulation module for model parameter optimization.

[0013] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the function of the control device in the hazardous chemical storage tank leakage detection device as described above.

[0014] Thirdly, the present invention provides an electronic device comprising: a processor; a memory connected to the processor; the memory storing instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the function of the control device in the hazardous chemical storage tank leak detection device as described above.

[0015] The present invention offers several advantages. By combining spatiotemporal alignment and fusion analysis of multi-source data with physical model simulation verification, it achieves intelligent identification and graded response of leakage characteristics, effectively solving the false alarm problem caused by the isolation of video surveillance and gas sensing data in existing detection methods. The data preprocessing module uses a precise clock protocol to perform microsecond-level synchronization and spatial registration of infrared thermal imaging video streams and photoacoustic spectral gas concentration readings, eliminating spatiotemporal differences in multi-source heterogeneous data and providing standardized input for subsequent analysis. The visual feature extraction module uses an improved target detection model to identify leakage-related visual features, while the gas concentration feature extraction module identifies abnormal fluctuations through time-series analysis. Both are processed in parallel, but the spatiotemporal context perception module constructs a spatial field model to analyze signal correlation. The adaptive decision fusion module dynamically adjusts the weights of different evidence sources using DS evidence theory, the digital twin simulation module performs leakage scenario deduction and real-time data comparison verification based on a physical model, and the linkage control execution module integrates the outputs of multiple modules to execute a graded response strategy. This multi-level collaborative mechanism significantly improves the accuracy and reliability of leakage detection in complex transportation environments, and gradually reduces the false alarm rate through continuous learning and feedback optimization. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 is a system architecture diagram of a hazardous chemical storage tank leakage detection device, medium and electronic equipment provided by the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Firstly, referring to Figure 1, the present invention provides a leak detection device for a hazardous chemical storage tank, comprising a physical monitoring device arranged at the location of the storage tank and a control device communicatively connected to the physical monitoring device; the physical monitoring device includes a high-definition video acquisition unit with infrared thermal imaging function and a gas sensing array employing photoacoustic spectroscopy technology; the control device includes: a data preprocessing module, used to perform time stamp alignment and spatial registration on the video stream acquired by the high-definition video acquisition unit and the gas concentration readings acquired by the gas sensing array, and to perform noise reduction processing to output standardized data; a visual feature extraction module, used to receive video data from the standardized data, apply a target detection model to identify leak-related visual features, and output visual detection results with confidence levels; and a gas concentration feature extraction module, used to receive video data from the standardized data. The system receives gas concentration data, analyzes time-series concentration data, identifies abnormal fluctuations, and outputs gas concentration characteristics. A spatiotemporal context awareness module receives visual detection results and gas concentration characteristics, constructs a spatial field model to analyze the spatiotemporal correlation between visual and gas signals, and outputs a correlation assessment result. An adaptive decision fusion module receives visual detection results, gas concentration characteristics, and correlation assessment results, performs uncertainty reasoning, dynamically adjusts the weights of different evidence sources, and outputs a joint decision result. A digital twin simulation module receives the joint decision result to perform leak scenario simulation, predicts diffusion patterns through simulation, and compares and verifies them with real-time monitoring data, outputting verification results. A linkage control execution module integrates the joint decision result and verification result to execute a tiered response strategy to trigger alarms or handling procedures.

[0020] The hazardous chemical storage tank leak detection device achieves accurate identification and response to leak risks through the collaborative operation of multiple modules. The high-definition video acquisition unit in the physical monitoring equipment captures changes in the temperature field distribution on the tank surface based on infrared thermal imaging principles, while the photoacoustic spectroscopy gas sensing array monitors fluctuations in the concentration of specific gases through molecular absorption spectroscopy characteristics. These two types of heterogeneous data are integrated through a data preprocessing module in the control device. This module uses a precision clock protocol to achieve hardware-level synchronization between video frames and gas sampling data, and completes the mapping and alignment between visual space and physical space based on the coordinate transformation relationship established by the calibration board, forming a standardized data packet with a spatiotemporal index.

[0021] After receiving video sequences from standardized data, the visual feature extraction module extracts multi-scale features using an improved YOLOv7 model that incorporates a cross-stage local attention mechanism. It then fuses shallow and deep feature information using a path aggregation network and finally outputs a leak area detection box with confidence using a non-maximum suppression algorithm. Simultaneously, the gas concentration feature extraction module analyzes concentration time-series data using a gated recurrent unit memory network, dynamically adjusts the anomaly detection threshold using a particle swarm optimization algorithm, and performs hardware self-diagnosis through attenuation analysis of the sensor response curve. It outputs gas concentration features including trend indicators and reliability parameters.

[0022] The spatiotemporal context awareness module constructs a discretized spatial field model with an octree structure, uses Kriging interpolation to fill monitoring blind spots, and utilizes a spatiotemporal graph attention network to analyze the diffusion patterns of leakage characteristics in three-dimensional space. This network integrates physical coordinates, monitoring values, and timestamp information for its nodes, while edge weights are dynamically calculated based on fluid dynamics parameters. An improved dynamic time warping algorithm is then used to quantify the spatiotemporal correlation between visual sequences and gas concentration sequences, ultimately outputting a multi-source signal correlation assessment result.

[0023] The adaptive decision fusion module adopts the DS evidence theory framework, treating visual detection confidence, gas concentration fluctuation indicators, and correlation assessment results as independent evidence sources. It constructs basic probability assignments through fuzzy membership functions and introduces a conflict evidence redistribution algorithm to handle highly conflicting scenarios. During evidence synthesis, the Jousselme distance metric is used to measure the degree of conflict, enabling dynamic weight fusion of multi-source evidence and outputting a joint decision probability distribution.

[0024] The digital twin simulation module establishes a physical model of the storage tank based on the parametric finite element method. It then couples the component transport equations using large eddy simulation (LES) for computational fluid dynamics and generates multiple leakage scenario simulations using the Monte Carlo method. Simulation data and real-time monitoring data are compared morphologically using a dynamic time warping algorithm, and then extended Kalman filtering is applied to correct model parameters in real time, thus completing the virtual-real mapping verification of the leakage scenarios.

[0025] The coordinated control execution module determines the graded response levels based on the joint decision probability distribution and simulation verification results, using a voting mechanism combined with confidence weighting. A finite state machine model manages the state transitions of the disposal process, with each state having a timing constraint mechanism. The disposal program library pre-sets instruction sets for ventilation control, sealing device activation, etc. During execution, the disposal effect data collected optimizes the parameters of each module through a feedback loop, forming a closed-loop control system. For example, when a suspected leak is detected, the system simultaneously initiates gas concentration trend analysis and visual diffusion pattern matching. Only after the digital twin simulation verification is passed will the corresponding level of sealing device activation command be triggered.

[0026] Specifically, in the hazardous chemical storage tank leakage detection device of the present invention, the data preprocessing module is configured to: perform microsecond-level synchronization processing on the video frames acquired by the high-definition video acquisition unit and the gas concentration sampling data acquired by the gas sensor array using a precision clock protocol to obtain synchronization data; based on the synchronization data, establish the transformation relationship between the camera coordinate system and the sensor array coordinate system through a calibration board, and calculate the intrinsic and extrinsic parameters using the Zhang Zhengyou calibration method to complete spatial registration and obtain registration data; perform multi-scale decomposition of the video data in the registration data using Gaussian pyramids, perform frequency band division of the gas data in the registration data using wavelet packet transform, reconstruct the signal after processing the transform coefficients through a threshold function, complete the noise reduction processing, and output a standardized data packet with a spatiotemporal index.

[0027] The data preprocessing module standardizes multi-source heterogeneous data through a multi-level processing flow. During the timestamp alignment stage, the module employs a precision clock protocol to perform hardware-level clock synchronization between the video frames from the high-definition video acquisition unit and the concentration sampling data from the gas sensor array. This multi-level synchronization mechanism eliminates timing misalignment issues caused by differences in sensor sampling frequencies, establishing a unified time reference for subsequent spatiotemporal correlation analysis. For example, when a sudden increase in gas concentration is detected, the system can immediately backtrack to the corresponding infrared thermal imaging video frame for comparison.

[0028] The spatial registration process, based on the coordinate system transformation relationship established by the calibration board, uses the Zhang Zhengyou calibration method to calculate the camera's intrinsic and extrinsic parameters. This process establishes a mapping relationship between the two-dimensional image pixel coordinates and the three-dimensional physical space coordinates, enabling the leak area identified in the video to be accurately located to the actual spatial position of the gas sensor array. In actual deployment, the calibration board needs to cover the key monitoring area of ​​the storage tank, and sufficient calibration samples are obtained through multi-angle shooting to establish a high-precision coordinate transformation model.

[0029] Noise reduction processing is optimized separately for the different characteristics of video and gas data. For video data, a Gaussian pyramid is used for multi-scale decomposition. Adaptive thresholding is applied at each level of the pyramid to effectively separate noise caused by ambient lighting variations from actual leakage characteristics. For gas concentration data, wavelet packet transform is applied for frequency band division. Based on the energy distribution characteristics of the leakage signal in the frequency domain, effective frequency bands are retained while environmental interference is suppressed. The threshold function is designed considering sensor characteristics, and the signal is reconstructed after nonlinear compression of the transform coefficients.

[0030] After three levels of processing, the data is encapsulated into standardized data packets with spatiotemporal indexes. The spatiotemporal index includes the data acquisition timestamp and corresponding spatial coordinates, enabling subsequent modules to quickly retrieve monitoring data within a specific spatiotemporal range. This standardized format provides a unified input interface for the visual feature extraction module and the gas concentration feature extraction module, avoiding parsing errors caused by differences in data formats. For example, when the system needs to analyze the area surrounding a sensor node, it can directly and quickly locate the corresponding video frame region and concentration data sequence using the spatial index.

[0031] The module also establishes a feedback mechanism. When the visual feature extraction module identifies leakage characteristics in a specific area, it can adjust the area of ​​interest of the data preprocessing module in reverse through spatiotemporal indexing. This dynamic optimization mechanism can prioritize the processing of monitoring data in high-value areas under limited computing resources, improving system response efficiency. In actual operation, the data preprocessing module continuously monitors the data quality of each sensor node, marks abnormal sensors, and initiates data compensation algorithms to maintain the overall monitoring reliability of the system.

[0032] Specifically, the visual feature extraction module of the hazardous chemical storage tank leakage detection device of the present invention is configured to: receive video data in a standardized data packet; process the video data using an object detection model, which is a neural network model based on YOLOv7 that introduces a cross-stage local attention mechanism and adds an adaptive spatial feature pyramid structure to the neck network; extract multi-scale feature maps from the video data using the object detection model; fuse deep and shallow features in the multi-scale feature maps using a path aggregation network; generate bounding box coordinates and class confidence based on the fused features; process the bounding box coordinates and class confidence using a non-maximum suppression algorithm to remove overlapping detections and output a visual detection result with confidence; simultaneously generate feature importance evaluation results during the extraction of multi-scale feature maps; feed the feature importance evaluation results back to the data preprocessing module; and adaptively adjust the threshold function in the noise reduction process according to the feature importance evaluation results.

[0033] After receiving gas concentration data from the standardized data packet from the data preprocessing module, the gas concentration feature extraction module initiates a multi-dimensional analysis process. The module employs a gated recurrent unit (ROU) long short-term memory network to process the concentration time-series data. This network structure selectively retains historical information through a gating mechanism, effectively capturing the long-term dependencies of gas concentration changes. A sliding time window mechanism dynamically adjusts the input sequence length based on gas diffusion characteristics; for example, the window is shortened during vehicle acceleration to quickly respond to concentration abrupt changes, and extended during constant-speed driving to analyze trend patterns.

[0034] During the concentration time-series data analysis, the module employs an improved particle swarm optimization algorithm to dynamically update the abnormal fluctuation detection threshold. This algorithm introduces an adaptive adjustment strategy for inertial weights, maintaining a larger weight in the early stages of iteration to enhance global search capabilities, and reducing the weight in the later stages to improve local refinement accuracy. This dynamic threshold mechanism can adapt to concentration baseline fluctuations under different environmental conditions; for example, it automatically increases the threshold to avoid false alarms when passing through industrial areas, and decreases the threshold to improve detection sensitivity in clean areas.

[0035] The synchronously running sensor health monitoring module analyzes the attenuation characteristics of the response curves of each node in the gas sensor array. By establishing a sensor response model and comparing the deviation between the theoretical attenuation curve and the actual measurement data, sensor performance degradation caused by aging or contamination is identified. For the identified abnormal sensor nodes, the system uses a spatial interpolation algorithm based on neighboring node data for data compensation and marks the reliability level of the data in the gas concentration characteristics.

[0036] Based on labeled and compensated gas concentration data, concentration time-series data analysis results, and sensor health status monitoring results, the module generates comprehensive gas concentration characteristics. Quantitative indicators of concentration change trends include parameters such as gradient change rate and fluctuation frequency spectrum characteristics, while sensor reliability assessment parameters integrate dimensions such as health score and confidence weight. For example, when an increased response delay is detected at a certain sensor node, the system will correspondingly reduce the feature weight of that node's data and simultaneously activate a backup sensor for cross-validation.

[0037] This processing flow forms an adaptive closed-loop optimization mechanism. When the system detects an abnormal concentration pattern in a specific area, it adjusts the size of the sliding time window and the sensitivity of the anomaly detection threshold through a feature feedback mechanism. This dynamic adjustment capability enables the system to adapt to complex environmental changes during transportation, such as automatically enhancing the monitoring of local concentration accumulation in tunnels and improving the accuracy of diffusion pattern identification in open areas.

[0038] Specifically, the gas concentration feature extraction module of the hazardous chemical storage tank leakage detection device of the present invention is configured to: receive gas concentration data in a standardized data packet; model and analyze the gas concentration time series data using a gated cyclic unit long short-term memory network, and dynamically adjust the length of the input sequence through a sliding time window mechanism; during the analysis of the concentration time series data, dynamically update the abnormal fluctuation detection threshold using a particle swarm optimization algorithm with an adaptive adjustment strategy incorporating inertial weights to identify concentration anomalies; simultaneously analyze the attenuation characteristics of the sensor response curves in the gas sensor array to monitor the sensor health status; mark and compensate the sensor data identified as abnormal based on the results of the sensor health status monitoring; and generate gas concentration features based on the marked and compensated gas concentration data, the analysis results of the concentration time series data, and the results of the sensor health status monitoring. The gas concentration features include quantitative indicators of concentration change trends and sensor reliability assessment parameters.

[0039] After receiving gas concentration data from the standardized data packet from the data preprocessing module, the gas concentration feature extraction module initiates a multi-dimensional analysis process. The module employs a gated recurrent unit (ROU) long short-term memory network to process the concentration time-series data. This network structure selectively retains historical information through a gating mechanism, effectively capturing the long-term dependencies of gas concentration changes. A sliding time window mechanism dynamically adjusts the input sequence length based on gas diffusion characteristics; for example, the window is shortened during vehicle acceleration to quickly respond to concentration abrupt changes, and extended during constant-speed driving to analyze trend patterns.

[0040] During the concentration time-series data analysis, the module employs an improved particle swarm optimization algorithm to dynamically update the abnormal fluctuation detection threshold. This algorithm introduces an adaptive adjustment strategy for inertial weights, maintaining a larger weight in the early stages of iteration to enhance global search capabilities, and reducing the weight in the later stages to improve local refinement accuracy. This dynamic threshold mechanism can adapt to concentration baseline fluctuations under different environmental conditions; for example, it automatically increases the threshold to avoid false alarms when passing through industrial areas, and decreases the threshold to improve detection sensitivity in clean areas.

[0041] The synchronously running sensor health monitoring module analyzes the attenuation characteristics of the response curves of each node in the gas sensor array. By establishing a sensor response model and comparing the deviation between the theoretical attenuation curve and the actual measurement data, sensor performance degradation caused by aging or contamination is identified. For the identified abnormal sensor nodes, the system uses a spatial interpolation algorithm based on neighboring node data for data compensation and marks the reliability level of the data in the gas concentration characteristics.

[0042] Based on labeled and compensated gas concentration data, concentration time-series data analysis results, and sensor health status monitoring results, the module generates comprehensive gas concentration characteristics. Quantitative indicators of concentration change trends include parameters such as gradient change rate and fluctuation frequency spectrum characteristics, while sensor reliability assessment parameters integrate dimensions such as health score and confidence weight. For example, when an increased response delay is detected at a certain sensor node, the system will correspondingly reduce the feature weight of that node's data and simultaneously activate a backup sensor for cross-validation.

[0043] This processing flow forms an adaptive closed-loop optimization mechanism. When the system detects an abnormal concentration pattern in a specific area, it adjusts the size of the sliding time window and the sensitivity of the anomaly detection threshold through a feature feedback mechanism. This dynamic adjustment capability enables the system to adapt to complex environmental changes during transportation, such as automatically enhancing the monitoring of local concentration accumulation in tunnels and improving the accuracy of diffusion pattern identification in open areas.

[0044] Specifically, the hazardous chemical storage tank leakage detection device of the present invention includes a spatiotemporal context perception module configured to: receive visual detection results and gas concentration characteristics; construct a three-dimensional spatial field model using an octree data structure for discretization based on the spatial location information and monitoring values ​​included in the visual detection results and gas concentration characteristics; fill in the blank areas of the spatial monitoring points using a Kriging interpolation algorithm during the construction of the three-dimensional spatial field model; analyze the spatial diffusion law and temporal evolution trend of leakage characteristics in the three-dimensional spatial field model using a spatiotemporal graph attention network, wherein the node features of the spatiotemporal graph attention network include physical coordinates, monitoring values ​​from the visual detection results and gas concentration characteristics, and timestamps, and the edge weights are calculated based on the distance between nodes and preset fluid dynamic parameters; calculate the similarity between the visual sequence and the gas concentration sequence under spatiotemporal transformation using an improved dynamic time warping algorithm; and output a correlation evaluation result based on the analysis results and similarity calculation results of the spatiotemporal graph attention network.

[0045] After receiving visual detection results from the visual feature extraction module and gas concentration features from the gas concentration feature extraction module, the spatiotemporal context awareness module initiates a multi-source data correlation analysis process. First, based on the spatial location information and monitoring values ​​included in the visual detection results and gas concentration features, the module constructs a three-dimensional spatial field model using an octree data structure for discretization. The octree data structure recursively divides the three-dimensional space into a hierarchical voxel grid. Each voxel node stores the aggregated monitoring data value for the corresponding spatial region. For example, the surface of a storage tank is divided into several cubic units, each unit integrating infrared thermal imaging temperature data and gas concentration readings. This discretization method supports efficient spatial range querying and dynamic updates, facilitating subsequent analysis of the distribution characteristics of leakage features.

[0046] During the construction of the three-dimensional spatial field model, the module fills in the blank areas of the spatial monitoring points using the Kriging interpolation algorithm. Kriging interpolation, based on the assumption of spatial autocorrelation, uses a variogram model to quantify the spatial dependencies between monitoring points, providing optimal unbiased estimation for areas without sensor coverage. For example, in areas with insufficient sensor coverage, such as the top of a storage tank, the algorithm generates a continuous concentration field based on the gas concentration gradient trends of surrounding nodes, eliminating the impact of monitoring blind spots on the analysis results.

[0047] The module employs a spatiotemporal graph attention network to analyze the spatial diffusion patterns and temporal evolution trends of leakage characteristics in a three-dimensional spatial field model. The node features of the spatiotemporal graph attention network include physical coordinates, monitoring values ​​from visual detection results and gas concentration characteristics, and timestamps—three-dimensional information. Edge weights are dynamically calculated based on the Euclidean distance between nodes and preset fluid dynamic parameters. The attention mechanism emphasizes the influence of nodes near the leakage source by calculating the correlation weights between nodes. For example, when an abnormal gas concentration is detected at a tank valve, the network increases its attention to adjacent nodes in that area, accurately capturing the diffusion path of the leaked substance along the tank surface.

[0048] An improved dynamic time warping algorithm is used to calculate the similarity between visual sequences and gas concentration sequences under spatiotemporal transformations. The improved algorithm introduces a weight constraint mechanism to limit local scaling on the time axis, avoiding matching distortion caused by differences in sensor sampling frequencies. For example, when analyzing smoke diffusion video sequences and gas concentration fluctuation sequences, the algorithm dynamically adjusts the time window to align key event points, quantifying the temporal coupling between visual and gas signals.

[0049] Based on the analysis results and similarity calculation results of the spatiotemporal graph attention network, the module outputs a correlation evaluation result. The correlation evaluation result integrates multiple dimensions such as spatial diffusion consistency and temporal synchronization, and uses fuzzy logic fusion to generate a quantitative score. For example, when the direction of smoke diffusion shown in the visual sequence is highly consistent with the direction of the gas concentration gradient, and the time difference between their peak occurrences is within milliseconds, the system outputs a high correlation score, providing a reliable basis for subsequent decision fusion. The analysis process forms a closed-loop optimization mechanism, with the correlation evaluation result fed back to the front-end feature extraction module to dynamically adjust the monitoring focus area.

[0050] Specifically, the hazardous chemical storage tank leakage detection device of the present invention includes an adaptive decision fusion module configured to: receive visual inspection results, gas concentration characteristics, and correlation assessment results; employ DS evidence theory to perform uncertainty reasoning on the evidence sources, with the basic probability allocation function of DS evidence theory designed based on fuzzy membership functions; use confidence data from the visual inspection results, fluctuation indicators from the gas concentration characteristics, and correlation assessment results as evidence sources; use the correlation assessment results as evidence weight adjustment factors in the calculation of the basic probability allocation function; apply evidence synthesis rules to synthesize evidence, with the evidence synthesis rules introducing a conflict evidence redistribution algorithm; use Jousselme distance to measure the degree of conflict between evidence during the evidence synthesis process; and generate a joint decision probability distribution based on the results of evidence synthesis using the evidence synthesis rules.

[0051] The adaptive decision fusion module receives visual detection results from the visual feature extraction module, gas concentration features from the gas concentration feature extraction module, and correlation evaluation results from the spatiotemporal context awareness module, and then initiates the multi-source evidence fusion process. The module employs the DS evidence theory framework for uncertainty reasoning. This theory handles uncertain multi-source information by establishing a basic probability assignment function. The basic probability assignment function is designed based on a fuzzy membership function, mapping continuous feature values ​​to the degree of support for different hypothetical propositions. For example, it converts the visual detection confidence score into a basic probability assignment for the proposition that a leak event occurred.

[0052] The module uses confidence data from visual inspection results, fluctuation indicators from gas concentration characteristics, and correlation assessment results as three independent sources of evidence. Visual inspection confidence data reflects the reliability of the video analysis model's judgment of leakage characteristics; gas concentration fluctuation indicators reflect the degree of anomaly in concentration changes; and correlation assessment results characterize the consistency level between visual and gas signals in the spatiotemporal dimensions. These three sources of evidence provide a basis for judging leakage risk from different dimensions.

[0053] During the calculation of the basic probability assignment function, the module uses the correlation assessment results as an evidence weight adjustment factor. When the visual sequence and the gas concentration sequence exhibit a high spatiotemporal correlation, the correlation assessment results automatically enhance the basic probability assignment value of the corresponding evidence source; when the two types of signals show a significant divergence, their weight influence is reduced. This dynamic adjustment mechanism enables the system to adaptively cope with complex environmental interference, such as reducing the weight ratio of relying solely on visual evidence when a vehicle passes through a highly reflective area.

[0054] The module applies evidence synthesis rules to fuse multi-source evidence, and the synthesis rules adopted introduce a conflict evidence redistribution algorithm. When there is a high degree of conflict between different evidence sources, existing DS theory may produce counterintuitive results. However, the conflict evidence redistribution algorithm maintains the rationality of the reasoning results by identifying the degree of conflict and distributing the conflict quality proportionally to uncertain propositions. For example, when visual inspection shows a high-confidence leak but the gas concentration is not abnormal, the algorithm will not simply reject any evidence, but will allocate the conflicting part to the proposition category of "to be further verified".

[0055] During the evidence synthesis process, the module uses the Jousselme distance to measure the degree of conflict between pieces of evidence. The Jousselme distance quantifies the magnitude of contradiction between different evidence sources by calculating the degree of difference between evidence bodies. This distance metric comprehensively considers the probability distribution differences of evidence across all propositions, and more accurately reflects the degree of inconsistency between pieces of evidence than simply calculating the quality of conflict. When the distance exceeds a set threshold, the system triggers a weighted adjustment mechanism to reduce the fusion weight of highly conflicting evidence.

[0056] Based on the evidence synthesis results obtained from applying the evidence synthesis rules, the module generates a joint decision probability distribution. This probability distribution includes confidence assignments for various propositions such as normal state, suspected leakage, and confirmed leakage, and provides an overall uncertainty measure. For example, it might output a probability distribution with "normal state confidence 0.15, suspected leakage confidence 0.70, confirmed leakage confidence 0.10, and uncertainty 0.05," providing a quantitative basis for subsequent response decisions. The fusion process forms a closed-loop optimization, and the joint decision results are fed back to the front-end feature extraction module to guide the adjustment of sensor data acquisition strategies.

[0057] Specifically, the digital twin simulation module of the hazardous chemical storage tank leakage detection device of the present invention is configured to: receive joint decision results; establish a physical model of the storage tank and medium based on the parametric finite element method; determine the boundary condition parameters of the physical model through experimental data inversion; perform computational fluid dynamics simulation on the established and parametric physical model using the large eddy simulation method, with the computational fluid dynamics simulation coupling component transport equations and energy equations; generate multiple sets of simulation scenarios based on the joint decision results using the Monte Carlo algorithm to perform leakage scenario deduction; compare the simulation data generated by the leakage scenario deduction with the real-time monitoring data curves collected by the physical monitoring equipment using the dynamic time warping algorithm; correct the simulation parameters of the physical model in real time based on the comparison results of the simulation data and the real-time monitoring data curves using extended Kalman filtering to obtain the corrected simulation parameters; and perform simulation verification based on the corrected simulation parameters to generate verification results.

[0058] After receiving the joint decision results from the adaptive decision fusion module, the digital twin simulation module initiates a physical model-driven leakage scenario verification process. The module establishes a physical model of the storage tank and the medium based on the parametric finite element method. This model uses parametric modeling technology to transform the tank's geometry, material properties, and the medium's physical characteristics into adjustable computational parameters. The physical model employs the finite element discretization method to decompose the continuous medium into a finite number of elements, establishing a numerical solution format for the governing equations on each element, forming a computable three-dimensional digital entity.

[0059] The boundary condition parameters of the physical model are determined by inverting experimental data. This process utilizes monitoring data and operational records from historical leakage cases, employing optimization algorithms to deduce the actual values ​​of the model's boundary conditions. During the inversion process, the actual monitored concentration distribution and temperature field data are used as objective functions. Iterative calculations minimize the differences between simulated and measured data, thereby calibrating key parameters such as the model's inlet velocity and wall conditions. This data-driven approach effectively solves the model distortion problem caused by relying solely on theoretical formulas.

[0060] Large eddy simulation (LES) was employed to perform computational fluid dynamics simulations on the established and parameterized physical model. This method directly analyzes large-scale eddy structures through filtering operations to simulate small-scale eddy effects, achieving a balance between computational accuracy and efficiency. The simulation process coupled component transport equations and energy equations. The component transport equations describe the convection and diffusion process of leaked substances in the flow field, while the energy equations characterize the heat exchange behavior between the fluid and the tank, forming a multi-physics coupled leakage diffusion simulation environment.

[0061] Based on the joint decision-making results, multiple simulation scenarios are generated using the Monte Carlo algorithm to infer leakage scenarios. The Monte Carlo method generates a statistically representative set of virtual leakage scenarios by randomly sampling uncertainty parameters in the joint decision-making results, such as the probability distribution of leakage location and the range of leakage rate fluctuations. Each scenario corresponds to a specific set of initial conditions and boundary parameters, and the probability space distribution of leakage possibilities is constructed through batch simulations.

[0062] A dynamic time warping algorithm is employed to compare simulation data generated from leakage scenario deduction with real-time monitoring data curves collected by physical monitoring equipment. This algorithm uses dynamic programming to find the optimal nonlinear alignment path between the two time series, eliminating matching errors caused by differences in sampling frequency or time shifts in event occurrence. During the comparison, the algorithm simultaneously considers the morphological similarity and temporal correlation of the data curves, quantifying the degree of agreement between the simulation and measured data in both the time and value domains.

[0063] Extended Kalman filtering (EKF) is used to correct the simulation parameters of the physical model in real time based on the comparison between simulation data and real-time monitoring data curves. EKF linearizes the nonlinear system model, incorporating the simulated system state and observed data into a unified estimation framework, and continuously adjusts the model parameters through a prediction-correction loop. The correction process focuses on optimizing key parameters affecting diffusion, such as the turbulent diffusion coefficient and surface adsorption rate, so that the simulated behavior gradually approximates the actual physical process.

[0064] Simulation verification is performed based on the corrected simulation parameters, generating verification results. The verification process employs a cross-validation strategy, using a portion of historical data as the training set for parameter correction, and the remaining data as the test set to evaluate the model's predictive ability. Verification results include indicators such as model confidence scores and prediction error distribution statistics, while also outputting support or rejection conclusions for the joint decision-making results, thus providing a physical mechanism-level verification of the front-end detection results. The simulation verification process forms a closed-loop optimization; the corrected model parameters are fed back to the adaptive decision fusion module to optimize the evidence weight allocation strategy for subsequent joint decisions.

[0065] Specifically, the hazardous chemical storage tank leakage detection device of the present invention has a linkage control execution module configured to: receive joint decision-making results and verification results; determine graded response levels based on the joint decision-making results and verification results, with each graded response level corresponding to a specific combination of disposal procedures; perform decision consistency judgment on the joint decision-making results and verification results using a voting mechanism combined with confidence weighting, with the decision consistency judgment using a dynamic consistency threshold set based on historical data; manage the state transition of the disposal process using a finite state machine model, with each state of the finite state machine model having a timeout rollback mechanism; select and trigger the corresponding disposal procedure from the disposal procedure library based on the determined graded response level and decision consistency judgment results, the disposal procedure library including ventilation system control commands, leakage sealing device activation commands, and emergency communication trigger commands; after triggering the disposal procedure, collect disposal effect data and feed the disposal effect data back to the visual feature extraction module, gas concentration feature extraction module, spatiotemporal context perception module, adaptive decision fusion module, and digital twin simulation module for model parameter optimization.

[0066] After receiving the joint decision results from the adaptive decision fusion module and the verification results from the digital twin simulation module, the linkage control execution module initiates a tiered response control process. The module determines the tiered response level based on the probability distribution in the joint decision results and the confidence score in the verification results. The tiered response levels are based on a preset threshold mapping relationship, discretizing continuous probability values ​​into multiple risk levels, each corresponding to a specific combination of handling procedures. For example, when the joint decision results show a leakage probability exceeding 0.8 and the verification confidence score is higher than 0.9, the highest level response is triggered; when the probability value is in the range of 0.4 to 0.6 and the verification confidence score is moderate, a medium-level response is initiated.

[0067] A voting mechanism combined with confidence-weighted grading is employed to assess the consistency between the joint decision-making results and the verification results. The voting mechanism treats the outputs of multiple modules as independent voting sources, assigning weights based on the historical accuracy of each module. Confidence-weighted grading uses fuzzy logic to fuse the standard deviation of the joint decision-making results and the confidence interval width of the verification results to calculate the overall decision consistency score. The decision consistency assessment uses a dynamically set consistency threshold based on historical data. This threshold is periodically updated using a sliding window statistical method to adapt to changes in risk characteristics under different transportation environments. For example, the consistency threshold requirement is automatically increased in industrial zone road sections to reduce the risk of false alarms.

[0068] A finite state machine model is used to manage state transitions in the handling process. The finite state machine defines state nodes such as standby, warning, response, and recovery, with a timeout rollback mechanism set for each state. When the system remains in a specific state for more than a preset time, it automatically rolls back to the previous safe state and triggers a diagnostic procedure. State transition conditions are dynamically adjusted based on the graded response level and the decision consistency judgment result. For example, when the consistency score is below a threshold, even if the graded response level is high, the system will prioritize transitioning to the warning state rather than directly triggering handling.

[0069] Based on the determined tiered response level and the decision consistency assessment, the corresponding response procedure is selected and triggered from the response procedure library. The library includes standardized operation sets such as ventilation system control commands, leak sealing device activation commands, and emergency communication trigger commands. Procedure selection is implemented using a rule engine, taking the tiered response level as input and activating the corresponding command sequence through pattern matching. For example, a medium-level response triggers ventilation system control commands and a local area alarm, while a high-level response additionally activates leak sealing devices and all-personnel emergency communication.

[0070] After the response procedure is triggered, the module collects response effect data through a sensor network, including indicators such as the rate of gas concentration decline and the temperature field stabilization time. This response effect data is linked to the original decision context via a spatiotemporal index, forming a closed-loop feedback chain. The data is fed back to the visual feature extraction module, gas concentration feature extraction module, spatiotemporal context perception module, adaptive decision fusion module, and digital twin simulation module for model parameter optimization. The optimization process employs an incremental learning algorithm, dynamically adjusting feature extraction sensitivity, evidence fusion weights, and simulation boundary conditions based on the response effect, improving the system's response accuracy in similar scenarios. The control process forms a self-evolution mechanism, continuously accumulating response cases to optimize the response strategy.

[0071] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the function of the control device in the hazardous chemical storage tank leakage detection device as described above.

[0072] Thirdly, the present invention provides an electronic device comprising: a processor; a memory connected to the processor; the memory storing instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the function of the control device in the hazardous chemical storage tank leak detection device as described above.

[0073] In the actual operating environment of hazardous chemical transport vehicles, this invention achieves accurate identification of leakage risks through multi-module collaboration. Taking a tanker truck traveling through an industrial area at night as an example, an infrared thermal imaging high-definition video acquisition unit continuously captures the temperature field distribution on the tank surface, while a photoacoustic spectroscopy gas sensing array simultaneously monitors the surrounding gas concentration. The data preprocessing module achieves microsecond-level synchronization between video frames and gas sampling through a precise clock protocol, eliminating timing deviations caused by differences in sampling frequencies. The spatial registration stage, based on the coordinate transformation relationship established by the calibration board, maps the two-dimensional image coordinates to three-dimensional physical space, enabling the accurate location of the corresponding sensor position for abnormal areas identified in the video.

[0074] When the system detects fluctuations in gas concentration, the visual feature extraction module analyzes the corresponding infrared video sequence using an improved YOLOv7 model. The model employs a cross-stage local attention mechanism to enhance its ability to focus on leak features, and a path aggregation network fuses multi-scale features to identify areas of temperature anomalies. Simultaneously, the gas concentration feature extraction module analyzes the concentration time-series data through a gated recurrent unit, dynamically adjusts the analysis scale based on vehicle speed using a sliding time window, and adaptively updates the anomaly detection threshold using a particle swarm optimization algorithm to distinguish between genuine leaks and industrial exhaust gas interference.

[0075] The spatiotemporal context awareness module constructs an octree-structured spatial field model and fills in monitoring blind spots using Kriging interpolation. A spatiotemporal graph attention network analyzes the correlation between gas diffusion direction and the smoke trajectory in the video, while a dynamic time warping algorithm quantifies the spatiotemporal correlation of multi-source signals. For example, when the gas concentration gradient direction is detected to be highly consistent with the vapor diffusion path in the video, the system improves the correlation score.

[0076] The adaptive decision fusion module integrates multi-source information using DS evidence theory. When video analysis shows a high-temperature area but the gas concentration is normal, the module reduces the weight of visual evidence based on the correlation assessment results to avoid misjudging engine waste heat as a leak. The digital twin simulation module establishes a physical model of the storage tank based on parametric finite element method, simulates leakage scenarios through large eddy simulation, and uses extended Kalman filtering to correct model parameters in real time to approximate the actual physical process.

[0077] The coordinated control execution module executes tiered responses based on the joint decision probability distribution. When the consistency score output by multiple modules exceeds a dynamic threshold, the system triggers the corresponding level of handling procedure. A finite state machine manages the state transitions of the response process, with a timeout rollback mechanism for each state to prevent system stagnation. Handling effect data is fed back to each analysis module, and model parameters are optimized through incremental learning to form a closed-loop optimization. For example, after handling is completed, the system adjusts the sensitivity parameters of the gas feature extraction module based on the concentration decline rate to improve the detection accuracy of subsequent similar scenarios.

[0078] This multi-level collaborative mechanism effectively overcomes the limitations of single sensors being susceptible to environmental interference. During the transition from dawn to dusk, the system distinguishes between actual leaks and changes in light and shadow through spatiotemporal correlation analysis; when passing through contaminated areas, it verifies and eliminates external interference through physical simulation. By continuously learning characteristic patterns under different environments, the system gradually establishes intelligent detection capabilities adapted to complex transportation scenarios, significantly reducing false alarm rates while maintaining high sensitivity.

[0079] This invention effectively solves the false alarm problem caused by the isolation of video and gas sensor data by constructing a multi-level data fusion and collaborative decision-making mechanism. In complex transportation environments, the system uses a data preprocessing module to perform microsecond-level timestamp alignment and spatial registration of infrared thermal imaging video streams and photoacoustic spectral gas concentration data, establishing a unified spatiotemporal coordinate system. This basic data integration eliminates the inconsistencies caused by differences in acquisition frequency and spatial location of multi-source heterogeneous data, laying the foundation for subsequent correlation analysis.

[0080] When transport vehicles pass through industrial areas or other disruptive environments, the visual feature extraction module and the gas concentration feature extraction module process standardized data in parallel. The improved YOLOv7 model identifies leak-related visual features through a cross-stage local attention mechanism, and a long short-term memory network analyzes the temporal fluctuations in gas concentration. The two are analyzed independently but are correlated through a spatiotemporal context awareness module. This module constructs a three-dimensional spatial field model and uses a spatiotemporal graph attention network to analyze the correlation between visual signals and gas signals in terms of diffusion patterns and temporal evolution. For example, when abnormal fluctuations in gas concentration are detected, the system simultaneously analyzes the video sequence of the corresponding spatial region. If the gas diffusion direction is inconsistent with the smoke trajectory in the video, it is determined to be environmental interference rather than a real leak.

[0081] The adaptive decision fusion module dynamically integrates multi-source evidence using DS evidence theory. When video analysis shows high-confidence leakage characteristics but the gas concentration is not abnormal, the module reduces the weight of visual evidence based on the spatiotemporal correlation assessment results to avoid misjudging condensate vapor reflection as a leak. This dynamic weight adjustment mechanism enables the system to adapt to interference characteristics under different environmental conditions, such as automatically reducing reliance on single visual evidence in strong light reflection scenarios.

[0082] The digital twin simulation module enhances decision-making reliability through physical model verification. Based on a tank model built using parametric finite element method, leakage diffusion modes are simulated using large eddy simulation and compared with real-time monitoring data for verification. When uncertainties exist in multi-module joint decision-making, the simulation module provides supplementary verification at the physical mechanism level, forming a dual verification mechanism combining virtual and real data. For example, when video and gas data conflict, the simulation module analyzes the probability of leakage through computational fluid dynamics, providing a physical basis for the final decision.

[0083] This invention establishes a closed-loop optimization mechanism. The linkage control execution module executes tiered responses based on the outputs of multiple modules, and the treatment effect data is fed back to each analysis module for parameter self-adjustment. By continuously learning characteristic patterns under different environments, the system gradually builds anti-interference capabilities. In complex scenarios such as vehicles passing through polluted areas and changes in day and night lighting, the system effectively distinguishes between actual leaks and environmental interference through cross-verification of multi-source evidence, significantly reducing the false alarm rate.

[0084] Embodiment 1 of this invention: In a real-world application scenario where hazardous chemical transport vehicles pass through industrial areas during the day, the environment is subject to interference from various industrial exhaust gases. Existing single-gas detection methods are prone to false alarms due to their cross-sensitivity characteristics. In this embodiment, an infrared thermal imaging high-definition video acquisition unit deployed at the storage tank continuously captures the temperature field distribution on the tank surface, while a photoacoustic spectroscopy gas sensing array synchronously monitors the surrounding gas concentration. The data preprocessing module achieves microsecond-level synchronization between video frames and gas sampling data through a precise clock protocol, and completes spatial registration using the coordinate transformation relationship established by the calibration board. When the system detects gas concentration fluctuations, the visual feature extraction module analyzes the corresponding video sequence using an improved YOLOv7 model, finding no abnormal temperature rise or smoke features on the tank surface. The three-dimensional spatial field model constructed by the spatiotemporal context perception module shows that the gas diffusion direction is not clearly related to the geometric center of the tank, and the adaptive decision fusion module reduces the weight of gas evidence based on the correlation evaluation results. The digital twin simulation module verifies the consistency between the gas diffusion trajectory and the industrial pollution source through computational fluid dynamics simulation, and finally, the linkage control execution module maintains normal monitoring status, successfully avoiding false alarms caused by industrial exhaust gases.

[0085] Embodiment 2 of this invention: During nighttime transportation, condensation occurs on the tank surface due to temperature differences. Existing video analysis methods easily misjudge condensation reflection as a leak. In this embodiment, the infrared thermal imaging unit captures local temperature changes, but the gas concentration feature extraction module does not detect abnormal fluctuations. The visual feature extraction module enhances feature focusing through a cross-stage attention mechanism, and the multi-scale features fused by the path aggregation network show that the temperature change boundary is blurred and does not conform to leak characteristics. The spatiotemporal context perception module analyzes and finds that the temperature fluctuations have a periodic pattern, which matches the phenomenon of condensation evaporation. The adaptive decision fusion module uses gas concentration stability as strong evidence and adjusts the weight allocation by combining spatiotemporal correlation analysis. The digital twin simulation module generates multiple sets of condensation scenario simulations using the Monte Carlo method, and the simulation data highly matches the measured temperature curves. The linkage control execution module, based on the consistency judgment of multiple modules, only triggers the auxiliary dehumidification device and does not activate the leak handling procedure, forming a precise response to environmental disturbances.

[0086] Embodiment 3 of this invention addresses the issue of liquid sloshing in tanks on winding mountain roads where frequent acceleration and braking by vehicles cause liquid to slosh. Existing pressure sensors are prone to generating false leakage signals. In this embodiment, multi-source monitoring data is spatiotemporally aligned. The visual feature extraction module identifies no leakage features on the tank surface, while the gas concentration feature extraction module detects that concentration changes are synchronized with vehicle motion. The spatiotemporal context perception module analyzes the strong correlation between vibration signals and motion acceleration using a dynamic time warping algorithm, and the adaptive decision fusion module uses motion state parameters as contextual evidence. The digital twin simulation module simulates the liquid sloshing pattern using a coupled dynamics model, verifying that pressure fluctuations are within normal operating conditions. The system manages the handling process using a finite state machine, recording operating data while maintaining monitoring for model optimization, thus avoiding false alarms caused by vehicle motion.

[0087] Embodiment 4 of this invention: In a rainy season transportation scenario, rainwater impacting the tank surface generates abnormal acoustic signals. This embodiment integrates acoustic sensor data and extracts the characteristic frequency band of raindrop impact through wavelet packet transform. The visual feature extraction module simultaneously analyzes the rain line trajectory in the video data, and the gas concentration feature extraction module confirms no leakage-related gas anomalies. The spatiotemporal context perception module establishes a sound-visual-gas multimodal correlation model, identifying that the sound source location is consistent with the rainwater droplet distribution. The adaptive decision fusion module uses DS evidence theory to synthesize multi-source information, and the digital twin simulation module verifies the match between signal characteristics and rainfall intensity through acoustic simulation. The system activates an environmental adaptive monitoring mode, dynamically adjusting the detection threshold to maintain reliable operation under heavy rain conditions.

[0088] Embodiment 5 of the present invention: Under continuous braking conditions on long downhill sections, the heating of the braking system may interfere with infrared monitoring. In this embodiment, the visual feature extraction module distinguishes between the thermal radiation from the tank surface and the heat source of the brake drum through multispectral analysis, and the gas concentration feature extraction module detects the characteristic gas of braking dust. The spatiotemporal context perception module combines GPS positioning information to construct a terrain-thermal field correlation model and identify the correlation between heat source distribution and braking frequency. The adaptive decision fusion module introduces vehicle operating parameters as auxiliary evidence, and the digital twin simulation module verifies through thermodynamic simulation that the temperature field change conforms to the braking heating law. The system triggers a graded response mechanism, activating the auxiliary verification process while maintaining the operation of the main monitoring system, ensuring the reliability of detection under complex conditions.

Claims

1. A leak detection device for a hazardous chemical storage tank, characterized in that, This includes physical monitoring equipment located at the storage tank location and control devices that are connected in communication with the physical monitoring equipment; The physical monitoring equipment includes a high-definition video acquisition unit with infrared thermal imaging capabilities and a gas sensor array employing photoacoustic spectroscopy technology. The control device includes: a data preprocessing module for timestamping and spatially registering the video stream acquired by the high-definition video acquisition unit and the gas concentration readings acquired by the gas sensor array, performing noise reduction, and outputting standardized data; a visual feature extraction module for receiving video data from the standardized data, applying a target detection model to identify leak-related visual features, and outputting visual detection results with confidence levels; and a gas concentration feature extraction module for receiving gas concentration data from the standardized data, analyzing concentration time-series data and identifying abnormal fluctuations, and outputting gas concentration characteristics; and spatiotemporal... The context-aware module receives visual detection results and gas concentration characteristics, constructs a spatial field model to analyze the correlation between visual signals and gas signals in the spatiotemporal dimensions, and outputs a correlation assessment result. The adaptive decision fusion module receives visual detection results, gas concentration characteristics, and correlation assessment results, performs uncertainty reasoning, dynamically adjusts the weights of different evidence sources, and outputs a joint decision result. The digital twin simulation module receives the joint decision result to perform leakage scenario simulation, predicts the diffusion mode through simulation, and compares and verifies it with real-time monitoring data, outputting a verification result. The linkage control execution module integrates the joint decision result and the verification result, executes a graded response strategy to trigger an alarm or handling procedure.

2. The hazardous chemical storage tank leakage detection device according to claim 1, characterized in that, The data preprocessing module is configured to: use a precision clock protocol to perform microsecond-level synchronization processing on the video frames acquired by the high-definition video acquisition unit and the gas concentration sampling data acquired by the gas sensor array to obtain synchronization data; based on the synchronization data, establish the transformation relationship between the camera coordinate system and the sensor array coordinate system through a calibration board, and use Zhang Zhengyou calibration method to calculate intrinsic and extrinsic parameters to complete spatial registration and obtain registration data; perform multi-scale decomposition of the video data in the registration data using Gaussian pyramid, apply wavelet packet transform to divide the frequency band of the gas data in the registration data, reconstruct the signal after processing the transform coefficients through a threshold function, complete the noise reduction processing, and output a standardized data packet with a spatiotemporal index.

3. The hazardous chemical storage tank leakage detection device according to claim 2, characterized in that, The visual feature extraction module is configured to: receive video data from standardized data packets; apply an object detection model to process the video data. The object detection model is a neural network model that introduces a cross-stage local attention mechanism on the basis of YOLOv7, and adds an adaptive spatial feature pyramid structure to the neck network. Multi-scale feature maps are extracted from video data using an object detection model; shallow and deep features from the multi-scale feature maps are fused using a path aggregation network. Based on the fused features, bounding box coordinates and class confidence scores are generated; non-maximum suppression algorithm is applied to process the bounding box coordinates and class confidence scores, remove overlapping detections, and output visual detection results with confidence scores; During the extraction of multi-scale feature maps, feature importance evaluation results are generated simultaneously. The feature importance assessment results are fed back to the data preprocessing module; The data preprocessing module adaptively adjusts the threshold function in the noise reduction process based on the feature importance assessment results.

4. The hazardous chemical storage tank leakage detection device according to claim 3, characterized in that, The gas concentration feature extraction module is configured to: receive gas concentration data from standardized data packets; model and analyze the gas concentration time-series data using a gated cyclic unit long short-term memory network, and dynamically adjust the input sequence length through a sliding time window mechanism; during the analysis of the concentration time-series data, dynamically update the abnormal fluctuation detection threshold using a particle swarm optimization algorithm with an adaptive adjustment strategy incorporating inertial weights to identify concentration anomalies; and simultaneously analyze the attenuation characteristics of the sensor response curves in the gas sensing array to monitor the sensor health status. Based on the results of sensor health status monitoring, sensor data identified as abnormal are marked and compensated. Based on the analysis results of the marked and compensated gas concentration data, concentration time series data, and sensor health status monitoring results, gas concentration characteristics are generated. The gas concentration characteristics include quantitative indicators of concentration change trends and sensor reliability assessment parameters.

5. The hazardous chemical storage tank leakage detection device according to claim 4, characterized in that, The spatiotemporal context awareness module is configured to receive visual detection results and gas concentration characteristics; Based on the spatial location information and monitoring values ​​included in the visual detection results and gas concentration characteristics, a three-dimensional spatial field model is constructed using an octree data structure for discretization. During the construction of the three-dimensional spatial field model, Kriging interpolation is used to fill in the blank areas of the spatial monitoring points. A spatiotemporal graph attention network is employed to analyze the spatial diffusion patterns and temporal evolution trends of leakage characteristics in the three-dimensional spatial field model. The node features of the spatiotemporal graph attention network include physical coordinates, monitoring values ​​from the visual detection results and gas concentration characteristics, and timestamps. Edge weights are calculated based on the distance between nodes and preset fluid dynamic parameters. An improved dynamic time warping algorithm is used to calculate the similarity between the visual sequence and the gas concentration sequence under spatiotemporal transformations. Based on the analysis results and similarity calculation results of the spatiotemporal graph attention network, the correlation evaluation results are output.

6. The hazardous chemical storage tank leakage detection device according to claim 5, characterized in that, The adaptive decision fusion module is configured to: receive visual detection results, gas concentration characteristics, and correlation assessment results; use DS evidence theory to perform uncertainty reasoning on the evidence sources, and the basic probability assignment function of DS evidence theory is designed based on the fuzzy membership function; and use the confidence data in the visual detection results, the fluctuation index in the gas concentration characteristics, and the correlation assessment results as evidence sources. In the calculation of the basic probability allocation function, the correlation evaluation result is used as the evidence weight adjustment factor; evidence is synthesized by applying evidence synthesis rules, which introduce conflict evidence redistribution algorithms; during the evidence synthesis process, the Jousselme distance is used to measure the degree of conflict between evidence; and a joint decision probability distribution is generated based on the results of evidence synthesis using the evidence synthesis rules.

7. The hazardous chemical storage tank leakage detection device according to claim 6, characterized in that, The digital twin simulation module is configured to: receive joint decision results; establish a physical model of the tank and medium based on the parametric finite element method; determine the boundary condition parameters of the physical model through experimental data inversion; and perform computational fluid dynamics simulation on the established and parametric physical model using the large eddy simulation method, with the computational fluid dynamics simulation coupling the component transport equation and the energy equation. Based on the joint decision-making results, multiple sets of simulation scenarios are generated using the Monte Carlo algorithm to simulate leakage scenarios. The dynamic time warping algorithm was used to compare the simulation data generated from the leakage scenario with the real-time monitoring data curves collected by the physical monitoring equipment. By using extended Kalman filtering, the simulation parameters of the physical model are corrected in real time based on the comparison results of simulation data and real-time monitoring data curves, and the corrected simulation parameters are obtained. Simulation verification is then performed based on the corrected simulation parameters to generate verification results.

8. The hazardous chemical storage tank leakage detection device according to claim 7, characterized in that, The linkage control execution module is configured to: receive joint decision results and verification results; determine graded response levels based on joint decision results and verification results, with each graded response level corresponding to a specific combination of handling procedures; and perform a decision consistency judgment on the joint decision results and verification results using a voting mechanism combined with confidence weighting, with the decision consistency judgment using a dynamic consistency threshold set based on historical data. A finite state machine model is used to manage the state transitions of the response process, with a timeout rollback mechanism set for each state of the finite state machine model. Based on the determined graded response level and decision consistency judgment results, the corresponding response procedure is selected and triggered from the response procedure library, which includes ventilation system control commands, leak sealing device activation commands, and emergency communication trigger commands. After triggering the response procedure, response effect data is collected and fed back to the visual feature extraction module, gas concentration feature extraction module, spatiotemporal context perception module, adaptive decision fusion module, and digital twin simulation module for model parameter optimization.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the functions of the control device in the hazardous chemical storage tank leakage detection device as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, include: processor; Memory connected to the processor; The memory stores instructions that can be executed by the processor to enable the processor to perform the functions of the control device in the hazardous chemical storage tank leak detection device as described in any one of claims 1 to 8.

Images