AI Vision-Based Leakage Monitoring System and Method for Key Connection Points in Pipelines in Pipelines

Abstract

This application relates to the fields of computer vision and industrial safety monitoring, and discloses an AI vision-based system and method for monitoring leakage at key connection points of pipelines in utility tunnels. The system includes a data acquisition module, an intelligent analysis module, a 3D modeling module, a spatiotemporal verification module, and an alarm prediction module. Through dynamic fusion of visible light, infrared, and pressure data, combined with neural network optimization driven by physical simulation parameters, it achieves real-time detection and precise 3D positioning of leakage areas. Furthermore, through optical flow tracking and fluid dynamics verification, it filters out false alarms caused by non-physical laws. Based on Bayesian decision-making and diffusion equation prediction, it generates a leakage risk heat map and triggers tiered alarms. This invention solves the technical problems of low leakage detection accuracy, high false alarm rate, and inaccurate positioning in complex utility tunnel scenarios. It possesses advantages such as high environmental adaptability, strong anti-interference capability, and forward-looking risk prediction, and is suitable for intelligent safety monitoring of underground utility tunnels and petrochemical pipelines.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and industrial safety monitoring technology, specifically to a system and method for monitoring leakage at key connection points of pipelines in utility tunnels based on AI vision. Background Technology

[0002] In the field of pipeline leakage monitoring, traditional methods mainly rely on manual inspections and fixed threshold sensor detection, which suffer from poor environmental adaptability and high false alarm rates. Although vision-based detection technology can identify some leakage characteristics, the reliability of the detection results decreases significantly under complex lighting changes, pipeline surface reflections, and dynamic background interference.

[0003] While existing deep learning solutions have improved detection accuracy in some scenarios, they are heavily reliant on training data and lack generalization ability when dealing with small sample scenarios involving diverse pipe materials and varying leakage patterns. Furthermore, multi-sensor fusion solutions are difficult to deploy on a large scale in industrial settings due to high hardware costs and complex data alignment. Additionally, existing technologies generally lack in-depth verification of the physical laws governing leakage, making it impossible to effectively distinguish between real leaks and dynamic environmental interference, thus limiting the practicality and reliability of detection systems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI vision-based system and method for monitoring leakage at key connection points of pipelines in utility tunnels. This solves the problems of high false alarm rate, low positioning accuracy, and lack of verification of physical laws in existing leakage detection technologies in complex utility tunnel environments.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a leakage monitoring system for key connection points of pipelines in utility tunnels based on AI vision, comprising:

[0006] The data acquisition module is used to simultaneously acquire visible light images, infrared thermal imaging data, and pressure sensing data to form a multimodal dataset.

[0007] The intelligent analysis module performs feature fusion analysis on the multimodal dataset based on a dynamic heterogeneous neural network, and outputs detection results including the coordinates of the leakage area and the leakage probability value.

[0008] The 3D modeling module receives the coordinates of the leakage area from the detection results, performs spatial mapping using a 3D model of the pipe gallery constructed using the neural radiation field, and corrects the temperature distribution field by combining real-time infrared thermal imaging data, outputting a set of leakage points with three-dimensional coordinates.

[0009] The spatiotemporal verification module receives the set of leakage points and executes the following:

[0010] The diffusion rate in the time dimension is calculated by tracking the movement trajectory of the leakage area using the multimodal optical flow method.

[0011] By dynamically adjusting the fluid dynamics diffusion equations based on the pipe material parameters in the neural radiation field model, the diffusion mode in the spatial dimension is verified.

[0012] When the diffusion rate in the time dimension and the diffusion pattern in the spatial dimension meet the preset physical law matching degree, a leakage event that has passed the verification is generated.

[0013] The alarm prediction module receives the verified leakage events and their corresponding leakage probability values, and executes the following through a Bayesian decision model:

[0014] When the leakage probability value exceeds the dynamic optimization threshold and the verification matching degree reaches the confidence interval, a graded alarm is triggered.

[0015] Based on the hydrodynamic diffusion equation, leakage paths are predicted, and early warning information including time-space diffusion heatmaps is generated.

[0016] Preferably, the data acquisition module includes:

[0017] Visible light camera, infrared thermal imager, pressure sensor, and laser spectrometer, among which:

[0018] The laser spectrometer is configured to detect the concentration of gas components around the pipeline and generate data on abnormal gas distribution.

[0019] The visible light camera and the infrared thermal imager achieve pixel-level spatial alignment through a coaxial gimbal, forming optical-thermal joint sensing data;

[0020] The pressure sensor acquires time-series data of pressure fluctuations inside the pipeline;

[0021] The multimodal data consists of spatiotemporally aligned optical images, thermodynamic images, pressure fluctuation sequences, and gas concentration matrices.

[0022] Preferably, the intelligent analysis module includes:

[0023] The dynamic heterogeneous neural network unit consists of a first branch network and a second branch network:

[0024] The first branch network is built on a lightweight convolutional network and is used to process visible light images and infrared thermal imaging data in real time.

[0025] The second branch network is trained based on data generated by meta-learning framework and physical simulation, and is used for leakage detection in scenarios with a small training sample size;

[0026] The dynamic weight allocation unit dynamically adjusts the output weights of the two branch networks based on the physical simulation confidence of the second branch network, the infrared anomaly intensity value calculated by the deviation between the thermal imaging data and the reference temperature field, and the fluctuation amplitude of the pressure sensing data.

[0027] Preferably, the data generated by the physical simulation includes:

[0028] Simulation data of diffusion paths of different pipe materials in leakage scenarios;

[0029] Data on the mapping relationship between the thermodynamic parameters of pipe materials and their infrared radiation characteristics.

[0030] Preferably, the 3D modeling module includes:

[0031] The neural radiation field modeling unit is used to construct a 3D neural radiation field model of the pipe gallery based on multi-view visible light and infrared data, wherein:

[0032] The neural radiation field modeling unit combines the thermal radiation characteristics of the pipe material to correct the temperature distribution of the 3D model.

[0033] The temperature distribution is generated by dynamically adjusting the coefficient and fusing real-time infrared data with material thermodynamic parameters.

[0034] The leak point location unit is used to map the leak detection results output by the intelligent analysis module to the 3D model, and to track the movement trajectory of the leak area using optical flow method to calculate the matching degree between the leak point and the 3D model.

[0035] Preferably, the calculation process of the dynamic adjustment coefficient includes:

[0036] Construct a nonlinear function f based on a multilayer perceptron, whose input layer receives the real-time pipeline pressure P. pipe With ambient temperature T env The normalized parameter;

[0037] The function f is optimized and trained using a meta-learning framework. The training dataset includes:

[0038] The thermal conductivity matrix of the pipe material generated by physical simulation;

[0039] Temperature field distribution data under different pressure conditions;

[0040] Temperature-pressure correlation data of leakage events from historical maintenance records;

[0041] In real-time calculation, the current pressure sensor reading and temperature sensor reading are standardized and preprocessed, and then input into the trained function f;

[0042] The output layer generates dynamic adjustment coefficients through the Sigmoid activation function, which are used to weight and fuse real-time infrared data with material thermodynamic parameters.

[0043] Preferably, the spatiotemporal verification module includes:

[0044] The time dimension verification unit is used to track the movement trajectory of the leakage area using optical flow method, calculate the leakage diffusion rate, and compare it with the preset allowable diffusion rate threshold of the pipe material.

[0045] The spatial dimension verification unit verifies the diffusion pattern of the leakage area based on the fluid dynamics diffusion equation, wherein the parameters of the diffusion equation are dynamically adjusted according to the pipe material information and real-time pressure data in the neural radiation field model.

[0046] The consistency determination unit determines and filters out false alarms when the verification results of the time dimension and the space dimension are logically conflicting.

[0047] Preferably, the optical flow tracking method employs a multimodal optical flow fusion algorithm, including:

[0048] Calculate the optical flow field based on grayscale gradient for visible light images;

[0049] Calculate the optical flow field based on the temperature gradient from infrared thermal imaging data;

[0050] By dynamically weighting and fusing two optical flow fields, the motion trajectory of the leakage area is generated.

[0051] Preferably, the alarm prediction module includes:

[0052] Bayesian decision-making unit, execution:

[0053] Based on the historical leakage event database, the alarm prior probability is dynamically updated. When the real-time detected leakage probability value exceeds the dynamically adjusted threshold, a high-confidence alarm signal is generated.

[0054] The sensitivity range of the dynamic adjustment threshold is adaptively adjusted based on the fluctuation range of the current ambient temperature and equipment operating pressure in the utility tunnel.

[0055] Diffusion prediction unit, execute:

[0056] By combining the three-dimensional pipe parameters provided by the neural radiation field model, a leakage diffusion path prediction map for a future preset time period is generated through fluid dynamics simulation.

[0057] Output a visualized heatmap with labeled diffusion direction, radius of influence, and risk level;

[0058] The dynamic threshold optimization unit executes:

[0059] Construct a multi-dimensional state space that includes ambient temperature and humidity, pressure fluctuation frequency, and historical false alarm records;

[0060] The dynamic adjustment threshold is iteratively optimized using a reinforcement learning algorithm to balance the conflicting metrics of real-time false alarm rate and false negative rate.

[0061] The alarm execution unit, in response to the high-confidence alarm signal, performs a tiered response action:

[0062] Level 1 Response: Activate the on-site audible and visual alarm devices in the utility tunnel and generate an emergency maintenance work order;

[0063] Level 2 Response: Push early warning information to mobile terminals and mark video streams of key monitoring areas.

[0064] This invention also provides a method for monitoring leakage at key connection points of pipelines in utility tunnels based on AI vision, comprising the following steps:

[0065] Simultaneously acquire visible light images, infrared thermal imaging data, and pressure sensing data of the utility tunnel area to generate spatiotemporally aligned multimodal data packets;

[0066] The multimodal data packets are input into a dynamic heterogeneous neural network for feature fusion, and detection results containing the coordinates and probability values ​​of the leakage area are generated in real time.

[0067] The detection results are mapped to a 3D pipe gallery model constructed from a neural radiation field, and the temperature field distribution is corrected by combining real-time infrared data to output a three-dimensional set of leak point coordinates.

[0068] The movement trajectory of the leakage area is tracked by multimodal optical flow method, and the spatiotemporal consistency of the leakage diffusion path is verified based on the hydrodynamic diffusion equation, generating verification results.

[0069] When the verification result meets the preset confidence conditions, a graded alarm is triggered based on the leakage probability value, and the leakage diffusion trend is predicted based on the corrected 3D model.

[0070] This invention provides a system and method for monitoring leakage at key connection points of pipelines in utility tunnels based on AI vision.

[0071] It has the following beneficial effects:

[0072] 1. This invention fuses multimodal data through a dynamic heterogeneous neural network, significantly improving detection accuracy while ensuring real-time performance. Lightweight network branches achieve millisecond-level response, while high-precision branches are optimized based on physical simulation data, effectively solving the problem of missed detections in small sample scenarios and making it suitable for rapid leak identification in complex industrial environments.

[0073] 2. The dynamic fusion mechanism of multimodal data in this invention can adapt to changes in illumination, surface reflection, and temperature fluctuations. By dynamically adjusting the weights of infrared and visible light data, it ensures stable detection of leakage areas even under extreme conditions such as low illumination and high reflectivity, thus improving the system's environmental robustness.

[0074] 3. The dynamic fusion mechanism of multimodal data in this invention can adapt to changes in illumination, surface reflection, and temperature fluctuations. By dynamically adjusting the weights of infrared and visible light data, it ensures stable detection of leakage areas even under extreme conditions such as low illumination and high reflectivity, thus improving the system's environmental robustness.

[0075] 4. This invention utilizes a dual verification mechanism based on optical flow tracking and fluid dynamics to deeply integrate AI detection results with physical laws, effectively filtering out false alarms caused by non-physical laws. This mechanism can distinguish between real leakage and dynamic interference (such as equipment movement or temporary foreign objects), significantly improving the reliability of detection results.

[0076] 5. This invention achieves intelligent adaptation of alarm triggering logic through Bayesian decision-making models and dynamic threshold optimization. Combined with fluid dynamics diffusion prediction, the system not only provides real-time alarms but also predicts leakage diffusion trends, providing forward-looking decision support for the safe operation and maintenance of utility tunnels and enhancing risk prevention and control capabilities. Attached Figure Description

[0077] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0078] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0079] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] Please see the appendix Figure 1 This invention provides an AI vision-based leakage monitoring system for critical connection points in utility tunnels, aiming to achieve high-precision leakage monitoring of key connection points within the tunnel. This system can be widely applied in complex environments such as underground integrated utility tunnels, oil pipelines, water supply pipelines, and chemical pipelines, enabling accurate leakage detection, three-dimensional positioning, trend prediction, and intelligent alarms, thereby improving pipeline safety and maintenance efficiency.

[0081] like Figure 1 As shown, the AI ​​vision-based pipeline critical connection point leakage monitoring system may include:

[0082] Data acquisition module: used to acquire visible light images, infrared thermal imaging data, pressure sensing data, and laser spectral analysis data;

[0083] Intelligent analysis module: Based on dynamic heterogeneous neural network (DHNN), it performs leakage detection on multimodal data and calculates the leakage probability;

[0084] 3D Modeling Module: Employs Neural Radiation Field (NeRF) technology to construct a 3D model of the utility tunnel environment, and combines the results from the intelligent analysis module to locate leakage points;

[0085] Spatiotemporal verification module: Combining optical flow tracing algorithm and fluid dynamics analysis, it verifies the consistency of the detected leakage area in time and space dimensions;

[0086] Alarm prediction module: It uses a Bayesian decision algorithm to trigger alarms and predicts the leakage diffusion trend based on a fluid dynamics model.

[0087] The following is a detailed description of each module in the system of this invention, providing a comprehensive explanation of the specific implementation principles, technical details, and processes of each module.

[0088] The data acquisition module is used to simultaneously acquire visible light images, infrared thermal imaging data, pressure sensing data, and gas composition data of the pipe gallery area, and to preprocess the multimodal data.

[0089] For example, the following sensors are arranged around the key connection points of the utility tunnel: a visible light camera (resolution ≥1920×1080, frame rate 30fps), an infrared thermal imager (wavelength range 8-14μm), a pressure sensor (accuracy ±0.1MPa), and a laser spectrometer (detection accuracy 0.1ppm). Each sensor is connected via a hardware synchronization unit to ensure a timestamp error of less than 1ms. Preferably, the visible light camera and the infrared thermal imager are coaxially mounted to eliminate viewing angle deviation.

[0090] First, contrast-limited adaptive histogram equalization (CLAHE) is performed on the visible light image to eliminate image noise under low-light conditions in the pipe gallery. Specifically, the image is segmented into 8×8 sub-regions, with a histogram cropping threshold of 0.03 for each sub-region, and the global image is reconstructed using bilinear interpolation. The signal-to-noise ratio (SNR) of the processed image is improved to over 30dB, effectively extracting texture features from the leakage area.

[0091] Infrared thermal imaging data undergoes temperature normalization to eliminate interference from ambient temperature fluctuations. The calculation formula is as follows:

[0092]

[0093] Among them, TIR The original temperature value, μ T The historical average temperature, σ T Let |T| be the standard deviation. Preferably, when |T| is... norm When the temperature reaches 2.5, it is determined to be an abnormal temperature, triggering the marking of the leakage candidate area.

[0094] The formula for calculating the rate of change of pipeline pressure in real time is:

[0095]

[0096] Among them, P t The current pressure value is represented by Δt, which is the sampling interval (default 1 second). If ΔP exceeds the threshold θP (set to 20% of the historical fluctuation range), a pressure anomaly event is generated. Furthermore, θP is dynamically adjusted based on the pipe material; for example, θP for steel pipes is 15% lower than that for polyethylene pipes.

[0097] The time axes of data from different sensors are aligned using the Dynamic Time Warping (DTW) algorithm. The cost matrix D is defined as follows:

[0098] D(i,j)=|X i -Y j |+min(D(i-1,j),D(i,j-1),D(i-1,j-1))

[0099] Among them, X i and Y j These are time series points for visible light and infrared data, respectively. Preferably, a sliding window mechanism (window length 5s) is used to achieve real-time alignment, with a computation delay of less than 50ms.

[0100] The gas concentration data acquired by the laser spectrometer undergoes noise reduction using a Kalman filter and is spatially matched with infrared temperature anomaly regions. For example, when the methane concentration... And the corresponding region T norm When the value is greater than 2.0, it is determined to be a gas leakage event.

[0101] The preprocessed multimodal data is transmitted to the intelligent analysis module via Gigabit Ethernet. The data packet format is defined as a quadruple:

[0102] DataPacket= <Image,T norm ,ΔP,C gas >

[0103] Where Image is the enhanced visible light / infrared fused image, and C gas This is the gas concentration vector.

[0104] This implementation method achieves low-latency and high-precision acquisition of utility tunnel environmental data through synchronous acquisition and preprocessing by multiple sensors.

[0105] The intelligent analysis module is used to perform leakage detection and probability calculation on the multimodal data acquired by the data acquisition module. Its core includes a dynamic heterogeneous neural network unit and a dynamic weight allocation unit.

[0106] For example, the first branch network adopts the lightweight MobileNetV4 architecture, taking a fused visible light and infrared image (size 256×256×3) as input and outputting a heatmap of the leakage area. Specifically, the network layer configuration is: 5 depthwise separable convolutional layers (kernel size 3×3), an SE attention module (compression ratio 16), a global average pooling layer, and a Softmax classification layer. This branch achieves an inference latency of less than 30ms on an NVIDIA Jetson AGX Xavier edge device.

[0107] The second branch network is built based on the Model-Independent Meta-Learning (MAML) framework, with the input being a leakage dataset generated from physical simulation. Preferably, the physical simulation data simulates the leakage diffusion process of different pipe materials (steel, polyethylene, PVC) using the finite element method, generating multimodal data including temperature, pressure, and fluid velocity fields. The network structure is ResNet-34, and the initial parameters are updated during training using second-order gradient optimization.

[0108]

[0109] Where α is the meta-learning rate (default 0.01), L task This is the task loss function.

[0110] The dynamic weight allocation unit is based on the physical simulation results (S) sim ), infrared anomaly intensity (I IR ) and pressure data (P pipr Calculate branch weights:

[0111]

[0112] Wherein, φ(·) is a nonlinear function optimized through meta-learning (implemented using a two-layer fully connected network), and ∈ is a small constant (1e-6) to prevent division by zero. Preferably, the infrared anomaly intensity value is obtained by calculating the standardized deviation between the current detection area temperature and the historical average temperature of the same location using statistical analysis methods. Preferably, when w1>0.7, the second branch network is activated for high-precision re-inspection.

[0113] The output feature map of the first branch network (size 16×16×512) and the feature map of the second branch network (size 16×16×512) are fused by dynamic weighting:

[0114] F fusion =w1·F branch1 +(1-w1)·F bramch2

[0115] The fused feature map is then reduced in dimension by a 1×1 convolutional layer to output the leakage probability value P. leak ∈[0,1]. Furthermore, when P leak A high-confidence leakage event is generated when the value is >0.85.

[0116] Taking steel pipes as an example, the physical simulation process includes:

[0117] 1. Create a 3D model of the pipeline and set the material parameters (elastic modulus 200 GPa, Poisson's ratio 0.3);

[0118] 2. Simulate liquid leakage caused by a tiny crack (0.1 mm in size) in COMSOL Multiphysics;

[0119] 3. Output simulation data including temperature gradient field (ΔT=5-15℃) and flow velocity field (0.1-1.2m / s).

[0120] This data is used for few-shot training of the second branch network (only 50 sets of simulation data are needed for each material class).

[0121] For example, 3D leakage models of steel, polyethylene, and PVC pipes were established using finite element simulation software. First, the crack size (range 0.1-2.0 mm), fluid type (water / oil / gas), and pressure parameters (0.5-5.0 MPa) were set. The diffusion path was calculated by solving the Navier-Stokes equations.

[0122]

[0123] Where ρ is the fluid density, μ is the dynamic viscosity, v is the velocity field, and p is the pressure field. The output data includes the diffusion path coordinates (x, y) over time. t ,y t ,z t and flow velocity v t This forms the spatiotemporal feature dataset for training the second branch network.

[0124] The thermal conductivity λ and specific heat capacity c of the pipe material were determined experimentally. p And the surface emissivity ε, to establish a quantitative relationship with infrared radiation intensity:

[0125] E IR =εσT 4 +(1-ε)E amb

[0126] Where σ is the Stefan-Boltzmann constant, T is the surface temperature, and E is the surface temperature. amb The ambient radiation intensity is used. Preferably, the mapping relationship between steel (ε = 0.85) and polyethylene (ε = 0.92) is embedded into the simulation model to generate infrared temperature field data that matches the material.

[0127] Multiphysics Coupled Simulation Process:

[0128] 1. Geometric modeling: Construct a 3D model of the pipe with cracks in COMSOL (example crack length 1.5mm);

[0129] 2. Parameter assignment: Input the thermodynamic parameters of the material (steel: λ=45.8W / (m·K), c p = 465 J / (kg·K);

[0130] 3. Boundary condition settings: Define fluid inlet pressure (example value 2.5MPa) and ambient temperature (25℃);

[0131] 4. Solving the equations: Solve the fluid dynamics equations and the heat conduction equations simultaneously.

[0132] 5. Data Output: Generates point clouds containing diffusion paths and temperature gradient fields. and infrared radiation intensity E IR The simulation dataset.

[0133] This implementation method uses diffusion path data generated through physical simulation to accurately reflect the differences in leakage patterns of pipes made of different materials.

[0134] The 3D modeling module is used to construct a neural radiation field model of the pipe gallery environment and to accurately locate leakage points. Its core includes a neural radiation field modeling unit and a leakage point location unit.

[0135] For example, multi-view visible light cameras and infrared thermal imagers are used to acquire environmental image data of the pipe gallery, which is then input into a neural radiation field (NeRF) model for 3D reconstruction. The model generates RGB color and temperature distributions from any viewpoint using volume rendering equations.

[0136]

[0137] Where σ is the density function, c is the color function, and T IR M is the actual temperature measured by the infrared sensor. 材质 γ represents the thermodynamic parameters of the material (steel: 0.85, polyethylene: 0.92), and γ is the dynamic adjustment coefficient.

[0138] The dynamic adjustment coefficient γ is generated using a multilayer perceptron (MLP) optimized by meta-learning, with the real-time pipeline pressure P as input.pipe With ambient temperature T env :

[0139] γ=f MLP (P pipe ,T env ;θ)

[0140] Where θ represents the network parameters, obtained through pre-training on a physical simulation dataset. Preferably, when P pipe >2.0MPa and T env At temperatures above 40℃, the value of γ is reduced to 0.3-0.5 to suppress infrared noise interference.

[0141] Specifically, the calculation process for the dynamic adjustment coefficient includes:

[0142] Construct a nonlinear function f based on a multilayer perceptron, whose input layer receives the real-time pipeline pressure P. pipe With ambient temperature T env The normalized parameter;

[0143] The function f is optimized and trained using a meta-learning framework. The training dataset includes:

[0144] The thermal conductivity matrix of the pipe material generated by physical simulation;

[0145] Temperature field distribution data under different pressure conditions;

[0146] Temperature-pressure correlation data of leakage events from historical maintenance records;

[0147] In real-time calculation, the current pressure sensor reading and temperature sensor reading are standardized and preprocessed, and then input into the trained function f;

[0148] The output layer generates dynamic adjustment coefficients through the Sigmoid activation function, which are used to weight and fuse real-time infrared data with material thermodynamic parameters.

[0149] The leak location unit receives the leak area coordinates (x, y) output by the intelligent analysis module. i ,y i The motion trajectory is tracked using optical flow. A matching index is defined as follows:

[0150]

[0151] Among them, F NeRF A three-dimensional temperature field rendered for the NeRF model, F IR This is the actual temperature value measured by the infrared sensor. When D... match A value less than 0.15 is considered a valid leak point; otherwise, it is marked as a false detection.

[0152] This implementation method achieves high-precision reconstruction of the three-dimensional temperature field of the utility tunnel through neural radiation field modeling and dynamic temperature correction.

[0153] The spatiotemporal verification module is used to verify the physical consistency of leakage detection results in the time and space dimensions. Its core includes a time dimension verification unit, a space dimension verification unit, and a consistency determination unit.

[0154] For example, the optical flow field based on grayscale gradient is calculated for a visible light image, and the Horn-Schunck algorithm is used to solve for the velocity vector (u). v ,v v ):

[0155] I x u v +I y v v +I t =0

[0156] Among them, I x ,I y For spatial gradient, I t The time gradient is used to calculate the temperature gradient optical flow field (u) from infrared thermal imaging data. v ,v v By minimizing temperature change error:

[0157]

[0158] By dynamically weighting and fusing two optical flow fields, a comprehensive motion trajectory of the leakage area is generated:

[0159] u=αu v +(1-α)u t v=αv v +(1-α)v t

[0160] Where, α=σ(P pipe / P max ), σ(·) is the Sigmoid function, P max This represents the maximum pressure that the pipeline can withstand.

[0161] Calculate the diffusion rate of the leakage area based on the optical flow field:

[0162]

[0163] v leak With respect to the permissible diffusion rate threshold v of the pipe material th Comparison. For example, steel pipe v th =0.2m / s, polyethylene pipe v th = 0.5 m / s. When v leak>1.5v th It is marked as an exception.

[0164] The thermal conductivity λ of the pipe material and the real-time pressure P are based on the output of the neural radiation field model. pipe Dynamically adjust the parameters of the diffusion equation:

[0165]

[0166] Wherein, the diffusion coefficient D = k1λ + k2P pipe k1 = 0.03 and k2 = 0.12 are empirical constants. Solving the equation yields the theoretical diffusion region A. theory , and detection area A detect Calculate overlap:

[0167]

[0168] When IoU < 0.6, it is determined to be spatially inconsistent.

[0169] If the time dimension verification is abnormal (v) leak >1.5v th Furthermore, insufficient spatial overlap (IoU < 0.6) triggers the consistency determination unit. For example, when there is a logical conflict between the two (e.g., high-speed diffusion but high spatial matching), a Bayesian decision model is used:

[0170]

[0171] Here, H0 represents the false positive hypothesis. The detection result is filtered when P(false positive) > 0.8.

[0172] This implementation method enables cross-sensor verification of leakage trajectory through multimodal optical flow fusion and hydrodynamic constraints.

[0173] The alarm prediction module is used to comprehensively assess leakage risks and generate early warning signals. Its core components include a Bayesian decision unit, a diffusion prediction unit, and a dynamic threshold optimization unit.

[0174] Specifically, the alarm prediction module includes:

[0175] Bayesian decision-making unit, execution:

[0176] Based on the historical leakage event database, the alarm prior probability is dynamically updated. When the real-time detected leakage probability value exceeds the dynamically adjusted threshold, a high-confidence alarm signal is generated.

[0177] The sensitivity range of the dynamic adjustment threshold is adaptively adjusted based on the fluctuation range of the current ambient temperature and equipment operating pressure in the utility tunnel.

[0178] Diffusion prediction unit, execute:

[0179] By combining the three-dimensional pipe parameters provided by the neural radiation field model, a leakage diffusion path prediction map for a future preset time period is generated through fluid dynamics simulation.

[0180] Output a visualized heatmap with labeled diffusion direction, radius of influence, and risk level;

[0181] The dynamic threshold optimization unit executes:

[0182] Construct a multi-dimensional state space that includes ambient temperature and humidity, pressure fluctuation frequency, and historical false alarm records;

[0183] The dynamic adjustment threshold is iteratively optimized using a reinforcement learning algorithm to balance the conflicting metrics of real-time false alarm rate and false negative rate.

[0184] The alarm execution unit, in response to the high-confidence alarm signal, performs a tiered response action:

[0185] Level 1 Response: Activate the on-site audible and visual alarm devices in the utility tunnel and generate an emergency maintenance work order;

[0186] Level 2 Response: Push early warning information to mobile terminals and mark video streams of key monitoring areas.

[0187] First, a prior probability distribution is constructed based on historical leakage data, and modeled using a Beta distribution:

[0188] P(H1)~Beta(α,β)

[0189] Where α represents the number of historical leaks and β represents the number of normal events. When a leak event is detected in real time, the posterior probability is updated using Bayes' theorem:

[0190]

[0191] Wherein, P(X|H1) is the leakage probability P output by the intelligent analysis module. leak The calculation assumes no leakage (H0). An alarm is triggered when P(H1|X) > 0.9.

[0192] Based on the pipe material parameter λ and real-time pressure P provided by the neural radiation field model pipe Solve the convection-diffusion equation to predict the diffusion range in the next 30 minutes:

[0193]

[0194] Wherein, the diffusion coefficient D = 0.05λ, and the velocity field v = k·P pipe •n (k = 0.12, n is the pipe normal vector). The finite difference method is used to discretize and solve the problem, and the output is a heat map of pollutant concentration distribution.

[0195] The alarm threshold θ is dynamically adjusted using the Q-learning algorithm. alarm The state space is defined as S = {T} env ,P pipe The action space is defined as the threshold adjustment Δθ ∈ {-0.1, 0, +0.1}. The reward function is designed as follows:

[0196] R=w1(1-FPR)+w2(1-FNR)-w3|Δθ|

[0197] Where FPR is the false positive rate, FNR is the false negative rate, and the weighting coefficients are w1 = 0.6, w2 = 0.3, and w3 = 0.1. The Q-table is updated every 24 hours, with a learning rate η = 0.2 and a discount factor γ = 0.9.

[0198] This implementation method combines Bayesian decision-making with fluid dynamics prediction to achieve accurate leakage alarm.

[0199] Please see the appendix Figure 2 The present invention also provides a method for monitoring leakage at key connection points of pipelines in utility tunnels based on AI vision, comprising the following steps:

[0200] S1. Multimodal data collaborative acquisition;

[0201] Environmental data is collected synchronously using visible light cameras, infrared thermal imagers, and pressure sensors deployed at key connection points in the utility tunnel. Visible light cameras capture detailed surface textures of the pipes, infrared thermal imagers detect areas of abnormal temperature, and pressure sensors monitor internal pressure fluctuations in real time. Data from multiple sensors is synchronized via a hardware-level clock to ensure timestamp alignment, providing consistent input for subsequent analysis.

[0202] S2, Dynamic Heterogeneous Neural Network Detection;

[0203] The collected multimodal data is input into a dynamic heterogeneous neural network:

[0204] The lightweight branch processes visible light and infrared fused images in real time to quickly locate suspected leakage areas;

[0205] The high-precision branch uses leakage data generated by physical simulation (simulating the diffusion path of pipes of different materials) to perform detailed analysis on suspected areas;

[0206] The dynamic weight allocation adaptively adjusts the output weights of the two branches based on the infrared anomaly intensity, pressure fluctuations, and simulation results, balancing real-time performance and detection accuracy.

[0207] S3, 3D Leakage Point Location and Modeling;

[0208] A three-dimensional model of the pipe gallery is constructed using Neural Radiation Field (NeRF) technology, and the leakage coordinates output by the intelligent analysis module are mapped to three-dimensional space. By fusing real-time infrared temperature data with the thermodynamic parameters of the pipe material (such as the thermal conductivity of steel pipes), the temperature distribution of the model is corrected, achieving millimeter-level precise location of the leakage point.

[0209] S4, Spatiotemporal consistency check;

[0210] Verification combining optical flow tracing and hydrodynamic analysis:

[0211] Optical flow tracking: Based on the fusion of visible light and infrared data, the optical flow field is used to calculate the movement trajectory of the leakage area and verify the rationality of the diffusion rate in the time dimension;

[0212] Fluid dynamics verification: Dynamically adjust the parameters of the diffusion equation based on pipeline material and pressure data, predict the leakage diffusion path, and match the spatial overlap with the detection area to screen out false alarms that are not based on physical laws.

[0213] S5. Trigger alarms and predict leakage diffusion paths based on Bayesian decision-making;

[0214] Based on a Bayesian decision model that integrates historical data and real-time detection results, the posterior probability of leakage events is calculated, and alarm thresholds are dynamically optimized to reduce false alarms. Simultaneously, a fluid dynamics model is used to predict diffusion paths in future timeframes, generating heat maps to mark high-risk areas and guide emergency response.

[0215] This method, through the collaborative innovation of multimodal fusion detection, three-dimensional physical modeling, interdisciplinary verification, and adaptive decision-making, constructs a full-chain leakage monitoring system from data perception to intelligent decision-making. It solves the shortcomings of traditional methods in terms of accuracy, real-time performance, and environmental adaptability, and provides reliable technical support for the safe operation and maintenance of utility tunnels.

[0216] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A leakage monitoring system for key connection points of pipelines in utility tunnels based on AI vision, characterized in that, include: The data acquisition module is used to simultaneously acquire visible light images, infrared thermal imaging data, and pressure sensing data to form a multimodal dataset. The intelligent analysis module performs feature fusion analysis on the multimodal dataset based on a dynamic heterogeneous neural network, and outputs detection results including the coordinates of the leakage area and the leakage probability value. The 3D modeling module receives the coordinates of the leakage area from the detection results, performs spatial mapping using a 3D model of the pipe gallery constructed using the neural radiation field, and corrects the temperature distribution field by combining real-time infrared thermal imaging data, outputting a set of leakage points with three-dimensional coordinates. The spatiotemporal verification module receives the set of leakage points and executes the following: The diffusion rate in the time dimension is calculated by tracking the movement trajectory of the leakage area using the multimodal optical flow method. By dynamically adjusting the fluid dynamics diffusion equations based on the pipe material parameters in the neural radiation field model, the diffusion mode in the spatial dimension is verified. When the diffusion rate in the time dimension and the diffusion pattern in the spatial dimension meet the preset physical law matching degree, a leakage event that has passed the verification is generated. The alarm prediction module receives the verified leakage events and their corresponding leakage probability values, and executes the following through a Bayesian decision model: When the leakage probability value exceeds the dynamic optimization threshold and the verification matching degree reaches the confidence interval, a graded alarm is triggered. Based on the hydrodynamic diffusion equation, leakage paths are predicted, and early warning information including time-space diffusion heatmaps is generated.

2. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 1, characterized in that, The data acquisition module includes: Visible light camera, infrared thermal imager, pressure sensor, and laser spectrometer, among which: The laser spectrometer is configured to detect the concentration of gas components around the pipeline and generate data on abnormal gas distribution. The visible light camera and the infrared thermal imager achieve pixel-level spatial alignment through a coaxial gimbal, forming optical-thermal joint sensing data; The pressure sensor acquires time-series data of pressure fluctuations inside the pipeline; The multimodal data consists of spatiotemporally aligned optical images, thermodynamic images, pressure fluctuation sequences, and gas concentration matrices.

3. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 1, characterized in that, The intelligent analysis module includes: The dynamic heterogeneous neural network unit consists of a first branch network and a second branch network: The first branch network is built on a lightweight convolutional network and is used to process visible light images and infrared thermal imaging data in real time. The second branch network is trained based on data generated by meta-learning framework and physical simulation, and is used for leakage detection in scenarios with a small training sample size; The dynamic weight allocation unit dynamically adjusts the output weights of the two branch networks based on the physical simulation confidence of the second branch network, the infrared anomaly intensity value calculated by the deviation between the thermal imaging data and the reference temperature field, and the fluctuation amplitude of the pressure sensing data.

4. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 3, characterized in that, The data generated by the physical simulation includes: Simulation data of diffusion paths of different pipe materials in leakage scenarios; Data on the mapping relationship between the thermodynamic parameters of pipe materials and their infrared radiation characteristics.

5. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 1, characterized in that, The 3D modeling module includes: The neural radiation field modeling unit is used to construct a 3D neural radiation field model of the pipe gallery based on multi-view visible light and infrared data, wherein: The neural radiation field modeling unit combines the thermal radiation characteristics of the pipe material to correct the temperature distribution of the 3D model. The temperature distribution is generated by dynamically adjusting the coefficient and fusing real-time infrared data with material thermodynamic parameters. The leak point location unit is used to map the leak detection results output by the intelligent analysis module to the 3D model, and to track the movement trajectory of the leak area using optical flow method to calculate the matching degree between the leak point and the 3D model.

6. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 5, characterized in that, The calculation process of the dynamic adjustment coefficient includes: Construct a nonlinear function f based on a multilayer perceptron, whose input layer receives the real-time pipeline pressure P. pipe With ambient temperature T env The normalized parameter; The function f is optimized and trained using a meta-learning framework. The training dataset includes: The thermal conductivity matrix of the pipe material generated by physical simulation; Temperature field distribution data under different pressure conditions; Temperature-pressure correlation data of leakage events from historical maintenance records; In real-time calculation, the current pressure sensor reading and temperature sensor reading are standardized and preprocessed, and then input into the trained function f; The output layer generates dynamic adjustment coefficients through the Sigmoid activation function, which are used to weight and fuse real-time infrared data with material thermodynamic parameters.

7. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 1, characterized in that, The spatiotemporal verification module includes: The time dimension verification unit is used to track the movement trajectory of the leakage area using optical flow method, calculate the leakage diffusion rate, and compare it with the preset allowable diffusion rate threshold of the pipe material. The spatial dimension verification unit verifies the diffusion pattern of the leakage area based on the fluid dynamics diffusion equation, wherein the parameters of the diffusion equation are dynamically adjusted according to the pipe material information and real-time pressure data in the neural radiation field model. The consistency determination unit determines and filters out false alarms when the verification results of the time dimension and the space dimension are logically conflicting.

8. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 7, characterized in that, The optical flow tracking method employs a multimodal optical flow fusion algorithm, including: Calculate the optical flow field based on grayscale gradient for visible light images; Calculate the optical flow field based on the temperature gradient from infrared thermal imaging data; By dynamically weighting and fusing two optical flow fields, the motion trajectory of the leakage area is generated.

9. The AI ​​vision-based leakage monitoring system for key connection points of pipelines in utility tunnels according to claim 1, characterized in that, The alarm prediction module includes: Bayesian decision-making unit, execution: Based on the historical leakage event database, the alarm prior probability is dynamically updated. When the real-time detected leakage probability value exceeds the dynamically adjusted threshold, a high-confidence alarm signal is generated. The sensitivity range of the dynamic adjustment threshold is adaptively adjusted based on the fluctuation range of the current ambient temperature and equipment operating pressure in the utility tunnel. Diffusion prediction unit, execute: By combining the three-dimensional pipe parameters provided by the neural radiation field model, a leakage diffusion path prediction map for a future preset time period is generated through fluid dynamics simulation. Output a visualized heatmap with labeled diffusion direction, radius of influence, and risk level; The dynamic threshold optimization unit executes: Construct a multi-dimensional state space that includes ambient temperature and humidity, pressure fluctuation frequency, and historical false alarm records; The dynamic adjustment threshold is iteratively optimized using a reinforcement learning algorithm to balance the conflicting metrics of real-time false alarm rate and false negative rate. The alarm execution unit, in response to the high-confidence alarm signal, performs a tiered response action: Level 1 Response: Activate the on-site audible and visual alarm devices in the utility tunnel and generate an emergency maintenance work order; Level 2 Response: Push early warning information to mobile terminals and mark video streams of key monitoring areas.

10. A method for monitoring leakage at critical connection points of pipelines in utility tunnels based on AI vision, applied to the AI ​​vision-based leakage monitoring system for critical connection points of pipelines in utility tunnels as described in any one of claims 1-9, characterized in that, Includes the following steps: Simultaneously acquire visible light images, infrared thermal imaging data, and pressure sensing data of the utility tunnel area to generate spatiotemporally aligned multimodal data packets; The multimodal data packets are input into a dynamic heterogeneous neural network for feature fusion, and detection results containing the coordinates and probability values ​​of the leakage area are generated in real time. The detection results are mapped to a 3D pipe gallery model constructed from a neural radiation field, and the temperature field distribution is corrected by combining real-time infrared data to output a three-dimensional set of leak point coordinates. The movement trajectory of the leakage area is tracked by multimodal optical flow method, and the spatiotemporal consistency of the leakage diffusion path is verified based on the hydrodynamic diffusion equation, generating verification results. When the verification result meets the preset confidence conditions, a graded alarm is triggered based on the leakage probability value, and the leakage diffusion trend is predicted based on the corrected 3D model.

Images