Methods, systems, and inspection robots for identifying defects in gas pipelines in utility tunnels

Abstract

This invention belongs to the field of gas pipeline fault prediction technology. It discloses a method, system, and inspection robot for identifying defects in gas pipelines in utility tunnels. The method includes: synchronously sensing multi-dimensional time-series data of the gas pipeline in the utility tunnel, including gas pipeline time-series data, combustible gas concentration time-series data, and environmental time-series data; performing structured processing and feature extraction on the multi-dimensional time-series data to output a multi-modal feature data set for defect identification and risk assessment; performing joint modeling based on the multi-modal feature data set to obtain a fused interactive embedding vector, and identifying coupled abnormal states with causal correlation or synergistic performance between structural defects and trace gas leaks based on the fused interactive embedding vector; and identifying early defect types in the gas pipeline in the utility tunnel based on the fused interactive embedding vector and the coupled abnormal states. This application significantly improves the accuracy and response speed of early defect identification in gas pipelines in utility tunnels.

Description

Technical Field

[0001] This invention relates to the field of gas pipeline fault prediction technology, and more specifically, to a method, system, and inspection robot for identifying defects in gas pipelines in utility tunnels. Background Technology

[0002] As a crucial infrastructure for energy operations, the operational safety of gas pipelines in utility tunnels has gradually become a focus of public safety attention. Especially in long-term service environments, gas pipelines are prone to early structural defects such as wall thinning, surface pitting, and structural cracks due to various complex factors including damp heat corrosion, load fatigue, airflow disturbance, and operational fluctuations. These defects can lead to minor gas leaks, the accumulation of localized combustible gases, and potentially cause significant safety hazards.

[0003] However, existing methods for pipeline defect identification mostly rely on single-modal data sources, such as structural strain, gas concentration, or infrared thermal imaging, lacking the ability to model the coupling relationship between structural, gas, and environmental signals. This makes it difficult to effectively address scenarios with complex environmental disturbances and subtle defect manifestations. Furthermore, most existing methods treat structural anomalies and gas leaks as two independent discrimination tasks, failing to identify key coupling mechanisms such as structural anomalies inducing trace gas leaks or the leak process further promoting structural deterioration, leading to inaccurate early identification. This limits the timely response and differentiated handling of complex hidden dangers in gas pipelines, making it difficult to meet the practical requirements of high safety and high reliability operation of urban gas systems.

[0004] Therefore, there is an urgent need for a technological means with multi-dimensional perception and intelligent recognition capabilities to achieve early detection and risk prediction of gas pipeline defects. In view of this, this invention proposes a method, system, and inspection robot for identifying gas pipeline defects in utility tunnels to solve the aforementioned problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a method for identifying defects in gas pipelines in utility tunnels, comprising:

[0006] Synchronous sensing of multi-dimensional time-series data of gas pipelines in utility tunnels, including gas pipeline time-series data, combustible gas concentration time-series data, and environmental time-series data;

[0007] Structured processing and feature extraction are performed on multidimensional time-series data to output a multimodal feature data set that can be used for defect identification and risk assessment;

[0008] Joint modeling is performed based on a multimodal feature dataset to obtain a fused interactive embedding vector. Based on the fused interactive embedding vector, coupled anomalous states with causal correlation or synergistic performance between structural defects and trace gas leaks are identified. The coupled anomalous states include structural anomalies inducing gas micro-leaks or local leaks promoting structural deterioration.

[0009] Based on the fusion of interactive embedded vectors and coupled abnormal states, the early defect types of gas pipelines in utility tunnels are identified.

[0010] Furthermore, the method for identifying early defect types in the gas pipeline of the utility tunnel includes:

[0011] The coupling anomaly state is matched with a pre-built coupling anomaly state value matching table to obtain the coupling anomaly state value corresponding to the coupling anomaly state.

[0012] By inputting the coupled abnormal state values ​​and the fused interactive embedding vector into the defect type diagnosis model, the early defect types of the gas pipeline in the utility tunnel are obtained.

[0013] Furthermore, the method for obtaining the fused interaction embedding vector includes:

[0014] Full alignment is performed on the data in the multimodal feature dataset along the time-space dimension to obtain a multimodal feature vector set;

[0015] Initialize a cross-attention multimodal fusion network, which includes two layers of cross-attention encoders and one layer of multilayer perceptron;

[0016] In the first-layer cross-attention encoder, the structural feature vector is used as the query vector and the gas feature vector is used as the key-value pair. The structure-gas fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism and then converted into a structure-gas fusion vector.

[0017] In the second-layer cross-attention encoder, the structure-gas fusion vector is used as the query vector and the environmental feature vector is used as the key-value pair. The structure-gas-environment fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism.

[0018] Each row of the structure-gas-environment fusion attention weight matrix is ​​converted into a fusion interaction vector; the fusion interaction vector is then sequentially input into a multilayer perceptron and subjected to a nonlinear transformation using the ReLU activation function, finally outputting a deeply fused fusion interaction embedding vector; the fusion interaction embedding vector contains nonlinear correlation information of the three modalities of structural features, gas features and environmental features existing at the same spatial or temporal point.

[0019] Furthermore, the method for obtaining the coupling abnormal state includes:

[0020] The fusion interaction embedding vector is input into the coupling anomaly identification model to obtain the coupling anomaly diagnosis result.

[0021] Furthermore, the method for obtaining the multimodal feature data set includes:

[0022] Structural features are extracted from multidimensional time-series data of gas pipelines to obtain pipeline structural feature data;

[0023] Leakage behavior features are extracted from the time-series data of combustible gas concentration to obtain combustible gas feature data.

[0024] Environmental feature data is obtained by extracting environmental data change trend features from environmental time series data;

[0025] Pipeline structural feature data, combustible gas feature data, and environmental feature data are used to construct a multimodal feature data set.

[0026] Furthermore, the method for obtaining the combustible gas characteristic data includes:

[0027] For the gas concentration time series within each detection period, after preprocessing with time-series denoising, wavelet downsampling, and outlier correction, the slope of concentration change, the magnitude of local concentration surges, and the concentration drift trend index are calculated, and combustible gas characteristic data are constructed.

[0028] Furthermore, the method for obtaining the pipeline structure feature data includes:

[0029] For pipeline wall thickness variation data in the time series data of gas pipelines, based on the wall thickness value sequence output by the phased array ultrasonic sensor, combined with the robot's real-time pose data and inspection path coordinates, a wall thickness gradient feature matrix reflecting the spatial distribution trend of abnormal wall thickness is constructed through interpolation reconstruction, coordinate registration, and sequence segmentation. The real-time pose data includes axial displacement, attitude angle, and rotation information. Based on the wall thickness gradient feature matrix, the maximum wall thickness reduction rate, corrosion area, overall wall thickness fluctuation degree, and fluctuation anomaly clustering identifier are extracted. For pipeline surface temperature data, image normalization, pseudo-color mapping, and temperature gradient calculation are performed on the original infrared image sequence to form the hot spot area and maximum temperature difference value.

[0030] The pipeline structural feature data are constructed by taking the maximum wall thickness reduction rate, corrosion area, overall wall thickness fluctuation degree, fluctuation anomaly clustering indicator, hot spot area index, maximum temperature difference value, heat distribution dispersion and structural image feature set.

[0031] Furthermore, the method for obtaining the environmental feature data includes:

[0032] For each of the five types of environmental physical parameters in the environmental time series data, corresponding time variation curves are constructed to form a time series function set with time as the horizontal axis and the physical quantity measurement value as the vertical axis. For the time series function of each environmental physical parameter, the corresponding first derivative is calculated by the sliding window difference method or the numerical differentiation method to obtain the set of first derivative change rates reflecting the rate of change of the environmental physical parameter.

[0033] The set of first-order derivative rates of change corresponding to temperature, humidity, wind speed, airflow direction, and air pressure is normalized and aligned according to the sampling timestamp to construct a multi-dimensional environmental change rate vector sequence in a unified format, which serves as environmental feature data.

[0034] A gas pipeline defect identification system for utility tunnels, used in the gas pipeline defect identification method for utility tunnels, includes:

[0035] The data acquisition module is used to synchronously sense multi-dimensional time-series data of gas pipelines in the utility tunnel. The multi-dimensional time-series data includes gas pipeline time-series data, combustible gas concentration time-series data, and environmental time-series data.

[0036] The first processing module is used to perform structured processing and feature extraction on multidimensional time-series data, and output a set of multimodal feature data that can be used for defect identification and risk assessment.

[0037] The coupling anomaly determination module performs joint modeling based on a multimodal feature data set to obtain a fused interactive embedding vector, and identifies coupled anomaly states where there is a causal correlation or synergistic performance between structural defects and trace gas leaks based on the fused interactive embedding vector; the coupled anomaly states include structural anomalies inducing gas micro-leaks or local leaks promoting structural deterioration.

[0038] The defect type identification module identifies early defect types in gas pipelines in utility tunnels based on fused interactive embedded vectors and coupled abnormal states.

[0039] A pipeline defect identification and inspection robot for gas pipelines in a utility tunnel includes a memory and a processor, as well as a computer program stored in the memory and running in the processor. When the processor executes the computer program, it implements the pipeline defect identification method for gas pipelines in the utility tunnel.

[0040] Compared with existing technologies, the technical effects and advantages of the gas pipeline defect identification method, system, and inspection robot proposed in this invention are as follows:

[0041] This application achieves intelligent identification and evolution risk assessment of early complex defects in gas pipelines in utility tunnels by constructing a three-dimensional cross-sensing and coupled reasoning mechanism involving pipeline structure, gas response, and environmental disturbance. Compared to existing detection methods that rely solely on single physical signals or lack heterogeneous feature fusion capabilities, this application first uses multimodal sensing components to collect real-time structural monitoring data such as pipeline wall thickness changes, surface thermal imaging, and image morphology, and jointly acquires combustible gas concentration changes and environmental disturbance factors to form a high-resolution, multi-source heterogeneous feature input set. Subsequently, a cross-attention multimodal fusion network is used to construct attention weight matrices between structure and gas, and between structure and gas and environment, extracting deep interactive features of structure-induced response and environmental modulation influence, and outputting a fused interactive embedding vector. This effectively identifies coupled abnormal states with causal correlation or synergistic performance under complex disturbances and unclear early signs.

[0042] Furthermore, this application, based on the fusion of interactive embedded vectors and coupled abnormal state numerical values, inputs them into a defect type diagnostic model, realizing fine-grained defect type identification under two typical coupling mechanisms: structural anomalies inducing gas micro-leakage and local leakage promoting structural deterioration. This includes early key hidden danger characteristics such as pitting corrosion on the inner wall, micro-cracks due to lack of fusion in welds, accelerated corrosion sections on the outer wall, and stress concentration thermal crack initiation zones. This effectively improves the timeliness and accuracy of gas pipeline defect early warning, providing strong technical support for the inherently safe operation of gas pipelines in utility tunnels. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the gas pipeline defect identification system in the utility tunnel according to Embodiment 1 of the present invention;

[0044] Figure 2 This is a flowchart of the defect identification method for gas pipelines in a utility tunnel according to Embodiment 2 of the present invention;

[0045] Figure 3 Flowchart of the method for obtaining the embedded vector of the fusion interaction;

[0046] Figure 4 This is a flowchart illustrating the method for obtaining multimodal feature data sets. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0048] Example 1:

[0049] Please see Figure 1 As shown in the figure, this embodiment discloses a gas pipeline defect identification system for utility tunnels, including a data acquisition module, a first processing module, a coupled anomaly determination module, and a defect type discrimination module. Each module is connected by wired and / or wireless means to realize data transmission.

[0050] In the gas pipeline defect identification method, system, and inspection robot provided in this application, the inspection robot includes a main body. The main body carries various sensing and processing components within the system, enabling high-frequency, high-coverage, and high-stability automated inspections of gas pipelines within enclosed, narrow, or structurally complex underground utility tunnels. The main body constitutes the mobile execution platform of this system. Its main tasks include: autonomous navigation along a preset route or dynamic path; stably carrying multimodal sensing devices and collecting raw sensing data; real-time processing of body posture changes and adaptive compensation for complex terrain to ensure the integrity and continuity of data collection; and providing high-quality, spatiotemporally accurate data input for the subsequent defect type discrimination module.

[0051] To achieve the above functions, the main body of the inspection robot includes the following components:

[0052] The inspection robot is equipped with a drive wheel assembly and a differential control system to provide mobility along the floor or track of the utility tunnel. The drive wheel assembly preferably uses rubber-coated high-grip tires, combined with adjustable motor speed output, to achieve forward, backward, turning, and U-turn movements, adapting to the narrow bends, ramps, lateral slopes, and micro-obstacle paths commonly found inside the utility tunnel.

[0053] To ensure the navigation safety of the inspection robot in areas with dense obstacles or limited visibility, the robot is equipped with an obstacle avoidance device, which includes, but is not limited to, lidar, infrared distance sensors, and ultrasonic probes. These sensors are arranged in a ring around the front and sides of the robot. By scanning the geometry of the path in front in real time and constructing point cloud data, the robot uses obstacle avoidance algorithms to adjust its local path, avoiding collisions with gas pipelines, supports, cable trays, and other facilities.

[0054] To further improve the posture stability and path tracking accuracy of the inspection robot during movement, an attitude perception and navigation unit is embedded in the main module of the robot. This unit integrates an inertial measurement unit (IMU), an electronic compass, and a wheel speedometer. The IMU includes a three-axis gyroscope and a three-axis accelerometer. The IMU acquires the robot's pitch, roll, and yaw angles in real time and performs attitude fusion estimation using a Kalman filter algorithm. The electronic compass determines the robot's heading angle relative to the tunnel axis, and the wheel speedometer, combined with an encoder, outputs the current linear velocity and steering speed. All sensor data is input to the motion controller, which drives the motor output through closed-loop control logic to achieve precise tracking of the specified path.

[0055] Considering that utility tunnel environments often involve water accumulation, slippery surfaces, and uneven slopes, the inspection robot is further equipped with a slope compensation mechanism and a water wading detection component. The slope compensation mechanism adjusts the drive force distribution in real time through dual-motor torque adjustment to prevent slippage uphill or loss of control downhill. The water wading detection component, based on distributed electrode pairs or water-sensitive capacitors, senses the contact status between the robot's chassis and the accumulated water in real time. Once the liquid level exceeds a set threshold, it can trigger deceleration, steering avoidance, or alarm shutdown.

[0056] In addition, the inspection robot has a built-in battery management system for real-time monitoring, charging and discharging scheduling, and over-temperature protection of the robot's overall power supply system. The battery management system includes high-capacity lithium battery cells, a power distribution board, and temperature sensors. Combined with software management algorithms, it supports continuous inspection and features remaining power estimation and a low-battery return-to-base mechanism. The battery management system also provides independent power supply stability for various high-energy-consuming sensing devices.

[0057] The main module of the inspection robot in this application also features multi-sensor mounting capability and structural compatibility design. The top plate of the inspection robot is equipped with a multi-degree-of-freedom mounting bracket, supporting modular installation of the structural detection module, image acquisition module, gas detection module, and computing / communication module. The wiring channels, interface standards, and heat dissipation strategies between each module are systematically preset through pre-embedded slots and electromagnetic shielding layers, improving the overall system stability and anti-interference capability. The structural detection module includes an ultrasonic phased array, the image acquisition module includes a visible light and infrared fusion camera, and the gas detection module includes a laser-based CH4 analyzer.

[0058] The data acquisition module is used to synchronously sense multi-dimensional time-series data of the gas pipeline in the utility tunnel. This multi-dimensional time-series data specifically includes gas pipeline time-series data, combustible gas concentration time-series data, and environmental time-series data. These three data sets together form the perceptual basis for identifying gas pipeline defects and leaks. Each data dimension has a clear physical source and sensing target, providing crucial input for subsequent structural defect modeling, gas anomaly detection, and dynamic risk assessment.

[0059] The time-series data of the gas pipeline refers to monitoring data obtained by sampling the structural morphology and physical state of the gas pipeline body at fixed points or continuously over time. This type of monitoring data includes pipeline wall thickness variation data and pipeline surface temperature data, etc. The above data is acquired through multimodal sensing components configured on the inspection robot.

[0060] In this embodiment, preferably, the pipe wall thickness variation data refers to two-dimensional cross-sectional data acquired by a phased array ultrasonic sensor to quantify the spatial distribution of gas pipeline wall thickness. The phased array ultrasonic sensor is mounted on the bottom or side of the inspection robot, with the probe acoustically coupled to the outer wall of the pipe. By controlling the excitation delay of multiple array elements, the sound beam is focused and scanned at different depths and angles within the pipe cross-section. When the ultrasonic wave propagates within the pipe, it will produce significant reflection at the interface between the inner wall and the internal gas. The system receives the reflected signal, measures the echo time difference, and combines this with the sound velocity information in the material to calculate the propagation path length between the outer and inner walls, i.e., the pipe wall thickness variation data.

[0061] The pipe surface temperature data is acquired by an infrared thermal imager mounted on the top or side support of the inspection robot. The infrared thermal imager is used to collect thermal radiation images of the pipe's outer surface in the 8-14 μm wavelength band. This temperature data has not yet been used to create a thermal distribution map; subsequent image stitching and anomaly area extraction will be performed using coordinate information.

[0062] In summary, the gas pipeline time-series data enables multi-angle perception and fusion representation of the pipeline's structural state at different levels and dimensions, which can significantly improve the accuracy of structural anomaly identification and the scope of business coverage, and provide high-resolution structural prior information for subsequent defect-leakage joint analysis and prediction models.

[0063] The combustible gas concentration time-series data refers to the concentration records over time obtained by sampling the concentration of combustible components in the pipe gallery space at fixed points or continuously, including typical flammable and explosive gases such as methane, hydrogen sulfide, and hydrogen. The combustible gas concentration time-series data can be acquired by gas detection sensors installed at the front or side of the inspection robot. These sensors preferably employ electrochemical gas-sensitive elements or high-precision gas analyzers based on the principle of tunable laser absorption spectroscopy (TDLAS). During data acquisition, ambient gas is driven into the sampling chamber by a built-in micro-pump, and its concentration is detected. Each detection result is linked to the sampling time and location, forming spatiotemporally labeled concentration time-series data. The main function of the combustible gas concentration time-series data is to assist in identifying early concentration fluctuation characteristics during trace gas leaks, especially before structural defects lead to obvious eruptions. It can provide early warning of hidden leaks based on parameters such as concentration mutation rate, gradient anomalies, or persistent drift.

[0064] The environmental time-series data refers to the environmental data that changes over time by sampling the background physical states of the gas pipeline and its surrounding air environment, such as temperature, humidity, wind speed, airflow direction, and air pressure, at fixed points or continuously. This environmental data is acquired in real time through environmental sensing units installed on the inspection robot, specifically including temperature and humidity sensors, wind speed and direction sensors, barometers, and acceleration and vibration sensors. The environmental time-series data is mainly used for background compensation and anomaly causation determination of structural and gas concentration data: for example, when a sudden increase in gas concentration is detected, if it is accompanied by a sudden change in airflow speed, it can be inferred that the anomaly may be affected by wind field disturbance; furthermore, in areas with significantly increased temperatures, thermal imaging data may be at risk of misjudgment, requiring error correction in conjunction with thermal conductivity background parameters. Therefore, while ensuring the reliability of various data types, the environmental time-series data also provides a constraint basis for the comprehensive correlation analysis of multimodal sensing results.

[0065] In summary, by simultaneously acquiring time-series data of gas pipelines, time-series data of combustible gas concentration, and environmental time-series data from multiple sources, this invention can achieve unified perception and collaborative analysis of structural status and leakage signs, significantly improving the coverage, accuracy, and response speed of early defect identification and dynamic leakage early warning of gas pipelines in utility tunnels, and constructing a perception framework adapted to complex underground working conditions.

[0066] The first processing module is used to perform structured processing and feature extraction on the multidimensional time-series data acquired by the data acquisition module, and output a multimodal feature data set that can be used for defect identification and risk assessment.

[0067] like Figure 4 As shown, the method for obtaining the multimodal feature data set includes:

[0068] Structural features are extracted from multidimensional time-series data of gas pipelines to obtain pipeline structural feature data;

[0069] Leakage behavior features are extracted from the time-series data of combustible gas concentration to obtain combustible gas feature data.

[0070] Environmental feature data is obtained by extracting environmental data change trend features from environmental time series data;

[0071] Pipeline structural feature data, combustible gas feature data, and environmental feature data are used to construct a multimodal feature data set.

[0072] The method for obtaining the pipeline structure feature data includes:

[0073] For pipeline wall thickness variation data, the system, based on the wall thickness value sequence output by the phased array ultrasonic sensor, combines the robot's real-time pose data and inspection path coordinates, and constructs a wall thickness gradient feature matrix reflecting the spatial distribution trend of abnormal wall thickness through interpolation reconstruction, coordinate registration, and sequence segmentation. The real-time pose data includes axial displacement, attitude angle, and rotation information. Based on the wall thickness gradient feature matrix, the system extracts the maximum wall thickness reduction rate, corrosion area, overall wall thickness fluctuation degree, and fluctuation anomaly clustering identifier. For pipeline surface temperature data, the system performs image normalization, pseudo-color mapping, and temperature gradient calculation on the original infrared image sequence to form the hot spot area and maximum temperature difference value.

[0074] The pipeline structural feature data are constructed by taking the maximum wall thickness reduction rate, corrosion area, overall wall thickness fluctuation degree, fluctuation anomaly clustering indicator, hot spot area index, maximum temperature difference value, heat distribution dispersion and structural image feature set.

[0075] It should be noted that, in the embodiments of this application, the wall thickness gradient feature matrix is ​​a two-dimensional structural feature representation used to characterize the variation trend of the gas pipeline wall thickness along the spatial distribution direction. Its core lies in constructing a gradient response matrix that reflects corrosion areas, thinning trends, and concentrated structural deterioration by performing spatial coordinate calibration and gradient calculation on multi-point wall thickness variation data acquired by phased array ultrasonic sensors. This feature matrix serves as an important input for structural anomaly identification and can be used to generate several key structural indicators, including the maximum wall thickness thinning rate, corrosion area, overall wall thickness fluctuation degree, and fluctuation anomaly clustering indicators.

[0076] Specifically, the method for constructing the wall thickness gradient feature matrix includes the following steps:

[0077] During the inspection robot's movement, the system acquires wall thickness measurements from the phased array ultrasonic sensors at a preset fixed sampling period and obtains robot pose data in real time, including axial displacement, attitude angle, and rotation angle. Based on this pose data, the wall thickness measurement value of each measuring point is mapped to the spatial coordinate system of the gas pipeline, forming a set of wall thickness data points with three-dimensional spatial labels. To ensure spatial continuity and directional consistency between measuring points in structural calculations, a polar coordinate expansion model with the pipeline axis as the X-axis, the circumferential direction as the θ-axis, and the radial direction as the R-axis is preferred to construct a unified spatial projection reference.

[0078] The wall thickness data points are interpolated at equal intervals and spatially meshed to construct a two-dimensional mesh matrix. Each matrix element corresponds to the wall thickness measurement value of the pipe at the axial position and circumferential angle, in millimeters (mm). Based on this, first-order difference calculations are performed on adjacent mesh points along the axial and circumferential directions to obtain the gradient of the wall thickness measurement value along the axial and circumferential directions, forming a two-dimensional gradient matrix, which is the wall thickness gradient feature matrix, used to quantify the magnitude and direction of wall thickness variation at different locations.

[0079] It should be noted that the wall thickness gradient feature matrix can not only reveal the wall thickness reduction trend of the pipeline in different spatial regions, but also serve as the calculation basis for corrosion distribution characteristics, crack propagation direction and structural integrity degradation trend, providing high-resolution input basis for subsequent corrosion identification, thinning area extraction and structural deterioration cluster analysis.

[0080] The method for extracting the maximum wall thickness reduction rate is as follows: by comparing the wall thickness measurement value at each measuring point with the wall thickness reference value (design wall thickness or historical healthy state wall thickness), the relative thinning percentage is calculated, and the largest relative thinning percentage is selected as the maximum wall thickness reduction rate.

[0081] The method for calculating the relative thinning percentage is as follows:

[0082] ;

[0083] in, This represents a relative thinning percentage. This serves as a reference value for wall thickness. This represents the wall thickness measurement. The relative thinning percentage is used to assess the severity of localized corrosion and is one of the key inputs for subsequent remaining life prediction and risk grading.

[0084] The method for obtaining the area of ​​the corroded region is as follows: Regions in the wall thickness gradient feature matrix whose wall thickness is lower than a set wall thickness threshold (for example, the wall thickness threshold can be set to 80% of the wall thickness reference value) are connected to form a corroded connected region. The sum of the physical areas corresponding to the mesh cells of all corroded connected regions is the area of ​​the corroded region. This indicator can be used to quantify the spatial influence range of defects.

[0085] The method for obtaining the overall wall thickness fluctuation degree is as follows: calculate the wall thickness fluctuation variance of the wall thickness measurements at all measuring points in the wall thickness gradient feature matrix, and calculate the mean value of all wall thickness fluctuation variances to obtain the overall wall thickness fluctuation degree. The overall wall thickness fluctuation degree is used to measure the wall thickness consistency level of the entire detection section.

[0086] The method for obtaining the fluctuation anomaly cluster identifier includes: recording measurement points where the wall thickness fluctuation variance is greater than the overall wall thickness fluctuation degree as fluctuation anomaly points; connecting adjacent fluctuation anomaly points to form fluctuation anomaly regions; and calculating the total physical area corresponding to the grid cells of the fluctuation anomaly regions, which is recorded as the fluctuation anomaly region area; if the fluctuation anomaly region area is greater than a preset fluctuation anomaly region area threshold, then marking the corresponding fluctuation anomaly region as a fluctuation anomaly cluster identifier; the fluctuation anomaly region area threshold can be set by those skilled in the art based on the total fluctuation anomaly region area. For example, the average value of the total fluctuation anomaly region area can be set as the fluctuation anomaly region area threshold.

[0087] Methods for performing image normalization, pseudo-color mapping, and temperature gradient calculation on the original infrared image sequence to form the hot spot area and maximum temperature difference include:

[0088] The system performs normalization processing on the original infrared image. Preferably, a linear normalization method is used to map the gray value range in each frame of the image to [0,1], eliminating the brightness difference between images, so that subsequent feature extraction focuses on relative temperature difference rather than absolute brightness change.

[0089] The normalized image is converted into a color image after pseudo-color mapping, enhancing the visual distinguishability of temperature gradients. The pseudo-color mapping method preferably uses the "Jet" or "Thermal" color scheme, with low-temperature regions corresponding to cool tones and high-temperature regions to warm tones, facilitating image segmentation and hotspot extraction. Subsequently, the system calculates the temperature gradient map of the image based on gradient operators, identifies regions of abrupt changes in thermal edges, and extracts closed high-temperature blocks using region connectivity analysis, constructing a set of candidate hotspot regions.

[0090] For each candidate hotspot region, the system converts the pixel area into an actual spatial area based on the image resolution and marks it as the hotspot area. Furthermore, it calculates the difference between the highest and lowest temperatures in each image frame, which is the maximum temperature difference value for the current frame.

[0091] The method for obtaining the combustible gas characteristic data includes:

[0092] For the gas concentration time series within each detection period, the system preprocesses it through time-series denoising, wavelet downsampling, and outlier correction, then calculates the concentration change slope, the magnitude of local concentration spikes, and the concentration drift trend index, constructing combustible gas characteristic data. The time-series denoising, wavelet downsampling, and outlier correction are conventional preprocessing techniques and will not be elaborated upon here.

[0093] The method for calculating the slope of the concentration change includes:

[0094] ;

[0095] in, The slope of the concentration change. To monitor the gas concentration value at the (i+1)th time point within the monitoring period, To monitor the gas concentration value at the i-th time point within the monitoring period, The sampling time interval between the (i+1)th time node and the ith time node.

[0096] The calculation method for the magnitude of the sudden increase in local concentration includes:

[0097] A local analysis window length is preset, and the time series data of combustible gas concentration is divided into sliding windows according to the local analysis window length to obtain Q time windows. The difference between the highest and lowest gas concentration values ​​in each time window is calculated to obtain the gas concentration change amplitude corresponding to each time window. The largest gas concentration change amplitude is selected as the local concentration surge amplitude.

[0098] The calculation method for the concentration drift trend index includes:

[0099] The variance of the gas concentration change amplitude is calculated for the gas concentration change amplitude over Q time windows, which yields the concentration drift trend index.

[0100] It should be noted that, in the embodiments of this application, in order to achieve early identification and dynamic warning of potential gas pipeline leakage, the system extracts features from the time series data of combustible gas concentration and constructs a set of feature indicators including the slope of concentration change, the magnitude of local concentration surge, and the concentration drift trend index. These indicators play key roles in short-term change response, sudden behavior capture, and medium- and long-term trend tracking during the leakage identification process, and together constitute a highly robust feature description system for trace leakage symptoms.

[0101] The concentration change slope measures the rate of change of gas concentration per unit time, and its physical meaning lies in reflecting the intensity of the time response to sudden changes in gas concentration. When a gas pipeline structure is damaged or its seal fails, the internal gas escapes into the pipe gallery space, causing the local concentration to rise rapidly in a very short time. At this time, the concentration change slope will show a significantly positive value, much higher than the slow changes caused by normal ventilation or background fluctuations. Therefore, this indicator is suitable for sensitively capturing the initial stage of a leak event and is an important triggering factor for triggering subsequent proactive safety response mechanisms.

[0102] The magnitude of a sudden increase in local concentration measures the maximum range of concentration fluctuations within a sliding time window. Its core function is to identify whether a significant concentration jump event occurs within a short period of time. This indicator can compensate for the insufficient response of the slope in instantaneous disturbance scenarios, and is particularly suitable for detecting sudden concentration surge signals caused by intermittent gas releases, pressure fluctuations, and local disturbance gas cloud drift, thereby improving the system's ability to identify discontinuous leakage patterns, such as short-jet and intermittent leaks.

[0103] The concentration drift trend index is primarily used to assess the stability and fluctuation trend of gas concentration over time. Its calculation method is based on the variance of concentration change amplitudes within multiple time windows, thus reflecting whether the gas concentration is continuously accumulating or changing periodically. When microcracks or fatigue leaks occur in pipelines, the escaping gas may not reach the high concentration required for an immediate alarm, but its concentration curve will exhibit a strong cumulative drift trend or periodic disturbance characteristics. In this case, the concentration drift trend index can quantify this latent leak characteristic, providing an important criterion for early identification of long-term, low-intensity leaks.

[0104] In summary, the three indicators of concentration change slope, local concentration surge magnitude, and concentration drift trend index, constructed in this application, form a feature characterization path for gas leakage behavior from three dimensions: "abrupt response → extreme value disturbance → cumulative trend." This not only significantly improves the coverage of different leakage types but also provides data support and judgment basis for subsequent defect-leakage correlation modeling, leakage level classification, and active inspection path adjustment, thereby enhancing the overall intelligent perception and dynamic decision-making capabilities of the system under complex operating conditions.

[0105] The method for obtaining the environmental feature data includes:

[0106] For five types of environmental physical parameters—temperature, humidity, wind speed, airflow direction, and air pressure—time-varying curves are constructed from the environmental time-series data, forming a set of time-series functions with time on the horizontal axis and the measured value of the physical quantity on the vertical axis. For each environmental parameter's time-series function, the corresponding first derivative is calculated using the sliding window difference method or numerical differentiation method, yielding a set of first derivative rate of change reflecting the rate of change of that physical quantity. Specifically, the system calculates the sets of first derivative rate of change for temperature, humidity, wind speed, airflow direction, and air pressure, respectively, to quantify the dynamic fluctuation characteristics of each physical quantity over time.

[0107] The five sets of first-order derivative change rates are normalized and aligned according to the sampling timestamps to construct a multidimensional environmental change rate vector sequence in a unified format, which serves as the environmental feature data of this application.

[0108] It should be noted that, in the embodiments of this application, in order to obtain the dynamic change rate of environmental physical quantities in the time dimension, the system performs first-order derivative calculations on each environmental time-series function, preferably using the sliding window difference method. The sliding window difference method is a commonly used numerical method that approximates the derivative by performing difference calculations on adjacent data points in a continuous time series. It is widely used in environmental monitoring, industrial process control, and time-series data analysis, and is a standard data processing method known and familiar to those skilled in the art. This method is also applicable to the derivative calculation of other environmental time-series data such as temperature, humidity, wind speed, airflow direction, and air pressure, ultimately forming a set of first-order derivative change rates corresponding to five types of environmental physical quantities, providing a standardized input format for subsequent construction of environmental feature data. Through the above derivative processing, real-time identification of rapid fluctuations, abnormal changes, or cumulative drift trends of environmental variables can be achieved, improving the system's adaptability to background interference factors.

[0109] In this application, the environmental feature data is not only used for background correction in the subsequent abnormal interference identification and multimodal fusion process, but also serves as an important reference for the system to determine whether the leakage signal is triggered by drastic environmental changes, thereby improving the overall identification model's discrimination accuracy and stability under complex working conditions.

[0110] The coupling anomaly detection module performs joint modeling based on a multimodal feature dataset to obtain a fused interactive embedding vector. Based on this vector, it identifies coupled anomaly states where structural defects and trace gas leaks exhibit causal correlation or synergistic behavior. These coupled anomaly states include structural anomalies inducing micro-gas leaks or localized leaks promoting structural degradation. Unlike traditional methods based on independent single-modal detection, this module constructs an intermodal feature correlation mechanism to intelligently identify complex conditions such as "structural degradation inducing leaks" or "leakage paths guided by structural defects," and outputs risk classification results with coupling attributes. In situations with complex environmental interference and unclear early signs of defects, it accurately reveals the potential causal or synergistic relationship between structural failure and micro-leaks, providing reliable technical support for developing refined maintenance plans and rapid risk mitigation measures.

[0111] like Figure 3 As shown, the method for obtaining the fusion interaction embedding vector includes:

[0112] Full alignment is performed on the data in the multimodal feature dataset in the time-space dimension to obtain a multimodal feature vector group, which includes structural feature vectors, gas feature vectors, and environmental feature vectors. Full alignment is used to ensure that each modal feature has a consistent label index at the same sampling time and the same detection coordinate.

[0113] Initialize a cross-attention multimodal fusion network, which includes two layers of cross-attention encoders and one layer of multilayer perceptron;

[0114] In the first-layer cross-attention encoder, the structural feature vector is used as the query vector and the gas feature vector is used as the key-value pair. The structure-gas fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism, and the structure-gas fusion attention weight matrix is ​​converted into a structure-gas fusion vector. The structure-gas fusion vector is used to characterize the gas response signal induced at the structural level, measure the response weight of the structural feature dimension to different gas feature dimensions, and characterize the potential gas leakage response characteristics induced at the structural level.

[0115] In the second-layer cross-attention encoder, the structure-gas fusion vector is used as the query vector and the environmental feature vector is used as the key-value pair. The structure-gas-environment fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism. The structure-gas-environment fusion attention weight matrix is ​​used to characterize the sensitivity of the structure-gas linkage signal to the environmental background state.

[0116] Each row of the structure-gas-environment fusion attention weight matrix is ​​converted into a fusion interaction vector. Each fusion interaction vector reflects the coupling and correlation pattern of the structure, gas, and environment in multidimensional space at the current moment. The fusion interaction vectors are sequentially input into a multilayer perceptron and subjected to a nonlinear transformation using the ReLU activation function, ultimately outputting a deeply fused fusion interaction embedding vector. The fusion interaction embedding vector contains nonlinear correlation information of the three modalities at the same spatial / temporal point, which is an important foundation for subsequent coupling determination and risk classification. The ReLU activation function is an activation function in deep learning.

[0117] Methods for obtaining elements in the structure-gas fusion attention weight matrix include:

[0118] ;

[0119] in, Let be the element corresponding to the i-th row and j-th column of the structure-gas fusion attention weight matrix, representing the attention weight allocation value of the structural features to the gas features between the i-th dimension of the structural feature vector and the j-th dimension of the gas feature vector. If The larger the value, the stronger the correlation between the i-th term of the structural feature vector and the j-th term of the gas feature vector, and the greater the weight of the gas information in the structure-gas fusion result. For vector dimensions, It refers to an exponential function with the natural constant e as its base. This represents the query vector obtained by linearly transforming the i-th dimension of the structural feature vector. This represents the bond vector obtained by linear transformation of the j-th dimension of the gas eigenvector. This represents the value vector of the j-th dimension of the gas feature vector after linear transformation. The numerator represents the matching strength between the i-th structural feature and the j-th gas feature; the denominator is the sum of the scores over all gas feature dimensions.

[0120] In the denominator This is the summation variable used to normalize all gas feature dimensions. Essentially, it iterates through all possible j values ​​to ensure the sum of the attention weights is 1, satisfying the probability normalization requirement. In other words, the denominator... It is a process of exponentially mapping and summing the attention scores of all components in the gas feature vector, which is used to normalize the individual attention scores in the current molecule.

[0121] In a preferred embodiment of this application, to achieve metric modeling of the potential correlation between structural features and gas features, the system introduces a cross-attention mechanism to model and weightedly fuse the information flow between structural and gas features. This mechanism takes structural feature vectors and gas feature vectors as inputs, and generates corresponding query vectors, key vectors, and value vectors through linear transformation mappings, respectively. The vector dimension is uniformly set to [value missing]. .

[0122] The numerator is the similarity index mapping result calculated by scaling the dot product between the i-th dimension of the current structural feature and the j-th dimension of the gas feature, reflecting the semantic association strength between the two in the embedding space. A larger similarity value indicates a stronger coupling relationship between the structural defect and gas fluctuations in that dimension. The denominator is a normalization term, mathematically expressed as the sum of the similarity indices over all dimensions of the gas feature vector. This normalization term originates from the Softmax function, designed to transform a set of arbitrary real values ​​(i.e., unconstrained similarity scores) into a set of non-negative real numbers with probabilistic semantics. Furthermore, to avoid gradient instability caused by numerical expansion of the dot product result in high-dimensional space, this application employs a standard scaling factor. Scaling the dot product result is a well-known design practice for attention mechanisms in the field of knowledge.

[0123] In summary, the overall structure of the formula not only has a clear physical interpretation—that is, a weighting function of the attention distribution of a certain dimension of the structural features on each dimension of the gas features—but also meets the mathematical requirements of numerical stability and trainability. It can be directly used to calculate the coupling perception strength between the structure-gas multimodal features, thereby providing a highly reliable information fusion basis for the subsequent defect type discrimination module.

[0124] Methods for obtaining elements in the structure-gas-environment fusion attention weight matrix include:

[0125] ;

[0126] in, The element in the x-th row and y-th column of the structure-gas-environment fusion attention weight matrix represents the attention weight of the x-th dimension of the structure-gas-environment fusion feature for the y-th dimension of the environmental feature. This represents the query vector corresponding to the x-th dimension of the structure-gas fusion feature; Let represent the key vector of the y-th dimension in the environmental features. This represents the value vector of the y-th dimension in the environmental features.

[0127] It should be noted that the second-layer cross-attention encoder uses the structure-gas fusion vector (the row elements in the structure-gas fusion attention weight matrix) as the query vector and the environmental feature vector as the key-value pair to construct the structure-gas-environment fusion attention weight matrix. Its specific calculation method is the same as that of the first-layer cross-attention encoder, still using the scaling dot product attention mechanism, and using softmax to ensure normalized semantics, thereby realizing the deep fusion and linkage of multimodal information.

[0128] In a preferred embodiment of this application, a second-layer cross-attention encoder is used to further introduce environmental background factors to dynamically regulate the structure-gas fusion characteristics, thereby improving the model's resistance to external disturbances and focusing on real micro-leakage collaborative anomaly events caused by structural defects.

[0129] The method for obtaining the coupling abnormal state includes:

[0130] The fused interactive embedding vector is input into the coupling anomaly identification model to obtain the coupling anomaly diagnosis result; the coupling anomaly diagnosis result includes structural anomalies inducing gas micro-leakage or local leakage that promotes structural deterioration;

[0131] The results of the coupling anomaly diagnosis are taken as the coupling anomaly state.

[0132] The training method for the coupled anomaly detection model includes:

[0133] A coupling anomaly identification dataset is pre-collected, which includes OH group coupling anomaly identification data and corresponding coupling anomaly diagnosis results, where OH is a positive integer greater than 0, and the coupling anomaly identification data includes fused interaction embedding vectors; the coupling anomaly identification dataset is divided into a training set and a validation set, where the training set is used to learn the parameters of the coupling anomaly identification model, and the validation set is used to evaluate the generalization ability of the coupling anomaly identification model to avoid overfitting;

[0134] During the training of the coupled anomaly detection model, a dynamic learning rate adjustment strategy and an early stopping mechanism are combined to minimize the cross-entropy loss function as the optimization objective. When the performance of the validation set meets the preset requirements, training is automatically stopped to ensure the convergence effect of the coupled anomaly detection model. The coupled anomaly detection model is implemented based on a neural network model architecture. The neural network model consists of an input layer, a hidden layer, and an output layer. The input layer receives coupled anomaly detection data and transforms it into a high-dimensional feature vector. The hidden layer uses activation functions to extract complex nonlinear patterns from the coupled anomaly detection data. The output layer calculates the probability distribution of coupled anomaly diagnosis results through a softmax activation function and outputs the coupled anomaly diagnosis result corresponding to the highest probability.

[0135] It should be noted that, in a preferred embodiment of this application, the system identifies two coupled abnormal states—structural anomalies inducing gas micro-leakage and localized leakage promoting structural degradation—based on fused interactive embedding vectors. This enables causal attribution and evolution path determination of potential pipeline risks. The fused interactive embedding vectors are generated by a cross-attention multimodal fusion network. These vectors simultaneously encode the dynamic interaction relationships between structural features, gas features, and environmental background features. By modeling the saliency dependencies between each feature dimension through a cross-modal attention mechanism, the system semantically reflects key linkage patterns at the level: what structural state changes might lead to what gas responses, or whether specific concentration disturbances are accompanied by signs of structural degradation.

[0136] Specifically, if the fused interactive embedding vector exhibits a structure-dominated change with a lagging response to sudden increases in gas concentration, it can be inferred that a structural anomaly induced a trace leak. Conversely, if the concentration drift trend is significant, the local hotspot response is enhanced, and the structural fluctuations exhibit asymmetric expansion, it can be inferred that a local leak promotes structural degradation. This modeling approach, based on spatiotemporal evolution and cross-attention coupling, overcomes the limitations of existing technologies that rely solely on a single threshold or rule engine to determine anomaly risks, and can significantly improve the accuracy of identifying early latent anomalies and coupled risk sources.

[0137] Compared to existing technologies that typically rely on independent structural defect identification models and gas leak detection, lacking the ability to fuse multi-source data and determine causal paths, the coupled anomaly identification mechanism based on fused interactive embedding vectors provided in this application has significant advantages. Specifically, this application can uniformly process multimodal feature inputs, improving the model's adaptability to multi-source information fusion in complex environments; it has the ability to model the causal relationship between structures and leaks, and can output more instructive risk attribution results, thereby improving the overall intelligence level and risk management capabilities of the inspection system.

[0138] The defect type identification module identifies early defect types in gas pipelines within utility tunnels based on fused interactive embedding vectors and coupled abnormal states. These early defect types are structural defects in the gas pipelines that exhibit early signs of defects and potential evolution risks.

[0139] The method for identifying early defect types in the gas pipeline corridor includes:

[0140] The coupling anomaly state is matched with a pre-built coupling anomaly state value matching table to obtain the coupling anomaly state value corresponding to the coupling anomaly state.

[0141] By inputting the coupled abnormal state values ​​and the fused interactive embedding vector into the defect type diagnosis model, the early defect types of the gas pipeline in the utility tunnel are obtained.

[0142] The numerical matching table for the coupling anomaly states is shown in Table 1:

[0143] Table 1. Numerical Matching Table for Coupled Abnormal States

[0144] Coupling Abnormal State Coupled abnormal state values Structural anomalies induce micro-gas leaks 1 Localized leaks promote structural deterioration 2

[0145] The training method for the defect type diagnostic model includes:

[0146] A defect type diagnosis dataset is pre-collected, which includes QX groups of defect type diagnosis data and the early defect types of gas pipelines in the pipe gallery corresponding to the QX groups of defect type diagnosis data, where QX is a positive integer greater than 0. The defect type diagnosis data includes coupled abnormal state values ​​and fused interactive embedding vectors. The defect type diagnosis dataset is divided into a training set and a validation set. The training set is used to learn the parameters of the defect type diagnosis model, and the validation set is used to evaluate the generalization ability of the defect type diagnosis model to avoid overfitting.

[0147] During the training of the defect type diagnosis model, a dynamic learning rate adjustment strategy and an early stopping mechanism are combined to minimize the cross-entropy loss function as the optimization objective. When the performance of the validation set meets the preset requirements, training is automatically stopped to ensure the convergence effect of the defect type diagnosis model. The defect type diagnosis model is implemented based on a neural network model architecture. The neural network model consists of an input layer, a hidden layer, and an output layer. The input layer receives the defect type diagnosis data and transforms it into a high-dimensional feature vector. The hidden layer uses an activation function to extract the complex nonlinear patterns of the defect type diagnosis data. The output layer calculates the probability distribution of early defect types of the gas pipeline in the pipeline corridor through the softmax activation function, and outputs the early defect type of the gas pipeline corridor corresponding to the highest probability as the prediction result.

[0148] In a preferred embodiment of this application, the method for identifying early-stage defect types in gas pipelines within utility tunnels effectively achieves intelligent identification of early-stage latent defect types in complex utility tunnel scenarios by jointly analyzing coupled abnormal state values ​​and fusing interactive embedding vectors. Based on the classification information of coupled abnormal states, this application introduces the synergistic interaction pattern between structural defects and gas leaks into the defect type diagnostic model, enabling the diagnostic process to consider both causal inference and multimodal feature expression, significantly improving the accuracy and specificity of defect identification.

[0149] Specifically, the system first matches the identified coupled anomaly states with a preset coupled anomaly state value matching table to obtain the coupled anomaly state value corresponding to the current anomaly state. Preferably, if the current state is a structural anomaly inducing a micro-leak of gas, it is marked as value 1; if the current state is a local leak promoting structural degradation, it is marked as value 2. This value label can be used as prior knowledge to participate in model discrimination, helping to guide the model to focus on key feature patterns under different types of anomalies.

[0150] Based on this, the system inputs the coupled anomalous state value and the previously generated fusion interactive embedding vector into the defect type diagnostic model. The fusion interactive embedding vector is a deep semantic representation formed by processing three types of features—structure, gas, and environment—through a two-level cross-attention mechanism, possessing high sensitivity and discriminative power for multimodal co-evolution processes. The coupled anomalous state value provides the model with a clear direction for reasoning, i.e., under what causal context defect classification is carried out. Therefore, this dual-input strategy achieves a fusion of data-driven and knowledge-guided modeling, exhibiting good generalization ability in complex scenarios.

[0151] In this application, the defect type diagnosis model is preferably able to identify pitting corrosion on the inner wall, microcracks due to lack of fusion in the weld, accelerated corrosion zone on the outer wall, and stress concentration thermal crack initiation zone; pitting corrosion on the inner wall and microcracks due to lack of fusion in the weld belong to gas micro-leakage induced by structural anomalies; accelerated corrosion zone on the outer wall and stress concentration thermal crack initiation zone belong to local leakage that promotes structural deterioration.

[0152] To further clarify, pitting corrosion refers to the formation of localized micro-pits or small erosion areas on the inner wall of a pipeline. It is often caused by stress corrosion or the action of chemical media, creating leakage channels that are difficult to detect with traditional visual inspection. This defect is usually accompanied by localized thinning and fluctuations in trace gas concentrations. Weld incomplete fusion microcracks refer to localized incomplete fusion or cracks at the welded parts of the pipeline. Initially, this manifests as a decrease in structural integrity and a tendency to cause leaks.

[0153] To further clarify, the accelerated corrosion zone refers to the long-term action of corrosive components in the leaked gas on the outer wall of the pipeline, leading to corrosion on the outer side of the pipe, increased wall thickness fluctuations, and improved temperature dispersion. The stress concentration thermal crack initiation zone refers to the initial zone of high-temperature induced structural stress cracks formed due to localized heat accumulation caused by leakage, superimposed on stress concentration in structural components, which is prone to developing into fracture hazards.

[0154] In summary, by fusing coupled abnormal state numerical values ​​and fused interactive embedding vectors for defect type identification, not only can surface defects be identified, but more importantly, early latent defects under the structure-leakage dual-factor interaction mechanism can be discovered. This achieves an intelligent discrimination method based on the coupling mechanism, breaking through the existing single-modal diagnostic accuracy limitations and demonstrating significant advantages in ensuring the safe operation of gas pipelines in utility tunnels.

[0155] Example 2:

[0156] Please see Figure 2 As shown, this embodiment provides a method for identifying defects in gas pipelines in utility tunnels, including:

[0157] Synchronous sensing of multi-dimensional time-series data of gas pipelines in utility tunnels. The multi-dimensional time-series data specifically includes gas pipeline time-series data, combustible gas concentration time-series data, and environmental time-series data.

[0158] Structured processing and feature extraction are performed on multidimensional time-series data to output a multimodal feature data set that can be used for defect identification and risk assessment;

[0159] Joint modeling is performed based on a multimodal feature dataset to obtain a fused interactive embedding vector. Based on the fused interactive embedding vector, coupled anomalous states with causal correlation or synergistic performance between structural defects and trace gas leaks are identified. The coupled anomalous states include structural anomalies inducing gas micro-leaks or local leaks promoting structural deterioration.

[0160] Based on the fusion of interactive embedded vectors and coupled abnormal states, the early defect types of gas pipelines in utility tunnels are identified.

[0161] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0162] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying defects in gas pipelines in utility tunnels, characterized in that, include: Synchronous sensing of multi-dimensional time-series data of gas pipelines in utility tunnels, including gas pipeline time-series data, combustible gas concentration time-series data, and environmental time-series data; Structured processing and feature extraction are performed on multidimensional time-series data to output a multimodal feature data set that can be used for defect identification and risk assessment; Joint modeling is performed based on a multimodal feature dataset to obtain a fused interactive embedding vector. Based on the fused interactive embedding vector, coupled anomalous states with causal correlation or synergistic performance between structural defects and trace gas leaks are identified. The coupled anomalous states include structural anomalies inducing gas micro-leaks or local leaks promoting structural deterioration. The method for obtaining the fusion interaction embedding vector includes: Full alignment is performed on the data in the multimodal feature dataset along the time-space dimension to obtain a multimodal feature vector set; Initialize a cross-attention multimodal fusion network, which includes two layers of cross-attention encoders and one layer of multilayer perceptron; In the first-layer cross-attention encoder, the structural feature vector is used as the query vector and the gas feature vector is used as the key-value pair. The structure-gas fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism and then converted into a structure-gas fusion vector. In the second-layer cross-attention encoder, the structure-gas fusion vector is used as the query vector and the environmental feature vector is used as the key-value pair. The structure-gas-environment fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism. Each row of the structure-gas-environment fusion attention weight matrix is ​​converted into a fusion interaction vector; the fusion interaction vector is then sequentially input into a multilayer perceptron and subjected to a nonlinear transformation using the ReLU activation function, finally outputting a deeply fused fusion interaction embedding vector; the fusion interaction embedding vector contains nonlinear correlation information of the three modalities of structural features, gas features and environmental features existing at the same spatial or temporal point; The method for obtaining the coupling abnormal state includes: The fused interactive embedding vector is input into the coupling anomaly identification model to obtain the coupling anomaly diagnosis result; the coupling anomaly diagnosis result is used as the coupling anomaly state. Based on the fusion of interactive embedded vectors and coupled abnormal states, the early defect types of gas pipelines in utility tunnels are identified.

2. The method for identifying defects in gas pipelines in utility tunnels according to claim 1, characterized in that, The method for identifying early defect types in the gas pipeline corridor includes: The coupling anomaly state is matched with a pre-built coupling anomaly state value matching table to obtain the coupling anomaly state value corresponding to the coupling anomaly state. By inputting the coupled abnormal state values ​​and the fused interactive embedding vector into the defect type diagnosis model, the early defect types of the gas pipeline in the utility tunnel are obtained.

3. The method for identifying defects in gas pipelines in utility tunnels according to claim 1, characterized in that, The method for obtaining the multimodal feature data set includes: Structural features are extracted from multidimensional time-series data of gas pipelines to obtain pipeline structural feature data; Leakage behavior features are extracted from the time-series data of combustible gas concentration to obtain combustible gas feature data. Environmental feature data is obtained by extracting environmental data change trend features from environmental time series data; Pipeline structural feature data, combustible gas feature data, and environmental feature data are used to construct a multimodal feature data set.

4. The method for identifying defects in gas pipelines in utility tunnels according to claim 3, characterized in that, The method for obtaining the combustible gas characteristic data includes: For the gas concentration time series within each detection period, after preprocessing with time-series denoising, wavelet downsampling, and outlier correction, the slope of concentration change, the magnitude of local concentration surges, and the concentration drift trend index are calculated, and combustible gas characteristic data are constructed.

5. The method for identifying defects in gas pipelines in utility tunnels according to claim 3, characterized in that, The method for obtaining the pipeline structure feature data includes: For pipeline wall thickness variation data in the time series data of gas pipelines, based on the wall thickness value sequence output by the phased array ultrasonic sensor, combined with the robot's real-time pose data and inspection path coordinates, a wall thickness gradient feature matrix reflecting the spatial distribution trend of abnormal wall thickness is constructed through interpolation reconstruction, coordinate registration, and sequence segmentation. The real-time pose data includes axial displacement, attitude angle, and rotation information. Based on the wall thickness gradient feature matrix, the maximum wall thickness reduction rate, corrosion area, overall wall thickness fluctuation degree, and fluctuation anomaly clustering identifier are extracted. For pipeline surface temperature data, image normalization, pseudo-color mapping, and temperature gradient calculation are performed on the original infrared image sequence to form the hot spot area and maximum temperature difference value. The pipeline structural feature data are constructed by taking the maximum wall thickness reduction rate, corrosion area, overall wall thickness fluctuation degree, fluctuation anomaly clustering indicator, hot spot area index, maximum temperature difference value, heat distribution dispersion and structural image feature set.

6. The method for identifying defects in gas pipelines in a utility tunnel according to claim 3, characterized in that, The method for obtaining the environmental feature data includes: For each of the five types of environmental physical parameters in the environmental time series data, corresponding time variation curves are constructed to form a time series function set with time as the horizontal axis and the physical quantity measurement value as the vertical axis. For the time series function of each environmental physical parameter, the corresponding first derivative is calculated by the sliding window difference method or the numerical differentiation method to obtain the set of first derivative change rates reflecting the rate of change of the environmental physical parameter. The set of first-order derivative rates of change corresponding to temperature, humidity, wind speed, airflow direction, and air pressure is normalized and aligned according to the sampling timestamp to construct a multi-dimensional environmental change rate vector sequence in a unified format, which serves as environmental feature data.

7. A gas pipeline defect identification system for utility tunnels, used to implement the gas pipeline defect identification method for utility tunnels as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to synchronously sense multi-dimensional time-series data of gas pipelines in the utility tunnel. The multi-dimensional time-series data includes gas pipeline time-series data, combustible gas concentration time-series data, and environmental time-series data. The first processing module is used to perform structured processing and feature extraction on multidimensional time-series data, and output a set of multimodal feature data that can be used for defect identification and risk assessment. The coupling anomaly determination module performs joint modeling based on a multimodal feature data set to obtain a fused interactive embedding vector, and identifies coupled anomaly states where there is a causal correlation or synergistic performance between structural defects and trace gas leaks based on the fused interactive embedding vector; the coupled anomaly states include structural anomalies inducing gas micro-leaks or local leaks promoting structural deterioration. The method for obtaining the fusion interaction embedding vector includes: Full alignment is performed on the data in the multimodal feature dataset along the time-space dimension to obtain a multimodal feature vector set; Initialize a cross-attention multimodal fusion network, which includes two layers of cross-attention encoders and one layer of multilayer perceptron; In the first-layer cross-attention encoder, the structural feature vector is used as the query vector and the gas feature vector is used as the key-value pair. The structure-gas fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism and then converted into a structure-gas fusion vector. In the second-layer cross-attention encoder, the structure-gas fusion vector is used as the query vector and the environmental feature vector is used as the key-value pair. The structure-gas-environment fusion attention weight matrix is ​​constructed through the scaling dot product attention mechanism. Each row of the structure-gas-environment fusion attention weight matrix is ​​converted into a fusion interaction vector; the fusion interaction vector is then sequentially input into a multilayer perceptron and subjected to a nonlinear transformation using the ReLU activation function, finally outputting a deeply fused fusion interaction embedding vector; the fusion interaction embedding vector contains nonlinear correlation information of the three modalities of structural features, gas features and environmental features existing at the same spatial or temporal point; The method for obtaining the coupling abnormal state includes: The fused interactive embedding vector is input into the coupling anomaly identification model to obtain the coupling anomaly diagnosis result; the coupling anomaly diagnosis result is used as the coupling anomaly state. The defect type identification module identifies early defect types in gas pipelines in utility tunnels based on fused interactive embedded vectors and coupled abnormal states.

8. A defect identification and inspection robot for gas pipelines in a utility tunnel, comprising a memory and a processor, and a computer program stored in the memory and running in the processor, characterized in that, When the processor executes the computer program, it implements the method for identifying defects in gas pipelines in pipe corridors as described in any one of claims 1-6.

Images