A water supply network leakage precision detection system

Abstract

This invention provides a precise leakage detection system for water supply networks, relating to the field of leakage detection technology. The system includes acquiring raw radar data covering the target area to be detected, generating suspected leakage points of interest (POIs), and generating dielectric constant characteristics as satellite feature data. It receives suspected leakage POIs and generates a robot inspection route by combining them with GIS data of the target area. The system performs leakage inspection according to the robot's route, collecting acoustic and optical multimodal data as robot feature data during the inspection process. The satellite feature data and robot feature data are then weighted and fused to obtain the leakage detection result. The beneficial effects are that system collaboration and data fusion solve the problem of disconnect between satellite and ground detection; the robot performs targeted detection based on satellite guidance, avoiding inefficient blind inspections and greatly improving efficiency; and the data fusion integrates macroscopic satellite features with fine on-site robot features, making the leakage location accuracy far higher than any single detection method.

Description

Technical Field

[0001] This invention relates to the field of leakage detection technology, and in particular to a precise leakage detection system for water supply networks. Background Technology

[0002] Pipeline leakage detection is a crucial step in ensuring the efficient use of water resources and the safe operation of pipeline networks. Currently, the mainstream technologies in this field mainly rely on satellite remote sensing or independent operation of ground-based detection equipment, and an effective collaborative detection system has not yet been formed.

[0003] In satellite remote sensing detection, existing technologies typically acquire backscattered data over a large area of ​​the Earth's surface using radar satellites. After processing, the soil dielectric constant is retrieved to preliminarily identify potential areas of leakage. However, this technology is limited by the resolution of satellite imagery and the accuracy of data interpretation. The identification results are usually for relatively large potential areas (e.g., points of interest with a radius of 100 to 200 meters), failing to pinpoint the exact location of leakage points. These preliminary screening results require further ground verification for confirmation.

[0004] For ground-based inspections, manual inspections using specialized equipment or ground-based mobile robots (such as wheeled or tracked robots) are typically employed. Manual verification is inefficient, with each suspected area taking several hours to inspect, and its positioning accuracy relies heavily on human experience, resulting in significant errors. Existing inspection robots largely depend on pre-set fixed routes or simple manual remote control. When performing large-scale pipeline network inspections, they lack guidance from satellite macroscopic screening results, still requiring traversal inspections and resulting in numerous blind inspection areas. The inspection efficiency is severely mismatched with the rapid, wide-area coverage capabilities of satellites.

[0005] In summary, the existing technologies exhibit a disconnect between satellite remote sensing and ground-based detection capabilities, specifically manifested in the following ways: First, there is a lack of technological collaboration; the lack of automated task coordination and path guidance mechanisms between satellite coarse screening results and ground-based fine inspection execution leads to a break in the process from suspected detection to on-site confirmation. Second, there is an imbalance in detection efficiency; the advantages of rapid, wide-area satellite screening cannot be complemented by the disadvantages of slow, small-scale ground-based verification, resulting in a long overall detection cycle and high resource consumption. Third, there is a gap in positioning accuracy; the suspected area provided by the satellite is too large, while ground equipment, without precise guidance, struggles to quickly focus on the core areas most likely to be missed, leading to insufficient final positioning accuracy of the missed points.

[0006] Therefore, there is an urgent need for a system that can organically combine wide-area screening by satellite remote sensing with precise detection by ground equipment, so as to achieve full-process collaboration from suspected area discovery to automatic planning of inspection paths, and then to on-site data collection and result fusion, thereby systematically improving the efficiency and accuracy of water supply network leakage detection. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a precise water supply network leakage detection system, comprising: a remote sensing initial screening subsystem, used to acquire raw radar data covering the target area to be detected, generate suspected leakage points of interest (POIs), and generate dielectric constant characteristics as satellite feature data; a collaborative control subsystem, connected to the remote sensing initial screening subsystem, used to receive the suspected leakage POIs and generate a robot inspection route by combining the GIS data of the target area; an inspection robot, connected to the collaborative control subsystem, used to perform leakage inspection according to the robot inspection route, and collect acoustic and optical multimodal data as robot feature data during the inspection process; and a data fusion subsystem, connected to the remote sensing initial screening subsystem and the inspection robot, used to perform weighted fusion of the satellite feature data and the robot feature data to obtain leakage detection results.

[0008] Preferably, the remote sensing initial screening subsystem includes: a data receiving module for acquiring raw radar data covering the target area from radar satellites; an image preprocessing module connected to the data receiving module for sequentially performing multi-view processing, radiometric correction, and geometric correction on the raw radar data to obtain a standard radar image; and a dielectric detection module for analyzing the dielectric constant features in the standard radar image to construct and match multi-band time-varying dielectric constant spectrum features, and then, in conjunction with pipeline geographical location information, extracting suspected leakage points of interest (POIs) and outputting satellite feature data containing POI coordinates, confidence levels, and corresponding dielectric constant features.

[0009] Preferably, the image preprocessing module includes: a multi-view processing unit, used to dynamically allocate weights for the pipeline body and its surroundings based on the grayscale characteristics of the pipeline area using an adaptive weighted multi-view algorithm, with 4 views in the azimuth direction and 2 views in the range direction, and then output a denoised radar image; a radiometric correction unit, connected to the multi-view processing unit, used to perform sensor error elimination on the denoised radar image, and to perform regional statistical correction for the backscattering differences of different underlying surfaces along the pipeline, and to establish a correction model by collecting ground object samples including dry soil, wet soil and the metal outer wall of the pipeline to eliminate the radiometric distortion of the denoised radar image; and a geometric correction unit, connected to the radiometric correction unit, used to convert the radiometrically corrected denoised radar image to the WGS-84 coordinate system, and to select ground control points along the pipeline, and to use a quadratic polynomial fitting algorithm to correct the geometric distortion of the denoised radar image, so that the error between the position of the pipeline in the image and the actual geographical coordinates is controlled within a preset error upper limit, and finally output a standard radar image.

[0010] Preferably, the dielectric detection module includes: a multi-band data acquisition unit, used to extract the backscattering coefficients of L, S, and C bands from the standard radar image, invert the multi-band soil dielectric constant, and record the dielectric constant variation data within a 24-hour time dimension, and construct a two-dimensional dielectric constant spectrum feature containing frequency band response dimension and time dimension based on the dielectric constant variation data; a water source feature matching unit, connected to the multi-band data acquisition unit, used to perform similarity matching between the two-dimensional dielectric constant spectrum feature and a preset water source dielectric spectrum feature library, the feature library containing spectral feature benchmarks for three types of water sources: water supply leakage, rainwater infiltration, and groundwater overflow; and a POI extraction unit, connected to the water source feature matching unit, used to extract dielectric anomaly areas within a set distance range on both sides of the pipeline axis and whose spectral features meet the water supply leakage criteria according to the matching results, using the center point of the dielectric anomaly area as the suspected leakage POI, and outputting the POI coordinates, confidence level, and corresponding dielectric constant feature as the satellite feature data.

[0011] Preferably, the collaborative control subsystem includes: a task planning module, used to receive suspected missed POIs generated by the remote sensing screening subsystem, and assign inspection tasks to the inspection robot according to the spatial distribution of the suspected missed POIs; a path optimization module, connected to the task planning module, used to integrate the GIS data of the target park with the coordinates of the suspected missed POIs, and generate the robot's inspection route using a path planning algorithm; and a status monitoring module, used to acquire the position, battery level, and data acquisition status of the inspection robot in real time, and generate and send an early warning command when the inspection robot deviates from the robot's inspection route by more than a preset deviation threshold, or when the battery level is lower than a preset battery level threshold.

[0012] Preferably, the acoustic-optical multimodal data includes acoustic signal features and infrared temperature difference features. The data fusion subsystem includes: a data access and preprocessing unit, used to receive satellite feature data from the remote sensing primary screening subsystem and acoustic signal features and infrared temperature difference features from the inspection robot, and perform standardization and noise reduction preprocessing respectively; a feature credibility evaluation unit, connected to the data access and preprocessing unit, used to calculate the single-modal credibility scores of the satellite feature data, the acoustic signal features, and the infrared temperature difference features respectively based on preset evaluation indicators related to pipeline leakage scenarios; and a dynamic feature fusion unit. The first unit, connected to the feature confidence evaluation unit, is used to determine the fusion weights of the satellite feature data, the acoustic signal features, and the infrared temperature difference features in the fusion process in real time through a dynamic attention weight allocation model, using the confidence scores of each single mode as input. It then performs fusion calculations on the weighted satellite feature data, the acoustic signal features, and the infrared temperature difference features to output a comprehensive leakage confidence value. The second unit, connected to the dynamic feature fusion unit, is used to compare the leakage confidence value with a preset confidence threshold, determine the leakage status based on the comparison result, and output the final leakage detection result.

[0013] Preferably, the data access and preprocessing unit includes: a satellite data preprocessing subunit, used to perform Z-Score normalization processing on the dielectric constant outliers in the satellite feature data from the remote sensing primary screening subsystem; an acoustic signal preprocessing subunit, used to perform frame averaging processing on the MFCC coefficients in the acoustic signal features from the inspection robot, and filter out noise frames with a signal-to-noise ratio lower than a preset threshold; and an infrared data preprocessing subunit, used to perform sliding window smoothing processing on the infrared temperature difference features from the inspection robot to suppress instantaneous temperature fluctuations.

[0014] Preferably, the feature credibility assessment unit includes: a satellite data credibility assessment subunit, used to calculate the single-mode credibility score of satellite feature data based on the duration of the satellite dielectric constant anomaly and its distance from the pipeline axis; an acoustic signal credibility assessment subunit, used to calculate the single-mode credibility score of acoustic signal features based on the matching degree and signal-to-noise ratio of acoustic signal features with a preset leak sound sample library; and an infrared data credibility assessment subunit, used to calculate the single-mode credibility score of infrared temperature difference features based on the fluctuation stability of infrared temperature difference and the correlation between its acquisition location and suspected leak POI.

[0015] Preferably, the dynamic feature fusion unit includes: a comprehensive confidence calculation subunit, used to receive the single-modal confidence scores of the acoustic signal features and the infrared temperature difference features, and synthesize the two based on a preset weight ratio to calculate the comprehensive confidence score of the robot features; an adaptive weight allocation subunit, connected to the comprehensive confidence calculation subunit, used to receive the single-modal confidence scores of the satellite feature data and the comprehensive confidence scores of the robot features, and dynamically allocate the fusion weights of the satellite feature data and the robot feature data according to their relative size relationship; and a confidence generation subunit, connected to the adaptive weight allocation subunit, used to perform weighted calculations with the allocated fusion weights and the corresponding satellite feature data and robot feature data respectively, fuse the weighted results and input them into the fully connected layer of the neural network for processing, and finally output a leakage confidence value in the range of 0 to 1.

[0016] Preferably, the allocation logic of the fusion weight in the adaptive weight allocation subunit is configured as follows: the higher the comprehensive credibility score of the robot features, the higher the fusion weight allocated to the robot feature data, and the lower the fusion weight allocated to the satellite feature data.

[0017] The above technical solution has the following advantages or beneficial effects:

[0018] 1. By automatically connecting satellite coarse screening and robot fine inspection through the collaborative control subsystem, a closed-loop workflow of satellite discovery-path planning-robot verification is constructed, which fundamentally changes the situation of isolated data and broken processes between satellite and ground equipment.

[0019] 2. The inspection robot performs targeted inspections directly under the guidance of satellite POIs, avoiding inefficient blind inspections of the entire area. This allows the satellite's rapid wide-area screening capabilities to complement the robot's on-site verification capabilities, significantly shortening the overall inspection cycle and improving resource efficiency.

[0020] 3. By using the data fusion subsystem to perform weighted fusion of macroscopic features from satellites and fine features from the robot on-site, the advantages of both are combined, resulting in a final accuracy of the location of the leak point that is much higher than that of the satellite coarse screening range and also better than the unguided autonomous robot detection. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a precise water supply network leakage detection system, which is a preferred embodiment of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment; other embodiments that conform to the spirit of the present invention may also fall within the scope of the present invention.

[0023] In a preferred embodiment of the present invention, based on the above-mentioned problems existing in the prior art, a precise detection system for water supply network leakage is provided, such as... Figure 1 As shown, it includes: a remote sensing initial screening subsystem 1, used to acquire raw radar data covering the target area to be detected, generate suspected missing points of interest (POIs), and generate dielectric constant characteristics as satellite feature data; a collaborative control subsystem 2, connected to the remote sensing initial screening subsystem 1, used to receive suspected missing POIs and generate robot inspection routes by combining them with GIS data of the target area; an inspection robot 3, connected to the collaborative control subsystem 2, used to perform missing inspections according to the robot inspection routes, and collect acoustic and optical multimodal data as robot feature data during the inspection process; and a data fusion subsystem 4, connected to the remote sensing initial screening subsystem 1 and the inspection robot 3, used to perform weighted fusion of satellite feature data and robot feature data to obtain missing detection results.

[0024] It should be noted that in this invention, "POI" is an abbreviation for "Point of Interest," meaning "point of interest" or "key point." In the context of this system, it specifically refers to "suspected missed points" with precise spatial locations identified through initial remote sensing screening. Each POI contains latitude and longitude coordinates, confidence level, and associated feature data, serving as core spatial data connecting satellite remote sensing screening with precise inspection by ground robots.

[0025] Specifically, this embodiment uses the leakage detection of a water supply network in a certain industrial park as an example to illustrate the detailed implementation process of this system:

[0026] 1. The remote sensing primary screening subsystem performs coarse screening of satellite data.

[0027] The remote sensing initial screening subsystem 1 initiates and coordinates with L-band radar satellites (such as L-SAR01A) to acquire raw radar data covering the target area in full polarization mode.

[0028] Subsequently, the remote sensing primary screening subsystem 1 processes the raw radar data, for example, by sequentially performing multi-view processing (using 4 views in the azimuth direction and 2 views in the range direction to suppress speckle noise), radiometric correction (eliminating the response error of the sensor itself), and geometric correction (converting and registering the image to the WGS-84 coordinate system) to improve data quality.

[0029] Based on the processed data, the system inverts the soil dielectric constant using the backscattering coefficient and, combined with the spectral characteristics of the dielectric constant across multiple frequency bands and time dimensions, effectively distinguishes between soil moisture anomalies caused by water supply pipeline leakage and interference from other water sources (such as rainwater and groundwater). The system marks the identified suspected water supply leakage areas with their center points as suspected leakage points of interest (POIs) and delineates a suspected area with a radius of 100-150 meters for each POI. Simultaneously, it extracts the dielectric constant characteristics of this area as satellite feature data.

[0030] Finally, the system encapsulates and outputs the above processing results (including POI coordinates, confidence level, dielectric constant characteristics, etc.).

[0031] 2. The collaborative control subsystem plans the robot's inspection tasks and paths:

[0032] The collaborative control subsystem 2 receives POI files from the remote sensing primary screening subsystem 1.

[0033] The system automatically assigns inspection tasks to standby inspection robots within the park based on the number and spatial distribution of Points of Interest (POIs) to avoid task overlap.

[0034] Next, the system integrates high-precision pipeline GIS data, terrain data, and received satellite POI coordinates from the park, and uses a path planning algorithm to generate the optimal robot inspection route. The core design of this route is to plan a high-density sampling path (e.g., sampling interval of 0.5 meters) in the core area of ​​the POI (e.g., within a radius of 50 meters), while using a conventional density in the outer area. This guides the robot to prioritize covering the areas most likely to be missed, thereby shortening the ineffective travel distance.

[0035] Throughout the inspection process, the system receives real-time data on the robot's location, battery level, and operational status. If the robot deviates from its planned route by more than a threshold or its battery is too low, the system will automatically issue a warning to ensure continuous task execution.

[0036] 3. Inspection robots perform detailed on-site inspections:

[0037] In this embodiment, the inspection robot 3 adopts a quadruped robot platform with strong terrain adaptability, and is equipped with a leak detector, dual-light camera, lidar, high-precision positioning and edge computing equipment.

[0038] The inspection robot 3 begins to move according to the inspection route issued by the collaborative control subsystem 2. Its high-precision positioning system enables it to reach the target POI area with centimeter-level accuracy.

[0039] During its journey, the robot continues to work: its onboard leak detector stays close to the ground to collect the sound signals of leaking water; dual-light cameras simultaneously acquire visible light and infrared images; and lidar perceives environmental obstacles in real time, providing data for the robot's autonomous obstacle avoidance and climbing.

[0040] The robot performs real-time preprocessing (such as noise reduction) on the collected sound signals, images and other data, collects and extracts effective robot feature data (such as the MFCC coefficient of the sound signal and the infrared temperature difference), and transmits it back in real time through the wireless network.

[0041] 4. Data fusion subsystem 4 completes the leakage confirmation:

[0042] The data fusion subsystem 4 receives satellite feature data (i.e. dielectric constant anomalies) from the remote sensing primary screening subsystem 1 and robot feature data from the inspection robot 3.

[0043] The system performs preprocessing such as standardization on the two types of data to prepare for fusion.

[0044] Subsequently, the system employs a multimodal fusion model based on an attention mechanism to perform weighted fusion of the two types of feature data. This model dynamically assigns weights based on the credibility of each type of data (e.g., the persistence of anomalies in satellite data, the signal-to-noise ratio of robot data, etc.), thereby highlighting the features of high credibility sources during fusion, improving the robustness of the judgment, and calculating a comprehensive leakage confidence value (range 0-1).

[0045] The system presets a judgment threshold (e.g., 0.9). When the confidence level is higher than this threshold, the location is confirmed as a leak point, and its precise coordinates are recorded (positioning error ≤ 1 meter), thus obtaining the final leak detection result. If it is in the middle range (e.g., 0.7-0.9), it is marked as pending review. If it is lower than the threshold, the possibility of leak is excluded.

[0046] Through the orderly coordination and deep cooperation of the above four subsystems, this embodiment realizes a complete closed loop from large-scale rapid screening to precise path guidance, then to fine on-site perception, and finally to data fusion decision-making, effectively solving the core problems of low efficiency, inaccurate positioning, and difficulty in coordination in traditional methods.

[0047] In a preferred embodiment of the present invention, such as Figure 1As shown, the remote sensing initial screening subsystem 1 includes: a data receiving module 11, used to acquire raw radar data covering the target area from radar satellites; an image preprocessing module 12, connected to the data receiving module 11, used to perform multi-view processing, radiometric correction and geometric correction on the raw radar data in sequence to improve data quality and obtain a standard radar image; and a dielectric detection module 13, used to analyze the dielectric constant characteristics in the standard radar image, construct and match multi-band time-varying dielectric constant spectrum characteristics to distinguish water supply leakage from other water source interference, and then combine pipeline geographical location information to extract suspected leakage POIs and output satellite feature data containing POI coordinates, confidence level and corresponding dielectric constant characteristics.

[0048] In a preferred embodiment of the present invention, the image preprocessing module 12 includes: a multi-view processing unit 121, used to dynamically allocate weights for the pipe body and the pipe periphery based on the grayscale characteristics of the pipe region using an adaptive weighted multi-view algorithm, based on 4 views in the azimuth direction and 2 views in the range direction, so as to retain the linear characteristics of the pipe body while suppressing speckle noise in the original radar data, and outputting a denoised radar image; and a radiation correction unit 122, connected to the multi-view processing unit 121, used to perform sensor error elimination on the denoised radar image and to address the backscattering differences of different underlying surfaces along the pipe. Regional statistical correction is performed by collecting ground object samples, including dry soil, moist soil, and the outer metal wall of the pipeline, to establish a correction model to eliminate radiation distortion caused by differences in the reflection of the underlying surface. Geometric correction unit 123, connected to radiation correction unit 122, is used to convert the radiation-corrected image to the WGS-84 coordinate system, select ground control points along the pipeline, and use a quadratic polynomial fitting algorithm to correct the geometric distortion of the radar image, so that the error between the position of the pipeline in the image and the actual geographic coordinates is controlled within the preset error upper limit, and finally output a standard radar image.

[0049] Specifically, in the pipeline inspection scenario of this embodiment, it is necessary to supplement targeted processing details based on the special characteristics of pipeline inspection to improve the adaptability of radar images to pipeline leakage detection. The specific processing process needs to be optimized around the requirements of low noise, high positioning accuracy, and clear linear features in the pipeline scene. The optimization aspects include: in the multi-view processing stage, based on 4 views in the azimuth direction and 2 views in the range direction, an adaptive weighted multi-view algorithm is adopted to dynamically adjust the weights according to the grayscale features of the pipeline area, assigning higher weights to the pipeline body (linear feature) area and normal weights to the surrounding soil (uniform area). While suppressing speckle noise (noise suppression rate ≥60%), it avoids blurring of key linear features such as pipeline interfaces and cracks caused by multi-view averaging, ensuring the clarity of subsequent leakage-related feature extraction; in the radiation correction stage, in addition to basic sensor error elimination (such as dark current correction, gain non-uniformity correction)... In addition to the positive aspect, regional statistical corrections are needed to address the differences in backscattering from different underlying surfaces such as soil and concrete pipe corridors along the pipeline. This involves collecting backscattering coefficient samples from typical features around the pipeline (such as dry soil, wet soil, and the metal outer wall of the pipeline) to establish a correction model, avoiding distortion of radiation values ​​in leak areas (such as wet soil) due to differences in underlying surface reflection. In the geometric correction stage, when converting to the WGS-84 coordinate system, optimization is required by combining GPS positioning data carried by the robot with ground control points (GCPs) along the pipeline. This involves selecting valves and marker posts every 50m along the pipeline as GCPs, obtaining their precise WGS-84 coordinates, and correcting geometric distortions in the radar image (such as offsets caused by distance migration and attitude errors) using a quadratic polynomial fitting algorithm. This ensures that the position of the pipeline in the image has an error of ≤0.5m from the actual geographic coordinates, meeting the requirements for leak point location.

[0050] All three processing methods must be executed without exception, and they must follow a fixed sequence of "multi-view processing → radiometric correction → geometric correction" to form a logically connected processing chain: First, multi-view processing suppresses speckle noise in the original radar data, reducing noise interference with subsequent corrections while preserving key pipeline features; then, radiometric correction is performed based on the denoised image to ensure that the difference in radiometric values ​​between the leaking area (such as moist soil caused by leakage) and the normal area truly reflects the actual situation, laying the data foundation for subsequent temperature / humidity anomaly correlation; finally, geometric correction binds the radiometrically corrected image to the actual geographic coordinates, ensuring that the subsequently identified leakage features can accurately correspond to the actual location of the pipeline, avoiding incorrect leakage point location due to coordinate deviations. If the order is adjusted (e.g., geometric correction first, then multi-view processing), the high noise of the original data will increase the control point matching error of geometric correction, or radiometric distortion will affect the weight allocation of multi-view processing, ultimately reducing the practical value of the radar image.

[0051] In a preferred embodiment of the present invention, such as Figure 1As shown, the dielectric detection module 13 includes: a multi-band data acquisition unit 131, used to extract the backscattering coefficients of L, S, and C bands from standard radar images, invert the multi-band soil dielectric constant, and record the dielectric constant change data in the 24-hour time dimension, and construct a two-dimensional dielectric constant spectrum feature containing the frequency band response dimension and the time dimension based on the dielectric constant change data; a water source feature matching unit 132, connected to the multi-band data acquisition unit 131, used to perform similarity matching between the two-dimensional dielectric constant spectrum feature and the preset water source dielectric spectrum feature library, which contains the spectrum feature benchmarks of three types of water sources: water supply leakage, rainwater infiltration, and groundwater overflow; and a POI extraction unit 133, connected to the water source feature matching unit 132, used to extract dielectric anomaly areas within a set distance range on both sides of the pipeline axis and whose spectrum features meet the water supply leakage criteria according to the matching results, take the center point of the dielectric anomaly area as the suspected leakage POI, and output the POI coordinates, confidence level, and corresponding dielectric constant features as satellite feature data.

[0052] Specifically, targeted optimization designs were made for the "spectral" feature construction, water source differentiation logic, and POI accurate extraction rules for the pipeline water supply leakage scenario in this embodiment.

[0053] The specific processing involves three steps: multi-dimensional spectral feature acquisition, water source feature database matching, and POI targeted extraction. The core is to achieve accurate screening of water supply leakage through dynamic features and scene constraints.

[0054] The first step is multi-band dielectric constant spectrum acquisition. The dielectric anomaly detection unit does not only acquire the dielectric constant at a single time point and in a single frequency band, but simultaneously acquires dielectric constant data in three key frequency bands (L-band, S-band, and C-band) and records the dielectric constant variation curves over 24 hours, forming a two-dimensional frequency-time spectral feature. The frequency band dimension reflects the difference in the dielectric constant's response to electromagnetic waves (e.g., in soil moist areas caused by water leakage, the dielectric constant fluctuates more significantly in the C-band), while the time dimension reflects the stability of the dielectric constant (e.g., in water leakage, the dielectric constant remains stable at 40-60 due to continuous seepage, while in rainwater infiltration shows a short-term increase followed by a gradual decrease).

[0055] The second step is water source spectral feature matching and differentiation. First, a dedicated database of dielectric spectral features of water sources along the pipeline is established, which stores the spectral feature benchmarks of three typical water sources: water supply leakage (dielectric constant 40-60, multi-frequency band synchronous stability, 24-hour fluctuation ≤5), rainwater infiltration (dielectric constant 35-55, short-term rapid increase followed by slow decrease within 1-3 hours, no synchronicity in frequency band response), and groundwater overflow (dielectric constant 20-30, smooth multi-frequency band response, 24-hour fluctuation ≤3). Then, the real-time collected two-dimensional dielectric constant spectral features are compared with the benchmarks in the water source dielectric spectral feature database. When the similarity is ≥85% and meets the characteristics of continuous stability + multi-frequency band synchronous response, it is judged as a suspected water supply leakage (excluding rainwater and groundwater interference).

[0056] The third step is the directional extraction of suspected leakage POIs. Combining the geometric correction results of radar images (WGS-84 coordinate system), the dielectric anomaly areas identified as suspected water supply leakage are spatially correlated with the actual pipeline route. Only the anomaly areas within 30 meters on both sides of the pipeline axis are retained. The center point of this area is used as the suspected leakage POI. The suspected area is dynamically delineated according to the pipeline burial depth (e.g., 100-meter radius for burial depth of 1-3 meters, 150-meter radius for burial depth of 3-5 meters). This avoids missing any areas (ensuring coverage of the possible spread of leakage) and also prevents too many invalid areas (excluding dielectric anomalies far from the pipeline).

[0057] In a preferred embodiment of the present invention, such as Figure 1 As shown, the collaborative control subsystem 2 includes: a task planning module 21, used to receive suspected missing POIs generated by the remote sensing screening subsystem, and assign inspection tasks to the inspection robot according to the spatial distribution of the suspected missing POIs, generating a task mapping relationship to avoid task overlap; a path optimization module 22, connected to the task planning module 21, used to integrate the GIS data of the target park with the coordinates of the suspected missing POIs, and generate the robot inspection route using a path planning algorithm; and a status monitoring module 23, used to obtain the position, power level, and data acquisition status of the inspection robot in real time; when the inspection robot deviates from the robot inspection route by more than a preset deviation threshold, or the power level is lower than a preset power level threshold, an early warning command is generated and sent.

[0058] Specifically, in this embodiment, the task planning module 21 receives the satellite coarse screening POI file, automatically assigns tasks to the quadruped robot (1 robot corresponds to 3-5 POI areas), and generates a POI area-robot task mapping table to avoid task overlap.

[0059] The path optimization module 22 integrates park GIS data (pipeline location, terrain) and satellite POI coordinates to plan robot paths, prioritizing coverage of core POI areas (with POI as the center, intensive sampling within a 50-meter radius, with an interval of 0.5 meters; regular sampling within a 50-150-meter radius, with an interval of 1 meter), thus reducing ineffective movement distances.

[0060] The status monitoring module 23 receives real-time data on the robot's position, battery level, and sensor data collection status. When the robot deviates from the planned path or the battery level drops below 20%, it automatically sends an early warning command to ensure continuous task execution.

[0061] Specifically, in this embodiment, the path optimization module optimizes the pipeline inspection scenario to obtain a three-dimensional correlation and fusion logic of "pipeline network-terrain-POI" to optimize the inspection path. The specific fusion process needs to be developed around the accuracy and efficiency requirements of pipeline inspection.

[0062] First, a unified coordinate calibration is performed. Using the WGS-84 coordinate system of the park's GIS data as a reference, three or more fixed markers within the park (such as pipeline valve wells and road corner stakes) are selected as ground control points to correct the error between satellite POI coordinates and GIS coordinates (ensuring coordinate deviation ≤ 0.3m). This avoids misalignment between POIs and actual pipeline locations due to coordinate system differences. Second, a hierarchical association model is constructed. The pipeline in the GIS data is segmented into main pipes, branch pipes, and interfaces. Each segment is bound to satellite POIs within a 500m radius along the pipeline. The suspected leakage risk level (e.g., high / medium / low) corresponding to the POI is marked in the GIS layer. At the same time, terrain data (e.g., slopes with a gradient > 15°, waterlogged areas) is converted into accessibility labels and associated with the POI areas. Finally, dynamic weight fusion is implemented. When the risk level of a POI is upgraded (e.g., high-risk POI), the weight of that POI in the fused data is automatically increased to ensure that subsequent path planning prioritizes high-value detection targets.

[0063] Furthermore, in this embodiment, the process of planning the robot path needs to be customized in conjunction with the constraints of the pipeline inspection scenario, specifically divided into four steps:

[0064] The first step is environmental modeling. Based on the fused GIS and POI data, a rasterized environmental map that the robot can recognize is constructed. The raster size is set to 0.2m × 0.2m. The raster attributes are labeled as passable areas (such as flat roads and pipelines), obstacle areas (such as buildings and steep slopes with a gradient > 20°), core POI areas (within a 50-meter radius), and regular POI areas (within a 50-150-meter radius), which clarifies the spatial constraints of path planning.

[0065] The second step is cost function optimization. The core cost function is designed as a base cost plus an additional cost. The base cost is the straight-line distance between grids, and the additional cost includes two parts: terrain additional cost (e.g., cost coefficient of 1.5 for slope areas and 1.0 for flat areas) and POI coverage additional cost (coverage weight of 1.8 for core POI areas, 1.2 for regular POI areas, and 0.8 for non-POI areas). The cost difference guides the algorithm to prioritize paths that cover high-value POIs and have low accessibility.

[0066] The third step is path search and pruning. The algorithm takes the robot's current position as the starting point and the inspection end point as the target point. During the search process, it calculates the cost and estimated distance of each candidate node in real time, prioritizes expanding nodes with low cost, and prunes invalid paths that are far away from the pipeline network and POI area (such as paths that deviate from the pipeline network axis by more than 10 meters) to avoid detours.

[0067] The fourth step is path smoothing. The obtained polyline path is smoothed and optimized using a third-order Bézier curve to ensure that the path curvature meets the robot's movement limits (turning radius ≥ 0.5m) and avoid sampling interruption caused by sharp turns.

[0068] More specifically, the effect of prioritizing coverage of the core POI area is achieved through a three-layer technical approach: density control, path sorting, and real-time verification. This ensures that the core area detection is thorough and efficient. First, sampling density is dynamically configured. During the path planning phase, the grid node density in the core POI area (within a 50-meter radius) is increased to twice that of the regular area, corresponding to a reduction in the robot sampling interval from 1 meter to 0.5 meters. Simultaneously, a core area sampling trigger command is set in the path code. When the robot's positioning data shows it has entered the core area, the sampling parameters are automatically switched to ensure denser collection of missed features. Second, path priority is prioritized. Before generating the global path, a closed-loop coverage path (such as a spiral or rectangular path centered on the POI) is planned for the core POI area. First, to ensure no blind spots in the core area, the closed-loop paths of multiple core POIs are connected through the shortest path, and finally the path of the regular POI area is connected to avoid invalid movement caused by back-and-forth trips. Second, real-time coverage verification is performed. The positioning module (such as RTK-GPS) on the robot records the location coordinates of the sampled positions in real time and compares them with the preset core area sampling point list. If the sampling point missing rate in a certain area is found to be greater than 5%, a supplementary sampling command is triggered, and a local supplementary sampling path is automatically generated until the core area sampling coverage reaches 100%. In addition, the core area movement speed is reduced to 60% of that in the regular area (e.g., from 0.8m / s to 0.5m / s) to ensure that the sampling equipment has enough time to capture the missing signals, further ensuring the detection accuracy of the core area.

[0069] In a preferred embodiment of the present invention, the acousto-optic multimodal data includes acoustic signal characteristics and infrared temperature difference characteristics, such as... Figure 1As shown, the data fusion subsystem 4 includes: a data access and preprocessing unit 41, used to receive satellite feature data from the remote sensing screening subsystem and acoustic signal features and infrared temperature difference features from the inspection robot, and to perform standardization and noise reduction preprocessing on them respectively; a feature credibility evaluation unit 42, connected to the data access and preprocessing unit 41, used to calculate the single-modal credibility scores of satellite feature data, acoustic signal features and infrared temperature difference features respectively based on preset evaluation indicators related to pipeline leakage scenarios; a dynamic feature fusion unit 43, connected to the feature credibility evaluation unit 42, used to take the single-modal credibility scores as input, and to determine the fusion weights of satellite feature data, acoustic signal features and infrared temperature difference features in real time through a dynamic attention weight allocation model, and to perform fusion calculation on the weighted satellite feature data, acoustic signal features and infrared temperature difference features, and output a comprehensive leakage confidence value; and a leakage judgment unit 44, connected to the dynamic feature fusion unit 43, used to compare the leakage confidence value with a preset confidence threshold, determine the leakage status according to the comparison result and output the final leakage detection result.

[0070] In a preferred embodiment of the present invention, such as Figure 1 As shown, the data access and preprocessing unit 41 includes: a satellite data preprocessing subunit 411, used to perform Z-Score normalization processing on the dielectric constant outliers in the satellite feature data from the remote sensing primary screening subsystem; an acoustic signal preprocessing subunit 412, used to perform frame averaging processing on the MFCC coefficients in the acoustic signal features from the inspection robot, and filter out noise frames with a signal-to-noise ratio lower than a preset threshold; and an infrared data preprocessing subunit 413, used to perform sliding window smoothing processing on the infrared temperature difference features from the inspection robot to suppress instantaneous temperature fluctuations.

[0071] In a preferred embodiment of the present invention, such as Figure 1 As shown, the feature credibility assessment unit 42 includes: a satellite data credibility assessment subunit 421, used to calculate the single-mode credibility score of satellite feature data based on the duration of the satellite dielectric constant anomaly and its distance from the pipeline axis; an acoustic signal credibility assessment subunit 422, used to calculate the single-mode credibility score of acoustic signal features based on the matching degree and signal-to-noise ratio of acoustic signal features with a preset leak sound sample library; and an infrared data credibility assessment subunit 423, used to calculate the single-mode credibility score of infrared temperature difference features based on the fluctuation stability of infrared temperature difference and the correlation between its acquisition location and suspected leak POI.

[0072] In a preferred embodiment of the present invention, such as Figure 1As shown, the dynamic feature fusion unit 43 includes: a comprehensive confidence calculation subunit 431, used to receive the single-modal confidence scores of the acoustic signal features and the infrared temperature difference features, and synthesize the two based on a preset weight ratio to calculate the comprehensive confidence score of the robot features; an adaptive weight allocation subunit 432, connected to the comprehensive confidence calculation subunit 431, used to receive the single-modal confidence scores of the satellite feature data and the comprehensive confidence scores of the robot features, and dynamically allocate the fusion weights of the satellite feature data and the robot feature data according to their relative size relationship; and a confidence generation subunit 433, connected to the adaptive weight allocation subunit 432, used to perform weighted calculations with the allocated fusion weights and the corresponding satellite feature vectors and robot feature vectors respectively, fuse the weighted results and input them into the fully connected layer of the neural network for processing, and finally output a leakage confidence value in the range of 0 to 1.

[0073] In a preferred embodiment of the present invention, the allocation logic of the fusion weight in the adaptive weight allocation subunit is configured as follows: the higher the comprehensive credibility score of the robot features, the higher the fusion weight allocated to the robot feature data, and the lower the fusion weight allocated to the satellite feature data.

[0074] Specifically, in this embodiment, for the pipeline leakage detection scenario, three types of features—satellite dielectric constant anomalies, robot acoustic signal MFCC coefficients, and infrared temperature difference—are adapted to the scenario. The specific logic for highlighting high-reliability features through dynamic weight adjustment and closed-loop reliability evaluation includes:

[0075] The model employs a four-step process—feature preprocessing, credibility assessment, dynamic attention allocation, and fusion computation—to accurately extract high-credibility features and output leakage detection results. The specific process is as follows:

[0076] The first step is multimodal feature preprocessing and standardization. First, the three types of input features are unified in dimensionality and noise filtered. Abnormal dielectric constant values ​​in satellite data (e.g., 3-5 for normal soil, 40-60 for leaky areas) are standardized using Z-Score to eliminate the impact of numerical range differences on fusion. The MFCC coefficients (13-dimensional feature vectors) of the robot's acoustic signal are processed by frame averaging to retain core acoustic features (e.g., high-frequency energy peaks of leaking sounds) and remove short-term noise frames (frames with a signal-to-noise ratio <10dB). The infrared temperature difference (temperature difference between the leaky area and its surroundings) is smoothed using a sliding window (window size 5 frames) to avoid interference from instantaneous temperature fluctuations. After preprocessing, all three types of features are converted into 256-dimensional feature vectors, laying the data foundation for subsequent fusion.

[0077] The second step is real-time reliability assessment of single-modal features. The model has a built-in reliability assessment module, which sets assessment indicators for the characteristics of each type of feature: For satellite dielectric constant anomalies, a reliability score (0-1 point) is calculated based on the duration of the anomaly (≥3 sampling periods for high reliability) and the distance from the pipeline network (<10 meters from the pipeline axis for high reliability); for robot acoustic signal MFCC coefficients, a score is calculated based on the voiceprint matching degree (≥85% matching degree with the leak sound sample library for high reliability) and the signal-to-noise ratio (≥15dB for high reliability); for infrared temperature difference, a score is calculated based on temperature difference stability (temperature difference fluctuation ≤0.5℃ / minute for high reliability) and correlation with POI (located in the POI area for high reliability). This step provides a quantitative basis for attention weight allocation, avoiding indiscriminate fusion of low reliability features.

[0078] The third step is a dynamic attention weight allocation based on credibility. The model does not use a fixed weight of 0.3 for satellite and 0.7 for robot. Instead, it dynamically adjusts the weight based on the single-modal credibility score: Let the credibility of satellite features be S, the credibility of robot sound signal be A, and the infrared temperature difference be T. First, calculate the comprehensive credibility of robot dual features (A×0.6+T×0.4, since sound and infrared data are closer to the leakage site). Then, the weights are adaptively adjusted using the formulas "Satellite weight = 0.3×S / (S+A×0.6+T×0.4) and Robot total weight = 0.7×(A×0.6+T×0.4) / (S+A×0.6+T×0.4)". For example, when the acoustic signal-to-noise ratio is high (A=0.9), the infrared temperature difference is stable (T=0.8), and the duration of the satellite dielectric constant anomaly is short (S=0.3), the satellite weight will drop to 0.15, and the robot's total weight will rise to 0.85, prioritizing the high reliability characteristics of the on-site data collection. If the satellite dielectric constant anomaly is clear (S=0.9), but the robot data is affected by environmental interference (A=0.4, T=0.5), the satellite weight can rise to 0.4 to compensate for the insufficiency of the on-site data.

[0079] The fourth step involves multimodal feature weighted fusion and leakage confidence output. The dynamically adjusted weights are multiplied by the corresponding feature vectors to obtain a weighted fused feature vector, which is then input into a fully connected layer for nonlinear transformation. The final output is a leakage confidence value of 0-1: a confidence value ≥ 0.85 is considered "high-confidence leakage"; 0.7-0.85 is "medium-confidence leakage"; and < 0.7 is considered "no leakage". Simultaneously, the model outputs the weight contribution percentage of each feature, clarifying "which type of feature is the core basis for leakage judgment," providing direction for subsequent manual review and further ensuring the reliability of the results.

[0080] Furthermore, to make the technical solution and advantages of the present invention clearer, the following uses the leakage detection of a water supply network in a university campus as an example to describe in detail the specific implementation process of the present invention.

[0081] Step 1: Data Acquisition and Processing of the Remote Sensing Primary Screening Subsystem

[0082] 1. Data reception: Activate the remote sensing preliminary screening subsystem, coordinate with L-band radar satellites (such as L-SAR01A) to receive raw radar data covering the campus area, with a data volume of approximately 50GB.

[0083] 2. Image Preprocessing: The remote sensing initial screening subsystem sequentially performs multi-view processing, radiometric correction, and geometric correction on the raw radar data. Multi-view processing suppresses speckle noise while paying particular attention to preserving the linear characteristics of the pipeline; radiometric correction performs regional statistical corrections for different underlying surfaces along the pipeline; geometric correction transforms the image to the WGS-84 coordinate system and performs fine correction using ground control points, ultimately outputting a high-quality standard radar image.

[0084] 3. Dielectric Detection and POI Generation: The remote sensing primary screening subsystem analyzes standard radar images to retrieve the soil dielectric constant. By constructing and matching multi-band time-varying dielectric constant spectral features, it effectively distinguishes water supply leakage from other water source interferences, extracting 11 suspected leakage POIs. Each POI has a suspected area with a radius of approximately 150 meters, with confidence levels between 0.75 and 0.88. Simultaneously, the corresponding dielectric constant features are extracted as satellite feature data.

[0085] Step 2: Task planning and path optimization of the collaborative control subsystem

[0086] 1. Task allocation: The collaborative control subsystem receives the POI file from the remote sensing primary screening subsystem, and assigns inspection tasks to the two inspection robots according to the spatial distribution of the 11 POIs, generating a task mapping relationship to avoid overlap.

[0087] 2. Path Optimization: The collaborative control subsystem integrates campus GIS data (including pipeline location and terrain) with POI coordinates to generate robot inspection routes. This route prioritizes coverage of the core POI areas (intensified sampling within a 50-meter radius, with an interval of 0.5 meters; regular sampling within a 50-150-meter radius, with an interval of 1 meter).

[0088] 3. Status monitoring initialization: The collaborative control subsystem establishes a monitoring list, sets early warning thresholds (such as deviation from the path > 2 meters, battery level < 20%), and starts real-time monitoring of the robot's position, battery level, and working status.

[0089] Step 3: On-site inspection by the inspection robot

[0090] The inspection robot starts according to the route issued by the collaborative control subsystem. Its high-precision positioning module integrates POI coordinates for initial position calibration. After the robot enters the core area of ​​the POI, the sound leak detector automatically releases and closes to the ground to collect sound signals, while the infrared camera collects temperature images. After noise reduction processing, the collected sound signals are used to extract sound signal features (MFCC coefficients), which, together with the infrared temperature difference features, are uploaded in real time as robot feature data.

[0091] Step 4: Data fusion subsystem leakage confirmation

[0092] 1. Data access and preprocessing: The data fusion subsystem receives satellite feature data (dielectric constant anomalies) and robot feature data (acoustic signal features, infrared temperature difference features), and performs preprocessing such as standardization and noise reduction respectively.

[0093] 2. Feature credibility assessment: The data fusion subsystem calculates the single-mode credibility scores of satellite feature data, acoustic signal features and infrared temperature difference features based on preset evaluation indicators related to pipeline leakage scenarios (such as abnormal duration, acoustic signature matching degree, temperature difference stability, etc.).

[0094] 3. Dynamic Feature Fusion: The data fusion subsystem takes the confidence scores of each single modality as input, determines the weights of the three types of features in the fusion through a dynamic attention weight allocation model, performs fusion calculation on the weighted features, and outputs a comprehensive leakage confidence value (0.93 in this example).

[0095] 4. Leakage detection: The data fusion subsystem compares the leakage confidence value (0.93) with the preset threshold (0.9) and determines that there is a leakage point with a positioning error of about 0.8 meters, thus obtaining the final leakage detection result.

[0096] Step 5: Summarizing and Processing Test Results

[0097] The collaborative control subsystem aggregates all results and generates a detection report. For POIs with confidence levels in the pending verification range, the robot is instructed to perform secondary encrypted sampling and supplement data, then re-integrate and determine the results. After the task is completed, the robot automatically returns. Managers arrange repairs based on the accurate leakage detection results and verify the repair effectiveness through robot re-inspection.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. A precise detection system for water supply network leakage, characterized in that, include: The remote sensing initial screening subsystem is used to acquire raw radar data covering the target area to be detected, generate suspected missing points of interest (POIs), and generate dielectric constant characteristics as satellite feature data. The collaborative control subsystem, connected to the remote sensing screening subsystem, is used to receive the suspected missing POIs and generate a robot inspection route by combining the GIS data of the target park. The inspection robot is connected to the collaborative control subsystem and is used to perform leakage inspections according to the robot's inspection route. During the inspection process, it collects acoustic and optical multimodal data as robot feature data. A data fusion subsystem, connecting the remote sensing primary screening subsystem and the inspection robot, is used to perform weighted fusion of the satellite feature data and the robot feature data to obtain the leakage detection result; The remote sensing primary screening system includes: The data receiving module is used to acquire raw radar data covering the target area from radar satellites; An image preprocessing module, connected to the data receiving module, is used to sequentially perform multi-view processing, radiometric correction, and geometric correction on the raw radar data to obtain a standard radar image. A dielectric detection module is used to analyze the dielectric constant features in the standard radar image to construct and match multi-band time-varying dielectric constant spectrum features. The dielectric detection module includes: The multi-band data acquisition unit is used to extract the backscattering coefficients of the L, S, and C bands from the standard radar image, invert the multi-band soil dielectric constant, and record the dielectric constant variation data in the 24-hour time dimension. Based on the dielectric constant variation data, a two-dimensional dielectric constant spectrum feature containing the frequency band response dimension and the time dimension is constructed. The water source feature matching unit is connected to the multi-band data acquisition unit and is used to perform similarity matching between the two-dimensional dielectric constant spectrum feature and the preset water source dielectric spectrum feature library. The feature library contains the spectrum feature benchmarks of three types of water sources: water supply leakage, rainwater infiltration and groundwater overflow.

2. The water supply network leakage precision detection system according to claim 1, characterized in that, The image preprocessing module includes: The multi-view processing unit is used to output a denoised radar image by dynamically allocating the weights of the pipeline body and the pipeline periphery based on the grayscale characteristics of the pipeline area, on the basis of 4 views in the azimuth direction and 2 views in the range direction, using an adaptive weighted multi-view algorithm. The radiation correction unit, connected to the multi-view processing unit, is used to perform sensor error elimination on the denoised radar image and perform regional statistical correction for the backscattering differences of different underlying surfaces along the pipeline. It also eliminates the radiation distortion of the denoised radar image by collecting ground object samples including dry soil, wet soil and the outer metal wall of the pipeline. The geometric correction unit, connected to the radiometric correction unit, is used to convert the radiometrically corrected denoised radar image to the WGS-84 coordinate system, select ground control points along the pipeline, and use a quadratic polynomial fitting algorithm to correct the geometric distortion of the denoised radar image, so that the error between the position of the pipeline in the image and the actual geographical coordinates is controlled within a preset error upper limit, and finally outputs a standard radar image.

3. The water supply network leakage precision detection system according to claim 1, characterized in that, The dielectric detection module also includes: The POI extraction unit, connected to the water source feature matching unit, is used to extract dielectric anomaly regions within a set distance range on both sides of the pipeline axis and whose spectral features meet the water supply leakage criteria, based on the matching results. The center point of the dielectric anomaly region is taken as the suspected leakage POI, and the POI coordinates, confidence level and corresponding dielectric constant features are output as the satellite feature data.

4. The water supply network leakage precision detection system according to claim 1, characterized in that, The collaborative control subsystem includes: The task planning module is used to receive suspected missing points of interest (POIs) generated by the remote sensing primary screening subsystem and assign inspection tasks to the inspection robot according to the spatial distribution of the suspected missing points of interest. The path optimization module, connected to the task planning module, is used to integrate the GIS data of the target park with the coordinates of the suspected missing POIs and generate the robot inspection route using a path planning algorithm. The status monitoring module is used to acquire the location, battery level, and data acquisition status of the inspection robot in real time. When it is detected that the inspection robot deviates from the inspection route by more than a preset deviation threshold, or the battery level is lower than a preset battery level threshold, an early warning command is generated and sent.

5. The water supply network leakage precision detection system according to claim 1, characterized in that, The acousto-optic multimodal data includes acoustic signal characteristics and infrared temperature difference characteristics; The data fusion subsystem includes: The data access and preprocessing unit is used to receive satellite feature data from the remote sensing primary screening subsystem and acoustic signal features and infrared temperature difference features from the inspection robot, and to perform standardization and noise reduction preprocessing respectively. The feature credibility assessment unit, connected to the data access and preprocessing unit, is used to calculate the single-mode credibility scores of the satellite feature data, the acoustic signal features and the infrared temperature difference features respectively based on preset assessment indicators related to pipeline leakage scenarios. The dynamic feature fusion unit, connected to the feature confidence evaluation unit, is used to take the confidence scores of each single mode as input, determine the fusion weights of the satellite feature data, the acoustic signal features and the infrared temperature difference features in the fusion process in real time through the dynamic attention weight allocation model, and perform fusion calculation on the weighted satellite feature data, acoustic signal features and infrared temperature difference features to output a comprehensive leakage confidence value. The leakage determination unit, connected to the dynamic feature fusion unit, is used to compare the leakage confidence value with a preset confidence threshold, determine the leakage status based on the comparison result, and output the final leakage detection result.

6. The water supply network leakage precision detection system according to claim 5, characterized in that, The data access and preprocessing unit includes: The satellite data preprocessing subunit is used to perform Z-Score normalization on the dielectric constant outliers in the satellite feature data from the remote sensing primary screening subsystem. The acoustic signal preprocessing subunit is used to perform frame averaging on the MFCC coefficients in the acoustic signal features from the inspection robot and filter out noise frames with a signal-to-noise ratio lower than a preset threshold. The infrared data preprocessing subunit is used to perform sliding window smoothing on the infrared temperature difference characteristics from the inspection robot to suppress instantaneous temperature fluctuations.

7. The water supply network leakage precision detection system according to claim 5, characterized in that, The feature credibility evaluation unit includes: The satellite data credibility assessment subunit is used to calculate the single-mode credibility score of satellite feature data based on the duration and location of the satellite dielectric constant anomaly and its distance from the pipeline axis. The acoustic signal credibility assessment subunit is used to calculate the single-mode credibility score of the acoustic signal features based on the matching degree between the acoustic signal features and the preset leakage acoustic sample library and its signal-to-noise ratio. The infrared data credibility assessment subunit is used to calculate the single-mode credibility score of infrared temperature difference characteristics based on the fluctuation stability of infrared temperature difference and the correlation between its acquisition location and suspected missing points of interest (POIs).

8. The water supply network leakage precision detection system according to claim 5, characterized in that, The dynamic feature fusion unit includes: The comprehensive credibility calculation subunit is used to receive the single-mode credibility scores of the acoustic signal features and the infrared temperature difference features, and synthesize the two based on a preset weight ratio to calculate the comprehensive credibility score of the robot features. An adaptive weight allocation subunit, connected to the comprehensive credibility calculation subunit, is used to receive the single-modal credibility score of the satellite feature data and the comprehensive credibility score of the robot features, and dynamically allocate the fusion weights of the satellite feature data and the robot feature data according to their relative size relationship. The confidence generation subunit, connected to the adaptive weight allocation subunit, is used to perform weighted calculations on the allocated fusion weights and the corresponding satellite feature data and robot feature data, respectively. After fusing the weighted results, the results are input into the fully connected layer of the neural network for processing, and finally output a leakage confidence value in the range of 0 to 1.

9. The water supply network leakage precision detection system according to claim 8, characterized in that, The allocation logic of the fusion weight in the adaptive weight allocation subunit is configured as follows: the higher the comprehensive credibility score of the robot features, the higher the fusion weight allocated to the robot feature data, and the lower the fusion weight allocated to the satellite feature data.

Images