AI identification method and system for leakage area of drainage pipe network

Abstract

The invention relates to the technical field of leakage detection, in particular to a drainage pipe network leakage area AI identification method and system, and the method comprises the following steps: collecting sound pressure changes through an acoustic collection device, extracting candidate leakage sections, collecting images through an inspection robot, screening training samples, analyzing brightness and profile changes, adjusting the training weight, and identifying the leakage position. And calculating a boundary coincidence area ratio, and adjusting model parameters in combination with a pipe section material. According to the method, leakage candidate sections are compared and screened based on continuous fluctuation of a sound pressure multi-frequency channel, the problem of frequent occurrence of misrecognition under noise interference is avoided, a complexity index is constructed by combining a texture direction and a span in an image, and sample screening is performed; the gray-scale curved surface constructed in the non-abnormal area in the image is compared with the profile trend of the abnormal area to identify the abnormal structure, the extraction capability of irregular leakage features is enhanced, the leakage boundary and material coding data are superposed to evaluate the identification interference degree of the mixed section, and the adaptability and stability of model parameters under different materials are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of leakage detection technology, and in particular to an AI-based method and system for identifying leakage areas in drainage pipe networks. Background Technology

[0002] Leak detection technology encompasses various methods and devices for identifying and monitoring potential leaks in liquid or gas transport systems. It is widely used in oil and gas pipelines, water supply and drainage systems, heating systems, and industrial production pipelines. It focuses on using sensor detection, signal analysis, and pipeline network operation status monitoring to determine and locate the position, amount, and development trend of potential or existing leaks in pipeline systems. This involves sensor detection methods, acoustic detection methods, gas tracing methods, pressure analysis methods, image recognition methods, and intelligent data analysis methods. By combining multiple signal sources or operating parameters, it monitors changes in pipeline structure or fluid state to extract and identify leak characteristics. Among these, drainage... The AI-based method for identifying leakage areas in water pipe networks refers to the process of identifying leakage areas in underground pipes within urban drainage systems. It involves acquiring internal pipe image data using image acquisition equipment, constructing a target recognition model based on feature extraction methods, extracting and comparing surface features of the pipe structure, and determining the leakage location by identifying morphological features such as cracks, water stains, and subsidence in the images. A convolutional neural network model is then used to learn multi-layer features from the training data, while a classifier is employed to classify and label abnormal areas, resulting in a region identification map for leakage areas in the drainage network. The process encompasses three steps: image data acquisition, feature region labeling, and model training and inference, and falls under the subcategory of automatic recognition based on image processing.

[0003] Traditional drainage network leakage area identification technology uses image feature extraction and model classification as the core process for leakage identification. It relies on single obvious features such as cracks and water stains in the image for judgment. When the image acquisition area does not contain obvious shapes or local features are weakened due to lighting conditions, leakage is prone to being missed or misjudged. There is no spatial mapping relationship between the continuous numbering of pipe nodes and the order of image frames, so image position alignment correction is not possible. This can easily cause the leakage area to deviate from its original position on the actual pipe section. Furthermore, it ignores the interference of different pipe materials on the boundary shape of the identification model output. In sections with mixed heterogeneous materials, type confusion often occurs, resulting in inconsistency in the system's identification boundary and a decrease in classification accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-based method and system for identifying leakage areas in drainage pipe networks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an AI-based method for identifying leakage areas in drainage pipe networks, comprising the following steps:

[0006] S1: Using acoustic acquisition equipment, collect sound pressure change data of drainage pipe section in multiple frequency channels, analyze time-series curves at multiple frequencies, compare sound pressure amplitude changes in adjacent time periods, analyze multi-frequency fluctuation trends, and combine equipment location records to generate candidate leakage sections and generate candidate leakage section information.

[0007] S2: Extract the information of the candidate leakage section, use the inspection robot to collect pipe section images, divide the image blocks, calculate the grayscale texture direction and distribution span, construct the texture complexity index, compare the difference between the image and the sample set, screen training samples, and obtain the training sample screening results.

[0008] S3: Based on the training sample screening results, analyze the brightness distribution of non-abnormal areas in the image, construct a reference gray-level trend surface, calculate the gray-level profile change, identify image areas with reverse gray-level trend changes, calculate the degree of concentration of profile change and adjust the training weight of the target area to obtain the recognition model training results.

[0009] S4: Using the training results of the recognition model, based on the mapping relationship between the numbering order of the image frames and the cumulative length of the movement path of the drainage network inspection robot, analyze the offset trend, adjust the position of the image frames, identify the leakage location, and obtain the mapping coordinates of the leakage location.

[0010] As a further aspect of the present invention, the candidate leakage section information includes frequency difference sequence, frequency fluctuation synchronization, and equipment positioning information; the training sample screening results include the number of image texture directions, texture distribution span, and texture complexity index; the recognition model training results include abnormal region boundaries, grayscale profile change trends, and profile change concentration; and the leakage location mapping coordinates include image frame index, movement path segment number, and structural reference node position.

[0011] As a further aspect of the present invention, the step of obtaining the candidate leakage section information is as follows:

[0012] S111: Using acoustic acquisition equipment, acquire sound pressure data of drainage pipe sections at multiple spatial locations, set multiple frequency channel identifiers, construct a time-series sound pressure sequence for each frequency channel, compare the sound pressure amplitude values ​​of adjacent time segments in the time-series sequence in turn, obtain the sound pressure change direction of each frequency channel between multiple time segments, and generate a sound pressure change direction sequence.

[0013] S112: Based on the sound pressure change direction sequence, analyze the number of frequency groups with the same continuous change direction in each frequency channel, calculate the degree of synchronous change of multiple frequencies in adjacent time windows, calculate the relative fluctuation consistency index of frequency segments, and generate fluctuation synchronization analysis results.

[0014] S113: Based on the fluctuation synchronization analysis results, match the acoustic sampling location index, detect and record the candidate leakage sections, record the corresponding timestamp and spatial number, and obtain the candidate leakage section information.

[0015] As a further aspect of the present invention, the step of obtaining the training sample screening results is as follows:

[0016] S211: Obtain the information of the candidate leakage section, collect the corresponding pipe section image, divide it into multiple image blocks, calculate the number of grayscale texture directions and the grayscale direction span of each block, and use the number of directions and span to form a coverage index to generate texture direction distribution coverage.

[0017] S212: Call the texture direction distribution coverage, analyze the gray intensity extreme value of the block in each direction angle, calculate the maximum, minimum and span of gray value under multiple direction angles and form a difference benchmark, calculate the gray value change asymmetry index of each block, and establish gray texture structure deviation value.

[0018] S213: Based on the grayscale texture structure deviation value, determine the texture structure difference characteristics, construct a texture difference ranking priority matrix, and filter the samples to obtain the training sample filtering results.

[0019] As a further aspect of the present invention, the step of obtaining the training result of the recognition model is as follows:

[0020] S311: Based on the training sample selection results, extract the brightness, texture, and edge features of the sample images, detect and mark the locations of abnormal regions, analyze the brightness distribution data of non-abnormal regions, and generate a brightness distribution benchmark dataset.

[0021] S312: Based on the brightness distribution benchmark dataset, analyze the brightness distribution of non-abnormal areas in the sample image, construct a benchmark gray-level trend surface, calculate the direction change value of the gray-level profile of the target area, compare the difference with the direction value of the benchmark surface, calculate the concentration value of the direction difference of the profile, screen the gray-level trend reverse change area, and obtain the index value of the reverse change area.

[0022] S313: Call the index value of the reverse change area, count the number of continuous segments of gray-scale change in the profile path direction, calculate the segment concentration value and adjust the training weight of the target area, train the leakage location recognition model, and obtain the recognition model training result.

[0023] As a further aspect of the present invention, the step of obtaining the leakage location mapping coordinates is as follows:

[0024] S411: Obtain the training results of the recognition model, and establish a mapping between each image frame and the actual inspection path based on the image frame number order and the cumulative length of the drainage pipe network inspection robot's movement path to obtain image frame path mapping data.

[0025] S412: Based on the image frame path mapping data, compare the spatial direction difference between the path length curve of the image frame and the preset pipe segment structure path, analyze the spatial offset trend of continuous frame segments, determine the offset distribution of each frame and the structure path, and obtain the spatial offset distribution value.

[0026] S413: Call the spatial offset distribution value, and based on the alignment adjustment result of the image frame in the spatial structure, combine the leakage identification information of the identification model to calibrate the actual spatial coordinates of the leakage point and obtain the leakage location mapping coordinates.

[0027] As a further aspect of the present invention, the method further includes:

[0028] S5: Call the leak location mapping coordinates, calculate the ratio of the overlapping area of ​​the leak type boundary to the total area for the leak boundary contour layer output in the recognition image, combine the pipe segment material information, compare the degree of mixing, evaluate the degree of interference of the pipe segment material type on the recognition result, adjust the recognition model parameters, and obtain the model feedback optimization record.

[0029] The model feedback optimization record includes the boundary overlap area, material information classification, and identification parameter adjustment content.

[0030] As a further aspect of the present invention, the step of obtaining the model feedback optimization record is as follows:

[0031] S511: Call the leakage location mapping coordinates, obtain the leakage boundary contour layer data in the recognition image, filter multiple leakage type areas, calculate the boundary overlap area between each type, and obtain the leakage boundary overlap area value.

[0032] S512: Based on the overlapping area value of the leakage boundary, calculate the area ratio of the overlapping area of ​​the boundary between various leakage types to the total area of ​​their respective boundaries, and arrange the area ratios in correspondence with the pipe segment material category information to obtain material and area ratio data.

[0033] S513: Call the material and area ratio data, compare the mixing degree of the lower boundary types of various pipe section materials, analyze the interference of materials on the leakage identification results, adjust the identification model parameters, and obtain the model feedback optimization record.

[0034] A drainage network leakage area AI recognition system, wherein the drainage network leakage area AI recognition system is used to execute the above-mentioned drainage network leakage area AI recognition method, the system comprising:

[0035] The sound pressure analysis module uses acoustic acquisition equipment to collect sound pressure variation data of drainage pipe sections in multiple frequency channels, analyzes the time-series curves at multiple frequencies, compares the sound pressure amplitude changes in adjacent time periods, analyzes the multi-frequency fluctuation trend, and combines the equipment location records to generate candidate leakage sections and generate candidate leakage section information.

[0036] The image filtering module extracts information on the candidate leakage sections, uses the inspection robot to collect images of the pipe sections, divides the image blocks, calculates the grayscale texture direction and distribution span, constructs a texture complexity index, compares the differences between the images and the sample set, filters training samples, and obtains the training sample filtering results.

[0037] Based on the training sample screening results, the grayscale recognition module analyzes the brightness distribution of non-abnormal areas in the image, constructs a baseline grayscale trend surface, calculates grayscale profile changes, identifies image regions with reverse grayscale trend changes, calculates the degree of concentration of profile changes and adjusts the training weights of the target region to obtain the recognition model training results.

[0038] The path matching module uses the training results of the recognition model to analyze the offset trend, adjust the position of the image frame, identify the leakage location, and obtain the leakage location mapping coordinates based on the mapping relationship between the numbering order of the image frames and the cumulative length of the movement path of the drainage pipe network inspection robot.

[0039] The boundary interference module calls the leak location mapping coordinates, calculates the ratio of the overlapping area of ​​the leak type boundary to the total area for the leak boundary contour layer output in the recognition image, compares the degree of mixing with the pipe segment material information, evaluates the degree of interference of the pipe segment material type on the recognition result, adjusts the recognition model parameters, and obtains the model feedback optimization record.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0041] In this invention, candidate leakage sections are screened based on continuous fluctuation comparison of sound pressure multi-frequency channels, avoiding the frequent problem of misidentification under noise interference. Complexity index is constructed by combining texture direction and span in the image and sample screening is performed. The gray-scale curved surface constructed in non-abnormal areas in the image is compared with the profile trend of abnormal areas to identify structural anomalies, enhancing the ability to extract irregular leakage features. The degree of interference in the identification of mixed segments is evaluated by superimposing leakage boundary and material coding data, ensuring the adaptability and stability of model parameters under different materials. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0043] Figure 2 This is a flowchart of the process for obtaining candidate leakage section information according to the present invention;

[0044] Figure 3 This is a flowchart of the process for obtaining the training sample screening results of the present invention;

[0045] Figure 4 This is a flowchart of the process for obtaining the training results of the recognition model of the present invention;

[0046] Figure 5 This is a flowchart of the process for obtaining the mapping coordinates of the leakage location according to the present invention;

[0047] Figure 6 The flowchart below shows the process for obtaining model feedback optimization records in this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0050] Please see Figure 1 This invention provides a technical solution: an AI-based method for identifying leakage areas in drainage pipe networks, comprising the following steps:

[0051] S1: Using acoustic acquisition equipment, collect sound pressure change data of drainage pipe section in multiple frequency channels, analyze time-series curves at multiple frequencies, compare sound pressure amplitude changes in adjacent time periods, analyze multi-frequency fluctuation trends, and combine equipment location records to generate candidate leakage sections and generate candidate leakage section information.

[0052] S2: Extract candidate leakage section information, use inspection robot to collect pipe section images, divide image blocks, calculate grayscale texture direction and distribution span, construct texture complexity index, compare the difference between the image and the sample set, screen training samples, and obtain training sample screening results.

[0053] S3: Based on the training sample screening results, analyze the brightness distribution of non-abnormal areas in the image, construct a baseline gray-level trend surface, calculate the gray-level profile change, identify image areas with reverse gray-level trend changes, calculate the degree of concentration of profile change and adjust the training weights of the target area to obtain the recognition model training results.

[0054] S4: Using the training results of the recognition model, based on the mapping relationship between the numbering order of the image frames and the cumulative length of the movement path of the drainage network inspection robot, analyze the offset trend, adjust the position of the image frames, identify the leakage location, and obtain the mapping coordinates of the leakage location.

[0055] S5: Call the leak location mapping coordinates, calculate the ratio of the overlapping area of ​​the leak type boundary to the total area for the leak boundary contour layer output in the recognition image, combine the pipe segment material information, compare the degree of mixing, evaluate the degree of interference of the pipe segment material type on the recognition results, adjust the recognition model parameters, and obtain the model feedback optimization record.

[0056] The candidate leakage section information includes frequency difference sequence, frequency fluctuation synchronization, and equipment positioning information. The training sample selection results include the number of image texture directions, texture distribution span, and texture complexity index. The recognition model training results include abnormal region boundaries, grayscale profile change trends, and profile change concentration. The leakage location mapping coordinates include image frame index, movement path segment number, and structural reference node position. The model feedback optimization record includes boundary overlap area, material information classification, and recognition parameter adjustment content.

[0057] Please see Figure 2 The steps for obtaining candidate leakage section information are as follows:

[0058] S111: Using acoustic acquisition equipment, acquire sound pressure data of drainage pipe sections at multiple spatial locations, set multiple frequency channel identifiers, construct a time-series sound pressure sequence for each frequency channel, compare the sound pressure amplitude values ​​of adjacent time segments in the time-series sequence in turn, obtain the sound pressure change direction of each frequency channel between multiple time segments, and generate a sound pressure change direction sequence.

[0059] Five acoustic acquisition nodes, numbered P1 to P5, were deployed at spatial locations inside the drainage pipe section. Each node was equipped with a sound pressure sensor to synchronously collect sound pressure data over different time periods. The data acquisition frequency for each node was 2kHz, and the sampling period was set to 0.1 seconds, for a total of 10 seconds, resulting in 100 frames of sound pressure data. Ten frequency channels were set for each frame, corresponding to frequencies from 200Hz to 2kHz, with a frequency interval of 200Hz. Each channel was numbered... For All raw sound pressure data matrices were indexed in two dimensions by position and channel. First, the sampled data for each frequency channel was normalized by subtracting the minimum value from each channel's data and then dividing by the difference between the maximum and minimum values ​​to obtain the normalized sound pressure value. Where i is the channel number and t is the time window number, for example, let i be the channel number and t be the time window number. The original sound pressure levels (SPL) of frames 1 to 5 at node P3 in channel P3 are [0.18, 0.21, 0.20, 0.22, 0.25] Pa, with a minimum of 0.18 and a maximum of 0.25. The normalized result is [(0.18-0.18) / (0.25-0.18), (0.21-0.18) / (0.07),…]≈[0, 0.4286, 0.2857, 0.5714, 1]. This sequence is used to construct the temporal SPL sequence for channel f4. Then, the difference between the normalized SPL amplitude values ​​in each pair of adjacent time frames is determined. For example, at t=1 and t=2... A positive difference is denoted as an increase (+1), resulting in a sequence direction of [+1,-1,+1,+1], which represents the change direction sequence of the frequency channel in a continuous time frame. Repeating this operation yields the change direction sequences of all frequency channels in each time period. These sequences are then aggregated and stored using node positions as the basic index, forming a multidimensional sound pressure direction sequence matrix, which serves as the basic data source for subsequent wave synchronization analysis.

[0060] S112: Based on the sound pressure change direction sequence, analyze the number of frequency groups with the same continuous change direction in each frequency channel, calculate the degree of synchronous change of multiple frequencies within adjacent time windows, and use the formula:

[0061] ;

[0062] Calculate the relative fluctuation consistency index of the frequency band and generate fluctuation synchronization analysis results;

[0063] in, It is a frequency band relative fluctuation consistency index, obtained by statistically averaging the consistency of the direction of sound pressure amplitude change among frequency channels over a continuous time period. This is the frequency channel index, representing the sequence number of the single frequency channel currently being analyzed. The total number of frequency channels represents the total number of frequency channels involved in the analysis. This is the time window index, indicating the starting time window number for the current analysis. Frequency channel index Index in time window The normalized sound pressure level (SPL) value is obtained by subtracting the minimum value of the channel across all time windows from the original SPL value and dividing by the difference between the maximum and minimum values. Frequency channel index Index in the next time window The normalized sound pressure level value is below. Frequency channel index The corresponding normalized frequency weight value is obtained by dividing the channel frequency by the mean of all channel frequencies. The arithmetic mean of the normalized frequency weights for all frequency channels is calculated by applying the values ​​to all channels. Summing and then dividing by get;

[0064] Based on the obtained sound pressure direction sequence, a frequency band fluctuation consistency analysis is performed on each frequency channel. First, the difference between the sound pressure change amplitudes of each group of adjacent time frames is calculated. Let the normalized sound pressure value array of the channel at a certain time t be: ; and the next time step t+1 is:

[0065] First, calculate the percentage change: ;For example: Then calculate the frequency normalization value. Given a frequency channel array of [200, 400, ..., 2000] Hz, an average frequency of 1100 Hz, and a normalized value of frequency divided by the average value, e.g., f1 = 200 Hz, then: ;average value Substitute the values ​​into the formula to calculate the weighted value for each channel item: ;

[0066] For example, channel 1:

[0067] ;

[0068] Repeat this operation, calculating for each of the 10 channels one by one. Let the final result array be:

[0069] ;

[0070] The average value is calculated as follows:

[0071] ;

[0072] The frequency band relative fluctuation consistency index refers to whether the direction and amplitude of sound pressure changes in different frequency channels exhibit a coordinated trend in physical behavior within two adjacent time segments, and is a dimensionless index value output through quantitative means. Its core meaning is to measure whether "multiple spatially close frequency channels simultaneously respond to the same disturbance mechanism," reflecting the consistency of frequency responses excited by potential leakage disturbances. This index does not rely on changes in absolute sound pressure values, but rather establishes a stable judgment dimension by normalizing fluctuation amplitude and frequency concentration, eliminating discrete changes caused by background noise, equipment errors, discontinuous excitation, etc. Its practical effect is reflected in: being able to locate and identify the frequency resonance response formed by intermittent, non-continuous, and micro-amplitude leakage disturbances in drainage pipe networks, and screening out frequency bands with physical coupling significance, significantly improving the sensitivity and anti-interference capability of leakage detection. This value is the frequency band relative fluctuation consistency index. In practical applications, we set the judgment threshold to 0.25. A value <0.25 is considered to indicate insufficient consistency in the direction of fluctuation; conversely, a value greater than 0.25 is considered a segment with strong synchronization. This threshold has been validated through extensive experiments. The effectiveness of the threshold was evaluated using the data shown in the table below, across 50 sample pipe segments:

[0073] Table 1. Fluctuation Consistency Threshold Sensitivity Test Table

[0074] Table 1 lists some of the test data. When the value changes to around 0.25, the recognition sensitivity and the false judgment rate are at an intersection point, so 0.25 is chosen as the optimal judgment threshold.

[0075] S113: Based on the wave synchronization analysis results, match the acoustic sampling location index, detect and record the candidate leakage sections, record the corresponding timestamp and spatial number, and obtain the candidate leakage section information;

[0076] In obtaining each time period After determining the index, the system first matches the index value with its corresponding time frame number and spatial location number, recording the time frame number as a vertical index. For example, the sampling time corresponding to the 18th frame in the current detection period is 1.8 seconds, and this data comes from the spatial node location numbered P4. Then, the effective frequency band range under that frame is filtered. From the channel direction synchronization results obtained in the previous calculation process, the set of channels with a consistency in change direction higher than the threshold of 0.25 is selected, such as those in f1 to f... 10 If only the directional synchronization of channels f3 to f8 is higher than the threshold among the 10 channels, then this set is marked as the effective frequency band range, and the corresponding frame is added to the recording structure. The value is calculated, and the flag status is set to "yes" to indicate that the current time and spatial location are candidate leakage sections. Then, the corresponding data from frames 20 and 23 are processed. For example, frame 20 corresponds to position P3. If the frequency is below the threshold, the effective frequency band range is an empty set, and the flag status is set to "No"; frame 23 corresponds to position P5. If the frequency exceeds the threshold and the effective frequency band range is f4 to f9, it is marked as "Yes". Throughout the process, each record includes a time frame number, spatial location number, and effective frequency band range. The numerical values ​​and whether a label is marked are stored in a unified candidate segment information array in chronological order, forming a mapping set from a three-dimensional index of time-space-frequency to leakage candidate labels. This mapping set will be called in subsequent steps as an input index in the image analysis stage, used to quickly locate the image frame number corresponding to the candidate segment from the video frames captured by the inspection robot. These image frames are then input into texture direction distribution calculation and brightness reference surface analysis, thereby maintaining the consistency and traceability of acoustic and visual detection data, and ensuring that the subsequent training process is only processed when both acoustic and visual features point to the same segment.

[0077] Please see Figure 3 The steps to obtain the training sample selection results are as follows:

[0078] S211: Obtain candidate leakage section information, collect corresponding pipe section images, divide them into multiple image blocks, calculate the number of grayscale texture directions and grayscale direction span of each block, and use the number of directions and span to form a coverage index to generate texture direction distribution coverage.

[0079] After obtaining information on candidate leakage sections, an inspection robot is used to acquire real-time images of the target drainage pipe section. The acquisition resolution is set to 640×480 pixels, and the entire image is evenly divided into 25 image blocks in 5 rows and 5 columns, with each block measuring 128×96 pixels. For each block, the number of grayscale texture directions and the grayscale direction span are calculated. The number of grayscale directions is obtained as follows: first, 12 direction templates are set in each block (angles from 0° to 165°, one direction every 15°). Then, the number of different gradient direction changes in each direction is counted. If the grayscale change in a direction has 3 gradient intensities (e.g., increasing, decreasing, and stable), the number of directions is counted as 3. The sum of all directions is the number of directions. The quantity and grayscale span are obtained by statistically analyzing the maximum difference in grayscale intensity along that direction, i.e., the difference between the maximum and minimum grayscale values ​​in the same direction. Assuming that block A has a maximum grayscale value of 220 and a minimum grayscale value of 110 in a 15° direction, the grayscale span is 110. The quantity and span of the 12 directions for each block are summed, and the product of the two is used as the coverage index. The coverage index is calculated as "total number of grayscale directions × average grayscale span". Assuming that the total number of directions for a certain block A is 24 and the average span is 85, the coverage index is 24 × 85 = 2040. This value will be used to generate the texture direction distribution coverage rate so as to compare the grayscale texture complexity with other image blocks and provide basic input data for subsequent image filtering and asymmetry index calculation.

[0080] S212: Analyze the texture direction distribution coverage, determine the extreme values ​​of grayscale intensity in each direction angle of the block, calculate the maximum, minimum, and span of grayscale values ​​under multiple direction angles, and establish a difference benchmark using the following formula:

[0081] ;

[0082] Calculate the asymmetry index of grayscale change for each block and establish the grayscale texture structure deviation value;

[0083] in, For the first The asymmetry index of grayscale changes in image blocks is obtained by analyzing the grayscale changes of the first image block in the image. Normalized analysis of grayscale differences from multiple directional angles within the block yielded the following results. The first division in the image Block index, The first calculated in each block Angle index for each direction. For the first Image blocks in the first The maximum grayscale value in each direction is obtained by counting the grayscale values ​​of all pixels in that direction and selecting the maximum value. For the first Image blocks in the first The minimum grayscale value in each direction is obtained by counting the grayscale values ​​of all pixels in that direction and selecting the minimum value. For the first Image blocks in the first The grayscale span in a direction is the grayscale difference between the maximum and minimum values. This is a normalization compensation term for the denominator, a constant set to avoid division by zero errors. It is generally dynamically adjusted based on the smallest unit of grayscale value. The set number of texture directions represents the number of different directions within each image block, which is determined by the preset number of direction templates;

[0084] Using the generated texture orientation distribution coverage, extract the maximum grayscale value for each of the 12 orientation angles within each image block. Minimum grayscale value And calculate the grayscale span. Taking the 7th block of the image as an example, at an angle of 30°, the maximum grayscale value is 210 and the minimum grayscale value is 90. Using the denominator compensation term And set the number of directions Substitute this data into the formula for calculating the asymmetry index of grayscale change:

[0085] ;

[0086] Substitute the direction value array of block 7, and let each direction... Given the values ​​[120, 100, 95, 80, 75, 60, 110, 130, 90, 85, 100, 105], calculate the corresponding values. Similarly, substituting into the formula, we first calculate the numerator:

[0087] ;

[0088] Example using the first direction:

[0089] ;

[0090] Summing in the same way in the other directions, the total sum of the numerators is approximately:

[0091] ;

[0092] Denominator:

[0093] ;

[0094] Substituting into the formula, we get:

[0095] ;

[0096] The grayscale variation asymmetry index is a quantitative indicator used to measure the degree of imbalance in intensity and span of grayscale distribution in different directions within an image block. A higher index value indicates stronger inconsistency in grayscale variation intensity across multiple directions, more complex texture structure, and significant directional feature differences within the image block. This parameter reflects the multi-directional asymmetry of textures in anomalous images such as potential cracks, erosion, and deposition, serving as an important basis for identifying typical leakage feature areas. During the training sample selection process, this index is used to construct a difference ranking matrix, thereby retaining image samples with more discriminative texture performance and eliminating low-difference images with homogeneous texture structures and insensitive changes, thus improving the quality of the training set and the recognition ability of subsequent models. The final grayscale variation asymmetry index for the 7th block is approximately 35.692. The advantage of this formula lies in the introduction of a quantization term based on the square root of the square of the directional span. This makes the direction with the larger the gray-scale difference and the more drastic the directional span contribute more to the overall asymmetry index, thus enhancing the quantification ability of deviation features.

[0097] S213: Based on the grayscale texture structure deviation value, determine the texture structure difference characteristics, construct a texture difference ranking priority matrix, and filter the samples to obtain the training sample filtering results;

[0098] After obtaining the grayscale texture structure deviation value of each image block, a deviation value set is established for the entire image block. A deviation value threshold of 30 is set. Based on the aforementioned block 7, the asymmetry index is 35.692, which is higher than the threshold; therefore, it is marked as a region with significant structural deviation. Conversely, if the grayscale change asymmetry index of block 12 is only 23.210, it is marked as having insignificant deviation. The sorting matrix is ​​constructed as follows: all blocks are sorted from high to low according to their asymmetry index, and each block is assigned a deviation level. The level classification criteria are: index values ​​greater than 40 are level one, 30-40 are level two, 20-30 are level three, and less than 20 are level four. Each level is assigned a corresponding priority screening weight, with weight ratios of 1.0 and 0. The weights 0.8, 0.6, and 0.3 will be combined with the sample labels to participate in the training sample selection process. Finally, image blocks with an asymmetry index of level 2 or above and a coverage index higher than the preset median value (based on 2040) will be selected as valid samples. In S212, the deviation value of 35.692 will be introduced. This value is higher than the screening threshold of 30 and the ranking level is level 2. Combined with the coverage index of 2400 of the image block, which is higher than the benchmark value of 2040, it is concluded that the block will be retained as a training sample. Conversely, if a block has a deviation value of 18.2 and a coverage index of only 1600, both of which fail to meet the standards, it will be removed from the training set. Finally, a sample set with deviation value and coverage weighted labels will be formed as the input data for subsequent model training.

[0099] Table 2. Sample set of image patch texture deviation and coverage index (example only)

[0100] Table 2 lists the actual combinations of deviation values ​​and coverage metrics in some image block samples, and generates the final training sample selection results based on this.

[0101] Please see Figure 4 The steps to obtain the training results of the recognition model are as follows:

[0102] S311: Based on the training sample selection results, extract the brightness, texture, and edge features of the sample images, detect and mark the locations of abnormal regions, analyze the brightness distribution data of non-abnormal regions, and generate a brightness distribution benchmark dataset.

[0103] Based on the sample image selection results, image data of the selected sample blocks were extracted. For each image block, brightness value, texture direction count, and edge features were independently extracted. Brightness values ​​were obtained using a linear grayscale reading method, extracting grayscale levels between 0 and 255 for each pixel in the 8-bit image, and generating a grayscale matrix for each block. Texture features were obtained through local direction count statistics. Assuming the image is divided into 25 blocks, each block was detected using 12 directional templates. Gradient abrupt changes in grayscale values ​​in each direction were extracted and their counts were statistically analyzed to form a texture direction count matrix. Edge features were obtained by extracting horizontal and vertical edges using the Sobel operator and then synthesizing the intensity to obtain an edge amplitude map. Three sets of edge amplitude maps were generated for each image block. The feature vector is then used to perform anomaly detection on the image. Pixel regions with brightness differences exceeding the local mean ±40 and edge amplitudes exceeding 120 are marked as anomaly regions. For example, if there are 640 pixels in block 8, and 78 of them simultaneously meet the criteria for brightness difference and edge amplitude, then this region is defined as an anomaly region. The remaining 562 pixels are classified as non-anomaly regions. For these non-anomaly regions, their brightness values ​​are statistically analyzed and a brightness histogram is constructed, divided into 8 brightness level segments. The pixel frequency of each segment is calculated to form a brightness frequency distribution table. Based on this frequency distribution, a brightness distribution benchmark dataset is established, which is used as a reference for the direction change value when S312 constructs the benchmark grayscale trend surface.

[0104] S312: Based on the brightness distribution benchmark dataset, analyze the brightness distribution of non-abnormal regions in the sample images, construct a benchmark gray-level orientation surface, calculate the direction change value of the gray-level profile of the target region, and compare the difference with the direction value of the benchmark surface using the following formula:

[0105] ;

[0106] Calculate the concentration value of the difference in the profile direction, screen out the areas of reverse change in grayscale trend, and obtain the index value of the reverse change area;

[0107] in, The value representing the concentration of differences in the cross-sectional direction is obtained by multiplying the directional differences by the pixel length, then weighting and normalizing them, and finally summing the results. For the target area The normalized value of the grayscale profile direction change at a location is obtained by extracting grayscale values ​​along the profile path of the target area, calculating the direction change, and then normalizing the result. The reference grayscale orientation surface in the first The normalized orientation value of the location is obtained by extracting the orientation value and normalizing it after establishing a surface based on the brightness distribution of non-abnormal regions. For the first The normalized value of the pixel length corresponding to the grayscale difference at a given location is obtained by measuring and normalizing the span of the grayscale abrupt change edge at that point. For the first The normalized texture density value at a location is obtained by calculating and normalizing the density of texture directions in the image at that point. For the first The normalized edge gradient magnitude is obtained by extracting image edges using the Sobel or Canny edge operator, calculating the gradient intensity, and then normalizing it. This refers to the index of the pixel position being analyzed along the grayscale profile path. This represents the total number of pixel positions involved in the calculation within the selected path.

[0108] Based on a baseline dataset of brightness distribution, image segments with stable brightness changes in non-abnormal regions are selected. A piecewise linear regression model is used to fit the grayscale change trend along the image's x-axis, constructing a baseline grayscale trend surface. Grayscale values ​​are extracted from regions in the target image along the profile path to form a grayscale profile line. The profile path length is set to 60 pixels, the corresponding path direction is the image's principal axis, and the sampling point is located in the 12th block. The resulting array of normalized values ​​for the target grayscale profile change is as follows: The normalized directional values ​​of the reference surface are Calculate the directional difference item by item Simultaneously, the grayscale mutation length of each point is extracted and normalized to... In actual image processing, if the grayscale abrupt change length at point v=1 is 3 pixels, and the maximum abrupt change length is set to 8 pixels, then... The same logic applies to the remaining points; texture density... The calculation is performed by normalizing the number of textures in different directions appearing within the 8-neighborhood of a point in the image. For example, if the number of directions at this point is 5 and the number of template directions is 12, then... Edge gradient magnitude The gradient strength at this point using the Sobel operator is 85, and the maximum gradient strength of the image is 150. Substitute the above parameters into the following formula:

[0109] ;

[0110] Let the path length be After summing the parameters of all pixels v, we get:

[0111] ;

[0112] ;

[0113] final:

[0114] ;

[0115] The profile direction difference concentration value represents the intensity of the spatial concentration of the difference between the brightness profile direction and the reference grayscale surface direction in a target area of ​​an image. It is a comprehensive index reflecting the abrupt clustering of grayscale profile changes in that area. Specifically, it considers not only the number of direction reversals but also the spatial clustering trend of these changes. Its calculation combines three types of information: grayscale difference, texture density, and edge amplitude, allowing the value to simultaneously reflect the suddenness of structural changes and the complex background of the image itself. A larger value indicates a large number of concentrated directional abrupt changes in the area, making it suitable as a strong candidate marker region for leakage areas; a smaller value indicates scattered changes or significant influence from complex textures, resulting in insufficient reliability, and the training weights should be reduced accordingly. This value indicates a significant difference between the path profile direction change and the reference surface. Based on an empirical threshold setting standard, when... The system was identified as having a reverse trend region. The current result meets the judgment criteria, and the index value of the reverse change region is 0.8996.

[0116] S313: Call the reverse change area index value, count the number of continuous segments of gray-scale change in the profile path direction, calculate the segment concentration value and adjust the training weight of the target area, train the leakage location recognition model, and obtain the recognition model training result.

[0117] The reverse change region index value is 0.8996. Image regions with this value greater than the threshold of 0.75 are selected as key analysis areas. The number of continuous gray-level change segments is counted along the profile path. Continuous change is defined as: gray-level segments with consistent gray-level change direction, gradient difference greater than 0.05, and a continuous span of more than 3 pixels. The path is set from v=1 to v=60. Five continuous segments meeting the criteria are detected in this segment, with segment lengths of 4, 5, 3, 6, and 3 pixels respectively. The concentration of the profile is calculated by dividing the number of segments by the total number of pixels in the path. Combined with the average length of the segments The training sample weights are set as follows: The weight range is set to [0,1]. The current value is in the low to medium range and belongs to the training intensity level L3. During the model training phase, this sample is input into the model dataset with a weight of 0.3148. The number of training rounds is reduced to 1 / 3 of the standard sample. Finally, the weight input and feature vector are merged and used to train the leakage identification model. The final training result is used for the next stage of spatial coordinate mapping and path difference analysis.

[0118] Table 3. Calculation of Profile Direction Indicators and Training Weights

[0119] As shown in Table 3, the concentration of differences in the cross-sectional direction The value and the continuity feature of the segment jointly influence the setting of training weights, ultimately determining the importance of the sample in model training.

[0120] Please see Figure 5 The steps to obtain the mapping coordinates of the leakage location are as follows:

[0121] S411: Obtain the training results of the recognition model, and establish a mapping between each image frame and the actual inspection path based on the image frame number order and the cumulative length of the drainage pipe network inspection robot's movement path to obtain image frame path mapping data.

[0122] After obtaining the training results of the recognition model, the corresponding image sequence is organized into a frame set according to the numbering order, and each frame is labeled with a number according to the generation order, with the numbering range set to frames 101 to 105. During the image acquisition process, the inspection robot moves at a constant speed and records its current position. For each frame captured, the system synchronously records the cumulative length of the robot's movement path in meters. Assuming a capture interval of 0.5 seconds and an average robot speed of 1 m / s, the path length difference between adjacent frames is 0.5 meters. Therefore, the corresponding path lengths for image frames 101 to 105 are 12.5 m, 13.0 m, 13.5 m, 14.0 m, and 14.5 m, respectively. m. Using this information, a one-to-one correspondence is constructed between the image frame number and the cumulative length of the inspection path. Then, a mapping index between the image frame and the physical structure path of the drainage pipeline is constructed. The image frame is regarded as a spatial profile snapshot, the frame number is used as the index primary key, and the cumulative path length is used as the horizontal spatial scale. The key image areas extracted from the image frame are mapped to the actual path coordinate system of the pipeline network, thereby obtaining the image frame path mapping data, which serves as the core basic data structure for subsequent spatial offset analysis and location calibration.

[0123] S412: Based on image frame path mapping data, compare the spatial orientation difference between the path length curve of the image frame and the preset pipe segment structure path, analyze the spatial offset trend of continuous frame segments, determine the offset distribution of each frame and the structure path, and obtain the spatial offset distribution value.

[0124] Based on image frame path mapping data, frame segment path curves are constructed in order of image frame number. Five image frames (101-105) are used to construct the actual path curve. The x-coordinate of each path point represents the cumulative path length, and the y-coordinate represents the horizontal center positioning line of the pipe in the central region of the image. The vertical distance of the pipe center from the trajectory line in each frame is calculated through image matching. The image center line is set as the standard path direction line. If the extracted pipe center in a frame image shows an upward offset in the image coordinate system, the corresponding spatial offset is positive; if it shows a downward offset, it is negative. The length of one pixel in the image is set to 0.01 m. The offset detection results are shown in Table 4. For frame 101, the detected offset pixel is 2, so its spatial offset value is 0.02 m; for frame 103, the offset pixel is 8, so its spatial offset value is 0.08 m. The spatial offset values ​​of all frames are obtained sequentially using this method. Then, the offset trend analysis of consecutive frame segments is performed. The difference in offset direction between each pair of adjacent image frames is extracted, and it is determined whether there is a continuous trend. If the spatial offset direction is the same in three consecutive frames and the offset difference is greater than 0.03... If the offset value is m, it is considered to be a stable spatial offset segment. If the offset values ​​of frames 103, 104, and 105 are 0.08 m, 0.03 m, and 0.10 m respectively, then the direction is not continuous and the offset difference is greater than 0.05, which is judged as no continuous trend. If the offset values ​​of frames 101, 102, and 103 are 0.02 m, 0.04 m, and 0.08 m respectively, all of which are positive values ​​and the trend is increasing, it is judged as an upward offset trend. This frame segment is recorded as "spatial offset segment - upward", and a spatial offset distribution value is generated as the basis for calculating the spatial alignment adjustment between the image frame and the structural path.

[0125] S413: Call the spatial offset distribution value, and based on the alignment adjustment results of the image frame in the spatial structure, combine the leakage identification information of the recognition model to calibrate the actual spatial coordinates of the leakage point and obtain the leakage location mapping coordinates.

[0126] The spatial offset distribution value is used to align the spatial positions of image frames 101 to 105 within the drainage network structure path. The cumulative path length and spatial offset value of each frame are merged into two-dimensional coordinates to generate the spatial coordinates of the leakage point. The cumulative path length is defined as the X-axis coordinate, and the spatial offset value as the Y-axis coordinate. In the example, frame number 103 corresponds to a cumulative path length of 13.5 m and a spatial offset of 0.08 m, so its corresponding spatial coordinates are (13.58, 0.15). The offset value is fine-tuned after locating the center grayscale segment in the image frame. Assuming the leakage center block in the image is offset by 2 pixels in the offset direction, the corresponding fine-tuning value is 0.02 m. The final Y-coordinate offset value is 0.08 + 0.02 = 0.10 m, therefore the leakage point location is (13.58, 0.15). Based on this, the leakage segments marked in the identification model are matched with the image frame numbers, and the image frames with leakage segments in the marked image numbers are selected and merged with the location coordinates to form a leakage location mapping coordinate dataset. This dataset can be output to the system layer as the final result for spatial decision-making applications such as leakage map annotation and on-site maintenance alignment. Its structure is shown in Table 4.

[0127] Table 4 Spatial Mapping and Offset Analysis Table

[0128] As shown in Table 4, each frame of image is mapped to spatial coordinate points through path length and image structure offset, forming the spatial calibration results of leakage points on the drainage pipe network structure.

[0129] Please see Figure 6 The steps for obtaining model feedback optimization records are as follows:

[0130] S511: Call the leakage location mapping coordinates, obtain the leakage boundary contour layer data in the recognition image, filter multiple leakage type areas, calculate the boundary overlap area between each type, and obtain the leakage boundary overlap area value.

[0131] The system calls the mapping coordinates of the leakage location to locate the corresponding frame in the identified image and reads the boundary contour layer stored in the image's layered structure. Each leakage area is labeled with a closed curve within the layer, and each area's boundary corresponds to a polygonal region, existing as a set of coordinate points. Each leakage area is also labeled with a type, such as "Type A" (e.g., high-frequency abrupt leakage) or "Type B" (e.g., low-contrast diffuse leakage). After extracting all labeled boundaries, the system performs intersection calculations on the boundary contours of different types of leakage within the same segment to obtain the overlapping areas between polygons. Boolean intersection operations are then used to calculate the overlapping closed areas between each type. For example, in a reinforced concrete pipe segment, type... The boundary area of ​​type A is 0.85 m², and that of type B is 0.65 m². The overlapping area of ​​the two is 0.42 m². Using the same method, leakage areas were extracted and overlap calculations were performed on the three pipe sections of polyethylene and ductile iron. The following three sets of data records were obtained: polyethylene pipe section A = 0.62 m², B = 0.57 m², overlap = 0.33 m²; ductile iron pipe section A = 0.91 m², B = 0.78 m², overlap = 0.58 m². After integrating the area records of the leakage contours of each material with the overlapping areas, a set of boundary overlapping area values ​​was formed, which served as the data source for the area ratio statistics in the next stage.

[0132] S512: Based on the overlapping area value of leakage boundaries, calculate the area ratio of the overlapping area of ​​multiple leakage types to the total area of ​​their respective boundaries, arrange the area ratios in correspondence with the pipe segment material category information, and obtain material and area ratio data.

[0133] Based on the set of overlapping area values, the area ratio of each leakage type to the overlapping area is calculated. The formula for calculating the area ratio is set as "overlapping area / total area of ​​each type of boundary". The results are uniformly output to three decimal places for subsequent comparative analysis. The calculation is performed on reinforced concrete pipe sections. The overlap ratio for type A is: The overlap ratio of type B is Polyethylene pipe section type A is Type B is Ductile iron pipe section type A is Type B is After the calculation is completed, each pipe segment material is used as the main index field, and arranged in a one-to-one correspondence with the boundary overlap ratio results of the two leakage types to form a joint structure table of material and area ratio, which serves as the input basis for subsequent interference identification analysis. The complete data structure record is shown in the table below:

[0134] Table 5: Analysis of the Overlap Ratio between Material and Leakage Boundary

[0135] As shown in Table 5, reinforced concrete pipe sections have a higher degree of overlap in Type B region, while the overlap ratio of ductile iron Type B boundary exceeds 0.74, and polyethylene has a lower overall degree of overlap.

[0136] S513: Call the material and area ratio data, compare the mixing degree of the lower boundary types of various pipe section materials, analyze the interference of materials on the leakage identification results, adjust the identification model parameters, and obtain the model feedback optimization record;

[0137] By calling up the material and boundary overlap ratio data, and comparing the differences in overlap ratios between type A and type B under different materials, the degree of boundary type mixing is analyzed. The mixing degree index is defined as the average of the overlap ratios of the two types. If the average value exceeds 0.60, the material is considered to have a strong tendency for boundary mixing; if it is less than 0.50, the boundary is considered to be clearly distinguishable. The mixing degree of reinforced concrete is calculated as follows: (0.494+0.646) / 2=0.570, polyethylene is (0.532+0.579) / 2=0.556, and ductile iron is (0.637+0.744) / 2=0.691. Based on this, it was determined that ductile iron caused the highest boundary confusion in model recognition, followed by reinforced concrete, and polyethylene caused the least. The analysis results were input into the model feedback optimization module. By adjusting the edge coincidence penalty coefficient in the boundary type feature extraction, the confidence of type B boundary in ductile iron pipe segments was reduced by 10%, type A was increased by 5%, type B in reinforced concrete was reduced by 5%, and the polyethylene type remained unchanged. The adjusted parameters were used as the optimization feedback record after this round of training so that they could enter the next round of iteration or deployment update configuration. The complete optimization parameter update record table will be retained in the system log for subsequent backtracking calls.

[0138] A drainage pipe network leakage area AI recognition system, which is used to execute the above-mentioned drainage pipe network leakage area AI recognition method, includes:

[0139] The sound pressure analysis module uses acoustic acquisition equipment to collect sound pressure variation data of drainage pipe sections in multiple frequency channels, analyzes the time-series curves at multiple frequencies, compares the sound pressure amplitude changes in adjacent time periods, analyzes the multi-frequency fluctuation trend, and combines the equipment location records to generate candidate leakage sections and generate candidate leakage section information.

[0140] The image filtering module extracts information on candidate leakage sections, uses an inspection robot to collect images of pipe sections, divides image blocks, calculates grayscale texture direction and distribution span, constructs texture complexity index, compares the differences between images and sample sets, filters training samples, and obtains training sample filtering results.

[0141] The grayscale recognition module analyzes the brightness distribution of non-abnormal areas in the image based on the training sample screening results, constructs a baseline grayscale trend surface, calculates grayscale profile changes, identifies image areas with reverse grayscale trend changes, calculates the degree of concentration of profile changes and adjusts the training weights of the target area to obtain the recognition model training results.

[0142] The path matching module uses the training results of the recognition model to analyze the offset trend, adjust the position of the image frame, identify the leakage location, and obtain the leakage location mapping coordinates based on the mapping relationship between the numbering order of the image frames and the cumulative length of the movement path of the drainage pipe network inspection robot.

[0143] The boundary interference module calls the leak location mapping coordinates, calculates the ratio of the overlapping area of ​​the leak type boundary to the total area for the leak boundary contour layer output in the recognition image, combines the pipe segment material information, compares the degree of mixing, evaluates the degree of interference of the pipe segment material type on the recognition results, adjusts the recognition model parameters, and obtains the model feedback optimization record.

[0144] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An AI-based method for identifying leakage areas in drainage pipe networks, characterized in that, Includes the following steps: S1: Using acoustic acquisition equipment, collect sound pressure change data of drainage pipe section in multiple frequency channels, analyze time-series curves at multiple frequencies, compare sound pressure amplitude changes in adjacent time periods, analyze multi-frequency fluctuation trends, and combine equipment location records to generate candidate leakage sections and generate candidate leakage section information. S2: Extract the information of the candidate leakage section, use the inspection robot to collect pipe section images, divide the image blocks, calculate the grayscale texture direction and distribution span, construct the texture complexity index, compare the difference between the image and the sample set, screen training samples, and obtain the training sample screening results. S3: Based on the training sample screening results, analyze the brightness distribution of non-abnormal areas in the image, construct a reference gray-level trend surface, calculate the gray-level profile change, identify image areas with reverse gray-level trend changes, calculate the degree of concentration of profile change and adjust the training weight of the target area to obtain the recognition model training results. S4: Using the training results of the recognition model, based on the mapping relationship between the numbering order of the image frames and the cumulative length of the movement path of the drainage network inspection robot, analyze the offset trend, adjust the position of the image frames, identify the leakage location, and obtain the mapping coordinates of the leakage location.

2. The AI ​​identification method for leakage areas in drainage pipe networks according to claim 1, characterized in that, The candidate leakage section information includes frequency difference sequence, frequency fluctuation synchronization, and equipment positioning information. The training sample screening results include the number of image texture directions, texture distribution span, and texture complexity index. The recognition model training results include abnormal region boundaries, grayscale profile change trends, and profile change concentration. The leakage location mapping coordinates include image frame index, movement path segment number, and structural reference node position.

3. The AI ​​identification method for leakage areas in drainage pipe networks according to claim 1, characterized in that, The steps for obtaining the candidate leakage section information are as follows: S111: Using acoustic acquisition equipment, acquire sound pressure data of drainage pipe sections at multiple spatial locations, set multiple frequency channel identifiers, construct a time-series sound pressure sequence for each frequency channel, compare the sound pressure amplitude values ​​of adjacent time segments in the time-series sequence in turn, obtain the sound pressure change direction of each frequency channel between multiple time segments, and generate a sound pressure change direction sequence. S112: Based on the sound pressure change direction sequence, analyze the number of frequency groups with the same continuous change direction in each frequency channel, calculate the degree of synchronous change of multiple frequencies in adjacent time windows, calculate the relative fluctuation consistency index of frequency segments, and generate fluctuation synchronization analysis results. S113: Based on the fluctuation synchronization analysis results, match the acoustic sampling location index, detect and record the candidate leakage sections, record the corresponding timestamp and spatial number, and obtain the candidate leakage section information.

4. The AI ​​identification method for leakage areas in drainage pipe networks according to claim 3, characterized in that, The steps for obtaining the training sample screening results are as follows: S211: Obtain the information of the candidate leakage section, collect the corresponding pipe section image, divide it into multiple image blocks, calculate the number of grayscale texture directions and the grayscale direction span of each block, and use the number of directions and span to form a coverage index to generate texture direction distribution coverage. S212: Call the texture direction distribution coverage, analyze the gray intensity extreme value of the block in each direction angle, calculate the maximum, minimum and span of gray value under multiple direction angles and form a difference benchmark, calculate the gray value change asymmetry index of each block, and establish gray texture structure deviation value. S213: Based on the grayscale texture structure deviation value, determine the texture structure difference characteristics, construct a texture difference ranking priority matrix, and filter the samples to obtain the training sample filtering results.

5. The AI ​​identification method for leakage areas in drainage pipe networks according to claim 4, characterized in that, The steps for obtaining the training results of the recognition model are as follows: S311: Based on the training sample selection results, extract the brightness, texture, and edge features of the sample images, detect and mark the locations of abnormal regions, analyze the brightness distribution data of non-abnormal regions, and generate a brightness distribution benchmark dataset. S312: Based on the brightness distribution benchmark dataset, analyze the brightness distribution of non-abnormal areas in the sample image, construct a benchmark gray-level trend surface, calculate the direction change value of the gray-level profile of the target area, compare the difference with the direction value of the benchmark surface, calculate the concentration value of the direction difference of the profile, screen the gray-level trend reverse change area, and obtain the index value of the reverse change area. S313: Call the index value of the reverse change area, count the number of continuous segments of gray-scale change in the profile path direction, calculate the segment concentration value and adjust the training weight of the target area, train the leakage location recognition model, and obtain the recognition model training result.

6. The AI ​​identification method for leakage areas in drainage pipe networks according to claim 5, characterized in that, The steps for obtaining the mapping coordinates of the leakage location are as follows: S411: Obtain the training results of the recognition model, and establish a mapping between each image frame and the actual inspection path based on the image frame number order and the cumulative length of the drainage pipe network inspection robot's movement path to obtain image frame path mapping data. S412: Based on the image frame path mapping data, compare the spatial direction difference between the path length curve of the image frame and the preset pipe segment structure path, analyze the spatial offset trend of continuous frame segments, determine the offset distribution of each frame and the structure path, and obtain the spatial offset distribution value. S413: Call the spatial offset distribution value, and based on the alignment adjustment result of the image frame in the spatial structure, combine the leakage identification information of the identification model to calibrate the actual spatial coordinates of the leakage point and obtain the leakage location mapping coordinates.

7. The AI ​​identification method for leakage areas in drainage pipe networks according to claim 1, characterized in that, The method further includes: S5: Call the leak location mapping coordinates, calculate the ratio of the overlapping area of ​​the leak type boundary to the total area for the leak boundary contour layer output in the recognition image, combine the pipe segment material information, compare the degree of mixing, evaluate the degree of interference of the pipe segment material type on the recognition result, adjust the recognition model parameters, and obtain the model feedback optimization record. The model feedback optimization record includes the boundary overlap area, material information classification, and identification parameter adjustment content.

8. The AI ​​identification method for leakage areas in drainage pipe networks according to claim 7, characterized in that, The steps for obtaining the model feedback optimization record are as follows: S511: Call the leakage location mapping coordinates, obtain the leakage boundary contour layer data in the recognition image, filter multiple leakage type areas, calculate the boundary overlap area between each type, and obtain the leakage boundary overlap area value. S512: Based on the overlapping area value of the leakage boundary, calculate the area ratio of the overlapping area of ​​the boundary between various leakage types to the total area of ​​their respective boundaries, and arrange the area ratios in correspondence with the pipe segment material category information to obtain material and area ratio data. S513: Call the material and area ratio data, compare the mixing degree of the lower boundary types of various pipe section materials, analyze the interference of materials on the leakage identification results, adjust the identification model parameters, and obtain the model feedback optimization record.

9. An AI-based system for identifying leakage areas in drainage pipe networks, characterized in that: The system is used to implement the AI ​​identification method for leakage areas in drainage pipe networks as described in any one of claims 1-8, and the system includes: The sound pressure analysis module uses acoustic acquisition equipment to collect sound pressure variation data of drainage pipe sections in multiple frequency channels, analyzes the time-series curves at multiple frequencies, compares the sound pressure amplitude changes in adjacent time periods, analyzes the multi-frequency fluctuation trend, and combines the equipment location records to generate candidate leakage sections and generate candidate leakage section information. The image filtering module extracts information on the candidate leakage sections, uses the inspection robot to collect images of the pipe sections, divides the image blocks, calculates the grayscale texture direction and distribution span, constructs a texture complexity index, compares the differences between the images and the sample set, filters training samples, and obtains the training sample filtering results. Based on the training sample screening results, the grayscale recognition module analyzes the brightness distribution of non-abnormal areas in the image, constructs a baseline grayscale trend surface, calculates grayscale profile changes, identifies image regions with reverse grayscale trend changes, calculates the degree of concentration of profile changes and adjusts the training weights of the target region to obtain the recognition model training results. The path matching module uses the training results of the recognition model to analyze the offset trend, adjust the position of the image frame, identify the leakage location, and obtain the leakage location mapping coordinates based on the mapping relationship between the numbering order of the image frames and the cumulative length of the movement path of the drainage pipe network inspection robot. The boundary interference module calls the leak location mapping coordinates, calculates the ratio of the overlapping area of ​​the leak type boundary to the total area for the leak boundary contour layer output in the recognition image, compares the degree of mixing with the pipe segment material information, evaluates the degree of interference of the pipe segment material type on the recognition result, adjusts the recognition model parameters, and obtains the model feedback optimization record.

Images