A method and system for intelligent inspection of concrete impermeability

Abstract

This application relates to the field of concrete testing technology, and in particular to an intelligent inspection method and system for concrete impermeability. The method includes: S1, acquiring a global low-resolution image of a group of concrete sample columns to determine a set of coordinate points for abnormal areas with macroscopic anomalies; S2, based on the set of coordinate points for the abnormal areas, extracting spectral features and applying sound signal excitation to each abnormal area to determine the impermeability defect characteristic information corresponding to each abnormal area; S3, further establishing an independent spatiotemporal evolution file for the impermeability defects corresponding to each abnormal area, repeating steps S1 and S2 during periodic inspections to construct dynamic feature evolution information for each defect area; S4, based on the dynamic feature evolution information of each defect area, predicting the future development trend of the defect area, and generating a heatmap predicting the impermeability life of the concrete sample columns and an overall risk rating report. This application enables the prediction of the impermeability life of concrete, supporting predictive maintenance decisions.

Description

Technical Field

[0001] This application relates to the field of concrete testing technology, and in particular to an intelligent inspection method and system for concrete impermeability. Background Technology

[0002] In the field of concrete impermeability testing, traditional methods mainly rely on manual visual observation of leakage traces or the use of single image recognition technology to determine the surface of concrete samples. Common testing methods include visual inspection, analysis of two-dimensional images captured by color or infrared cameras, and surface anomaly detection algorithms based on single-modal convolutional neural networks.

[0003] The limitations of the above-mentioned concrete impermeability testing technology in practical applications are mainly manifested in: insufficient ability to detect leakage in its early stages, limited diagnostic depth of internal defects, lack of accumulation and analysis of spatiotemporal evolution data, and efficiency bottlenecks in large-scale sample testing. Summary of the Invention

[0004] This application provides a method and system for intelligent inspection of concrete impermeability to solve the above-mentioned problems.

[0005] Firstly, this application provides a method for intelligent inspection of concrete impermeability, the method comprising:

[0006] S1. Obtain a global low-resolution image of the concrete sample column group, analyze the global low-resolution image, and determine the set of coordinate points of the abnormal area where macroscopic anomalies exist.

[0007] S2. Based on the set of coordinate points of the abnormal area, perform spectral feature extraction and sound signal excitation on each abnormal area to determine the anti-permeability defect feature information corresponding to each abnormal area.

[0008] S3. Based on the anti-seepage defect feature information, establish an independent spatiotemporal evolution archive for the anti-seepage defect corresponding to each abnormal area. During periodic inspections, repeat steps S1 and S2 to store the anti-seepage defect feature information at different time nodes into the corresponding spatiotemporal evolution archive, thereby constructing dynamic feature evolution information for each defect area.

[0009] S4. Based on the dynamic feature evolution information of each defect area, use the long short-term memory network model to learn its evolution law, predict the future development trend of the defect area, and generate a heat map of the impermeability life prediction of the concrete sample column and an overall risk rating report.

[0010] The above technical solutions enable rapid localization of macroscopic anomalies, significantly reducing full-area scanning time and improving inspection efficiency; simultaneous acquisition of spectral and acoustic features overcomes the limitations of single sensors, improving the accuracy and robustness of internal defect detection; spatiotemporal evolution archives continuously record defect evolution data, providing a complete time-series basis for trend analysis; and LSTM-based prediction models can capture nonlinear evolution patterns, enabling quantitative prediction and risk visualization of impermeability life, supporting predictive maintenance decisions, and extending the service life of concrete structures.

[0011] Optionally, the step of acquiring a global low-resolution image of the concrete sample column group, analyzing the global low-resolution image, and determining the set of coordinate points of the abnormal area where macroscopic anomalies exist includes:

[0012] The global low-resolution image is acquired using a bionic compound eye vision sensor, wherein the bionic compound eye vision sensor is composed of an array of multiple micro-eye optical units, each micro-eye optical unit pointing in a different direction, thereby achieving one-time imaging of a group of concrete sample columns within a large field of view.

[0013] The global low-resolution image is input into a preset lightweight convolutional neural network model for analysis. The preset lightweight convolutional neural network model is trained to identify macroscopic abnormal patterns related to leakage from complex background textures, including color abnormal areas, irregular wet spot contours, and surface aggregated bubble areas.

[0014] The preset lightweight convolutional neural network model outputs the center point coordinates or bounding box coordinates of all identified abnormal regions, forming the set of coordinate points of the abnormal regions.

[0015] Optionally, the step of extracting spectral features and exciting acoustic signals for each anomalous region to determine the anti-permeability defect feature information corresponding to each anomalous region includes:

[0016] The abnormal region is scanned using a hyperspectral imaging sensor to obtain its continuous spectral cubic data in the visible to near-infrared bands.

[0017] A transient physical excitation is applied to the abnormal area by a mechanical striking device, and the echo signal generated by the transient physical excitation is collected by a high-precision microphone.

[0018] The continuous spectral cubic data is subjected to dimensionality reduction and feature extraction to quantify the ratio of depth to reflectance of specific absorption bands that are closely related to the moisture content and density of concrete, thereby forming a hyperspectral feature vector.

[0019] The echo signal is preprocessed and its features are extracted, and it is converted into an acoustic feature vector characterizing the impedance characteristics of the internal structure.

[0020] The hyperspectral feature vector is concatenated with the acoustic feature vector to obtain a multimodal feature vector, and the multimodal feature vector is classified and judged to determine the anti-permeability defect feature information.

[0021] Optionally, the step of performing feature classification on the multimodal feature vector to determine the impermeability defect feature information includes:

[0022] The multimodal feature vectors are input into a pre-trained multimodal fusion classifier;

[0023] The multimodal fusion classifier performs at least one of the following judgments based on the input features: determining whether the abnormal region has internal defects, identifying the specific type of the internal defects, and assessing the severity level of the internal defects;

[0024] The types of internal defects include capillary water seepage, microcracks, through cracks, and internal cavities, and the severity level is classified according to the size, depth, and activity of the defects.

[0025] Optionally, the training process of the multimodal fusion classifier includes:

[0026] Collect labeled concrete sample data with known defect types and grades, the labeled concrete sample data including paired hyperspectral data and echo signal data;

[0027] The hyperspectral and acoustic feature vectors are extracted separately and then fused to construct a training dataset.

[0028] Using the training dataset, an end-to-end multilayer perceptron network based on an attention mechanism is trained with cross-entropy as the loss function. The attention mechanism is used to adaptively assign different weights to features from different modalities during the fusion process, thereby optimizing the decision boundary of the classifier.

[0029] Optionally, the construction of dynamic feature evolution information for each defect region includes:

[0030] Create an independent database entry as its spatiotemporal evolution archive for each of the anti-permeability defect feature information confirmed in step S2;

[0031] In each inspection cycle, when the corresponding defect area is visited again, steps S1 and S2 are executed, and the newly acquired anti-seepage defect feature information of the area, together with its timestamp and environmental temperature and humidity data, is used as a new data record point.

[0032] The data recording points are appended and stored to the corresponding spatiotemporal evolution archive in chronological order to construct the dynamic feature evolution information, and the detection data at discrete time points are linked together to form a dynamic sequence describing the changes in the physicochemical properties of the defect region over time.

[0033] Optionally, the step of learning the evolution pattern of each defective region based on its dynamic feature evolution information and using a long short-term memory network model to predict the future development trend of the defective region includes:

[0034] The dynamic feature evolution information of each defect region is constructed into regular time series data, which is used as the input of the long short-term memory network model;

[0035] The Long Short-Term Memory Network model learns and memorizes the long-term dependencies and short-term fluctuation patterns of the dynamic feature evolution information over time through its internal gating mechanism.

[0036] The training objective of the Long Short-Term Memory Network model is to minimize the error between its prediction of future values ​​of defect features and the actual measured values ​​in subsequent inspections.

[0037] Based on the trained Long Short-Term Memory network model, and using the dynamic feature evolution information from the previous N time points as input, the key state parameters of the defective region at the next M time points are recursively predicted.

[0038] The key state parameters include at least: the equivalent size of the defect, the average moisture content of the region, and the defect activity level;

[0039] By comparing the key state parameters with the corresponding preset critical thresholds, the expected time point at which the corresponding defect area reaches a dangerous state is determined, which is the local impermeability life of the current defect area.

[0040] Optionally, the generated heat map of the impermeability life prediction and the overall risk rating report of the concrete sample column include:

[0041] For each pixel on the surface of the concrete sample column, the local impermeability life value is assigned according to the defect area to which it belongs.

[0042] All the local impermeability life values ​​are mapped to a color space, where different colors or shades of color represent different life ranges, thus forming a continuous color mapping map on the surface of the entire concrete sample column, namely the impermeability life prediction heat map.

[0043] Based on the prediction results of all defect areas, several key indicators were extracted, including: the shortest local impermeability life, the number of high-risk defect areas, and the average evolution acceleration of defects.

[0044] A risk rating matrix is ​​established, mapping the values ​​of the key indicators to a preset risk level range, and generating an overall risk rating report that includes the final risk level, the location of the main risk sources, and predictive maintenance recommendations.

[0045] Optionally, the method further includes:

[0046] Perform trend analysis on the time-series data stored in the spatiotemporal evolution archive;

[0047] Quantify each type of parameter in the impermeability defect characteristic information, including moisture index, defect activity index, first derivative or moving average over time;

[0048] Based on the results of the trend analysis, the next inspection plan for the defect area is dynamically adjusted. For active defects with a feature change rate exceeding the threshold, the inspection cycle is automatically shortened, while for stable defects, the inspection cycle is extended, thus achieving adaptive optimization of inspection resources.

[0049] Secondly, this application provides an intelligent inspection system for concrete impermeability, the system comprising:

[0050] An anomaly screening module is used to acquire a global low-resolution image of a group of concrete sample columns, analyze the global low-resolution image, and determine the set of coordinate points of anomaly areas where macroscopic anomalies exist.

[0051] The defect analysis module is used to extract spectral features and excite sound signals for each abnormal region based on the set of coordinate points of the abnormal region, and determine the anti-permeability defect feature information corresponding to each abnormal region.

[0052] The evolution analysis module is used to establish an independent spatiotemporal evolution archive for each abnormal area based on the anti-seepage defect feature information. During periodic inspections, steps S1 and S2 are repeated to store the anti-seepage defect feature information at different time points into the corresponding spatiotemporal evolution archive, thereby constructing dynamic feature evolution information for each defect area.

[0053] The report output module is used to learn the evolution law of each defect area based on the dynamic feature evolution information of each defect area using a long short-term memory network model, predict the future development trend of the defect area, and generate a heat map of the impermeability life prediction of the concrete sample column and an overall risk rating report. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a schematic diagram illustrating an application scenario provided in one embodiment of this application;

[0056] Figure 2 A flowchart of an intelligent inspection method for concrete impermeability provided in an embodiment of this application;

[0057] Figure 3 This is a structural schematic diagram of an intelligent inspection system for concrete impermeability provided in an embodiment of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0059] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0060] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0061] The limitations of existing concrete impermeability testing technologies in practical applications are mainly manifested in: insufficient ability to detect leakage in its early stages, limited diagnostic depth for internal defects, lack of accumulation and analysis of spatiotemporal evolution data, and efficiency bottlenecks in large-scale sample testing.

[0062] Based on this, this application provides a method and system for intelligent inspection of concrete impermeability. By combining multimodal sensing with intelligent algorithms, it achieves rapid localization of macroscopic anomalies, significantly reduces the scanning time of the entire area, and improves inspection efficiency. Simultaneous acquisition of spectral and acoustic features overcomes the limitations of single sensors, improving the accuracy and robustness of internal defect detection. The spatiotemporal evolution archive continuously records defect evolution data, providing a complete time series basis for trend analysis. The LSTM-based prediction model can capture nonlinear evolution laws, realize quantitative prediction of impermeability life and risk visualization, support predictive maintenance decisions, and extend the service life of concrete structures.

[0063] Figure 1 This application provides an illustration of an application scenario. When testing the impermeability of concrete, the method provided in this application enables quantitative prediction and risk visualization of impermeability life, supports predictive maintenance decisions, and extends the service life of concrete structures.

[0064] Specifically, the method provided in this application can be applied to any server. The server interacts with a bionic compound eye vision sensor, acquiring a global low-resolution image from the sensor. Through a combination of multimodal sensing and intelligent algorithms, rapid localization of macroscopic anomalies is achieved, significantly reducing full-area scanning time and improving inspection efficiency. Simultaneous acquisition of spectral and acoustic features overcomes the limitations of a single sensor, improving the accuracy and robustness of internal defect detection. A spatiotemporal evolution archive continuously records defect evolution data, providing a complete temporal basis for trend analysis. An LSTM-based prediction model captures nonlinear evolution patterns, enabling quantitative prediction and risk visualization of impermeability life, supporting predictive maintenance decisions, extending the service life of concrete structures, and ultimately providing the impermeability life prediction heatmap and overall risk rating report to relevant personnel to clarify the concrete's impermeability performance. Specific implementation methods can be found in the following embodiments.

[0065] Figure 2 This is a flowchart illustrating an intelligent inspection method for concrete impermeability according to an embodiment of this application. The method of this embodiment can be applied to the server in the above scenario. Figure 2 As shown, the method includes:

[0066] S1. Obtain a global low-resolution image of the concrete sample column group, analyze the global low-resolution image, and determine the set of coordinate points of the abnormal area where macroscopic anomalies exist.

[0067] S2. Based on the set of coordinate points of the abnormal area, perform spectral feature extraction and sound signal excitation on each abnormal area to determine the anti-permeability defect feature information corresponding to each abnormal area.

[0068] S3. Based on the anti-seepage defect feature information, establish an independent spatiotemporal evolution archive for the anti-seepage defect corresponding to each abnormal area. During periodic inspections, repeat steps S1 and S2 to store the anti-seepage defect feature information at different time nodes into the corresponding spatiotemporal evolution archive, thereby constructing dynamic feature evolution information for each defect area.

[0069] S4. Based on the dynamic feature evolution information of each defect area, use the long short-term memory network model to learn its evolution law, predict the future development trend of the defect area, and generate a heat map of the impermeability life prediction of the concrete sample column and an overall risk rating report.

[0070] Technical Background and Working Principle: The impermeability of concrete structures is directly related to their durability and safety. Traditional inspection methods mainly rely on manual visual observation or single sensors (such as optical cameras) for surface inspection, which are difficult to capture subtle changes in internal defects and cannot achieve long-term tracking and prediction. With the development of IoT and AI technologies, intelligent inspection has become possible. This embodiment uses a periodic inspection robot to simulate biological perception mechanisms, combining multimodal sensing and time-series analysis to achieve a complete process from macroscopic anomaly location to internal defect feature extraction and dynamic evolution prediction. Its core working principle is as follows: First, biomimetic vision is used to quickly screen macroscopic anomaly areas, avoiding the waste of resources from full-area high-resolution scanning; then, hyperspectral imaging and acoustic excitation are used to simultaneously acquire the spectral response and structural impedance characteristics of the material's interior, forming multimodal features to accurately describe defects; by establishing a spatiotemporal evolution archive, discrete inspection data are concatenated into a time series to capture the evolution trend of defects; finally, a Long Short-Term Memory (LSTM) network is used to model the time-series data, learn the long-term dependencies of defect evolution, thereby predicting future development trends and generating a visual report. This method upgrades point detection to continuous monitoring, enabling proactive early warning and life assessment of anti-seepage performance.

[0071] Technical Solution and Component Functions: The technical solution of this embodiment is executed by a periodic inspection robot, specifically including four steps: S1. A bionic compound eye vision sensor acquires a global low-resolution image of the concrete sample column group at once, and a lightweight convolutional neural network model is used to analyze the image to identify macroscopic anomaly patterns such as color anomalies, wet spot contours, or surface bubbles, and outputs the coordinate set of the center point or bounding box of the anomaly area; S2. For each anomaly area, a hyperspectral imaging sensor is used to scan and acquire spectral cube data in the visible to near-infrared bands. At the same time, a transient physical excitation is applied through a mechanical tapping device, and an echo signal is collected using a high-precision microphone. Subsequently, the spectral data is dimensionality reduced (e.g., principal component analysis) to extract hyperspectral feature vectors related to moisture content and density. Time-frequency feature extraction is performed on the echo signal to obtain acoustic feature vectors, and the two are concatenated into a multi-dimensional array. After the modal feature vectors are input into a multimodal fusion classifier, the existence, type (e.g., capillary seepage, microcracks, etc.) and severity level of defects are determined. S3: An independent spatiotemporal evolution archive database entry is created for each confirmed defect. In subsequent inspection cycles, S1 and S2 are repeated, appending newly acquired defect feature information along with timestamps and environmental temperature and humidity data to construct a dynamic feature evolution information sequence. S4: The time-series data of each defect is normalized and input into a long short-term memory network model. Through a gating mechanism, the model learns the evolution pattern, predicts future key state parameters (e.g., equivalent defect size, average regional moisture content, activity level), and compares them with preset thresholds to determine the local seepage resistance life. Finally, the life value is mapped to a color space to generate a seepage resistance life prediction heatmap. The shortest life, the number of high-risk defects, and other indicators are combined to generate an overall risk rating report through a risk rating matrix. The collaborative functions of each component achieve a closed-loop processing from data acquisition, feature fusion, time-series recording to prediction output.

[0072] Beneficial Effects: This embodiment achieves the following beneficial effects by combining multimodal sensing with intelligent algorithms: rapid localization of macroscopic anomalies, significantly reducing full-area scanning time and improving inspection efficiency; simultaneous acquisition of spectral and acoustic features, overcoming the limitations of single sensors and improving the accuracy and robustness of internal defect detection; continuous recording of defect evolution data in a spatiotemporal evolution archive, providing a complete time-series basis for trend analysis; and LSTM-based prediction models that can capture nonlinear evolution laws, enabling quantitative prediction and risk visualization of impermeability life, supporting predictive maintenance decisions, and extending the service life of concrete structures.

[0073] In some embodiments, a biomimetic compound eye vision sensor is used to acquire a global low-resolution image. The biomimetic compound eye vision sensor consists of an array of multiple micro-eye optical units, each pointing in a different direction, to achieve a single imaging of a group of concrete sample columns within a large field of view. The global low-resolution image is input into a preset lightweight convolutional neural network model for analysis. The preset lightweight convolutional neural network model is trained to identify macroscopic abnormal patterns related to leakage from complex background textures, including color abnormal areas, irregular wet spot contours, and surface aggregated bubble areas. The preset lightweight convolutional neural network model outputs the center point coordinates or bounding box coordinates of all identified abnormal areas, forming a set of abnormal area coordinate points.

[0074] Technical Background and Working Principle: In concrete seepage prevention inspections, quickly and comprehensively locating macroscopic anomalies on the surface is the primary step. Traditional inspections using single-lens cameras suffer from limited field of view, requiring multiple shots and image stitching, resulting in low inspection efficiency and the potential for missing defects due to stitching errors. This embodiment draws inspiration from the wide-angle imaging principle of insect compound eyes, using a biomimetic compound eye vision sensor to achieve a single large field of view imaging, overcoming the field of view limitations of traditional cameras. Its working principle involves an array of multiple micro-eye optical units, each pointing in a different direction, capturing local images in parallel, and fusing them at the hardware or software level into a global low-resolution image covering the entire sample column group. Subsequently, a lightweight convolutional neural network model is used to efficiently analyze this image. This model, trained on a large number of concrete surface anomaly images, can accurately identify macroscopic anomaly patterns related to leakage from complex background textures (such as aggregate shadows and formwork joints), thereby achieving rapid and accurate location of anomaly areas.

[0075] Technical Solution and Component Functions: The technical solution of this embodiment specifically includes: a biomimetic compound eye vision sensor consisting of multiple micro-eye optical units arranged in an array at a specific angle. Each unit uses a microlens with a fixed focal length, and the overall field of view covers approximately 120°, enabling the acquisition of a global low-resolution panoramic image of a group of concrete sample columns in a single operation. The acquired image is transmitted to a computing unit and input into a preset lightweight convolutional neural network model. This model typically includes an input layer, several convolutional layers employing depthwise separable convolutions, pooling layers for dimensionality reduction and feature extraction, and an output layer responsible for generating the center point coordinates or bounding box coordinates of the abnormal region, ultimately forming a set of coordinate points for the abnormal region. The biomimetic compound eye sensor is responsible for large-area rapid imaging, while the lightweight convolutional neural network undertakes the core computational task of intelligently identifying and locating abnormal regions from complex scenes.

[0076] Beneficial effects: This embodiment combines biomimetic optical design with deep learning recognition, achieving two core benefits: improved inspection efficiency: single-shot imaging avoids multiple moves and stitching, significantly shortening the initial scanning time; accurate anomaly localization: the lightweight neural network reduces computational resource consumption while ensuring accuracy, making real-time analysis on the mobile robot platform possible, and providing accurate target areas for subsequent fine inspection.

[0077] In some embodiments, an abnormal region is scanned using a hyperspectral imaging sensor to acquire its continuous spectral cubic data in the visible to near-infrared bands; a transient physical excitation is applied to the abnormal region using a mechanical tapping device, while the echo signal generated by the transient physical excitation is acquired using a high-precision microphone; the continuous spectral cubic data is subjected to dimensionality reduction and feature extraction to quantify the ratio of depth to reflectance of a specific absorption band closely related to the water content and density of concrete, forming a hyperspectral feature vector; the echo signal is preprocessed and its features are extracted to convert it into an acoustic feature vector characterizing the impedance properties of the internal structure; the hyperspectral feature vector and the acoustic feature vector are concatenated to obtain a multimodal feature vector, and the multimodal feature vector is classified and judged to determine the characteristic information of impermeability defects.

[0078] Technical Background and Working Principle: The impermeability defects in concrete are not only manifested on the surface, but more importantly, in changes in its internal porosity, moisture content, and microstructure. Single-dimensional visual inspection of the surface or internal acoustic inspection is insufficient to fully characterize the physicochemical nature of the defects. For example, vision cannot perceive internal moisture, while acoustics is insensitive to the chemical composition of the material. This embodiment achieves multi-dimensional perception of defects by simultaneously extracting spectral features and exciting acoustic signals. Its core working principle is as follows: Hyperspectral imaging, by acquiring continuous spectral data from the visible to near-infrared bands, can capture the absorption and reflection characteristics of specific chemical components in concrete, such as moisture and hydroxides, to light, thereby quantifying moisture content and density; while the stress waves generated by transient acoustic excitation (such as mechanical impact) propagate within the material, and the frequency attenuation and velocity characteristics of their echo signals are closely related to the material's elastic modulus, internal porosity, and cracks, effectively characterizing structural impedance. Fusing information obtained from these two different physical principles allows for cross-validation and a more comprehensive revelation of the type and severity of defects.

[0079] Technical Solution and Component Functions: The technical solution of this embodiment specifically includes: After locating the abnormal area, a hyperspectral imaging sensor performs a fine scan of the area to acquire continuous spectral cubic data covering the 400-1000 nm band; simultaneously, a mechanical impact device (such as an electromagnetic impact hammer) applies a transient physical excitation to the abnormal area with adjustable energy (such as 0.5-2 J), while a high-precision broadband piezoelectric microphone collects the generated echo signal at a sampling rate of not less than 48 kHz; subsequently, the data processing unit performs dimensionality reduction processing such as principal component analysis on the spectral cubic data, extracts the ratio of depth to reflectance of a specific absorption band closely related to the water content and density of concrete, and forms a hyperspectral feature vector; after preprocessing the acquired echo signal such as filtering and denoising, it extracts key indicators such as the time-frequency peak value and signal attenuation rate to form an acoustic feature vector characterizing the impedance characteristics of the internal structure; finally, the hyperspectral feature vector and the acoustic feature vector are concatenated to obtain a comprehensive multimodal feature vector for subsequent classifier to determine defects.

[0080] Beneficial Effects: This embodiment achieves the following significant benefits through the synergistic detection of spectroscopy and acoustics: Complementary Detection Dimensions: It simultaneously acquires information on the chemical composition (spectroscopy) and physical structure (acoustics) of the material, overcoming the limitations of single-modal detection; Enhanced Diagnostic Depth: It can effectively distinguish between different defects such as surface dampness and internal seepage, and micro-cracks and through-cracks, greatly improving the accuracy and reliability of defect identification and assessment; Laying the Foundation for Fusion Classification: The provided multimodal feature vectors contain richer defect information, providing a solid data foundation for subsequent high-precision classification and judgment.

[0081] In some embodiments, feature classification judgment is performed on the multimodal feature vector to determine the anti-seepage defect feature information, including: inputting the multimodal feature vector into a pre-trained multimodal fusion classifier; the multimodal fusion classifier performs at least one of the following judgments based on the input features: determining whether there are internal defects in the abnormal area, identifying the specific type of internal defects, and assessing the severity level of internal defects; wherein, the types of internal defects include capillary seepage, microcracks, through cracks, and internal cavities, and the severity level is classified according to the size, depth, and activity of the defects.

[0082] Technical Background and Working Principle: After obtaining the multimodal feature vectors fused from spectral and acoustic data, the core challenge is how to accurately and automatically determine the attributes of defects. Traditional single-modal classifiers or simple feature concatenation methods struggle to effectively handle differences in the distribution, dimensions, and importance of features across different modalities, potentially leading to decision bias. This embodiment introduces a multimodal fusion classifier based on an attention mechanism to simulate the cognitive process of human experts dynamically focusing on different cues (such as paying more attention to moisture features in spectral data or void responses in acoustic data) during judgment. Its working principle is that the attention mechanism automatically learns and assigns different weights to feature subsets from hyperspectral and acoustic modalities, highlighting the information most relevant to the current defect judgment during the model fusion stage and suppressing noise or unimportant features. This optimizes the classifier's decision boundary, achieving a refined and integrated judgment of defect existence, type, and severity level.

[0083] Technical Solution and Component Functions: The technical solution of this embodiment specifically includes: inputting the multimodal feature vector obtained in the aforementioned steps into a pre-trained multimodal fusion classifier; the classifier first processes the input vector through an attention mechanism module, which calculates a set of attention coefficients for the spectral feature sub-vector and the acoustic feature sub-vector respectively, and generates a weighted fusion feature vector by weighted summation; thereafter, the weighted feature is fed into a multilayer perceptron network, which is usually composed of several fully connected layers, and gradually abstracts features through nonlinear transformation; finally, in the output layer, the network outputs three key judgment results simultaneously: a binary flag indicating whether there is an internal defect in the abnormal region, a multi-class label indicating the specific type of defect (including capillary seepage, microcracks, through cracks and internal cavities), and a numerical or grade label assessing the severity level of the defect, which integrates the equivalent size, inferred depth and evolutionary activity of the defect.

[0084] Beneficial Effects: This embodiment achieves the following core benefits by introducing an attention-based multimodal fusion classifier: More accurate judgment: Adaptive weight allocation strengthens the role of key modal features and improves the ability to distinguish easily confused defects (such as surface wet spots and internal water seepage); More comprehensive output: The existence confirmation, type identification, and level assessment of defects are completed in one process, providing a complete state profile and simplifying the subsequent data processing flow; Enhanced decision interpretability: By analyzing the attention weights, the main feature sources on which the model's decisions are based can be traced, providing clues for result verification.

[0085] In some embodiments, labeled concrete sample data with known defect types and levels are collected. The labeled concrete sample data includes paired hyperspectral data and echo signal data. The hyperspectral feature vector and acoustic feature vector are extracted and fused to construct a training dataset. Using the training dataset, an attention-based multilayer perceptron network is trained end-to-end with cross-entropy as the loss function. The attention mechanism is used to adaptively assign different weights to features from different modalities during the fusion process, thereby optimizing the decision boundary of the classifier.

[0086] Technical Background and Working Principle: Effective fusion of multimodal data is key to improving defect classification accuracy. Traditional methods typically involve simple splicing or averaging of hyperspectral and acoustic features, neglecting the varying contributions of different modal features to the classification task. This is particularly relevant in concrete impermeability defect detection, where hyperspectral features primarily reflect material chemical composition (e.g., moisture content), while acoustic features characterize internal structural impedance (e.g., density). The importance of both in defect identification dynamically changes with defect type and environmental conditions. Attention mechanisms, by simulating human cognitive processes, can dynamically assess the importance of different modal features and assign adaptive weights during fusion. This allows the classifier to focus more on the feature modalities most relevant to the current defect type, thereby optimizing the decision boundary, improving classification performance, and reducing noise interference.

[0087] Technical Solution and Component Functions: The training process first collects labeled concrete sample data, including paired hyperspectral data (spectral cubes in the 400–1000 nm band obtained from scanning anomalous regions using a hyperspectral imaging sensor) and echo signal data (from transient physical excitation applied by a mechanical impact device and acquired by a high-precision microphone). Hyperspectral feature vectors are extracted using principal component analysis (PCA) to reduce dimensionality, retaining the top 10 principal components to capture spectral features related to moisture content and density. Acoustic feature vectors are extracted using time-frequency analysis to extract peak frequencies, attenuation rates, and other indicators to characterize internal structural impedance. The two feature vectors are concatenated to form a multimodal feature vector, constructing the training dataset. The attention-based multilayer perceptron network consists of an input layer, an attention layer, and a fully connected layer. The attention layer calculates the attention weights for the hyperspectral and acoustic subvectors, normalizes them using the softmax function, and then weights them for feature fusion. The fully connected layer comprises two hidden layers (128 and 64 neurons respectively) and an output layer, outputting the probability distribution of defect types (capillary seepage, microcracks, through-cracks, internal cavities) and their levels. Training uses a cross-entropy loss function and the Adam optimizer for end-to-end learning to minimize the error between the predictions and the true labels, iteratively optimizing the network parameters.

[0088] Beneficial effects: By introducing an attention mechanism, this training method can adaptively fuse multimodal features, significantly improving the classifier's accuracy in identifying concrete impermeability defects, while also enhancing the model's generalization ability in complex scenarios. The optimized decision boundary reduces misclassification, provides reliable input for subsequent defect evolution prediction, and mitigates performance degradation caused by feature redundancy or modal conflicts.

[0089] In some embodiments, an independent database entry is created for each anti-seepage defect feature information confirmed in step S2 as its spatiotemporal evolution archive; in each inspection cycle, when the corresponding defect area is accessed again, steps S1 and S2 are executed, and the newly acquired anti-seepage defect feature information of the area, together with its timestamp and environmental temperature and humidity data, is used as a new data recording point; the data recording point is appended and stored in the corresponding spatiotemporal evolution archive in chronological order to construct dynamic feature evolution information, and the detection data at discrete time points are linked into a dynamic sequence describing the changes of the physicochemical properties of the defect area over time.

[0090] Technical Background and Working Principle: The evolution of concrete impermeability defects is a typical spatiotemporal dynamic process. Traditional detection methods usually only record single detection results, which cannot effectively track the evolution trend and rate of defects. By establishing an independent spatiotemporal evolution profile for each defect, and organizing and storing discrete, periodic detection data (including multimodal feature vectors and environmental parameters) according to time series, a complete trajectory of defect state changes over time can be constructed. This time-series data organization method upgrades static snapshot detection to dynamic process monitoring, providing an indispensable data foundation for subsequent learning of defect evolution patterns and life prediction based on time series analysis, thereby enabling a forward-looking assessment of the health status of concrete structures.

[0091] Technical Solution and Component Functions: The system assigns a unique identifier (Defect ID) to each confirmed defect and creates an independent entry in the spatiotemporal evolution archive database. The database adopts a relational table structure, with the Defect ID as the primary key and fields including feature vectors (storing hyperspectral and acoustic fusion features), timestamps (recording the precise detection time), and ambient temperature and relative humidity (synchronously collected by integrated sensors). In each inspection cycle, the robot revisits the known defect location, accurately locates it using compound eye vision sensors, and triggers the hyperspectral imaging and acoustic excitation detection process to acquire new anti-permeability defect feature information. The system automatically packages the new data with the current timestamp and environmental parameters into a single data record point and appends it to the corresponding defect's archive through transactional write operations, ensuring the temporal continuity and integrity of the data, ultimately constructing a dynamic feature evolution information sequence indexed by time.

[0092] Beneficial effects: By establishing and maintaining spatiotemporal evolution archives, continuous and systematic tracking of the entire life cycle of each concrete impermeability defect is achieved, transforming isolated detection points into dynamic sequences rich in temporal correlation information. This not only provides data support for accurately analyzing defect evolution patterns (such as acceleration, stabilization, or decay), but also enables machine learning-based prediction models to learn reliable patterns from historical data, significantly improving the accuracy and reliability of impermeability life prediction.

[0093] In some embodiments, the dynamic feature evolution information of each defect region is constructed as regular time series data and used as input to a long short-term memory network model. The long short-term memory network model learns and memorizes the long-term dependencies and short-term fluctuation patterns of the dynamic feature evolution information in the time dimension through its internal gating mechanism. The training objective of the long short-term memory network model is to minimize the error between its prediction of the future value of the defect features and the actual measurement value in subsequent inspections. Based on the trained long short-term memory network model, the key state parameters of the defect region at the next M time points are recursively predicted using the dynamic feature evolution information of the previous N time points as input. The key state parameters include at least: the equivalent size of the defect, the average water content of the region, and the defect activity level. By comparing the key state parameters with the corresponding preset critical thresholds, the expected time point when the corresponding defect region reaches a dangerous state is determined, which is the local impermeability life of the current defect region.

[0094] Technical Background and Working Principle: The evolution of concrete defects is a typical nonlinear temporal process, influenced by multiple factors such as material aging and environmental loads, exhibiting complex memory effects and delayed responses. Traditional statistical models or simple regression methods struggle to effectively capture this long-term dependency. Long Short-Term Memory (LSTM) networks, as a special type of recurrent neural network, selectively memorize important historical information and forget redundant content through their sophisticated gating mechanisms (input gate, forget gate, output gate), thus simultaneously learning both long-term trends and short-term fluctuations in the sequence. This makes LSTM particularly suitable for extracting evolutionary patterns from the spatiotemporal evolution archives of defects and for making multi-step predictions of key future state parameters, providing a quantitative basis for life assessment.

[0095] Technical Solution and Component Functions: The prediction system first arranges the data record points (including multimodal feature vectors and environmental parameters) in the spatiotemporal evolution archive of each defect in chronological order to construct a regular time series dataset. After standardization, this sequence is used as the input to the LSTM model. The LSTM model adopts a two-layer stacked structure, with each layer containing 64 memory units, and the input time step N is set to 6 (i.e., using the data from the 6 most recent inspections). Model training uses the features of the previous N time steps as input and the actual measurement value of the subsequent M-th (M=3) time step as the objective, employing a mean squared error loss function and the Adam optimizer for supervised learning. During prediction, the real data from the previous N time points are used as the initial input to generate the first prediction; then, this prediction value is fed back to the network to recursively generate the prediction sequence for the next M steps, obtaining the evolution trajectory of the defect's equivalent size, regional average water content, and defect activity level; finally, linear interpolation is used to determine the time point when each key parameter first exceeds the preset danger threshold, and the earliest time is taken as the local impermeability lifetime of the defect.

[0096] Beneficial effects: Using LSTM models for defect evolution prediction can accurately capture the temporal dynamic characteristics of concrete defect development, enabling quantitative prediction of local impermeability life. This method transforms experience-based qualitative judgment into data-driven quantitative prediction, providing a crucial time window for preventative maintenance and significantly improving the predictability and scientific rigor of concrete structure health management.

[0097] In some embodiments, for each pixel on the surface of the concrete sample column, a local impermeability life value is assigned according to the defect area to which it belongs; all local impermeability life values ​​are mapped to a color space, where different colors or shades of color represent different life ranges, thereby forming a continuous color mapping map on the entire surface of the concrete sample column, i.e., an impermeability life prediction heatmap; by combining the prediction results of all defect areas, several key indicators are extracted, including: the shortest local impermeability life, the number of high-risk defect areas, and the average evolution acceleration of defects; a risk rating matrix is ​​established, and the values ​​of the key indicators are mapped to preset risk level ranges, generating an overall risk rating report that includes the final risk level, the location of the main risk sources, and predictive maintenance recommendations.

[0098] Technical Background and Working Principle: Traditional concrete permeability assessment relies primarily on discrete test data and the experience-based judgment of professionals, lacking intuitive visualization and a systematic risk assessment framework. The permeability life prediction heatmap transforms numerical life prediction results into a continuous color distribution, achieving spatial visualization of defect development trends and enabling rapid identification and location of high-risk areas. Simultaneously, the overall risk rating report integrates multiple key indicators to construct a multi-dimensional assessment system from local to global perspectives, transforming complex prediction data into actionable decision support information. This combination of visualization and quantitative assessment significantly improves the efficiency and scientific rigor of concrete structure health management.

[0099] Technical Solution and Component Functions: The system first assigns a corresponding lifespan value to each pixel on the surface of the concrete sample column based on the predicted local impermeability lifespan of each defect area using a spatial interpolation algorithm. Then, it uses the HSV color space for color mapping, where hue linearly changes from 0° (red, representing a lifespan ≤ 1 year) to 120° (green, representing a lifespan ≥ 10 years), forming a heatmap for impermeability lifespan prediction. Simultaneously, the system automatically calculates three key indicators: shortest local impermeability lifespan (the minimum lifespan of all pixels), number of high-risk defect areas (the number of defect areas with a lifespan less than 2 years), and average defect evolution acceleration (calculated based on the lifespan change rate of the most recent three inspection cycles). These indicators are input into a preset 3×3 risk rating matrix, where rows correspond to indicator levels (high, medium, low), and columns correspond to risk levels (I, II, III). The final risk level is determined through cross-mapping, and an overall risk rating report containing the coordinates of major risk sources and targeted maintenance recommendations is automatically generated.

[0100] Beneficial effects: By generating a heat map predicting the impermeability life and an overall risk rating report, this method enables an intuitive visualization of the health status of concrete structures and a systematic risk assessment. It transforms complex prediction data into easily understandable graphics and hierarchical information, significantly improving the judgment efficiency of on-site engineers, providing a scientific basis for preventative maintenance decisions, and achieving a leap from local defect analysis to overall structural assessment.

[0101] In some embodiments, trend analysis is performed on the time-series data stored in the spatiotemporal evolution archive; each type of parameter in the anti-seepage defect characteristic information is quantified, including moisture index, defect activity index, first derivative or moving average over time; based on the results of the trend analysis, the next inspection plan for the defect area is dynamically adjusted; for active defects whose characteristic change rate exceeds the threshold, their inspection cycle is automatically shortened; for stable defects, their inspection cycle is extended, thereby achieving adaptive optimization of inspection resources.

[0102] Technical Background and Working Principle: In modern building facility management, Building Information Modeling (BIM) has become the core platform for carrying digital information about physical assets. However, traditional BIM models mainly contain geometric and basic attribute information, lacking the ability to dynamically perceive and predict structural health status. Integrating impermeability life data obtained based on multimodal detection and artificial intelligence prediction into BIM achieves seamless fusion of physical world detection data and virtual models. By establishing coordinate mapping relationships, two-dimensional predictive heat maps are accurately applied to the three-dimensional component surfaces, and risk assessment results are written into the model attributes. This upgrades the BIM model from a static "as-built archive" to a dynamic "health monitoring dashboard," providing key technical support for the full lifecycle management of infrastructure based on digital twins.

[0103] Technical Solution and Component Functions: The system first establishes an accurate spatial mapping table by calibrating the coordinate correspondence between feature points on the surface of concrete sample columns and the corresponding component surfaces in the BIM model. Next, it outputs the heat map of the predicted seepage life as a GeoTIFF format texture image with spatial coordinate reference, while simultaneously converting the overall risk rating report into structured data in JSON format. This data is then uploaded to the model database via the BIM platform's RESTful API (such as the Autodesk Forge API or BIM 360 API). In the BIM modeling environment (such as Revit), a dedicated plugin is developed to read the uploaded data, assigning the heat map texture as a diffuse map to the material properties of the corresponding concrete component, achieving 3D visualization of the prediction results. Simultaneously, the plugin writes key indicators from the report, such as the final risk level and the shortest local seepage life, into the component's shared parameter set. Based on these parameter values, different color overlays are set through view filters (e.g., red for high risk, yellow for medium risk), automatically highlighting areas of different risk levels in the BIM 3D view, ultimately generating a digital twin model that integrates the prediction results.

[0104] Beneficial effects: By integrating the predicted impermeability lifespan into the Building Information Model (BIM), the health status of concrete structures has been transformed from two-dimensional planar analysis to three-dimensional spatial positioning, enabling seamless integration of detection, prediction, and management decisions within a unified platform. This integration not only significantly improves the informatization level of facility management but also provides intuitive and reliable decision-making basis for preventive maintenance, resource optimization, and life-cycle cost control based on digital twins.

[0105] Figure 3 This is a structural schematic diagram of an intelligent inspection system for concrete impermeability provided in an embodiment of this application, as shown below. Figure 3 As shown, the concrete impermeability intelligent inspection system 300 of this embodiment includes: an anomaly screening module 301, a defect analysis module 302, an evolution analysis module 303, and a report output module 304.

[0106] Anomaly screening module 301 is used to acquire a global low-resolution image of the concrete sample column group, analyze the global low-resolution image, and determine the set of coordinate points of the abnormal area where macroscopic anomalies exist.

[0107] The defect analysis module 302 is used to extract spectral features and excite sound signals for each abnormal region based on the set of coordinate points of the abnormal region, and determine the anti-permeability defect feature information corresponding to each abnormal region.

[0108] The evolution analysis module 303 is used to establish an independent spatiotemporal evolution archive for each abnormal area based on the anti-seepage defect feature information. During periodic inspections, steps S1 and S2 are repeated to store the anti-seepage defect feature information at different time nodes into the corresponding spatiotemporal evolution archive, thereby constructing dynamic feature evolution information for each defect area.

[0109] The report output module 304 is used to learn the evolution law of each defect area by using a long short-term memory network model based on the dynamic feature evolution information of each defect area, predict the future development trend of the defect area, and generate a heat map of the impermeability life prediction of the concrete sample column and an overall risk rating report.

[0110] Optionally, the anomaly filtering module 301 is specifically used for:

[0111] The global low-resolution image is acquired using a bionic compound eye vision sensor, wherein the bionic compound eye vision sensor is composed of an array of multiple micro-eye optical units, each micro-eye optical unit pointing in a different direction, thereby achieving one-time imaging of a group of concrete sample columns within a large field of view.

[0112] The global low-resolution image is input into a preset lightweight convolutional neural network model for analysis. The preset lightweight convolutional neural network model is trained to identify macroscopic abnormal patterns related to leakage from complex background textures, including color abnormal areas, irregular wet spot contours, and surface aggregated bubble areas.

[0113] The preset lightweight convolutional neural network model outputs the center point coordinates or bounding box coordinates of all identified abnormal regions, forming the set of coordinate points of the abnormal regions.

[0114] Optionally, the defect analysis module 302 is specifically used for:

[0115] The abnormal region is scanned using a hyperspectral imaging sensor to obtain its continuous spectral cubic data in the visible to near-infrared bands.

[0116] A transient physical excitation is applied to the abnormal area by a mechanical striking device, and the echo signal generated by the transient physical excitation is collected by a high-precision microphone.

[0117] The continuous spectral cubic data is subjected to dimensionality reduction and feature extraction to quantify the ratio of depth to reflectance of specific absorption bands that are closely related to the moisture content and density of concrete, thereby forming a hyperspectral feature vector.

[0118] The echo signal is preprocessed and its features are extracted, and it is converted into an acoustic feature vector characterizing the impedance characteristics of the internal structure.

[0119] The hyperspectral feature vector is concatenated with the acoustic feature vector to obtain a multimodal feature vector, and the multimodal feature vector is classified and judged to determine the anti-permeability defect feature information.

[0120] Optionally, when the defect analysis module 302 performs feature classification on the multimodal feature vector to determine the impermeability defect feature information, it is specifically used for:

[0121] The multimodal feature vectors are input into a pre-trained multimodal fusion classifier;

[0122] The multimodal fusion classifier performs at least one of the following judgments based on the input features: determining whether the abnormal region has internal defects, identifying the specific type of the internal defects, and assessing the severity level of the internal defects;

[0123] The types of internal defects include capillary water seepage, microcracks, through cracks, and internal cavities, and the severity level is classified according to the size, depth, and activity of the defects.

[0124] Optionally, in the defect analysis module 302, the training process of the multimodal fusion classifier is specifically used for:

[0125] Collect labeled concrete sample data with known defect types and grades, the labeled concrete sample data including paired hyperspectral data and echo signal data;

[0126] The hyperspectral and acoustic feature vectors are extracted separately and then fused to construct a training dataset.

[0127] Using the training dataset, an end-to-end multilayer perceptron network based on an attention mechanism is trained with cross-entropy as the loss function. The attention mechanism is used to adaptively assign different weights to features from different modalities during the fusion process, thereby optimizing the decision boundary of the classifier.

[0128] Optionally, the evolution analysis module 303 is specifically used for:

[0129] Create an independent database entry as its spatiotemporal evolution archive for each of the anti-permeability defect feature information confirmed in step S2;

[0130] In each inspection cycle, when the corresponding defect area is visited again, steps S1 and S2 are executed, and the newly acquired anti-seepage defect feature information of the area, together with its timestamp and environmental temperature and humidity data, is used as a new data record point.

[0131] The data recording points are appended and stored to the corresponding spatiotemporal evolution archive in chronological order to construct the dynamic feature evolution information, and the detection data at discrete time points are linked together to form a dynamic sequence describing the changes in the physicochemical properties of the defect region over time.

[0132] Optionally, the report output module 304 is specifically used for:

[0133] The dynamic feature evolution information of each defect region is constructed into regular time series data, which is used as the input of the long short-term memory network model;

[0134] The Long Short-Term Memory Network model learns and memorizes the long-term dependencies and short-term fluctuation patterns of the dynamic feature evolution information over time through its internal gating mechanism.

[0135] The training objective of the Long Short-Term Memory Network model is to minimize the error between its prediction of future values ​​of defect features and the actual measured values ​​in subsequent inspections.

[0136] Based on the trained Long Short-Term Memory network model, and using the dynamic feature evolution information from the previous N time points as input, the key state parameters of the defective region at the next M time points are recursively predicted.

[0137] The key state parameters include at least: the equivalent size of the defect, the average moisture content of the region, and the defect activity level;

[0138] By comparing the key state parameters with the corresponding preset critical thresholds, the expected time point at which the corresponding defect area reaches a dangerous state is determined, which is the local impermeability life of the current defect area.

[0139] Optionally, the report output module 304, when generating the heat map for predicting the impermeability life of the concrete sample column and the overall risk rating report, is specifically used for:

[0140] For each pixel on the surface of the concrete sample column, the local impermeability life value is assigned according to the defect area to which it belongs.

[0141] All the local impermeability life values ​​are mapped to a color space, where different colors or shades of color represent different life ranges, thus forming a continuous color mapping map on the surface of the entire concrete sample column, namely the impermeability life prediction heat map.

[0142] Based on the prediction results of all defect areas, several key indicators were extracted, including: the shortest local impermeability life, the number of high-risk defect areas, and the average evolution acceleration of defects.

[0143] A risk rating matrix is ​​established, mapping the values ​​of the key indicators to a preset risk level range, and generating an overall risk rating report that includes the final risk level, the location of the main risk sources, and predictive maintenance recommendations.

[0144] Optionally, the system 300 further includes an inspection optimization module 305, specifically used for:

[0145] Perform trend analysis on the time-series data stored in the spatiotemporal evolution archive;

[0146] Quantify each type of parameter in the impermeability defect characteristic information, including moisture index, defect activity index, first derivative or moving average over time;

[0147] Based on the results of the trend analysis, the next inspection plan for the defect area is dynamically adjusted. For active defects with a feature change rate exceeding the threshold, the inspection cycle is automatically shortened, while for stable defects, the inspection cycle is extended, thus achieving adaptive optimization of inspection resources.

[0148] The system in this embodiment can be used to execute the methods of any of the above embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

Claims

1. A method for intelligent inspection of concrete impermeability, characterized in that, Periodic inspection robots used for permeability testing of batches of concrete sample columns include: S1. Obtain a global low-resolution image of the concrete sample column group, analyze the global low-resolution image, and determine the set of coordinate points of the abnormal area where macroscopic anomalies exist. S2. Based on the set of coordinate points of the abnormal area, perform spectral feature extraction and sound signal excitation on each abnormal area to determine the anti-permeability defect feature information corresponding to each abnormal area. S3. Based on the anti-seepage defect feature information, establish an independent spatiotemporal evolution archive for the anti-seepage defect corresponding to each abnormal area. During periodic inspections, repeat steps S1 and S2 to store the anti-seepage defect feature information at different time nodes into the corresponding spatiotemporal evolution archive, thereby constructing dynamic feature evolution information for each defect area. S4. Based on the dynamic feature evolution information of each defect area, use the long short-term memory network model to learn its evolution law, predict the future development trend of the defect area, and generate a heat map of the anti-seepage life prediction of the concrete sample column and an overall risk rating report. The step of extracting spectral features and exciting acoustic signals for each anomalous region to determine the anti-permeability defect feature information corresponding to each anomalous region includes: The abnormal region is scanned using a hyperspectral imaging sensor to obtain its continuous spectral cubic data in the visible to near-infrared bands. A transient physical excitation is applied to the abnormal area by a mechanical striking device, and the echo signal generated by the transient physical excitation is collected by a high-precision microphone. The continuous spectral cubic data is subjected to dimensionality reduction and feature extraction to quantify the ratio of depth to reflectance of specific absorption bands that are closely related to the moisture content and density of concrete, thereby forming a hyperspectral feature vector. The echo signal is preprocessed and its features are extracted, and it is converted into an acoustic feature vector characterizing the impedance characteristics of the internal structure. The hyperspectral feature vector is concatenated with the acoustic feature vector to obtain a multimodal feature vector, and the multimodal feature vector is classified and judged to determine the anti-permeability defect feature information.

2. The method according to claim 1, characterized in that, The process of acquiring a global low-resolution image of the concrete sample column group, analyzing the global low-resolution image, and determining the set of coordinate points of the abnormal areas where macroscopic anomalies exist includes: The global low-resolution image is acquired using a bionic compound eye vision sensor, wherein the bionic compound eye vision sensor is composed of an array of multiple micro-eye optical units, each micro-eye optical unit pointing in a different direction, thereby achieving one-time imaging of a group of concrete sample columns within a large field of view. The global low-resolution image is input into a preset lightweight convolutional neural network model for analysis. The preset lightweight convolutional neural network model is trained to identify macroscopic abnormal patterns related to leakage from complex background textures, including color abnormal areas, irregular wet spot contours, and surface aggregated bubble areas. The preset lightweight convolutional neural network model outputs the center point coordinates or bounding box coordinates of all identified abnormal regions, forming the set of coordinate points of the abnormal regions.

3. The method according to claim 2, characterized in that, The step of performing feature classification and judgment on the multimodal feature vector to determine the impermeability defect feature information includes: The multimodal feature vectors are input into a pre-trained multimodal fusion classifier; The multimodal fusion classifier performs at least one of the following judgments based on the input features: determining whether the abnormal region has internal defects, identifying the specific type of the internal defects, and assessing the severity level of the internal defects; The types of internal defects include capillary water seepage, microcracks, through cracks, and internal cavities, and the severity level is classified according to the size, depth, and activity of the defects.

4. The method according to claim 3, characterized in that, The training process of the multimodal fusion classifier includes: Collect labeled concrete sample data with known defect types and grades, the labeled concrete sample data including paired hyperspectral data and echo signal data; The hyperspectral and acoustic feature vectors are extracted separately and then fused to construct a training dataset. Using the training dataset, an end-to-end multilayer perceptron network based on an attention mechanism is trained with cross-entropy as the loss function. The attention mechanism is used to adaptively assign different weights to features from different modalities during the fusion process, thereby optimizing the decision boundary of the classifier.

5. The method according to claim 3, characterized in that, The dynamic feature evolution information for constructing each defect region includes: Create an independent database entry as its spatiotemporal evolution archive for each of the anti-permeability defect feature information confirmed in step S2; In each inspection cycle, when the corresponding defect area is visited again, steps S1 and S2 are executed, and the newly acquired anti-seepage defect feature information of the area, together with its timestamp and environmental temperature and humidity data, is used as a new data record point. The data recording points are appended and stored to the corresponding spatiotemporal evolution archive in chronological order to construct the dynamic feature evolution information, and the detection data at discrete time points are linked together to form a dynamic sequence describing the changes in the physicochemical properties of the defect region over time.

6. The method according to claim 5, characterized in that, The method of learning the evolution pattern of each defective region based on its dynamic feature evolution information and using a long short-term memory network model to predict its future development trend includes: The dynamic feature evolution information of each defect region is constructed into regular time series data, which is used as the input of the long short-term memory network model; The Long Short-Term Memory Network model learns and memorizes the long-term dependencies and short-term fluctuation patterns of the dynamic feature evolution information over time through its internal gating mechanism. The training objective of the Long Short-Term Memory Network model is to minimize the error between its prediction of future values ​​of defect features and the actual measured values ​​in subsequent inspections. Based on the trained Long Short-Term Memory network model, and using the dynamic feature evolution information from the previous N time points as input, the key state parameters of the defective region at the next M time points are recursively predicted. The key state parameters include at least: the equivalent size of the defect, the average moisture content of the region, and the defect activity level; By comparing the key state parameters with the corresponding preset critical thresholds, the expected time point at which the corresponding defect area reaches a dangerous state is determined, which is the local impermeability life of the current defect area.

7. The method according to claim 6, characterized in that, The heat map predicting the impermeability life of the generated concrete sample column and the overall risk rating report include: For each pixel on the surface of the concrete sample column, the local impermeability life value is assigned according to the defect area to which it belongs. All the local impermeability life values ​​are mapped to a color space, where different colors or shades of color represent different life ranges, thus forming a continuous color mapping map on the surface of the entire concrete sample column, namely the impermeability life prediction heat map. Based on the prediction results of all defect areas, several key indicators were extracted, including: the shortest local impermeability life, the number of high-risk defect areas, and the average evolution acceleration of defects. A risk rating matrix is ​​established, mapping the values ​​of the key indicators to a preset risk level range, and generating an overall risk rating report that includes the final risk level, the location of the main risk sources, and predictive maintenance recommendations.

8. The method according to claim 7, characterized in that, The method further includes: Perform trend analysis on the time-series data stored in the spatiotemporal evolution archive; Quantify each type of parameter in the impermeability defect characteristic information, including moisture index, defect activity index, and first derivative or moving average over time; Based on the results of the trend analysis, the next inspection plan for the defect area is dynamically adjusted. For active defects with a feature change rate exceeding the threshold, the inspection cycle is automatically shortened, while for stable defects, the inspection cycle is extended, thus achieving adaptive optimization of inspection resources.

9. A concrete impermeability inspection system, characterized in that, The method applied to any one of claims 1-8 includes: An anomaly screening module is used to acquire a global low-resolution image of a group of concrete sample columns, analyze the global low-resolution image, and determine the set of coordinate points of anomaly areas where macroscopic anomalies exist. The defect analysis module is used to extract spectral features and excite sound signals for each abnormal region based on the set of coordinate points of the abnormal region, and determine the anti-permeability defect feature information corresponding to each abnormal region. The evolution analysis module is used to establish an independent spatiotemporal evolution archive for each abnormal area based on the anti-seepage defect feature information. During periodic inspections, steps S1 and S2 are repeated to store the anti-seepage defect feature information at different time points into the corresponding spatiotemporal evolution archive, thereby constructing dynamic feature evolution information for each defect area. The report output module is used to learn the evolution law of each defect area based on the dynamic feature evolution information of each defect area using a long short-term memory network model, predict the future development trend of the defect area, and generate a heat map of the impermeability life prediction of the concrete sample column and an overall risk rating report.

Images