Building outer wall hollowing identification method and system based on pulsed eddy current thermal imaging

Abstract

The invention provides a building outer wall hollowing identification method and system based on pulsed eddy current thermal imaging, and the method comprises the steps: enabling a pulsed eddy current excitation device to act on the surface of a building outer wall, applying transient pulsed magnetic field excitation to a metal component in a wall body, and enabling the metal component to sense eddy current and serve as an internal heat source to generate Joule heat; synchronously acquiring the transient pulsed magnetic field excitation by using an infrared thermal imager to obtain a heat map sequence formed by internal heat conduction on the surface of the outer wall of the building, and preprocessing the heat map sequence; and on the basis of the preprocessed heat map sequence, solving a heat conduction inverse problem to reconstruct a wall internal thermal attribute feature map, and on the basis of the wall internal thermal attribute feature map, completing the identification of the building external wall hollowing. According to the invention, deep, accurate, quantitative and automatic detection of building outer wall hollowing is realized, and reliability, efficiency and applicability are obviously better than those of the prior art.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing technology for building quality, specifically relating to a method and system for identifying hollow areas in building exterior walls based on pulsed eddy current thermal imaging. Background Technology

[0002] Hollow areas in the cladding or plaster layers of building exterior walls are a common quality defect, posing a risk of detachment and seriously threatening people's lives and property. Traditional detection methods mainly rely on manual tapping, which is inefficient, highly subjective, and heavily dependent on the experience of the inspectors. Existing infrared thermal imaging methods typically use sunlight or hot air guns as heat sources, identifying hollow areas by observing surface temperature differences. However, these methods are greatly affected by environmental factors (sunlight, weather), and heat is only transferred from the surface, limiting their ability to detect deep or small-sized hollow areas. Pulsed eddy current thermal imaging technology has been used for defect detection in metallic materials, but its application in non-metallic, non-conductive building materials (such as ceramic tiles and cement mortar) faces fundamental challenges because these materials are inherently non-conductive and cannot directly induce eddy currents. This invention, through innovative exciter design and detection principles, successfully applies this technology to the inspection of building exterior walls, overcoming the bottlenecks of existing technologies. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide an automated method and system for identifying hollow areas in building exterior walls that is unaffected by ambient light and weather, has a large detection depth, high accuracy, and can achieve rapid scanning.

[0004] To achieve the above objectives, the present invention provides the following solution: A method for identifying hollow areas in building exterior walls based on pulsed eddy current thermal imaging includes: The pulsed eddy current excitation device is applied to the surface of the building's exterior wall, and a transient pulsed magnetic field excitation is applied to the metal components inside the wall, causing the metal components to induce eddy currents and generate Joule heat as an internal heat source. The transient pulse magnetic field excitation is synchronously acquired using an infrared thermal imager to obtain a thermal image sequence formed on the surface of the building exterior wall due to internal heat conduction, and the thermal image sequence is preprocessed. Based on the preprocessed heat map sequence, the internal thermal property feature map of the wall is reconstructed by solving the inverse heat conduction problem. Based on the internal thermal property feature map of the wall, the hollowness of the building exterior wall is identified.

[0005] Preferably, the method for preprocessing the heatmap sequence includes: Inter-frame registration is performed on the heat map sequence, and the original radiation data is converted into absolute temperature values ​​according to preset emissivity, ambient temperature and ambient humidity parameters to obtain a standardized heat map sequence. A fast Fourier transform is performed on the temperature-time curve of each pixel in the standardized heatmap sequence to convert it from the time domain to the frequency domain, generating a series of phase data cubes with different frequency components. The phase data cube is reconstructed into a two-dimensional matrix, and principal component analysis is performed on the two-dimensional matrix to obtain a spatial score map; Based on the spatial score map, background thermal field stripping is performed, and a known intact region is introduced as a reference for spatiotemporal difference processing to obtain a difference sequence. The difference sequence is concatenated with the phase data cube to reconstruct a new multimodal feature sequence, thus completing the preprocessing of the heatmap sequence.

[0006] Preferred methods for identifying hollow areas in building exterior walls include: Based on the partial differential equation of heat conduction, the problem of detecting hollow areas is transformed into the inverse problem of heat conduction for reconstructing the characteristic map of the internal thermal properties of the wall. Based on the preprocessed heat map sequence, the reconstructed thermal property feature map of the wall interior is obtained using an end-to-end inversion network based on deep learning. Pixel areas in the thermal property feature map of the wall interior that are below a preset statistical threshold are identified as hollow areas, thus obtaining a hollow area distribution map of the building exterior wall.

[0007] Preferably, the method for obtaining the characteristic map of the reconstructed internal thermal properties of the wall includes: A physical constraint network is constructed by using the U-Net++ architecture with dense jump connections as the backbone and embedding a physical activation layer with a built-in heat conduction positive problem solver at the end. The multimodal feature sequence is used as the input of the physical constraint network. A composite loss function that integrates data loss, physical residual loss and sparsity loss is used to perform multi-task physical-driven optimization on the physical constraint network to obtain the optimized physical constraint network. Using existing parametric hollow wall finite element simulation data, a two-stage transfer training was performed on the optimized physical constraint network to obtain a hollow recognition model. Based on the hollow wall identification model, a feature map of the internal thermal properties of the wall is obtained.

[0008] Preferably, in the physical constraint network, the U-Net++ architecture with dense skip connections is used to simultaneously capture full-scale features from micro-texture to macro-structure; The physical activation layer is used to take the thermal property feature distribution initially predicted by the network as input, and combined with known boundary conditions, to simulate the theoretical thermal response sequence that the wall surface should produce under the initially predicted thermal property feature distribution using a differential operator constructed based on the finite difference method. The theoretical thermal response sequence is then compared with the actual collected and preprocessed real thermal image sequence to obtain the physical residual. The known boundary conditions include the location of the pulse heat source.

[0009] Preferred methods for obtaining a distribution map of hollow areas in building exterior walls include: An adaptive threshold algorithm based on statistical process control is used to initially segment the feature map of thermal properties inside the wall, thereby enabling robust extraction of abnormal regions in different walls. Mathematical morphology methods are used to post-process robust anomaly regions of different walls, and noise is filtered out and the geometry and connectivity of hollow regions are optimized by opening and closing operations. Quantitative feature extraction and severity analysis are performed on the optimized hollow connected regions, and the area and location of the hollow connected regions are calculated. The hollow connected regions are classified according to the degree of deviation of the thermal attribute values ​​of the hollow connected regions, and quantitative analysis results are obtained. The quantitative analysis results are fused with the pre-acquired visible light image to obtain a distribution map of building exterior wall hollowness that includes coordinates, area, and level.

[0010] The present invention also provides a building exterior wall hollowness identification system based on pulsed eddy current thermal imaging, for implementing the method, comprising: The pulsed eddy current excitation module is used to apply the pulsed eddy current excitation device to the surface of the building exterior wall, and apply transient pulsed magnetic field excitation to the metal components inside the wall, so that the metal components induce eddy currents and generate Joule heat as an internal heat source. The thermal image sequence acquisition module is used to synchronously acquire the transient pulse magnetic field excitation using an infrared thermal imager to obtain a thermal image sequence formed on the surface of the building exterior wall due to internal heat conduction, and to preprocess the thermal image sequence. The hollow wall identification module is used to reconstruct the internal thermal property feature map of the wall by solving the inverse heat conduction problem based on the preprocessed heat map sequence, and to identify the hollow wall of the building exterior based on the internal thermal property feature map.

[0011] Preferably, the heat map sequence acquisition module includes: The standardized unit is used to perform inter-frame registration of the heat map sequence and convert the original radiation data into absolute temperature values ​​according to preset emissivity, ambient temperature and ambient humidity parameters to obtain a standardized heat map sequence. The phase data acquisition unit is used to perform a fast Fourier transform on the temperature-time curve of each pixel in the standardized heat map sequence, converting it from the time domain to the frequency domain, and generating a series of phase data cubes with different frequency components. The principal component analysis unit is used to reconstruct the phase data cube into a two-dimensional matrix and perform principal component analysis on the two-dimensional matrix to obtain a spatial score map. The differential processing unit is used to perform background thermal field stripping based on the spatial score map, and introduce a known intact region as a reference for spatiotemporal differential processing to obtain a differential sequence. The sequence reconstruction unit is used to concatenate the difference sequence with the phase data cube to reconstruct a new multimodal feature sequence, thereby completing the preprocessing of the heatmap sequence.

[0012] Compared with existing technologies, the beneficial effects of this invention are as follows: Compared with traditional external heating methods (such as hot air guns and sunlight), this invention generates an internal heat source by stimulating internal metal components, and the heat wave is conducted from the inside to the outside. This has a unique advantage in detecting deep voids between the decorative layer and the structural layer, solving the problem that traditional methods are insensitive to deep defects. Based on the inverse problem of heat conduction, the internal thermal property feature map is reconstructed, realizing a leap from "surface temperature observation" to "internal structure inversion". It can accurately locate the position, shape and size of voids, and even assess their severity, far exceeding the accuracy of manual tapping and conventional infrared thermography. Attached Figure Description

[0013] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of a method for identifying hollow areas in building exterior walls based on pulsed eddy current thermal imaging, according to an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Example 1: like Figure 1 As shown, a method for identifying hollow areas in building exterior walls based on pulsed eddy current thermal imaging includes: S1: The pulsed eddy current excitation device is applied to the surface of the building's exterior wall, applying a transient pulsed magnetic field excitation to the metal components inside the wall. This induces eddy currents in the metal components, causing them to generate Joule heat as an internal heat source. Specifically, the electromagnetic-thermal energy conversion chain designed in this embodiment actively transforms the inherent non-detection target metal components inside the building's exterior wall into a controllable transient internal heat source located near the defect, thereby fundamentally changing the external-to-inside thermal excitation path and detection mechanism in traditional infrared thermal imaging detection. This process begins with the optimized configuration of the excitation device and its parameters. The pulsed eddy current excitation device—typically composed of a high-frequency signal generator, a power amplifier, and a specifically designed excitation coil (such as using an E-type magnetic core structure to enhance the magnetic field converging effect)—is placed parallel to or in slight contact with the surface of the exterior wall to be tested. A rigorously optimized single short-duration electrical pulse is applied to the coil, with its key parameters, especially the frequency, limited to the range of 100kHz to 1MHz, and the pulse width adjustable between 0.1ms and 1000ms. The choice of this high-frequency band is the key to achieving efficient energy conversion: it ensures that the alternating magnetic field can effectively penetrate the non-conductive building finish layer, while inducing sufficiently strong eddy currents on the surface of the metal components based on the skin effect principle. Simultaneously, it avoids the problem of excessive heat concentration on the metal surface due to excessively high frequencies, which would hinder subsequent heat conduction. When this transient pulsed magnetic field penetrates the wall finish layer and acts on the internal metal components (such as plaster wire mesh, insulation anchors, or concrete reinforcement), it induces strong closed eddy currents in the metal components based on Faraday's law of electromagnetic induction. As these eddy currents flow through metal materials with a certain resistivity, they convert the electromagnetic energy carried by the pulsed magnetic field into Joule heat almost instantaneously (on the order of milliseconds to seconds) according to the Joule-Lenz law. This energy conversion process causes the temperature of the metal components to rise rapidly in a very short time, thus creatively shaping it into a transient, "embedded" planar heat source located inside the wall. Ultimately, the directly heated metal component acts as an ideal heat wave emitter, generating a transient heat wave that diffuses into the surrounding wall materials (cement mortar, brickwork, insulation layer, etc.). This heat wave propagates outward primarily through conduction, and its conduction path is modulated by the internal structure of the wall. Specifically, in the hollow defect area, the air gap between the metal component and the exterior wall finish layer constitutes a high thermal resistance layer, significantly hindering heat flow upward to the outer surface. This results in a spatiotemporally specific surface temperature field distribution in the subsequent infrared thermal image sequence, distinct from the intact area. This provides a clear physical mechanism and extremely high signal-to-noise ratio for subsequent high-precision hollow identification based on the inverse problem of heat conduction.

[0018] S2: Use an infrared thermal imager to synchronously acquire transient pulse magnetic field excitation, obtain a thermal image sequence formed on the surface of the building exterior wall due to internal heat conduction, and preprocess the thermal image sequence.

[0019] A further implementation method for preprocessing the heatmap sequence includes: Inter-frame registration is performed on the thermal image sequence, and the original radiance data is converted into absolute temperature values ​​according to preset emissivity, ambient temperature, and ambient humidity parameters to obtain a standardized thermal image sequence. Specifically, high-precision inter-frame registration is performed on the original thermal image sequence to eliminate pixel-level spatial drift caused by minor vibrations of equipment or the environment. Subsequently, based on the precisely measured material emissivity, ambient temperature, and humidity parameters on site, the original radiance data captured by the camera is rigorously converted into absolute temperature values ​​in Kelvin to generate a standardized thermal image sequence. This step is a crucial leap from qualitative thermal imaging to quantitative thermal analysis, ensuring direct integration of data with the physical model and guaranteeing the consistency of data collected at different times and from different devices.

[0020] A Fast Fourier Transform (FFT) is performed on the temperature-time curve of each pixel in the standardized heatmap sequence to transform it from the time domain to the frequency domain, generating a series of phase data cubes with different frequency components. Specifically, this process creatively abandons the dependence on the amplitude spectrum and instead extracts and constructs a series of phase data cubes with different frequency components. Phase data is naturally immune to interference such as uneven surface emissivity and uneven heating, and its phase delay is directly related to the depth of heat wave propagation. This greatly suppresses surface noise and encodes depth information in the frequency dimension, making it possible to detect deep voids.

[0021] The phase data cube is reconstructed into a two-dimensional matrix, and principal component analysis (PCA) is performed on the matrix to obtain a spatial score map. Based on the spatial score map, background thermal field stripping is performed, and known intact regions are introduced as references for spatiotemporal differencing to obtain a difference sequence. Statistical feature blind source separation is then performed on the generated phase data cube. It is then reconstructed into a two-dimensional matrix in the space-frequency dimension (rows represent spatial pixels, and columns represent phase values ​​at different frequencies), and principal component analysis is performed on it. PCA automatically decouples independent thermal response modes from phase information at dozens of frequencies. Low-order principal components typically carry the background thermal field composed of the overall structure and uniform thermal diffusion, while high-order principal components capture the weak but consistent anomalous frequency responses caused by local defects. Extracting the spatial score maps corresponding to these high-order principal components essentially completes a statistically based intelligent stripping of the background thermal field, significantly improving the signal-to-noise ratio.

[0022] Building upon this foundation, a physically-guided differential enhancement strategy is introduced. Using a known intact region as a physical reference, the high-order principal component score map or phase map at a specific frequency obtained from the above processing is spatiotemporally differentially analyzed with the average response of this reference region. This operation effectively cancels common-mode noise and global environmental disturbances, and dramatically amplifies the minute differences in thermal response between defective and intact regions. Consequently, in the generated differential sequence, defects are highlighted as significant outliers, while the uniform background is suppressed.

[0023] The differential sequence is concatenated with the phase data cube to reconstruct a new multimodal feature sequence, completing the preprocessing of the heatmap sequence. Specifically, multimodal feature fusion is performed. The differentially enhanced sequence is concatenated with representative frequency components from the original phase data cube along the channel dimension, reconstructing a novel, information-rich multimodal feature sequence. This sequence integrates depth-sensitive phase features, statistically enhanced defect contour features, and physical differential contrast features, forming a high-dimensional feature set that is highly sensitive to internal hollow defects. This innovative data reconstruction process provides high-quality input for subsequent intelligent recognition algorithms, offering "pre-focusing" and "multi-angle insight," fundamentally improving the accuracy and robustness of the entire system's recognition.

[0024] S3: Based on the preprocessed heat map sequence, the internal thermal property feature map of the wall is reconstructed by solving the inverse heat conduction problem. Based on the internal thermal property feature map of the wall, the hollowness of the building exterior wall is identified.

[0025] A further implementation method for identifying hollow areas in building exterior walls includes: S31: Based on the partial differential equation of heat conduction, the problem of detecting hollow areas is transformed into the inverse problem of heat conduction for reconstructing the characteristic map of the internal thermal properties of the wall. S32: Based on the preprocessed heat map sequence, a reconstructed feature map of the internal thermal properties of the wall is obtained using an end-to-end inversion network based on deep learning; a further implementation method includes: A physically constrained network is constructed using a U-Net++ architecture with dense skip connections as the backbone, and a physically active layer with a built-in thermal conduction problem solver is embedded at the end. A further implementation method involves using the U-Net++ architecture with dense skip connections within the physically constrained network to simultaneously capture full-scale features from micro-texture to macro-structure. Specifically, the network backbone, through its encoder-decoder structure and dense cross-layer connections, achieves full reuse and fusion of features at different resolutions. The encoder path downsamples layer by layer, capturing the macro-contextual semantic information of the wall's thermal response; while the decoder path gradually recovers spatial details through upsampling. The crucial dense skip connections directly transmit the high-resolution, detailed features (such as tiny temperature abrupt change edges) of each encoder layer to all deeper decoders, ensuring that full-scale features from micro-texture to macro-structure are simultaneously preserved and used for final decision-making. This guarantees that the retrieved thermal property feature map has extremely high spatial accuracy and sensitivity to small-sized defects.

[0026] The physical activation layer takes the initially predicted thermal property distribution as input and, combined with known boundary conditions, uses a differential operator based on the finite difference method to forward simulate the theoretical thermal response sequence that the wall surface should produce under the initially predicted thermal property distribution. This theoretical thermal response sequence is then compared with the actual collected and preprocessed real thermal map sequence to obtain the physical residual. The known boundary conditions include the location of the pulse heat source. Specifically, after U-Net++ outputs the initially predicted thermal property distribution map, the creative process enters the embedded verification stage of physical laws. This initial prediction result is immediately fed into the physical activation layer at the end of the network. This layer is a hard-constraint module with a built-in differentiable, simplified heat conduction forward problem solver. This solver, based on the finite difference method, discretizes the continuous heat conduction partial differential equation into linear algebraic operations. The Physical Activation Layer uses the thermal property distribution map, which may not perfectly conform to physical laws and is initially predicted by U-Net++, as its calculated material parameter field. It then combines this with known boundary conditions—particularly the spatial location and time function of the instantaneous volumetric heat source formed by pulsed eddies in the metal component, and the initial temperature field—for forward simulation. By performing this simplified numerical simulation, the layer outputs a theoretical sequence of thermal responses to the wall surface. Next, the system calculates and feeds back the physical residuals. The theoretical thermal response sequence simulated by the Physical Activation Layer is compared point-by-point with a standard heat map sequence that has been collected and rigorously preprocessed, under the same spatiotemporal coordinates. The difference between the two is quantified as the physical residual. This residual, along with the traditional data residual between the network prediction and the training labels, and the regularization loss that encourages sparsity in the void distribution, constitutes a multi-task joint loss function. During training, the physical residual is backpropagated through the differentiable operator within the Physical Activation Layer, acting like a precise "physics coach" to directly guide the weight updates of the U-Net++ backbone network.

[0027] Ultimately, this structure forces deep learning networks to move beyond being merely "black box" fitters and become optimization engines constrained by first principles of physics. Through iterative training, the network learns to adjust its parameters so that its predicted thermal property distribution maps minimize the data discrepancies with the true labels while allowing its theoretical responses, derived from physical laws, to approximate real surface temperature measurements. This ensures that even in complex scenarios where training data is not fully covered, the network can produce physically plausible and interpretable inversion results, greatly enhancing the model's generalization ability and reliability in the real world.

[0028] Multimodal feature sequences are used as input to the physical constraint network. A composite loss function, fusing data loss, physical residual loss, and sparsity loss, is employed to perform multi-task physical-driven optimization of the physical constraint network, resulting in an optimized network. Existing parametric hollow-wall finite element simulation data is used to perform two-stage transfer training on the optimized network to obtain a hollow recognition model. Specifically, after obtaining the multimodal feature sequences, they are used as input to the physical constraint network, and a composite loss function fusing data-driven and physical laws is employed to perform multi-task collaborative optimization. Data loss terms (such as smoothing L1 loss) are responsible for calculating the pixel-level difference between the thermal property map predicted by the network and the "real" label provided by high-fidelity simulation or experimental calibration, ensuring that the network output closely matches the target value numerically, which is the foundation of model accuracy. The core innovation lies in the physical residual loss term, L_physics. It calculates the mean square error between the theoretical thermal response sequence simulated by the physical activation layer and the standardized heat map sequence collected in practice. This introduces the physical conservation law represented by the partial differential equation of heat conduction as a strong constraint into the optimization process, forcing the network's predictions to be consistent with the observed physical phenomena, thus significantly improving its generalization ability and the reasonableness of the prediction results. The sparsity loss term, L_sparsity, is based on the prior knowledge that the spatial distribution of voids is sparse. It applies L1 norm regularization to the prediction map, encouraging the model to produce solutions that are mostly uniform, with concentrated defects and clear boundaries, effectively suppressing the generation of background noise and false defects. By minimizing these three losses simultaneously through backpropagation, the network parameters are guided to a state that achieves an optimal balance between data fitting ability and physical consistency, thus obtaining the optimized physical constraint network.

[0029] To achieve a robust transition from "ideal simulation" to "complex reality," this invention employs a two-stage physics-guided transfer training strategy. In the first stage, pre-training is performed on a large-scale parametric hollow-wall finite element simulation dataset. This dataset generates a massive amount of "thermal property truth map - surface thermal response sequence" data pairs by randomly varying the depth, size, and shape of the hollows and the thermal properties of the wall material. In this stage, the network primarily relies on L_data and L_physics for learning, grasping the core physical laws governing the interaction between heat waves and complex structures from the near-infinite and perfect simulation data, establishing a robust initial mapping from surface thermal response to internal properties. In the second stage, a small amount of real wall detection data, rigorously calibrated using precision instruments (such as endoscopy and impact testing), is used to fine-tune the pre-trained model. In this stage, L_physics plays a crucial role; it acts as a "physical anchor point," preventing the model from overfitting to flawed labels when real data contains labeling errors or unknown noise, and guiding the model to find the optimal solution between the real data distribution and physical laws. Ultimately, through these two stages of relay training, we obtained a highly mature and reliable hollow wall identification model that fully absorbs the physical regularity of simulation data and the complexity of real data, achieving accurate, quantitative, and automated diagnosis of hollow walls in building exteriors.

[0030] Based on the hollow wall identification model, the thermal property feature map of the wall interior is obtained.

[0031] S33: Identify pixel areas in the thermal property feature map of the wall interior that are below the preset statistical threshold as hollow areas, and obtain a distribution map of hollow areas in the building exterior wall.

[0032] A further embodiment of the method for obtaining a distribution map of hollow areas in building exterior walls includes: S311: An adaptive thresholding algorithm based on statistical process control is used to initially segment the thermal property feature map inside the wall, achieving robust extraction of anomalous regions for different walls. Specifically, the adaptive thresholding algorithm based on statistical process control is used to initially segment the thermal property feature map inside the wall. This method treats the entire feature map as a statistical population, calculating the mean (μ) and standard deviation (σ) of all its pixel values. Subsequently, the segmentation threshold is dynamically set to... , where k is an empirically adjustable parameter (typically between 1.5 and 3). The innovation of this algorithm lies in its ability to automatically adapt to baseline drift in thermal properties caused by different wall materials, environmental conditions, and excitation intensities, without requiring a preset fixed threshold. It extracts anomalous regions by identifying "significantly low value areas" relative to the current statistical characteristics of the wall itself, thus achieving robust initial screening of anomalous regions for different walls.

[0033] S312: Mathematical morphology methods are used to post-process robust anomaly regions of different walls. Opening and closing operations are used to filter out noise and optimize the geometry and connectivity of hollow areas. Specifically, an opening operation of "erosion followed by dilation" is performed sequentially to filter out isolated pseudo-defects caused by noise. Then, a closing operation of "dilation followed by erosion" is performed to fill small holes caused by noise or inversion errors within the same hollow area and to mend narrow fractures. This process effectively optimizes the geometry and spatial connectivity of hollow areas, making the segmentation results more morphologically consistent with physical facts (hollow areas are usually continuous regions with a certain area), thus obtaining optimized connected hollow areas.

[0034] S313: Quantitative feature extraction and severity analysis are performed on the optimized hollow connected regions. The area and location of the hollow connected regions are calculated, and the thermal attribute values ​​of the hollow connected regions are graded based on their deviation, obtaining quantitative analysis results. For each independent connected region, the algorithm automatically calculates its pixel area, centroid coordinates (for localization), perimeter, and other geometric features. More importantly, severity is graded based on the deviation of its thermal attribute values: the average thermal attribute value of all pixels in the region is calculated, and based on the difference between the average thermal attribute value and the adaptive threshold T, it is divided into different levels such as "slight," "moderate," and "severe," thereby obtaining structured quantitative analysis results.

[0035] S314: The quantitative analysis results are fused with pre-acquired visible light images to obtain a building exterior wall void distribution map containing coordinates, area, and severity level. Specifically, the quantitative analysis results are ultimately fused with pre-acquired and highly registered visible light images. On the registered visible light base map, the precise outlines of the voids are highlighted using semi-transparent layers of different colors (e.g., green, yellow, red) according to their severity level. Simultaneously, a structured diagnostic report is generated, which not only includes a visualized void distribution map but also details the number, center coordinates, projected area, and severity level of each void in a data table format. This innovative process ultimately delivers a building exterior wall void distribution map integrating intuitive visual positioning and precise quantitative data, successfully completing the ultimate transformation from non-destructive testing signals to engineering decision-making information that directly guides maintenance.

[0036] Example 2 This invention also provides a building exterior wall hollowness identification system based on pulsed eddy current thermal imaging, and a method for implementing this system, including: The pulsed eddy current excitation module is used to apply the pulsed eddy current excitation device to the surface of the building exterior wall, and apply transient pulsed magnetic field excitation to the metal components inside the wall, so that the metal components induce eddy currents and generate Joule heat as an internal heat source. The thermal image sequence acquisition module is used to synchronously acquire transient pulse magnetic field excitation using an infrared thermal imager to obtain a thermal image sequence formed on the surface of the building exterior wall due to internal heat conduction, and to preprocess the thermal image sequence; The hollow wall identification module is used to reconstruct the internal thermal property feature map of the wall by solving the inverse heat conduction problem based on the preprocessed heat map sequence, and then to identify hollow walls in the building exterior.

[0037] A further implementation method is that the heatmap sequence acquisition module includes: The standardized unit is used to perform inter-frame registration of the heat map sequence and convert the original radiation data into absolute temperature values ​​according to preset emissivity, ambient temperature and ambient humidity parameters to obtain a standardized heat map sequence. The phase data acquisition unit is used to perform a fast Fourier transform on the temperature-time curve of each pixel in the standardized heat map sequence, converting it from the time domain to the frequency domain, and generating a series of phase data cubes with different frequency components. The principal component analysis unit is used to reconstruct the phase data cube into a two-dimensional matrix and perform principal component analysis on the two-dimensional matrix to obtain a spatial score map. The differential processing unit is used to strip the background thermal field based on the spatial score map and introduce a known intact region as a reference for spatiotemporal differential processing to obtain a differential sequence. The sequence reconstruction unit is used to concatenate the difference sequence with the phase data cube to reconstruct a new multimodal feature sequence, thus completing the preprocessing of the heatmap sequence.

[0038] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for identifying the hollowing of the building outer wall based on the pulsed eddy current thermal imaging, characterized in that, The method comprises the following steps: The pulse eddy current excitation device is applied to the surface of the building outer wall to apply a transient pulse magnetic field excitation to the metal member inside the wall body, so that the metal member induces eddy current and generates Joule heat as an internal heat source; An infrared thermal imager is used to synchronously collect the transient pulse magnetic field excitation to obtain a thermal map sequence formed by internal heat conduction of the building outer wall surface, and the thermal map sequence is preprocessed; Based on the preprocessed thermal map sequence, the thermal conduction inverse problem is solved to reconstruct a wall body internal thermal property feature map, and the building outer wall hollowing is identified based on the wall body internal thermal property feature map.

2. The method of claim 1, wherein, The method for preprocessing the thermal map sequence comprises the following steps: The thermal map sequence is inter-frame registered, and the original radiation data is converted into absolute temperature values according to preset emissivity, ambient temperature and ambient humidity parameters to obtain a standardized thermal map sequence; The temperature-time curve of each pixel point in the standardized thermal map sequence is subjected to fast Fourier transform to convert from time domain to frequency domain, and a phase data cube under a series of different frequency components is generated; The phase data cube is reconstructed into a two-dimensional matrix, and principal component analysis is performed on the two-dimensional matrix to obtain a spatial score map; Based on the spatial score map, background thermal field stripping is performed, and a known intact area is introduced as a reference for spatio-temporal difference processing to obtain a difference sequence; The difference sequence and the phase data cube are spliced and reconstructed into a new multi-modal feature sequence to complete the preprocessing of the thermal map sequence.

3. The method of claim 2, wherein, The method for identifying the building outer wall hollowing comprises the following steps: Based on the heat conduction partial differential equation, the hollowing detection problem is converted into a heat conduction inverse problem of reconstructing the wall body internal thermal property feature map; Based on the preprocessed thermal map sequence, a deep learning-based end-to-end inversion network is used to obtain the reconstructed wall body internal thermal property feature map; The pixel region in the wall body internal thermal property feature map that is lower than a preset statistical threshold is identified as a hollowing to obtain a building outer wall hollowing distribution map.

4. The method of claim 3, wherein, The method for obtaining the reconstructed wall body internal thermal property feature map comprises the following steps: A U-Net++ architecture with dense skip connection is adopted as a backbone, and a physical activation layer with a built-in heat conduction forward problem solver is embedded at the end to construct a physical constraint network; The multi-modal feature sequence is taken as an input of the physical constraint network, a compound loss function that fuses data loss, physical residual loss and sparsity loss is used to perform multi-task physical driving optimization on the physical constraint network, and an optimized physical constraint network is obtained; Existing parameterized hollowing-wall finite element simulation data are used to perform two-stage migration training on the optimized physical constraint network to obtain a hollowing identification model; Based on the hollowing identification model, the wall body internal thermal property feature map is obtained.

5. The method of claim 4, wherein, In the physical constraint network, the U-Net++ architecture with dense skip connection is used to simultaneously capture full-scale features from micro textures to macro structures. The physical activation layer is configured to input a network preliminary predicted thermal property feature distribution as an input, combine known boundary conditions, simulate a theoretical thermal response sequence that should be generated on the wall surface under the preliminary predicted thermal property feature distribution through a differential operator constructed based on a finite difference method, and compare the theoretical thermal response sequence with an actual collected and preprocessed real thermal map sequence to obtain a physical residual; wherein the known boundary conditions include a pulse heat source position.

6. The method of claim 5, wherein, The method for obtaining the building outer wall hollow distribution map comprises: An adaptive threshold algorithm based on statistical process control is used to perform initial segmentation on the wall internal thermal property feature map, so as to realize robust abnormal region extraction of different walls; A mathematical morphology method is used to post-process the robust abnormal regions of different walls, and an opening and closing operation is used to filter noise and optimize the geometric shape and connectivity of the hollow regions; Quantitative feature extraction and severity analysis are performed on the optimized hollow connected regions, the area and position of the hollow connected regions are calculated, the hollow connected regions are classified by using a thermal property value deviation degree of the hollow connected regions, and a quantitative analysis result is obtained; The quantitative analysis result is fused with a pre-collected visible light map to obtain a building outer wall hollow distribution map containing coordinates, area and grades.

7. A system for identifying the hollowing of the building outer wall based on the pulsed eddy current thermal imaging, for implementing the method according to any one of claims 1-6, characterized in that, Comprise: A pulse eddy current excitation module is configured to apply a pulse eddy current excitation device to the surface of a building outer wall to apply a transient pulse magnetic field excitation to a metal component inside the wall, so that the metal component induces eddy current and generates Joule heat as an internal heat source; A thermal map sequence acquisition module is configured to use an infrared thermal imager to synchronously acquire the transient pulse magnetic field excitation to obtain a thermal map sequence formed on the surface of the building outer wall due to internal heat conduction, and to pre-process the thermal map sequence; A hollow identification module is configured to reconstruct a wall internal thermal property feature map by solving a heat conduction inverse problem based on the pre-processed thermal map sequence, and to complete identification of a building outer wall hollow based on the wall internal thermal property feature map.

8. The system of claim 7, wherein, The thermal map sequence acquisition module comprises: A standardization unit is configured to perform inter-frame registration on the thermal map sequence, and convert original radiation data into absolute temperature values according to preset emissivity, ambient temperature and ambient humidity parameters to obtain a standardized thermal map sequence; A phase data acquisition unit is configured to perform fast Fourier transform on a temperature-time curve of each pixel point in the standardized thermal map sequence to convert from time domain to frequency domain and generate a phase data cube under a series of different frequency components; A principal component analysis unit is configured to reconstruct the phase data cube into a two-dimensional matrix and perform principal component analysis on the two-dimensional matrix to obtain a spatial score map; A difference processing unit is configured to perform background thermal field stripping based on the spatial score map, and introduce a known intact area as a reference to perform spatio-temporal difference processing to obtain a difference sequence; A sequence reconstruction unit is configured to splice and reconstruct the difference sequence and the phase data cube into a new multi-modal feature sequence to complete pre-processing of the thermal map sequence.

Images