Outer wall hollowing detection method and system

Abstract

The invention belongs to the field of hollowing detection, and provides an outer wall hollowing detection method and system, and the method comprises the steps: carrying out the multi-mode detection of an outer wall, and obtaining an infrared thermal imaging image and an ultrasonic detection signal; performing image enhancement processing on the infrared thermal imaging image; respectively carrying out feature extraction on the infrared thermal imaging image and the ultrasonic echo image by using a convolutional neural network; extracting and fusing high-dimensional feature vectors, and aligning an infrared thermal imaging image with an ultrasonic echo image through an image registration algorithm and a spatial transformation model; and classifying the fused feature vectors through a multi-layer sensor according to the data aligned in the step, and outputting the specific position, size and severity of the hollowing area. According to the scheme, the outer wall hollowing detection can be efficiently carried out.

Description

Technical Field

[0001] The invention belongs to the field of hollowing detection, and in particular relates to an exterior wall hollowing detection method and system. Background Art

[0002] Exterior wall hollowing is a common quality problem in building structures, especially in the exterior wall coating or decorative layer of high-rise buildings. The hollowing phenomenon is usually caused by the failure of the bond between the exterior wall coating, insulation layer or finishing layer and the base layer, resulting in delamination, shedding, etc. This problem not only affects the beauty of the building, but may also bring serious safety hazards. Especially in high-rise buildings, the hollowing and shedding of exterior wall decorative materials such as tiles, marble or paint may pose a safety threat to pedestrians and surrounding facilities. Therefore, timely detection and repair of exterior wall hollowing is crucial to ensure the safety of buildings.

[0003] Traditional methods for detecting hollow walls mainly rely on manual tapping and listening. Inspectors tap the surface of the exterior wall and listen to the sound to determine whether there is a hollow wall. Although this method is simple and intuitive, it has obvious limitations: manual detection is inefficient, the coverage area is limited, and it is greatly affected by the experience of the inspectors, and the size and location of the hollow area cannot be quantified. With the expansion of building scale and the diversification of decorative materials, traditional detection methods are difficult to meet the needs of large-scale and refined detection.

[0004] In recent years, non-destructive testing technology has been widely used in building structure inspection. Infrared thermal imaging technology and ultrasonic testing technology, as common non-destructive testing methods, can effectively identify hollowing of exterior walls. Infrared thermal imaging uses temperature differences to detect thermal anomalies on the wall surface, and identifies hollow areas by analyzing temperature changes in different areas of the wall surface. Ultrasonic testing uses the reflection and attenuation characteristics of sound waves to determine whether there are hollows inside the wall. However, due to the diversity of exterior wall materials and complex structures, a single detection method is often difficult to guarantee the accuracy and comprehensiveness of the detection. For example, infrared thermal imaging is greatly affected by ambient temperature, and the signal attenuation of ultrasonic testing in multi-layer material structures will also affect the accuracy.

[0005] Therefore, how to integrate multiple nondestructive testing technologies and give full play to their respective advantages has become a research hotspot in exterior wall hollowing detection technology. By combining infrared thermal imaging and ultrasonic testing, not only can the dual capture of surface and internal information be achieved, but also the feature extraction and fusion of different detection data can be performed through deep learning technology to improve the accuracy and robustness of detection. However, given the heterogeneity of multimodal data, how to spatially align these data is a technical difficulty. Summary of the invention

[0006] In order to solve the problems in the prior art, the present invention provides a method for detecting hollowing of an exterior wall, the method comprising the following steps: Step 1, data collection: perform multimodal detection on the exterior wall to obtain infrared thermal imaging images and ultrasonic detection signals respectively. Infrared thermal imaging is used to detect the temperature distribution on the wall surface, and ultrasonic detection is used to reflect the sound wave reflection characteristics inside the wall. Step 2, data preprocessing: perform image enhancement processing on the infrared thermal imaging image to improve the image contrast, and extract the edge features of the wall through the edge detection algorithm; perform denoising processing on the ultrasonic detection signal and convert it into a two-dimensional echo map; Step 3, feature extraction and fusion: Use convolutional neural networks to extract features from infrared thermal imaging images and ultrasonic echograms respectively. The extracted features include edges, textures, and echo intensity related to hollowing. The feature vectors of different modalities are integrated through weighted fusion to form a fused high-dimensional feature vector. Step 4, spatial alignment of data: extract and fuse high-dimensional feature vectors, align the infrared thermal imaging image with the ultrasonic echo map through image registration algorithm and spatial transformation model, so that different modality data are presented in a unified coordinate system to achieve spatial consistency of data; Step 5, hollow area determination: Based on the data after step alignment, the fused feature vector is classified through a multi-layer perceptron to output the specific location, size and severity of the hollow area.

[0007] Another aspect of the present invention further provides an exterior wall hollowing detection system, the system comprising the following modules: Data acquisition module: Perform multi-modal detection on the exterior wall to obtain infrared thermal imaging images and ultrasonic detection signals. Infrared thermal imaging is used to detect the temperature distribution on the wall surface, and ultrasonic detection is used to reflect the sound wave reflection characteristics inside the wall. Data preprocessing module: Perform image enhancement processing on infrared thermal imaging images to improve image contrast, and extract edge features of the wall through edge detection algorithm; perform denoising processing on ultrasonic detection signals and convert them into two-dimensional echo images; Feature extraction and fusion module: Convolutional neural networks are used to extract features from infrared thermal imaging images and ultrasonic echograms. The extracted features include edges, textures, and echo intensity related to hollowing. The feature vectors of different modes are integrated through weighted fusion to form a fused high-dimensional feature vector. Data spatial alignment module: extracts and fuses high-dimensional feature vectors, aligns infrared thermal imaging images with ultrasonic echograms through image registration algorithms and spatial transformation models, and presents different modal data in a unified coordinate system to achieve spatial consistency of data; Hollow area determination module: Based on the data after step alignment, the fused feature vector is classified through a multi-layer perceptron to output the specific location, size and severity of the hollow area.

[0008] The exterior wall hollowing detection method and system of the present invention have the following beneficial effects: By combining infrared thermal imaging and ultrasonic testing technology, comprehensive collection of external wall surface and internal hollowing information is achieved, and the surface temperature distribution and internal sound wave reflection characteristics can be captured simultaneously. By utilizing the complementarity of these two non-destructive testing methods, the comprehensiveness and accuracy of the detection are improved, overcoming the limitations of a single detection method.

[0009] The present invention uses convolutional neural networks to extract high-level features from infrared thermal imaging images and ultrasonic echograms, and uses a weighted fusion method to integrate the features of different modes to form a high-dimensional feature vector. This process can effectively extract key features related to hollowing and enhance the ability to identify hollow areas in complex building structures.

[0010] By using image registration algorithms and non-rigid transformation models, infrared thermal imaging images and ultrasonic echograms can be accurately aligned in a unified coordinate system, solving the problem of heterogeneity of multimodal data. Spatial alignment ensures the consistency of multimodal data analysis and provides a reliable spatial basis for subsequent detection steps.

[0011] Based on the multi-layer perceptron model, the fused feature vector can be automatically classified and output the specific location, size and severity of the hollow area, realizing the automatic and intelligent judgment of the hollow problem, reducing the subjective errors of manual operation and greatly improving the detection efficiency.

[0012] In summary, the present invention can greatly improve the efficiency, accuracy and automation level of exterior wall hollowing detection, and has important application value and broad market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0014] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0015] Below, the invention is preferably described in conjunction with the accompanying drawings and specific implementation methods.

[0016] This embodiment solves the above problem through the following steps: In one embodiment, reference Figure 1 The present invention provides a method for detecting hollowing defects in exterior walls of buildings, which relates to a technical solution for detecting hollowing defects in exterior walls of buildings by combining multimodal nondestructive testing technology. The method aims to accurately identify the separation area between the exterior wall coating, finishing layer or insulation layer and the base layer by collecting, processing, fusing and analyzing multiple detection signals of the exterior wall, so as to achieve accurate judgment of the location, size and severity of the hollowing.

[0017] Step 1, data collection: Perform multimodal detection on the exterior wall to obtain infrared thermal imaging images and ultrasonic detection signals respectively. Infrared thermal imaging is used to detect the temperature distribution on the wall surface, and ultrasonic detection is used to reflect the sound wave reflection characteristics inside the wall.

[0018] First, use infrared thermal imaging equipment to scan the surface of the exterior wall and obtain infrared thermal imaging images. Infrared thermal imaging equipment is used to detect the temperature distribution of the exterior wall surface. Its principle is to capture the temperature difference between different areas of the exterior wall surface to reflect the bonding condition between the wall surface coating, decorative layer or insulation layer and the base layer. Areas with abnormal temperatures may represent hollow areas, because hollow areas usually have poor bonding and reduced thermal conductivity, resulting in a difference in surface temperature from the surrounding normal areas. Therefore, infrared thermal imaging images provide a preliminary thermal basis for identifying hollow areas on the exterior wall surface.

[0019] Secondly, use ultrasonic testing equipment to perform ultrasonic scanning on the exterior wall to obtain ultrasonic testing signals. Ultrasonic testing equipment reflects the structural characteristics of the interior of the wall by emitting sound waves to the exterior wall and receiving the reflected sound wave signals. The propagation speed and reflection characteristics of sound waves between different material layers are different. The hollow area will cause abnormal sound wave reflection or large energy attenuation due to the internal gaps or debonding. Ultrasonic testing signals can accurately reflect the bonding state between the internal layers of the exterior wall, thereby identifying potential hollow areas inside the wall. In particular, ultrasonic testing is suitable for identifying deep-seated hollowing problems, which supplements the limitation of infrared thermal imaging that cannot penetrate the interior of the wall.

[0020] Through the above steps, the temperature distribution image of the exterior wall surface and the sound wave reflection signal inside the wall are obtained respectively. Infrared thermal imaging is used to provide the thermal characteristics of the exterior wall surface, and ultrasonic detection is used to provide the sound wave propagation characteristics inside the exterior wall. These two types of data reflect the physical state of the exterior wall surface and different layers inside, respectively, and provide basic data support for subsequent data processing, feature extraction, fusion and spatial alignment.

[0021] Step 2, data preprocessing: perform image enhancement processing on the infrared thermal imaging image to improve the image contrast, and extract the edge features of the wall through the edge detection algorithm; perform denoising processing on the ultrasonic detection signal and convert it into a two-dimensional echo image.

[0022] First, the infrared thermal imaging images are preprocessed using image enhancement technology. By adjusting the brightness and contrast of the image, the visibility of the temperature difference in the image is improved, so that the temperature distribution characteristics of the wall surface are more clearly presented. At the same time, in order to further extract geometric information that is helpful for subsequent feature extraction and analysis, the preprocessed infrared thermal imaging images are processed using an edge detection algorithm. Specifically, the edge features of the wall surface are extracted through classic edge detection algorithms such as Canny or Sobel, especially focusing on the edges of areas where hollows may exist, so as to obtain a feature image reflecting the outline of the wall structure. This process helps to highlight the geometric features of potential hollow areas and improve the resolution of the image and the efficiency of extracting effective information.

[0023] Secondly, the ultrasonic detection signal is denoised to reduce the impact of environmental noise and other interference factors on signal quality. Through signal processing techniques such as filters or wavelet transforms, the noise components in the ultrasonic signal are removed to retain the effective echo information related to the internal structure of the wall. The denoised ultrasonic signal is further converted into a two-dimensional echo map, which intuitively reflects the reflection intensity and signal attenuation of the ultrasonic wave between different layers inside the wall through the mapping of the time-distance model. The generation of this two-dimensional echo map facilitates the spatial alignment and feature fusion of ultrasonic data and infrared thermal imaging images in subsequent steps.

[0024] The ultrasonic signal is further converted into a two-dimensional echo map through the mapping of the time-distance model, including: The denoised ultrasonic signal is mainly a time domain signal, that is, the sound wave propagation time and the received echo signal. In order to convert it into a two-dimensional echo map that can intuitively represent the internal structure of the wall, it is first necessary to map time to depth. This step is based on the propagation speed of ultrasonic waves in the medium: , Among them, d is the propagation depth of ultrasound in the medium, v is the propagation speed of ultrasound in the wall material, t is the echo time, and 2 represents the round-trip path of ultrasound (from emission to reflection and then back to the receiver).

[0025] By calculating the arrival time t of the reflected wave, the depth d of each reflection point can be determined, thereby converting the time domain signal into the depth information inside the wall.

[0026] The depth information obtained in the above steps is combined with the lateral displacement to generate a two-dimensional echo map. The specific process is as follows: Horizontal scanning: The ultrasonic detection equipment moves or scans on the wall surface, and obtains echo information at different horizontal positions (x-axis) by transmitting and receiving sound waves multiple times. The corresponding depth echo signal is recorded at each horizontal position.

[0027] Two-dimensional data matrix construction: Multiple one-dimensional echo signals (corresponding to depth information at different lateral positions) obtained by lateral scanning are combined into a two-dimensional data matrix, where the vertical axis represents the depth information (d-axis) and the horizontal axis represents the lateral displacement of the detection equipment on the wall surface (x-axis).

[0028] Image rendering: The two-dimensional matrix is ​​color-coded or gray-scaled according to the echo intensity (amplitude or energy of the signal). Locations with greater echo intensity (such as internal structure interfaces or hollow areas) will be displayed as high-light or dark, and vice versa. A two-dimensional echo map can be generated to intuitively display the structural characteristics between different layers inside the wall.

[0029] Through the processing of this step, the edge features of the infrared thermal imaging image and the two-dimensional echo map of the ultrasonic detection signal are effectively extracted and optimized, providing high-quality input data for feature extraction, data fusion and accurate identification of hollow areas in subsequent steps, ensuring detection accuracy and data reliability.

[0030] Step 3, feature extraction and fusion: Use convolutional neural networks to extract features from infrared thermal imaging images and ultrasonic echo images respectively. The extracted features include edges, textures, and echo intensities related to hollowing. The feature vectors of different modalities are integrated through weighted fusion to form a fused high-dimensional feature vector.

[0031] First, a pre-trained convolutional neural network is used to extract features from infrared thermal imaging images. The convolutional neural network consists of several convolutional layers, pooling layers, and fully connected layers, and the input is the infrared thermal imaging image that has been preprocessed. The convolutional layer extracts high-level features related to hollowing, such as local temperature differences and edge features. For example, the thermal anomaly area on the wall surface is usually the potential location of the hollowing area. The pooling layer is used to reduce the dimension of the feature map while retaining the key features to improve the computational efficiency of the model. Finally, the extracted high-level features are converted into feature vectors through the fully connected layer. The feature vector includes multi-dimensional information reflecting the temperature difference, edge and texture features in the infrared thermal imaging image, providing input for subsequent data fusion.

[0032] Next, another convolutional neural network is used to extract features from the ultrasonic echogram. This network is similar to the network structure used to process infrared thermal imaging, but its convolution kernel is designed to adapt to the characteristics of time series signals or two-dimensional signals, and can capture hollowing-related features such as echo intensity and reflection pattern in ultrasonic signals that reflect the internal structure of the wall. In particular, the echo of ultrasound in the hollow area usually shows significant attenuation or abnormal reflection characteristics, so the convolutional neural network can extract these high-level features that are directly related to the hollowing and convert them into feature vectors. The feature vector mainly includes information such as echo intensity and attenuation pattern.

[0033] After the feature extraction of infrared thermal imaging images and ultrasonic echograms is completed, the feature fusion step is entered. First, the infrared thermal imaging feature vector and ultrasonic feature vector obtained are weighted to reflect the importance of different modal data to hollowing detection. The weighting coefficient can be determined through experience or by training and optimizing experimental data of different building materials. For example, for weaker wall structures, ultrasonic signals may be more effective than infrared thermal imaging in detecting hollowing areas. At this time, the weight of ultrasonic features is higher. Then, the features from infrared and ultrasonic detection are fused by concatenating or weighted averaging feature vectors to form the final high-dimensional feature vector.

[0034] The fused high-dimensional feature vector contains the temperature difference information of the wall surface, the echo intensity information of the wall interior, and the respective edge and texture features, so that the vector can fully characterize the existence of hollowing. For example, if a certain detection area shows abnormal temperature in infrared thermal imaging and significant echo attenuation in ultrasonic detection, the fused feature vector will reflect the high confidence that the area is a hollow area.

[0035] Through this feature extraction and fusion step, the extracted high-dimensional feature vector has the complementarity and spatial consistency of different modal data, providing more accurate and reliable feature information for the subsequent determination of hollow areas.

[0036] Step 4, spatial alignment of data: extract and fuse high-dimensional feature vectors, align the infrared thermal imaging image with the ultrasonic echo map through image registration algorithm and spatial transformation model, so that different modal data are presented in a unified coordinate system to achieve spatial consistency of data.

[0037] First, the infrared thermal imaging image and the ultrasonic echogram are geometrically aligned through the preliminary registration step. Since the infrared thermal imaging image and the ultrasonic echogram belong to different modalities, direct alignment may have spatial differences. Therefore, the present invention uses a feature point-based image registration algorithm for preliminary alignment. The specific steps are as follows: Using the high-dimensional feature vectors generated in the feature extraction and fusion steps, key feature points are extracted from the infrared thermal imaging image and the ultrasonic echo map. The feature points in the infrared thermal imaging image mainly include the temperature anomaly edges of the hollow area, and the feature points in the ultrasonic echo map include the abnormal sound wave reflection area (such as strong echo or attenuation abnormal points). These feature points identify the possible locations of hollows in the two modal data.

[0038] The extracted infrared image feature points are matched with the ultrasonic image feature points by using algorithms such as SIFT (Scale Invariant Feature Transform) or SURF (Speeded Up Robust Features). This process is based on the local descriptors of the feature points, and finds point pairs with similar physical features in the two images, thereby establishing a cross-modal matching relationship.

[0039] After the preliminary matching is completed, the infrared thermal imaging image and the ultrasonic echo map are geometrically registered using a rigid transformation model. Rigid transformation includes translation, rotation, and scale transformation, and the optimal registration parameters are determined by minimizing the distance between the feature points of the two modalities. This step roughly aligns the images of different modalities to ensure that they are presented in roughly the same coordinate system.

[0040] After the preliminary geometric alignment, the fine spatial alignment stage is entered. Since the acquisition of different modal data may have non-rigid deformation (such as the angle of the detection equipment or the curved surface characteristics of the exterior wall surface), the present invention further uses a non-rigid transformation model for spatial alignment, specifically: Using non-rigid transformation methods such as Thin Plate Spline (TPS), local deformation adjustment is performed on infrared thermal imaging images and ultrasonic echo images. The model can adapt to local nonlinear changes and achieve fine alignment of images by interpolating between key feature points. Non-rigid transformation not only considers global coordinate consistency, but also adjusts local geometric shapes, so that the two modal data can remain consistent under complex surface deformation.

[0041] To ensure the accuracy of spatial alignment, the present invention introduces an optimization-based registration algorithm, which minimizes the matching error between the registered infrared thermal imaging image and the ultrasonic echo map by iteratively optimizing the error function. The error function is usually based on the maximization of feature point matching error, image gradient similarity or mutual information. After several iterative optimizations, a fine spatial alignment result can be obtained.

[0042] For example, in practical applications, if the edge of the temperature anomaly in a certain area of ​​the infrared thermal imaging image and the acoustic wave attenuation of the corresponding area in the ultrasonic echo map are obvious, then through the above registration steps, the temperature anomaly area of ​​the infrared image can be accurately aligned with the attenuation area of ​​the ultrasonic image, so that the two pieces of information can be presented simultaneously in a unified coordinate system. At this point, whether it is the surface temperature difference shown in the infrared thermal imaging or the internal structure attenuation shown in the ultrasonic image, they can be comprehensively analyzed in the same space, so as to accurately identify the location and size of the hollow area.

[0043] Through the processing of this step, the data of infrared thermal imaging image and ultrasonic echo map are successfully aligned, ensuring the consistency of different modal data in space, and providing a reliable spatial basis for the accurate determination of hollow areas in subsequent steps. At the same time, through the fine spatial alignment process, the effective fusion and analysis of different modal feature vectors in a unified coordinate system are guaranteed.

[0044] Step 5, hollow area determination: Based on the data after step alignment, the fused feature vector is classified through a multi-layer perceptron to output the specific location, size and severity of the hollow area.

[0045] First, the high-dimensional feature vector after spatial alignment and feature fusion is input into the multi-layer perceptron model. The multi-layer perceptron model consists of an input layer, multiple hidden layers and an output layer. The input layer receives the fused feature vector, which contains information such as temperature difference and edge features in infrared thermal imaging images, and sound wave reflection intensity and attenuation mode in ultrasonic echograms. Each hidden layer processes the input features through a nonlinear activation function (such as ReLU or Sigmoid), extracts high-level feature representations in different dimensions, and gradually enhances the recognition ability of hollow areas.

[0046] Next, the hidden layer in the multilayer perceptron model further processes the input feature vector through a fully connected operation, extracting high-level representations related to hollow features layer by layer. This process can identify complex feature interaction patterns related to hollowing through the model's nonlinear learning ability. For example, if a certain area shows significant temperature anomalies in infrared thermal imaging images and strong sound wave reflection attenuation in ultrasonic echo images, the model will be able to capture the association of these composite features through the multilayer perceptron, thereby accurately identifying the area as a hollow area.

[0047] After the feature processing is completed, it enters the output layer of the model. The output layer is responsible for classifying the fused feature vectors into multiple categories, including "hollow area" and "non-hollow area", and further outputs the specific attribute information of the hollow area. Specifically, the output layer gives the following information through the classification results of the fused feature vector: Hollow drum position: The multilayer perceptron determines the specific coordinate position of the hollow drum area based on the input feature vector. By ensuring the consistency of the coordinate system aligned with the aforementioned space, the output hollow drum position is consistent with the actual position of the infrared thermal imaging and ultrasonic detection signals.

[0048] Hollow size: Based on the classification results of the multilayer perceptron model, the area or volume of the hollow region is calculated. This result can provide the precise range of the hollow region by inferring the spatial range of the relevant area in the feature vector.

[0049] Hollowing severity: Outputs the severity of the hollowing area, which is determined based on the significance of the temperature difference in infrared thermal imaging and the strength of ultrasonic echo attenuation. The severity of hollowing can be divided into multiple levels: "mild hollowing", "moderate hollowing" and "severe hollowing". The model outputs the specific hollowing level of the area through a comprehensive analysis of the input features.

[0050] Finally, the multi-layer perceptron classification model of this step generates detailed detection results for the hollow area of ​​the exterior wall, outputs the specific location, size and severity of the hollow area, and ensures the accuracy and reliability of the detection. At the same time, this step can automatically determine the hollowing situation of a large range of exterior walls, reduce manual intervention, and improve detection efficiency and accuracy.

[0051] On the other hand, the present invention also provides an exterior wall hollowing detection system, comprising: Data acquisition module: Perform multi-modal detection on the exterior wall to obtain infrared thermal imaging images and ultrasonic detection signals. Infrared thermal imaging is used to detect the temperature distribution on the wall surface, and ultrasonic detection is used to reflect the sound wave reflection characteristics inside the wall. Data preprocessing module: Perform image enhancement processing on infrared thermal imaging images to improve image contrast, and extract edge features of the wall through edge detection algorithm; perform denoising processing on ultrasonic detection signals and convert them into two-dimensional echo images; Feature extraction and fusion module: Convolutional neural networks are used to extract features from infrared thermal imaging images and ultrasonic echograms. The extracted features include edges, textures, and echo intensity related to hollowing. The feature vectors of different modes are integrated through weighted fusion to form a fused high-dimensional feature vector. Data spatial alignment module: extracts and fuses high-dimensional feature vectors, aligns infrared thermal imaging images with ultrasonic echograms through image registration algorithms and spatial transformation models, and presents different modal data in a unified coordinate system to achieve spatial consistency of data; Hollow area determination module: Based on the data after step alignment, the fused feature vector is classified through a multi-layer perceptron to output the specific location, size and severity of the hollow area.

[0052] The prior art mentioned in the above background technology section and specific embodiment section of the present invention can be used as part of the present invention to understand the meaning of some technical features or parameters.

Claims

1. A method for detecting hollowing of an exterior wall, characterized in that: The method comprises the following steps: Step 1, data collection: perform multimodal detection on the exterior wall to obtain infrared thermal imaging images and ultrasonic detection signals respectively. Infrared thermal imaging is used to detect the temperature distribution on the wall surface, and ultrasonic detection is used to reflect the sound wave reflection characteristics inside the wall. Step 2, data preprocessing: perform image enhancement processing on the infrared thermal imaging image to improve the image contrast, and extract the edge features of the wall through the edge detection algorithm; perform denoising processing on the ultrasonic detection signal and convert it into a two-dimensional echo map; Step 3, feature extraction and fusion: Use convolutional neural networks to extract features from infrared thermal imaging images and ultrasonic echograms respectively. The extracted features include edges, textures, and echo intensity related to hollowing. The feature vectors of different modalities are integrated through weighted fusion to form a fused high-dimensional feature vector. Step 4, spatial alignment of data: extract and fuse high-dimensional feature vectors, align the infrared thermal imaging image with the ultrasonic echo map through image registration algorithm and spatial transformation model, so that different modality data are presented in a unified coordinate system to achieve spatial consistency of data; Step 5, hollow area determination: Based on the data after step alignment, the fused feature vector is classified through a multi-layer perceptron to output the specific location, size and severity of the hollow area.

2. The method for detecting hollowing of an exterior wall according to claim 1, characterized in that: The conversion into a two-dimensional echogram form includes: Mapping time to depth; The ultrasonic detection equipment moves or scans on the wall surface, and obtains echo information at different lateral positions by transmitting and receiving sound waves multiple times. The corresponding depth echo signal is recorded at each lateral position. The multiple one-dimensional echo signals obtained by the lateral scanning are combined into a two-dimensional data matrix, where the vertical axis represents the depth information and the horizontal axis represents the lateral displacement of the detection device on the wall surface; The two-dimensional matrix is ​​color-coded or gray-scaled according to the echo intensity. The position with high echo intensity will be displayed as high-brightness or dark-color, and vice versa. At this time, a two-dimensional echo map can be generated.

3. The method for detecting hollowing of an exterior wall according to claim 1, characterized in that: The method of using a convolutional neural network to extract features from the infrared thermal imaging image and the ultrasonic echograph includes: For infrared thermal imaging images, a pre-trained convolutional neural network is used for feature extraction. The convolutional neural network consists of several convolutional layers, pooling layers and fully connected layers. The input is the infrared thermal imaging image that has been preprocessed. The convolutional layer extracts high-level features related to local temperature differences, edge features and hollowing. The pooling layer is used to reduce the dimension of the feature map while retaining the key features. The fully connected layer converts the extracted high-level features into feature vectors. For the ultrasonic echogram, another convolutional neural network is used for feature extraction. Its convolution kernel is designed to adapt to the characteristics of time series signals or two-dimensional signals. The convolutional neural network can extract these high-level features that are directly related to hollowing and convert them into feature vectors.

4. The method for detecting hollowing of an exterior wall according to claim 1, characterized in that: Extract and fuse high-dimensional feature vectors, align infrared thermal imaging images with ultrasonic echograms through image registration algorithms and spatial transformation models, present different modal data in a unified coordinate system, and achieve spatial consistency of data, including: The infrared thermographic image and the ultrasonic echogram are geometrically aligned through a preliminary registration step; After the preliminary matching is completed, the infrared thermal imaging image and the ultrasonic echo map are geometrically registered using a rigid transformation model; Using non-rigid transformation method, local deformation adjustment is performed on infrared thermal imaging image and ultrasonic echo map; Based on the optimal registration algorithm, the error function is iteratively optimized to minimize the matching error between the registered infrared thermal imaging image and the ultrasonic echograph.

5. The method for detecting hollowing of an exterior wall according to claim 4, characterized in that: The geometric alignment of the infrared thermography image and the ultrasound echogram is performed through a preliminary registration step including: Using the high-dimensional feature vector generated in the feature extraction and fusion steps, key feature points are extracted from the infrared thermal imaging image and the ultrasonic echogram respectively; The extracted infrared image feature points and ultrasonic image feature points are matched by using SIFT or SURF algorithm.

6. An external wall hollowing detection system, characterized in that: The system includes the following modules: Data acquisition module: Perform multi-modal detection on the exterior wall to obtain infrared thermal imaging images and ultrasonic detection signals. Infrared thermal imaging is used to detect the temperature distribution on the wall surface, and ultrasonic detection is used to reflect the sound wave reflection characteristics inside the wall. Data preprocessing module: Perform image enhancement processing on infrared thermal imaging images to improve image contrast and extract edge features of the wall through edge detection algorithm; De-noising the ultrasonic detection signal and converting it into a two-dimensional echograph; Feature extraction and fusion module: Convolutional neural networks are used to extract features from infrared thermal imaging images and ultrasonic echograms. The extracted features include edges, textures, and echo intensity related to hollowing. The feature vectors of different modes are integrated through weighted fusion to form a fused high-dimensional feature vector. Data spatial alignment module: extracts and fuses high-dimensional feature vectors, aligns infrared thermal imaging images with ultrasonic echograms through image registration algorithms and spatial transformation models, and presents different modal data in a unified coordinate system to achieve spatial consistency of data; Hollow area determination module: Based on the data after step alignment, the fused feature vector is classified through a multi-layer perceptron to output the specific location, size and severity of the hollow area.

7. The exterior wall hollowing detection system according to claim 6, characterized in that: The conversion into a two-dimensional echogram form includes: Mapping time to depth; The ultrasonic detection equipment moves or scans on the wall surface, and obtains echo information at different lateral positions by transmitting and receiving sound waves multiple times. The corresponding depth echo signal is recorded at each lateral position. The multiple one-dimensional echo signals obtained by the lateral scanning are combined into a two-dimensional data matrix, where the vertical axis represents the depth information and the horizontal axis represents the lateral displacement of the detection device on the wall surface; The two-dimensional matrix is ​​color-coded or gray-scaled according to the echo intensity. The position with high echo intensity will be displayed as high-brightness or dark-color, and vice versa. At this time, a two-dimensional echo map can be generated.

8. The exterior wall hollowing detection system according to claim 6, characterized in that: The method of using a convolutional neural network to extract features from the infrared thermal imaging image and the ultrasonic echograph includes: For infrared thermal imaging images, a pre-trained convolutional neural network is used for feature extraction. The convolutional neural network consists of several convolutional layers, pooling layers and fully connected layers. The input is the infrared thermal imaging image that has been preprocessed. The convolutional layer extracts high-level features related to local temperature differences, edge features and hollowing. The pooling layer is used to reduce the dimension of the feature map while retaining the key features. The fully connected layer converts the extracted high-level features into feature vectors. For the ultrasonic echogram, another convolutional neural network is used for feature extraction. Its convolution kernel is designed to adapt to the characteristics of time series signals or two-dimensional signals. The convolutional neural network can extract these high-level features that are directly related to hollowing and convert them into feature vectors.

9. The exterior wall hollowing detection system according to claim 6, characterized in that: Extract and fuse high-dimensional feature vectors, align infrared thermal imaging images with ultrasonic echograms through image registration algorithms and spatial transformation models, present different modal data in a unified coordinate system, and achieve spatial consistency of data, including: The infrared thermographic image and the ultrasonic echogram are geometrically aligned through a preliminary registration step; After the preliminary matching is completed, the infrared thermal imaging image and the ultrasonic echo map are geometrically registered using a rigid transformation model; Using non-rigid transformation method, local deformation adjustment is performed on infrared thermal imaging image and ultrasonic echo map; Based on the optimal registration algorithm, the error function is iteratively optimized to minimize the matching error between the registered infrared thermal imaging image and the ultrasonic echograph.

10. The exterior wall hollowing detection system according to claim 9, characterized in that: The geometric alignment of the infrared thermography image and the ultrasound echogram is performed through a preliminary registration step including: Using the high-dimensional feature vector generated in the feature extraction and fusion steps, key feature points are extracted from the infrared thermal imaging image and the ultrasonic echogram respectively; The extracted infrared image feature points and ultrasonic image feature points are matched by using SIFT or SURF algorithm.

Images