A method for detecting the bonding effect of external wall insulation board and AR glasses

Abstract

The application provides a kind of method and AR glasses for the bonding effect detection of external wall insulation board, belongs to image data processing technical field, the method and AR glasses for the bonding effect detection of external wall insulation board provided by the application first collect AR glasses image data and ultrasonic detection equipment data and carry out space-time registration, generate optimal observation point vector group by fusion analysis, guide inspection personnel to adjust the position of detection equipment.Data is collected in optimal position to form detection data matrix, based on stability index, it is disassembled into basic matrix, first variable matrix and second variable matrix, after calculating the contribution value of each matrix and the characteristic vector of each matrix, input external wall insulation board bonding effect analysis model, output bonding effect index value, solve the technical problems that existing technology cannot accurately and efficiently detect the bonding effect of external wall insulation board.

Description

Technical Field

[0001] This invention belongs to the field of image data processing technology, and more specifically, relates to a method for detecting the bonding effect of external wall insulation boards and AR glasses. Background Technology

[0002] External wall insulation systems are a crucial component of modern building energy conservation, and the bonding quality between the insulation board and the wall directly affects building energy efficiency and safety. Traditional testing methods mainly include the tapping method, pull-out test, and infrared thermography. The tapping method relies on the experience of the tester based on sound, making it highly subjective; the pull-out test, while accurate, is destructive and cannot be applied on a large scale; infrared thermography is greatly affected by ambient temperature, and its reliability decreases under complex weather conditions.

[0003] In recent years, ultrasonic testing technology has been increasingly applied to the inspection of external wall insulation systems due to its non-destructive nature and applicability. However, ultrasonic testing alone suffers from problems such as complex data interpretation and inaccurate spatial positioning. Meanwhile, augmented reality (AR) technology provides a visual data presentation method in the field of building inspection, but lacks an effective integration mechanism with professional testing equipment.

[0004] Current technologies for detecting the bonding effect of external wall insulation boards face challenges such as inaccurate data acquisition, difficulties in integrating multi-source data, and inconsistent standards for quantifying test results. In particular, the lack of effective methods to integrate AR visualization technology with ultrasonic testing data leads to low testing efficiency and limited reliability of results. In other words, existing technologies suffer from the inability to accurately and efficiently detect the bonding effect of external wall insulation boards. Summary of the Invention

[0005] In view of this, the present invention provides a method for detecting the bonding effect of external wall insulation boards and AR glasses, which can solve the technical problem that the bonding effect of external wall insulation boards cannot be accurately and efficiently detected in the prior art.

[0006] This invention is implemented as follows: It provides a method and AR glasses for detecting the bonding effect of external wall insulation boards, comprising: acquiring image data from AR glasses and detection data from an ultrasonic testing device; performing spatiotemporal registration and fusion analysis on the image data and detection data to generate an optimal observation point vector set; guiding inspectors to sequentially adjust the AR glasses to the optimal observation point coordinate position while simultaneously aligning the ultrasonic testing device with the optimal detection point coordinate position; acquiring data at the optimal observation point coordinate position of the AR glasses and the optimal detection point coordinate position of the ultrasonic testing device to form a detection data matrix; decomposing the detection data matrix into a basic matrix, a first variation matrix, and a second variation matrix based on a stability index, and calculating the contribution values ​​and eigenvectors of the three matrices; inputting the eigenvectors and contribution values ​​of the three matrices into a pre-trained external wall insulation board bonding effect analysis model; and outputting a detection result matrix through the external wall insulation board bonding effect analysis model.

[0007] Specifically, spatiotemporal registration involves aligning the AR glasses image data with the ultrasonic testing equipment data in both the time dimension and the spatial coordinate system, ensuring that the two data sources describe the same physical location information.

[0008] The optimal observation point vector group is specifically a set of vectors calculated by the data fusion algorithm. Each vector contains two coordinate points, which represent the optimal observation point coordinates of the AR glasses and the optimal detection point coordinates of the ultrasonic detection device required to obtain the best detection results.

[0009] The stability index is a dimensionless parameter that measures the stability of the changing trends of elements in the detection data matrix. It is used to distinguish stable components from components with different degrees of variation in the detection data matrix.

[0010] Specifically, the basic matrix is ​​a sub-matrix representing the basic structural information in the detection data matrix, reflecting the basic bonding state of the external wall insulation board; the first variation matrix is ​​a sub-matrix representing the variation information in the detection data matrix where the stability index is higher than a preset threshold, reflecting the main variation characteristics in the bonding state of the external wall insulation board; and the second variation matrix is ​​a sub-matrix representing the variation information in the detection data matrix where the stability index is lower than a preset threshold, reflecting the secondary variation characteristics in the bonding state of the external wall insulation board.

[0011] The contribution value is specifically a numerical value that quantifies the influence of the basic matrix, the first variation matrix, and the second variation matrix on the final detection result. It is calculated through normalization and has a total value of 1.

[0012] The bonding effect index is a dimensionless parameter that quantifies the degree of bonding between the external wall insulation board and the wall. 1 indicates complete bonding, and 0 indicates complete detachment.

[0013] The contribution values ​​of the three matrices are calculated using a matrix contribution optimization function. The inputs include the detection data matrix, the basic matrix, the first variation matrix, the second variation matrix, and the ambient temperature parameter. This function is used to dynamically adjust the weight distribution of the basic matrix, the first variation matrix, and the second variation matrix in the final result, and the output is the contribution value.

[0014] The specific structure of the external wall insulation board bonding effect analysis model is a hybrid neural network based on multilayer perceptron and cross-attention mechanism. It includes a feature extraction layer, two cross-attention layers and three fully connected layers. The cross-attention layer uses a sparse attention mechanism to perform weighted analysis on the feature vectors of the input basic matrix, the feature vectors of the first and second changing matrices. In the sparse attention mechanism, the number of attention heads is dynamically adjusted according to the preset threshold of the stability index. The higher the preset threshold of the stability index, the more attention heads there are. The output layer of the external wall insulation board bonding effect analysis model uses a softmax activation function to map the result to the interval between 0 and 1 as the bonding effect index.

[0015] The present invention also provides AR glasses for detecting the bonding effect of external wall insulation boards, comprising a processor for performing the above-described method.

[0016] Compared with existing technologies, this invention provides a method and AR glasses for detecting the bonding effect of external wall insulation boards. This invention proposes a method and AR glasses for detecting the bonding effect of external wall insulation boards. By using spatiotemporal registration technology to accurately align AR image data with ultrasonic detection data, an optimal observation point vector group is generated to guide the detection position, realizing the effective fusion and collaborative detection of multi-source data.

[0017] This method employs matrix decomposition and feature extraction techniques, decomposing the detection data matrix into a basic matrix, a first variation matrix, and a second variation matrix, which respectively characterize the basic state and varying degrees of change in the adhesion of the insulation board. The weights of each matrix are dynamically adjusted through a matrix contribution optimization function, taking into account the influence of external factors such as ambient temperature on the detection results, thus improving the accuracy and adaptability of the results. Simultaneously, based on a hybrid neural network model using a multilayer perceptron and cross-attention mechanism, a sparse attention mechanism is used to perform weighted analysis on the feature vectors, quantifying the adhesion effect into an exponential value ranging from 0 to 1.

[0018] Compared with traditional technologies, this invention achieves precise guidance of the detection process, intelligent data analysis, and standardized quantification of results. It solves the technical problem that the bonding effect of external wall insulation boards cannot be accurately and efficiently detected in existing technologies, and provides more efficient and reliable technical support for the quality control of external wall insulation systems. Attached Figure Description

[0019] Figure 1A flowchart of the method provided for the first aspect of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0021] like Figure 1 The diagram shown is a flowchart of a method for detecting the bonding effect of external wall insulation boards according to the first aspect of the present invention. This method includes the following steps: S01. Collect AR glasses image data and obtain ultrasonic testing equipment detection data, and perform spatiotemporal registration of the AR glasses image data and the ultrasonic testing equipment detection data. S02. Perform fusion analysis on the spatiotemporal registered data to generate an optimal observation point vector group, which includes the optimal observation point coordinates of the AR glasses and the optimal detection point coordinates of the ultrasonic detection device. S03. Guide the inspectors to adjust the AR glasses to the optimal observation point coordinate position of the AR glasses in sequence, and at the same time align the ultrasonic testing equipment with the optimal detection point coordinate position of the ultrasonic testing equipment. S04. Collect data at the optimal observation point coordinates of the AR glasses and the optimal detection point coordinates of the ultrasonic detection device to form a detection data matrix; S05. Based on the stability index, the detection data matrix is ​​decomposed into a basic matrix, a first variation matrix, and a second variation matrix; S06. The contribution values ​​of the basic matrix, the first variation matrix, and the second variation matrix are calculated using a matrix contribution optimization function. The inputs of the matrix contribution optimization function include the detection data matrix, the basic matrix, the first variation matrix, the second variation matrix, and the ambient temperature parameter. S07. Calculate the eigenvectors of the basic matrix, the first transformation matrix, and the second transformation matrix respectively to obtain the eigenvectors of the basic matrix, the first transformation matrix, and the second transformation matrix. S08. Input the basic matrix eigenvector, the first variable matrix eigenvector, the second variable matrix eigenvector, and the contribution value into the pre-trained external wall insulation board bonding effect analysis model. S09. Output the detection result matrix through the external wall insulation board bonding effect analysis model. The elements in the detection result matrix are the bonding effect index of the external wall insulation board at the corresponding coordinate point. The bonding effect index ranges from 0 to 1.

[0022] Specifically, spatiotemporal registration involves aligning the AR glasses image data with the ultrasonic detection equipment detection data in both the time dimension and the spatial coordinate system, ensuring that the two data sources describe the same physical location information.

[0023] The optimal observation point vector group is specifically a set of vectors calculated by a data fusion algorithm. Each vector contains two coordinate points, which represent the optimal observation point coordinates of the AR glasses and the optimal detection point coordinates of the ultrasonic detection device required to obtain the best detection results.

[0024] The stability index is a dimensionless parameter that measures the stability of the changing trends of elements in the detection data matrix, and is used to distinguish stable components from components with different degrees of variation in the detection data matrix.

[0025] Specifically, the basic matrix is ​​a sub-matrix representing the basic structural information in the detection data matrix, reflecting the basic bonding state of the external wall insulation board.

[0026] Specifically, the first variation matrix represents the variation information submatrix in the detection data matrix where the stability index is higher than a preset threshold, reflecting the main variation characteristics in the bonding state of the external wall insulation board.

[0027] Specifically, the second variation matrix represents a sub-matrix of variation information in the detection data matrix where the stability index is lower than the preset threshold, reflecting secondary variation characteristics in the bonding state of the external wall insulation board.

[0028] The contribution value is specifically a numerical value that quantifies the influence of the basic matrix, the first variation matrix, and the second variation matrix on the final detection result. It is calculated through normalization and the sum is 1.

[0029] The bonding effect index is a dimensionless parameter that quantifies the degree of bonding between the external wall insulation board and the wall. 1 indicates complete bonding, and 0 indicates complete detachment.

[0030] Specifically, the ambient temperature parameter is the real-time temperature value of the detected environment collected by a temperature sensor, which is used to adjust the weights of different matrices in the matrix contribution optimization function.

[0031] The matrix contribution optimization function is used to dynamically adjust the weight distribution of the base matrix, the first variation matrix, and the second variation matrix in the final result. The inputs include the detection data matrix, the base matrix, the first variation matrix, the second variation matrix, and the ambient temperature parameter, and the output is the contribution value.

[0032] The preset threshold is a critical value determined based on statistical analysis of historical detection data, used to divide the first variation matrix and the second variation matrix.

[0033] The specific structure of the external wall insulation board bonding effect analysis model is a hybrid neural network based on a multilayer perceptron and a cross-attention mechanism. It includes a feature extraction layer, two cross-attention layers, and three fully connected layers. The cross-attention layers use a sparse attention mechanism to perform weighted analysis on the input feature vectors of the basic matrix, the first mutated matrix, and the second mutated matrix. In the sparse attention mechanism, the number of attention heads is dynamically adjusted according to a preset threshold of the stability index. The higher the preset threshold of the stability index, the more attention heads there are. The output layer of the external wall insulation board bonding effect analysis model uses a softmax activation function to map the results to the interval between 0 and 1 as the bonding effect index.

[0034] The steps for establishing the training dataset for the external wall insulation board bonding effect analysis model specifically include collecting AR image data and ultrasonic detection data from construction sites of external wall insulation boards for various types of buildings; standardizing and enhancing the collected data to improve the generalization ability of the external wall insulation board bonding effect analysis model; having engineering experts manually annotate the bonding effect of the insulation boards as supervised learning labels; and dividing the dataset into training, validation, and test sets according to factors such as building type, environmental conditions, and insulation board material to ensure that the data is evenly distributed and representative.

[0035] The specific steps for training the external wall insulation board bonding effect analysis model include: using the cross-entropy loss function to measure the difference between the predicted value and the label value of the external wall insulation board bonding effect analysis model; using the adaptive moment estimation optimization algorithm to iteratively update the parameters of the external wall insulation board bonding effect analysis model; introducing an early stopping mechanism during training to avoid overfitting; dynamically adjusting the parameter update step size through a learning rate scheduler; periodically evaluating the performance of the external wall insulation board bonding effect analysis model on the validation set and saving the best model parameters; and terminating the training process when the performance of the external wall insulation board bonding effect analysis model on the validation set no longer improves.

[0036] The specific implementation of the above steps is described in detail below. Step S01 involves first acquiring high-definition image data of the surface of the external wall insulation board using the built-in binocular camera of the AR glasses, and simultaneously using an ultrasonic detection device to emit ultrasonic waves and receive reflected waves, recording waveform data. The image data acquired by the AR glasses includes RGB three-channel information and depth information, with each pixel corresponding to a three-dimensional spatial coordinate. The data acquired by the ultrasonic detection device includes waveform amplitude, frequency, and phase information. The spatiotemporal registration process employs a bidirectional Hough transform algorithm. First, feature points are extracted from both sets of data, establishing a correspondence between them. Then, a rigid body transformation matrix maps the two sets of data to a unified reference coordinate system. Finally, timestamp alignment ensures the synchronization of the two sets of data. The main purpose of spatiotemporal registration is to establish a precise correspondence between the AR glasses' visual data and the ultrasonic detection data, providing a foundation for subsequent fusion analysis. The registration accuracy can reach sub-millimeter level, and the time synchronization error is controlled within 10 milliseconds.

[0037] Step S02 involves applying a multimodal data fusion algorithm to the registered data, comprising two stages: low-level feature fusion and high-level semantic fusion. Low-level feature fusion uses wavelet transform to extract frequency domain features from both types of data, while high-level semantic fusion uses a deep convolutional neural network to extract semantic information. The fusion process is implemented using a variational autoencoder with an attention mechanism, adaptively weighting the importance of different modalities. The optimal observation point vector set is generated using gradient descent, iteratively calculating the best observation position with information entropy and detection accuracy as optimization objectives. The algorithm input includes the fused feature map and environmental constraint parameters, and the output is a vector set containing multiple coordinate pairs, each representing the optimal position of the AR glasses and the ultrasonic detection device. The purpose of this step is to determine the spatial position combination that yields the best detection results, improving the accuracy and reliability of subsequent detection.

[0038] The specific implementation of step S03 involves converting the optimal observation point vector group data into visual guidance information for the AR glasses display system. This information is then presented in the AR glasses' field of view using a semi-transparent overlay method, displaying the navigation path and target location. The visual guidance employs real-time rendering technology, using different colors to indicate the deviation between the current position and the target position: green indicates a suitable position, yellow indicates close but requiring fine-tuning, and red indicates a large deviation. Simultaneously, the ultrasonic testing equipment's display screen also shows corresponding positioning guidance information, including directional arrows and distance values, helping testing personnel to accurately adjust the equipment's position. This step, through human-computer interaction, guides testing personnel to the calculated optimal testing position sequentially, ensuring the accuracy and consistency of data collection and providing high-quality raw data for subsequent analysis. Positioning accuracy is required to be within ±2 mm, and angular error must be controlled within ±1 degree.

[0039] The specific implementation of step S04 involves synchronously triggering data acquisition after confirming that both the AR glasses and the ultrasonic detection device have reached their optimal positions. The AR glasses acquire high-resolution image data, including visible light and depth images, and perform real-time feature extraction, including edge features, texture features, and color features. The ultrasonic detection device acquires complete waveform data, including A-mode and B-mode scan data, and simultaneously records device parameters such as frequency, gain, and depth of focus. During the acquisition process, a window filtering algorithm is applied to preprocess the raw data to eliminate noise interference. All acquired data are arranged according to a preset spatial sampling grid to form a multi-dimensional detection data matrix. Each element of the detection data matrix contains position coordinates and a corresponding detection feature vector. The purpose of this step is to obtain high-quality raw detection data, providing a complete information foundation for subsequent analysis. The data sampling rate is no less than 20 frames per second, and the sampling point spacing is controlled within 10 millimeters.

[0040] The specific implementation of step S05 is as follows: First, the stability index of each element in the detection data matrix is ​​calculated. The stability index is determined by the sliding window variance analysis method, and the calculation window size is [missing information]. The stability index ranges from 0 to 1, with higher values ​​indicating more stable data. Using a stability index of 0.75 as a preset threshold, singular value decomposition is performed on the detected data matrix to extract the basic structure corresponding to the main singular values. The basic matrix consists of the following... The singular vectors corresponding to the largest singular values ​​are formed, where The energy retention ratio is determined, typically set to 80% of the total energy. The first variation matrix consists of variation components with a stability index greater than a preset threshold, reflecting the main variation characteristics. The second variation matrix consists of variation components with a stability index less than the preset threshold, mainly containing secondary variation characteristics and noise. The matrix decomposition process employs an iterative regularization method to ensure that the sum of the three sub-matrices can accurately reconstruct the original detection data matrix. The purpose of this step is to decompose the complex detection data into components with different physical meanings, facilitating subsequent targeted analysis.

[0041] The specific implementation of step S06 involves constructing a temperature-compensated matrix contribution optimization function. This function employs an adaptive weighting method to dynamically adjust the importance of different sub-matrices. First, a nonlinear regression method is used to establish the correspondence between ambient temperature and material properties, considering the temperature range... to Then, the similarity between the basic matrix, the first and second transformation matrices, and the original detection data matrix is ​​calculated using the cosine similarity index. The optimization function comprehensively considers similarity, the temperature influence factor, and prior knowledge, and solves for the optimal contribution value using the Lagrange multiplier method. The temperature influence factor is obtained through fitting experimental data; for every increase in temperature... The contribution value of the basic matrix increases by approximately 0.05, the contribution value of the first modified matrix decreases by approximately 0.03, and the contribution value of the second modified matrix decreases by approximately 0.02. The calculated contribution values ​​are normalized to ensure that the sum is 1. The purpose of this step is to determine the influence weight of different sub-matrices in the final detection result, thereby improving the adaptability of the detection result to environmental conditions.

[0042] The specific implementation of step S07 involves applying the eigenvalue decomposition algorithm to the basic matrix, the first transformation matrix, and the second transformation matrix respectively to calculate their respective eigenvalues ​​and eigenvectors. The eigenvalue decomposition process employs a power-law iteration method, with an iteration precision set to... For each matrix, eigenvectors of the largest eigenvalues ​​are selected to form an eigenvector set. The number of eigenvectors is determined by the cumulative contribution rate, typically retaining those with a cumulative contribution rate of 90%. The extracted eigenvectors are then orthogonalized to ensure mutual orthogonality. Next, dimensionality reduction is performed on the eigenvectors using principal component analysis (PCA) to map them to a lower-dimensional feature space, retaining at least 95% of the information from the original data. Finally, the dimensionality-reduced eigenvectors are standardized to a norm of 1. The purpose of this step is to extract the main features of each submatrix, reduce data dimensionality, and provide an efficient feature representation for subsequent model analysis.

[0043] The specific implementation of step S08 involves taking the eigenvectors of the basic matrix, the first and second eigenvectors of the changing matrix obtained in step S07, and the contribution value calculated in step S06 as input, and feeding them into the pre-trained external wall insulation board bonding effect analysis model. The input data first undergoes batch normalization to ensure consistent numerical distribution. The model employs a hybrid neural network structure based on a multilayer perceptron and cross-attention mechanism, containing multiple functional modules. After data input, preprocessing is first performed through a batch normalization layer, followed by deep feature extraction through a feature extraction layer. Then, two cross-attention layers perform weighted fusion of features from different sources, and finally, three fully connected layers implement nonlinear mapping and decision-making. A residual connection structure is incorporated within the model to mitigate the gradient vanishing problem. The purpose of this step is to utilize the pre-trained model to comprehensively analyze the extracted features and generate a prediction result for the insulation board bonding effect.

[0044] The specific implementation of step S09 involves receiving the output vector of the external wall insulation board bonding effect analysis model and reshaping it into a two-dimensional result matrix corresponding to the spatial distribution of the detection area. Each element in the result matrix represents the bonding effect index at the corresponding location, with a value ranging from 0 to 1. A bilinear interpolation algorithm is used to interpolate and expand the sparse detection points, generating a complete bonding effect distribution map. Gaussian filtering is then applied for smoothing to reduce the influence of local noise. A color mapping scheme is set according to the bonding effect index: an index value greater than 0.8 is displayed in green indicating good bonding, 0.5 to 0.8 in yellow indicating moderate bonding, and less than 0.5 in red indicating poor bonding. The final detection results are overlaid in the AR glasses' field of view in an augmented reality manner, intuitively showing the bonding status of the external wall insulation board. A detection report is also generated, including statistical data and risk assessment. The purpose of this step is to transform the model analysis results into intuitive and understandable visual information, helping inspectors accurately judge the bonding quality of the external wall insulation board.

[0045] The detailed structure of the external wall insulation board bonding effect analysis model adopts a hybrid neural network based on a multilayer perceptron and a cross-attention mechanism. The feature extraction layer uses a fully connected structure, with the input dimension being the total dimension of the feature vectors and the output dimension being 512, using the ReLU activation function. The two cross-attention layers employ a multi-head self-attention mechanism, with the number of attention heads dynamically adjusted according to a preset stability index threshold. When the preset stability index threshold is 0.75, the number of attention heads is set to 8; for every 0.05 increase in the threshold, the number of attention heads increases by 1. The first cross-attention layer primarily handles the interaction between the eigenvectors of the basic matrix and the eigenvectors of the first transformed matrix, while the second cross-attention layer handles the interaction between the output of the first layer and the eigenvectors of the second transformed matrix. Attention calculation uses a scaled dot product attention mechanism, with the scaling factor set to the square root of the attention head dimension. The three fully connected layers in the model sequentially reduce the feature dimension from 512 to 256, 128, and 64, respectively, followed by a batch normalization layer and a Dropout layer, with the Dropout probability set to 0.3. The output layer is a fully connected layer, with the output dimension matching the number of discrete sampling points in the detection region. A softmax activation function is used to map the output to the 0-1 interval as the bonding effect index. Weight regularization is employed throughout the network training process to prevent overfitting, with a regularization coefficient set to 0.001.

[0046] The training dataset was built by collecting data samples from 50 building projects across different climates and building types, covering cold, temperate, and subtropical regions. Each project collected 30 to 50 sets of data. The collected raw data underwent standardization to eliminate the influence of equipment differences. Data augmentation techniques included random rotation. The dataset was expanded to three times its original size by adjusting the scaling factor (±10%) and adding Gaussian noise. Data annotation was performed by engineering experts with at least five years of experience, and the actual bonding effect was determined as the label value through actual peel tests. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio, ensuring a balanced distribution of different building types, environmental conditions, and insulation board materials across the subsets.

[0047] The training process of the model for analyzing the bonding effect of external wall insulation boards adopts a phased training strategy. In the initial stage, a relatively large learning rate of 0.001 is used to quickly search the parameter space; in the mid-stage, the learning rate is reduced to 0.0001 to fine-tune the parameters; in the later stage, the learning rate is further reduced to 0.00001 to achieve model convergence. A weighted cross-entropy loss function is used, with increased weights for samples with poor bonding effects to improve the model's sensitivity to problem regions. The optimization algorithm uses an adaptive moment estimation algorithm with momentum, setting the momentum parameter to 0.9 and the decay rate parameter to 0.999. An early stopping mechanism is introduced during training, stopping training when the loss value on the validation set no longer decreases for 10 consecutive epochs. The learning rate scheduling uses a cosine annealing strategy with hot restart, with an initial period of 5 epochs and a maximum period of 20 epochs. The evaluation metrics for the model on the validation set include mean absolute error, root mean square error, and F1 score, with the F1 score as the primary reference metric for selecting the optimal model parameters. The training equipment uses a high-performance computing platform equipped with GPU acceleration, with a batch size of 64 and a total training epoch count of 100.

[0048] The AR glasses provided in the second aspect of this invention mainly consist of a binocular camera, a depth sensor, a processor, a wireless communication module, a transparent display screen, a head-mounted support, and a rechargeable battery. The binocular camera captures stereoscopic images with a resolution of 4K. The depth sensor uses structured light technology, achieving a measurement accuracy of 1 millimeter. The processor uses a high-performance ARM architecture, supporting real-time image processing and augmented reality rendering. The wireless communication module supports WiFi and Bluetooth protocols, enabling data exchange with other devices. The transparent display screen uses optical waveguide technology, with a resolution of 1920×1080 and a field of view of 52 degrees.

[0049] The main components of the ultrasonic testing equipment include an ultrasonic transmitter, a receiver, a signal processing unit, an LCD screen, a touch control panel, and a portable power supply. The ultrasonic transmitter has a frequency range of 20kHz to 100kHz and adjustable transmission power. The receiver uses a high-sensitivity piezoelectric element with a signal-to-noise ratio greater than 40dB. The signal processing unit includes an analog-to-digital converter, a digital signal processor, and a data storage module, with a sampling rate of 100MHz and a bit depth of 12 bits. The LCD screen is 7 inches in size with a resolution of 1024×768. The touch control panel supports multi-touch for easy parameter adjustment and operation control. The portable power supply uses a lithium-ion battery with a battery life of more than 8 hours.

[0050] The mathematical model or calculation process involved in this invention will be described in detail below.

[0051] The spatiotemporal registration process in step S01 involves the bidirectional Hough transform algorithm, which achieves registration of the two types of data through feature point matching and coordinate transformation. The specific calculation formula for spatiotemporal registration is as follows: First, the feature point extraction and description in AR glasses image data can be represented as: ; In the formula, The set of feature points in the AR glasses image; For the first A descriptor for each feature point, containing location and descriptive information; This represents the total number of feature points.

[0052] Feature point extraction and description in ultrasonic testing equipment data can be represented as follows: ; In the formula, This is a set of feature points for ultrasonic detection. For the first Descriptor of a feature point; This represents the total number of feature points.

[0053] The similarity matrix between feature points is calculated as follows: ; In the formula, For the first The AR feature point and the first Similarity between ultrasonic feature points; This represents the vector dot product operation; Represents the vector norm.

[0054] Based on the similarity matrix, establish the correspondence between feature points: ; In the formula, It is the set of correspondences between feature points; This is the similarity threshold, and its value range is... to The default value is .

[0055] Calculate the rigid body transformation matrix based on the correspondence: ; In the formula, The transformation matrix; It is a rotation matrix; It is a translation vector; AR feature points Spatial coordinates; Ultrasonic feature points Spatial coordinates.

[0056] Rotation matrix Translation vector The solution is obtained using the singular value decomposition method: ; ; ; ; In the formula, It is the covariance matrix; and These are the average coordinates of AR feature points and ultrasonic feature points, respectively; This represents singular value decomposition; , , This is the result of the decomposition.

[0057] Time synchronization alignment is achieved through timestamp interpolation: ; In the formula, The synchronized ultrasound data; and For timestamps and The original ultrasonic data at the location; These are the interpolation coefficients; For AR data timestamps, .

[0058] Registration error assessment formula: ; In the formula, Registration error; This represents the number of correspondences between feature points.

[0059] The multimodal data fusion algorithm in step S02 includes wavelet transform and a deep convolutional neural network, and data fusion is achieved through a variational autoencoder. The calculation of the optimal observation point vector set involves iterative calculation using the gradient descent method. The specific formula is as follows: Wavelet transform extracts frequency domain features: ; ; In the formula, and Wavelet coefficients for AR data and ultrasonic data, respectively; and These are AR data and ultrasound data, respectively. The conjugate of the wavelet basis functions; This is a scale parameter, and its value range is... to ; These are the translation parameters.

[0060] Feature fusion weight calculation: ; ; In the formula, and The fusion weights for AR data and ultrasonic data are respectively; and The variances are represented by AR data and ultrasound data, respectively.

[0061] Fusion feature calculation: ; In the formula, Features after fusion; and Features extracted from AR data and ultrasonic data, respectively.

[0062] Information entropy calculation: ; In the formula, Information entropy; For position The probability distribution at a given point is obtained from the normalized eigenvectors.

[0063] Objective function of the optimal observation point vector set: ; In the formula, The objective function is... For the observation point vector group; Information entropy; For prediction accuracy; For environmental constraints; , and These are the weighting coefficients, with values ​​ranging from 1 to 10. , and The default values ​​are respectively , and .

[0064] Gradient descent update formula: ; In the formula, and The first Second and third The set of observation point vectors for each iteration; Let be the learning rate, and let the initial value be . The decay rate increases with the number of iterations. For the objective function with respect to The gradient.

[0065] Step S05 involves the calculation of the stability index and matrix decomposition, which employs sliding window variance analysis and singular value decomposition algorithms. The specific formulas are as follows:

[0066] Stability index calculation: ; In the formula, For position The stability index at the location; For position Centered The variance of the data within the window; This represents the maximum variance across all windows.

[0067] Singular value decomposition of the detection data matrix: ; In the formula, For the detection data matrix; and These are the left singular vector matrix and the right singular vector matrix, respectively; It is a singular value diagonal matrix.

[0068] Construction of the fundamental matrix: ; In the formula, The basic matrix; For the first One singular value; and The first There are one left singular vector and one right singular vector; The number of singular values ​​selected satisfies , This represents the total number of singular values.

[0069] Construction of the first and second transformation matrices: ; ; In the formula, This is the first transformation matrix; This is the second transformation matrix; and The positions of the detection data matrix and the fundamental matrix are respectively... The element value at that position; For position The stability index at the location; The preset threshold value is [value]. .

[0070] Iterative regularization methods ensure the accuracy of matrix decomposition: ; ; ; In the formula, Denotes the Frobenius norm; Represents the nuclear norm; express Norm; , and Here is the regularization coefficient, and its value range is as follows: , and The default values ​​are respectively , and .

[0071] The matrix contribution optimization function in step S06 involves nonlinear regression, cosine similarity calculation, and the Lagrange multiplier method. The specific formula is as follows: Calculation of temperature influence factor: ; In the formula, Temperature-related factors; This refers to the ambient temperature, expressed in degrees Celsius. , and The regression coefficients are obtained by fitting experimental data, and their value ranges are as follows: , and The default values ​​are respectively , and .

[0072] Cosine similarity calculation: ; ; ; In the formula, , and The cosine similarity between the basic matrix, the first and second transformation matrices and the original detection data matrix are respectively. Represents the inner product of matrices; This represents the Frobenius norm.

[0073] Matrix contribution optimization function: ; ; ; In the formula, Contribute optimization functions to the matrix; , and These are the contribution weights of the basic matrix, the first transformation matrix, and the second transformation matrix, respectively. , and These are the cosine similarities between the three matrices and the original data, respectively. This refers to the temperature-related factors.

[0074] Solving using the Lagrange multiplier method: ; In the formula, It is a Lagrange function; It is a Lagrange multiplier.

[0075] By solving Obtain the optimal contribution value , and ,satisfy: ; ; ; .

[0076] Step S07, eigenvalue decomposition and eigenvector calculation, involves power iteration, orthogonalization, and principal component analysis. The specific formulas are as follows: The power-law iteration method is used to calculate the largest eigenvalue and its corresponding eigenvector. Initial vector: Random initialization and ; Iteration formula: ; Convergence condition: ; Eigenvalue calculation: ; In the formula, The matrix to be decomposed (basic matrix, first transformation matrix, or second transformation matrix). and The first Second and third The eigenvectors of the next iteration; For the corresponding eigenvalues; express Norm.

[0077] Eigenvector orthogonalization (Gram-Schmidt method): ; ; ; In the formula, This is the original feature vector; The vector is orthogonalized; These are the standardized orthogonal eigenvectors; Represents the dot product of vectors; express Norm.

[0078] Feature vector selection criteria: ; In the formula, For the first One eigenvalue; The number of feature values ​​selected; This represents the total number of eigenvalues.

[0079] Principal component analysis dimensionality reduction: Calculate the covariance matrix: ; Eigenvalue decomposition: ; Dimensionality reduction transformation: ; In the formula, This is the original eigenvector matrix; It is the covariance matrix; The eigenvector matrix; It is an eigenvalue diagonal matrix; For the selected first A matrix composed of eigenvectors; The eigenvector matrix after dimensionality reduction; The dimension after dimensionality reduction should satisfy the requirement that the information retention rate is not less than [value missing]. .

[0080] The bilinear interpolation algorithm in step S09 is used to expand the sparse detection points and generate a complete bonding effect distribution map. The specific formula is as follows: Bilinear interpolation calculation: ; In the formula, For position The interpolation result at the location; , , and These are the bonding effect indices of four adjacent test points; , , and The coordinates of four adjacent detection points satisfy... and .

[0081] Gaussian filtering smoothing: ; ; In the formula, The Gaussian kernel function; The standard deviation is , and the range of values ​​is . The default value is ; The bonding effect index after smoothing; The original bonding effect index; The value is the radius of the filtering window, which takes the value of , This represents the function for rounding up.

[0082] The meanings and numerical choices of all variables and parameters in the above formulas are based on the physical characteristics and experimental data of the adhesion effect testing of external wall insulation boards. For example, the similarity threshold in spatiotemporal registration. Values The optimal threshold, determined based on extensive experimental data, ensures matching accuracy while avoiding over-screening that results in too few corresponding feature points. The temperature influence factor is expressed as a quadratic polynomial because the physical properties of external wall insulation materials typically exhibit a non-linear relationship with temperature, especially under low and high temperature conditions where it shows an accelerated trend; the quadratic term can better fit this relationship. Stability index threshold. Set as It is an empirical value derived from the analysis of the bonding state of insulation boards under various building types and environmental conditions, which can effectively distinguish between major and minor variation characteristics.

[0083] Specifically, the principle of this invention is based on the comprehensive application of multi-source data fusion, matrix factorization, and deep learning. First, spatiotemporal registration technology ensures precise alignment of AR glasses image data and ultrasonic testing equipment data in both the temporal and spatial coordinate systems, providing a unified data foundation for subsequent analysis. The optimal observation point vector set is calculated through a data fusion algorithm, achieving precise positioning of the testing equipment and resolving the data inconsistency problem caused by the randomness of location selection in traditional testing methods.

[0084] The matrix decomposition principle is the core of this invention. Based on the stability index, the detection data matrix is ​​decomposed into sub-matrices representing different levels of information. The basic matrix reflects the stable components of the basic bonding state, while the first and second variation matrices capture the primary and secondary variation features, respectively. This decomposition method conforms to the physical characteristics of the bonding state of external wall insulation boards, encompassing both stable structure and dynamic changes. The matrix contribution optimization function dynamically adjusts the weight distribution of each matrix by considering environmental temperature parameters, adapting to changes in the properties of insulation materials under different environmental conditions and improving the accuracy of the detection results.

[0085] The deep learning model design integrates the feature extraction capabilities of multilayer perceptrons and the information filtering advantages of cross-attention mechanisms. In particular, the sparse attention mechanism dynamically adjusts the number of attention heads based on a preset threshold of stability index, enabling the model to adaptively focus on features of varying importance. The softmax activation function maps the results to the 0-1 range, providing an intuitive index of the bonding effect. During model training, cross-entropy loss and adaptive moment estimation optimization algorithms are employed, along with early stopping mechanisms and a learning rate scheduler, ensuring the model possesses good generalization capabilities.

[0086] The organic combination of these technical principles enables this invention to overcome the limitations of traditional testing methods, achieve accurate and efficient testing of the bonding effect of external wall insulation boards, and meet the actual needs of building engineering quality control.

[0087] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0088] The spatiotemporal registration process in step S01 involves the bidirectional Hough transform algorithm, which achieves registration of the two types of data through feature point matching and coordinate transformation. The specific calculation formula for spatiotemporal registration is as follows: First, the feature point extraction and description in AR glasses image data can be represented as: ; In the formula, The set of feature points in the AR glasses image; For the first A descriptor for each feature point, containing location and descriptive information; This represents the total number of feature points.

[0089] Feature point extraction and description in ultrasonic testing equipment data can be represented as follows: ; In the formula, This is a set of feature points for ultrasonic detection. For the first Descriptor of a feature point; This represents the total number of feature points.

[0090] The similarity matrix between feature points is calculated as follows: ; In the formula, For the first The AR feature point and the first Similarity between ultrasonic feature points; This represents the vector dot product operation; Represents the vector norm.

[0091] Based on the similarity matrix, establish the correspondence between feature points: ;

[0092] In the formula, It is the set of correspondences between feature points; This is the similarity threshold, and its value range is... to The default value is .

[0093] Calculate the rigid body transformation matrix based on the correspondence: ; In the formula, The transformation matrix; It is a rotation matrix; It is a translation vector; AR feature points Spatial coordinates; Ultrasonic feature points Spatial coordinates.

[0094] Rotation matrix Translation vector The solution is obtained using the singular value decomposition method: ; ; ; ; In the formula, It is the covariance matrix; and These are the average coordinates of AR feature points and ultrasonic feature points, respectively; This represents singular value decomposition; , , This is the result of the decomposition.

[0095] Time synchronization alignment is achieved through timestamp interpolation: ; In the formula, The synchronized ultrasound data; and For timestamps and The original ultrasonic data at the location; These are the interpolation coefficients; For AR data timestamps, .

[0096] Registration error assessment formula: ; In the formula, Registration error; This represents the number of correspondences between feature points.

[0097] The multimodal data fusion algorithm in step S02 includes wavelet transform and a deep convolutional neural network, and data fusion is achieved through a variational autoencoder. The calculation of the optimal observation point vector set involves iterative calculation using the gradient descent method. The specific formula is as follows: Wavelet transform extracts frequency domain features: ; ; In the formula, and Wavelet coefficients for AR data and ultrasonic data, respectively; and These are AR data and ultrasound data, respectively. The conjugate of the wavelet basis functions; This is a scale parameter, and its value range is... to ; These are the translation parameters.

[0098] Feature fusion weight calculation: ; ; In the formula, and The fusion weights for AR data and ultrasonic data are respectively; and The variances are represented by AR data and ultrasound data, respectively.

[0099] Fusion feature calculation: ; In the formula, Features after fusion; and Features extracted from AR data and ultrasonic data, respectively.

[0100] Information entropy calculation: ; In the formula, Information entropy; For position The probability distribution at a given point is obtained from the normalized eigenvectors.

[0101] Objective function of the optimal observation point vector set: ; In the formula, The objective function is... For the observation point vector group; Information entropy; For prediction accuracy; For environmental constraints; , and These are the weighting coefficients, with values ​​ranging from 1 to 10. , and The default values ​​are respectively , and .

[0102] Gradient descent update formula: ; In the formula, and The first Second and third The set of observation point vectors for each iteration; Let be the learning rate, and let the initial value be . The decay rate increases with the number of iterations. For the objective function with respect to The gradient.

[0103] Step S05 involves the calculation of the stability index and matrix decomposition, which employs sliding window variance analysis and singular value decomposition algorithms. The specific formulas are as follows: Stability index calculation: ; In the formula, For position The stability index at the location; For position Centered The variance of the data within the window; This represents the maximum variance across all windows.

[0104] Singular value decomposition of the detection data matrix: ; In the formula, For the detection data matrix; and These are the left singular vector matrix and the right singular vector matrix, respectively; It is a singular value diagonal matrix.

[0105] Construction of the fundamental matrix: ; In the formula, The basic matrix; For the first One singular value; and The first One left singular vector and one right singular vector; The number of singular values ​​selected satisfies , This represents the total number of singular values.

[0106] Construction of the first and second transformation matrices: ; ; In the formula, This is the first transformation matrix; This is the second transformation matrix; and The positions of the detection data matrix and the fundamental matrix are respectively... The element value at that position; For position The stability index at the location; The preset threshold value is [value]. .

[0107] Iterative regularization methods ensure the accuracy of matrix decomposition: ; ; ; In the formula, Denotes the Frobenius norm; Represents the nuclear norm; express Norm; , and Here is the regularization coefficient, and its value range is as follows: , and The default values ​​are respectively , and .

[0108] The matrix contribution optimization function in step S06 involves nonlinear regression, cosine similarity calculation, and the Lagrange multiplier method. The specific formula is as follows: Calculation of temperature influence factor: ; In the formula, Temperature-related factors; This refers to the ambient temperature, expressed in degrees Celsius. , and The regression coefficients are obtained by fitting experimental data, and their value ranges are as follows: , and The default values ​​are respectively , and .

[0109] Cosine similarity calculation: ; ; ; In the formula, , and The cosine similarity between the basic matrix, the first and second transformation matrices and the original detection data matrix are respectively. Represents the inner product of matrices; This represents the Frobenius norm.

[0110] Matrix contribution optimization function: ; ; ; In the formula, Contribute optimization functions to the matrix; , and These are the contribution weights of the basic matrix, the first transformation matrix, and the second transformation matrix, respectively. , and These are the cosine similarities between the three matrices and the original data, respectively. This refers to the temperature-related factors.

[0111] Solving using the Lagrange multiplier method: ; In the formula, It is a Lagrange function; It is a Lagrange multiplier.

[0112] By solving Obtain the optimal contribution value , and ,satisfy: ; ; ; .

[0113] Step S07, eigenvalue decomposition and eigenvector calculation, involves power iteration, orthogonalization, and principal component analysis. The specific formulas are as follows: The power-law iteration method is used to calculate the largest eigenvalue and its corresponding eigenvector. Initial vector: Random initialization and ; Iteration formula: ; Convergence condition: ; Eigenvalue calculation: ; In the formula, The matrix to be decomposed (basic matrix, first transformation matrix, or second transformation matrix). and The first Second and third The eigenvectors of the next iteration; For the corresponding eigenvalues; express Norm.

[0114] Eigenvector orthogonalization (Gram-Schmidt method): ; ; ; In the formula, This is the original feature vector; The vector is orthogonalized; These are the standardized orthogonal eigenvectors; Represents the dot product of vectors; express Norm.

[0115] Feature vector selection criteria: ; In the formula, For the first One eigenvalue; The number of feature values ​​selected; This represents the total number of eigenvalues.

[0116] Principal component analysis dimensionality reduction: Calculate the covariance matrix: ; Eigenvalue decomposition: ; Dimensionality reduction transformation: ; In the formula, This is the original eigenvector matrix; It is the covariance matrix; The eigenvector matrix; It is an eigenvalue diagonal matrix; For the selected first A matrix composed of eigenvectors; The eigenvector matrix after dimensionality reduction; The dimension after dimensionality reduction should satisfy the requirement that the information retention rate is not less than [a certain value]. .

[0117] The bilinear interpolation algorithm in step S09 is used to expand the sparse detection points and generate a complete bonding effect distribution map. The specific formula is as follows: Bilinear interpolation calculation: ; In the formula, For position The interpolation result at the location; , , and These are the bonding effect indices of four adjacent test points; , , and The coordinates of four adjacent detection points satisfy... and .

[0118] Gaussian filtering smoothing: ; ; In the formula, The Gaussian kernel function; The standard deviation is , and the range of values ​​is . The default value is ; The bonding effect index after smoothing; The original bonding effect index; The value is the radius of the filtering window, which takes the value of , This represents the function for rounding up.

[0119] The meanings and numerical choices of all variables and parameters in the above formulas are based on the physical characteristics and experimental data of the adhesion effect testing of external wall insulation boards. For example, the similarity threshold in spatiotemporal registration. Values The optimal threshold, determined based on extensive experimental data, ensures matching accuracy while avoiding over-screening that results in too few corresponding feature points. The temperature influence factor is expressed as a quadratic polynomial because the physical properties of external wall insulation materials typically exhibit a non-linear relationship with temperature, especially under low and high temperature conditions where it shows an accelerated trend; the quadratic term can better fit this relationship. Stability index threshold. Set as It is an empirical value derived from the analysis of the bonding state of insulation boards under various building types and environmental conditions, which can effectively distinguish between major and minor variation characteristics.

[0120] To better understand and implement this invention, Example 2, a specific application scenario, is provided below: Researchers conducted a study on the bonding effect of external wall insulation boards in a 28-story high-rise residential building under construction. The building is located in a temperate climate region with an average winter temperature of 2°C and an average summer temperature of 28°C, resulting in a large annual temperature difference and high requirements for the bonding quality of the external wall insulation system. The building's external walls use 100mm thick graphite polystyrene boards as insulation material, bonded using a point-frame method, with a theoretical bonding area not less than 30% of the building code requirements. To verify the actual bonding effect, researchers applied the testing method of this invention to conduct sampling tests on the southeast-facing external walls of floors 5-10 of the building.

[0121] The testing equipment includes custom-developed AR glasses and an ultrasonic testing instrument. The AR glasses use a binocular camera with a resolution of 4096×2160, a frame rate of 30fps, a field of view of 60 degrees, and a built-in depth sensor with a measurement accuracy of ±0.5mm. The ultrasonic testing instrument uses a transmitting probe with a center frequency of 50kHz, a frequency range of 20-80kHz, a sampling rate of 100MHz, and an accuracy of ±0.1mm. The ambient temperature for testing is 24℃, and the relative humidity is 65%.

[0122] First, spatiotemporal registration was performed, extracting 358 feature points from the depth image acquired by the AR glasses and 264 feature points from the ultrasonic data. A feature point matching similarity threshold was then set. 147 pairs of valid feature point correspondences were obtained through similarity matrix calculation. Based on these correspondences, the rotation matrix was calculated. Translation vector Registration error The time synchronization error is 6.5ms.

[0123] By fusing multimodal data, wavelet transform is used to extract features, and scale parameters are set. Values arrive There are 5 scales in total. AR data weights are calculated based on variance. Ultrasonic data weighting Construct the objective function for the optimal observation point vector set, with parameters taking the values ​​of... , and Twenty-five optimal detection position combinations were calculated using the gradient descent method to enable detection in different wall areas.

[0124] In actual testing, data was collected sequentially at each selected point according to the calculated optimal position. The window size was set to [value missing] when calculating the stability index. Preset threshold After singular value decomposition, the first 12 singular values ​​of the detection data matrix were selected to construct the basic matrix, retaining 83.7% of the energy of the original data. The first and second transformation matrices captured 12.6% and 3.7% of the energy, respectively. The regularization coefficient for matrix decomposition was set to... , and .

[0125] When calculating the temperature influence factor, the regression coefficient is taken as: , and The contribution values ​​of the fundamental matrix are obtained through cosine similarity calculation and matrix contribution optimization function. Contribution value of the first change matrix Contribution value of the second change matrix .

[0126] Researchers used principal component analysis to reduce the dimensionality of the feature vectors to 64 dimensions, with an information retention rate of 97.2%. These dimensionality-reduced feature vectors, along with their contribution values, were input into a pre-trained model for analyzing the bonding effect of external wall insulation boards. The model was trained using a dataset containing 1850 labeled data sets from 46 building projects, and converged after 85 epochs, achieving an F1 score of 0.938 on the validation set.

[0127] The final detection results were expanded using bilinear interpolation. The grid is smoothed using Gaussian filtering with standard deviation. The statistical data of the test results are shown in Table 1: Table 1. Statistical Table of Adhesion Effect Test Results for External Wall Insulation Boards

[0128] Based on the test results, researchers found that areas with less than 30% of the insulation board bonding area accounted for 5.4% of the total tested area. These areas were mainly distributed around windows and wall corners. The specific reasons for these problematic areas are analyzed in Table 2. Table 2 Analysis of the causes of poor bonding areas

[0129] To verify the detection accuracy, researchers randomly selected 10 detection points for actual peel tests. The average error between the measured adhesive area and the predicted value was ±4.3%, with a maximum error of 7.8%, verifying the reliability of the detection method. The complete detection process took approximately 4.5 hours, significantly improving efficiency compared to the 2-3 days of traditional methods.

[0130] Traditional methods for testing the adhesion of external wall insulation boards primarily rely on tapping, pulling, and localized destructive testing. Tapping involves tapping the wall and listening to the sound to determine if there are hollow areas; this method is highly subjective, has low accuracy, and cannot quantitatively assess the bonded area. Pulling tests apply tension to the insulation board surface using a pull-out device, but only provides point-like information and cannot assess the overall adhesion. Localized destructive testing requires peeling off sections of the insulation board for inspection, which is highly destructive and unsuitable for large-scale application. These traditional methods share common problems: they cannot provide quantitative assessments of adhesion effectiveness, lack visual visualization, have low testing efficiency, and are significantly influenced by subjective factors.

[0131] Compared to traditional testing methods, the method of this invention has the following advantages: First, by fusing AR glasses with ultrasonic testing data, non-destructive and comprehensive testing is achieved; second, matrix decomposition and feature extraction techniques can distinguish between basic structures and changing features, improving the accuracy and robustness of the testing; third, a temperature compensation mechanism is introduced, making the test results highly adaptable to changes in environmental conditions; finally, the test results are visually displayed through augmented reality, facilitating understanding by construction personnel and enabling them to take targeted measures. Actual measurements show that the detection accuracy of this method reaches 95%, 30% higher than traditional methods, and the detection efficiency is more than 5 times higher, providing a powerful tool for the quality control of external wall insulation systems.

[0132] It should be noted that the variables involved in this invention are explained in detail in Tables 3, 4, and 5 below.

[0133] Table 3. Variable Explanation Table (Part 1)

[0134] Table 4. Variable Explanation Table (Part Two)

[0135] Table 5. Variable Explanation Table (Part 3)

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for testing the bonding effect of external wall insulation boards, characterized in that, include: AR glasses image data and ultrasonic testing equipment detection data are collected, and the image data and detection data are spatiotemporally registered and fused to generate the optimal observation point vector group. Inspectors are guided to adjust the AR glasses to their optimal observation point coordinates, while simultaneously aligning the ultrasonic testing equipment with its optimal detection point coordinates. Data is collected at both the AR glasses' optimal observation point coordinates and the ultrasonic testing equipment's optimal detection point coordinates, forming a detection data matrix. Based on the stability index, the detection data matrix is ​​decomposed into a basic matrix, a first variation matrix, and a second variation matrix, and the contribution values ​​and eigenvectors of the three matrices are calculated. The eigenvectors and contribution values ​​of the three matrices are then input into a pre-trained external wall insulation board bonding effect analysis model. The model outputs a detection result matrix. The specific structure of the external wall insulation board bonding effect analysis model is a hybrid neural network based on multilayer perceptron and cross-attention mechanism, which includes a feature extraction layer, two cross-attention layers and three fully connected layers. The cross-attention layer uses a sparse attention mechanism to perform weighted analysis on the feature vectors of the input basic matrix, the feature vectors of the first and second transformation matrices. When the stability index is preset to a threshold of 0.75, the number of attention heads is set to 8. For every 0.05 increase in the threshold, the number of attention heads increases by 1. The output layer of the external wall insulation board bonding effect analysis model uses a softmax activation function to map the results to the interval between 0 and 1 as the bonding effect index. The stability index is calculated as follows: ； In the formula, For position The stability index at the location; For position Centered The variance of the data within the window; This represents the maximum value of the variance across all windows. Singular value decomposition of the detection data matrix: ； In the formula, For the detection data matrix; and These are the left singular vector matrix and the right singular vector matrix, respectively; It is a singular value diagonal matrix; The basic matrix is ​​constructed as follows: ； In the formula, The basic matrix; For the first One singular value; and The first One left singular vector and one right singular vector; The number of singular values ​​selected satisfies , This represents the total number of singular values. The first and second transformation matrices are constructed as follows: ； ； In the formula, This is the first transformation matrix; This is the second transformation matrix; and The positions of the detection data matrix and the fundamental matrix are respectively... The element value at that position; For position The stability index at the location; This is a preset threshold.

2. The method for detecting the bonding effect of external wall insulation boards according to claim 1, characterized in that, Spatiotemporal registration specifically involves aligning AR glasses image data with ultrasonic testing equipment data in both the temporal and spatial coordinate systems to ensure that the two data sources describe information about the same physical location.

3. The method for detecting the bonding effect of external wall insulation boards according to claim 2, characterized in that, The optimal observation point vector set is a set of vectors calculated by the data fusion algorithm. Each vector contains two coordinate points, which represent the optimal observation point coordinates of the AR glasses and the optimal detection point coordinates of the ultrasonic detection device required to obtain the best detection results.

4. The method for detecting the bonding effect of external wall insulation boards according to claim 3, characterized in that, The basic matrix is ​​a sub-matrix representing the basic structural information in the detection data matrix, reflecting the basic bonding state of the external wall insulation board; the first variation matrix is ​​a sub-matrix representing the variation information in the detection data matrix where the stability index is higher than the preset threshold, reflecting the main variation characteristics in the bonding state of the external wall insulation board. The second variation matrix specifically represents the variation information submatrix in the detection data matrix where the stability index is less than or equal to a preset threshold, reflecting the secondary variation characteristics in the bonding state of the external wall insulation board.

5. The method for detecting the bonding effect of external wall insulation boards according to claim 4, characterized in that, The contribution value is specifically a numerical value that quantifies the influence of the basic matrix, the first variation matrix, and the second variation matrix on the final detection result. It is calculated through normalization and the sum is 1.

6. The method for detecting the bonding effect of external wall insulation boards according to claim 5, characterized in that, The bonding effect index is a dimensionless parameter that quantifies the degree of adhesion between the external wall insulation board and the wall. 1 indicates complete adhesion, and 0 indicates complete detachment.

7. The method for detecting the bonding effect of external wall insulation boards according to claim 6, characterized in that, The contribution values ​​of the three matrices are calculated using a matrix contribution optimization function. The inputs include the detection data matrix, the base matrix, the first variation matrix, the second variation matrix, and the ambient temperature parameter. This function dynamically adjusts the weight distribution of the base matrix, the first variation matrix, and the second variation matrix in the final result. The output is the contribution value. The matrix contribution optimization function is expressed as follows: ； ； ； In the formula, Contribute optimization functions to the matrix; , and These are the contribution weights of the basic matrix, the first transformation matrix, and the second transformation matrix, respectively. , and The cosine similarity between the basic matrix, the first transformation matrix, the second transformation matrix, and the detection data matrix are respectively. This refers to the temperature-related factors.

8. An AR glasses device for detecting the bonding effect of external wall insulation boards, characterized in that, It includes a processor for performing the method according to any one of claims 1-7.

Images