A method and system for detecting hollow walls based on multi-excitation response and spatial fusion

Abstract

This invention discloses a method and system for detecting wall hollowness based on multi-excitation response and spatial fusion, belonging to the field of non-destructive testing technology. The method includes: acquiring multi-type excitation response signals from multiple measuring points on the wall to be tested, including response signals under transient impact excitation and frequency sweep excitation; preprocessing the response signal of each measuring point and constructing a unified representation to obtain the frequency domain amplitude spectrum and time-frequency energy distribution; extracting multi-dimensional features of each measuring point and constructing a neighborhood set based on spatial coordinates; fusing the multi-dimensional features with neighborhood information to generate spatial enhancement features; using the measuring points as graph nodes and the adjacency relationships in the neighborhood set as graph edges, inputting the spatial enhancement features into a graph attention network, and regressing to output a continuous value of the hollowness degree of each measuring point, thus completing the hollowness assessment. This invention achieves stable detection and continuous quantitative assessment of wall hollowness.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing technology, specifically relating to a method and system for detecting hollow walls based on multi-excitation response and spatial fusion. Background Technology

[0002] Hollow spots in walls are a common defect in building quality inspection. If not addressed promptly, they can easily lead to tile detachment, plaster cracking, and even structural damage. Current detection methods mainly rely on manual tapping, where operators judge by listening to differences in the echoes. This method is highly subjective, has poor repeatability, and makes it difficult to quantitatively describe the extent and severity of hollow spots.

[0003] In recent years, some studies have attempted to use a single excitation (such as impact elastic waves or ultrasound) combined with signal processing technology for hollow wall identification. However, as a multi-layered heterogeneous structure, the response of the wall is affected by the coupling of multiple factors such as the finishing layer, mortar layer, and base layer. The single excitation method is difficult to fully excite the dynamic characteristics of the hollow area. Especially in lightweight block walls, the excitation energy is easily absorbed by the structure, the response signal is weak and easily drowned out by noise, resulting in insufficient detection sensitivity.

[0004] Meanwhile, most existing detection methods based on vibration or acoustic response analyze individual measurement points independently, ignoring the spatial correlation between these points—hollow areas often exhibit continuous distribution, and the response characteristics of adjacent measurement points have natural spatial consistency. Traditional point-by-point threshold judgment or single-point classification models cannot utilize this spatial neighborhood information, leading to problems such as isolated noise points or discontinuous boundaries in the detection results.

[0005] In addition, the actual field testing environment has interference factors such as inconsistent excitation, differences in sensor coupling, and local defects in the wall, making it difficult to stably characterize the hollow state with a single acquisition or a single feature. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a method and system for detecting wall hollowness based on multi-excitation response and spatial fusion. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a method for detecting wall hollowness based on multi-excitation response and spatial fusion, comprising:

[0008] Collect multi-type excitation response signals from multiple measuring points on the wall to be tested. The multi-type excitation response signals include at least the response signal under transient impact excitation and the response signal under frequency sweep excitation.

[0009] The multi-type excitation response signals at each measurement point are preprocessed and a unified characterization is constructed to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measurement point.

[0010] Based on the frequency domain amplitude spectrum and the time-frequency energy distribution, multidimensional features of each measurement point are extracted, and a neighborhood set is constructed according to the spatial coordinates of the measurement point. The multidimensional features are then fused with the neighborhood information to generate spatial enhancement features for each measurement point.

[0011] Using each measurement point as a graph node and the adjacency relationships in the neighborhood set as graph connection edges, the spatial enhancement features are input into the graph attention network as node features. The graph attention network aggregates neighbor node information and regresses to output a continuous value of the hollowness degree of each measurement point.

[0012] The degree of hollowness at each measuring point is evaluated based on the continuous values ​​of the hollowness degree.

[0013] In one embodiment of the present invention, the step of preprocessing the multi-type excitation response signals at each measuring point and constructing a unified characterization to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measuring point includes:

[0014] For each measurement point, the response signal acquired each time is sequentially subjected to bandpass filtering, baseline correction, amplitude normalization, and outlier removal to obtain the preprocessed effective signal;

[0015] Perform Fast Fourier Transform on multiple valid signals at the same measurement point to obtain the frequency domain amplitude spectrum of each sample, and then fuse the frequency domain amplitude spectra of each sample according to frequency points to obtain the frequency domain amplitude spectrum of the measurement point.

[0016] Perform short-time Fourier transform or continuous wavelet transform on multiple valid signals at the same measurement point to obtain multiple time-frequency energy distributions; fuse the multiple time-frequency energy distributions to obtain the time-frequency energy distribution of the measurement point.

[0017] In one embodiment of the present invention, the step of extracting multidimensional features of each measurement point based on the frequency domain amplitude spectrum and the time-frequency energy distribution includes:

[0018] Based on the frequency domain amplitude spectrum, frequency domain features and amplitude energy features of each measurement point are extracted; the frequency domain features include one or more of the following: dominant frequency, spectral centroid, and spectral bandwidth; the amplitude energy features include one or more of the following: peak amplitude and spectral energy.

[0019] Based on the time-frequency energy distribution, time-domain correlation features are extracted for each measurement point; the time-domain correlation features include one or more of echo delay and energy attenuation features.

[0020] Based on pre-constructed benchmark feature values, the difference features of each measurement point relative to the area without voids are extracted; the difference features include one or more of the following: main frequency shift, relative amplitude change, and energy difference;

[0021] The frequency domain features, amplitude energy features, time domain correlation features, and difference features are combined to obtain the multidimensional features of each measurement point.

[0022] In one embodiment of the present invention, the reference feature value is obtained by the following steps:

[0023] In a known area without voids, multiple reference measurement points are selected. Based on the frequency domain amplitude spectrum of each reference measurement point, the single-point reference main frequency, single-point reference peak amplitude, and single-point reference spectral energy of each reference measurement point are extracted respectively.

[0024] The reference frequency is obtained by averaging the single-point reference main frequency of all reference measurement points; the reference peak amplitude is obtained by averaging the single-point reference peak amplitude of all reference measurement points; and the reference spectral energy is obtained by averaging the single-point reference spectral energy of all reference measurement points.

[0025] The reference main frequency, the reference peak amplitude, and the reference spectral energy are used as the reference characteristic values.

[0026] In one embodiment of the present invention, the step of constructing a neighborhood set based on the spatial coordinates of the measuring point, fusing the multidimensional features with the neighborhood information, and generating spatially enhanced features for each measuring point includes:

[0027] Based on the spatial coordinates of the measurement points, a neighborhood set for each measurement point is constructed;

[0028] For each measurement point, calculate the mean of the multidimensional features of all measurement points in its neighborhood set to obtain the neighborhood mean feature of that measurement point.

[0029] For each measurement point, calculate the difference between the multidimensional features of that measurement point and the mean features of its neighborhood to obtain the feature difference of that measurement point;

[0030] The multidimensional features, neighborhood mean features, and feature differences of the measurement point are concatenated to obtain the spatial enhancement features of the measurement point.

[0031] In one embodiment of the present invention, the graph attention network includes:

[0032] The input layer receives the spatial augmentation features of each measurement point, forming a node feature matrix.

[0033] The normalization layer, connected after the input layer, is used to perform layer normalization on the spatial enhancement features of each measurement point and output a normalized feature matrix.

[0034] The first graph attention layer, connected after the normalization layer, is used to perform graph attention calculation on the normalized feature matrix using multiple attention heads. The outputs of the multiple attention heads are concatenated and then activated by the LeakyReLU activation function to obtain the first graph attention feature output by the first graph attention layer.

[0035] The second graph attention layer, connected after the first graph attention layer, is used to perform graph attention calculation on the attention features of the first graph using multiple attention heads, average the outputs of the multiple attention heads, and then activate them through the ELU activation function to obtain the attention features of the second graph.

[0036] The regression head, connected after the attention layer of the second graph, includes a fully connected layer and a sigmoid activation function, used to map the attention features of the second graph to normalized hollowness values;

[0037] The post-processing unit, connected after the regression head, is used to map the normalized hollowness value to a continuous range of 0~3 through scale transformation, so as to obtain a continuous hollowness value for each measuring point.

[0038] In one embodiment of the present invention, it further includes:

[0039] A training dataset is constructed using a preset detection region as samples; each sample includes: a node feature matrix, a neighborhood set, and a true hollowness label for each measurement point; the node feature matrix is ​​composed of the spatial enhancement features of each measurement point;

[0040] The graph attention network is trained using mean squared error as the loss function and minimizing the error between the true hollowness label and the predicted continuous hollowness value as the optimization objective.

[0041] In one embodiment of the present invention, the step of evaluating the degree of hollowness at each measuring point based on the continuous value of hollowness includes:

[0042] Obtain a preset set of discrete level thresholds, wherein the discrete level thresholds include a first threshold, a second threshold, and a third threshold, and the first threshold is less than the second threshold and the third threshold;

[0043] When the continuous value of the hollowness degree is less than or equal to the first threshold, the measuring point is determined to have no hollowness; when the continuous value of the hollowness degree is greater than the first threshold and less than or equal to the second threshold, the measuring point is determined to have mild hollowness; when the continuous value of the hollowness degree is greater than the second threshold and less than or equal to the third threshold, the measuring point is determined to have moderate hollowness; when the continuous value of the hollowness degree is greater than the third threshold, the measuring point is determined to have severe hollowness.

[0044] Secondly, the present invention also provides a wall hollowing detection system based on multi-excitation response and spatial fusion, comprising:

[0045] The acquisition module is used to acquire multi-type excitation response signals from multiple measuring points on the wall to be tested. The multi-type excitation response signals include at least the response signal under transient impact excitation and the response signal under frequency sweep excitation.

[0046] The characterization construction module is used to preprocess the multi-type excitation response signals of each measurement point and construct a unified characterization to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measurement point.

[0047] The feature enhancement module is used to extract multi-dimensional features of each measurement point based on the frequency domain amplitude spectrum and the time-frequency energy distribution, construct a neighborhood set according to the spatial coordinates of the measurement point, and fuse the multi-dimensional features with the neighborhood information to generate spatial enhancement features for each measurement point.

[0048] The graph attention regression module is used to take each measurement point as a graph node and the adjacency relationship in the neighborhood set as the graph connection edge. It inputs the spatial enhancement feature as node feature into the graph attention network, aggregates the neighbor node information through the graph attention network, and regresses and outputs the continuous value of the hollowness degree of each measurement point.

[0049] The hollowness assessment module is used to assess the degree of hollowness at each measuring point based on the continuous value of the hollowness degree.

[0050] Thirdly, the present invention also provides an electronic device, comprising:

[0051] processor;

[0052] Memory, used to store one or more programs;

[0053] Wherein, when the one or more programs are executed by the processor, the processor implements the method described in the first aspect.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] 1. This invention effectively suppresses the influence of inconsistent excitation, sensor coupling differences, and local interference on a single acquisition by collecting response signals from two types of excitation: transient impact excitation and frequency sweep excitation. It also performs multiple repeated acquisitions and signal fusion at each measuring point, thus solving the problem that a single excitation method cannot fully excite the characteristics of the hollow area and achieving stable and uniform characterization of the wall response signal. Compared to a single excitation method, this invention can obtain high-frequency response information from transient impact and broadband resonance characteristics from frequency sweep excitation. The two excitations complement each other, significantly improving the excitation efficiency and detection sensitivity of hollow features.

[0056] 2. This invention extracts multi-dimensional frequency domain, time domain, energy and difference features based on benchmark feature values, and constructs neighborhood mean and feature difference by combining the spatial coordinates of the measurement points to form spatial enhancement features. It effectively utilizes the spatial correlation between measurement points, significantly improves the regional continuity and boundary accuracy of hollow drum detection, and avoids the problem of isolated noise points caused by point-by-point judgment in the prior art.

[0057] 3. This invention introduces a graph attention network, which uses measurement points as nodes and spatial neighborhoods as edges. It adaptively aggregates neighbor information through an attention mechanism and outputs a hollowness value in a continuous range of 0 to 3 from end to end, which can effectively reflect the severity of hollowness.

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0059] Figure 1 This is a flowchart of a wall hollow detection method based on multi-excitation response and spatial fusion provided in an embodiment of the present invention;

[0060] Figure 2 This is another flowchart of the wall hollow detection method based on multi-excitation response and spatial fusion provided in the embodiments of the present invention;

[0061] Figure 3 This is a schematic diagram of the wall hollow detection system based on multi-excitation response and spatial fusion provided in an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0064] Figures 1-2 This is a flowchart of a wall hollow detection method based on multi-excitation response and spatial fusion provided in an embodiment of the present invention. Figures 1-2 As shown, this embodiment of the invention provides a method for detecting wall hollowness based on multi-excitation response and spatial fusion, including:

[0065] S1. Collect multi-type excitation response signals from multiple measuring points on the wall to be tested. The multi-type excitation response signals include at least the response signals under transient impact excitation and the response signals under frequency sweep excitation.

[0066] Among them, the multi-type excitation response signals include at least the response signals under transient impact excitation and the response signals under frequency sweep excitation.

[0067] Specifically, a regular grid of measuring points is constructed on the surface of the wall to be inspected. The spacing of the measuring point grid can be determined according to the requirements of inspection accuracy and efficiency: too small a spacing will result in too many measuring points and prolonged inspection time; too large a spacing will result in insufficient spatial resolution and difficulty in accurately locating the boundary of the hollow area.

[0068] For each measuring point, multiple types of excitation are applied and response signals are acquired. Transient impact excitation can be applied using an impact hammer. Under transient impact excitation, the high-frequency components in the response signal are more sensitive to small and surface voids. Frequency sweep excitation can be applied using a vibrator, and the frequency sweep range can be set according to the wall structure and detection depth requirements. Under frequency sweep excitation, the resonant frequency characteristics in the response signal are more sensitive to deep and large-area voids. Data is acquired multiple times (e.g., 5-10 times) at each measuring point to improve data robustness.

[0069] Taking the interior wall of a residential building as an example, its area is The surface is flat and without a finishing layer. Measuring points are evenly distributed in a grid at 0.2m intervals, totaling... There are 10 measuring points, and the coordinates of each measuring point are recorded. , where x is the horizontal coordinate (e.g., with the left edge of the wall as the origin and the rightward direction as the positive direction), and y is the vertical coordinate (e.g., with the bottom edge of the wall as the origin and the upward direction as the positive direction), so as to construct the spatial neighborhood in subsequent steps.

[0070] For transient impact excitation, a steel impact hammer with a diameter of 1.5 cm and a 2 mm thick rubber coating on the hammerhead is used. The impact point of the hammer is fixed at the center of the measuring point, and the impulse of each strike is controlled at 5 N·s. The force sensor integrated in the impact hammer outputs a trigger signal for the synchronous acquisition system's startup and triggering. For frequency sweep excitation, a miniature vibrator (model PCB086C03, frequency response range 10 Hz~3000 Hz) is used and attached 2 cm away from the measuring point. The vibrator is driven by a signal generator, with a frequency sweep range of 50 Hz~2000 Hz and a sweep period of 2 s. A logarithmic frequency sweep method is used, which is beneficial for fully exciting the low-frequency resonance characteristics of the wall.

[0071] Transient impact excitation and frequency sweep excitation were applied sequentially to each measuring point. The transient impact response was picked up using an accelerometer (sensitivity 100mV / g), and the frequency sweep response was synchronously acquired using the same accelerometer, with the accelerometer position consistent with that used for the transient impact response. The sampling rate was set to 10kHz, and the sampling duration was 3s. Each measuring point was repeatedly sampled 6 times for each type of excitation. It should be noted that a certain time interval was required between each two acquisitions to ensure that the wall vibration completely decayed.

[0072] S2. Preprocess the multi-type excitation response signals of each measuring point and construct a unified characterization to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measuring point.

[0073] In one embodiment, step S2 includes:

[0074] S21. For each measurement point, the response signal acquired each time is sequentially subjected to bandpass filtering, baseline correction, amplitude normalization, and outlier removal to obtain the preprocessed effective signal.

[0075] S22. Perform Fast Fourier Transform on multiple valid signals at the same measurement point to obtain the frequency domain amplitude spectrum of each sample, and fuse the frequency domain amplitude spectra of each sample according to frequency points to obtain the frequency domain amplitude spectrum of the measurement point.

[0076] S23. Perform short-time Fourier transform or continuous wavelet transform on multiple valid signals at the same measurement point to obtain multiple time-frequency energy distributions; fuse the multiple time-frequency energy distributions to obtain the time-frequency energy distribution of the measurement point.

[0077] In this embodiment, the response signals acquired at each measurement point are first preprocessed sequentially using bandpass filtering, baseline correction, amplitude normalization, and outlier removal. Bandpass filtering removes low-frequency drift and high-frequency noise, preserving signal components within the effective frequency band. Baseline correction eliminates the DC component in the signal, causing it to fluctuate around the zero baseline. Amplitude normalization normalizes the signal amplitudes from different acquisitions or measurement points to the same scale, eliminating the influence of amplitude variations caused by inconsistent striking force; maximum value normalization or energy normalization can be used. Outlier removal identifies and removes abnormal sampling signals caused by operational errors or sudden interference; common methods include median filtering and threshold discrimination methods (such as the 3σ criterion).

[0078] For example, bandpass filtering is performed on each acquired response signal. An FIR bandpass filter of order 100 is designed, with a passband range of 50Hz to 1800Hz. Then, the time mean of the filtered response signal is subtracted to eliminate the influence of the DC component on signal analysis. Energy normalization is used to normalize the energy of each baseline-corrected response signal to 1. Finally, outlier removal is performed: the mean and standard deviation of the peak amplitude of the six samples at the measurement point are calculated, and outliers deviating from the mean by more than three times the standard deviation are removed. Typically, the remaining valid data points K' are greater than or equal to 5. If the valid data for a measurement point is insufficient, the measurement point needs to be reacquired.

[0079] Furthermore, a Fast Fourier Transform is performed on multiple preprocessed valid signals at the same measurement point to obtain the frequency domain amplitude spectrum of each sample. The amplitude spectrum reflects the energy distribution of the signal at different frequency components. The frequency domain amplitude spectra of each sample at the same measurement point are then fused according to frequency points. The fusion method can be an arithmetic mean (with equal weights for each sample) or a weighted mean (with weights determined by the signal-to-noise ratio or signal energy) to obtain the frequency domain amplitude spectrum of that measurement point. Simultaneously, a Short-Time Fourier Transform or a Continuous Wavelet Transform is performed on multiple valid signals at the same measurement point to obtain the time-frequency energy distribution of each sample. The time-frequency energy distributions of multiple samples are then fused (e.g., by point-by-point averaging or energy accumulation) to obtain the time-frequency energy distribution of that measurement point.

[0080] For example, a 4096-point Fast Fourier Transform (FFT) is performed on each preprocessed effective signal segment. To reduce spectral leakage, a Hanning window is applied for windowing, and a 50% overlap rate is used to improve frequency resolution, thereby obtaining the amplitude spectrum. Then, the frequency domain amplitude spectra of all effective signals within the same measurement point are arithmetically averaged over frequency points to obtain the stable frequency domain response at that measurement point. At this point, the frequency resolution is approximately 2.44 Hz, and there are approximately 800 frequency points within the effective frequency band of 50 Hz to 1800 Hz.

[0081] For time-frequency characterization, Short-Time Fourier Transform (STFT) is used for time-frequency analysis: the window length is set to 256 points, and the overlap rate is set to 75%. After transformation, the time-frequency energy distribution is obtained. The time-frequency energy distribution of the same measurement point is obtained by averaging the time-frequency energy distributions sampled multiple times within the same measurement point.

[0082] After the above processing, two sets of unified response characterization data are generated for each measuring point, namely frequency domain amplitude spectrum data and time-frequency energy distribution data.

[0083] S3. Based on the frequency domain amplitude spectrum and time-frequency energy distribution, extract the multidimensional features of each measurement point, construct a neighborhood set according to the spatial coordinates of the measurement point, and fuse the multidimensional features with the neighborhood information to generate the spatial enhancement features of each measurement point.

[0084] First, frequency domain features and amplitude energy features are extracted based on the frequency domain amplitude spectrum. Frequency domain features include one or more of the following: dominant frequency, spectral centroid, and spectral bandwidth. Amplitude energy features include one or more of the following: peak amplitude and spectral energy. These features can characterize the response characteristics of the wall at the measuring point from different perspectives. In the hollow area, the vibration characteristics of the wall will change significantly due to the loss of the base layer constraint, mainly manifested as a decrease in dominant frequency, an increase in peak amplitude, and an increase in spectral energy.

[0085] Furthermore, based on the time-frequency energy distribution, time-domain correlation features of each measurement point are extracted, including one or more of echo delay and energy attenuation features. Specifically, echo delay refers to the time corresponding to the maximum energy peak in the time-frequency energy distribution, reflecting the propagation time of the excitation signal from the impact point to the sensor. In hollow areas, the echo delay is usually longer due to the reduced sound velocity or extended path. Energy attenuation features refer to the rate of energy attenuation of the signal over time. In hollow areas, the attenuation is usually faster due to increased energy dissipation.

[0086] Based on pre-constructed benchmark feature values, the difference features of each measuring point relative to the void-free region are extracted. In one embodiment, the benchmark feature values ​​can be obtained by the following steps: Select multiple benchmark measuring points in the known void-free region; based on the frequency domain amplitude spectrum of each benchmark measuring point, extract the single-point benchmark dominant frequency, single-point benchmark peak amplitude, and single-point benchmark spectral energy of each benchmark measuring point; take the average of the single-point benchmark dominant frequencies of all benchmark measuring points to obtain the benchmark dominant frequency; take the average of the single-point benchmark peak amplitudes of all benchmark measuring points to obtain the benchmark peak amplitude; take the average of the single-point benchmark spectral energy of all benchmark measuring points to obtain the benchmark spectral energy; use the benchmark dominant frequency, benchmark peak amplitude, and benchmark spectral energy as benchmark feature values.

[0087] In this embodiment, the areas without voids are first confirmed through manual tapping or destructive verification. Multiple reference measurement points are then selected from these known void-free areas. Based on the frequency domain amplitude spectra of these reference measurement points, the reference dominant frequency, reference peak amplitude, and reference spectral energy are calculated.

[0088] ;

[0089] ;

[0090] ;

[0091] In the formula, , , These represent the reference main frequency, reference peak amplitude, and reference spectral energy, respectively. Here is the index of the reference measurement point, the number of reference measurement points, and the main frequency of the reference measurement point. Peak amplitude of the benchmark measurement point Spectral energy of the benchmark measurement point , Indicates the reference measurement point The frequency domain amplitude spectrum.

[0092] Optionally, the difference characteristics include one or more of the following: dominant frequency offset, relative amplitude change, and energy difference. The dominant frequency offset is obtained by subtracting the reference dominant frequency from the dominant frequency at the measuring point. Hollow areas typically cause a decrease in dominant frequency, therefore the dominant frequency offset is negative. The relative amplitude change is obtained by dividing the peak amplitude at the measuring point by the peak amplitude at the reference point. The amplitude in hollow areas is usually higher, therefore the ratio of the peak amplitude at the measuring point to the peak amplitude at the reference point is greater than 1. The energy difference is obtained by subtracting the reference spectral energy from the spectral energy at the measuring point. Since the energy in hollow areas is usually increased, the difference is positive. These difference characteristics have good robustness and can eliminate reference differences under different wall types and different testing conditions.

[0093] By combining the above frequency domain features, amplitude energy features, time domain correlation features, and difference features, multidimensional features of each measurement point are obtained.

[0094] For example, eight benchmark points were selected near the four corners and center of the wall, verified by manual tapping to be free of hollow areas. These points are numbered i_ref = {5, 12, 48, 67, 105, 143, 182, 200}. Taking benchmark point i_ref = 5 as an example, the key data of its frequency domain amplitude spectrum in the range of 50~1500Hz are shown in Table 1:

[0095] Table 1

[0096] 200 0.12 300 0.35 400 0.82 500 0.41 600 0.18

[0097] The main frequency at reference measurement point i_ref = 5 can be calculated from the data shown in Table 1. Peak amplitude Spectral energy .

[0098] The above calculation process was repeated for all 8 benchmark points, and the benchmark characteristic values ​​of each benchmark point are shown in Table 2.

[0099] Table 2

[0100] 5 400 0.82 1.010 12 390 0.79 0.985 48 410 0.85 1.034 67 395 0.81 1.002 105 405 0.83 1.021 143 385 0.78 0.976 182 400 0.84 1.028 200 395 0.80 0.994

[0101] Furthermore, the reference dominant frequency was calculated based on the frequency domain amplitude spectrum of the eight reference measurement points: ,

[0102] Reference peak amplitude:

[0103] and reference spectrum energy: .

[0104] Furthermore, a neighborhood set is constructed based on the spatial coordinates of the measurement points. The spatial coordinates of the measurement points can be planar coordinates. It can also be three-dimensional coordinates. The neighborhood set is determined using either the K-nearest neighbor algorithm or a distance threshold algorithm. For example, when defining the neighborhood set using the K-nearest neighbor algorithm, the K nearest neighbors in terms of spatial Euclidean distance are selected, where K=4 (four neighbors) or K=8 (eight neighbors). When constructing the neighborhood using the distance threshold algorithm, all neighbors with a radius less than or equal to a preset threshold are selected. It should be noted that for boundary neighbors, the number of neighbors may be less than the number of internal neighbors. Therefore, the size of the neighborhood set should remain variable in subsequent calculations, and the mean should be calculated by dividing by the actual number of neighbors.

[0105] For each measuring point, the mean of the multidimensional features of all measuring points within its neighborhood set is calculated to obtain the neighborhood mean feature. The neighborhood mean feature reflects the overall trend of multidimensional features within the local region of that measuring point, which can smooth local noise and enhance regional consistency. Simultaneously, the difference between the multidimensional features of that measuring point and the neighborhood mean feature is calculated to obtain the feature difference. It should be understood that if the feature difference is positive, it indicates that the features of that measuring point are higher than the neighborhood mean, possibly corresponding to the center or boundary of a hollow area; conversely, if the feature difference is negative, it indicates that the features of that measuring point are lower than the neighborhood mean, possibly corresponding to the edge of a normal area or a hollow area.

[0106] Finally, the multidimensional features, neighborhood mean features, and feature differences of the measurement points are concatenated to form spatially enhanced features. In one specific implementation of this invention, the multidimensional features, neighborhood mean features, and feature differences are all 10-dimensional, resulting in a 30-dimensional spatially enhanced feature after concatenation. This spatially enhanced feature simultaneously includes single-point information, local trends, and local anomaly information.

[0107] S4. Using each measurement point as a graph node and the adjacency relationship in the neighborhood set as the graph connection edge, the spatial enhancement feature is input as the node feature into the graph attention network. The graph attention network aggregates the neighbor node information and regresses to output the continuous value of the hollowness degree of each measurement point.

[0108] Graph attention networks are a type of neural network used to process graph-structured data. They adaptively aggregate neighbor information by assigning different weights to different neighbors of each node through an attention mechanism.

[0109] In one embodiment, the graph attention network includes:

[0110] The input layer receives the spatial augmentation features of each measurement point, forming a node feature matrix. For example, the wall to be detected has... There are [number] measurement points, and the spatial augmentation feature dimension of each measurement point is [dimension]. Then the dimension of the node feature matrix is .

[0111] The normalization layer, connected after the input layer, is used to perform layer normalization on the spatial enhancement features of each measurement point and output a normalized feature matrix.

[0112] The first graph attention layer, connected after the normalization layer, is used to perform graph attention calculation on the normalized feature matrix using multiple attention heads. The outputs of the multiple attention heads are concatenated and then activated by the LeakyReLU activation function to obtain the first graph attention feature output by the first graph attention layer.

[0113] The second graph attention layer, connected after the first graph attention layer, is used to perform graph attention calculation on the attention features of the first graph using multiple attention heads. The outputs of the multiple attention heads are averaged and then activated by the ELU activation function to obtain the attention features of the second graph.

[0114] The regression head, connected after the attention layer of the second graph, includes a fully connected layer and a sigmoid activation function, which is used to map the attention features of the second graph to normalized hollowness values;

[0115] The post-processing unit, connected after the regression head, is used to map the normalized hollowness value to a continuous range of 0~3 through scale transformation, so as to obtain the continuous hollowness value of each measuring point.

[0116] Optionally, during the training of the graph attention network, a training dataset is constructed using a preset detection region as samples. Each sample includes: a node feature matrix, a neighborhood set, and a true hollowness label for each measurement point. The node feature matrix consists of spatial enhancement features for each measurement point. The graph attention network is trained using mean squared error as the loss function and minimizing the error between the true hollowness label and the predicted continuous hollowness value as the optimization objective. Here, the "preset detection region" refers to a continuous region on the wall to be detected, defined by natural boundaries such as room or wall dividing lines, or by artificial design. The measurement points within this region constitute a complete graph structure sample, and the regions do not overlap.

[0117] In this embodiment, both the first and second graph attention layers are equipped with 4 attention heads, using the Adam optimizer with an initial learning rate of 0.001 and 200 training rounds.

[0118] The mean squared error loss function is expressed as follows:

[0119] ;

[0120] In the formula, For sample index, Labels indicating the true degree of hollowness in the sample. This represents a continuous value indicating the predicted degree of hollowness in the sample. Indicates the number of samples.

[0121] S5. Evaluate the degree of hollowness at each measuring point based on the continuous values ​​of the hollowness degree.

[0122] In one embodiment, a preset set of discrete level thresholds is obtained. The discrete level thresholds include a first threshold, a second threshold, and a third threshold, wherein the first threshold is less than the second threshold and less than the third threshold; for example, the first threshold is 0.5, the second threshold is 1.5, and the third threshold is 2.5. When the continuous value of the hollowness degree is less than or equal to the first threshold, the measuring point is determined to have no hollowness; when the continuous value of the hollowness degree is greater than the first threshold and less than or equal to the second threshold, the measuring point is determined to have mild hollowness; when the continuous value of the hollowness degree is greater than the second threshold and less than or equal to the third threshold, the measuring point is determined to have moderate hollowness; and when the continuous value of the hollowness degree is greater than the third threshold, the measuring point is determined to have severe hollowness.

[0123] In this embodiment, the evaluation results can be presented in a visual manner, such as generating a heat map of the wall hollow distribution, using different colors to represent different degrees of hollowness; or generating a test report, listing the coordinates of all measuring points, continuous values ​​of hollowness degree, and discrete levels.

[0124] Figure 3 This is a schematic diagram of the wall hollow detection system based on multi-excitation response and spatial fusion provided in an embodiment of the present invention. Figure 3 As shown, this embodiment of the invention also provides a wall hollowing detection system based on multi-excitation response and spatial fusion, comprising:

[0125] The acquisition module 310 is used to acquire multi-type excitation response signals from multiple measuring points on the wall to be tested. The multi-type excitation response signals include at least the response signal under transient impact excitation and the response signal under frequency sweep excitation.

[0126] The characterization construction module 320 is used to preprocess the multi-type excitation response signals of each measurement point and construct a unified characterization to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measurement point;

[0127] The feature enhancement module 330 is used to extract multi-dimensional features of each measurement point based on the frequency domain amplitude spectrum and time-frequency energy distribution, and construct a neighborhood set according to the spatial coordinates of the measurement point, and fuse the multi-dimensional features with the neighborhood information to generate spatial enhancement features for each measurement point.

[0128] The graph attention regression module 340 is used to take each measurement point as a graph node and the adjacency relationship in the neighborhood set as the graph connection edge, input the spatial enhancement feature as the node feature into the graph attention network, aggregate the neighbor node information through the graph attention network, and regress the continuous value of the hollowness degree of each measurement point.

[0129] The hollowness assessment module 350 is used to assess the degree of hollowness at each measuring point based on continuous values ​​of hollowness degree.

[0130] This invention also provides an electronic device, such as... Figure 4As shown, it includes a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404.

[0131] Memory 403 is used to store computer programs;

[0132] When processor 401 executes the program stored in memory 403, it performs the following steps:

[0133] Collect multi-type excitation response signals from multiple measuring points on the wall to be tested. The multi-type excitation response signals include at least the response signal under transient impact excitation and the response signal under frequency sweep excitation.

[0134] The multi-type excitation response signals at each measurement point are preprocessed and a unified characterization is constructed to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measurement point.

[0135] Based on the frequency domain amplitude spectrum and the time-frequency energy distribution, multidimensional features of each measurement point are extracted, and a neighborhood set is constructed according to the spatial coordinates of the measurement point. The multidimensional features are then fused with the neighborhood information to generate spatial enhancement features for each measurement point.

[0136] Using each measurement point as a graph node and the adjacency relationships in the neighborhood set as graph connection edges, the spatial enhancement features are input into the graph attention network as node features. The graph attention network aggregates neighbor node information and regresses to output a continuous value of the hollowness degree of each measurement point.

[0137] The degree of hollowness at each measuring point is evaluated based on the continuous values ​​of the hollowness degree.

[0138] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0139] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0140] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0141] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0142] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0143] For the embodiments of the device / electronic device / storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments.

[0144] It should be noted that the device, electronic device, and storage medium in the embodiments of the present invention are respectively the system, electronic device, and storage medium for applying the above-mentioned wall hollow detection method based on multi-excitation response and spatial fusion. Therefore, all embodiments of the above-mentioned wall hollow detection method based on multi-excitation response and spatial fusion are applicable to the device, electronic device, and storage medium, and can achieve the same or similar beneficial effects.

[0145] The terminal device provided by the embodiments of the present invention can display proper nouns and / or fixed phrases for users to select, thereby reducing user input time and improving user experience.

[0146] This terminal device exists in various forms, including but not limited to:

[0147] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and are primarily designed to provide voice and data communication. These terminals include smartphones (e.g., iPhones), multimedia phones, feature phones, and low-end phones.

[0148] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, such as the iPad.

[0149] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes audio and video players (such as iPods), handheld game consoles, e-book readers, as well as smart toys and portable car navigation devices.

[0150] (4) Other electronic devices with data interaction functions.

[0151] In the description of this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0152] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0153] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus (devices), or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects, all of which are collectively referred to herein as "modules" or "systems." Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The computer program may be stored / distributed in a suitable medium, provided with or as part of other hardware, or may take other distribution forms, such as via the Internet or other wired or wireless telecommunications systems.

[0154] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0155] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0156] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0157] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting hollow walls based on a multi-excitation response and spatial fusion neural network, characterized in that, include: Collect multi-type excitation response signals from multiple measuring points on the wall to be tested. The multi-type excitation response signals include at least the response signal under transient impact excitation and the response signal under frequency sweep excitation. The multi-type excitation response signals at each measurement point are preprocessed and a unified characterization is constructed to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measurement point. Based on the frequency domain amplitude spectrum and the time-frequency energy distribution, multidimensional features of each measurement point are extracted, and a neighborhood set is constructed according to the spatial coordinates of the measurement point. The multidimensional features are then fused with the neighborhood information to generate spatial enhancement features for each measurement point. Using each measurement point as a graph node and the adjacency relationships in the neighborhood set as graph connection edges, the spatial enhancement features are input into the graph attention network as node features. The graph attention network aggregates neighbor node information and regresses to output a continuous value of the hollowness degree of each measurement point. The degree of hollowness at each measuring point is assessed based on the continuous values ​​of the hollowness. The step of extracting multidimensional features of each measurement point based on the frequency domain amplitude spectrum and the time-frequency energy distribution includes: extracting frequency domain features and amplitude energy features of each measurement point based on the frequency domain amplitude spectrum; the frequency domain features include one or more of the following: dominant frequency, spectral centroid, and spectral bandwidth; the amplitude energy features include one or more of the following: peak amplitude and spectral energy; extracting time domain correlation features of each measurement point based on the time-frequency energy distribution; the time domain correlation features include one or more of the following: echo delay and energy attenuation features; extracting difference features of each measurement point relative to the region without voids based on pre-constructed benchmark feature values; the difference features include one or more of the following: dominant frequency offset, relative amplitude change, and energy difference; and combining the frequency domain features, the amplitude energy features, the time domain correlation features, and the difference features to obtain the multidimensional features of each measurement point. The step of constructing a neighborhood set based on the spatial coordinates of the measuring points and fusing the multidimensional features with the neighborhood information to generate spatially enhanced features for each measuring point includes: constructing a neighborhood set for each measuring point based on the spatial coordinates of the measuring points; for each measuring point, calculating the mean of the multidimensional features of all measuring points in its neighborhood set to obtain the neighborhood mean feature of the measuring point; for each measuring point, calculating the difference between the multidimensional feature of the measuring point and the neighborhood mean feature to obtain the feature difference of the measuring point; and concatenating the multidimensional feature, neighborhood mean feature, and feature difference of the measuring point to obtain the spatially enhanced feature of the measuring point.

2. The method according to claim 1, characterized in that, The steps of preprocessing the multi-type excitation response signals at each measurement point and constructing a unified characterization to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measurement point include: For each measurement point, the response signal acquired each time is sequentially subjected to bandpass filtering, baseline correction, amplitude normalization, and outlier removal to obtain the preprocessed effective signal; Perform Fast Fourier Transform on multiple valid signals at the same measurement point to obtain the frequency domain amplitude spectrum of each sample, and then fuse the frequency domain amplitude spectra of each sample according to frequency points to obtain the frequency domain amplitude spectrum of the measurement point. Perform short-time Fourier transform or continuous wavelet transform on multiple valid signals at the same measurement point to obtain multiple time-frequency energy distributions; fuse the multiple time-frequency energy distributions to obtain the time-frequency energy distribution of the measurement point.

3. The method according to claim 1, characterized in that, The benchmark feature value is obtained through the following steps: In a known area without voids, multiple reference measurement points are selected. Based on the frequency domain amplitude spectrum of each reference measurement point, the single-point reference main frequency, single-point reference peak amplitude, and single-point reference spectral energy of each reference measurement point are extracted respectively. The average value of the single-point reference dominant frequency for all reference measurement points is used to obtain the reference dominant frequency; The average value of the single-point peak amplitude of all benchmark measurement points is used to obtain the benchmark peak amplitude. The average value of the single-point reference spectrum energy of all reference measurement points is used to obtain the reference spectrum energy. The reference main frequency, the reference peak amplitude, and the reference spectral energy are used as the reference characteristic values.

4. The method according to claim 1, characterized in that, The graph attention network includes: The input layer receives the spatial augmentation features of each measurement point, forming a node feature matrix. The normalization layer, connected after the input layer, is used to perform layer normalization on the spatial enhancement features of each measurement point and output a normalized feature matrix. The first graph attention layer, connected after the normalization layer, is used to perform graph attention calculation on the normalized feature matrix using multiple attention heads. The outputs of the multiple attention heads are concatenated and then activated by the LeakyReLU activation function to obtain the first graph attention feature output by the first graph attention layer. The second graph attention layer, connected after the first graph attention layer, is used to perform graph attention calculation on the attention features of the first graph using multiple attention heads, average the outputs of the multiple attention heads, and then activate them through the ELU activation function to obtain the attention features of the second graph. The regression head, connected after the attention layer of the second graph, includes a fully connected layer and a sigmoid activation function, used to map the attention features of the second graph to normalized hollowness values; The post-processing unit, connected after the regression head, is used to map the normalized hollowness value to a continuous range of 0~3 through scale transformation, so as to obtain a continuous hollowness value for each measuring point.

5. The method according to claim 4, characterized in that, Also includes: A training dataset is constructed using preset detection regions as samples; Each sample includes: a node feature matrix, a neighborhood set, and a true hollowness label for each measurement point; the node feature matrix is ​​composed of the spatial enhancement features of each measurement point. The graph attention network is trained using mean squared error as the loss function and minimizing the error between the true hollowness label and the predicted continuous hollowness value as the optimization objective.

6. The method according to claim 1, characterized in that, The steps for evaluating the degree of hollowness at each measuring point based on the continuous values ​​of hollowness include: Obtain a preset set of discrete level thresholds, wherein the discrete level thresholds include a first threshold, a second threshold, and a third threshold, and the first threshold is less than the second threshold and the third threshold; When the continuous value of the hollowness degree is less than or equal to the first threshold, the measuring point is determined to have no hollowness; when the continuous value of the hollowness degree is greater than the first threshold and less than or equal to the second threshold, the measuring point is determined to have mild hollowness; when the continuous value of the hollowness degree is greater than the second threshold and less than or equal to the third threshold, the measuring point is determined to have moderate hollowness; when the continuous value of the hollowness degree is greater than the third threshold, the measuring point is determined to have severe hollowness.

7. A wall hollowing detection system based on multi-excitation response and spatial fusion neural network, characterized in that, To implement the method according to any one of claims 1 to 6, comprising: The acquisition module is used to acquire multi-type excitation response signals from multiple measuring points on the wall to be tested. The multi-type excitation response signals include at least the response signal under transient impact excitation and the response signal under frequency sweep excitation. The characterization construction module is used to preprocess the multi-type excitation response signals of each measurement point and construct a unified characterization to obtain the frequency domain amplitude spectrum and time-frequency energy distribution of each measurement point. The feature enhancement module is used to extract multi-dimensional features of each measurement point based on the frequency domain amplitude spectrum and the time-frequency energy distribution, construct a neighborhood set according to the spatial coordinates of the measurement point, and fuse the multi-dimensional features with the neighborhood information to generate spatial enhancement features for each measurement point. The graph attention regression module is used to take each measurement point as a graph node and the adjacency relationship in the neighborhood set as the graph connection edge. It inputs the spatial enhancement feature as node feature into the graph attention network, aggregates the neighbor node information through the graph attention network, and regresses and outputs the continuous value of the hollowness degree of each measurement point. The hollowness assessment module is used to assess the degree of hollowness at each measuring point based on the continuous value of the hollowness degree.

8. An electronic device, characterized in that, include: processor; Memory, used to store one or more programs; When the one or more programs are executed by the processor, the processor performs the method according to any one of claims 1 to 6.

Images