Full-automatic wall surface acquisition device and intelligent detection method for wall surface hollowing

Abstract

The invention provides a full-automatic wall surface acquisition device and a wall surface hollowing intelligent detection method, the device comprises a support structure for providing stable structure support and a control module, the support structure is provided with a traction and knocking module, a fan and a microphone, the traction and knocking module is used for knocking a wall surface, the fan is used for driving the microphone to rotate, and the microphone is used for driving the fan to rotate. The draught fan is used for providing negative pressure wind power so that the supporting structure can be tightly attached to the surface of the outer wall of a building, the microphone is used for recording wall surface knocking signals, and the control module is in signal connection with the traction and knocking module, the draught fan and the microphone, controls the traction and knocking module, the draught fan and the microphone and obtains collected signals of the microphone. According to the device, standardized and automatic knocking and signal acquisition can be realized, and the consistency and comparability of data are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural health detection, and in particular relates to a fully automatic wall surface acquisition device and an intelligent wall hollowing detection method. Background Art

[0002] Wall hollowing is a common quality issue in buildings. If not promptly detected and repaired, it can lead to exterior wall collapse and safety accidents. Traditional manual tapping methods rely on direct contact with the wall. Hollowing detection in high-rise buildings often requires scaffolding or the use of hanging baskets, which not only increases inspection costs but also poses significant safety risks. Furthermore, manual inspections at high altitudes are affected by factors such as wind and operational stability, which can easily lead to missed inspections or misjudgments, reducing detection reliability. Therefore, manual tapping and listening methods rely on subjective experience, resulting in poor efficiency and consistency, making them difficult to adapt to high-rise or large-scale inspection needs.

[0003] With the development of intelligent detection technology, methods based on acoustic signal analysis are gaining popularity. While current non-contact detection methods based on ultrasound or infrared thermal imaging can improve detection accuracy to a certain extent, they are limited by factors such as ambient temperature, wall material, and equipment cost, making them difficult to implement on a large scale in practical projects. Furthermore, existing machine learning methods for acoustic signal processing and classification still suffer from unstable feature extraction and significant noise-induced classification accuracy, making them inadequate for hollowing detection in complex environments. Consequently, existing detection methods still lack accuracy, robustness, and portability.

[0004] Therefore, designing an efficient and standardized hollow drum signal acquisition device and combining it with optimized signal processing and classification algorithms has become the core issue of current research. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the existing technology, solve the problem of hollowing of building walls, and propose a fully automatic wall acquisition device and an intelligent wall hollowing detection method.

[0006] The technical solutions for achieving the purpose of the present invention are:

[0007] A fully automatic wall surface acquisition device includes a support structure and a control module for providing stable structural support. A traction and knocking module, a fan and a microphone are installed on the support structure. The traction and knocking module is used to knock on the wall. The fan is used to provide negative pressure wind force to make the support structure close to the surface of the building's exterior wall. The microphone is used to record the wall knocking signal. The control module is connected to the traction and knocking module, fan and microphone signals to control the traction and knocking module, fan and microphone to obtain the acquisition signal of the microphone.

[0008] Preferably, the traction and knocking module includes a swing arm motor, a chrome-plated optical axis cross bar, a bearing, a chrome-plated optical axis longitudinal rod, and a hammer head device. The swing arm motor is installed in the supporting structure, and its output end is connected to the chrome-plated optical axis longitudinal rod. The lower end of the chrome-plated optical axis longitudinal rod is located below the supporting structure and is connected to the hammer head device. The upper end of the chrome-plated optical axis longitudinal rod is connected to the chrome-plated optical axis cross bar. The chrome-plated optical axis cross bar is rotatably connected to the upper side of the supporting structure through a bearing. The swing arm motor drives the chrome-plated optical axis longitudinal rod to swing, and then drives the hammer head device to knock.

[0009] Preferably, the upper end of the chrome-plated optical axis longitudinal rod is bolted to the chrome-plated optical axis transverse rod.

[0010] Preferably, the bearings are installed at symmetrical positions at both ends of the upper side of the supporting structure.

[0011] Preferably, the support structure is a rectangular parallelepiped frame with dimensions of 200×100×100 cm and is made of metal.

[0012] Preferably, the swing arm motor adopts a 24V DC motor, the arm length of the swing arm motor is 77 mm, and the swing arm angle is 30 degrees.

[0013] Preferably, the diameters of the chrome-plated optical axis crossbar, bearings, and chrome-plated optical axis longitudinal rod are 14 mm, and the hammer head device is made of metal and has a diameter of 35 mm.

[0014] Preferably, the swing arm motor is installed at a central position inside the support structure.

[0015] Preferably, the fan is installed on the outer surface of the supporting structure.

[0016] An intelligent wall hollowing detection method, comprising:

[0017] Step 1: Collect the audio signal generated by knocking on the wall through a fully automatic wall acquisition device;

[0018] Step 2: Using Bayesian optimized variational mode decomposition and integrated empirical mode decomposition to perform noise reduction on the audio signal and extract audio signal features;

[0019] Step 3: Based on the audio signal features, classification is performed through the trained Bayesian optimization support vector machine and the ensemble learning model to complete hollow drum detection.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] This paper has developed an automated acquisition device that can precisely control the force, angle, and position of the knocks, ensuring the standardization of hollowing signal acquisition, reducing human error, and improving detection accuracy and efficiency. Compared with traditional manual detection, this automated device is more stable and consistent, and is particularly suitable for hollowing detection in high-rise buildings.

[0022] In terms of signal processing, the present invention uses a noise reduction algorithm to remove environmental noise, ensuring the accuracy of the hollow drum signal. By extracting features from the signal, key parameters that effectively reflect the hollow drum characteristics are obtained, providing a reliable basis for subsequent classification. The present invention also uses a majority voting machine learning classification algorithm to achieve automatic recognition and classification of hollow drums, with an accuracy rate of 99.31%. Through experimental comparison, the optimal model was selected, improving classification accuracy and system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a front view of the device of the present invention.

[0024] Figure 2 It is a side view of the device of the present invention.

[0025] Figure 3 Flow chart of the method of the present invention.

[0026] Figure 4 Optimizing VMD flow chart for Bayesian algorithm.

[0027] Figure 5 This is the flow chart of the EEMD algorithm.

[0028] Figure 6 The marble signal is reconstructed after denoising. Figure 6 (a) is the reconstructed image of the marble wall hollow signal after denoising. Figure 6 Middle (b) is the reconstructed image of the marble wall non-hollow signal after denoising.

[0029] Figure 7 This is the MFCC+Mel mixed feature flow chart.

[0030] Figure 8 This is an example of the MFCC+Mel fusion feature map of marble wall audio. Figure 8 (a) is an example of the MFCC+Mel fusion feature map of the marble wall hollowing signal. Figure 8 (b) is an example of the MFCC+Mel fusion feature map of the non-hollow signal of the marble wall.

[0031] Figure 9 This is the flow chart of the majority voting algorithm model.

[0032] Figure 10 Confusion matrix for the majority voting model. DETAILED DESCRIPTION

[0033] Combine Figure 1 The present invention provides a fully automatic wall surface acquisition device, comprising:

[0034] Traditional hollowing detection methods have significant limitations in high-rise buildings. Because manual tapping relies on direct contact with the wall, hollowing detection in high-rise buildings often requires the construction of scaffolding or the use of hanging baskets, which not only increases testing costs but also poses significant safety risks. Furthermore, manual testing at high altitudes is affected by factors such as wind and operator stability, which can easily lead to missed detections or misjudgments, reducing detection reliability.

[0035] In response to the above problems, this embodiment designs and manufactures a fully automatic hollow drum signal acquisition device that can accurately control the force, angle, and position of the knock, improving the stability and consistency of the signal acquisition. The device is suitable for different types of wall materials, reduces human errors, and improves detection efficiency. In particular, it can significantly improve safety and ease of operation in high-rise buildings, providing a more efficient and accurate solution for building hollow drum detection. The detailed design diagram of the device is shown in the figure below. Figure 1 and Figure 2 The device can be divided into several parts, including the support structure 6, the traction and knocking module, the fan 1, the microphone 2, and the control system module. The details of each module are shown below.

[0036] (1) Support structure

[0037] The support structure 6 is a rectangular parallelepiped frame that provides stable structural support to prevent the device from shaking or deforming during operation. The support structure 6 measures 200 × 100 × 100 cm and is made of metal, providing sufficient strength and rigidity.

[0038] (2) Traction and knocking module

[0039] The traction and percussion module consists of a swing arm motor unit 3, a chrome-plated optical axis longitudinal rod 7, a chrome-plated optical axis cross rod 4, a bearing 5, and a hammer head device 8. Its main function is to control the hammer head 8 to perform rhythmic percussion to generate a hollow drum signal and transmit it to the microphone 2. The swing arm motor 3 is installed in the center of the support structure module 6 and is mainly responsible for providing traction to pull the hammer head device 8 to percuss the wall. The chrome-plated optical axis cross rod 4 is connected to the support structure 6 via a bearing 5, and the chrome-plated optical axis cross rod 4 can rotate around the bearing 5. The bearing 5 is connected to the support structure 6 and is located symmetrically on both sides of the upper side of the support structure 6. The chrome-plated optical axis longitudinal rod 7 is connected to the chrome-plated optical axis cross rod 4 by bolt connection. The hammer head device 8 is connected to the chrome-plated optical axis longitudinal rod 7 by bolt connection. The hammer head device 8, the chrome-plated optical axis longitudinal rod, and the chrome-plated optical axis cross rod are connected as a whole and can rotate freely around the bearing 5. The power is provided by the swing arm motor unit 3. The swing arm motor device 3 can accurately pull the hammer head device 8, and the hammer head device 8 is driven to move by the rotation of the motor 3 to achieve precise knocking.

[0040] The diameter of the chrome-plated optical axis longitudinal rod 7 and the chrome-plated optical axis transverse rod 4 is 14 mm, and they are mainly used for fixing and guiding to ensure that the striking motion of the hammer head remains in a straight line, so that the swing arm motor remains stable during the striking process, reduces lateral deviation, and avoids the impact of lateral deviation on the detection results.

[0041] The hammer head device 8 is made of metal and has a diameter of 35 mm. Appropriate mass is considered during design to ensure that the striking force is moderate, which can effectively stimulate the hollow drum signal without damaging the wall structure.

[0042] The swing arm motor device 3 uses a 24V DC motor. The arm length of the swing arm motor device 3 is 77mm, the swing arm angle is 30 degrees, and the swing frequency of the swing arm motor under normal conditions is 5 times / minute. The movement speed of the swing arm can be controlled by the speed regulator to adapt to the detection requirements of different walls.

[0043] The bearing 5 has an outer diameter of 35 mm and an inner diameter of 14 mm and is used to reduce friction and improve the operating efficiency of the swing arm motor and traction system.

[0044] The rectangular frame of the supporting structure provides overall structural support, while the bearings and optical axis ensure that the moving parts (such as the swing arm motor and traction system) can operate smoothly, avoiding the impact of friction or resistance on the accuracy of the knocking.

[0045] The chrome-plated optical axis longitudinal rod 7 and the chrome-plated optical axis transverse rod 4 can rotate freely around the bearings. During the use of the device, it is lowered along the wall by a rope. During the lowering process, the gravity of the device itself and the wind force provided by the fan will perfectly fit the entire device to the wall, thereby ensuring the stability of the knocking process. The function of the swing arm motor is to provide power for the chrome-plated optical axis and the hammer head, which are connected by a traction rope and the hammer head. As the swing arm motor rotates, the wall is knocked. The audio signal generated by the knocking is collected by the microphone, and then the collected signals are uniformly noise-reduced and their features are extracted. Finally, they are classified by a machine learning algorithm.

[0046] (3) Fan device

[0047] The fan 1 is set on the front of the supporting structure 6, and mainly provides a negative pressure wind force for the device during use; the core component of the fan 1 is the YNF200-2T fan, whose main function is to provide continuous and stable wind force, so that the device can be close to the surface of the building's exterior wall, ensuring that there will be no slipping or tilting when knocking, thereby ensuring the accuracy of signal acquisition. After the device is hung and lowered to the detection area, the fan 1 is started, and the strong airflow makes the device stick to the wall. At the same time, combined with the gravity of the device itself, it ensures that the equipment is stable and does not shake during the detection process. The adsorption effect of the fan 1 can also reduce the impact of the uneven wall surface on the detection accuracy to a certain extent, so that the knocking force acts evenly on the wall, improving the stability and repeatability of the detection.

[0048] (4) Microphone

[0049] Microphone 2, mounted on the underside of support structure module 6, records tapping signals from the wall, collecting the resulting sound and converting it into a digital signal for subsequent analysis. This module uses a VM10PRO high-sensitivity microphone, which can capture subtle sound changes and maintain a high signal-to-noise ratio even in complex environments.

[0050] When hammerhead device 8 strikes the wall, the wall emits different echo signals. Microphone 2 collects these sounds, converts them into electrical signals, and stores them in the control system after signal amplification and filtering. Because the sound wave propagation characteristics of hollow areas and non-hollow areas are different, subsequent signal analysis can be used to determine whether the wall is hollow.

[0051] During operation, the control system activates the swing arm motor 3, which drives the hammer 8 to strike the wall at a set force and frequency. If hollow areas are present, the hammer 8 will produce a reverberation of a specific frequency, while areas without hollow areas will have a different acoustic signature. This signal is then collected by a microphone and transmitted to the control system for recording and analysis.

[0052] Combine Figure 3 , this embodiment also provides a fully automatic wall acquisition device and an intelligent detection method for wall hollowing, including signal acquisition, signal processing and learning model output detection results; in terms of signal processing, Bayesian optimized variational mode decomposition (BO-VMD) and ensemble empirical mode decomposition (EEMD) are used to reduce noise and improve the separability of signal features. Subsequently, Mel spectrum (Mel) and Mel cepstral coefficient (MFCC) features are extracted, and the classification performance is enhanced by feature-level fusion. Finally, Bayesian optimized support vector machine (BO-SVM) and ensemble learning model are used for classification to achieve high-precision hollowing detection. This method can effectively overcome the limitations of traditional detection methods and provide efficient, standardized and intelligent solutions for building hollowing detection. The solution specifically includes:

[0053] 1. Marble wall signal collection and evaluation indicators

[0054] Data collection was conducted strictly in accordance with standardized procedures to ensure that the collected hollowing signals were highly representative and usable. Using the collection device designed in this invention, hollowing signals from marble walls were collected. A total of 1,445 samples were processed, including 721 hollowing samples and 724 non-hollowing samples on the marble wall. Details are shown in Table 1:

[0055] Table 1 Sample details

[0056]

[0057] The calculation methods of various indicators are shown in Table 2. Since the present invention discusses the noise reduction and classification problems, the signal-to-noise ratio (SNR), root mean square error (RMSE), precision, recall and F1 value are selected as evaluation indicators.

[0058] Table 2 Evaluation indicators

[0059]

[0060]

[0061] For noise reduction, the signal-to-noise ratio (SNR) is a key indicator for evaluating signal quality, which represents the ratio of signal strength to noise strength. A high SNR means that the ratio of useful components in the signal to background noise is high, which indicates that the clarity and quality of the signal are good, where Ps is the power of the signal and Pn is the power of the noise. The root mean square error (RMSE) quantifies the error before and after signal processing, reflecting the gap between the processed signal and the original signal. A lower RMSE means that the signal after noise reduction is closer to the original signal and the error is smaller, where n is the number of samples and y is the number of samples. i is the ith actual value, is the i-th predicted value, is the squared error for each sample.

[0062] For classification problems, precision, recall, and F1-score are calculated using macro-averaging. TP represents the number of audio samples that are both positive and predicted; FP represents the number of audio samples that are negative but predicted as positive; and FN represents the number of audio samples that are positive but predicted as negative.

[0063] 2 Bayesian algorithm optimization VMD-EEMD denoising algorithm

[0064] Variational mode decomposition (VMD) is an improved signal decomposition method proposed by Konstantin Dragomiretskiy and Dominik Zosso in 2014. Compared to traditional empirical mode decomposition (EMD), VMD is more stable when processing nonlinear and nonstationary signals and can reduce the influence of noise and endpoint effects. Its core concept is to decompose the signal into several intrinsic mode functions (IMFs) with the narrowest bandwidth through an optimization algorithm, thereby achieving more accurate signal separation.

[0065] The goal of VMD is to decompose the input signal f(t) into several intrinsic mode functions Uk(t), each with a central frequency of ωk. The VMD algorithm achieves this goal by solving the following variational problem:

[0066]

[0067] Where: Uk(t) is the kth mode function; ωk is the center frequency of the kth mode function; δ(t) is the Dirac function; j is a virtual unit.

[0068] Combine Figure 4 , the specific implementation steps of the VMD algorithm are as follows:

[0069] (1) Initialize the modal function {u k} and the center frequency {ω k}, set the Lagrange multiplier λ and penalty parameter α.

[0070] (2) For each mode k, perform the following update:

[0071] 1. Update the modal function:

[0072] Removing other modes from the signal yields the residual signal:

[0073] f k =f-∑ ik u i (2)

[0074] where f k Represents the residual signal corresponding to the kth modal component, usually k = 1, 2, ... n.

[0075] Calculate the spectrum of the mode function via Fourier transform:

[0076]

[0077] Among them U k Represents the modal function corresponding to the kth modal component, usually k = 1, 2, ... n, and the inverse Fourier transform obtains the modal function U in the time domain k .

[0078] 2. Update the center frequency

[0079] Update the center frequency based on the spectrum of the mode function:

[0080]

[0081] where w k Represents the center frequency corresponding to the kth modal component, usually k = 1, 2, ...n.

[0082] (3) Update the Lagrange multiplier to ensure that the decomposition result satisfies the constraints:

[0083] λ=λ+τ(f-∑ k U k ) (5)

[0084] (4) Check whether the changes in the modal function and center frequency are less than the preset threshold. If the conditions are met, stop the iteration.

[0085] Experimental results show that the effect of VMD decomposition is highly sensitive to parameter selection. The K value determines the number of decomposition layers. Too large or too small will affect the accuracy of signal decomposition; the penalty factor α affects the bandwidth of the modal component. Improper value selection may cause over-decomposition or under-decomposition of the signal, affecting the reconstruction quality. Since the optimal parameters of different audio signals vary, manual adjustment is costly. Therefore, how to optimize K and α to adapt to different signals becomes the key to improving the VMD decomposition effect. To this end, the present invention introduces Bayesian optimization, which improves the stability and accuracy of VMD decomposition by efficiently searching for the optimal parameter combination by constructing a proxy model and intelligently selecting sampling points.

[0086] The Bayesian optimization algorithm is used to optimize the number of VMD decomposition layers K, penalty sub-alpha, and time step tau. The flowchart is as follows: Figure 4 shown.

[0087] VMD parameter optimization decomposition, the parameter optimization results are shown in Table 3:

[0088] Table 3 VMD parameter optimization results

[0089]

[0090] The analysis results show that Bayesian optimization can effectively avoid the subjectivity and uncertainty brought by traditional empirical parameter setting, so that VMD can achieve better decomposition effect under different signal conditions.

[0091] EEMD (Ensembled Empirical Mode Decomposition) is an improvement to the EMD method. It was proposed by Wu and Huang in 2009 to alleviate the mode aliasing problem of EMD. Its core idea is to add white noise to the original signal, perform EMD decomposition multiple times, and average the results to reduce the impact of mode aliasing. Through this method, EEMD can improve the stability and accuracy of IMF decomposition, making it more reliable when processing nonlinear and non-stationary signals. Figure 5 , the steps of EEMD are as follows:

[0092] (1) Add white noise

[0093] Different white noises are added to the original signal x(t) to generate multiple noisy signals xi(t)=x(t)+wi(t), where wi(t) is the white noise signal added for the i-th time.

[0094] (2)EMD decomposition

[0095] Perform EMD decomposition on each noisy signal xi(t) to obtain several IMFs:

[0096]

[0097] where IMFi,j(t) is the jth IMF of the i-th noisy signal, typically i = 1, 2, …, n; j = 1, 2, …, n. ri(t) is the residual component, and N is the number of IMFs.

[0098] (3) The jth IMF component IMF obtained for all noisy signals i,j (t) and average to get the final j-th IMF:

[0099]

[0100] Where M is the number of times white noise is added, usually i = 1, 2, ... n; j = 1, 2, ... n.

[0101] The present invention further screens the IMF components obtained by VMD decomposition through EEMD, and the flow chart is as follows: Figure 7 As shown:

[0102] EEMD uses the IMF components obtained from VMD decomposition as input, and obtains stable IMFs through noise addition and multiple decompositions. Then, the correlation between each IMF and the original signal is calculated, and the key IMFs are screened for signal reconstruction. Using the Pearson correlation coefficient to screen IMFs can effectively remove noise, retain important features, reduce redundancy, and improve computational efficiency. The screened reconstructed signal is closer to the original signal, improving fidelity, while also enhancing analysis and detection capabilities in feature extraction and pattern recognition, making signal patterns and trends clearer. The signal visualization results of the marble wall after signal reconstruction are shown below. Figure 6 shown.

[0103] It can be seen that after VMD and EEMD processing, the noise of the signal is effectively removed, the reconstructed signal is smoother, the waveform is more regular, and there is no significant high-frequency noise interference.

[0104] A wavelet soft threshold denoising model, a wavelet hard threshold denoising model, an adaptive wavelet denoising model, and a Kalman filter denoising model were constructed and compared with the BO-VEM-EEMD denoising model of the present invention. The comparison results are shown in Tables 4 and 5:

[0105] Table 4 Marble wall is not hollow

[0106]

[0107] Table 5 Marble wall is not hollow

[0108]

[0109] Tables 4 and 5 compare the SNR and RMSE of different denoising methods in the marble wall hollowing signal and non-hollowing signal. The results show that the BO-VMD-EEMD model has the best denoising effect, with an SNR of 24.1 and an RMSE of 0.0016 for the hollowing signal, and an SNR of 80.72 and an RMSE of only 2.679×10 6 , all significantly outperforming the other methods. Wavelet hard thresholding performed second, with SNRs of 9.48 and 12.57, and RMSEs of 0.0087 and 0.0053, respectively. Wavelet soft thresholding and adaptive wavelet denoising were weaker, and Kalman filtering performed the worst. Overall, BO-VMD-EEMD demonstrated significant advantages in improving signal-to-noise ratio and reducing error.

[0110] 3 Marble knocking signal feature extraction

[0111] Next, we extract MFCC and Mel features from the collected audio signal and fuse them to more comprehensively reflect the effective information in the sound signal. MFCC features simulate the human hearing characteristics through the linear transformation of the logarithmic energy spectrum, while Mel features are based on the nonlinear perception of the human ear to frequency. Combining the two, we can more accurately extract the characteristics of the hollow drum signal. The specific process is as follows: Figure 7 shown.

[0112] The fused features are then visualized, and the fusion diagram of the hollow and non-hollow features of the marble wall is shown in the figure below. Figure 8 shown.

[0113] Combining Mel-spectrogram features with MFCC features can simultaneously preserve the detailed information of the sound signal and extract more compact and decorrelated features, thereby improving the expressiveness of the features. This fusion method helps enhance the model's adaptability to noise and signal distortion, improve classification and recognition accuracy, and enhance the model's generalization ability in different environments.

[0114] 4 Majority Voting Algorithm Classification Model

[0115] The present invention uses a majority voting classification algorithm based on ensemble learning to classify and process the collected signals. Specifically, the majority voting classifier integrates four models: BO-SVM, SVM, grid search optimized SVM, and chaotic particle swarm optimized SVM to classify and identify the data set.

[0116] Majority voting is an ensemble learning method that combines the predictions of multiple models to improve overall classification accuracy and stability. Its core concept is to leverage the strengths of multiple base learners to mitigate the limitations of individual models, thereby improving generalization. In this voting mechanism, multiple models independently predict the test sample, and ultimately a comprehensive decision is made based on the results of each model.

[0117] Majority voting is divided into two methods: hard voting and soft voting. Hard voting directly counts the number of predictions for each category and selects the category with the most occurrences as the final result; soft voting is based on the weighted average of the category probabilities output by each model, and selects the category with the highest average probability as the final prediction. In contrast, soft voting can make more full use of the confidence information of each model, so the present invention adopts a soft voting mechanism. The majority voting model based on BO-SVM, SVM, grid search optimized SVM, and chaotic particle swarm optimized SVM has the following overall process: Figure 9 shown.

[0118] The specific steps of the majority voting model proposed in this invention are as follows:

[0119] The image data was first grayscaled and resized to 128×128 pixels. The data was then flattened into a one-dimensional array, normalized, and divided into training and test sets. Four SVM models were constructed as base learners: a standard SVM, a chaotic PSO-optimized SVM, a grid search-optimized SVM, and a Bayesian-optimized SVM. Model parameters were optimized using different methods. A soft voting mechanism was then used to integrate the four base learners into an ensemble classifier, which was trained using the training set. Finally, model performance was evaluated on the test set, with accuracy calculated and a classification report generated to comprehensively analyze the classification results.

[0120] After training and testing, the majority voting model achieved a recognition accuracy of 99.31%. The optimal fitness achieved during training for the chaotic particle swarm optimization SVM model was 1, corresponding to an optimal penalty coefficient c of 477.93. The optimal penalty coefficient achieved during training for the grid search optimization SVM model was 1, with an optimal kernel parameter of 0.001. During training for the Bayesian optimization algorithm model, the optimal kernel function type was linear, with an optimal penalty coefficient c of 747.12 and an optimal kernel parameter g of 0.184. The detailed recognition rates for the models are shown in Table 6; a comparison of the recognition accuracy of the various models is shown in Table 7.

[0121] Table 6 Recognition effect of majority voting model

[0122]

[0123] Table 7 Accuracy of each model

[0124]

[0125] The majority voting ensemble model performed best, achieving an accuracy of 99.31%, significantly outperforming each individual model and demonstrating the advantages of ensemble learning. The Bayesian optimized SVM came in second, achieving an accuracy of 98.36%, demonstrating strong parameter optimization capabilities. The standard SVM and the Chaos PSO optimized SVM had similar performance, at 91.75% and 92.73%, respectively. The grid search optimized SVM performed relatively poorly, at only 92.16%, indicating that its parameter search efficiency is inferior to that of Bayesian optimization. Overall, ensemble models and efficient optimization methods can effectively improve classification performance.

[0126] The confusion matrix of the majority voting model is as follows Figure 10 As shown:

[0127] The confusion matrix of the majority voting classification model shows that the model achieves 99.31% accuracy, further validating its superior performance in the marble signal classification task. By integrating the strengths of multiple base learners, the model is able to efficiently and accurately distinguish between hollow and non-hollow signals and exhibits good robustness. Experimental results demonstrate the feasibility of the majority voting algorithm for wall hollowing detection, as well as its excellent engineering application value and stability.

[0128] This paper proposes a fully automated hollow drum signal acquisition device, combining optimized signal processing and classification algorithms to build a complete hollow drum detection process. The device can achieve standardized and automated percussion and signal acquisition, ensuring data consistency and comparability.

[0129] The above description is only an application implementation method of the patent of the present invention, but the protection scope of the patent of the present invention is not limited to this, and the scope of rights of the patent of the present invention cannot be limited by this. Any equivalent changes made based on the technical solution of the patent of the present invention should be covered by the protection scope of the patent of the present invention.

Claims

1. A fully automatic wall acquisition device, characterized in that: It includes a support structure and a control module for providing stable structural support. A traction and knocking module, a fan and a microphone are installed on the support structure. The traction and knocking module is used to knock on the wall. The fan is used to provide negative pressure wind to make the support structure close to the surface of the building's exterior wall. The microphone is used to record the wall knocking signal. The control module is connected to the traction and knocking module, fan and microphone signals to control the traction and knocking module, fan and microphone to obtain the microphone's collection signal.

2. A fully automatic wall surface acquisition device according to claim 1, characterized in that: The traction and knocking module includes a swing arm motor, a chrome-plated optical axis cross bar, a bearing, a chrome-plated optical axis longitudinal rod, and a hammer head device. The swing arm motor is installed in the supporting structure, and its output end is connected to the chrome-plated optical axis longitudinal rod. The lower end of the chrome-plated optical axis longitudinal rod is located below the supporting structure and is connected to the hammer head device. The upper end of the chrome-plated optical axis longitudinal rod is connected to the chrome-plated optical axis cross bar. The chrome-plated optical axis cross bar is rotatably connected to the upper side of the supporting structure through a bearing. The swing arm motor drives the chrome-plated optical axis longitudinal rod to swing, thereby driving the hammer head device to knock.

3. The fully automatic wall surface acquisition device according to claim 2, characterized in that: The upper end of the chrome-plated optical axis longitudinal rod is connected with the chrome-plated optical axis transverse rod by bolts.

4. The fully automatic wall surface acquisition device according to claim 2, characterized in that: The bearings are installed at symmetrical positions at both ends of the upper side of the supporting structure.

5. The fully automatic wall surface acquisition device according to claim 2, characterized in that: The supporting structure is a rectangular parallelepiped frame with a size of 200×100×100 cm and is made of metal.

6. The fully automatic wall surface acquisition device according to claim 5, characterized in that: The swing arm motor adopts a 24V DC motor, the arm length of the swing arm motor is 77mm, and the swing arm angle is 30 degrees.

7. The fully automatic wall surface acquisition device according to claim 6, characterized in that: The diameters of the chrome-plated optical axis crossbar, bearings, and chrome-plated optical axis longitudinal rod are 14 mm, and the hammer head device is made of metal with a diameter of 35 mm.

8. The fully automatic wall surface acquisition device according to claim 2, characterized in that: The swing arm motor is installed at the center position inside the supporting structure.

9. The fully automatic wall surface acquisition device according to claim 1, characterized in that: The fan is installed on the outer surface of the supporting structure.

10. An intelligent wall hollowing detection method based on the fully automatic wall surface acquisition device according to any one of claims 1 to 9, characterized in that: include: Step 1: Collect the audio signal generated by knocking on the wall through a fully automatic wall acquisition device; Step 2: Using Bayesian optimized variational mode decomposition and integrated empirical mode decomposition to perform noise reduction on the audio signal and extract audio signal features; Step 3: Based on the audio signal features, classification is performed through the trained Bayesian optimization support vector machine and ensemble learning model to complete hollow drum detection.

Images