A wall hollow soundprint feature recognition method and system based on deep learning

Abstract

The application provides a wall hollow sound print feature recognition method and system based on deep learning, relates to the technical field of deep learning, and identifies a knocking sound segment to extract basic acoustic characteristics by collecting mixed acoustic signals in a wall knocking process in real time. On this basis, a confidence degree judgment mechanism is introduced, a result is directly output when the confidence degree reaches a preset threshold, and when the confidence degree is insufficient, a deep verification process is started, noise separation and feature fine extraction are performed on the mixed signals, and finally, the detection result is generated. The problems of low artificial detection efficiency, strong subjectivity, and the problems of intelligent detection equipment in complex construction site environments, such as being easily disturbed by background noise and sudden impact noise, leading to a decrease in recognition accuracy, are effectively solved.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, specifically to a method and system for identifying the acoustic signature features of hollow walls based on deep learning. Background Technology

[0002] In construction projects, the quality inspection of wall plastering or tile installation is crucial, with "hollow spots" being a core inspection indicator, referring to areas where the finishing layer separates from the base wall. Traditional inspection methods rely on experienced workers tapping the wall with a hollow spot hammer and judging the presence of hollow spots by the sound characteristics. However, this manual method is inefficient, and the results are limited by personal experience and subjective judgment, making it difficult to guarantee consistency and accuracy.

[0003] To overcome the aforementioned limitations and improve the objectivity and automation of testing, intelligent voiceprint recognition detection equipment has emerged. This equipment automatically identifies hollow areas by analyzing the sounds produced when tapping on a wall. During the final acceptance phase, inspectors use a device integrating a standardized tapping head and a high-fidelity microphone, moving at a constant speed along the wall and continuously tapping. The device's built-in processor collects the tapping echoes in real time and inputs them into a pre-trained deep learning model. Based on a massive amount of standard samples, this model can extract key identification information from the sound signals, determine whether the tapping location is solid or hollow, and display the results on a screen in real time. It can even mark hollow areas on a digital map of the wall. In an ideal laboratory environment, the equipment achieves extremely high accuracy.

[0004] However, actual construction sites are far more complex than laboratories. First, other trades often operate simultaneously on the same or adjacent floors, such as drilling, cutting, and dismantling / assembling metal scaffolding, generating continuous high-decibel background noise. This noise, along with the effective impact signal, is collected by the microphone, forming a mixed signal that interferes with the model's recognition ability, leading to a decrease in accuracy.

[0005] Secondly, construction sites also generate a large amount of sudden, high-energy impact noise, such as sledgehammers striking steel bars and heavy objects falling. These sounds are short in duration and high in instantaneous intensity, and their time-frequency characteristics are similar to the sounds of impact from testing equipment. The system not only needs to separate the valid signal from the continuous background noise, but also needs to distinguish between intentional impacts by testing personnel and environmental impact noise, avoiding misinterpreting the latter as a valid detection signal.

[0006] Furthermore, the operator's actions also introduce uncertainty. Although the equipment is designed with standardized tapping heads, different users' habits and physical conditions can lead to slight deviations in tapping force, angle, and frequency. Additionally, the wall surface itself is not perfectly homogeneous; tapping near structural columns, wall edges, or tile joints produces a different sound compared to the central area. If the model only learns from samples under standard force and position, it may make judgment errors in the face of variations in actual operation. The most challenging scenario is when a strong impact noise occurs nearby at the moment the operator taps the wall, such as a metal pipe falling from the neighboring building. In this case, the signal collected by the microphone is composed of three superimposed parts: the effective wall echo distorted by changes in tapping force and position, the persistent background noise, and the sudden impact noise that highly overlaps with the effective signal. The resulting signal waveform is exceptionally complex, deviating significantly from the standard pattern learned during model training. The model struggles to correctly separate information reflecting the internal structural state of the wall, potentially misjudging a solid wall as hollow due to the superimposed noise, thus greatly reducing the reliability of the test results. Summary of the Invention

[0007] This application provides a deep learning-based method and system for recognizing acoustic features of hollow walls, aiming to solve the problems of low efficiency, subjective experience limitations, and reduced accuracy caused by noise interference in complex construction site environments in traditional methods for detecting hollow walls.

[0008] On the one hand, this application provides a deep learning-based method for recognizing the acoustic signature features of hollow walls, including: The system collects mixed acoustic signals generated during wall tapping in real time, identifies the tapping sound segments in the mixed acoustic signals, and extracts the basic acoustic features of the tapping sound segments. Based on the aforementioned basic acoustic features and a pre-trained lightweight fast judgment model, a preliminary judgment of the wall condition is made at the current tapping point, and a judgment confidence level is generated based on the preliminary judgment result; the wall condition includes one of solid and hollow. If the confidence level reaches a preset confidence threshold, the final detection result is generated directly based on the preliminary judgment result. If the confidence level does not reach the preset confidence threshold, the deep verification process is initiated to perform noise separation and fine feature extraction on the mixed acoustic signal to generate the final detection result.

[0009] Optionally, if the confidence level does not reach a preset confidence threshold, the deep verification process is initiated, and the steps of generating the final detection result after noise separation and fine feature extraction of the mixed acoustic signal include: The mixed acoustic signal is subjected to noise separation processing to obtain a clean wall response signal; Acquire the impact force information, and adjust the amplitude compensation of the wall response signal according to the impact force information to obtain the target wall response signal, so as to compensate for the distortion caused by inconsistent impact actions; Fine acoustic information reflecting the internal structural characteristics of the wall is extracted from the response signal of the target wall, and correlation analysis is performed by combining the basic acoustic characteristics of adjacent impact points with the wall condition to obtain the correlation analysis results. Based on the detailed acoustic information and the correlation analysis results, a final judgment result on the wall condition is generated.

[0010] Optionally, after performing a preliminary judgment on the wall condition at the current tapping point based on the basic acoustic features and a pre-trained lightweight fast judgment model, and generating a judgment confidence score based on the preliminary judgment result; after the step of the wall condition including either solid or hollow, the method further includes: Obtain information on the wall material type of the current detection area, and query the acoustic response characteristics and noise sensitivity parameters corresponding to the wall material type information based on the wall material type information; Based on the acoustic response characteristics and the noise sensitivity parameters, the preset confidence threshold is adjusted so that the preset confidence threshold matches the acoustic behavior characteristics of the current wall material. The acoustic behavior characteristics are the inherent acoustic response patterns exhibited by the wall material after being struck, including the propagation characteristics of the striking sound, the energy attenuation rate, and the spectral distribution pattern.

[0011] Optionally, the step of acquiring the impact force information and adjusting the wall response signal based on the impact force information to obtain the target wall response signal, in order to compensate for the distortion caused by inconsistent impact actions, includes: Obtain the magnitude of the striking force, the striking angle deviation, the striking contact time, and the striking contact area; The overall amplitude of the wall response signal is compensated based on the magnitude of the striking force. The spectral distribution of the wall response signal is adjusted based on the tapping angle deviation and the tapping contact time; The local energy distribution of the wall response signal is corrected based on the impact contact area to obtain the target wall response signal.

[0012] Optionally, the step of extracting fine acoustic information reflecting the internal structural characteristics of the wall from the target wall response signal includes: The target wall response signal is subjected to multi-scale spectral decomposition to obtain the energy distribution in different frequency ranges; Time-domain waveform feature analysis is performed on the target wall response signal to obtain the envelope shape parameters and transient impact characteristics of the target wall response signal; The energy distribution, envelope shape parameters, and transient impact characteristics are compared with pre-stored typical acoustic feature patterns to identify the structural tendency of the target wall response signal and to evaluate the degree of distinction between the structural tendency and the actual hollow defects; wherein, the structural tendency is the tendency of the internal structure type of the wall reflected by the target wall response signal, including hollow defect tendency or solid wall structure tendency. Based on the structural tendency and degree of differentiation, fine acoustic information reflecting the internal structural characteristics of the wall is generated.

[0013] Optionally, the step of performing a correlation analysis by combining the basic acoustic characteristics of adjacent impact points with the wall condition to obtain the correlation analysis results includes: Acquire acoustic information of the set of adjacent striking points within a preset range around the current striking point; The confidence level of the acoustic information of each striking point in the set of adjacent striking points is obtained, and the confidence level includes high confidence level, medium confidence level and low confidence level. Based on the judgment confidence level, a contribution weight is assigned to each tap point in the set of adjacent tap points. The contribution weight includes a high contribution weight, a medium contribution weight, and a low contribution weight. Tap points with high judgment confidence are assigned a high contribution weight, tap points with medium judgment confidence are assigned a medium contribution weight, and tap points with low judgment confidence are assigned a low contribution weight. The fine acoustic information of the current tapping point is fused with the weighted acoustic information of the adjacent tapping points to generate a weighted correlation analysis result. The wall status is then analyzed based on the weighted correlation analysis result to obtain the correlation analysis result.

[0014] Optionally, the step of performing multi-scale spectral decomposition on the target wall response signal to obtain the energy distribution within different frequency ranges includes: A preliminary analysis of the target wall response signal was performed to obtain the non-stationarity and nonlinear characteristics of the signal. Based on the degree of nonstationarity and the nonlinear characteristics, the wavelet basis function and the number of decomposition layers are dynamically selected. Based on the wavelet basis function and the number of decomposition layers, wavelet transform is performed on the target wall response signal to obtain the spectral coefficients at different scales; The energy distribution within different frequency ranges is calculated based on the spectral coefficients.

[0015] Optionally, the step of performing time-domain waveform feature analysis on the target wall response signal to obtain the envelope shape parameters and transient impact characteristics of the target wall response signal includes: The target wall response signal is divided into time-varying feature windows, and the length and position of the time-varying feature windows are adaptively adjusted according to the instantaneous energy and frequency change rate of the signal. Within each time-varying feature window, the local envelope shape parameters of the signal are calculated, including the local rise time, local decay time, and local peak duration. Within each time-varying feature window, the local transient impulse characteristics of the signal are calculated, including the local impulse factor and local kurtosis. Based on the changing trends of the local envelope shape parameters and the local transient impact characteristics within different time-varying feature windows, the envelope shape parameters and transient impact characteristics of the target wall response signal are identified.

[0016] Optionally, the step of identifying the envelope shape parameters and transient impact characteristics of the target wall response signal based on the changing trends of the local envelope shape parameters and the local transient impact characteristics within different time-varying feature windows includes: The changing trend is segmented, and within each segment, the local envelope shape parameters and the local transient impact characteristics are smoothed to eliminate local abnormal fluctuations, thus obtaining the smoothing result for each segment. Calculate the feature description quantity of each segment based on the smoothing result of each segment; Weights are assigned to the feature descriptions of the segments, and weighted fusion is performed to obtain a weighted fusion result. The weights are determined based on the length of the segments and the stability of the features. Based on the weighted fusion results, the envelope shape parameters and transient impact characteristics of the wall response signal are identified.

[0017] On the other hand, this application provides a deep learning-based wall hollow soundprint feature recognition system, which includes: The signal acquisition and processing module is used to collect mixed acoustic signals generated during the wall-tapping process in real time, identify the tapping sound segments in the mixed acoustic signals, and extract the basic acoustic features of the tapping sound segments. The rapid judgment and confidence assessment module is used to make a preliminary judgment on the wall condition at the current tapping point based on the basic acoustic features and a pre-trained lightweight rapid judgment model, and to generate a judgment confidence level based on the preliminary judgment result; wherein, the wall condition includes one of solid and hollow. The preliminary judgment result output module is used to directly generate the final detection result based on the preliminary judgment result when the judgment confidence reaches a preset confidence threshold; A deep verification module is used to initiate a deep verification process when the confidence level does not reach a preset confidence threshold. This process performs noise separation and fine feature extraction on the mixed acoustic signal to generate the final detection result. This application relates to a deep learning-based method and system for recognizing acoustic signatures of hollow walls. By real-time acquisition of mixed acoustic signals during wall tapping and identifying the tapping sound segments to extract basic acoustic features, a preliminary judgment of the wall's condition is achieved. Based on this, a confidence level mechanism is introduced. When the confidence level reaches a preset threshold, the result is directly output, effectively improving detection efficiency. When the confidence level is insufficient, a deep verification process is initiated to perform noise separation and fine feature extraction on the mixed signal, thereby generating the final detection result. This method effectively solves the problems of low efficiency and high subjectivity in manual detection in existing technologies, as well as the susceptibility of intelligent detection equipment to background noise and sudden impact noise interference in complex construction site environments, leading to a decrease in recognition accuracy. By employing a hierarchical judgment strategy, this application can effectively address signal distortion caused by variations in striking force, angle, and frequency, as well as wall heterogeneity. It avoids misinterpreting environmental impact sounds as valid detection signals, significantly improving the objectivity, consistency, and accuracy of wall hollowness detection. It maintains high reliability, especially in complex scenarios with noise pollution and signal distortion. Attached Figure Description

[0018] To illustrate this application more clearly, the accompanying drawings used in the embodiments will be briefly described below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0019] Figure 1 The diagram above illustrates a flowchart of a deep learning-based method for recognizing the acoustic signature of hollow walls. Figure 2 The diagram above illustrates a structural schematic of a deep learning-based wall hollow soundprint feature recognition system.

[0020] Figure labels: 100, Deep learning-based wall hollow soundprint feature recognition system; 10, Signal acquisition and processing module; 20, Rapid judgment and confidence assessment module; 30, Preliminary judgment result output module; 40, Deep verification module. Detailed Implementation

[0021] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] Traditional methods for detecting hollow walls, whether relying on subjective judgment based on human experience or on intelligent devices based on voiceprint recognition, all face challenges in complex construction site environments, including decreased accuracy, susceptibility to noise interference, and insufficient robustness to operational inconsistencies. Particularly in scenarios with high-decibel background noise and sudden impact noise, existing methods struggle to effectively separate pure wall response signals, leading to biased or even misjudged assessments of the wall's condition and severely impacting the reliability of the detection results.

[0024] like Figure 1 The diagram illustrates an exemplary flowchart of a deep learning-based method for recognizing the acoustic signature of hollow walls. This application proposes a deep learning-based method for recognizing the acoustic signature of hollow walls, comprising: S10, Real-time acquisition of mixed acoustic signals generated during wall tapping, identification of tapping sound segments in the mixed acoustic signals, and extraction of the basic acoustic features of the tapping sound segments; The mixed acoustic signal refers to the composite sound signal formed by the superposition of the wall response sound, ambient background noise, and other sudden noises generated during the wall-tapping process. The tapping segment specifically refers to the sound fragment in the mixed acoustic signal directly related to the wall-tapping event, serving as the basis for subsequent feature extraction and judgment. Basic acoustic features refer to parameters extracted from the tapping segment that can preliminarily reflect the acoustic characteristics of the wall, such as short-time energy, zero-crossing rate, and spectral centroid. These features are typically used for rapid, preliminary judgment.

[0025] S20, based on the basic acoustic features and the pre-trained lightweight fast judgment model, a preliminary judgment of the wall condition is made at the current tapping point, and a judgment confidence level is generated based on the preliminary judgment result; the wall condition includes one of solid and hollow. The lightweight, fast judgment model is a pre-trained deep learning model characterized by low computational resource consumption and fast inference speed, making it suitable for preliminary screening and judgment of a large number of tapping points. Wall condition refers to the internal structural condition of the wall, mainly including solid and hollow types. Solid indicates a dense, defect-free wall structure, while hollow indicates the presence of cavities or delamination within the wall. Judgment confidence is the model's level of confidence in the preliminary judgment result, usually expressed as a probability value, reflecting the model's assessment of the reliability of the current judgment result.

[0026] S30, if the confidence level reaches the preset confidence threshold, the final detection result is generated directly based on the preliminary judgment result; S40, if the confidence level does not reach the preset confidence threshold, the deep verification process is initiated, and the final detection result is generated after noise separation and fine feature extraction of the mixed acoustic signal.

[0027] The deep verification process is a more complex and sophisticated signal processing and feature extraction mechanism. It is activated when the initial judgment lacks sufficient confidence, aiming to improve the accuracy of the final judgment through deeper analysis. Noise separation refers to the process of removing environmental noise and interference signals from the mixed acoustic signal to obtain a purer wall response signal. Fine feature extraction refers to extracting deeper and more discernible acoustic features from the purified signal after noise separation. These features can more accurately reflect the subtle structural differences within the wall.

[0028] The wall hollow sound signature identification method of this application first acquires mixed acoustic signals generated during wall tapping in real time and performs preliminary processing on these signals. Specifically, a high-sensitivity microphone array or a single high-fidelity microphone can be used to capture the sound waves generated when the wall is tapped. The acquired signals are then sent to a signal processing unit, which can use methods such as energy threshold detection, short-time Fourier transform (STFT), or Mel-frequency cepstral coefficients (MFCC) to identify the tapping sound segments in the mixed acoustic signals that are related to the tapping event. For example, when the signal energy rises sharply in a short period of time and then decays, it can be identified as a tapping sound segment. Once the tapping sound segment is identified, basic acoustic features are extracted from it. These features can be time-domain features (such as peak amplitude, energy, and duration) or frequency-domain features (such as spectral centroid, bandwidth, and resonant frequency). These basic features aim to quickly capture the basic properties of the tapping sound, providing a basis for subsequent preliminary judgment.

[0029] Next, based on the extracted basic acoustic features and a pre-trained lightweight fast judgment model, a preliminary judgment of the wall condition at the current tapping point is made. This lightweight fast judgment model can be a small convolutional neural network (CNN), support vector machine (SVM), or decision tree model, characterized by high computational efficiency and the ability to run quickly on resource-constrained devices. The model receives the basic acoustic features as input and outputs a preliminary judgment of the wall condition at the current tapping point, namely, a judgment of "solid" or "hollow". Simultaneously, the model generates a judgment confidence score based on the preliminary judgment result. This confidence score is typically a probability value between 0 and 1, representing the model's confidence level in the current judgment result. For example, if the model judges it as "hollow" with a confidence score of 0.95, it means that the model is 95% confident that there is a hollow area at the tapping point.

[0030] Subsequently, the generated judgment confidence level is compared with a preset confidence threshold. The preset confidence threshold is a configurable parameter, for example, it can be set to 0.8. When the judgment confidence level reaches or exceeds the preset confidence threshold, it indicates that the model has high confidence in the initial judgment result. In this case, the final detection result is directly generated based on the initial judgment result and output. For example, if the initial judgment is "solid" and the confidence level is 0.9, which is higher than the threshold of 0.8, then the final detection result is "solid". This mechanism allows for rapid results even with high confidence levels, improving detection efficiency.

[0031] However, when the confidence level fails to reach the preset confidence threshold, it indicates that the preliminary judgment may contain uncertainties or risks. To ensure the accuracy of the detection, a deep verification process is initiated. In this process, the original mixed acoustic signal undergoes more complex noise separation and fine feature extraction. Noise separation can employ techniques such as blind source separation, spectral subtraction, and deep learning denoising networks, aiming to remove environmental noise and interference from the mixed signal as much as possible to obtain a pure wall response signal. Fine feature extraction extracts richer and more discriminative acoustic features from the pure wall response signal, such as multi-scale spectral features, temporal envelope features, and resonance mode analysis. These refined features can reveal subtle differences in the internal structure of the wall more deeply. Finally, based on these deeply processed and finely extracted features, a more reliable final detection result is generated.

[0032] The acoustic signature feature recognition method for wall hollowness in this application significantly improves the accuracy and robustness of wall hollowness detection in complex construction site environments by introducing a layered judgment mechanism. The method first acquires mixed acoustic signals generated during wall tapping in real time and uses signal processing technology to identify the tapping sound segments. Basic acoustic features are extracted from these tapping sound segments, which can quickly capture the basic attributes of the tapping sound. Subsequently, based on these basic acoustic features and combined with a pre-trained lightweight fast judgment model, a preliminary judgment of the wall state at the current tapping point is made, and a judgment confidence level is generated. This preliminary judgment stage aims to quickly filter out detection results with high confidence, thereby improving overall detection efficiency.

[0033] When the confidence level reaches a preset confidence threshold, indicating high reliability, the preliminary judgment is directly adopted as the final detection result. This mechanism avoids unnecessary complex calculations, enabling rapid conclusions in most cases. However, when the confidence level falls below the preset threshold, a deep verification process is initiated instead of outputting a result prematurely. In this process, the original mixed acoustic signal undergoes more refined noise separation to remove environmental noise and interference as much as possible, obtaining a pure wall response signal. Then, deeper and more discernible fine acoustic features are extracted from the pure wall response signal. These fine features more accurately reflect subtle structural differences within the wall, compensating for the shortcomings of the preliminary judgment. Finally, based on these deeply processed and finely extracted features, a more reliable and accurate final detection result is generated.

[0034] This layered judgment strategy enables the method of this application to balance detection efficiency and accuracy. When noise interference is low or signal characteristics are clear, a judgment can be made quickly; while in cases of complex noise or ambiguous signal characteristics, a deep verification mechanism is used for detailed analysis, effectively addressing the complex environment of construction sites and reducing the false judgment rate. Compared to traditional single-model judgment or purely manual detection methods, the method of this application can more effectively handle noise in mixed acoustic signals, improve robustness to inconsistencies in tapping actions, and ultimately provide more reliable wall hollow detection results.

[0035] In some embodiments, when the confidence level is determined to be below a preset confidence threshold, the deep verification process is initiated, and the steps of generating the final detection result after noise separation and fine feature extraction of the mixed acoustic signal include: The mixed acoustic signal is subjected to noise separation processing to obtain a clean wall response signal; Acquire the impact force information, and adjust the amplitude compensation of the wall response signal according to the impact force information to obtain the target wall response signal, so as to compensate for the distortion caused by inconsistent impact actions; Fine acoustic information reflecting the internal structural characteristics of the wall is extracted from the response signal of the target wall, and correlation analysis is performed by combining the basic acoustic characteristics of adjacent impact points with the wall condition to obtain the correlation analysis results. Based on the detailed acoustic information and the correlation analysis results, a final judgment result on the wall condition is generated.

[0036] Specifically, noise separation processing of the mixed acoustic signals aims to effectively remove environmental noise, background noise, and non-impact-generated noise from the acquired mixed acoustic signals, thereby obtaining a pure wall response signal generated solely by the wall's impact response. This can be achieved through various signal processing techniques, such as adaptive filtering, spectral subtraction, and wavelet denoising, to preserve the acoustic characteristics of the wall itself to the greatest extent possible.

[0037] The process involves acquiring the impact force information and adjusting the wall response signal based on this information to obtain the target wall response signal. The purpose is to compensate for distortions in the wall response signal caused by inconsistent striking actions (e.g., differences in striking force, angle, and contact time). The impact force information can include parameters such as striking force, striking angle, and striking contact time, which can be acquired in real-time using sensors (e.g., force sensors, accelerometers, and visual recognition systems). By adjusting the amplitude of the wall response signal, the amplitude of the response signals under different striking intensities can be compared, thereby eliminating errors caused by human operation and improving the accuracy of feature extraction.

[0038] In practical applications, extracting fine acoustic information reflecting the internal structural characteristics of a target wall from its response signal aims to deeply analyze the wall's subtle responses to impacts, revealing the essential characteristics of its internal structure (such as the presence of hollow defects). This fine acoustic information may include, but is not limited to, the signal's spectral distribution, energy attenuation characteristics, time-domain waveform envelope, resonant frequency, and damping coefficient. Simultaneously, correlation analysis is performed between the fundamental acoustic characteristics of adjacent impact points and the wall's condition, aiming to enhance the robustness of the judgment by utilizing spatial continuity and correlation. By analyzing information from surrounding impact points that have already been judged or preliminarily judged, strong contextual support can be provided for the judgment at the current point, avoiding potential misjudgments caused by isolated judgments.

[0039] Finally, based on detailed acoustic information and correlation analysis results, a final judgment on the wall condition is generated. This means that the final judgment does not rely solely on information from a single dimension, but rather integrates the depth acoustic characteristics of the current point with judgment information from its surrounding environment, thus forming a more comprehensive, accurate, and reliable judgment on the wall condition.

[0040] Through the above technical solutions, this application significantly improves the accuracy and reliability of wall condition identification under conditions of low confidence. Specifically, noise separation processing ensures the purity of the analyzed data, laying the foundation for subsequent fine feature extraction. The introduction of impact force information and amplitude compensation adjustment effectively overcome the signal distortion problem caused by inconsistent manual tapping operations, enabling a more objective assessment of the acoustic response of the wall itself, thereby avoiding misjudgments due to operational errors. In addition, the extraction of fine acoustic information allows for the capture of more subtle features of the internal structure of the wall, improving the ability to identify hollow defects. Furthermore, by combining the correlation analysis of adjacent tapping points, important spatial context information is provided for the judgment of the current point. Through multi-point collaborative verification, the randomness and uncertainty of single tapping point judgment are effectively reduced, significantly reducing the occurrence of missed and false alarms, making the final wall condition judgment result more robust and reliable.

[0041] For example, suppose that during a hollow detection test on a wall, a certain tapping point is initially identified as "solid" by a lightweight, fast judgment model, but its confidence level is only 65%, lower than the preset 70% confidence threshold. At this point, a deep verification process is initiated. First, the mixed acoustic signal collected at the tapping point undergoes noise separation processing, for example, using a deep learning-based noise suppression algorithm to filter out background noise from the construction site, obtaining a clean wall response signal. Next, using a force sensor and accelerometer integrated into the tapping tool, information such as the magnitude of the tapping force, the tapping angle, and the contact time are acquired in real time. Based on this tapping force information, the clean wall response signal is amplitude compensated and adjusted; for example, if the tapping force is too weak, the signal amplitude is appropriately amplified to simulate the response under standard tapping, thus obtaining the target wall response signal. Subsequently, refined acoustic information is extracted from this target wall response signal, such as obtaining the energy distribution of different frequency bands through wavelet packet decomposition, calculating the signal attenuation rate, and resonant frequency. Simultaneously, the system queries the basic acoustic characteristics and assessed wall conditions of adjacent tapping points within a 10-centimeter radius of the tapping point (e.g., two out of three surrounding points are identified as "hollow" with high confidence). The system then fuses and correlates the detailed acoustic information of the current point with the weighted information of these adjacent points. Finally, based on the detailed acoustic information of the current point and the correlation analysis results incorporating surrounding environmental information, the system comprehensively determines that the tapping point is "hollow" and outputs the final detection result.

[0042] In some embodiments, after making a preliminary judgment on the wall condition at the current tapping point based on the basic acoustic features and a pre-trained lightweight fast judgment model, and generating a judgment confidence score based on the preliminary judgment result; after the step of the wall condition including one of solid and hollow, the method further includes: Obtain information on the wall material type of the current detection area, and query the acoustic response characteristics and noise sensitivity parameters corresponding to the wall material type information based on the wall material type information; Based on the acoustic response characteristics and the noise sensitivity parameters, the preset confidence threshold is adjusted so that the preset confidence threshold matches the acoustic behavior characteristics of the current wall material. The acoustic behavior characteristics are the inherent acoustic response patterns exhibited by the wall material after being struck, including the propagation characteristics of the striking sound, the energy attenuation rate, and the spectral distribution pattern.

[0043] Specifically, obtaining information about the wall material type in the current detection area can be understood as identifying the constituent materials of the wall being detected through various methods. For example, this information can be manually entered by the operator, automatically identified by sensors integrated into the detection equipment (such as infrared sensors or ultrasonic sensors), or inferred by analyzing wall images. This wall material type information is a key input for subsequently dynamically adjusting the confidence threshold.

[0044] The system queries the acoustic response characteristics and noise sensitivity parameters corresponding to the wall material type. It utilizes a pre-built and stored database of acoustic characteristics for various common wall materials (e.g., concrete, brick, gypsum board, wood paneling). This database details the inherent propagation characteristics of each material's sound signature when struck (e.g., propagation speed), energy attenuation rate (the rate at which sound wave energy decreases with time and distance), and spectral distribution (the energy distribution of the striking sound at different frequencies). Furthermore, the database stores sensitivity parameters for each material to environmental noise, reflecting the degree to which the acoustic characteristics of a specific material are disturbed under different noise environments.

[0045] In practical applications, adjusting the preset confidence threshold based on acoustic response characteristics and noise sensitivity parameters can be understood as adaptively correcting the confidence threshold used for preliminary judgment based on the unique acoustic behavior characteristics of the current wall material. For example, if the current wall material is identified as gypsum board with strong sound absorption, its acoustic response signal upon impact may be relatively weak and decay rapidly. In this case, the preset confidence threshold can be appropriately lowered to avoid frequently misclassifying valid preliminary judgment results as uncertain due to insufficient signal strength, thereby reducing unnecessary in-depth verification processes. Conversely, if a concrete wall with a clear acoustic response and obvious characteristics is detected, the threshold can be appropriately increased to further improve the accuracy and reliability of the preliminary judgment. This dynamic adjustment mechanism aims to enable the threshold to better adapt to the acoustic characteristics of different materials, thereby optimizing the accuracy and efficiency of preliminary judgment. The acoustic behavior characteristics specifically refer to the inherent acoustic response laws exhibited by the wall material after being struck, such as the propagation speed of the impact sound within the material, the rate of sound wave energy decay with distance and time, and the energy distribution of the impact sound at different frequencies.

[0046] Through the above technical solution, this application can significantly improve the adaptability and accuracy of the wall hollow acoustic signature identification method under different wall material environments. Compared with the scheme using a fixed confidence threshold, this application avoids misjudgment or inefficiency caused by material differences by dynamically adjusting the threshold. Specifically, for materials with indistinct acoustic response signal characteristics, it can effectively reduce the probability of misjudging as "uncertain" and initiating deep verification, thereby improving the efficiency of preliminary judgment; for materials with obvious acoustic response characteristics, it can improve the accuracy of preliminary judgment and reduce missed judgments. Thus, the technical solution of this application makes the preliminary judgment of wall condition more accurate and efficient, especially in the face of complex and ever-changing actual construction environments, providing more reliable detection results and reducing overall detection costs and time.

[0047] In some embodiments, the step of acquiring the impact force information and adjusting the wall response signal based on the impact force information to obtain the target wall response signal, in order to compensate for the distortion caused by inconsistent impact actions, includes: Obtain the magnitude of the striking force, the striking angle deviation, the striking contact time, and the striking contact area; The overall amplitude of the wall response signal is compensated based on the magnitude of the striking force. The spectral distribution of the wall response signal is adjusted based on the tapping angle deviation and the tapping contact time; The local energy distribution of the wall response signal is corrected based on the impact contact area to obtain the target wall response signal.

[0048] Specifically, the magnitude of the striking force, the striking angle deviation, the striking contact time, and the striking contact area are obtained by capturing and quantifying these physical parameters in real time or near real time using sensors integrated into the striking device or external measuring devices. The magnitude of the striking force can be measured using force sensors or accelerometers, reflecting the intensity of the strike. The striking angle deviation refers to the degree of deviation of the striking tool from the normal direction of the wall surface, which can be estimated using an inertial measurement unit (IMU) or a vision sensor. The striking contact time refers to the duration of contact between the striking tool and the wall surface, which can be determined by capturing signal changes at the moment of contact using a high-sampling-rate vibration sensor or acoustic sensor. The striking contact area refers to the actual area of ​​contact between the striking tool and the wall surface, which can be obtained using pressure distribution sensors or image processing technology.

[0049] The purpose of compensating for the overall amplitude of the wall's response signal based on the magnitude of the striking force is to eliminate signal strength differences caused by varying striking force. For example, a larger striking force will result in a larger amplitude of the wall's response signal, and vice versa. By normalizing or proportionally adjusting the signal amplitude, signals under different striking forces can be made comparable.

[0050] Furthermore, the spectral distribution of the wall response signal is adjusted based on the tapping angle deviation and tapping contact time. This aims to correct frequency component distortion caused by changes in tapping posture and duration. For example, angled tapping or prolonged contact may amplify or attenuate certain frequency components, thereby altering the inherent spectral characteristics of the sound signature. By analyzing the relationship between these parameters and spectral changes, targeted compensation can be applied to the signal's frequency response, bringing it closer to the ideal spectral distribution.

[0051] Furthermore, correcting the local energy distribution of the wall response signal based on the impact contact area aims to rectify local energy concentration or dispersion caused by differences in the impact contact area. For example, point-like and area-like impacts will cause different propagation and reflection patterns of sound waves inside the wall, thus affecting the local energy distribution of the signal in the time domain. By correcting the local energy, it can be ensured that the energy distribution characteristics of the signal can more accurately reflect the internal structure of the wall.

[0052] Through the above technical solution, this application can more comprehensively and accurately compensate for the distortion of the wall response signal caused by inconsistent striking actions (including various factors such as the magnitude of striking force, striking angle deviation, striking contact time, and striking contact area). Compared with schemes that rely solely on single striking force information for amplitude compensation, this application significantly improves the purity and accuracy of the target wall response signal by making multi-dimensional and refined adjustments to the overall amplitude, spectral distribution, and local energy distribution of the signal. As a result, subsequent refined acoustic information extraction and wall condition judgment will be based on more reliable input data, thereby effectively improving the robustness and accuracy of wall hollow acoustic feature recognition and reducing the misjudgment rate caused by differences in operator striking techniques.

[0053] In some embodiments, the step of extracting fine acoustic information reflecting the internal structural characteristics of the wall from the target wall response signal includes: The target wall response signal is subjected to multi-scale spectral decomposition to obtain the energy distribution in different frequency ranges; Time-domain waveform feature analysis is performed on the target wall response signal to obtain the envelope shape parameters and transient impact characteristics of the target wall response signal; The energy distribution, envelope shape parameters, and transient impact characteristics are compared with pre-stored typical acoustic feature patterns to identify the structural tendency of the target wall response signal and to evaluate the degree of distinction between the structural tendency and the actual hollow defects; wherein, the structural tendency is the tendency of the internal structure type of the wall reflected by the target wall response signal, including hollow defect tendency or solid wall structure tendency. Based on the structural tendency and degree of differentiation, fine acoustic information reflecting the internal structural characteristics of the wall is generated.

[0054] The multi-scale spectral decomposition of the target wall response signal aims to obtain the energy distribution across different frequency ranges by breaking the signal down into multiple frequency components. This process can reveal the differences in the wall's response to impacting sound waves at different frequencies; for example, hollow areas may exhibit abnormal energy concentration or attenuation patterns within specific frequency ranges.

[0055] Furthermore, time-domain waveform feature analysis is performed on the target wall response signal to obtain the signal's envelope shape parameters and transient impact characteristics. Envelope shape parameters, such as rise time, decay time, and duration, reflect the damping characteristics and resonant behavior of the wall material; while transient impact characteristics quantify the energy transfer and decay rate at the moment of impact. These time-domain features are crucial for distinguishing the transient response differences between solid walls and hollow areas.

[0056] Specifically, the obtained energy distribution, envelope shape parameters, and transient impact characteristics are compared with pre-stored typical acoustic feature patterns. The purpose is to identify the structural tendency of the target wall's response signal and assess the degree to which this structural tendency distinguishes from actual hollow defects. The typical acoustic feature patterns are trained based on a large amount of acoustic data from known wall states (solid or hollow) and represent the inherent acoustic fingerprints of different wall structures. Through comparison, it can be determined whether the current signal is more inclined towards a solid structure or a hollow defect. The structural tendency can be understood as the bias in the internal structural type of the wall reflected by the target wall's response signal, specifically including a tendency towards hollow defects or a tendency towards a solid wall structure. The degree of distinguishability quantifies the matching degree or difference between the current signal and the typical pattern, thereby assessing the reliability of the judgment.

[0057] Therefore, based on the identified structural tendencies and the degree of differentiation assessed, refined acoustic information reflecting the internal structural characteristics of the wall is ultimately generated. This refined acoustic information is a quantitative representation that integrates multi-dimensional features, enabling a more accurate and comprehensive description of the physical structural state inside the wall.

[0058] Through the above technical solution, this application can extract richer and more refined acoustic features from the target wall response signal, including multi-scale spectral energy distribution, temporal envelope shape parameters, and transient impact characteristics. The acquisition of these refined features enables more accurate identification of the wall's structural tendency when faced with complex or ambiguous acoustic signals, and quantifies its distinction from actual hollow defects. This significantly improves the accuracy and reliability of in-depth verification of wall conditions under low confidence levels, effectively avoiding misjudgments or omissions, thereby enhancing the overall performance and robustness of wall hollow detection.

[0059] In some embodiments, the step of performing correlation analysis by combining the basic acoustic characteristics of adjacent impact points with the wall condition to obtain the correlation analysis results includes: Acquire acoustic information of the set of adjacent striking points within a preset range around the current striking point; The confidence level of the acoustic information of each striking point in the set of adjacent striking points is obtained, and the confidence level includes high confidence level, medium confidence level and low confidence level. Based on the judgment confidence level, a contribution weight is assigned to each tap point in the set of adjacent tap points. The contribution weight includes a high contribution weight, a medium contribution weight, and a low contribution weight. Tap points with high judgment confidence are assigned a high contribution weight, tap points with medium judgment confidence are assigned a medium contribution weight, and tap points with low judgment confidence are assigned a low contribution weight. The fine acoustic information of the current tapping point is fused with the weighted acoustic information of the adjacent tapping points to generate a weighted correlation analysis result. The wall status is then analyzed based on the weighted correlation analysis result to obtain the correlation analysis result.

[0060] Specifically, before performing correlation analysis, the acoustic information of a set of adjacent tapping points within a preset range around the current tapping point is first acquired. This acoustic information may include, but is not limited to, basic acoustic features, detailed acoustic information, and preliminary judgment results. Simultaneously, the judgment confidence level of the acoustic information of each tapping point in the set of adjacent tapping points is acquired. Judgment confidence level is an indicator that measures the reliability of the preliminary judgment result. It can be determined based on the probability value or similarity score output by the fast judgment model and is divided into different levels such as high judgment confidence, medium judgment confidence, and low judgment confidence. For example, when the judgment probability is higher than a certain threshold, it can be considered high judgment confidence; when the judgment probability is lower than another threshold, it can be considered low judgment confidence; and what falls between the two is medium judgment confidence.

[0061] Furthermore, based on the confidence level of each adjacent knock point, a corresponding contribution weight is assigned. Knock points with high confidence levels are assigned high contribution weights, indicating that their information is more reliable and should play a greater role in the association analysis; knock points with medium confidence levels are assigned medium contribution weights; and knock points with low confidence levels are assigned low contribution weights to reduce the noise or uncertainty they may introduce. These contribution weights can be dynamically adjusted according to a preset function or rule; for example, a linear or non-linear mapping relationship can be used to convert confidence values ​​into weight values.

[0062] Subsequently, the detailed acoustic information of the current tapping point is fused with the weighted acoustic information of neighboring tapping points. This fusion process can be a weighted average, weighted summation, or other form of feature fusion algorithm, aiming to comprehensively consider the high-precision information of the current point itself and the weighted reliable information of surrounding points. This generates a weighted correlation analysis result. Finally, based on this weighted correlation analysis result, a correlation analysis is performed on the wall condition, resulting in a more accurate and robust correlation analysis result.

[0063] Through the above technical solution, this application effectively solves the problem of decreased analysis accuracy caused by inconsistent reliability of preliminary judgments from adjacent tapping points during wall condition correlation analysis. By assigning different contribution weights to adjacent tapping points, high-confidence information is fully utilized while low-confidence information is reasonably suppressed, thereby significantly improving the accuracy and reliability of the correlation analysis results. This weighted fusion method enables more robust and accurate judgments when facing complex and variable wall structures and environmental noise, reducing the false judgment rate and improving the overall performance of wall hollow detection.

[0064] In some embodiments, the step of performing multi-scale spectral decomposition on the target wall response signal to obtain the energy distribution in different frequency ranges includes: A preliminary analysis of the target wall response signal was performed to obtain the non-stationarity and nonlinear characteristics of the signal. Based on the degree of nonstationarity and the nonlinear characteristics, the wavelet basis function and the number of decomposition layers are dynamically selected. Based on the wavelet basis function and the number of decomposition layers, wavelet transform is performed on the target wall response signal to obtain the spectral coefficients at different scales; The energy distribution within different frequency ranges is calculated based on the spectral coefficients.

[0065] Specifically, a preliminary analysis of the target wall response signal is conducted to obtain the signal's non-stationarity and nonlinear characteristics. Non-stationarity refers to the degree to which the signal's statistical characteristics, such as mean, variance, or spectrum, change over time. Wall impact response signals are typically non-stationary because energy is concentrated at the moment of impact and then rapidly decays; their frequency components may also change over time. Nonlinearity refers to the non-linear relationship between the signal's output and input, which may stem from the complex structure of the wall material or nonlinear effects during sound wave propagation. This preliminary analysis can be achieved using various signal processing techniques. For example, short-time Fourier transform can be used to analyze the signal's time-frequency distribution, or Hilbert-Huang transform can be used to extract instantaneous frequency and amplitude. Statistics such as kurtosis and skewness can also be calculated to assess non-stationarity and nonlinearity. The aim is to provide a reliable basis for the subsequent dynamic selection of wavelet transform parameters.

[0066] Furthermore, based on the obtained non-stationarity and nonlinear characteristics, the wavelet basis functions and the number of decomposition levels are dynamically selected. Wavelet basis functions are the core of wavelet transform; different basis functions have unique time-frequency localization characteristics. For example, the Daubechies wavelet is suitable for analyzing signals with tight support and smoothness, while the Morlet wavelet is more suitable for analyzing signals with oscillatory characteristics. Dynamic selection means that the wavelet basis function that best captures the signal's features can be automatically matched according to the actual characteristics of the signal. The number of decomposition levels determines the level of detail in the signal decomposition at different frequency scales; more levels result in higher frequency resolution but lower time resolution. Dynamically selecting the number of decomposition levels ensures that, under different signal characteristics, the details of the signal are fully revealed while avoiding redundant information caused by over-decomposition. This selection process can be based on a predefined set of rules, machine learning models, or optimization algorithms, performing optimal matching from a candidate wavelet basis function library and a range of decomposition levels based on preliminary analysis results.

[0067] Based on this, wavelet transform is performed on the target wall response signal using dynamically selected wavelet basis functions and decomposition levels, thereby obtaining spectral coefficients at different scales. As a time-frequency analysis method, wavelet transform can decompose a signal into wavelet coefficients at different frequency scales and time locations. These coefficients accurately reflect the energy and characteristics of the signal within a specific frequency range and at a specific time point. Through this adaptive wavelet transform, the transient and local characteristics of the wall response signal at different frequency scales can be captured more accurately.

[0068] Finally, the energy distribution within different frequency ranges is calculated based on the obtained spectral coefficients. The spectral coefficients are the direct output of the wavelet transform; by squaring and summing these coefficients, the energy distribution of the signal at different frequency ranges, i.e., different scales, can be calculated. This energy distribution can intuitively reflect the response intensity of the wall's internal structure to sound waves of different frequencies under impact. For example, hollow areas may exhibit abnormally high energy in a specific low-frequency band, while solid areas may have a more uniform energy distribution in higher-frequency bands.

[0069] Through the above technical solution, this application overcomes the limitations of traditional fixed-parameter multi-scale spectral decomposition methods when processing complex wall response signals. Due to the varying non-stationarity and nonlinear characteristics of wall impact response signals, the dynamic selection of wavelet basis functions and decomposition levels allows the spectral decomposition process to adaptively match the inherent characteristics of the signal, thereby significantly improving the extraction accuracy and robustness of energy distribution across different frequency ranges. This more accurate energy distribution information can more effectively highlight the subtle differences in acoustic response between hollow defects and solid walls, thus improving the accuracy of extracting fine acoustic information reflecting the internal structural characteristics of the wall from the target wall response signal, ultimately contributing to improved reliability in wall condition assessment.

[0070] In some embodiments, the step of performing time-domain waveform feature analysis on the target wall response signal to obtain the envelope shape parameters and transient impact characteristics of the target wall response signal includes: The target wall response signal is divided into time-varying feature windows, and the length and position of the time-varying feature windows are adaptively adjusted according to the instantaneous energy and frequency change rate of the signal. Within each time-varying feature window, the local envelope shape parameters of the signal are calculated, including the local rise time, local decay time, and local peak duration. Within each time-varying feature window, the local transient impulse characteristics of the signal are calculated, including the local impulse factor and local kurtosis. Based on the changing trends of the local envelope shape parameters and the local transient impact characteristics within different time-varying feature windows, the envelope shape parameters and transient impact characteristics of the target wall response signal are identified.

[0071] Specifically, when performing time-domain waveform feature analysis on the target wall response signal, the signal is first divided into multiple time-varying feature windows. The length and position of these windows are not fixed but adaptively adjusted based on the instantaneous energy and frequency change rate of the signal. For example, when the signal energy is high or the frequency changes drastically, the feature window can be set shorter to capture finer transient features; conversely, when the signal is relatively stable, the feature window can be set longer to obtain more stable overall features. This adaptive division method ensures optimal feature extraction results across different signal segments.

[0072] Secondly, within each time-varying feature window, the local envelope shape parameters of the signal are calculated. These parameters specifically include local rise time, local decay time, and local peak duration. Local rise time reflects the time required for the signal to rise from its initial point to its peak, local decay time represents the time required for the signal to fall from its peak to a certain level, and local peak duration describes the duration of the signal's peak region. These parameters can meticulously characterize the energy change process and sustained characteristics of the wall's response signal. Simultaneously, within each time-varying feature window, the signal's local transient impact characteristics are also calculated, including local impulse factor and local kurtosis. The local impulse factor measures the intensity of the impulse component in the signal, while local kurtosis reflects the sharpness or impact of the signal waveform. These parameters are crucial for identifying the unique acoustic responses caused by internal structural defects in the wall.

[0073] Finally, by analyzing the changing trends of the calculated local envelope shape parameters and local transient impact characteristics within different time-varying feature windows, the overall envelope shape parameters and transient impact characteristics of the target wall response signal can be identified. This trend analysis can comprehensively consider the local characteristics of the signal in different time periods, thus obtaining a more comprehensive and robust global feature description.

[0074] The above technical solution effectively addresses the limitations of traditional fixed-window analysis methods in processing non-stationary wall response signals, improving the accuracy and robustness of time-domain waveform feature extraction. This adaptive and multi-dimensional feature extraction method enables the acquired envelope shape parameters and transient impact characteristics to more realistically and comprehensively reflect the internal structural characteristics of the wall, thereby significantly improving the accuracy and reliability of wall surface hollow acoustic signature recognition.

[0075] In some embodiments, the step of identifying the envelope shape parameters and transient impact characteristics of the target wall response signal based on the changing trends of the local envelope shape parameters and the local transient impact characteristics within different time-varying feature windows includes: The changing trend is segmented, and within each segment, the local envelope shape parameters and the local transient impact characteristics are smoothed to eliminate local abnormal fluctuations, thus obtaining the smoothing result for each segment. Calculate the feature description quantity of each segment based on the smoothing result of each segment; Weights are assigned to the feature descriptions of the segments, and weighted fusion is performed to obtain a weighted fusion result. The weights are determined based on the length of the segments and the stability of the features. Based on the weighted fusion results, the envelope shape parameters and transient impact characteristics of the wall response signal are identified.

[0076] The segmentation of the aforementioned trend involves dividing the sequence of local envelope shape parameters and local transient impact characteristics obtained from different time-varying feature windows into several continuous sub-intervals based on their temporal order or characteristic change patterns. For example, segmentation can be based on fluctuations in signal energy, changes in frequency components, or preset time intervals. The aim is to decompose the complex overall trend into more easily analyzed and processed local segments, enabling optimized processing of signal characteristics at different stages.

[0077] Within each segment, the local envelope shape parameters and local transient impact characteristics are smoothed to eliminate local abnormal fluctuations, resulting in a smoothed result for each segment. Various techniques can be employed for smoothing, such as moving average filtering, Gaussian filtering, median filtering, or wavelet denoising. The aim is to remove random noise and transient interference from the data, making the changing trends of local features clearer and more stable, thereby improving the accuracy of subsequent feature extraction.

[0078] Based on the smoothing results of each segment, the feature descriptors of each segment are calculated. These feature descriptors are indicators that quantify and summarize the local envelope shape parameters and local transient impact characteristics after smoothing. For example, the mean, variance, slope, peak value, energy, or intensity of a specific frequency component can be calculated for each segment. The purpose is to condense the complex data of each segment into concise and representative values, facilitating subsequent comparison and fusion.

[0079] Weights are assigned to the feature descriptors of the segments, and weighted fusion is performed to obtain a weighted fusion result. The weights are determined based on the segment length and feature stability. For example, longer segments typically contain more information and may have more stable statistical characteristics, thus they can be assigned higher weights. Feature stability can be measured by the variance or standard deviation of the feature descriptors within a segment; a smaller variance indicates more stable features, and the weight can be increased accordingly. Weighted fusion is achieved by multiplying the feature descriptors of each segment by their corresponding weights and then summing and averaging the results. The purpose is to comprehensively consider the reliability and importance of different segments, so that the final fusion result can more comprehensively and accurately reflect the overall characteristics of the wall response signal.

[0080] Through the above technical solution, this application can significantly improve the accuracy and robustness of identifying the envelope shape parameters and transient impact characteristics of wall response signals. Compared with direct identification based on the original change trend, this technical solution effectively eliminates local abnormal fluctuations through piecewise smoothing processing, reducing the impact of noise on the identification results. Furthermore, by introducing a weight allocation mechanism based on segment length and feature stability, it ensures that more reliable and representative signal features play a greater role in the final identification, thereby avoiding misjudgments caused by local uncertainties. This makes the extracted fine acoustic information more accurate, thereby improving the accuracy and reliability of the deep verification process when the confidence level of the initial wall condition judgment result is low, providing solid technical support for the accurate judgment of the final wall condition.

[0081] This application also proposes a deep learning-based system for recognizing the acoustic signature features of hollow walls, such as... Figure 2 As shown, a deep learning-based wall hollow soundprint feature recognition system 100 includes: The signal acquisition and processing module 10 is used to acquire mixed acoustic signals generated during the wall-tapping process in real time, identify the tapping sound segments in the mixed acoustic signals, and extract the basic acoustic features of the tapping sound segments. The rapid judgment and confidence assessment module 20 is used to make a preliminary judgment on the wall condition at the current tapping point based on the basic acoustic features and a pre-trained lightweight rapid judgment model, and to generate a judgment confidence level based on the preliminary judgment result; wherein, the wall condition includes one of solid and hollow. The preliminary judgment result output module 30 is used to directly generate the final detection result based on the preliminary judgment result when the judgment confidence reaches a preset confidence threshold. The deep verification module 40 is used to initiate the deep verification process when the confidence level does not reach the preset confidence threshold, and generate the final detection result after performing noise separation and fine feature extraction on the mixed acoustic signal.

[0082] The deep learning-based acoustic signature recognition system for hollow walls proposed in this application significantly improves system performance by introducing a hierarchical judgment mechanism, combining rapid preliminary judgment with deep verification processes. Compared to existing technologies, this system can quickly provide results when confidence is high, improving detection efficiency; while when confidence is insufficient, it can activate the deep verification module for more refined noise separation and feature extraction, effectively addressing interference from high-decibel background noise and sudden impact noise, and reducing the false judgment rate. This intelligent hierarchical processing strategy enables the system to significantly improve detection accuracy and reliability in complex and variable on-site environments while ensuring detection efficiency, providing a more advanced and reliable solution for hollow wall detection. The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. For those skilled in the art, this application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for recognizing the acoustic signature features of hollow walls based on deep learning, characterized in that, include: The system collects mixed acoustic signals generated during wall tapping in real time, identifies the tapping sound segments in the mixed acoustic signals, and extracts the basic acoustic features of the tapping sound segments. Based on the aforementioned basic acoustic features and a pre-trained lightweight fast judgment model, a preliminary judgment of the wall condition is made at the current tapping point, and a judgment confidence level is generated based on the preliminary judgment result; the wall condition includes one of solid and hollow. If the confidence level reaches a preset confidence threshold, the final detection result is generated directly based on the preliminary judgment result. If the confidence level does not reach the preset confidence threshold, the deep verification process is initiated to perform noise separation and fine feature extraction on the mixed acoustic signal to generate the final detection result.

2. The method for identifying acoustic signatures of hollow walls according to claim 1, characterized in that, If the confidence level does not reach the preset confidence threshold, the deep verification process is initiated. The steps for generating the final detection result after noise separation and fine feature extraction of the mixed acoustic signal include: The mixed acoustic signal is subjected to noise separation processing to obtain a clean wall response signal; Acquire the impact force information, and adjust the amplitude compensation of the wall response signal according to the impact force information to obtain the target wall response signal, so as to compensate for the distortion caused by inconsistent impact actions; Fine acoustic information reflecting the internal structural characteristics of the wall is extracted from the response signal of the target wall, and correlation analysis is performed by combining the basic acoustic characteristics of adjacent impact points with the wall condition to obtain the correlation analysis results. Based on the detailed acoustic information and the correlation analysis results, a final judgment result on the wall condition is generated.

3. The method for identifying acoustic signatures of hollow walls according to claim 1, characterized in that, After the step of making a preliminary judgment on the wall condition at the current tapping point based on the basic acoustic features and a pre-trained lightweight fast judgment model, and generating a judgment confidence score based on the preliminary judgment result; wherein the wall condition includes one of solid and hollow, the method further includes: Obtain information on the wall material type of the current detection area, and query the acoustic response characteristics and noise sensitivity parameters corresponding to the wall material type information based on the wall material type information; Based on the acoustic response characteristics and the noise sensitivity parameters, the preset confidence threshold is adjusted so that the preset confidence threshold matches the acoustic behavior characteristics of the current wall material. The acoustic behavior characteristics are the inherent acoustic response patterns exhibited by the wall material after being struck, including the propagation characteristics of the striking sound, the energy attenuation rate, and the spectral distribution pattern.

4. The method for identifying acoustic signatures of hollow walls according to claim 2, characterized in that, The steps of acquiring the impact force information and adjusting the wall response signal based on the impact force information to obtain the target wall response signal, in order to compensate for the distortion caused by inconsistent impact actions, include: Obtain the magnitude of the striking force, the striking angle deviation, the striking contact time, and the striking contact area; The overall amplitude of the wall response signal is compensated based on the magnitude of the striking force. The spectral distribution of the wall response signal is adjusted based on the tapping angle deviation and the tapping contact time; The local energy distribution of the wall response signal is corrected based on the impact contact area to obtain the target wall response signal.

5. The method for identifying acoustic signatures of hollow walls according to claim 2, characterized in that, The step of extracting fine acoustic information reflecting the internal structural characteristics of the wall from the target wall response signal includes: The target wall response signal is subjected to multi-scale spectral decomposition to obtain the energy distribution in different frequency ranges; Time-domain waveform feature analysis is performed on the target wall response signal to obtain the envelope shape parameters and transient impact characteristics of the target wall response signal; The energy distribution, envelope shape parameters, and transient impact characteristics are compared with pre-stored typical acoustic feature patterns to identify the structural tendency of the target wall response signal and to evaluate the degree of distinction between the structural tendency and the actual hollow defects; wherein, the structural tendency is the tendency of the internal structure type of the wall reflected by the target wall response signal, including hollow defect tendency or solid wall structure tendency. Based on the structural tendency and degree of differentiation, fine acoustic information reflecting the internal structural characteristics of the wall is generated.

6. The method for identifying acoustic signatures of hollow walls according to claim 2, characterized in that, The step of performing correlation analysis by combining the basic acoustic characteristics of adjacent impact points with the wall condition to obtain the correlation analysis results includes: Acquire acoustic information of the set of adjacent striking points within a preset range around the current striking point; The confidence level of the acoustic information of each striking point in the set of adjacent striking points is obtained, and the confidence level includes high confidence level, medium confidence level and low confidence level. Based on the judgment confidence level, a contribution weight is assigned to each tap point in the set of adjacent tap points. The contribution weight includes a high contribution weight, a medium contribution weight, and a low contribution weight. Tap points with high judgment confidence are assigned a high contribution weight, tap points with medium judgment confidence are assigned a medium contribution weight, and tap points with low judgment confidence are assigned a low contribution weight. The fine acoustic information of the current tapping point is fused with the weighted acoustic information of the adjacent tapping points to generate a weighted correlation analysis result. The wall status is then analyzed based on the weighted correlation analysis result to obtain the correlation analysis result.

7. The method for identifying acoustic signatures of hollow walls according to claim 5, characterized in that, The step of performing multi-scale spectral decomposition on the target wall response signal to obtain the energy distribution in different frequency ranges includes: A preliminary analysis of the target wall response signal was performed to obtain the non-stationarity and nonlinear characteristics of the signal. Based on the degree of nonstationarity and the nonlinear characteristics, the wavelet basis function and the number of decomposition layers are dynamically selected. Based on the wavelet basis function and the number of decomposition layers, wavelet transform is performed on the target wall response signal to obtain the spectral coefficients at different scales; The energy distribution within different frequency ranges is calculated based on the spectral coefficients.

8. The method for identifying acoustic signature features of hollow walls according to claim 5, characterized in that, The step of performing time-domain waveform feature analysis on the target wall response signal to obtain the envelope shape parameters and transient impact characteristics of the target wall response signal includes: The target wall response signal is divided into time-varying feature windows, and the length and position of the time-varying feature windows are adaptively adjusted according to the instantaneous energy and frequency change rate of the signal. Within each time-varying feature window, the local envelope shape parameters of the signal are calculated, including the local rise time, local decay time, and local peak duration. Within each time-varying feature window, the local transient impulse characteristics of the signal are calculated, including the local impulse factor and local kurtosis. Based on the changing trends of the local envelope shape parameters and the local transient impact characteristics within different time-varying feature windows, the envelope shape parameters and transient impact characteristics of the target wall response signal are identified.

9. The method for identifying acoustic signatures of hollow walls according to claim 8, characterized in that, The step of identifying the envelope shape parameters and transient impact characteristics of the target wall response signal based on the changing trends of the local envelope shape parameters and the local transient impact characteristics within different time-varying characteristic windows includes: The changing trend is segmented, and within each segment, the local envelope shape parameters and the local transient impact characteristics are smoothed to eliminate local abnormal fluctuations, thus obtaining the smoothing result for each segment. Calculate the feature description quantity of each segment based on the smoothing result of each segment; Weights are assigned to the feature descriptions of the segments, and weighted fusion is performed to obtain a weighted fusion result. The weights are determined based on the length of the segments and the stability of the features. Based on the weighted fusion results, the envelope shape parameters and transient impact characteristics of the wall response signal are identified.

10. A deep learning-based system for recognizing the acoustic signature of hollow walls, characterized in that, The system includes: The signal acquisition and processing module is used to collect mixed acoustic signals generated during the wall-tapping process in real time, identify the tapping sound segments in the mixed acoustic signals, and extract the basic acoustic features of the tapping sound segments. The rapid judgment and confidence assessment module is used to make a preliminary judgment on the wall condition at the current tapping point based on the basic acoustic features and a pre-trained lightweight rapid judgment model, and to generate a judgment confidence level based on the preliminary judgment result; wherein, the wall condition includes one of solid and hollow. The preliminary judgment result output module is used to directly generate the final detection result based on the preliminary judgment result when the judgment confidence reaches a preset confidence threshold; The deep verification module is used to initiate the deep verification process when the confidence level does not reach the preset confidence threshold, and to generate the final detection result after performing noise separation and fine feature extraction on the mixed acoustic signal.

Images