Multi-classification detection method for wall hollowing
By fusing the Mel spectrum features and the Mel cepspectral coefficient features and optimizing the support vector machine using Bayesian algorithm, the MFCC+Mel-Bo-SVM model was established, which solved the limited detection range and misjudgment in wall hollow detection. The problem of high rate achieves higher detection accuracy and simplified detection process.
Patent Information
- Application Number
- CN202411053189.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the wall hollow detection, the prior art has problems such as limited detection range, strong subjectivity, high misjudgment rate, and lack of effective classification methods.
The fusion of Mel spectral features and Mel cepspectral coefficient features is adopted, combined with Bayesian algorithm optimization support vector machine, and the MFCC+Mel-Bo-SVM algorithm model is established to conduct multi-classification detection of wall hollows.
It improves the accuracy and robustness of the inspection, reduces the misjudgment rate, is suitable for a variety of wall materials, and simplifies the inspection process, enhancing the efficiency and safety of the inspection.
Smart Images

Figure CN119064477A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of wall hollowing detection, and in particular to a multi-classification detection method for wall hollowing. Background Art
[0002] Wall hollowing is a common problem in construction projects that needs to be solved urgently, especially in concrete construction and decoration. Hollowing not only affects the appearance of the building, but may also pose a threat to structural safety, leading to a series of quality problems during subsequent use. Therefore, how to effectively detect and classify wall hollowing problems has become an important part of construction project quality control.
[0003] Huang Jizhong [1] Ultrasonic imaging devices are used to image hollows in the outer wall enclosures of cultural relics, but their detection range and depth are limited, and they are also affected by environmental factors such as humidity, temperature, and noise. Hu Yuanxin et al. [2] It is proved that there is a difference between the percussion sound of hollow drum and non-hollow drum area, but there is a lack of systematic judgment method; Wang Junwei et al. [3] The accuracy of using the vibration characteristics of the sound of knocking on the wall to judge the hollow drum situation is still insufficient; Sun Qianyu et al. [4] Establish wall hollowing defect model; Yuan Lin et al. [5] VMD is used to detect hollowing inside wood, but it has not been verified for hollowing in different materials of building exterior walls.
[0004] Liu Sisi et al. [6] The optimized Mel cepstral coefficients were selected as features to identify abnormal noise of the window motor; With the vigorous development of machine learning, Wang Tianrui et al. [7] Using MFC to identify environmental sounds, defect detection methods based on machine learning have gradually been applied, which can achieve good results in processing images with a lot of noise and complex backgrounds. Many scholars at home and abroad regard defect detection as a classification problem, and many classification models are used in defect recognition. Shen Zhaoqing [8] et al. proposed using support vector machine (SVM) for defect recognition, Qu Ziyi [9] et al. proposed a multi-classification detection method for wall hollowing, support vector machine (SVM), combined with other algorithms to improve the detection accuracy. Saleem
[10] etc. use artificial neural networks to identify various defects.
[11] et al. combined artificial neural network and logistic regression algorithm to eliminate noise very effectively to identify cracks.
[0005] To sum up, although there are many methods for wall hollowing detection, there are still problems such as limited detection range, strong subjectivity, high misjudgment rate, and lack of effective classification methods. Therefore, it is necessary to provide a multi-classification detection method for wall hollowing to solve the above technical problems. Summary of the invention
[0006] The purpose of the present invention is to provide a multi-classification detection method for wall hollowing so as to solve the problems raised in the above-mentioned background technology, such as limited detection range of wall hollowing, strong subjectivity, high misjudgment rate and lack of classification method. The technical solution of the present invention aims at the technical problem that the existing technical solution is too single, and provides a solution that is significantly different from the existing technology.
[0007] To achieve the above object, the present invention provides the following technical solution: a multi-classification detection method for wall hollowing, comprising the following steps:
[0008] S100. Collect the hollow drum audio and non-hollow drum audio on the wall;
[0009] S200. Extracting Mel spectrum features and Mel cepstral coefficient features of audio;
[0010] S300. Fusing the extracted Mel spectrum feature and Mel cepstral coefficient feature to obtain MFCC+Mel fusion feature, where Mel is the Mel spectrum feature and MFCC is the Mel cepstral coefficient feature;
[0011] S400. Use the Bayesian algorithm to optimize the support vector machine, input the fusion features into the SVM algorithm model after Bayesian optimization, and establish the MFCC+Mel-Bo-SVM algorithm model;
[0012] S500. Based on the MFCC+Mel-Bo-SVM algorithm model, training is performed to obtain the classification result of wall hollowing.
[0013] In a further embodiment, the collected audio signal is pre-emphasized to enhance the high frequency components of the audio signal and improve the signal-to-noise ratio, and then frame processing is performed to capture the time-varying characteristics of the audio signal, and finally windowing pre-processing is performed to reduce edge effects.
[0014] In a further embodiment, a frequency domain signal is obtained by performing fast Fourier transform and short-time Fourier transform analysis on each frame of the audio signal preprocessed by windowing, so as to comprehensively analyze the frequency domain characteristics of the signal and extract the frequency components and time resolution.
[0015] In a further embodiment, after performing a fast Fourier transform or a short-time Fourier transform on each frame of the audio signal, the energy of each frequency component is calculated to obtain the square value of the frequency domain signal, and the calculated energy spectrum is filtered to capture the frequency domain features in the extracted audio signal.
[0016] In a further embodiment, the signal after short-time Fourier transform and energy spectrum calculation filtering can obtain MFCC features through discrete cosine transform; the signal after fast Fourier transform and energy spectrum calculation filtering can obtain MEL features through logarithmic processing.
[0017] In a further embodiment, the MEL feature and the MFCC feature are normalized and then connected in series at the frame level to obtain a mixed cepstrum feature, and the fused signal is visualized to obtain a fusion graph of hollow drum feature and non-hollow drum feature;
[0018] The normalization formula is:
[0019]
[0020] In formula (1), MFCC norm is the normalized MFCC feature, μMFCC and σMFCC are the mean and standard deviation of the MFCC feature respectively;
[0021] In formula (2), Mel norm is the normalized MEL feature, μMel and σMel are the mean and standard deviation of the Mel spectrum feature respectively;
[0022] The frame-level concatenation formula is:
[0023] Combined Features = [MFCC norm [Y]|Mel norm [Y]](3)
[0024] In formula (3), Combined Features is the fusion feature MFCC of MFCC and Mel norm [Y] and Mel norm [Y] are the MFCC features and Mel features after normalization of the Yth dimension.
[0025] In a further embodiment, a Bayesian algorithm is used to optimize support vector machine parameters to improve classification accuracy and reliability, the support vector machine maps data to a high-dimensional space through a kernel function to find the optimal classification plane, and Bayesian optimization is used to quickly and globally search for the best parameter combination.
[0026] In a further embodiment, the process of optimizing the support vector machine using the Bayesian algorithm is:
[0027] S401. The optimization ranges of the regularization parameter c and the radial basis kernel function parameter g are set to 0.1 to 1000 and 0.001 to 1, respectively. First, the initial c and g combination parameters are randomly selected within the optimization range;
[0028] S402. Use five-fold cross validation to evaluate the model performance with the accuracy of the sample test as the objective function f, and continuously update the model by adding the optimal model parameters that improve the objective function to the initial set;
[0029] S403. Set the iteration termination condition: when the number of iterations reaches the upper limit or the objective function reaches the optimal value, stop the iteration and output the optimal combination parameters c and g.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] 1. The present invention fuses the Mel spectrum feature (Mel) and the Mel cepstral coefficient feature (MFCC). Compared with the existing technology, it can combine the advantages of both and provide a more comprehensive audio feature representation. The Mel spectrum feature (Mel) provides detailed information of the spectrum, while the Mel cepstral coefficient feature (MFCC) provides compact features in the time domain. By fusing these two features, the way the human auditory system processes audio signals can be better simulated, and the robustness of the model to noise and distortion can be improved.
[0032] 2. The present invention optimizes the parameters c and g of the support vector machine (SVM) through the Bayesian optimization (Bo) algorithm, and can quickly and effectively find the best parameter combination. The Bayesian optimization uses a probabilistic proxy model and an acquisition function to guide the search by effectively searching the hyperparameter space, and uses the existing parameter information to determine the next evaluation point. Compared with the existing technology, it effectively avoids a large number of useless attempts and can usually find the optimal solution within a smaller number of sampling times, thereby improving the classification accuracy and generalization ability of the support vector machine (SVM) model.
[0033] 3. The present invention verifies the effectiveness of Bayesian optimization (Bo) and support vector machine (SVM) in the wall hollowing detection task through experiments, and the experimental results show that the support vector machine (SVM) model after Bayesian optimization makes the accuracy of the algorithm model (MFCC+Mel-Bo-SVM) on the test set reach 96.36%, and compared with models such as SVM, random forest, k-nearest neighbor algorithm, grid search optimized support vector machine (SVM), and chaotic particle swarm optimized support vector machine (SVM), the accuracy of the algorithm model (MFCC+Mel-Bo-SVM) is improved by 6.61%, 9.58%, 15.27%, 13.9%, and 5.02%, respectively, which improves the effectiveness and practicality of Bayesian optimization (Bo) and support vector machine (SVM) in the field of wall hollowing detection.
[0034] 4. The present invention, through the constructed algorithm model (MFCC+Mel-Bo-SVM) multi-classification model, classifies the hollowing and non-hollowing of cement walls, paint walls, marble walls, and tiles. Compared with various traditional methods, the accuracy rate is significantly improved to 96.36%.
[0035] 5. The technical solution proposed in the present invention is not only applicable to various wall materials such as cement walls, paint walls, marble walls, and tile walls, but also to other occasions where defect detection by sound is required. Compared with the existing technology, it increases the applicability of practical applications.
[0036] 6. The present invention simplifies the traditional hollow drum detection process and reduces the complexity and uncertainty of manual operation by effectively processing and extracting features of audio signals. Compared with the existing technology, this technical solution improves the efficiency of wall hollow drum detection, shortens the detection time, reduces the workload of manual detection, and simplifies the detection process.
[0037] 7. The present invention can effectively reduce the false positive rate by improving the detection accuracy, thereby reducing the potential safety hazards of building structures caused by missed detection or false detection. The technical solution can also be extended to other defect detection fields based on sound signals, thus having wide versatility.
[0038] In summary, the inventions and innovations in this application have significant advantages in practical applications, such as improving detection accuracy, reducing manual operations, shortening detection time, and enhancing safety, which all help to improve the technical level and efficiency of related industries. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 This is the flow chart of MFCC+Mel-Bo-SVM;
[0040] Figure 2It is the MFCC+Mel mixed feature flow chart;
[0041] Figure 3 Examples of MFCC+Mel fusion feature maps for various wall audios;
[0042] (a) is a schematic diagram of hollowing of cement wall.
[0043] (b) Schematic diagram of non-hollow cement wall.
[0044] (c) Schematic diagram of hollowing of paint wall.
[0045] (d) Schematic diagram of non-hollowing paint wall.
[0046] (e) Schematic diagram of the hollowing of the marble wall.
[0047] (f) Schematic diagram of a marble wall without hollowing.
[0048] (g) Schematic diagram of hollowing of tile wall.
[0049] (h) Schematic diagram of non-hollowing ceramic tile wall.
[0050] Figure 4 This is a schematic diagram of a hollow drum striking device;
[0051] Figure 5 Schematic diagram of BO optimization fitness curve;
[0052] Figure 6 Schematic diagram of hollow drum recognition results for each model;
[0053] (a) is a schematic diagram of hollowing of cement wall.
[0054] (b) Schematic diagram of non-hollow cement wall.
[0055] (c) Schematic diagram of hollowing of paint wall.
[0056] (d) Schematic diagram of non-hollowing paint wall.
[0057] (e) Schematic diagram of the hollowing of the marble wall.
[0058] (f) Schematic diagram of a marble wall without hollowing.
[0059] (g) Schematic diagram of hollowing of tile wall.
[0060] (h) Schematic diagram of non-hollowing tile wall.
[0061] Figure 7 This is a diagram of the confusion matrix; Figure 8 Flowchart of the Bayesian optimization algorithm. DETAILED DESCRIPTION
[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0063] like Figure 1 As shown, the present invention provides a multi-classification detection method for wall hollowing, the steps are as follows:
[0064] S100. Collect the hollow and non-hollow audio of the wall; common wall types include: cement wall, paint wall, marble wall, tile wall, etc.
[0065] In a further optimized embodiment, the percussion sound feature is used to extract the audio. The extraction of the sound feature can retain the effective part for audio classification, and the fusion of different sound features can more comprehensively reflect the useful information in the sound signal.
[0066] S200. Extract the Mel spectrum features (Mel) and Mel cepstral coefficient features (MFCC) of the audio.
[0067] In the field of sound processing, Mel-Frequency Cepstral Coefficients (MFCC) features occupy an important position. This feature extraction method is achieved by linearly transforming the logarithmic energy spectrum of the sound frequency on a nonlinear mel scale. The purpose of this is to better simulate the response of the human auditory system.
[0068] In addition to the Mel-scale Spectrum Features (MFCC), Mel-scale Spectrum Features (hereinafter referred to as Mel) is another important feature in sound processing. It is based on the nonlinear perception of frequency by the human auditory system, namely the mel scale. Although Mel-scale features (Mel) are usually discussed together with Mel-frequency cepstral coefficients (MFCC), they differ in extraction and processing methods. Hollow drum audio is also a type of sound signal. Therefore, by extracting MFCC features and MEL features from the collected sound signals and fusing them, the human auditory system can be better simulated, thereby improving the recognition accuracy of the model. The collected audio is extracted from Mel-scale features (Mel) and Mel-scale cepstral coefficient features (MFCC) and then fused. The fusion process is as follows: Figure 2 As shown, the specific fusion process is shown in step S300.
[0069] S300. Fuse the extracted Mel spectrum feature and Mel cepstral coefficient feature to obtain MFCC+Mel fusion feature, where Mel is the Mel spectrum feature and MFCC is the Mel cepstral coefficient feature.
[0070] S301. Pre-emphasis: Pre-emphasis is to make the high-frequency components more prominent by amplifying the high-frequency parts, just like "brightening" the high-frequency parts of the sound, which helps to enhance the high-frequency components and improve the signal-to-noise ratio;
[0071] S302. Framing: Framing is to cut a long audio signal into small segments, each of which lasts for a very short time. This can capture the characteristics of the audio signal changing over time and improve the precision of the analysis;
[0072] S303. Windowing: Windowing is to smooth each small segment of the frame to avoid the sudden change during cutting and make the edge smoother. This can reduce the edge effect and improve the accuracy of spectrum analysis. Through the above processing, the useful information in the audio signal can be enhanced, the influence of noise and distortion can be reduced, and a good foundation can be laid for subsequent feature extraction and pattern recognition. The signal adopts overlapping sampling with a frame length of 25ms and a frame length of 10ms in the overlapping area.
[0073] S304. Calculate the power spectrum by fast Fourier transform and short-time Fourier transform.
[0074] Fast Fourier transform (FFT) and short-time Fourier transform (STFT) are performed on each frame of signal to obtain a frequency domain signal in order to comprehensively analyze the frequency domain characteristics of the signal. FFT is used to convert the signal from the time domain to the frequency domain, and can extract the frequency components of the entire signal, but lacks time resolution.
[0075] The short-time Fourier transform (STFT) divides the signal into multiple short time periods for Fourier transform, thereby providing time resolution while retaining the frequency domain information. Therefore, by performing fast Fourier transform (FFT) and short-time Fourier transform (STFT) respectively, a more comprehensive frequency domain characteristic can be obtained, taking into account both frequency resolution and time resolution. All original signals are processed by frame division, and then the Mel-frequency cepstral coefficient feature (MFCC) and Mel-frequency spectrum feature (Mel) are extracted respectively, that is, all signal frames are used to calculate these two types of features. Through this processing method, different features can be extracted from the same signal, making full use of the information in the signal and improving the effect of subsequent analysis and recognition.
[0076] S305. Calculate the power spectrum of each frame and apply a Mel filter. Calculate the power spectrum of each frame and apply a Mel filter.
[0077] The energy spectrum of each frame of signal is calculated and filtered in order to better analyze the frequency domain characteristics of the signal. First, the signal is converted to the frequency domain through fast Fourier transform (FFT) or short-time Fourier transform (STFT), and the energy (power) of each frequency component, that is, the square value of the frequency domain signal, is calculated. Next, a Mel filter is applied to the calculated power spectrum to map the frequency domain signal to the Mel scale, which is a scale that is closer to the frequency resolution perceived by the human ear and can better capture the features in the audio signal. By calculating the power spectrum of each frame and applying a suitable filter, more representative frequency domain features in the signal can be extracted, thereby improving the effect of subsequent analysis and processing.
[0078] S306. Discrete cosine transform (DCT) and logarithm are used to obtain MFCC and Mel features.
[0079] After the signal is filtered by short-time Fourier transform (STFT) and energy spectrum calculation, the Mel-frequency cepstral coefficient feature (MFCC) can be obtained through discrete cosine transform (DCT). The function of discrete cosine transform (DCT) is to convert the signal from the frequency domain to the cepstral domain. On the one hand, it can compress the feature dimension and decorrelate, making the signal more compact and independent in the feature space, which helps to improve the recognition performance of the model. On the other hand, after the signal is filtered by fast Fourier transform (FFT) and energy spectrum calculation, the Mel-frequency spectrum feature (Mel) can be obtained through logarithmic processing. The role of logarithm is to convert multiplicative noise into additive noise and compress the dynamic range of the feature, making the feature changes in the signal more significant and easier to distinguish, thereby improving the subsequent processing and recognition effects. Through these two processing methods, more representative and robust audio features can be extracted.
[0080] S307. Mel spectrum features and MFCC features are fused in series at the frame level, and mixed cepstrum features are obtained after normalization.
[0081] Among them, the MFCC features and Mel spectrum features are normalized:
[0082]
[0083] In formula (1), MFCC norm is the normalized MFCC feature, μMFCC and σMFCC are the mean and standard deviation of the MFCC feature respectively;
[0084] In formula (2), Mel norm is the normalized MEL feature, μMel and σMel are the mean and standard deviation of the Mel spectrum feature, respectively.
[0085] The fusion feature is to concatenate the normalized MFCC features and Mel spectrum features at the frame level:
[0086] Combimed Features = [MFCC norm [Y][Mel norm [Y]] (3)
[0087] Among them, MFCC norm [Y] and Mel norm [Y] are the normalized MFCC and Mel spectrum features of the Yth dimension.
[0088] Through the above steps, the original signal is processed into a mixed feature containing Mel spectrum features and MFCC features, and is normalized.
[0089] The fused signal is a two-dimensional matrix that concatenates the Mel spectrum features (MEL) and the Mel cepstral coefficient features (MFCC). Each row of this matrix represents a frame, and each column represents a feature (including Mel spectrum features (Mel) and Mel cepstral coefficient features (MFCC) coefficients). The imshow function of Matplotlib is used for visualization. This function takes the two-dimensional matrix as input and draws a color image, in which the values of the matrix are mapped to the color space. The feature fusion diagram of the hollow and non-hollow walls of cement walls, paint walls, marble walls, and tiles is shown in the figure. Figure 3 (a), (b), (c), (d), (e), (f), (g) and (h) in the figure.
[0090] The fusion of Mel spectrum features (Mel) and Mel cepstral coefficient features (MFCC) can retain the spectral details and the compact features of decorrelation. These features retain the detailed spectral information of the audio signal. By capturing the energy of different frequency bands, the details of various frequency components in the audio signal can be reflected. Combining Mel spectrum features (Mel) and Mel cepstral coefficient features (MFCC), the detailed information in the spectrum graph is retained, and the compact and decorrelation features of Mel cepstral coefficient features (MFCC) are introduced. This multi-level information capture method makes the feature dimension richer and can cover more signal information. At the same time, the Mel spectrum graph provides detailed information in the frequency domain, and the Mel cepstral coefficient feature (MFCC) feature further extracts the time domain features of the signal through discrete cosine transform (DCT). The combination of the two enables the model to obtain information from both the time domain and the frequency domain, improving the comprehensiveness of the feature representation. The decorrelation of the Mel cepstral coefficient feature (MFCC) reduces the redundant information between features, so that the model can better adapt to the main components of the signal during training and is not easily disturbed by noise. The independence of this feature improves the robustness of the model to noise and distortion.
[0091] S400. Use the Bayesian algorithm to optimize the support vector machine, input the fusion features into the SVM algorithm model after Bayesian optimization, and establish the MFCC+Mel-Bo-SVM algorithm model.
[0092] The basic idea of support vector machine (SVM) is a commonly used classification algorithm. It maps nonlinear data to a high-dimensional feature space by using a kernel function, and performs inner product operations between vectors in this space. The goal of support vector machine (SVM) is to find an optimal classification hyperplane that can maximize the interval between data of different categories, thereby effectively performing classification and recognition. The two important parameters of the support vector machine (SVM) model are the radial basis kernel function parameter g and the regularization (penalty) coefficient c. g controls the width of the kernel function, and c controls the model complexity and tolerance for misclassification. Choosing appropriate c and g is crucial to improving the classification accuracy and reliability of the support vector machine (SVM) model.
[0093] Bayesian optimization is an efficient global optimization algorithm that uses Bayesian statistical theory to optimize the objective function. The core idea of this algorithm is to use the existing parameter information to determine the next evaluation point, thereby effectively searching the hyperparameter space and avoiding a large number of useless attempts. Specifically, the Bayesian optimization algorithm is mainly composed of two important parts: the probabilistic agent model and the acquisition function.
[0094] The probabilistic proxy model is used to predict the value of the objective function. It is based on historical data and Bayesian theory. Through this model, the objective function value under different parameter combinations can be estimated, thereby predicting the possible position of the optimal parameter combination.
[0095] The acquisition function determines the next evaluation point based on the prediction of the probabilistic proxy model. In Bayesian optimization, the acquisition function selects the parameter combination that can provide the most information for evaluation in order to better update the probabilistic proxy model. This selection process is based on the criterion of maximizing expected improvement, that is, selecting the parameter combination that is most likely to bring performance improvement.
[0096] Through this efficient search strategy, Bayesian optimization can usually find the optimal solution within a smaller number of sampling times, which makes it very effective in fields such as hyperparameter optimization. Especially when faced with complex optimization problems, it can quickly find a solution close to the global optimal solution, thereby improving the performance and stability of the model.
[0097] The Bayesian optimization process is as follows Figure 1 As shown, the Bayesian optimization algorithm process is as follows Figure 8 As shown;
[0099] S401. Set the optimization ranges of parameters c and g to 0.1-1000 and 0.001-1 respectively, and first randomly select initial c and g combination parameters within the optimization range.
[0100] The optimization range of parameter c is set to 0.1-1000 and the optimization range of parameter g is set to 0.001-1 in order to find the optimal parameter combination in the support vector machine (SVM) model and improve the model performance. Parameter c controls the misclassification tolerance. Smaller c values improve the generalization ability, while larger c values may lead to overfitting. Parameter g controls the influence range of the Gaussian kernel function. Smaller g values make the model smoother, while larger g values make the model complex but may overfit. By setting a wide optimization range, we can ensure that the best parameter combination is found in a wide parameter space, thereby improving the classification performance and generalization ability. Randomly selecting the initial parameter combination helps to avoid local optimal solutions and improve optimization efficiency.
[0101] S402. Use five-fold cross validation to evaluate the model performance with the accuracy of sample testing as the objective function f, and continuously update the model by adding the optimal model parameters that improve the objective function to the initial set.
[0102] S403. Set the iteration termination condition: when the number of iterations reaches the upper limit or the objective function reaches the optimal value, stop the iteration and output the optimal combination parameters (c and g).
[0103] After the Bayesian optimization algorithm, the linear kernel function was finally selected, which is suitable for the classification requirements in the hollow drum detection task. This selection is based on the optimization goal and maximizes the performance indicators of the model on the validation set. The optimal penalty coefficient c = 4.168 and the kernel parameter gamma = 0.247 ensure that the model can effectively distinguish between hollow drum and non-hollow drum audio signals. The simplicity of the linear kernel function combined with the optimized parameter combination improves the generalization ability and stability of the model in practical applications, thereby achieving the highest classification and recognition accuracy.
[0104] The support vector machine (SVM) machine learning method optimized by the Bayesian algorithm has not been widely mentioned in the field of wall hollowing detection, although it has been applied in public patents. In addition to application innovations in this field, improvements include improving parameter optimization efficiency, enhancing model generalization capabilities, improving detection accuracy, and improving model stability. By dynamically adjusting parameter search strategies and selecting the best kernel function type and parameter combination, the optimized support vector machine (SVM) model can more quickly and accurately distinguish between wall hollowing and non-hollowing states, providing a more effective solution for solving complex environments in practical applications.
[0105] S500. Based on the MFCC+Mel-Bo-SVM algorithm model, training is performed to obtain the classification results of wall hollowing and perform evaluation indicators.
[0106] The calculation methods of various indicators are shown in Table 1. Since the sound problem discussed is essentially a classification problem, precision, recall and F1 value are selected as evaluation indicators, among which F1 is obtained by the harmonic mean of precision and recall.
[0107] Table 1 Evaluation indicators
[0108]
[0109]
[0110] For multi-classification problems, precision, recall and F1 value are calculated using macro-average. TP represents the number of wall-knocking audios that are both positive and predicted, i.e., true positives; FP represents the number of wall-knocking audios that are actually negative but predicted as positive, i.e., false positives; and FN represents the number of wall-knocking audios that are actually positive but predicted as negative, i.e., false negatives.
[0111] Example 1 Experiment and Analysis
[0112] The data of this experiment is collected by recording with a microphone, and the wall is knocked using the knocking device shown in the figure. Figure 4 The device is a rectangular block as a whole. The T-shaped rod is connected to it through two bearings. The small ball is connected to the bottom of the rod by bolts to form a complete wall knocking device. Each time the wall is knocked, the device is pressed horizontally on the wall to be knocked, the hammer head is lifted to the highest point, and then released to let it swing freely to knock the wall. The advantage of this is that the force of each knock can be guaranteed to be the same, thereby eliminating the impact of different knocking forces on the data set.
[0113] Among them, the collected hollow drum audios are 262 cement walls, 226 paint walls, 330 marble walls, and 313 tiles. The collected non-hollow drum audios are 334 cement walls, 312 paint walls, 200 marble walls, and 216 tiles, totaling 2193 audio data, as shown in Table 2.
[0114] Table 2 Sample statistics
[0115]
[0116]
[0117] The collected data are subjected to mixed feature extraction of Mel cepstral coefficient feature (MFCC) + Mel spectrum feature (Mel), and converted into grayscale images. Grayscale images can reduce image volume, improve model processing speed, simplify model processing difficulty, and enhance image information. All images are adjusted to 64x64 pixels, and each image is flattened into a one-dimensional array. The data set is divided into a training set and a test set, of which the training set accounts for 80% and the test set accounts for 20%, and the training set and the test set are standardized.
[0118] After preprocessing the data, the training set is used to train the model, and then the test set is classified and identified.
[0119] The linear classification recognition result obtained by Bayesian optimization has the highest accuracy, so linear is selected for support vector machine (SVM) modeling. In the MFCC+Mel-Bo-SVM model parameter optimization, the Bayesian optimization (Bo) algorithm is used to optimize the two parameters of the support vector machine (SVM). This method performs global optimization in the parameter space to avoid falling into the local optimum of the parameters. Bayesian optimization (Bo) optimization can find the best parameter combination in a short time. Compared with the trial and error method, the empirical selection method, the grid search method, etc., the number of iterations is small (saving time);
[0120] Among them, the fitness curve of Bayesian optimization (Bo) is as follows Figure 5 shown.
[0121] like Figure 6 As shown in Figure 1, support vector machine (SVM), random forest, k-nearest neighbor algorithm, grid search optimized support vector machine (SVM), chaotic particle swarm optimized support vector machine (SVM) models are constructed respectively, and the recognition results of various types of hollow drums and non-hollow drums constitute training sets and prediction sets are shown in Figure 1. Figure 6 (a), (b), (c), (d), (e), (f), (g) and (h) in the figure.
[0122] from Figure 6 It can be clearly seen that the algorithm model (MFCC+Mel-Bo-SVM) has the best recognition effect for all samples, and the overall accuracy is shown in Table 3:
[0123] Table 3. Comparison of overall model accuracy
[0124]
[0125]
[0126] The overall accuracy of the algorithm in this paper is improved by 6.61%, 9.58%, 15.27%, 13.9% and 5.02% respectively compared with the support vector machine (SVM), random forest, k-nearest neighbor algorithm, grid search optimized support vector machine (SVM) and chaotic particle swarm optimized support vector machine (SVM) models.
[0127] like Figure 7 As shown in FIG. 1 , it can be further seen from the confusion diagram of the test set samples that the algorithm model (MFCC+Mel-Bo-SVM) has very good classification and recognition performance.
[0128] The contents not described in detail in this specification belong to the prior art known to the professional and technical personnel in this field. In the description of the present invention, unless otherwise specified, "multiple" means two or more; the orientation or position relationship indicated by the terms "upper", "lower", "left", "right", "inner", "outer", "front end", "rear end", "head", "tail" and the like is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", "third" and the like are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0129] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.
Claims
1. A multi-classification detection method for wall hollowing, comprising the following steps: S100. Collect the hollow drum audio and non-hollow drum audio on the wall; S200. Extracting Mel spectrum features and Mel cepstral coefficient features of audio; S300. Fusing the extracted Mel spectrum feature and Mel cepstral coefficient feature to obtain MFCC+Mel fusion feature, where Mel is the Mel spectrum feature and MFCC is the Mel cepstral coefficient feature; S400. Use the Bayesian algorithm to optimize the support vector machine, input the fusion features into the SVM algorithm model after Bayesian optimization, and establish the MFCC+Mel-Bo-SVM algorithm model; S500. Based on the MFCC+Mel-Bo-SVM algorithm model, training is performed to obtain the classification result of wall hollowing.
2. A multi-classification detection method for wall hollowing according to claim 1, characterized in that: The collected audio signal is pre-emphasized to enhance the high-frequency components of the audio signal and improve the signal-to-noise ratio. It is then framed to capture the time-varying characteristics of the audio signal. Finally, windowing preprocessing is performed to reduce edge effects.
3. A multi-classification detection method for wall hollowing according to claim 2, characterized in that: By performing fast Fourier transform and short-time Fourier transform analysis on each frame of the audio signal that has been preprocessed by windowing, a frequency domain signal is obtained to comprehensively analyze the frequency domain characteristics of the signal and extract the frequency components and time resolution.
4. A multi-classification detection method for wall hollowing according to claim 3, characterized in that: After performing fast Fourier transform or short-time Fourier transform on each frame of audio signal, the energy of each frequency component is calculated to obtain the square value of the frequency domain signal, and the calculated energy spectrum is filtered to capture the frequency domain features in the extracted audio signal.
5. A multi-classification detection method for wall hollowing according to claim 4, characterized in that: After the signal is filtered by short-time Fourier transform and energy spectrum calculation, the MFCC feature can be obtained by discrete cosine transform; after the signal is filtered by fast Fourier transform and energy spectrum calculation, the MEL feature can be obtained by taking the logarithm.
6. A multi-classification detection method for wall hollowing according to claim 5, characterized in that: By normalizing the MEL features and MFCC features, and then concatenating them at the frame level, a mixed cepstrum feature is obtained, and the fused signal is visualized to obtain a fusion graph of hollow drum features and non-hollow drum features; The normalization formula is: In formula (1), MFCC norm is the normalized MFCC feature, μMFCC and σMFCC are the mean and standard deviation of the MFCC feature respectively; In formula (2), Mel norm is the normalized MEL feature, μMel and σMel are the mean and standard deviation of the Mel spectrum feature respectively; The frame-level concatenation formula is: Combined Features=[MFCC norm [Y]|Mel norm [Y]](3) In formula (3), Combined Features is the fusion feature MFCC of MFCC and Mel norm [Y] and Mel norm [Y] are the MFCC features and Mel features after normalization of the Yth dimension.
7. A multi-classification detection method for wall hollowing according to claim 1, characterized in that: The Bayesian algorithm is used to optimize the support vector machine parameters to improve the classification accuracy and reliability. The support vector machine maps the data to the high-dimensional space through the kernel function to find the optimal classification plane, and the Bayesian optimization is used to quickly search the best parameter combination globally.
8. A multi-classification detection method for wall hollowing according to claim 7, characterized in that: The process of Bayesian algorithm optimization support vector machine is: S401. The optimization ranges of the regularization parameter c and the radial basis kernel function parameter g are set to 0.1 to 1000 and 0.001 to 1, respectively. First, the initial c and g combination parameters are randomly selected within the optimization range; S402. Use five-fold cross validation to evaluate the model performance with the accuracy of the sample test as the objective function f, and continuously update the model by adding the optimal model parameters that improve the objective function to the initial set; S403. Set the iteration termination condition: when the number of iterations reaches the upper limit or the objective function reaches the optimal value, stop the iteration and output the optimal combination parameters c and g.
Citation Information
Patent Citations
Wall hollowing detection method and device thereof
CN113671031A
Environmental sound classification analysis method and device and medium
CN114882909A
