A Classification Method for Acoustic Signals Based on Multi-resolution Analysis and Feature Separation
By combining wavelet multi-resolution analysis and feature separation index screening with support vector machine classification, the problems of insufficient interpretability and robustness in existing acoustic signal classification methods are solved, and higher accuracy and efficiency of acoustic signal classification are achieved.
Patent Information
- Application Number
- CN202410180892.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-18
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-02-18
AI Technical Summary
Among existing sound signal classification methods, deep learning methods have low interpretability and depend on the quality of training data, while statistical learning methods have poor robustness and are difficult to accurately classify different types of sound signals in noisy environments.
The acoustic signal is decomposed using wavelet multi-resolution analysis to extract multi-scale features. After feature separation index screening and dimensionality reduction, support vector machine is used for classification.
It improves the accuracy and interpretability of acoustic signal classification, reduces the impact of noise on classification, and enhances algorithm efficiency and classification accuracy.
Smart Images

Figure CN118194145B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for acoustic signal processing and classification, specifically to an acoustic signal classification method based on multi-resolution analysis and feature separation. Background Technology
[0002] Sound waves are elastic waves that propagate through a medium from rapidly vibrating sound sources. Due to the differences between the sound source and the medium through which the sound wave travels, the sound wave contains partial information about both the source and the medium. The mechanical energy of the sound wave can be converted into an electrical signal using acoustic sensors via the piezoelectric effect, enabling the measurement of the sound wave. The information contained within the sound signal can then be extracted through processing. Sound waves can be further classified into ultrasound, infrasound, and perceptible sound based on the different vibration speeds of the sound source. Ultrasound has a short wavelength, low diffraction, and good anisotropy, making it suitable for acoustic detection of various structures. Infrasound has a high speed and strong penetrating power, used for early warning of meteorological and geological disasters. However, acoustic sensors can receive a wide variety of sound waves, some containing useful information and others being useless noise. Therefore, to better extract the information contained in the sound signals, it is extremely important to perform more precise classification and analysis of the different types of sound signals received by the sensor.
[0003] Existing methods for classifying acoustic signals are largely based on machine learning, primarily divided into deep learning and statistical learning methods. Deep learning methods train deep networks through error propagation to classify different types of acoustic signals. However, their classification accuracy is highly dependent on the quality and quantity of training data, and their interpretability is weak, failing to adequately analyze and interpret the various acoustic characteristics of different types of acoustic signals. Statistical learning methods classify acoustic signals based on the statistical regularities of different signal features. Their classification accuracy is also affected by data quality, but they require less data and offer better interpretability because they classify signals based on the statistical regularities of signal features. Therefore, the quality of acoustic signal data is the primary factor affecting classification algorithms, and environmental noise is a significant factor influencing signal quality. Thus, measures must be taken to improve the quality of training data to enhance the classification accuracy of different types of acoustic signals. Summary of the Invention
[0004] To address the issues of low interpretability in deep learning methods and poor robustness in statistical learning, this invention provides a sound signal classification method based on multi-resolution analysis and feature separability. This method uses wavelet multi-resolution decomposition to decompose the sound signal, extracts multi-scale features, and selects and classifies features based on their separability at different decomposition scales, thereby improving the classification accuracy of different types of sound signals.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] A method for classifying acoustic signals based on multi-resolution analysis and feature separation includes the following steps:
[0007] Step 1: Load different types of acoustic signals S(t) to be classified, and perform multi-resolution decomposition of the acoustic signals using wavelet multi-resolution analysis to obtain sub-signals of the acoustic signals at different scales;
[0008] Step 2: Extract features from each sub-signal obtained by multi-resolution decomposition, extract 28-dimensional features, calculate the feature separability index ST of each sub-signal, and filter the features of each sub-signal based on the feature separability index ST to obtain the best feature set that best reflects the features of different acoustic signals.
[0009] Step 3: To reduce data redundancy and improve algorithm efficiency, the feature separability index ST is used to reduce the dimensionality of the optimal feature set, and a support vector machine is used to classify the dimensionality-reduced feature set to obtain accurate classifications of different types of acoustic signals.
[0010] Compared with the prior art, the present invention has the following advantages:
[0011] 1. This invention uses wavelet multi-resolution analysis to decompose acoustic signals containing random noise into sub-signals at different scales. The feature separability index constructed in this invention is used to quantitatively evaluate the features of each sub-signal, and then compares, analyzes and filters them. By analyzing and concentrating information at each scale of the acoustic signal, adaptive information enhancement can be performed on acoustic signals containing unknown noise, thereby improving the classification accuracy of subsequent classification algorithms.
[0012] 2. This invention studies and analyzes 28-dimensional features of acoustic signals, including time, frequency, and fractal domains. It allows for multi-faceted analysis of the acoustic characteristics of different types of acoustic signals and the aggregation of information from various scale components of the acoustic signal. Specifically, the feature separability index constructed in this invention reduces the dimensionality of the 28-dimensional features, thereby reducing the computational overhead associated with them.
[0013] 3. Most existing acoustic signal classification algorithms are based on deep learning, which has weak interpretability. They cannot obtain accurate classification models for acoustic signals that are difficult to collect. Furthermore, classification methods based on statistical learning cannot accurately classify acoustic signals contaminated by unknown sudden noise. The acoustic signal classification method proposed in this invention, based on multi-resolution analysis and feature separation, adaptively reduces the influence of noise and improves classification accuracy by decomposing the signal at multiple scales and using feature separation for feature selection. Attached Figure Description
[0014] Figure 1 This is a flowchart of the acoustic signal classification method based on multi-resolution analysis and feature separation of the present invention.
[0015] Figure 2 This is a schematic diagram of different types of sound signals.
[0016] Figure 3 This is a schematic diagram of the characteristic separability index of different types of acoustic signal data S(t).
[0017] Figure 4 This is a schematic diagram showing different separability characteristics. Detailed Implementation
[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.
[0019] This invention provides a sound signal classification method based on multi-resolution analysis and feature separability. First, wavelet multi-resolution analysis is used to decompose the sound signal to be classified into sub-signals at different scales. Then, to better identify and analyze the acoustic characteristics of different types of sound signals, features are extracted from each sub-signal obtained from the multi-resolution decomposition. A total of 28 features are extracted, including 22-dimensional time-domain features, 6-dimensional frequency-domain features, and 2 fractal-domain features. The separability index value of each sub-signal feature is calculated, and features are filtered based on the separability index to obtain the optimal feature set that best reflects the characteristics of the sound signal. Finally, to reduce data redundancy and improve algorithm efficiency, the optimal feature set is dimensionality-reduced using the separability index, and a support vector machine is used to classify the dimensionality-reduced feature set, resulting in accurate classifications of different types of sound signals. Figure 1 As shown, the specific steps are as follows:
[0020] Step 1: Load different types of acoustic signals S(t) to be classified, and perform multi-resolution decomposition of the acoustic signals using wavelet multi-resolution analysis to obtain sub-signals of the acoustic signals at different scales. The specific steps are as follows:
[0021] Step 11: Perform convolution operations on S(t) using the parent wavelet function and the mother wavelet function to obtain approximation coefficients and detail coefficients at different scales:
[0022]
[0023]
[0024] in, and It is the conjugate of the mother wavelet function and the father wavelet function, j is the level of the wavelet function, represents the decomposition scale, and k represents the time shift position of the wavelet function; Aj,k The low-frequency component corresponding to time k in the j-level decomposition of the signal reflects the overall trend of the signal; D j,k These are the detail coefficients at time k of the signal decomposition at level j, representing the high-frequency components of the signal;
[0025] Steps 1 and 2: Reconstruct the approximate signal AS for each decomposition scale using wavelet functions. j (t) and detail signal DS j (t):
[0026]
[0027]
[0028] As parameter j changes, AS j (t) and DS j (t) represents the sub-signals at different scales of the signal, which are decomposed into multiple sub-signals using multi-resolution analysis.
[0029] Step 2: Extract features from each sub-signal obtained by multi-resolution decomposition, extract 28-dimensional features, calculate the feature separability index ST for each sub-signal, and filter the features of each sub-signal based on the feature separability index ST to obtain the best feature set that best reflects the characteristics of different acoustic signals. The specific steps are as follows:
[0030] Step 21: To better identify different types of acoustic signals and analyze their acoustic characteristics, 28-dimensional features are extracted from sub-signals at each scale, including 15-dimensional time-domain features: root mean square (F1), waveform coefficients (F2), amplitude factor (F3), maximum value (F4), minimum value (F5), peak-to-peak value (F6), standard deviation (F7), average value (F8), kurtosis (F9), and average rectified value (F1). 10 Impact factor F 11 Marginal coefficient (F) 12 Energy (F) 13 ), skewness (F) 14 ), square root amplitude (F) 15 ); 5-dimensional frequency domain characteristics: center frequency (F 16 ), mean square frequency (F 17 ), root mean square frequency (F 18 ), frequency variance (F 19 ), frequency standard deviation (F 20 ); 6 entropy characteristics: approximate entropy (F 21 ), fuzzy entropy (F 22 ), Sample entropy (F) 23 ), permutation entropy (F 24 ), power spectral entropy (F 25 ), singular spectral entropy (F26 Two fractal features: box counting dimension (F) 27 Information dimension (F) 28 ).
[0031] Step 22: Calculate the feature separability index ST. The specific steps are as follows:
[0032] Step 221: Calculate the inter-class separability index ST b The inter-class separability index ST can be described by the sum of the distances between the centers of the eigenvalues of each class of sound signals. The sum of the absolute values of these center distances represents the inter-class separability index ST. b The following results were obtained:
[0033]
[0034] Where, N c TC represents the number of types of acoustic signals. i This represents the characteristic average value of the i-th sound signal.
[0035] Step 222: Represent the intra-class separability index ST of the signal by the standard deviation of each class of eigenvalues. w The intraclass separability index ST is obtained by summing these standard deviations. w As shown below:
[0036]
[0037] in, N represents the j-th eigenvalue of the i-th type of acoustic signal. i This indicates the number of data points for a particular type of acoustic signal.
[0038] Steps 2-3: Larger ST b and larger ST w This means that the feature can effectively represent the differences between different types of signals. Because the ratio form better incorporates the ST of the feature... b and ST w Therefore, ST is used. b and ST w The ratio of the features is used as a measure of overall feature separability. This index is called the feature separability index ST, as shown below:
[0039]
[0040] Steps 2 and 3: Feature selection is performed based on the feature separability index (ST) to obtain the optimal feature set. The specific steps are as follows:
[0041] Step 231: Construct vectors for the feature separability index ST of each sub-signal, and feature separability vector l of each sub-signal. sThe structure is as follows:
[0042]
[0043] Among them ST i s Let be the separability index of the i-th feature of the sub-signal s, and n be the wavelet decomposition level. n-level wavelet decomposition yields 2n sub-signals and 2n 28×1 feature separability vectors.
[0044] Step 232: For each of the 28-dimensional features, compare the ST values in the 2n sub-signals, and select the sub-signal features with the highest separability to form the optimal feature set.
[0045] Step 3: To reduce data redundancy and improve algorithm efficiency, the feature separability index ST is used to reduce the dimensionality of the optimal feature set, and a support vector machine is used to classify the dimensionality-reduced feature set to obtain accurate classifications of different types of acoustic signals. The specific steps are as follows:
[0046] Step 31: The feature separability index ST is calculated by combining the intra-class and inter-class distances of different types of acoustic signal features. It characterizes the impact and contribution of different features on classification accuracy. By selecting features with high ST values, redundancy is reduced and efficiency is improved while achieving better classification results. i Let i represent the feature vector of the i-th data point.
[0047] Step 32: Input the dimensionality-reduced feature set into a support vector machine for classification. Observe the classification accuracy under different dimensions to obtain the minimum feature dimension with the desired accuracy. Use this as the criterion for feature selection set dimension to complete the dimensionality reduction of the signal and input it into the support vector machine for classification to obtain the class label y.
[0048] y=ω T l
[0049] Where l is the feature vector of the input signal, ω T This is the transpose of the support vector machine parameters;
[0050] Step 33: The support vector machine parameters are trained as follows:
[0051]
[0052] sty i *(ω T l i )≥1, i=1,2,...,N
[0053] Where N is the number of training data points, y i Let i be the category label for the i-th data point.
[0054] Steps 3 and 4: When performing multi-signal classification tasks, the labels of various signals are continuously adjusted, and pairwise training is conducted to complete the multi-classification task.
[0055] Example:
[0056] Step 1: Load different types of acoustic signal data S(t), a total of 5711 acoustic signal data points were loaded, each data point is 4096 in length, the sampling rate is 5MHz, and five different acoustic signals are as follows: Figure 2 As shown. First, a suitable wavelet basis is selected. The db5 wavelet in the Daubechies wavelet sequence has excellent performance in terms of orthogonality and time-frequency support, and is therefore selected for use in this invention. Second, the db5 wavelet basis is used to perform three-level wavelet decomposition on different types of acoustic signal data S(t) to obtain the approximation coefficients and detail coefficients of S(t) at three scales. Then, the sub-signals of S(t) at each scale are reconstructed through the approximation coefficients and detail coefficients of each level of decomposition.
[0057] Step 2: Extract 28-dimensional features from the sub-signals at each scale obtained in Step 2, and calculate the separability index of each feature in all sub-signals. The ST values of different features in S(t) are shown below. Figure 3 As shown, different features exhibit varying degrees of separability, resulting in different impacts on classification. The energy feature has the highest ST value of 2958, the sample entropy has an ST value of 316, the waveform factor has an ST value of 148, and the root mean square ST value is the lowest at 3. The four features are as follows: Figure 4 As shown, the five acoustic signals are represented by Roman numerals I, II, III, IV, and V, respectively. As the ST value decreases, it becomes more difficult to distinguish different types of acoustic signals using this feature, demonstrating that the ST exponent can measure the impact of features on classification accuracy. The optimal set of 28 features is obtained by filtering the 28-dimensional features of sub-signals at each scale based on the ST value.
[0058] Step 3: Perform ST exponent dimensionality reduction and classification on the selected optimal feature set. To verify the effectiveness of ST value dimensionality reduction, Principal Component Analysis (PCA) was used to reduce the dimensionality of the feature set, and the results were compared with the classification accuracy of the Support Vector Machine (SVM) algorithm. The improved classification accuracy of this invention is verified by comparing the results with those of the SVM algorithm. The results are shown in Table 1. This invention can achieve a classification accuracy of up to 100%, far exceeding the accuracy of SVM. Furthermore, by calculating the feature separability index (ST) of each sub-signal in the multi-resolution decomposition, the information concentration of the signal at different scales and the performance of different acoustic features at different scales can be obtained, demonstrating interpretability and signal analysis capabilities far superior to deep learning.
[0059] Table 1. Comparison of classification accuracy between this algorithm and other algorithms under different feature dimensions.
[0060]
Claims
1. A method for classifying acoustic signals based on multi-resolution analysis and feature separation, characterized in that, The method includes the following steps: Step 1: Load different types of sound signals to be classified The acoustic signal is decomposed into sub-signals at different scales by using wavelet multi-resolution analysis. Step 2: Extract features from each sub-signal obtained by multi-resolution decomposition, extract 28-dimensional features, and calculate the feature separability index of each sub-signal. Based on the feature separability index Feature filtering is performed on each sub-signal to obtain the optimal feature set that best reflects the characteristics of different acoustic signals. The specific steps are as follows: Step 21: Extract 28-dimensional features from sub-signals at each scale, including 15-dimensional time-domain features: root mean square. waveform coefficients Amplitude factor maximum value minimum value peak-to-peak value Standard deviation ,average value kurtosis Average rectified value Impact factor Marginal coefficient ,energy skewness Square root amplitude 5-dimensional frequency domain characteristics: center frequency Mean square frequency Root mean square frequency Frequency variance Frequency standard deviation 6 entropies Feature: Approximate entropy Fuzzy entropy Sample entropy Permutation entropy Power spectral entropy Singular spectral entropy Two fractal features: box counting dimension Information dimensionality ; Step 22: Calculate the feature separability index The specific steps are as follows: Step 221: Calculate the inter-class separability index : in, The number of types of sound signals. Representing the The characteristic average value of acoustic signals; Step 222: Represent the intra-class separability index of the signal by the standard deviation of each class of eigenvalues. : in, Indicates the first The first type of sound signal 1 eigenvalue, The number of data points representing a certain type of acoustic signal; Steps two through three: Use and The ratio of features to total features is used as a measure of overall feature separability; this index is called the feature separability index. : ; Steps two and three: Based on the feature separability index Feature selection is performed to obtain the optimal feature set; Step 3: Use the feature separability index The optimal feature set is dimensionality reduced, and the reduced feature set is classified using a support vector machine to obtain accurate classifications of different types of acoustic signals.
2. The acoustic signal classification method based on multi-resolution analysis and feature separation according to claim 1, characterized in that, The specific steps of step one are as follows: Step 11: Using the parent wavelet function and the mother wavelet function to... Perform convolution operations to obtain approximation coefficients and detail coefficients at different scales: in, and It is the conjugate of the mother wavelet function and the father wavelet function. is the layer number of the wavelet function, and represents the decomposition scale. This indicates the time shift position of the wavelet function; Corresponding to signal Layer decomposition The low-frequency components at any given time reflect the overall trend of the signal; It is a signal Layer decomposition The time detail factor represents the high-frequency part of the signal; Steps 1 and 2: Reconstruct the approximate signal for each decomposition scale using wavelet functions. and detailed signals : As parameter j changes and The sub-signals representing the signal at different scales are decomposed into multiple sub-signals using multi-resolution analysis.
3. The acoustic signal classification method based on multi-resolution analysis and feature separation according to claim 1, characterized in that... The specific steps of steps two and three are as follows: Step 231: Calculate the feature separability index of each sub-signal. Construct vectors, feature separability vectors of each sub-signal The structure is as follows: , ,..., , in Sub-signal The The separability index of each feature, The wavelet decomposition level is denoted as . Layer wavelet decomposition Each sub-signal will receive indivual dimensional feature separability vector; Steps 2 and 3: Compare the 28 features in each dimension. In each sub-signal The values are then used to select the sub-signal features with the highest separability, forming the optimal feature set.
4. The acoustic signal classification method based on multi-resolution analysis and feature separation according to claim 1, characterized in that... The specific steps of step three are as follows: Step 31: Feature Separability Index By combining intra-class and inter-class distances of different types of acoustic signal features, the influence and contribution of different features on classification accuracy are characterized. This is achieved through the calculation of high-resolution acoustic signal features. The characteristics of the value are selected, and by Indicates the first Feature vectors of data points; Step 32: Input the dimensionality-reduced feature set into a support vector machine for classification. Observe the classification accuracy under different dimensions to obtain the minimum feature dimension with the desired accuracy. Use this as the criterion for feature selection set dimension to complete the signal dimensionality reduction and input it into the support vector machine for classification to obtain the class label. : in The input signal feature vector, This is the transpose of the support vector machine parameters; Step 33: The support vector machine parameters are trained as follows: in The number of training data points, For the first Category labels for each data point; Steps 3 and 4: When performing multi-signal classification tasks, the labels of various signals are continuously adjusted, and pairwise training is conducted to complete the multi-classification task.
Citation Information
Patent Citations
Multi-resolution singular-spectrum entropy and SVM based leakage acoustic emission signal identification method
CN105488520A
Acoustic signal mode classification method based on eigenvalue clustering
CN116230009A
