A pulse sound recognition method based on Hilbert-huang transform and Mel spectrum transform

Abstract

The application discloses a pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform, and comprises the following steps: training an MFCC model and a marginal spectrum model based on training pulse signals to generate MFCC coefficients and marginal spectrum coefficients, so as to construct a trained Gaussian mixture model; for a given pulse signal, performing Mel spectrum transform processing and HHT transform processing based on the pulse signal respectively to generate MFCC coefficients and marginal spectrum coefficients of the pulse signal; calculating Gaussian probability density functions of the MFCC coefficients and the marginal spectrum coefficients to generate a Gaussian mixture model of a pulse to be recognized; and performing matching based on the Gaussian mixture model of the pulse to be recognized and the trained Gaussian mixture model to determine a kind of pulse signal corresponding to a maximum probability. The application adopts Hilbert-Huang transform (HHT) to extract marginal spectrum features of a pulse signal, so that local time domain features of a nonlinear non-stationary sound signal can be effectively recognized, and recognition precision is improved.

Description

Technical Field

[0001] This application relates to the field of audio detection technology, and in particular to a pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform. Background Technology

[0002] Impulse sound refers to a sound signal that changes abruptly and lacks continuity within a short period. Common impulse sound signals include gunshots, cannon fire, thunder, and the sound of a door slamming. Traditional impulse sound recognition methods include theoretical approaches such as wavelet transform, autocorrelation, zero-crossing rate analysis, spectral analysis, and linear predictive analysis. Generally, traditional impulse sound recognition methods can achieve good recognition rates and effectively distinguish between target and interference signals. However, when the interference signal and the target signal have the same frequency range and their spectra overlap, traditional methods struggle to separate the interference signal from the target signal, resulting in a high false alarm rate and unsatisfactory recognition results. On the other hand, in recent years, new sound recognition methods have been proposed. A widely used method involves extracting the Mel-spectral coefficients of the sound signal and applying a Hidden Markov Model (HMM) to identify the target impulse sound signal. This method has been widely applied in speech recognition and pattern recognition. However, this method also has limitations; it cannot well express the local time-domain characteristics of nonlinear, non-stationary sound signals and cannot further accurately identify the details of time-frequency features.

[0003] Traditional recognition methods extract the Mel-spectral coefficients of sound signals and apply Hidden Markov Models (HMMs) to identify target impulse sound signals. While this method has been widely used in speech and pattern recognition, Mel-spectral coefficients cannot adequately represent the local time-domain characteristics of nonlinear, non-stationary sound signals, and cannot accurately identify the details of time-frequency features. Furthermore, the probabilistic recognition method using HMMs has high computational complexity in data training and probability calculation; different input parameters can cause fluctuations in the algorithm's recognition accuracy, resulting in a low recognition rate when dealing with non-stationary signals. Summary of the Invention

[0004] This application provides a pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform. The Hilbert-Huang transform (HHT) is used to extract the marginal spectral features of the pulse signal, so as to effectively identify the local time domain features of nonlinear and non-stationary sound signals and improve the recognition accuracy.

[0005] This application provides an embodiment of a pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform, including:

[0006] The MFCC model and marginal spectrum model are pre-trained based on training pulse signals to generate MFCC coefficients and marginal spectrum coefficients, so as to construct a well-trained Gaussian mixture model.

[0007] For a given pulse signal, Mel spectrum transformation and HHT transformation are performed on the pulse signal respectively to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal.

[0008] Calculate the Gaussian probability density function of the MFCC coefficients and marginal spectral coefficients to generate a Gaussian mixture model of the pulse to be identified;

[0009] Based on the Gaussian mixture model of the pulse to be identified and the trained Gaussian mixture model, a matching is performed to determine the type of pulse signal corresponding to the highest probability, which is the pulse type of the given pulse signal.

[0010] Optionally, the MFCC model and marginal spectral model are pre-trained based on training impulse signals to generate MFCC coefficients and marginal spectral coefficients, in order to construct a trained Gaussian mixture model, including:

[0011] A training sequence is constructed in advance by generating MFCC coefficients and marginal spectral coefficients based on the training pulse signal;

[0012] Based on the training sequence, the GMM likelihood is expressed as:

[0013]

[0014] Where λ represents the estimated parameter, X i Represents the training sequence;

[0015] The training process satisfies:

[0016] A new parameter is estimated using the EM algorithm. This improves the likelihood under the new model parameters.

[0017] Iterate until the model converges.

[0018] Optionally, performing Mel-ray spectral transform and HHT-ray transform on the pulse signal to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal includes:

[0019] After preprocessing the pulse signal, an FFT transformation is performed;

[0020] Calculate the spectral line energy for each frame of FFT-processed data;

[0021] For each frame of spectral energy, multiply it by the frequency domain response of the Merr filter and add them together to determine the energy passing through the Merr filter;

[0022] Based on the energy passing through the Merr filter, the DCT is calculated to determine the MFCC coefficients and the first difference spectral coefficients.

[0023] Optionally, performing Mel-ray spectral transform and HHT-ray transform on the pulse signal to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal includes:

[0024] The pulse signal is subjected to EMD filtering to obtain multiple IMF components;

[0025] Perform Hilbert transform on each IMF component;

[0026] Based on the results of the Hilbert transform and the Hilbert spectrum, the Hilbert marginal spectrum and instantaneous energy density level are determined, satisfying:

[0027]

[0028] Vh(ω)=h(ω t+1 )-h(ω t-1 )

[0029] Where h(ω) represents the marginal spectrum of the signal, Vh(ω) represents the first-order difference coefficient of the marginal spectrum, and H(ω,t) represents the Hilbert spectrum;

[0030] Marginal spectral coefficients are determined based on the marginal spectrum of the signal.

[0031] Optionally, calculating the Gaussian probability density function of the MFCC coefficients and marginal spectral coefficients to generate a Gaussian mixture model of the pulse to be identified includes:

[0032] The probability density function of an M-order Gaussian mixture model is defined to satisfy:

[0033]

[0034] Where X is a D-dimensional random vector; b i (X i ), i = 1, L, M are sub-distributions; ω i It is a mixed weight;

[0035] The MFCC coefficients, first difference spectral coefficients, marginal spectral coefficients, and the first difference coefficients of the marginal spectrum are used as sub-distribution b. i (X i ), to generate a Gaussian mixture model of the pulse to be identified.

[0036] Optionally, based on the Gaussian mixture model of the pulse to be identified and the trained Gaussian mixture model, a matching process is performed to determine the type of pulse signal corresponding to the highest probability, including:

[0037] Based on the maximum a posteriori probability of Bayesian theory, the pulse signal is identified as belonging to a category of the training data, satisfying the following conditions:

[0038] i = argmaxP(X / λ) i )

[0039] Where i represents the type of pulse signal identified, P(X / λ) i ) represents the maximum posterior probability.

[0040] This application also proposes a pulse sound recognition device based on Hilbert-Huang transform and Mel spectrum transform, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described above.

[0041] This application also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the impulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described above.

[0042] This application provides a pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform. The Hilbert-Huang transform (HHT) is used to extract the marginal spectral features of the pulse signal, so as to effectively identify the local time domain features of nonlinear and non-stationary sound signals and improve the recognition accuracy.

[0043] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0044] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0045] Figure 1 This is an example of the overall process of the pulse sound recognition method according to an embodiment of this application;

[0046] Figure 2 This is an example of the MFCC coefficient extraction process for the impulse sound recognition method in this application embodiment;

[0047] Figure 3This is an example of a marginal spectrum of the impulse sound recognition method according to an embodiment of this application. Detailed Implementation

[0048] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0049] This application provides a pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform, such as... Figure 1 As shown, the pulse signal recognition scheme in this application embodiment is divided into two main parts: a data training part and a model matching part.

[0050] First, a standard impulse signal is trained. Based on Mel spectral theory, MFCC coefficients and their first-order spectral difference coefficients are generated. Then, based on Hilbert-Huang theory, marginal spectral coefficients and their first-order spectral difference coefficients are generated. The two types of coefficients are combined and the Gaussian probability density function of each coefficient is calculated to generate a trained Gaussian mixture model.

[0051] Secondly, model matching is performed on the unknown pulse signal. After signal preprocessing (windowing and framing), the unknown pulse signal undergoes two types of spectral transformations to generate MFCC coefficients, marginal spectral coefficients, and corresponding first-order difference spectral coefficients; and the Gaussian probability density function of each coefficient is calculated to generate a Gaussian mixture model of the pulse to be identified.

[0052] Finally, the trained model is matched with the model to be identified, the maximum likelihood probability of the match is calculated, and the type of pulse signal corresponding to the maximum probability is found. This type is the pulse type of the pulse signal to be identified, and the identification is complete. The specific steps include the following:

[0053] The MFCC model and marginal spectrum model are trained in advance by generating MFCC coefficients and marginal spectrum coefficients based on the training pulse signal, so as to construct a trained Gaussian mixture model.

[0054] This application describes the data training of Gaussian Mixture Models (GMMs). Training a GMM involves determining the model's parameters λ based on a given set of training data and a certain criterion. The most commonly used parameter estimation method is Maximum Likelihood (ML) estimation. In some embodiments, the MFCC model and marginal spectral model are pre-trained based on training impulse signals to generate MFCC coefficients and marginal spectral coefficients, thereby constructing a trained Gaussian Mixture Model.

[0055] A training sequence is constructed in advance by generating MFCC coefficients and marginal spectral coefficients based on the training pulse signal;

[0056] Based on the training sequence, the GMM likelihood is expressed as:

[0057]

[0058] Where λ represents the estimated parameter, X i Represents the training sequence;

[0059] The training process satisfies:

[0060] A new parameter is estimated using the EM (Expectation Maximization) algorithm. This improves the likelihood under the new model parameters. Iterate until the model converges.

[0061] The new model parameters are then used as the current parameters for training, and this iterative process continues until the model converges. In each iteration, the following re-estimation formula is used to ensure that the model likelihood monotonically increases:

[0062] The revaluation of mixed weights satisfies:

[0063]

[0064] The revaluation of the mean satisfies:

[0065]

[0066] The revaluation of variance satisfies:

[0067]

[0068] Wherein, the posterior probability of component i is:

[0069]

[0070] The steps of the EM (Expectation Maximization) parameter estimation algorithm are as follows:

[0071] Input: Observed variable X = {X j}, j=1,2,L,N,error ε,iteration number M.

[0072] Output: θ={(ω i ,μ i ,Σ i ),i=1,L,Q}.

[0073] 1) Initialize parameters, Let m = 1;

[0074] 2) Iteration:

[0075] Step E – Calculate the observed value X using the concept of statistical averaging. j The probability of coming from the i-th model:

[0076]

[0077] M-step -- Using the calculated probability estimates, calculate the model parameters obtained from the maximum likelihood:

[0078]

[0079]

[0080]

[0081] If m ≥ M, then skip to step 3); otherwise, calculate the parameter iteration error, if ||θ m -θ m-1 If ||≤ε, then jump to step 3); otherwise, m=m+1, return to step 1 in step 2) and continue iterating.

[0082] 3) Output the final output This is the estimated result, and the EM parameter estimation algorithm ends.

[0083] For a given pulse signal, Mel spectrum transformation and HHT transformation are performed on the pulse signal respectively to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal.

[0084] In some embodiments, performing Mel spectral transform and HHT transform processing on the pulse signal to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal includes:

[0085] After preprocessing the pulse signal, an FFT transform is performed. In some specific examples, preprocessing includes pre-emphasis, framing, and windowing. Pre-emphasis is used to compensate for the loss of high-frequency components and enhance them; windowing can be achieved using a Hamming window function.

[0086] The Fast Fourier Transform (FFT) transforms each frame of the signal, converting it from time-domain data to frequency-domain data.

[0087] X(i,k)=FFT[x i (m)]

[0088] Calculate the spectral line energy for each frame of FFT-processed data;

[0089] E(i,k)=[x i (k)] 2

[0090] For each frame of spectral energy, multiply it by the frequency domain response of the Merr filter and add them together to determine the energy passing through the Merr filter;

[0091]

[0092] Based on the energy passing through the Merr filter, the DCT is calculated to determine the MFCC coefficients and the first difference spectral coefficients. An exemplary procedure is as follows:

[0093] Calculate the DCT cepstral coefficients and the FFT cepstral of the sequence x(n). for:

[0094]

[0095] In the formula FT and FT -1 Let represent the Fourier transform and inverse Fourier transform. The DCT cepstral of sequence x(n) is:

[0096]

[0097] In the formula, parameter N is the length of sequence x(n); C(k) is the orthogonality factor, which can be expressed as:

[0098]

[0099] After taking the logarithm of the Merr filter energy, the DCT is calculated as follows:

[0100]

[0101] Δmfcc(i,n)=mfcc(i+1,n)-mfcc(i-1,n)

[0102] In the formula, S(i,m) represents the energy of the Merr filter; m represents the m-th Merr filter (there are M filters in total); i represents the i-th frame of data; n represents the spectral line after DCT; Δmfcc(i,n) is the first-order difference coefficient.

[0103] The above method generates MFCC coefficients and difference spectrum coefficients, which serve as characteristic coefficients of the Mel spectrum of the acoustic target and as features for identifying unknown acoustic targets in the acoustic target model.

[0104] The Hilbert-Huang Transform (HHT) is an effective method for analyzing and processing nonlinear and non-stationary signals. Through the definitions of Empirical Mode Decomposition (EMD) and Intrinsic Mode Functions (IMFs), HHT adaptively decomposes a complex signal into a finite set of IMFs whose instantaneous frequencies have physical meaning, can be amplitude or frequency modulated, and are distributed from high to low frequencies. The Hilbert transform results of these IMFs constitute a three-dimensional time-frequency-energy distribution spectrum revealing the time-varying characteristics of nonlinear and non-stationary signals. HHT requires no prior knowledge; the decomposition is adaptively dependent on the signal itself, and the decomposition has real physical meaning. HHT exhibits superior characteristics in analyzing and processing nonlinear and non-stationary signals. In some embodiments, performing Mel-spectrum transform and HHT transform processing on the pulse signal to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal includes:

[0105] The pulse signal is subjected to EMD filtering to obtain multiple IMF components. EMD (empirical mode decomposition) is a signal stabilization process that decomposes the fluctuations or trends at different scales in the signal step by step, producing a series of data sequences with different characteristic scales, each of which is an IMF component.

[0106] Empirical Mode Decomposition

[0107] For any time series X(t), its Hilbert transform can be obtained by the following definition:

[0108]

[0109] Based on the above equation, the analytic signal Z(t) is defined as follows:

[0110] Z(t)=X(t)+i·Y(t)=a(t)·exp(i·θ(t))

[0111] in,

[0112] The formula for calculating instantaneous frequency is:

[0113]

[0114] Using empirical mode decomposition, the instantaneous frequency is decomposed into a single component IMF that has the actual physical meaning.

[0115] The IMF components ensure the effectiveness of Hilbert analysis. The EMD method decomposes nonlinear, non-stationary signals into a set of IMFs. The EMD method determines the intrinsic modes of a signal based on its characteristic time scale and then decomposes them sequentially. It uses the time interval between consecutive extrema as the time scale definition for the intrinsic modes of the signal because it not only provides high time-frequency resolution but is also applicable to signals without zero-crossing points. The method for screening the intrinsic modes of a signal is given below:

[0116] First, the filtering process is performed using the envelopes formed by the local maxima and local minima of the signal, respectively. After determining all local extrema, the upper envelope is formed by connecting the cubic spline curves of the maxima, and similarly, the lower envelope is formed by connecting the cubic spline curves of the minima. In this way, all data points of the signal are located within the regions enclosed by the upper and lower envelopes.

[0117] Let m1(t) be the mean of the upper and lower envelopes of the original pulse signal s(t). Then the difference between s(t) and m1(t) is the first component, denoted as h1(t):

[0118] s1(t)-m1(t)=h1(t)

[0119] In the second screening, treating h1(t) as the original pulse signal, and using the same method, we can obtain:

[0120] h1(t)-m 1,1 (t)=h 1,1 (t)

[0121] Then repeat the screening process k times until h is reached. 1,k (t) satisfies the IMF conditions and is the first IMF component. This process is represented as follows:

[0122]

[0123] IMF 1,t (t)=h 1,k (t), so that imf 1,t (t) is the first IMF component selected from the original pulse signal s(t). In this embodiment, the selection at this level is referred to as inner layer selection.

[0124] The inner-layer filtering process relies solely on the feature time scale to first decompose the finest-scale local modes from the signal. This is achieved by using the following formula to... 1,t Separate s(t) from the other components of s(t):

[0125] s1(t)-imf1(t)=r1(t)

[0126] r1(t) includes s(t) except imf 1,t The remaining components of r1(t) are then treated as new signals to be decomposed, and the same inner-layer filtering process is applied to r1(t). Repeating the above steps, we obtain:

[0127]

[0128] The EMD method decomposes nonlinear, non-stationary signals into a set of IMF decomposition methods, and the decomposition results are as follows:

[0129]

[0130] For each IMF component, a Hilbert transform is performed. That is, after decomposition, non-IMF components are discarded, thus focusing more on low-energy, high-frequency components. The result of each Hilbert transform is expressed as:

[0131]

[0132] in,

[0133]

[0134]

[0135] H represents the Hilbert transform operator. The above equation clearly separates amplitude modulation and frequency modulation, breaking the limitations of constant amplitude and frequency in the Fourier transform. This allows HHT to be successfully applied to the processing and analysis of nonlinear and non-stationary signals.

[0136] The Hilbert spectrum is defined as follows:

[0137]

[0138] Based on the results of the Hilbert transform and the Hilbert spectrum, the Hilbert marginal spectrum and the instantaneous energy density level are determined. In this embodiment, energy density is represented by the square of the amplitude. Squaring the amplitude in the Hilbert spectrum yields the Hilbert energy spectrum. The Hilbert marginal spectrum h(ω) (the Hilbert Marginal Spectrum, or HMS for short) and the instantaneous energy density level IE(t) are defined as follows:

[0139] h(ω)=∫0 T H(ω,t)dt

[0140] Vh(ω)=h(ω t+1 )-h(ω t-1 )

[0141] Where T is the signal sampling time, h(ω) represents the marginal spectrum of the signal, Vh(ω) represents the first-order difference coefficient of the marginal spectrum, and H(ω,t) represents the Hilbert spectrum. h(ω) reflects the distribution of amplitude values ​​at each frequency point, representing the cumulative amplitude along the entire data span in a probabilistic sense, that is, the amplitude contribution of each frequency to the whole. Figure 3 An example of a marginal spectrum is shown. Finally, the marginal spectrum coefficients are determined based on the signal marginal spectrum. These coefficients are used as a feature coefficient of the marginal spectrum in the acoustic target, and in the acoustic target model, they serve as a feature for identifying unknown acoustic targets.

[0142] Calculate the Gaussian probability density function of the MFCC coefficients and marginal spectral coefficients to generate a Gaussian mixture model of the pulse to be identified.

[0143] Based on the Gaussian mixture model of the pulse to be identified and the trained Gaussian mixture model, a matching is performed to determine the type of pulse signal corresponding to the highest probability, which is the pulse type of the given pulse signal.

[0144] This application addresses the complexity issue of existing Hidden Markov Model (HMM) algorithms by employing a Gaussian Mixture Model (GMM). A GMM can be viewed as a continuously distributed HMM with one state. The GMM has lower algorithmic complexity than HMMs. Theoretically, a GMM can decompose any acoustic target signal. Using the Hilbert-Huang Transform (HHT), marginal spectral features of the pulse signal are extracted. In this embodiment, first-order spectral features of the acoustic target are added. These first-order spectral features reflect the changing trends of nonlinear, non-stationary signals. By introducing higher-order features, the accuracy of acoustic target recognition is improved.

[0145] A Gaussian Mixture Model (GMM) is a model with only one state, in which multiple Gaussian distribution functions exist. A GMM can be viewed as a continuously distributed Hidden Markov Model with one state. In some embodiments, calculating the Gaussian probability density functions of the MFCC coefficients and marginal spectral coefficients to generate a GMM for the pulse to be identified includes:

[0146] The probability density function of an M-order Gaussian mixture model is defined to satisfy:

[0147]

[0148] Where X is a D-dimensional random vector; b i (X i), i = 1, L, M are sub-distributions; ω i It uses mixed weights. Each sub-distribution is a D-dimensional joint Gaussian probability distribution, which can be represented as:

[0149]

[0150] In the formula, μ i It is the mean vector; Σ i It is a covariance matrix, and the mixed weight values ​​satisfy the following conditions:

[0151]

[0152] A complete Gaussian mixture model consists of a parameter mean vector, a covariance matrix, and mixture weights, and is represented as:

[0153] λ={ω i ,μ i ,Σ i}, i=1,L,M

[0154] For a given time series X = {X t The log-likelihood obtained using the GMM model for t = 1, 2, ..., T can be defined as:

[0155]

[0156] The MFCC coefficients, first difference spectral coefficients, marginal spectral coefficients, and the first difference coefficients of the marginal spectrum are used as sub-distribution b. i (X i ), to generate a Gaussian mixture model of the pulse to be identified.

[0157] In some embodiments, matching is performed based on the Gaussian mixture model of the pulse to be identified and a trained Gaussian mixture model to determine the type of pulse signal corresponding to the highest probability, including:

[0158] Maximum a posteriori probability based on Bayesian theory:

[0159]

[0160]

[0161] Since P(λ) i The prior probability of the unknown acoustic target is unknown. Assuming that the probability of the unknown acoustic target belonging to each class in the training set is equal, then identifying the class of the pulse signal belonging to the training data satisfies the following:

[0162] i = argmaxP(X / λ) i )

[0163] Where i represents the type of pulse signal identified, P(X / λ)i ) represents the maximum posterior probability.

[0164] In this embodiment, the feature coefficients of the unknown acoustic target are calculated with the training set. The maximum posterior probability of the acoustic feature coefficients in the training set is calculated, and the type corresponding to the maximum probability is calculated. This type is the category of the unknown acoustic target, and the identification is completed.

[0165] This application also proposes an application example of a pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform, taking a certain gunshot data as an example, including the following steps:

[0166] Step 1: Select gunshot data as training data; use the training data to generate Mel spectral coefficients. Use multiple sets of training data to generate MFCC coefficients one by one; generate MFCC coefficients and spectral difference coefficients for each set of data.

[0167] Step 2: Using multiple sets of training data, generate marginal spectral coefficients one by one; for each set of data, generate marginal spectral coefficients and spectral difference coefficients;

[0168] Step 3: GMM Model Parameter Estimation. Train the GMM model for the data using the coefficients stored in the array from the previous two steps, and then use EM parameter estimation.

[0169] Step 4: Process the gunshot data to be identified and generate identification coefficients. Using the data to be tested, generate the MFCC coefficients and marginal spectral coefficients of the data, as well as the spectral difference coefficients of the corresponding coefficients;

[0170] Step 5: Calculate the posterior probability and identify the target. Combine the coefficients of the test data with the GMM model data to calculate the maximum likelihood probability term of the test data, thereby identifying the target type of the test data. Identification is complete.

[0171] The applicant validated the method described in this application using MATLAB, and statistically analyzed the recognition rate and accuracy of each algorithm, as shown in Table 1 below:

[0172] Table 1. Statistics of algorithm recognition rates under different features

[0173]

[0174] The results show that the algorithm's recognition rate is significantly improved after introducing Hilbert marginal spectrum features, proving that this feature can effectively identify the local time-domain features of nonlinear and non-stationary acoustic signals. After introducing first-order difference spectrum features, the algorithm's recognition rate is further improved, proving that difference spectrum features are also an important identification feature of acoustic target signals and can effectively distinguish pulse signals with similar features.

[0175] The method in this application establishes a Gaussian mixture model (GMM) of the acoustic target, extracts feature parameters such as the marginal spectral coefficients, Mel spectral coefficients, and difference spectra of the acoustic target, realizes the identification of multi-dimensional feature space, further accurately identifies the details of time-frequency features, greatly improves the recognition rate of pulse acoustic targets, and overcomes the limitations of traditional methods.

[0176] This application also proposes a pulse sound recognition device based on Hilbert-Huang transform and Mel spectrum transform, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described above.

[0177] This application also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the impulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described above.

[0178] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0179] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0180] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0181] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.

Claims

1. A pulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform, characterized in that, include: The MFCC model and marginal spectrum model are pre-trained based on training pulse signals to generate MFCC coefficients and marginal spectrum coefficients, so as to construct a well-trained Gaussian mixture model. For a given pulse signal, Mel spectrum transformation and HHT transformation are performed on the pulse signal respectively to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal. Calculate the Gaussian probability density function of the MFCC coefficients and marginal spectral coefficients to generate a Gaussian mixture model of the pulse to be identified; Based on the Gaussian mixture model of the pulse to be identified and the trained Gaussian mixture model, a matching is performed to determine the type of pulse signal corresponding to the highest probability, which is the pulse type of the given pulse signal. The MFCC model and marginal spectral model are pre-trained based on training impulse signals to generate MFCC coefficients and marginal spectral coefficients, in order to construct a pre-trained Gaussian mixture model, including: A training sequence is constructed in advance by generating MFCC coefficients and marginal spectral coefficients based on the training pulse signal; Based on the training sequence, the GMM likelihood is expressed as: in, Indicates the estimated parameters, Represents the training sequence; The training process satisfies: A new parameter is estimated using the EM algorithm. This makes the likelihood under the new model parameters... ; Iterate until the model converges; The Gaussian probability density function for calculating the MFCC coefficients and marginal spectral coefficients to generate a Gaussian mixture model for the pulse to be identified includes: The probability density function of an M-order Gaussian mixture model is defined to satisfy: in, It is a D-dimensional random vector; , It is a sub-distribution; It is a mixed weight; The MFCC coefficients, first difference spectral coefficients, marginal spectral coefficients, and the first difference coefficients of the marginal spectrum are used as sub-distributions. To generate a Gaussian mixture model of the pulse to be identified.

2. The impulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described in claim 1, characterized in that, Based on the pulse signal, Mel spectrum transform and HHT transform are performed respectively to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal, including: After preprocessing the pulse signal, an FFT transformation is performed; Calculate the spectral line energy for each frame of FFT-processed data; For each frame of spectral energy, multiply it by the frequency domain response of the Merr filter and add them together to determine the energy passing through the Merr filter; Based on the energy passing through the Merr filter, the DCT is calculated to determine the MFCC coefficients and the first difference spectral coefficients.

3. The impulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described in claim 2, characterized in that, Based on the pulse signal, Mel spectrum transform and HHT transform are performed respectively to generate the MFCC coefficients and marginal spectral coefficients of the pulse signal, including: The pulse signal is subjected to EMD filtering to obtain multiple IMF components; Perform Hilbert transform on each IMF component; Based on the results of the Hilbert transform and the Hilbert spectrum, the Hilbert marginal spectrum and instantaneous energy density level are determined, satisfying: Where h(ω) represents the marginal spectrum of the signal, The first-order difference coefficients of the marginal spectrum are represented by H(ω,t), and the Hilbert spectrum is represented by H(ω,t). Marginal spectral coefficients are determined based on the marginal spectrum of the signal.

4. The impulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described in claim 1, characterized in that, Based on the Gaussian mixture model of the pulse to be identified and the trained Gaussian mixture model, a matching process is performed to determine the types of pulse signals corresponding to the highest probability, including: Based on the maximum a posteriori probability of Bayesian theory, the pulse signal is identified as belonging to a category of the training data, satisfying the following conditions: in, i Indicates the type of pulse signal identified. This represents the maximum posterior probability.

5. A pulse sound recognition device based on Hilbert-Huang transform and Mel spectrum transform, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the steps of the impulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the impulse sound recognition method based on Hilbert-Huang transform and Mel spectrum transform as described in any one of claims 1 to 4.

Images