Broiler chicken throat unhealthy state identification method based on acoustic characteristics

Abstract

The invention provides a broiler chicken throat unhealthy state recognition method based on acoustic features, and the method comprises the steps: collecting and carrying out the classification and marking of a plurality of broiler chicken voice segments, and obtaining the voice segments after classification and marking; de-noising processing, endpoint detection and framing processing are carried out on the voice segments after classification and marking, and recognized voice segments are obtained; extracting acoustic characteristic parameters in the recognition voice segments, performing matrix splicing on the acoustic characteristic parameters to obtain a characteristic parameter matrix, and outputting the characteristic parameter matrix; carrying out matrix fusion and dimension reduction processing on the extracted characteristic parameter matrix to obtain a low-dimensional matrix; inputting the low-dimensional matrix into a model for training to form an identification model; and inputting the to-be-detected data into the identification model for identification, and outputting an identification result. By improving the feature extraction mode and the type of the filter, the accuracy of acoustic feature recognition is improved, so that the model recognition precision and efficiency are improved, and the recognition cost is reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of breeding detection, and in particular relates to a method for identifying an unhealthy state of a broiler's throat based on acoustic features. Background Art

[0002] The calls of broiler chickens can often reflect the health status of their throats. By detecting and analyzing the audio characteristics of broiler calls, the status of their throats can be discovered in a timely manner, providing farmers with early warning information so that they can receive timely treatment and reduce their mortality rate due to illness.

[0003] Audio analysis technology can capture subtle changes in the calls of broilers. By combining a variety of technical means such as biosignal processing, computer vision and audio analysis, it can diagnose the health of broiler throats. Although audio technology has shown potential application value in broiler farming, its practical application still faces many challenges and limitations. At present, audio technology for broiler farming is still in the initial exploration stage, and the technical maturity is relatively low. Broilers of different breeds and different growth stages respond differently to audio signals. How to accurately design and adjust audio signals to meet the actual needs of broiler diagnosis and identification still requires further research and verification.

[0004] In addition, high-precision audio analysis technology requires high-quality audio playback equipment and audio analysis systems, but high-quality audio playback equipment and audio analysis systems are often expensive. For small and medium-sized broiler farms, the high cost of equipment may become a key factor restricting their adoption of audio technology. At the same time, the implementation of audio technology requires farmers to have certain technical knowledge and operating skills. Some farmers may find it difficult to master the key points of audio analysis technology, which affects its application effect.

[0005] Therefore, it is necessary to design a method for identifying the unhealthy state of broiler larynx based on acoustic features, which can use the lowest possible equipment cost and simple operation method to identify the healthy state of broilers through audio data, thereby improving the efficiency and accuracy of broiler larynx health status detection. Summary of the invention

[0006] The present invention relates to a method for identifying an unhealthy state of a broiler's larynx based on acoustic features, so as to at least solve the problem of low efficiency and accuracy in broiler's larynx health monitoring in the related art.

[0007] In order to achieve the above objectives, the present invention provides the following technical solutions: A method for identifying the unhealthy state of broiler throat based on acoustic features, the steps are as follows, including: S01, collecting multiple broiler chicken speech segments, and classifying and marking the speech segments to obtain classified and marked speech segments; S02, performing denoising and endpoint detection on the classified and marked speech segments to obtain recognized speech segments; S03, extracting acoustic feature parameters in the recognized speech segment, performing matrix splicing on the acoustic feature parameters to obtain a feature parameter matrix, and outputting the feature parameter matrix, wherein the acoustic feature parameters include WMFCC coefficients, WIMFCC coefficients and LPCC coefficients, and the extraction process of the WMFCC coefficients includes obtaining a spectrum representation of the recognized speech segment by wavelet transform, and inputting the spectrum representation of the recognized speech segment into different types of filter groups respectively, obtaining an energy value or other spectrum features of each frequency band divided by each filter group, and forming a frequency-energy matrix; S04, performing matrix fusion and dimensionality reduction processing on the feature parameter matrix to obtain a low-dimensional matrix; S05, inputting the low-dimensional matrix into the improved Gaussian mixture-hidden Markov model for training to form a recognition model; S06. Input the data to be tested into the recognition model for recognition, and output the recognition result.

[0008] Furthermore, the step of extracting acoustic features from the speech segment includes: Extract the WMFCC coefficients, the steps of extracting the WMFCC coefficients are: S11, pre-emphasize and perform frame-by-frame windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S12, performing discrete wavelet transform on each frame of the windowed signal to obtain a frequency spectrum representation of each frame. The purpose of performing discrete wavelet transform is to convert a time domain signal into a frequency domain signal; S13, inputting the spectrum representation of each frame into a group of Mel filter banks, performing nonlinear frequency division, performing matrix combination on the output energy and spectrum characteristics of each group of Mel filter banks to form a frequency-energy matrix, taking the logarithm of the frequency-energy matrix, and converting it into a logarithmic Mel spectrum; S14, performing complex wavelet transform on the logarithmic Mel spectrum to obtain WMFCC coefficients, normalizing the WMFCC coefficients, and outputting the WMFCC coefficients in the speech segment.

[0009] Furthermore, the complex wavelet transform introduces frequency domain features into the transform, combines audio features at different scales by a weighted combination method, and uses weights, number of scales and offsets in the weighted combination method to limit the audio features, which can more effectively capture the sound quality characteristics of the audio signal and improve the accuracy of audio signal analysis.

[0010] Furthermore, the Mel filter bank is a group of triangular bandpass filters equally spaced on the Mel frequency scale. The Mel filter bank can perform nonlinear division of frequency, and usually 24 Mel filters are used to cover the entire frequency range.

[0011] Furthermore, the types of the Mel filter bank include: An equal loudness filter bank, wherein the equal loudness filter bank sets dense low-frequency filters and relatively sparse high-frequency filters based on the loudness sensitivity of the human ear to different frequencies, so that the loudness perception of high frequency, medium frequency and low frequency is in the same frequency band range; A harmonic filter bank, wherein the harmonic filter bank is based on the harmonic characteristics of the pitch, takes the fundamental frequency as the basis, increases the center frequencies of several identical harmonic filters by different integer multiples, and sets a narrow frequency band near the center frequency of each filter to capture the main harmonic characteristics; A formant filter bank, wherein the formant filter bank sets the formant peak frequency to a common human voice resonance frequency based on the formant characteristics of speech, so that the filter is concentrated around the formant and can capture acoustic features at a specific frequency; A pitch filter bank, which is mainly used to capture the pitch characteristics of low to ultra-high frequency bands. The pitches are arranged logarithmically based on the starting frequency, usually in a multiple increasing manner, and a wider bandwidth is used near the pitch frequency to cover the pitch changes, which can cover a wide frequency domain from bass to treble; The spectral feature filter group is mainly used to capture the spectral center of gravity of the sound and its changes. It takes the low-frequency starting frequency as the benchmark, uniformly covers the spectrum in the form of exponential increase, and adjusts the coverage range in the vicinity of the pitch frequency in a logarithmic increase manner. The coverage range of the spectral feature filter group can completely cover the pitch changes from low frequency to ultra-high frequency, and is suitable for analyzing the overall spectral characteristics.

[0012] Furthermore, the process of extracting acoustic feature parameters from the speech segment also includes: Extract WIMFCC coefficients. The steps of extracting WIMFCC coefficients are as follows: S21, pre-emphasize and perform frame-by-frame windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S22, performing discrete wavelet transform on each frame of the windowed signal to obtain a frequency spectrum representation of each frame. The purpose of performing discrete wavelet transform is to convert a time domain signal into a frequency domain signal; S23, inputting the spectrum representation of each frame into a set of inverse Mel filter banks, performing nonlinear frequency division, performing matrix combination on the output energy and spectrum characteristics of each set of inverse Mel filter banks to form an inverse frequency-energy matrix, taking the logarithm of the inverse frequency-energy matrix, and converting it into a logarithmic Mel spectrum; S24, performing complex wavelet transform on the inverse logarithmic Mel spectrum to obtain WIMFCC coefficients, normalizing the WIMFCC coefficients, and outputting the WIMFCC coefficients in the speech segment.

[0013] Furthermore, the process of extracting acoustic feature parameters from the speech segment also includes: Extracting LPCC coefficients, the steps of extracting LPCC coefficients are: S31, performing windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S32, calculating the autocorrelation sequence of each frame of the windowed signal to obtain the similarity between the signal and its delayed version; S33, initializing the obtained autocorrelation sequence, and storing the obtained initialization sequence; S34, inputting the initialization sequence into the Levinson-Durbin recursive algorithm, and obtaining the complete LPC coefficients by gradually calculating the coefficients of each predictor; S35, performing logarithmic processing on the LPC coefficients, and transforming the processed LPC coefficients to obtain a representation of the LPC coefficients in the frequency domain; S36. Obtain the real part of the transformation result, and use the real part of the transformation result as the LPCC coefficient.

[0014] Furthermore, the method of performing matrix fusion on the characteristic parameter matrix includes performing feature dimension reduction processing on the characteristic parameter matrix using principal component analysis to obtain a low-dimensional matrix; The steps of principal component analysis include: S41, standardizing each row of the characteristic parameter matrix to obtain a standardized matrix, wherein the standardization method is to subtract the mean value of the row from the value of the matrix; S42, calculating the covariance matrix of the standardized matrix, and calculating the eigenvalues ​​and corresponding eigenvectors of the covariance matrix; S43, arranging the corresponding eigenvectors into a matrix by row according to the corresponding values ​​to form a complete component matrix, and taking the first N eigenvectors of the complete component matrix as the principal components; S44. Transpose the first N eigenvectors, construct and perform data projection, and obtain a low-dimensional matrix.

[0015] Furthermore, the low-dimensional matrix is ​​preferably a one-dimensional matrix.

[0016] Furthermore, the improved Gaussian mixture-hidden Markov model includes: A hidden Markov model, wherein the hidden Markov model is used to describe time series data, and the hidden Markov model includes a hidden state, a state transition probability matrix, an observation probability distribution, and an initial state distribution; A Gaussian mixture model is provided, wherein the Gaussian mixture model models the complex distribution of data by weighted sum of multiple Gaussian distributions, uses a shape-based clustering algorithm instead of a K-means algorithm for initial cluster division, and uses a grid search method to optimize the parameters of the number of states of a hidden Markov model and the number of Gaussian mixture components.

[0017] By introducing wavelet transform and complex wavelet transform into the extraction process of MFCC coefficients and IMFCC, the sound quality characteristics of audio signals can be captured more effectively, and the efficiency and accuracy of audio signal analysis can be improved; by adding different types of Mel filter groups to the extraction process of MFCC coefficients and IMFCC, the pitch changes from low frequency to ultra-high frequency can be fully detected, thereby improving the accuracy of audio signal analysis; by establishing a Gaussian mixture-hidden Markov model, optimizing the parameters of the number of states and the number of Gaussian mixture components of the hidden Markov model, and training the model according to the training set, the trained model can be used for existing farm machinery and equipment, reducing equipment costs and simplifying the operation method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] in: Figure 1 The flowchart of identifying the unhealthy state of broiler larynx based on acoustic features; Figure 2 Flowchart for training a model for identifying the unhealthy state of broiler larynx based on acoustic features; Figure 3 Flowchart for extracting Mel-frequency cepstral coefficients improved by wavelet; Figure 4 The flowchart for extracting three acoustic features; Figure 5 A schematic diagram of the Mel filter bank used in the WMFCC coefficient extraction process; Figure 6 Flowchart of dimensionality reduction for principal component analysis; Figure 7 Flowchart for endpoint detection; Figure 8 Spectrogram for cooing sounds; Fig. 9 This is a spectrogram of the sound of mucus accumulation in the throat; Fig.10 This is a spectrogram of a normal call. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of this application.

[0021] In order to identify the health status of broilers through their vocalization status, it is necessary to first identify the vocalization status of the larynx of each broiler in the farm. Based on this, the embodiment of the present disclosure provides a method for identifying the unhealthy status of the larynx of broilers based on acoustic features. Figure 1 to Figure 2 , the method comprising, S01, collecting multiple broiler chicken speech segments, and classifying and marking the speech segments to obtain classified and marked speech segments; S02, performing denoising and endpoint detection on the classified and marked speech segments to obtain recognized speech segments; S03, extracting acoustic feature parameters in the recognized speech segment, performing matrix splicing on the acoustic feature parameters to obtain a feature parameter matrix, and outputting the feature parameter matrix, wherein the acoustic feature parameters include WMFCC coefficients, WIMFCC coefficients and LPCC coefficients, and the extraction process of the WMFCC coefficients includes obtaining a spectrum representation of the recognized speech segment by wavelet transform, and inputting the spectrum representation of the recognized speech segment into different types of filter groups respectively, obtaining an energy value or other spectrum features of each frequency band divided by each filter group, and forming a frequency-energy matrix; S04, performing matrix fusion and dimensionality reduction processing on the extracted frequency-energy matrix to obtain a low-dimensional matrix; S05, inputting the low-dimensional matrix into the improved Gaussian mixture-hidden Markov model for training to form a recognition model; S06. Input the data to be tested into the recognition model for recognition, and output the recognition result.

[0022] The process of extracting acoustic feature parameters from the speech segment includes: Extract the WMFCC coefficients, the steps of extracting the WMFCC coefficients are: S11, pre-emphasize and perform frame-by-frame windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S12, performing discrete wavelet transform on each frame of the windowed signal to obtain a frequency spectrum representation of each frame. The purpose of performing discrete wavelet transform is to convert a time domain signal into a frequency domain signal; S13, inputting the spectrum representation of each frame into a group of Mel filter banks, performing nonlinear frequency division, performing matrix combination on the output energy and spectrum characteristics of each group of Mel filter banks to form a frequency-energy matrix, taking the logarithm of the frequency-energy matrix, and converting it into a logarithmic Mel spectrum; S14, performing complex wavelet transform on the logarithmic Mel spectrum to obtain WMFCC coefficients, normalizing the WMFCC coefficients, and outputting the WMFCC coefficients in the speech segment.

[0023] like Figure 3 As shown in the figure, the spectrum obtained by discrete wavelet transform is passed through a set of Mel filters (Mel filterbank). The Mel filter bank is a set of triangular bandpass filters equally spaced on the Mel frequency scale. The Mel filter bank can perform nonlinear division of frequency. Usually 24 Mel filters are used to cover the entire frequency range. The extraction process of acoustic features through different types of filters is as follows: Figure 4 shown.

[0024] The types of Mel filter banks include: An equal loudness filter bank, wherein the equal loudness filter bank sets dense low-frequency filters and sparse high-frequency filters based on the loudness sensitivity of the human ear to different frequencies, which conforms to the equal loudness curve, wherein the perceived loudness of low and high frequencies is lower than that of the mid-frequency band, so denser filters are needed to enhance the response of these areas, and the perceived loudness of high, mid-frequency and low frequencies is made to be within the same frequency band by setting the sparsity of the filters; The center frequency formula is : The gradually increasing frequency band design can use the logarithmic distribution formula

[0025] is the starting frequency, is the filter number, is a constant that controls the density of the filter. Take a smaller value to increase density, and take a larger value in the high frequency band to increase the interval. =5.

[0026] Frequency Range: Set the appropriate frequency band range above and below the center frequency to cover the frequency fluctuation range:

[0027] in It is a floating percentage of the frequency range, suitable for flexible coverage of different frequency bands. =0.2.

[0028] The number of filters in the equal loudness filter bank and their frequency ranges are shown in the following table:

[0029] A harmonic filter bank, wherein the harmonic filter bank is based on the harmonic characteristics of the pitch. The fundamental frequency and its integer multiple frequencies have an important influence on the timbre and pitch characteristics of the sound. Based on the fundamental frequency, the center frequencies of several identical harmonic filters are increased by different integer multiples, and narrow frequency bands are set near the center frequency of each filter to capture the main harmonic characteristics; The center frequency formula is :

[0030] in is the filter number, is the fundamental frequency. For example, the fundamental frequency of human voice can be around 100Hz, and appropriate adjustment can cover the target harmonic range. =100Hz.

[0031] Frequency range: Sets a narrow frequency band around the center frequency to capture the main harmonic characteristics. The frequency range can use a relatively narrow bandwidth, for example:

[0032]

[0033] in Indicates bandwidth, set here =20Hz (approximately 20% of the center frequency).

[0034] The number of filters in the harmonic filter bank and their covered frequency ranges are shown in the following table:

[0035] Formant filter banks are based on the formant characteristics of speech, which usually appear in the low and mid-frequency bands to enhance the resonance characteristics of the sound at specific frequencies. Formants are very important in the production of phonemes and are therefore crucial to human voice recognition and clarity.

[0036] Center Frequency Formula : Based on the fixed frequency band design, it can be directly set to a specific resonance peak frequency, such as 250Hz, 500Hz, etc. These are common human voice resonance frequencies. The formula is:

[0037] It is the low frequency reference, which is set to 250Hz here.

[0038] Frequency range: The bandwidth depends on the distribution of each resonance peak frequency. Usually the resonance peak bandwidth is around 200Hz, and the formula can be used:

[0039]

[0040] Focusing the filter around the resonance peak can capture the acoustic characteristics of this frequency band.

[0041] The number of filters in the formant filter bank and their frequency coverage range are shown in the following table:

[0042] A pitch filter bank, which is mainly used to capture the pitch characteristics of low to ultra-high frequency bands. The pitches are arranged logarithmically based on the starting frequency, usually in a multiple increasing manner, and a wider bandwidth is used near the pitch frequency to cover the pitch changes, which can cover a wide frequency domain from bass to treble; The center frequency formula is : Logarithmically arrange the tones, usually using a multiple increasing formula:

[0043] in is the starting frequency, which is suitable for the logarithmic distribution characteristics of the tone. Here the starting frequency is set to 100Hz.

[0044] Frequency Range: Use a wider bandwidth around the tone frequency to cover the tone variation. The frequency range can be selected as:

[0045]

[0046] The number of filters in the tone filter bank and their frequency coverage range are shown in the following table:

[0047] The spectral feature filter group is mainly used to capture the spectral center of gravity of the sound and its changes. It takes the low-frequency starting frequency as the benchmark, uniformly covers the spectrum in the form of exponential increase, and adjusts the coverage range in the vicinity of the pitch frequency in a logarithmic increase manner. The coverage range of the spectral feature filter group can completely cover the pitch changes from low frequency to ultra-high frequency, and is suitable for analyzing the overall spectral characteristics.

[0048] The center frequency formula is : Usually an exponential increase is used to evenly cover the spectrum

[0049] in is the low frequency starting frequency, here the low frequency starting frequency is set to 50Hz.

[0050] Frequency range: The frequency range increases logarithmically according to the importance and characteristics of the frequency band:

[0051]

[0052] in Set to the frequency range floating ratio, here the frequency range floating ratio is set to 0.1.

[0053] The number of filters in the tone filter bank and their frequency coverage range are shown in the following table:

[0054] The output energy of each filter group is formed into a frequency-energy matrix and input into the feature processing module. The logarithm of the frequency-energy matrix of each filter group is taken and converted into a logarithmic Mel spectrum. This step is to convert the nonlinear spectrum energy into a linear relationship that is more suitable for subsequent processing.

[0055] Among them, the logarithmic transformation is to be more in line with the perception of the human ear, and its formula is as follows:

[0056] Among them is Output energy, is a translation constant to avoid logarithmic zero.

[0057] The logarithmic Mel spectrum is subjected to complex wavelet transform to obtain the MFCC coefficients in the speech segment. During the wavelet transform process, the 2nd to 13th coefficients after the transform are taken as MFCC coefficients, and the 1st coefficient is usually discarded because it may contain non-speech information.

[0058] The extracted MFCC coefficients are normalized to eliminate the differences in amplitudes of different audio signals and improve the stability and accuracy of subsequent processing.

[0059] The complex wavelet transform introduces frequency domain features into the transform, combines audio features at different scales by a weighted combination method, and uses weights, scale numbers and offsets in the weighted combination method to limit audio features, which can more effectively capture the sound quality features of audio signals and improve the accuracy of audio signal analysis.

[0060] In the improved complex wavelet transform, the standard complex wavelet transform formula is:

[0061] in: is the complex wavelet transform coefficient, is the input signal, is the wavelet function, For scale, is the translation parameter.

[0062] In order to enhance the ability to capture sound quality, we introduce frequency domain features into the transformation. The improved formula is as follows:

[0063] in: is the improved complex wavelet transform coefficient, It is a frequency domain feature function, which represents the additional features of the audio signal. Here we choose loudness and clarity.

[0064] In order to specifically represent the loudness and clarity of the audio signal, we can use It is divided into two parts: one for the characteristics representing loudness and the other for the characteristics representing clarity.

[0065] Loudness can be measured by the amount of energy in a specific frequency range, usually using the following formula:

[0066] in: , It is in The energy of the filter bank output,

[0067] In order to combine the loudness and frequency characteristics, the frequency domain characteristic function of the loudness can be expressed as:

[0068] in: is the weight, which indicates the contribution of each frequency band to the loudness, and is usually weighted according to the sensitivity of the human ear to different frequencies.

[0069] Clarity is usually associated with high-frequency content and can be measured by the ratio of high-frequency energy to low-frequency energy:

[0070] in: is the instantaneous value of sharpness; HF represents the index set of the high-frequency filter bank; LF represents the index set of the low-frequency filter bank.

[0071] Combining clarity with frequency characteristics, it can be expressed as:

[0072] The final frequency domain characteristic function It can be expressed as:

[0073] The specific formula is:

[0074] Perform complex wavelet transform on the signal at different scales, and obtain the comprehensive result by weighted combination method:

[0075] in: is the comprehensive improved complex wavelet transform coefficient; are weights, reflecting the importance of different scales; is the scale quantity; are parameters for different scales and offsets.

[0076] Through the above improvement scheme, the improved complex wavelet transform can more effectively capture the sound quality characteristics of the audio signal. This method combines multi-scale analysis and frequency domain characteristics, which helps to improve the accuracy of audio signal analysis.

[0077] The logarithmic power matrix is ​​transformed into the cepstrum domain by applying the improved complex wavelet transform instead of the traditional discrete cosine transform. Compared with the discrete wavelet transform, the improved complex wavelet transform can not only capture the amplitude and phase information of the signal, but also effectively capture the sound quality of the audio, as well as the concentrated energy and compression information. The first n improved complex wavelet transform coefficients of the five filter groups corresponding to the number of filters of each filter group are retained respectively (n is the number of filters in each filter group). These five groups of low-order coefficients describe the main acoustic characteristics of the audio.

[0078] The Mel-frequency cepstral coefficient is based on the Mel frequency scale, and the linear prediction cepstral coefficient is based on linear prediction analysis. The inverse Mel-frequency cepstral coefficient (WIMFCC) is introduced to fully characterize the spectral characteristics of the high-frequency range.

[0079] The process of extracting acoustic feature parameters from the speech segment also includes: Extract WIMFCC coefficients. The steps of extracting WIMFCC coefficients are as follows: S21, pre-emphasize and perform frame-by-frame windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S22, performing discrete wavelet transform on each frame of the windowed signal to obtain a frequency spectrum representation of each frame. The purpose of performing discrete wavelet transform is to convert a time domain signal into a frequency domain signal; S23, inputting the spectrum representation of each frame into a group of inverse Mel filter banks, performing nonlinear frequency division, performing matrix combination on the output energy and spectrum characteristics of each group of inverse Mel filter banks to form an inverse frequency-energy matrix, taking the logarithm of the inverse frequency-energy matrix, and converting it into an inverse logarithmic Mel spectrum; S24, performing complex wavelet transform on the inverse logarithmic Mel spectrum to obtain WIMFCC coefficients, normalizing the WIMFCC coefficients, and outputting the WIMFCC coefficients in the speech segment.

[0080] The WIMFCC coefficients are sparsely distributed in the low-frequency part of the filter and densely distributed in the high-frequency part, which can characterize the spectral characteristics of the high-frequency range.

[0081] The process of extracting acoustic feature parameters from the speech segment also includes: Extracting LPCC coefficients, the steps of extracting LPCC coefficients are: S31, performing windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S32, calculating the autocorrelation sequence of each frame of the windowed signal to obtain the similarity between the signal and its delayed version; S33, initializing the obtained autocorrelation sequence, and storing the obtained initialization sequence; S34, inputting the initialization sequence into the Levinson-Durbin recursive algorithm, and obtaining the complete LPC coefficients by gradually calculating the coefficients of each predictor; S35, performing logarithmic processing on the LPC coefficients, and transforming the processed LPC coefficients to obtain a representation of the LPC coefficients in the frequency domain; S36. Obtain the real part of the transformation result, and use the real part of the transformation result as the LPCC coefficient.

[0082] The method for performing matrix fusion on the characteristic parameter matrix includes performing feature dimension reduction processing on the characteristic parameter matrix using principal component analysis to obtain a low-dimensional matrix; like Figure 5 The flowchart of dimensionality reduction of principal component analysis is shown, and the steps of principal component analysis include: S41, standardizing each row of the characteristic parameter matrix to obtain a standardized matrix, wherein the standardization method is to subtract the mean value of the row from the value of the matrix; S42, calculating the covariance matrix of the standardized matrix, and calculating the eigenvalues ​​and corresponding eigenvectors of the covariance matrix; S43, arranging the corresponding eigenvectors into a matrix by row according to the corresponding values ​​to form a complete component matrix, and taking the first N eigenvectors of the complete component matrix as the principal components; S44. Transpose the first N eigenvectors, construct and perform matrix projection, and obtain a low-dimensional matrix. For the convenience of value acquisition, the low-dimensional matrix is ​​preferably a one-dimensional matrix.

[0083] The specific details are as follows: subtract the mean of each feature in the original data set so that the mean of the new data set is 0. Divide the centered data by the standard deviation of each feature so that the scale of each feature is the same. Use the centered data to calculate the covariance matrix, perform eigenvalue decomposition on the covariance matrix, and obtain the eigenvalues ​​and corresponding eigenvectors. Sort the eigenvalues ​​from large to small, and select the first k largest eigenvalues. The eigenvectors corresponding to these eigenvalues ​​are the principal component directions we are looking for. Use the selected k eigenvectors as column vectors to form a projection matrix W. Project the standardized data into the k-dimensional subspace defined by W to obtain a low-dimensional matrix.

[0084] The improved Gaussian mixture-hidden Markov model includes: A hidden Markov model, wherein the hidden Markov model is used to describe time series data, and the hidden Markov model includes a hidden state, a state transition probability matrix, an observation probability distribution, and an initial state distribution; A Gaussian mixture model is provided, wherein the Gaussian mixture model models the complex distribution of data by weighted sum of multiple Gaussian distributions, uses a shape-based clustering algorithm instead of a K-means algorithm for initial cluster division, and uses a grid search method to optimize the parameters of the number of states of a hidden Markov model and the number of Gaussian mixture components.

[0085] A Hidden Markov Model (HMM) is a statistical model used to describe time series data. It consists of the following components: The hidden states of the system, which change with certain transition probabilities. The state transition probability matrix, which defines the transition probability from one state to another. The observation probability distribution, which is the probability distribution of the observed data generated under each hidden state. The initial state distribution, which is the probability distribution of the state of the system at the initial time.

[0086] Gaussian mixture model (GMM) is a probabilistic model for representing data. It models the complex distribution of data through the weighted sum of multiple Gaussian distributions. Each Gaussian distribution has its mean, covariance and weight. In the traditional Gaussian mixture model (GMM), the K-means algorithm provides the initial cluster center (i.e., the mean of the Gaussian component) through cluster analysis. This initialization method usually provides a reasonable starting point, thereby helping the EM algorithm to converge to a better solution faster. Here, the K-shape algorithm is used instead of the K-means algorithm to perform initial cluster division and obtain the cluster center of the time series. The K-shape algorithm focuses on the shape characteristics of the time series, while the traditional K-means is based on the Euclidean distance and is not suitable for processing shape changes in time series data. The steps of the K-shape algorithm are as follows: First, randomly select K time series as the initial cluster center. For each time series, calculate its shape similarity with all cluster centers, assign it to the cluster with the highest similarity, and calculate the new cluster center. The new cluster center is the average shape of all time series in the current cluster. Calculate the shape similarity of the new cluster center and update the cluster center. The assignment step and the update step are performed repeatedly until the change in cluster center is lower than the set threshold or the maximum number of iterations is reached.

[0087] The process of applying the multi-classification model based on the improved Gaussian mixture-hidden Markov model to train the data set to be predicted includes the following steps: The sound collection module receives the broiler sound signal, classifies and marks the received broiler sound signal, inputs the classified and marked speech segment into the sound processing module, performs pre-emphasis, noise reduction, frame windowing and endpoint detection processing, forms a recognized speech segment, regards the corresponding start frame and end frame as a combination, and a combination of multiple start frames and end frames arranged in chronological order; then, for each combination, a frequency-energy matrix is ​​formed by extracting features, and the frequency-energy matrix and the frequency domain features are input into the feature processing module together. The feature processing module forms a training set and a test set through logarithmic calculation, complex wavelet transform and screening, and the training set and the test set are input into a multi-classification model based on an improved Gaussian mixture-hidden Markov model to obtain a recognition result of a sound category of a broiler sound signal to be identified, performs grid search on the training result, and debugs the model parameters, thereby improving the recognition accuracy of the model.

[0088] The audio data is collected by using the Xinke V-01 audio collection device for collecting multiple broiler voice segments. The audio recording parameters are sampling accuracy 16 bits, sampling frequency 16KHz, and storage format wav. After the broilers eat, the audio collection device is placed close to the chicken cage and on the outer layer of the chicken cage to collect audio data. The collected audio data is screened, sorted and classified and marked, and the normal calls of broilers and the calls of broilers in unhealthy throats are classified and marked by artificial listening methods to obtain the classified and marked voice segments. Among them, the calls of broilers in unhealthy throats include the clucking calls of broilers and the calls of broilers with mucus accumulation in the throat.

[0089] The purpose of pre-emphasis on the classified and marked speech segments is to enhance the high-frequency components of the signal, which helps to balance the frequency attenuation caused by the vocal cords and lip effects in the audio signal. The pre-emphasis process can be described as:

[0090] in, is the nth sample of the input signal, is the nth sample of the output signal after pre-emphasis, is the pre-emphasis coefficient.

[0091] The pre-emphasized audio data is processed by frame windowing. The main purpose of frame windowing is to improve the signal processing effect by dividing the signal into smaller frames and applying a window function when processing a continuous signal. In this implementation, a Hamming window is selected, and the expression is as follows:

[0092] in: is the window length; The original sound signal without signal filtering contains a lot of noise interference, so it needs to be filtered to retain as many useful sound signals as possible. The spectrum energy distribution range of the cooing sound is from 2100 Hz to 400 Hz. Figure 7 As shown in the figure, the spectrum energy distribution range of the sound caused by mucus accumulation in the throat is from 2500 Hz to 600 Hz. Figure 8 As shown; the spectrum energy distribution range of normal calls is from 4500 Hz to 2000 Hz, such as Fig. 9 As shown in FIG. , combined with the improved spectral subtraction method, speech enhancement of sound data is achieved.

[0093] The improved spectral subtraction method is used for speech enhancement. The specific implementation method is as follows: Convert the pre-emphasized time domain signal to the frequency domain using short-time Fourier transform, as shown below:

[0094] In the formula is the spectrum representation, represents the frequency index, represents the timeframe index; Select the noise segment in the audio and calculate the average amplitude spectrum of the noise. The expression is as follows:

[0095] In the formula is the number of frames of the noise segment;

[0096] Where: is the exponential parameter of the amplitude spectrum subtraction; is the parameter that controls the improvement of the signal-to-noise ratio; is the enhanced speech signal; is the spectrum representation, represents the frequency index, represents the timeframe index; is the average amplitude spectrum of the audio signal noise; The steps for endpoint detection are as follows: Figure 6 As shown in the figure, its purpose is to distinguish the speech and non-speech areas in the sound signal. By calculating the average short-time energy Ms and the average zero-crossing rate Zs of the audio, set half of the average of the short-time energy of all frames as the high energy threshold MH, the energy mean of the silent part and half of the average of MH as the low energy threshold ML, and the zero-crossing rate threshold Zs. Take a higher short-time energy as the threshold MH. Using this threshold, we can first separate the voiced part of the speech. Then take a lower energy threshold ML. Using this threshold, we can search from the previously selected voiced part to both ends, and add the speech part of the lower energy segment to the speech segment to further expand the speech segment range. Since the two ends of the speech are consonants, they are also part of the speech, but the energy of the consonants is as low as the energy of the silent part, but the zero-crossing rate is much higher than that of the silent part. In order to distinguish the two, the speech segment that has been distinguished by short-time energy will continue to search to both ends. The part with a short-time zero-crossing rate greater than 3 times Zs is considered to be the unvoiced part of the speech. Add this part to the speech segment to get the final speech segment.

[0097] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for identifying the unhealthy state of broiler throat based on acoustic features, characterized in that: The steps are as follows, including, S01, collecting multiple broiler chicken speech segments, and classifying and marking the speech segments to obtain classified and marked speech segments; S02, performing denoising and endpoint detection on the classified and marked speech segments to obtain recognized speech segments; S03, extracting acoustic feature parameters in the recognized speech segment, performing matrix splicing on the acoustic feature parameters to obtain a feature parameter matrix, and outputting the feature parameter matrix, wherein the acoustic feature parameters include WMFCC coefficients, WIMFCC coefficients and LPCC coefficients, and the extraction process of the WMFCC coefficients includes obtaining a spectrum representation of the recognized speech segment by wavelet transform, and inputting the spectrum representation of the recognized speech segment into different types of filter groups respectively, obtaining an energy value or other spectrum features of each frequency band divided by each filter group, and forming a frequency-energy matrix; S04, performing matrix fusion and dimensionality reduction processing on the feature parameter matrix to obtain a low-dimensional matrix; S05, inputting the low-dimensional matrix into the improved Gaussian mixture-hidden Markov model for training to form a recognition model; S06. Input the data to be tested into the recognition model for recognition, and output the recognition result.

2. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 1, characterized in that: The process of extracting acoustic feature parameters from a speech segment includes: Extract the WMFCC coefficients, the steps of extracting the WMFCC coefficients are: S11, pre-emphasize and perform frame-by-frame windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S12, performing discrete wavelet transform on each frame of the windowed signal to obtain a frequency spectrum representation of each frame. The purpose of performing discrete wavelet transform is to convert a time domain signal into a frequency domain signal; S13, inputting the spectrum representation of each frame into a group of Mel filter banks, performing nonlinear frequency division, performing matrix combination on the output energy and spectrum characteristics of each group of Mel filter banks to form a frequency-energy matrix, taking the logarithm of the frequency-energy matrix, and converting it into a logarithmic Mel spectrum; S14, performing complex wavelet transform on the logarithmic Mel spectrum to obtain WMFCC coefficients, normalizing the WMFCC coefficients, and outputting the WMFCC coefficients in the speech segment.

3. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 2, characterized in that: The complex wavelet transform introduces frequency domain features into the transform, combines audio features at different scales by a weighted combination method, and uses weights, scale numbers and offsets in the weighted combination method to limit audio features, which can more effectively capture the sound quality features of audio signals and improve the accuracy of audio signal analysis.

4. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 3, characterized in that: The Mel filter bank is a group of triangular bandpass filters equally spaced on the Mel frequency scale. The Mel filter bank can perform nonlinear division of frequency, and usually 24 Mel filters are used to cover the entire frequency range.

5. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 4, characterized in that: The types of Mel filter banks include: An equal loudness filter bank, wherein the equal loudness filter bank sets dense low-frequency filters and relatively sparse high-frequency filters based on the loudness sensitivity of the human ear to different frequencies, so that the loudness perception of high frequency, medium frequency and low frequency is in the same frequency band range; A harmonic filter bank, wherein the harmonic filter bank is based on the harmonic characteristics of the pitch, takes the fundamental frequency as the basis, increases the center frequencies of several identical harmonic filters by different integer multiples, and sets a narrow frequency band near the center frequency of each filter to capture the main harmonic characteristics; A formant filter bank, wherein the formant filter bank sets the formant peak frequency to a common human voice resonance frequency based on the formant characteristics of speech, so that the filter is concentrated around the formant and can capture acoustic features at a specific frequency; A pitch filter bank, which is mainly used to capture the pitch characteristics of low to ultra-high frequency bands. The pitches are arranged logarithmically based on the starting frequency, usually in a multiple increasing manner, and a wider bandwidth is used near the pitch frequency to cover the pitch changes, which can cover a wide frequency domain from bass to treble; The spectral feature filter group is mainly used to capture the spectral center of gravity of the sound and its changes. It takes the low-frequency starting frequency as the benchmark, uniformly covers the spectrum in the form of exponential increase, and adjusts the coverage range in the vicinity of the pitch frequency in a logarithmic increase manner. The coverage range of the spectral feature filter group can completely cover the pitch changes from low frequency to ultra-high frequency, and is suitable for analyzing the overall spectral characteristics.

6. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 5, characterized in that: The process of extracting acoustic feature parameters from the speech segment also includes: Extract WIMFCC coefficients. The steps of extracting WIMFCC coefficients are as follows: S21, pre-emphasize and perform frame-by-frame windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S22, performing discrete wavelet transform on each frame of the windowed signal to obtain a frequency spectrum representation of each frame. The purpose of performing discrete wavelet transform is to convert a time domain signal into a frequency domain signal; S23, inputting the spectrum representation of each frame into a group of inverse Mel filter banks, performing nonlinear frequency division, performing matrix combination on the output energy and spectrum characteristics of each group of inverse Mel filter banks to form an inverse frequency-energy matrix, taking the logarithm of the inverse frequency-energy matrix, and converting it into an inverse logarithmic Mel spectrum; S24, performing complex wavelet transform on the inverse logarithmic Mel spectrum to obtain WIMFCC coefficients, normalizing the WIMFCC coefficients, and outputting the WIMFCC coefficients in the speech segment.

7. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 6, characterized in that: The process of extracting acoustic feature parameters from the speech segment also includes: Extracting LPCC coefficients, the steps of extracting LPCC coefficients are: S31, pre-emphasize and perform frame-by-frame windowing processing on the sound segment signal to obtain a windowed signal of each frame of the sound segment, wherein the windowing processing function adopts a Hening window; S32, calculating the autocorrelation sequence of each frame of the windowed signal to obtain the similarity between the signal and its delayed version; S33, initializing the obtained autocorrelation sequence, and storing the obtained initialization sequence; S34, inputting the initialization sequence into the Levinson-Durbin recursive algorithm, and obtaining the complete LPC coefficients by gradually calculating the coefficients of each predictor; S35, performing logarithmic processing on the LPC coefficients, and transforming the processed LPC coefficients to obtain a representation of the LPC coefficients in the frequency domain; S36. Obtain the real part of the transformation result, and use the real part of the transformation result as the LPCC coefficient.

8. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 7, characterized in that: The method for performing matrix fusion on the characteristic parameter matrix includes performing feature dimension reduction processing on the characteristic parameter matrix using principal component analysis to obtain a low-dimensional matrix; The steps of principal component analysis include: S41, standardizing each row of the characteristic parameter matrix to obtain a standardized matrix, wherein the standardization method is to subtract the mean value of the row from the value of the matrix; S42, calculating the covariance matrix of the standardized matrix, and calculating the eigenvalues ​​and corresponding eigenvectors of the covariance matrix; S43, arranging the corresponding eigenvectors into a matrix by row according to the corresponding values ​​to form a complete component matrix, and taking the first N eigenvectors of the complete component matrix as the principal components; S44. Transpose the first N eigenvectors, construct and perform data projection, and obtain a low-dimensional matrix.

9. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 8, characterized in that: The low-dimensional matrix is ​​preferably a one-dimensional matrix.

10. The method for identifying the unhealthy state of broiler larynx based on acoustic features according to claim 9, characterized in that: The improved Gaussian mixture-hidden Markov model includes: A hidden Markov model, wherein the hidden Markov model is used to describe time series data, and the hidden Markov model includes a hidden state, a state transition probability matrix, an observation probability distribution, and an initial state distribution; A Gaussian mixture model is provided, wherein the Gaussian mixture model models the complex distribution of data by weighted sum of multiple Gaussian distributions, uses a shape-based clustering algorithm instead of a K-means algorithm for initial cluster division, and uses a grid search method to optimize the parameters of the number of states of a hidden Markov model and the number of Gaussian mixture components.

Images