Newborn cry recognition system based on machine learning

Abstract

The invention discloses a newborn cry recognition system based on machine learning, which belongs to the technical field of intelligent newborn nursing and comprises a data acquisition module, a feature extraction module, a feature selection module, a newborn cry recognition module and a recognition result feedback module. According to the method, the improved badger optimization algorithm is adopted to perform feature selection on the cry audio feature set, key features beneficial to distinguishing different cry categories are automatically selected, manual intervention errors are avoided, feature selection can be better and flexibly adjusted according to individual differences, and it is ensured that the system has high recognition accuracy on different newborn cry; according to the method, the mixed kernel function support vector machine model combined with the K nearest neighbor algorithm is adopted to carry out newborn cry recognition, local similarity improvement between samples is captured, boundaries between different cry categories are accurately divided, and overlapping between the categories is avoided, so that different types of cry are recognized more accurately and efficiently, and the overall usability of cry recognition is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent neonatal care, and specifically refers to a neonatal cry recognition system based on machine learning. Background Art

[0002] Newborn cry recognition uses audio signal processing technology to extract the characteristics of newborn cries, and combines it with machine learning technology for further analysis to achieve automatic recognition of newborn cry categories. It aims to assist nursing staff in quickly judging the needs of newborns, and then respond to the needs of newborns in a timely manner, better improve the quality of care, and provide important technical support for the intelligentization of newborn care.

[0003] However, in the existing process of newborn cry recognition, the original audio of the cry is used as signal data. Under environmental conditions, it is easy to contain a large amount of redundant information and noise. When it is applied to the recognition of newborn cries, the large differences in the cries of newborns and the low specificity of the overall audio lead to increased computational complexity. The cry characteristics of different newborns vary due to individual differences. Therefore, the audio features of newborns often have to rely on manual intervention to be extracted. Traditional intelligent methods are difficult to adapt to the dynamic changes of data features and are more likely to ignore the individual characteristics of details, which further leads to the lack of automation and intelligence in the feature processing process of newborn cries, and also leads to technical problems such as poor recognition effect; the specific categories of newborn cries are relatively complex, and different categories have both similar features and highly different features. When traditional audio recognition models and methods are applied to the complex cry category recognition process of newborns, they are easily insensitive to subtle differences in cries, and it is even more difficult to effectively capture the subtle differences between different cry categories, resulting in poor accuracy in newborn cry recognition and more difficult to use and apply. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a newborn cry recognition system based on machine learning. In the existing newborn cry recognition process, the original audio of the cry is used as signal data. Under environmental conditions, it is easy to contain a large amount of redundant information and noise. When it is suitable for newborn cry recognition, the cry of newborns is quite different and the specificity of the overall audio is relatively low, which leads to increased computational complexity. The cry characteristics of different newborns vary due to individual differences. Therefore, the audio features of newborns often have to rely on manual intervention to be extracted. Traditional intelligent methods are difficult to adapt to the dynamic changes of data features and are more likely to ignore the individual characteristics of details, which further leads to the lack of automation and intelligence in the feature processing process of newborn cries, and also leads to technical problems such as poor recognition effect. This solution creatively adopts an improved honey badger optimization algorithm to perform feature selection on the cry audio feature set, automatically selects key features that help distinguish different cry categories, and avoids manual intervention errors. , and can flexibly adjust feature selection according to individual differences to ensure that the system has high recognition accuracy for different newborn cries; in view of the existing newborn cry recognition process, the specific categories of newborn cries are relatively complex, and different categories have both similar features and features with large differences. Traditional audio recognition models and methods, when applied to the complex cry category recognition process of newborns, are easily insensitive to subtle differences in cries, and it is more difficult to effectively capture the slight differences between different cry categories, resulting in poor accuracy in newborn cry recognition and more difficult to use and apply technical problems. This solution creatively uses a hybrid kernel function support vector machine model combined with a K-nearest neighbor algorithm for newborn cry recognition. By capturing the local similarity between samples, it can accurately divide the boundaries between different newborn cry categories, thereby avoiding overlap between categories, and then more accurately and efficiently identifying different types of cries, improving the overall usability of cry recognition and the level of intelligence in newborn care.

[0005] The technical solution adopted by the present invention is as follows: the newborn crying recognition system based on machine learning provided by the present invention comprises a data acquisition module, a feature extraction module, a feature selection module, a newborn crying recognition module and a recognition result feedback module;

[0006] The data acquisition module obtains a newborn crying signal through signal acquisition, obtains a newborn crying filtered signal by performing data preprocessing on the newborn crying signal, and sends the newborn crying filtered signal to the feature extraction module;

[0007] The feature extraction module extracts features from the cry filter signal of the newborn through a time domain feature extraction unit, a frequency domain feature extraction unit and an audio feature extraction unit to obtain a cry audio feature set, and sends the cry audio feature set to the feature selection module;

[0008] The feature selection module uses an improved honey badger optimization algorithm to perform feature selection on a crying audio feature set to obtain a crying audio key feature set, and sends the crying audio key feature set to a newborn crying recognition module;

[0009] The newborn cry recognition module uses a hybrid kernel function support vector machine model combined with a K-nearest neighbor algorithm to perform newborn cry recognition, obtain a newborn cry category, and send the newborn cry category to the recognition result feedback module;

[0010] The recognition result feedback module generates a newborn crying recognition report according to the newborn crying category.

[0011] Furthermore, in the data acquisition module, the signal acquisition is specifically performed by placing a recording device near the newborn to acquire the cry signal of the newborn;

[0012] The data preprocessing specifically includes segmenting the newborn crying signal into fixed lengths, smoothing each signal segment using a Hanning window function, and removing environmental noise from the signal using a bandpass filter to obtain a newborn crying filtered signal.

[0013] Furthermore, in the feature extraction module, a time domain feature extraction unit, a frequency domain feature extraction unit and an audio feature extraction unit are provided, including the following contents:

[0014] The time domain feature extraction unit extracts the time domain features by calculating the zero-crossing rate, energy value, root mean square and statistical matrix of the cry filtering signal of the newborn, and obtains the cry time domain feature set;

[0015] The zero-crossing rate is used to reflect the vibration characteristics of the signal, and the calculation formula is:

[0016] ;

[0017] Where ZCR is the zero crossing rate, N cry is the number of signal samples, specifically the number of filtered signal samples of newborn crying, n is the signal sample index, specifically the index of filtered signal samples of newborn crying, I(·) is the indicator function, when it satisfies When , the indicator function takes the value of 1, otherwise takes the value of 0, x[n] is the sample value of the n-th newborn cry filter signal, and x[n-1] is the sample value of the n-1-th newborn cry filter signal;

[0018] The energy value is used to measure the signal strength, and the calculation formula is:

[0019] ;

[0020] Where E is the energy value, specifically the energy value of the filtered signal of the newborn's cry;

[0021] The root mean square is used to reflect the stability and change trend of the signal amplitude, and the calculation formula is:

[0022] ;

[0023] In the formula, RMS is the root mean square, specifically the root mean square of the filtered signal of the newborn cry;

[0024] The statistical moments are used to reflect the statistical characteristics of the signal amplitude distribution, including the signal mean, standard deviation, skewness and peak value;

[0025] The frequency domain feature extraction unit calculates the Mel frequency cepstrum coefficient, autocorrelation value and power spectrum density of the newborn crying filter signal to extract the frequency domain features and obtain the crying frequency domain feature set;

[0026] The Mel frequency cepstrum coefficients are obtained by performing short-time Fourier transform on the filtered signal of the cry of the newborn, performing Mel frequency transform, and then performing discrete cosine transform to obtain the Mel frequency cepstrum coefficients. The calculation formula is:

[0027] ;

[0028] Where X[k] is the complex amplitude of the kth frequency point obtained by short-time Fourier transform, k is the signal frequency index, specifically the frequency index of the filtered signal of the newborn cry, is the Fourier transform basis function, j is the imaginary unit, e is the base of the natural logarithm, π is the circumference of a circle, and f mel is the Mel frequency obtained by Mel frequency transformation, log(·) is the logarithmic function, f is the signal frequency, specifically the frequency of the filtered signal of the newborn cry, MFCC c is the cth Mel frequency cepstral coefficient, specifically the cth Mel frequency cepstral coefficient of the newborn cry filter signal, c is the Mel frequency cepstral coefficient index, M is the number of Mel filters, m is the Mel filter index, E m is the energy value of the Mel band, cos(·) is the cosine function;

[0029] The autocorrelation value is used to analyze the periodicity and repeatability of the signal, and the calculation formula is:

[0030] ;

[0031] In the formula, The signal delay is The autocorrelation value at the time, specifically refers to the delay of the filtered signal of the newborn cry The autocorrelation value at It is the signal delay, specifically the delay of filtering the cry signal of a newborn baby;

[0032] The power spectrum density is used to describe the energy distribution of the signal at different frequencies;

[0033] The audio feature extraction unit specifically extracts audio features from the newborn crying filter signal through the time domain feature extraction unit and the frequency domain feature extraction unit to obtain a crying audio feature set.

[0034] Furthermore, in the feature selection module, there are an algorithm initialization unit, a fitness function construction unit, a prey odor intensity calculation unit, a honey badger behavior simulation unit, a density factor optimization unit, a prey position update unit and a honey badger position iterative update unit, including the following contents:

[0035] Algorithm initialization unit, which uses the good point set method to initialize the position of the honey badger in the search space and sets the maximum number of iterations and the initial prey position;

[0036] The calculation formula of the honey badger position is:

[0037] ;

[0038] In the formula, po i is the position of the i-th honey badger, which is used to represent the crying audio feature subset, po i1 is the position component of the ith honey badger in the first dimension, which is used to represent the selection state of the first cry audio feature of the ith honey badger. i2 is the position component of the ith honey badger in the second dimension, po iD is the position component of the ith honey badger in the Dth dimension, where D is the search space dimension, which is equal to the total number of crying audio features;

[0039] A fitness function construction unit constructs the fitness function according to the model error rate;

[0040] The fitness function is used to evaluate the quality of the honey badger's position, and the calculation formula is:

[0041] ;

[0042] Where F(·) is the fitness function, is the number of features selected for the i-th honey badger position, Error(po i ) is the model error rate, specifically the error rate of the hybrid kernel function support vector machine model combined with the K nearest neighbor algorithm, is the penalty coefficient, which is used to control the number of feature selections;

[0043] The prey odor intensity calculation unit is used to calculate the prey odor intensity. The calculation formula is:

[0044] ;

[0045] In the formula, St i is the prey odor intensity of the i-th honey badger, r1 is the first random number in the range [0,1], d i is the distance between the ith honey badger and its prey, S is the square distance between the ith honey badger and the i+1th honey badger, po i+1 is the position of the i+1th honey badger, po prey It is the prey position;

[0046] A honey badger behavior simulation unit is used to update the position of the honey badger, specifically by letting the honey badger randomly perform digging behavior or honey collecting behavior to update the position of the honey badger;

[0047] The calculation formula of the mining behavior is:

[0048] ;

[0049] ;

[0050] In the formula, is the updated position of the honey badger, U is a symbolic variable used to change the search direction, is a constant factor used to control the search range, r2 is the second random number in the range [0,1], is the density factor at the u-th iteration, r3 is the third random number in the range [0,1], r4 is the fourth random number in the range [0,1], and r5 is the fifth random number in the range [0,1];

[0051] The calculation formula of the honey collecting behavior is:

[0052] ;

[0053] Wherein, r6 is the sixth random number in the range of [0,1];

[0054] The density factor optimization unit uses chaotic mapping to update the density factor. The calculation formula is:

[0055] ;

[0056] In the formula, is the density factor at the u+1th iteration, u is the iteration index, is the bifurcation parameter;

[0057] The prey position updating unit specifically calculates the fitness value of each honey badger position through the fitness function, and selects the honey badger position with the best fitness value as the prey position;

[0058] The honey badger position iterative updating unit performs iterative updating of the honey badger position by repeatedly executing the prey odor intensity calculation unit, the honey badger behavior simulation unit, the density factor optimization unit and the prey position updating unit until a maximum number of iterations is reached, thereby obtaining the prey position as a key feature set of the crying audio.

[0059] Furthermore, in the newborn cry recognition module, a K nearest neighbor algorithm construction unit, a hybrid kernel function support vector machine construction unit, a model training unit and a recognition result generation unit are provided, including the following contents:

[0060] K-nearest neighbor algorithm building unit, used for local domain learning, specifically for local similarity measurement through K-nearest neighbor algorithm;

[0061] The calculation formula of the local similarity measure is:

[0062] ;

[0063] Where dis(·) is the local similarity metric function, specifically the Euclidean distance calculation function, fea a is the ath crying audio key feature sample, a is the first index of the crying audio key feature sample, fea b is the bth crying audio key feature sample, b is the second index of the crying audio key feature sample, the first index of the crying audio key feature sample is not equal to the second index of the crying audio key feature sample, N fea is the sample dimension of the key feature of the crying audio, z is the index of the sample dimension of the key feature of the crying audio, fea az is the value of the zth dimension of the ath crying audio key feature sample, fea bz is the value of the zth dimension of the bth crying audio key feature sample;

[0064] A hybrid kernel function support vector machine construction unit is used to model complex nonlinear relationships and classify them. Specifically, a Gaussian kernel function and a polynomial kernel function are selected as basic kernel functions, a hybrid kernel function is designed based on the basic kernel function, and a hybrid kernel function support vector machine is constructed.

[0065] The calculation formula of the Gaussian kernel function is:

[0066] ;

[0067] Where KF RBF (·) is the Gaussian kernel function, exp(·) is the exponential function, It is a hyperparameter of the Gaussian kernel function, which is used to control the width of the Gaussian kernel function;

[0068] The calculation formula of the polynomial kernel function is:

[0069] ;

[0070] Where KF Poly (·) is the polynomial kernel function, g is the constant term of the polynomial kernel function, and q is the polynomial order, which is used to control the complexity of high-dimensional feature interactions;

[0071] The calculation formula of the hybrid kernel function is:

[0072] ;

[0073] Where KF mix (·) is the mixing kernel function, is the Gaussian kernel function weight, is the weight of the polynomial kernel function;

[0074] The model training unit is specifically configured to construct a hybrid kernel function support vector machine model combined with the K nearest neighbor algorithm based on the key feature set of the crying audio, through the K nearest neighbor algorithm construction unit and the hybrid kernel function support vector machine construction unit, and perform model training to obtain a newborn crying recognition model;

[0075] The recognition result generating unit specifically uses the newborn cry recognition model to perform newborn cry recognition to obtain the newborn cry category.

[0076] Furthermore, in the recognition result feedback module, a newborn cry recognition report is generated according to the newborn cry category and fed back to the nursing staff in real time. The newborn cry recognition report includes newborn information, newborn cry category and nursing suggestions.

[0077] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0078] (1) In the existing process of newborn cry recognition, the original audio of the cry is used as signal data. Under environmental conditions, it is easy to contain a large amount of redundant information and noise. When it is applied to newborn cry recognition, the difference between newborn cries is large and the specificity of the overall audio is small, which increases the computational complexity. The cry characteristics of different newborns vary due to individual differences. Therefore, the audio features of newborns often have to rely on manual intervention to be extracted. Traditional intelligent methods are difficult to adapt to the dynamic changes of data features and are more likely to ignore the individual characteristics of details, which further leads to the lack of automation and intelligence in the feature processing process of newborn cries. At the same time, it also leads to technical problems such as poor recognition effect. This scheme creatively adopts the improved honey badger optimization algorithm to select the crying audio feature set, automatically selects the key features that help distinguish different crying categories, avoids manual intervention errors, and can better flexibly adjust the feature selection according to individual differences, ensuring that the system has high recognition accuracy for different newborn cries;

[0079] (2) In the existing process of newborn cry recognition, the specific categories of newborn cries are relatively complex, and different categories have both similar features and quite different features. When traditional audio recognition models and methods are applied to the complex cry category recognition process of newborns, they are easily insensitive to subtle differences in cries, and it is even more difficult to effectively capture the slight differences between different cry categories, resulting in poor accuracy in newborn cry recognition and more difficult to use and apply technical problems. This solution creatively adopts a hybrid kernel function support vector machine model combined with a K-nearest neighbor algorithm for newborn cry recognition. By capturing the local similarity between samples, it can accurately divide the boundaries between different newborn cry categories, thereby avoiding overlap between categories, and then more accurately and efficiently identifying different types of cries, improving the overall usability of cry recognition and the level of intelligence in newborn care. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 A schematic diagram of the framework of a newborn cry recognition system based on machine learning provided by the present invention;

[0081] Figure 2 It is a flowchart of the feature extraction module;

[0082] Figure 3 It is a flowchart of the feature selection module;

[0083] Figure 4 This is a flow chart of the newborn cry recognition module.

[0084] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0085] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0086] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0087] Example 1, see Figure 1 The newborn crying recognition system based on machine learning provided by the present invention comprises a data acquisition module, a feature extraction module, a feature selection module, a newborn crying recognition module and a recognition result feedback module;

[0088] The data acquisition module obtains a newborn crying signal through signal acquisition, obtains a newborn crying filtered signal by performing data preprocessing on the newborn crying signal, and sends the newborn crying filtered signal to the feature extraction module;

[0089] The feature extraction module extracts features from the cry filter signal of the newborn through a time domain feature extraction unit, a frequency domain feature extraction unit and an audio feature extraction unit to obtain a cry audio feature set, and sends the cry audio feature set to the feature selection module;

[0090] The feature selection module uses an improved honey badger optimization algorithm to perform feature selection on a crying audio feature set to obtain a crying audio key feature set, and sends the crying audio key feature set to a newborn crying recognition module;

[0091] The newborn cry recognition module uses a hybrid kernel function support vector machine model combined with a K-nearest neighbor algorithm to perform newborn cry recognition, obtain a newborn cry category, and send the newborn cry category to the recognition result feedback module;

[0092] The recognition result feedback module generates a newborn crying recognition report according to the newborn crying category.

[0093] Example 2, see Figure 1This embodiment is based on the above embodiment. In the data acquisition module, the signal acquisition is specifically performed by placing a recording device within 30 cm of the newborn and using a 44.1 KHz sampling frequency to acquire the signal to obtain the cry signal of the newborn;

[0094] The data preprocessing specifically includes segmenting the cry signal of the newborn baby with a fixed length of 50 milliseconds, smoothing each signal segment by a Hanning window function, removing the environmental noise in the signal by a bandpass filter, and obtaining a filtered cry signal of the newborn baby.

[0095] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the feature extraction module, a time domain feature extraction unit, a frequency domain feature extraction unit and an audio feature extraction unit are provided, and includes the following contents:

[0096] The time domain feature extraction unit extracts the time domain features by calculating the zero-crossing rate, energy value, root mean square and statistical matrix of the cry filtering signal of the newborn, and obtains the cry time domain feature set;

[0097] The zero-crossing rate is used to reflect the vibration characteristics of the signal, and the calculation formula is:

[0098] ;

[0099] Where ZCR is the zero crossing rate, N cry is the number of signal samples, specifically the number of filtered signal samples of newborn crying, n is the signal sample index, specifically the index of filtered signal samples of newborn crying, I(·) is the indicator function, when it satisfies When , the indicator function takes the value of 1, otherwise takes the value of 0, x[n] is the sample value of the n-th newborn cry filter signal, and x[n-1] is the sample value of the n-1-th newborn cry filter signal;

[0100] The energy value is used to measure the signal strength, and the calculation formula is:

[0101] ;

[0102] Where E is the energy value, specifically the energy value of the filtered signal of the newborn's cry;

[0103] The root mean square is used to reflect the stability and change trend of the signal amplitude, and the calculation formula is:

[0104] ;

[0105] In the formula, RMS is the root mean square, specifically the root mean square of the filtered signal of the newborn cry;

[0106] The statistical moments are used to reflect the statistical characteristics of the signal amplitude distribution, including the signal mean, standard deviation, skewness and peak value;

[0107] The frequency domain feature extraction unit calculates the Mel frequency cepstrum coefficient, autocorrelation value and power spectrum density of the newborn crying filter signal to extract the frequency domain features and obtain the crying frequency domain feature set;

[0108] The Mel frequency cepstrum coefficients are obtained by performing short-time Fourier transform on the filtered signal of the cry of the newborn, performing Mel frequency transform, and then performing discrete cosine transform to obtain the Mel frequency cepstrum coefficients. The calculation formula is:

[0109] ;

[0110] Where X[k] is the complex amplitude of the kth frequency point obtained by short-time Fourier transform, k is the signal frequency index, specifically the frequency index of the filtered signal of the newborn cry, is the Fourier transform basis function, j is the imaginary unit, e is the base of the natural logarithm, π is the circumference of a circle, and f mel is the Mel frequency obtained by Mel frequency transformation, log(·) is the logarithmic function, f is the signal frequency, specifically the frequency of the filtered signal of the newborn cry, MFCC c is the cth Mel frequency cepstral coefficient, specifically the cth Mel frequency cepstral coefficient of the newborn cry filter signal, c is the Mel frequency cepstral coefficient index, M is the number of Mel filters, m is the Mel filter index, E m is the energy value of the Mel band, cos(·) is the cosine function;

[0111] The autocorrelation value is used to analyze the periodicity and repeatability of the signal, and the calculation formula is:

[0112] ;

[0113] In the formula, The signal delay is The autocorrelation value at the time, specifically refers to the delay of the filtered signal of the newborn cry The autocorrelation value at It is the signal delay, specifically the delay of filtering the cry signal of a newborn baby;

[0114] The power spectrum density is used to describe the energy distribution of the signal at different frequencies;

[0115] The audio feature extraction unit specifically extracts audio features from the newborn crying filter signal through the time domain feature extraction unit and the frequency domain feature extraction unit to obtain a crying audio feature set.

[0116] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In the feature selection module, an algorithm initialization unit, a fitness function construction unit, a prey odor intensity calculation unit, a honey badger behavior simulation unit, a density factor optimization unit, a prey position update unit and a honey badger position iteration update unit are provided, including the following contents:

[0117] Algorithm initialization unit, which uses the good point set method to initialize the position of the honey badger in the search space and sets the maximum number of iterations and the initial prey position;

[0118] The calculation formula of the honey badger position is:

[0119] ;

[0120] In the formula, po i is the position of the i-th honey badger, which is used to represent the crying audio feature subset, po i1 is the position component of the ith honey badger in the first dimension, which is used to represent the selection state of the first cry audio feature of the ith honey badger. i1 When the value is 0, it means selecting the first cry audio feature of the i-th honey badger. i1 When the value is 1, it means that the first cry audio feature of the i-th honey badger is not selected. i2 is the position component of the ith honey badger in the second dimension, po iD is the position component of the ith honey badger in the Dth dimension, where D is the search space dimension, which is equal to the total number of crying audio features;

[0121] A fitness function construction unit constructs the fitness function according to the model error rate;

[0122] The fitness function is used to evaluate the quality of the honey badger's position, and the calculation formula is:

[0123] ;

[0124] Where F(·) is the fitness function, is the number of features selected for the i-th honey badger position, Error(po i ) is the model error rate, specifically the error rate of the hybrid kernel function support vector machine model combined with the K nearest neighbor algorithm, is the penalty coefficient, which is used to control the number of feature selections;

[0125] The prey odor intensity calculation unit is used to calculate the prey odor intensity. The calculation formula is:

[0126] ;

[0127] In the formula, Sti is the prey odor intensity of the i-th honey badger, r1 is the first random number in the range [0,1], d i is the distance between the ith honey badger and its prey, S is the square distance between the ith honey badger and the i+1th honey badger, po i+1 is the position of the i+1th honey badger, po prey It is the prey position;

[0128] A honey badger behavior simulation unit is used to update the position of the honey badger, specifically by letting the honey badger randomly perform digging behavior or honey collecting behavior to update the position of the honey badger;

[0129] The calculation formula of the mining behavior is:

[0130] ;

[0131] ;

[0132] In the formula, is the updated position of the honey badger, U is a symbolic variable used to change the search direction, is a constant factor used to control the search range, r2 is the second random number in the range [0,1], is the density factor at the u-th iteration, the initial value of the density factor is a random number in the range of (0,1), r3 is a third random number in the range of [0,1], r4 is a fourth random number in the range of [0,1], and r5 is a fifth random number in the range of [0,1];

[0133] The calculation formula of the honey collecting behavior is:

[0134] ;

[0135] Wherein, r6 is the sixth random number in the range of [0,1];

[0136] The density factor optimization unit uses chaotic mapping to update the density factor. The calculation formula is:

[0137] ;

[0138] In the formula, is the density factor at the u+1th iteration, u is the iteration index, is a bifurcation parameter, and the value of the bifurcation parameter is 3.99;

[0139] The prey position updating unit specifically calculates the fitness value of each honey badger position through the fitness function, and selects the honey badger position with the best fitness value as the prey position;

[0140] A honey badger position iterative updating unit, which iteratively updates the honey badger position by repeatedly executing the prey odor intensity calculation unit, the honey badger behavior simulation unit, the density factor optimization unit and the prey position updating unit until a maximum number of iterations is reached, and obtains the prey position as a key feature set of the crying audio;

[0141] By performing the above operations, in the existing newborn cry recognition process, the original audio of the cry exists as signal data. Under environmental conditions, it is easy to contain a large amount of redundant information and noise. When it is suitable for newborn cry recognition, the cry of newborns is quite different and the specificity of the overall audio is relatively small, which leads to increased computational complexity. The cry characteristics of different newborns vary due to individual differences. Therefore, the audio features of newborns often have to rely on manual intervention to be extracted. Traditional intelligent methods are difficult to adapt to the dynamic changes of data features and are more likely to ignore the individual characteristics of details, which further leads to the lack of automation and intelligence in the feature processing process of newborn cries, and also leads to technical problems such as poor recognition effect. This scheme creatively adopts the improved honey badger optimization algorithm to perform feature selection on the crying audio feature set, automatically selects key features that help distinguish different crying categories, avoids manual intervention errors, and can better flexibly adjust feature selection according to individual differences, ensuring that the system has high recognition accuracy for different newborn cries.

[0142] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In the newborn cry recognition module, a K nearest neighbor algorithm construction unit, a hybrid kernel function support vector machine construction unit, a model training unit and a recognition result generation unit are provided, including the following contents:

[0143] K-nearest neighbor algorithm building unit, used for local domain learning, specifically for local similarity measurement through K-nearest neighbor algorithm;

[0144] The calculation formula of the local similarity measure is:

[0145] ;

[0146] Where dis(·) is the local similarity metric function, specifically the Euclidean distance calculation function, fea a is the ath crying audio key feature sample, a is the first index of the crying audio key feature sample, fea b is the bth crying audio key feature sample, b is the second index of the crying audio key feature sample, the first index of the crying audio key feature sample is not equal to the second index of the crying audio key feature sample, N fea is the sample dimension of the key feature of the crying audio, z is the index of the sample dimension of the key feature of the crying audio, feaaz is the value of the zth dimension of the ath crying audio key feature sample, fea bz is the value of the zth dimension of the bth crying audio key feature sample;

[0147] A hybrid kernel function support vector machine construction unit is used to model complex nonlinear relationships and classify them. Specifically, a Gaussian kernel function and a polynomial kernel function are selected as basic kernel functions, a hybrid kernel function is designed based on the basic kernel function, and a hybrid kernel function support vector machine is constructed.

[0148] The calculation formula of the Gaussian kernel function is:

[0149] ;

[0150] Where KF RBF (·) is the Gaussian kernel function, exp(·) is the exponential function, It is a hyperparameter of the Gaussian kernel function, which is used to control the width of the Gaussian kernel function;

[0151] The calculation formula of the polynomial kernel function is:

[0152] ;

[0153] Where KF Poly (·) is the polynomial kernel function, g is the constant term of the polynomial kernel function, and q is the polynomial order, which is used to control the complexity of high-dimensional feature interactions;

[0154] The calculation formula of the hybrid kernel function is:

[0155] ;

[0156] Where KF mix (·) is the mixing kernel function, is the Gaussian kernel function weight, is the weight of the polynomial kernel function;

[0157] The model training unit is specifically configured to construct a hybrid kernel function support vector machine model combined with the K nearest neighbor algorithm based on the key feature set of the crying audio, through the K nearest neighbor algorithm construction unit and the hybrid kernel function support vector machine construction unit, and perform model training to obtain a newborn crying recognition model;

[0158] The recognition result generating unit specifically uses the newborn cry recognition model to perform newborn cry recognition to obtain newborn cry categories, wherein the newborn cry categories include hunger, sleepiness, pain, discomfort, anxiety and normal;

[0159] By performing the above operations, in view of the fact that in the existing newborn cry recognition process, the specific categories of newborn cries are relatively complex, and different categories have both similar features and features with large differences. Traditional audio recognition models and methods, when applied to the complex cry category recognition process of newborns, are easily insensitive to subtle differences in cries, and it is even more difficult to effectively capture the slight differences between different cry categories, resulting in poor accuracy in newborn cry recognition and more difficult to use and apply technical problems. This solution creatively adopts a hybrid kernel function support vector machine model combined with a K-nearest neighbor algorithm for newborn cry recognition. By capturing the local similarity between samples, it can accurately divide the boundaries between different newborn cry categories, thereby avoiding overlap between categories, and then more accurately and efficiently identifying different types of cries, improving the overall usability of cry recognition and the level of intelligence in newborn care.

[0160] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the recognition result feedback module, a newborn cry recognition report is generated according to the newborn cry category and fed back to the nursing staff in real time. The newborn cry recognition report includes newborn information, newborn cry category and nursing suggestions.

[0161] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0162] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0163] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A newborn cry recognition system based on machine learning, characterized by: It includes a data acquisition module, a feature extraction module, a feature selection module, a newborn crying recognition module and a recognition result feedback module; The data acquisition module obtains a newborn crying signal through signal acquisition, obtains a newborn crying filtered signal by performing data preprocessing on the newborn crying signal, and sends the newborn crying filtered signal to the feature extraction module; The feature extraction module extracts features from the cry filter signal of the newborn through a time domain feature extraction unit, a frequency domain feature extraction unit and an audio feature extraction unit to obtain a cry audio feature set, and sends the cry audio feature set to the feature selection module; The feature selection module uses an improved honey badger optimization algorithm to perform feature selection on a crying audio feature set to obtain a crying audio key feature set, and sends the crying audio key feature set to a newborn crying recognition module; The newborn cry recognition module uses a hybrid kernel function support vector machine model combined with a K-nearest neighbor algorithm to perform newborn cry recognition, obtain a newborn cry category, and send the newborn cry category to the recognition result feedback module; The recognition result feedback module generates a newborn crying recognition report according to the newborn crying category.

2. The newborn cry recognition system based on machine learning according to claim 1, characterized in that: In the feature selection module, there are algorithm initialization unit, fitness function construction unit, prey odor intensity calculation unit, honey badger behavior simulation unit, density factor optimization unit, prey position update unit and honey badger position iterative update unit, including the following contents: Algorithm initialization unit, which uses the good point set method to initialize the position of the honey badger in the search space and sets the maximum number of iterations and the initial prey position; The calculation formula of the honey badger position is: ； In the formula, po i is the position of the i-th honey badger, which is used to represent the crying audio feature subset, po i1 is the position component of the ith honey badger in the first dimension, which is used to represent the selection state of the first cry audio feature of the ith honey badger. i2 is the position component of the ith honey badger in the second dimension, po iD is the position component of the ith honey badger in the Dth dimension, where D is the search space dimension, which is equal to the total number of crying audio features; A fitness function construction unit constructs the fitness function according to the model error rate; The fitness function is used to evaluate the quality of the honey badger's position, and the calculation formula is: ； Where F(·) is the fitness function, is the number of features selected for the i-th honey badger position, Error(po i ) is the model error rate, specifically the error rate of the hybrid kernel function support vector machine model combined with the K nearest neighbor algorithm, is the penalty coefficient, which is used to control the number of feature selections; The prey odor intensity calculation unit is used to calculate the prey odor intensity. The calculation formula is: ； In the formula, St i is the prey odor intensity of the i-th honey badger, r1 is the first random number in the range [0,1], d i is the distance between the ith honey badger and its prey, S is the square distance between the ith honey badger and the i+1th honey badger, po i+1 is the position of the i+1th honey badger, po prey It is the prey position; A honey badger behavior simulation unit is used to update the position of the honey badger, specifically by letting the honey badger randomly perform digging behavior or honey collecting behavior to update the position of the honey badger; The density factor optimization unit uses chaotic mapping to update the density factor. The calculation formula is: ； In the formula, is the density factor at the u+1th iteration, u is the iteration index, is the bifurcation parameter, is the density factor at the u-th iteration, and the initial value of the density factor is a random number in the range of (0,1); The prey position updating unit specifically calculates the fitness value of each honey badger position through the fitness function, and selects the honey badger position with the best fitness value as the prey position; The honey badger position iterative updating unit performs iterative updating of the honey badger position by repeatedly executing the prey odor intensity calculation unit, the honey badger behavior simulation unit, the density factor optimization unit and the prey position updating unit until a maximum number of iterations is reached, thereby obtaining the prey position as a key feature set of the crying audio.

3. The newborn cry recognition system based on machine learning according to claim 2, characterized in that: In the honey badger behavior simulation unit, the calculation formula for the digging behavior is: ； ； In the formula, is the updated position of the honey badger, U is a symbolic variable used to change the search direction, is a constant factor used to control the search range, r2 is the second random number in the range [0,1], r3 is the third random number in the range [0,1], r4 is the fourth random number in the range [0,1], and r5 is the fifth random number in the range [0,1]; The calculation formula of the honey collecting behavior is: ； Wherein, r6 is the sixth random number in the range of [0,1].

4. The newborn cry recognition system based on machine learning according to claim 3, characterized in that: In the newborn cry recognition module, there are a K nearest neighbor algorithm construction unit, a hybrid kernel function support vector machine construction unit, a model training unit and a recognition result generation unit, including the following contents: K-nearest neighbor algorithm building unit, used for local domain learning, specifically for local similarity measurement through K-nearest neighbor algorithm; The calculation formula of the local similarity measure is: ； Where dis(·) is the local similarity metric function, specifically the Euclidean distance calculation function, fea a is the ath crying audio key feature sample, a is the first index of the crying audio key feature sample, fea b is the bth crying audio key feature sample, b is the second index of the crying audio key feature sample, the first index of the crying audio key feature sample is not equal to the second index of the crying audio key feature sample, N fea is the sample dimension of the key feature of the crying audio, z is the index of the sample dimension of the key feature of the crying audio, fea az is the value of the zth dimension of the ath crying audio key feature sample, fea bz is the value of the zth dimension of the bth crying audio key feature sample; A hybrid kernel function support vector machine construction unit is used to model complex nonlinear relationships and classify them. Specifically, a Gaussian kernel function and a polynomial kernel function are selected as basic kernel functions, a hybrid kernel function is designed based on the basic kernel function, and a hybrid kernel function support vector machine is constructed. The calculation formula of the Gaussian kernel function is: ； Where KF RBF (·) is the Gaussian kernel function, exp(·) is the exponential function, It is a hyperparameter of the Gaussian kernel function, which is used to control the width of the Gaussian kernel function; The calculation formula of the polynomial kernel function is: ； Where KF Poly (·) is the polynomial kernel function, g is the constant term of the polynomial kernel function, and q is the polynomial order, which is used to control the complexity of high-dimensional feature interactions; The calculation formula of the hybrid kernel function is: ； Where KF mix (·) is the mixing kernel function, is the Gaussian kernel function weight, is the weight of the polynomial kernel function; The model training unit is specifically configured to construct a hybrid kernel function support vector machine model combined with the K nearest neighbor algorithm based on the key feature set of the crying audio, through the K nearest neighbor algorithm construction unit and the hybrid kernel function support vector machine construction unit, and perform model training to obtain a newborn crying recognition model; The recognition result generating unit specifically uses the newborn cry recognition model to perform newborn cry recognition to obtain the newborn cry category.

5. The newborn cry recognition system based on machine learning according to claim 4, characterized in that: In the feature extraction module, there are a time domain feature extraction unit, a frequency domain feature extraction unit and an audio feature extraction unit, including the following contents: The time domain feature extraction unit extracts the time domain features by calculating the zero-crossing rate, energy value, root mean square and statistical matrix of the cry filtering signal of the newborn, and obtains the cry time domain feature set; The frequency domain feature extraction unit calculates the Mel frequency cepstrum coefficient, autocorrelation value and power spectrum density of the newborn crying filter signal to extract the frequency domain features and obtain the crying frequency domain feature set; The audio feature extraction unit specifically extracts audio features from the newborn crying filter signal through the time domain feature extraction unit and the frequency domain feature extraction unit to obtain a crying audio feature set.

6. The newborn cry recognition system based on machine learning according to claim 5, characterized in that: In the time domain feature extraction unit, the zero-crossing rate is used to reflect the signal vibration characteristics, and the calculation formula is: ； Where ZCR is the zero crossing rate, N cry is the number of signal samples, specifically the number of filtered signal samples of newborn crying, n is the signal sample index, specifically the index of filtered signal samples of newborn crying, I(·) is the indicator function, when it satisfies When , the indicator function takes the value of 1, otherwise takes the value of 0, x[n] is the sample value of the n-th newborn cry filter signal, and x[n-1] is the sample value of the n-1-th newborn cry filter signal; The energy value is used to measure the signal strength, and the calculation formula is: ； Where E is the energy value, specifically the energy value of the filtered signal of the newborn's cry; The root mean square is used to reflect the stability and change trend of the signal amplitude, and the calculation formula is: ； In the formula, RMS is the root mean square, specifically the root mean square of the filtered signal of the newborn cry; The statistical moments are used to reflect the statistical characteristics of the signal amplitude distribution, including the signal mean, standard deviation, skewness and peak value; In the frequency domain feature extraction unit, the Mel frequency cepstrum coefficient is obtained by performing short-time Fourier transform on the newborn crying filter signal, performing Mel frequency transform, and then performing discrete cosine transform to obtain the Mel frequency cepstrum coefficient. The calculation formula is: ； Where X[k] is the complex amplitude of the kth frequency point obtained by short-time Fourier transform, k is the signal frequency index, specifically the frequency index of the filtered signal of the newborn cry, is the Fourier transform basis function, j is the imaginary unit, e is the base of the natural logarithm, π is the circumference of a circle, and f mel is the Mel frequency obtained by Mel frequency transformation, log(·) is the logarithmic function, f is the signal frequency, specifically the frequency of the filtered signal of the newborn cry, MFCC c is the cth Mel frequency cepstral coefficient, specifically the cth Mel frequency cepstral coefficient of the newborn cry filter signal, c is the Mel frequency cepstral coefficient index, M is the number of Mel filters, m is the Mel filter index, E m is the energy value of the Mel band, cos(·) is the cosine function; The autocorrelation value is used to analyze the periodicity and repeatability of the signal, and the calculation formula is: ； In the formula, The signal delay is The autocorrelation value at the time, specifically refers to the delay of the filtered signal of the newborn cry The autocorrelation value at It is the signal delay, specifically the delay of filtering the cry signal of a newborn baby; The power spectrum density is used to describe the energy distribution of a signal at different frequencies.

7. The newborn cry recognition system based on machine learning according to claim 6, characterized in that: In the data acquisition module, the signal acquisition is specifically performed by placing a recording device near the newborn to acquire the cry signal of the newborn; The data preprocessing specifically includes segmenting the newborn crying signal into fixed lengths, smoothing each signal segment using a Hanning window function, and removing environmental noise from the signal using a bandpass filter to obtain a newborn crying filtered signal.

8. The newborn cry recognition system based on machine learning according to claim 8, characterized in that: In the recognition result feedback module, a newborn cry recognition report is generated according to the newborn cry category and fed back to the nursing staff in real time. The newborn cry recognition report includes newborn information, newborn cry category and nursing suggestions.

Images