Medical percussion sound intelligent classification system based on voiceprint recognition technology

Abstract

The invention relates to a medical percussion sound intelligent classification system based on a voiceprint recognition technology, and belongs to the technical field of medical treatment. Comprising a preprocessing module, a feature extraction module and an integrated classification module, noise and interference in percussion sound signals are effectively removed through combination of a CEEMDAN algorithm and a DTW algorithm, and the purity of the signals is improved. The advantages of manual and automatic dual feature extraction are fused, the features of the percussion sound signals are comprehensively captured, and the diagnosis accuracy is improved. And an integrated learning strategy is adopted, so that the classification accuracy and the system robustness are remarkably improved, and the system can still keep stable performance in a complex noise environment.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and mainly relates to an intelligent classification system for medical percussion sounds based on voiceprint recognition technology. Background Art

[0002] In the medical field, traditional percussion diagnosis methods rely on doctors' experience and subjective judgment. Especially when dealing with critical diseases such as tension pneumothorax, it is difficult to make timely and accurate diagnoses. Traditional percussion methods are highly subjective, and doctors need to accumulate experience over a long time, making it difficult to standardize operations in emergency scenarios. In addition, existing intelligent percussion systems mostly rely on single signal processing technologies, and there is room for improvement in the diagnostic accuracy in complex noise environments.

[0003] Existing intelligent percussion systems mostly use single signal processing technologies to extract features and classify percussion sound signals. These methods have improved the objectivity of diagnosis to a certain extent, but in complex noise environments, their accuracy and robustness are still challenged.

[0004] Main disadvantages of existing technical solutions:

[0005] 1. In complex noise environments, the robustness of existing systems is insufficient and they are easily interfered with.

[0006] 2. Single signal processing technologies are difficult to comprehensively capture the features of percussion sound signals, resulting in limited diagnostic accuracy.

[0007] 3. The system lacks an integrated learning strategy and it is difficult to further improve the classification accuracy and system stability. Summary of the Invention

[0008] The purpose of the present invention is to solve the problems of insufficient diagnostic accuracy and weak robustness of existing intelligent percussion systems in complex noise environments, and to provide an intelligent classification system for medical percussion sounds based on voiceprint recognition technology. By introducing an integrated learning strategy and multiple signal processing technologies, the ability to extract features of percussion sound signals and classification accuracy are improved, providing an efficient and accurate intelligent percussion solution for primary medical institutions.

[0009] To achieve the above purpose, the present invention provides the following solution:

[0010] An intelligent classification system for medical percussion sounds based on voiceprint recognition technology

[0011] It includes a preprocessing module, a feature extraction module, and an integrated classification module. Among them, the preprocessing module uses the CEEMDAN algorithm and the DTW algorithm to denoise and filter the original percussion sound signal, and realizes automatic audio clipping through volume threshold detection; the feature extraction module combines the advantages of manual extraction and automatic extraction based on wavelet transform and convolutional recurrent neural network, extracts the time-domain, frequency-domain, and time-frequency domain features of the percussion sound signal and fuses them; the integrated classification module uses multiple base classifiers to perform integrated learning through the hard voting mechanism of the Voting Classifier and outputs the optimal category.

[0012] Optionally, the preprocessing module uses the CEEMDAN algorithm to decompose the original percussion sound signal into multiple intrinsic mode functions (IMFs) and residuals; uses the DTW algorithm to compare the feature changes between different IMF components, and selects the IMF component most similar to the original signal for reconstruction to achieve denoising and filtering.

[0013] Optionally, the features manually extracted in the feature extraction module include the time-domain features of the audio waveform diagram:

[0014] Interquartile Range (IQR), Signal-to-Noise Ratio (SNR), the maximum amplitude of the audio

[0015] Max;

[0016] Frequency-domain features: Perform Fourier transform (usually fast Fourier transform FFT) on the audio signal to convert the time-domain signal into a frequency-domain signal, and then calculate based on the frequency-domain data: spectral slope, spectral peak factor, spectral roll-off point.

[0017] Optionally, the automatic extraction based on wavelet transform and convolutional recurrent neural network in the feature extraction module is to generate its time-scale energy spectrum by passing the audio signal through continuous wavelet transform (CWT):

[0018] Continuous wavelet transform:

[0019]

[0020] Among them:

[0021] x(y) is the audio signal;

[0022] a is the scale parameter, reflecting the frequency resolution;

[0023] b is the time translation parameter;

[0024] ψ(t) is the mother wavelet function, which determines the analysis characteristics. Selecting the "cmor" complex wavelet is suitable for frequency-time resolution;

[0025] ψ * is the complex conjugate of the mother wavelet, ensuring the energy integrity of the signal; W x

[0026] (a, b) are the wavelet coefficients.

[0027] Energy spectrum calculation:

[0028] S(a, b) = |W x (a, b)| 2

[0029] where:

[0030] S(a, b) is the energy spectrum, reflecting the energy distribution of the signal in the time-scale space;

[0031] |W x (a, b)| 2 is the squared modulus of the wavelet coefficient, representing the energy intensity of the signal.

[0032] Normalization and scaling:

[0033] Normalize the energy spectrum S(a, b) to the interval [0, 1] to improve training stability;

[0034] Through interpolation operations, adjust S′(a, b) to a fixed size (H, W):

[0035] S′(a, b) = zoom(S(a, b), (H / S a , W / S b ))

[0036] where:

[0037] S a , S b are the height and width of the original spectrum;

[0038] H, W are the target height and width.

[0039] Optionally, the feature extraction module further includes a classification model based on CRNN, and the architecture includes the following parts: convolutional module (CNN Layers), activation function (ReLU), pooling operation (Pooling), fully connected module (Fully Connected Layer), recurrent module (GRU Layers), classification layer (Classification Layer).

[0040] Optionally, it further includes calculating the error between the model prediction and the true label using the cross-entropy loss function:

[0041]

[0042] Wherein:

[0043] is the average loss value; N is the number of samples; C is the number of classes;

[0044] y ij is the true label (1 or 0) of the i-th sample belonging to the j-th class;

[0045] is the predicted probability that the i-th sample belongs to the j-th class.

[0046] Optionally, it further includes an Adam optimizer:

[0047] Parameter update rule:

[0048]

[0049] Wherein:

[0050] θ t is the parameter after the t-th update;

[0051] η is the learning rate;

[0052] m t is the first-order momentum, the exponentially weighted average of the gradient;

[0053] v t is the second-order momentum, the exponentially weighted average of the squared gradient;

[0054] is a small value to prevent the denominator from being zero.

[0055] Optionally, it further includes feature fusion: the artificial features and the dimension-reduced GRU features are normalized and concatenated to form a composite feature vector for each sample, and the fused feature vector is input into the ensemble classifier.

[0056] Optionally, the ensemble classification module adopts multiple classifiers such as random forest, logistic regression, support vector machine, XGBoost, K-nearest neighbor algorithm, naive Bayes, K-nearest neighbor algorithm, naive Bayes, etc. The multiple base classifiers apply typical machine learning methods to obtain the predicted classification effects of each classifier on the audio.

[0057] Optionally, the ensemble classification module adopts multiple base classifiers to perform ensemble learning through the hard voting mechanism of Voting Classifier and outputs the optimal class.

[0058] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention:

[0059] Through the combination of the CEEMDAN algorithm and the DTW algorithm, the present invention effectively removes the noise and interference in the percussion sound signal and improves the purity of the signal.

[0060] The present invention integrates the advantages of manual and automatic feature extraction, comprehensively captures the features of the percussion sound signal, and improves the accuracy of diagnosis.

[0061] The present invention adopts an ensemble learning strategy, significantly improves the accuracy of classification and the robustness of the system, and enables the system to maintain stable performance in a complex noise environment. Description of the Drawings

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0063] Figure 1 Chart of the original signal and IMFs in Embodiment 1 of the present invention;

[0064] Figure 2 Audio waveform diagram in Embodiment 1 of the present invention;

[0065] Figure 3 Frequency domain diagram of the extracted audio in Embodiment 1 of the present invention.

[0066] Figure 4 Functional module composition diagram of a medical percussion sound intelligent classification system based on voiceprint recognition technology of the present invention. Detailed Description of the Invention

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0068] The purpose of the present invention is to provide a medical percussion sound intelligent classification system based on voiceprint recognition technology. By introducing an ensemble learning strategy and various signal processing technologies, the feature extraction ability and classification accuracy of the percussion sound signal are improved.

[0069] To make the above objects, features, and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0070] Embodiment 1

[0071] An intelligent classification system for medical percussion sounds based on voiceprint recognition technology, which includes a preprocessing module, a feature extraction module, and an integrated classification module. The preprocessing module uses the CEEMDAN algorithm and the DTW algorithm to denoise and filter the original percussion sound signal, and realizes automatic audio clipping through volume threshold detection. The feature extraction module combines the advantages of manual extraction and automatic extraction based on wavelet transform and convolutional recurrent neural network to extract the time-domain, frequency-domain, and time-frequency domain features of the percussion sound signal. The integrated classification module uses multiple base classifiers to perform integrated learning through the hard voting mechanism of the Voting Classifier and outputs the optimal category.

[0072] Specifically, the preprocessing module:

[0073] I. Use the CEEMDAN algorithm to decompose the original percussion sound signal into multiple intrinsic mode functions (IMFs) and residuals.

[0074] Steps of CEEMDAD:

[0075] 1. Import necessary libraries: including importing libraries required for processing audio and images.

[0076] 2. Define the EMD function: The emd function implements the basic steps of empirical mode decomposition, finding the local maxima and minima of the signal. Calculate the upper and lower envelopes and take their average as the IMF. Subtract this IMF from the original signal to get the residual. Repeat the above process until the termination condition is met (such as reaching the maximum number of IMFs or the amplitude of the residual is less than a certain threshold).

[0077] 3. Process the audio and save the images: including processing all.wav audio files in a folder, using CEEMDAN to decompose the audio signal, calculating the adaptive correlation, selecting the IMFs with high correlation, and plotting the graphs of the original signal and IMFs, as Figure 1 shown, the threshold of the correlation coefficient is 0.5.

[0078] 4. Synthesize the processed audio: Finally, it is used to synthesize the selected IMFs into a new.wav format file.

[0079] II. Use the DTW algorithm to compare the feature changes between different IMF components, select the IMF component most similar to the original signal for reconstruction, and achieve denoising and filtering.

[0080] Set the volume threshold, detect the audio percussion sounds exceeding the threshold, and clip the audio into multiple small segments.

[0081] Use indicators such as zero-crossing rate and spectral flatness to eliminate pure noise segments.

[0082] Furthermore, the feature extraction module is divided into manual feature extraction, automatic feature extraction based on wavelet transform and convolutional recurrent neural network, and feature fusion; among which

[0083] Manual feature extraction:

[0084] Time-domain features: Extract the time-domain features of the audio waveform diagram

[0085] 1. Interquartile Range (IQR): The interquartile range is a method in descriptive statistics used to determine the gap between the third quartile (Q3) and the first quartile (Q1). Quartiles are three points that divide a data set into four equal parts, where Q1 is the 25th percentile and Q3 is the 75th percentile. IQR is a measure of the dispersion of data, similar to variance and standard deviation, but it is more often used as a robust statistic. The formula for calculating IQR is: IQR = Q3 - Q1.

[0086] 2. Signal-to-Noise Ratio (SNR): The signal-to-noise ratio is an important parameter for measuring signal quality, representing the ratio of the power of the useful signal to the power of the background noise. SNR is defined as the ratio of signal power to noise power and is usually expressed in decibels (dB). Its formula is:

[0087]

[0088] Where:

[0089] P signal represents the signal power;

[0090] P noise represents the noise power.

[0091] 3. Max: The maximum amplitude of the audio

[0092] Frequency-domain features: Perform a Fourier transform (usually the fast Fourier transform FFT) on the audio signal to convert the time-domain signal into a frequency-domain signal, and then calculate these features based on the frequency-domain data.

[0093] 1. Spectral Slope: The spectral slope is the slope obtained by fitting the spectral envelope through a linear regression method. The specific steps include: performing a Fourier transform on the audio signal to obtain the frequency-domain representation. Select a frequency range, usually the entire spectrum or a part after removing the low-frequency part to reduce the influence of the DC component. Perform a linear regression analysis on the spectral amplitude (usually the logarithmic amplitude) within the selected frequency range to obtain the slope value.

[0094] The calculation of the spectral slope can be expressed by the following formula:

[0095]

[0096] Wherein:

[0097] H(f) is the spectral amplitude at frequency f;

[0098] f1 and f2 are the lower and upper limits of the selected frequency range.

[0099] 2. Spectral Crest Factor: The spectral crest factor is a characteristic that describes the sharpness of the spectral distribution. It is defined as the ratio of the maximum value of the spectrum to the average value of the spectrum. The Fourier transform is performed on the audio signal to obtain the frequency-domain representation.

[0100] The calculation steps are as follows:

[0101]

[0102] Where the maximum value of the spectrum is max(|H(f)|), and the average value of the spectrum is mean(|H(f)|). The spectral crest factor reflects the sharpness of the spectrum, and a high value indicates the presence of prominent peaks in the spectrum.

[0103] 3. Spectral Roll-off: The spectral roll-off is a specific frequency point in the spectrum, at which the frequency components before this point accumulate to a certain proportion (usually 85%) of the total energy. The calculation steps are as follows:

[0104] Total energy of the spectrum:

[0105] E total = ∑|H(f)| 2

[0106] Find the frequency point corresponding to when the cumulative energy reaches 85%, which is the spectral roll-off point.

[0107] Spectral roll-off point:

[0108] Find f roll-off such that

[0109]

[0110] Automatically extract features based on wavelet transform and convolutional recurrent neural network:

[0111] In this embodiment, the multi-scale analysis characteristics of wavelet transform are utilized to obtain the time-frequency characteristics of the audio signal. The wavelet energy spectrum has good time resolution in the high-frequency band and excellent frequency resolution in the low-frequency band, which is particularly suitable for capturing non-stationary characteristics in audio signals. The time-scale energy spectrum generated by wavelet transform is used to replace the traditional frequency spectrum. Interpolation operations are used to adjust all input features to a unified size, simplifying subsequent model training. The recurrent neural network captures the global dependencies between time steps. Combined with a classification layer, multi-class prediction is achieved through Softmax activation.

[0112] Specifically, the implementation of the model and algorithm

[0113] (1) Wavelet transform

[0114] Wavelet transform is a mathematical tool for analyzing signals through time-scale analysis, which can provide localized information in both the time domain and the frequency domain simultaneously. The output wavelet energy spectrum is the energy distribution of the signal at different time points and frequency scales, and is suitable for analyzing non-stationary signals (such as audio).

[0115] The audio signal is passed through continuous wavelet transform (CWT) to generate its time-scale energy spectrum.

[0116] Continuous wavelet transform:

[0117]

[0118] Where:

[0119] x(t) is the audio signal;

[0120] a is the scale parameter, reflecting the frequency resolution;

[0121] b is the time shift parameter;

[0122] ψ(t) is the mother wavelet function, which determines the analysis characteristics. Choosing the "cmor" complex wavelet is suitable for frequency-time resolution.

[0123] ψ * is the complex conjugate of the mother wavelet, ensuring the energy integrity of the signal;

[0124] W x (a, b)) is the wavelet coefficient.

[0125] Energy spectrum calculation:

[0126] S(a, b) = |W x (a, b)| 2

[0127] Where:

[0128] S(a,b) is the energy spectrum, reflecting the energy distribution of the signal in the time-scale space;

[0129] |W x (a,b)| 2 is the squared modulus of the wavelet coefficient, representing the energy intensity of the signal.

[0130] Normalization and scaling:

[0131] Normalize the energy spectrum S(a,b) to the interval [0,1] to improve the training stability.

[0132] Through interpolation operation, adjust S(a,b) to a fixed size (H,W):

[0133] S’(a,b) = zoom(S(a,b),(H / S a ,W / S b ))

[0134] Where:

[0135] S a ,S b are the height and width of the original spectrum respectively;

[0136] H,W are the target height and width respectively.

[0137] Specifically, the model and algorithm implementation

[0138] (2) Deep learning model

[0139] CRNN combines the advantages of convolutional neural network (CNN) and recurrent neural network (RNN): CNN is good at extracting local spatial features and is suitable for processing two-dimensional image features of wavelet energy spectrum; RNN is used to capture dependencies in time series and performs excellently in wavelet time series features.

[0140] Design a classification model based on CRNN, and the architecture includes the following parts:

[0141] Convolution module (CNN Layers)

[0142] Contains three two-dimensional convolutional layers, which extract local features of low, medium, and high levels respectively.

[0143] Convolution calculation formula:

[0144]

[0145] Where:

[0146] X is the input feature map;

[0147] K is the convolution kernel (weight matrix);

[0148] F ij is the value of the feature map of the convolution output at position (i, j);

[0149] i and j are the row and column indices of the current pixel;

[0150] m and n are the row and column indices of the convolution kernel.

[0151] ① Activation function (ReLU)

[0152] ReLU(z) = max(0, z)

[0153] Where:

[0154] z is the convolution result;

[0155] ReLU is a non - linear activation function that retains positive values and suppresses negative values, improving the model's expressive ability.

[0156] ② Pooling operation (Pooling)

[0157] After each layer of convolution, the feature dimension is reduced through the ReLU activation function and the pooling operation.

[0158] Pooling formula:

[0159] P ij = max m,n [i + m, j + n]

[0160] Where P ij is the feature map after pooling.

[0161] ③ Fully connected module (Fully Connected Layer)

[0162] After the output of the convolution module is flattened into a vector, it is input into the fully connected layer:

[0163] y h = σ(W h x + b h )

[0164] Where:

[0165] y h is the output of the hidden layer;

[0166] σ is the ReLU activation function;

[0167] W h is the weight matrix;

[0168] b h is the bias vector;

[0169] x is the input feature vector.

[0170] ④ Recurrent Module (GRU Layers)

[0171] The gated recurrent unit (GRU) is used to capture temporal features. GRU is an optimized recurrent unit that reduces the problem of gradient vanishing in long time series.

[0172] GRU update formula:

[0173] z t = σ(W z x t + U z h t-1 )

[0174] r t = σ(W t x t + U t h t-1 )

[0175]

[0176] Where:

[0177] z t is the update gate;

[0178] r t is the reset gate;

[0179] h t-1 is the previous hidden state;

[0180] h t is the current hidden state;

[0181] denotes element-wise multiplication;

[0182] is the candidate hidden state.

[0183] ⑤ Classification Layer

[0184] Outputs class probabilities through the Softmax activation function:

[0185]

[0186] Where:

[0187] is the predicted probability of the i-th class;

[0188] z i is the linear transformation result of the classifier output.

[0189] (3) Loss Function

[0190] Calculate the error between the model prediction and the true label using the cross - entropy loss function:

[0191]

[0192] Where:

[0193] is the average loss value;

[0194] N is the number of samples;

[0195] C is the number of classes;

[0196] y ij is the true label (1 or 0) of the i - th sample belonging to the j - th class;

[0197] is the predicted probability of the i - th sample belonging to the j - th class.

[0198] (4) Adam optimizer

[0199] Parameter update rule:

[0200]

[0201] Where:

[0202] θ t is the parameter after the t - th update;

[0203] η is the learning rate;

[0204] m t is the first - order momentum (exponentially weighted average of the gradient);

[0205] v t is the second - order momentum (exponentially weighted average of the squared gradient);

[0206] : A small value to prevent the denominator from being zero.

[0207] Feature fusion:

[0208] For the trained CRNN model, select the hidden state h at the last time step of the GRU layer GRU as the deep feature extraction layer. The features of this layer contain both the time - frequency - space features extracted by the convolutional layer and the temporal dependence captured by the GRU, with high information density. Since the dimension of h GRU is relatively high, to reduce the computational complexity and prevent overfitting, perform principal component analysis (PCA) dimensionality reduction on it. Fit the PCA model on the training set and set the target dimension to be the same as the artificial feature dimension. PCA projects the high - dimensional data into a low - dimensional space through linear transformation, retaining the direction with the largest variance in the data.

[0209] Normalize and concatenate the artificial features with the GRU features after dimensionality reduction to form a composite feature vector for each sample. The fused feature vector is input into an ensemble classifier to improve classification accuracy.

[0210] 1. Construct a data matrix: Arrange the GRU features of all samples into a data matrix

[0211]

[0212] where:

[0213] N is the number of samples;

[0214] is the GRU feature of the nth sample.

[0215] 2. Centralization processing: Centralize the data matrix by subtracting the mean vector μ of the features:

[0216] |H centered = H - μ|

[0217] 3. Calculate the covariance matrix:

[0218]

[0219] 4. Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C to obtain eigenvalues and eigenvectors:

[0220] C = V∧V T

[0221] where:

[0222] Λ is a diagonal matrix containing eigenvalues;

[0223] V is the corresponding eigenvector matrix.

[0224] 5. Dimensionality reduction processing: Select the first k largest eigenvalues and their corresponding eigenvectors to form a dimensionality reduction matrix W:

[0225] W = [V1, V2, … V K

[0226] Project the high-dimensional feature h GRU into the low-dimensional space:

[0227] f GRU = W T (h GRU - μ)

[0228] Obtain the GRU feature vector f GRU .

[0229] ​6. Normalization: To make the artificial features and GRU features comparable in terms of numerical scale, both are normalized as follows:

[0230] Normalize each feature:

[0231]

[0232] where f i is the original feature value, μ is the mean of the feature values, and σ i is the standard deviation of the feature.

[0233] Normalize f manual and f CRNN respectively to obtain the normalized feature vectors f' manual and f' GRU .

[0234] 7. Feature concatenation: Concatenate the normalized artificial features and GRU features to form a comprehensive feature vector f fusion :

[0235] f fusion = [f' manual , f' GRU

[0236] This operation combines the two types of features, enabling the model to utilize the information of both artificial features and deep features simultaneously.

[0237] 8. Input to the ensemble classifier: Use the fused feature vector f fusion as the input and import it into the ensemble classifier for training and testing.

[0238] This embodiment also provides the application and performance evaluation of the model and algorithm

[0239] In multi-classification tasks, the following metrics are commonly used to evaluate the performance of the model: Confusion Matrix, Accuracy, Precision, Recall, F1-score, Support, etc. The following are the methods and formulas for calculating these metrics (taking the Confusion Matrix, Accuracy, and Precision as examples).

[0240] (1) Confusion Matrix

[0241] The Confusion Matrix is a C×C matrix, where C is the number of classes in the classification task. The rows represent the actual classes, and the columns represent the predicted classes.

[0242] Representation form:

[0243]

[0244] Among them:

[0245] CM: Confusion Matrix.

[0246] TP ii (True Positives): The number of samples in which class i is correctly classified.

[0247] FP ij (False Positives): The number of samples that are actually in class i but are misclassified as class j.

[0248] FN ij (False Negatives): The number of samples that are actually in class i but are misclassified as class j. (2) Accuracy

[0249] Accuracy measures the proportion of correct predictions made by the model across all classes.

[0250] Calculation formula:

[0251]

[0252] Numerator: The sum of the diagonal elements of the confusion matrix (total number of correctly classified samples).

[0253] Denominator: The total number of all samples.

[0254] (3) Precision

[0255] Precision measures how many of the samples predicted to be in a particular class are correct.

[0256] Calculation formula (for class i):

[0257]

[0258] (4) Macro-Averaged Precision: Taking precision as an example

[0259] Take the average of the precisions for all classes:

[0260]

[0261] (5) Weighted-Averaged Precision: Taking precision as an example

[0262] Take the weighted average of the precisions for all classes (weights are the class supports):

[0263]

[0264] The core of multi-class performance evaluation starts from the confusion matrix and calculates indicators such as precision, recall, and F1-score through formulas. The main process is as follows:

[0265] Generate a confusion matrix and record the prediction results for each class.

[0266] Calculate Precision, Recall, and F1-Score for each class.

[0267] Perform macro-averaging (treating each class equally) or weighted-averaging (considering the class sample ratio) on the indicators.

[0268] Select a suitable evaluation indicator according to the task requirements.

[0269] Furthermore, this embodiment also includes an integrated classification module: After obtaining all features, we select those that contain both manually extracted features and automatically extracted features by neural networks. By applying typical machine learning methods using multiple base classifiers such as Random Forest, Logistic Regression, Support Vector Machine, XGBoost, K-Nearest Neighbor Algorithm, Naive Bayes, we obtain the prediction and classification effects of each classifier on this audio.

[0270] For example: Random Forest is an ensemble learning method that conducts classification or regression prediction by constructing multiple decision trees. Its main features include being composed of multiple decision trees, each tree being independently constructed, and introducing randomness during the construction process. Each decision tree usually uses Bootstrap sampling to draw samples from the original training set, which means the same sample may be drawn multiple times, while some samples may not be drawn at all. This also increases the independence between trees.

[0271] Logistic Regression is a widely used statistical model mainly for binary classification problems. Although it has the word "regression" in its name, it is actually a classification algorithm. The goal of Logistic Regression is to predict the probability of an event occurring, which is mapped through a logistic function (also known as the Sigmoid function), and it can map any value to between 0 and 1, representing the probability.

[0272] Support Vector Machine (SVM) is a powerful classification model that is widely used in the field of statistical learning. The goal of SVM is to find an optimal hyperplane in the feature space to distinguish different classes. The following are some key features of SVM: Maximum Margin Principle: The core idea of SVM is to find a hyperplane that maximizes the distance (i.e., margin) to the nearest data points (support vectors). This hyperplane separates data points of different classes and has the maximum margin, thereby improving the generalization ability of the model. Kernel Trick: SVM can handle non-linearly separable data through the Kernel Trick. Common kernel functions include linear kernel, polynomial kernel, Radial Basis Function (RBF) kernel, and Sigmoid kernel, etc. The kernel function allows SVM to find the optimal hyperplane in a high-dimensional space without explicitly performing dimensional transformation. Soft Margin and Regularization: In practical applications, the data may not be completely linearly separable. SVM introduces the concept of soft margin, allowing some data points to violate the maximum margin principle, that is, allowing some points to fall into the wrong class. This is achieved by introducing slack variables and a regularization parameter C, where C controls the trade-off between the margin width and the classification error.

[0273] It should be noted that this article uses multiple models including but not limited to these models.

[0274] After obtaining the prediction classification effects of all models on this audio, we use the hard voting mechanism of the Voting Classifier to synthesize the prediction results of each classifier and output the optimal class.

[0275] The hard voting mechanism of the Voting Classifier is an ensemble learning method. Its core idea is to vote on the prediction results of multiple classifiers, and the result with the majority of votes is used as the final prediction result. The following are the main features and principles of the hard voting mechanism:

[0276] 1. Independent Prediction: In hard voting, each model in the ensemble independently predicts the same data point, and the output of each model is a class label.

[0277] 2. Collect Votes: Collect the prediction results of each model, which is equivalent to each model voting for the class it predicts.

[0278] 3. Count Votes: For each class, calculate the number of votes received.

[0279] 4. Determine the Winner: Select the class with the most votes as the final prediction result. If there are multiple classes with the same number of votes, usually the first class to reach this number of votes is selected, or other strategies are used to solve the tie problem.

[0280] Finally, the optimal choice is obtained through the form of hard voting of the Voting Classifier, the category of the audio is judged, and the entire process of the model is completed.

[0281] Embodiment 2:

[0282] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the medical percussion sound intelligent classification system based on voiceprint recognition technology in Embodiment 1.

[0283] Optionally, the above electronic device may be a server.

[0284] In addition, an embodiment of the present invention also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the medical percussion sound intelligent classification system based on voiceprint recognition technology in Embodiment 1.

[0285] Embodiments of the present invention relate to a method, a system, and a computer program product. The invention can exist in the form of a complete hardware implementation, a complete software implementation, or a combination of software and hardware. Specifically, the present invention can be implemented on a computer-usable storage medium, which contains computer-executable program codes, and these storage media include but are not limited to disk memories, CD-ROMs, optical memories, etc.

[0286] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that computer program instructions can implement each process and / or block in the flowcharts and / or block diagrams, as well as combinations thereof. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices, generating a machine, such that the instructions executed by these processors can implement one or more processes in the flowchart, or the functions specified by one or more blocks in the block diagram.

[0287] These computer program instructions can also be stored in a computer-readable storage medium, which can guide a computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the storage medium generate a manufactured product, which includes an instruction device and implements one or more processes in the flowchart, or the functions specified by one or more blocks in the block diagram.

[0288] These computer program instructions can be loaded onto a computer or other programmable data processing device, causing the instructions to perform a series of operational steps on the computer or other programmable device to generate a computer-implemented processing procedure, thereby implementing, through the instructions executed on the computer or other programmable device, the steps of the functions specified in one or more of the processes in the flowchart and / or one or more of the boxes in the block diagram. (There is no flowchart in the article. Is it necessary to add it?)

[0289] This specification describes the various embodiments in a progressive manner and emphasizes the differences between each embodiment and other embodiments. The same or similar parts in each embodiment are referred to each other.

[0290] Specific examples of the present invention are used to illustrate its principles and implementation manners, aiming to help understand the method and core idea of the present invention. For those skilled in the art, the specific implementation manners and application scopes may be adjusted according to the concepts of the present invention. Therefore, the content of this specification should not be regarded as a limitation to the present invention.

Claims

1. An intelligent classification system for medical percussion sounds based on voiceprint recognition technology, characterized in that, It includes a preprocessing module, a feature extraction module, and an integrated classification module. Among them, the preprocessing module uses the CEEMDAN algorithm and the DTW algorithm to denoise and filter the original percussion sound signal, and realizes automatic audio clipping through volume threshold detection; the feature extraction module combines the advantages of manual extraction and automatic extraction based on wavelet transform and convolutional recurrent neural network to extract the time-domain, frequency-domain, and time-frequency domain features of the percussion sound signal; the integrated classification module uses multiple base classifiers to perform integrated learning through the hard voting mechanism of the Voting Classifier and outputs the optimal category.

2. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 1, wherein The preprocessing module uses the CEEMDAN algorithm to decompose the original percussion sound signal into multiple intrinsic mode functions (IMFs) and residuals; uses the DTW algorithm to compare the feature changes between different IMF components, and selects the IMF component most similar to the original signal for reconstruction to achieve denoising and filtering.

3. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 1, characterized in that, The features manually extracted in the feature extraction module include the time-domain features of the audio waveform diagram: Interquartile Range (IQR), Signal-to-Noise Ratio (SNR), and the maximum amplitude Max of the audio. Frequency-domain features: Perform Fourier transform on the audio signal to convert the time-domain signal into a frequency-domain signal, and then calculate based on the frequency-domain data: spectral slope, spectral peak factor, spectral roll-off point.

4. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 1, wherein In the feature extraction module, automatic extraction based on wavelet transform and convolutional recurrent neural network. Pass the audio signal through continuous wavelet transform (CWT) to generate its time-scale energy spectrum: Continuous wavelet transform: Where: x(t) is the audio signal; a is the scale parameter, reflecting the frequency resolution; b is the time translation parameter; ψ(t) is the mother wavelet function, determining the analysis characteristics, and the "cmor" complex wavelet is selected to be suitable for frequency-time resolution; ψ * is the complex conjugate of the mother wavelet, ensuring the energy integrity of the signal; Energy spectrum calculation: S(a,b) = |W x (a,b)| 2 Among them: S(a, b) is the energy spectrum, reflecting the energy distribution of the signal in the time-scale space; |W x (a, b)| 2 is the squared modulus of the wavelet coefficient, representing the energy intensity of the signal; Normalization and scaling: Normalize the energy spectrum S(a,b) to [0,1] to improve the training stability; Adjust S(a,b) to a fixed size (H,W) through interpolation operation: S'(a,b) = zoom(S(a,b),(H / S a ,W / S b )) Where: S a , S b : the height and width of the original spectrum; H, W: the target height and width.

5. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 1 or 4, characterized in that, The feature extraction module also includes a classification model based on CRNN, and its architecture includes the following parts: convolutional module (CNNLayers), activation function (ReLU), pooling operation (Pooling), fully connected module (Fully Connected Layer), recurrent module (GRU Layers), classification layer (Classification Layer).

6. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 1 or 5, characterized in that, It also includes calculating the error between the model prediction and the true label using the cross-entropy loss function: Wherein: is the average loss value; N is the number of samples; C is the number of classes; y ij is the true label (1 or 0) of the i-th sample belonging to the j-th class; is the predicted probability of the i-th sample belonging to the j-th class.

7. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 6, wherein It also includes the Adam optimizer: Parameter update rule: Where: θ t is the parameter after the t-th update; η is the learning rate; m t is the first-order momentum, the exponentially weighted average of the gradients; v t is the second-order momentum, the exponentially weighted average of the squared gradients; is a small value to prevent the denominator from being zero.

8. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 7, characterized in that, It also includes feature fusion: The manual features and the dimension-reduced GRU features are normalized and concatenated to form a composite feature vector for each sample, and the fused feature vector is input into the integrated classifier.

9. The intelligent classification system for medical percussion sounds based on voiceprint recognition technology according to claim 1, characterized in that, The integrated classification module uses multiple base classifiers such as random forest, logistic regression, support vector machine, XGBoost, K-nearest neighbor algorithm, naive Bayes, K-nearest neighbor algorithm or naive Bayes to apply typical machine learning methods to obtain the prediction classification effects of each classifier on this audio.

Images