Method for analyzing cough sound by using disease characteristics to diagnose respiratory diseases

Abstract

The invention relates to the field of biological medicine, and discloses a method and system for analyzing cough sound by using disease characteristics to diagnose respiratory diseases, and the method comprises the steps: deploying a six-microphone annular array to achieve the precise positioning and triggering of a sound source; self-adaptive spectral subtraction and Wiener filtering cascade are adopted to enhance the audio; segmenting a cough segment based on energy envelope; fusing the Mel-cepstrum, the linear prediction residual error, the harmonic energy ratio and the transient zero-crossing rate to construct a pathological feature matrix; extracting local, medium-range and global time sequence features through a three-branch parallel convolutional network; inputting a disease specific classifier to discriminate asthma, pneumonia and laryngitis respectively, and applying a feature decoupling regular term to improve interpretability. The system correspondingly realizes the modularized processing flow. According to the method, the cough sound collection quality and the disease subtype recognition accuracy in a complex environment are improved, meanwhile, the thermodynamic diagram is output to assist clinical decision making, and the diagnosis credibility and practicability are enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of biological medicine, and particularly relates to a method for diagnosing respiratory diseases by analyzing cough sound with disease characteristics. BACKGROUND

[0002] With the deep integration of artificial intelligence and biosensing technology in the field of medical health, non-contact disease diagnosis methods based on sound signals are gradually becoming an important research direction for respiratory disease screening. Coughing, as a typical physiological response of respiratory diseases, its acoustic characteristics contain rich pathological information. By analyzing the frequency spectrum, time domain and nonlinear dynamic parameters of cough sound, early identification and classification of diseases such as asthma, chronic obstructive pulmonary disease and pneumonia can be achieved. The current mainstream method relies on high-fidelity microphones to collect audio in a quiet clinical environment, and performs feature extraction and diagnosis reasoning through a cloud model. Its technical path is based on the assumption of ideal signal-to-noise ratio and stable sound field, ignoring the complex interference characteristics of sound signals in real-life scenarios.

[0003] Among them, the cough sound diagnosis technology for family and community environment aims to realize non-invasive, continuous and privacy-protected respiratory health monitoring. This direction requires the diagnosis system to automatically capture cough events in the user's daily activity state, and to complete preliminary signal purification and feature encoding on the device side, thereby reducing the dependence on environmental control and manual intervention. However, existing technologies generally treat biosensors as independent data channels, only collecting and uploading raw signals, lacking multi-modal sensing coordination and front-end intelligent processing capabilities, resulting in serious lack of robustness of the system in complex acoustic scenarios.

[0004] The cough audio collected by existing technologies in noisy environments is easily mixed with non-related sound sources such as TV conversations, appliance operation and street traffic, causing serious distortion of key features such as mel-frequency cepstral coefficients and power spectral density, and increasing the misdiagnosis rate. At the same time, traditional noise reduction algorithms rely on fixed noise templates or single-microphone spectral subtraction, which cannot dynamically distinguish between cough sound sources and transient interference, and do not use physiological trigger context provided by laryngeal vibration and electromyographic signals, resulting in high missing detection rate of effective cough segments. If the raw audio is directly uploaded to the cloud for processing, it not only increases communication load and response delay, but also causes user privacy leakage risk. These problems are particularly prominent in high-frequency demand scenarios such as home chronic disease management, child respiratory tract monitoring and elderly health warning, and urgently need a front-end collaborative diagnosis architecture that integrates multi-modal sensing, embedded intelligent processing and privacy protection mechanisms. SUMMARY

[0005] The application provides a method for diagnosing respiratory diseases by analyzing cough sound with disease characteristics, which realizes precise spatial positioning and physical isolation of cough sound source in complex acoustic environment by constructing a front-end acquisition system with a multi-channel acoustic sensing array working in cooperation with an adaptive noise suppression model, then adopts a time-frequency domain joint enhancement algorithm to perform nonlinear filtering and spectrum reconstruction on the original audio signal, eliminates steady-state and non-steady-state interference components, and obtains a high-fidelity cough audio segment. On this basis, a deep time sequence feature extraction network based on prior knowledge guidance of pathophysiology is established, which fuses four types of acoustic parameters, including mel-frequency cepstral coefficients, linear prediction residual spectrum, harmonic energy ratio and transient zero-crossing rate, to construct a cough sound pathological feature vector space. Further, a respiratory disease specific classifier architecture is introduced, which is composed of a bronchospasm feature detection sub-network, a lung parenchyma exudation feature recognition sub-network and an upper respiratory tract obstruction feature discrimination sub-network, each sub-network corresponding to three typical respiratory system diseases, asthma, pneumonia and laryngitis, and realizing disease-oriented decoupling of the feature space and optimization of the discrimination boundary through a multi-task joint training mechanism, and finally outputting a disease probability distribution and confidence evaluation with clinical interpretability.

[0006] The application provides a method for diagnosing respiratory diseases by analyzing cough sounds with disease characteristics, which comprises: deploying a ring array structure composed of six omnidirectional microphones, each microphone is uniformly distributed at an interval of 60° angle on a rigid circular ring support with a diameter of 30 cm, the sampling frequency is set to 44100 Hz, and the quantization bit number is 16 bits; the time difference of sound source arriving at each microphone is calculated by a beam forming algorithm, and a three-dimensional sound source positioning function is constructed, when the positioning result converges in the preset human oral cavity space coordinate range and the duration is greater than 0.5 s, a cough event capture mechanism is triggered; an adaptive spectral subtraction and wiener filtering cascade processing procedure is started, first, the original multi-channel signal is subjected to short-time Fourier transform to obtain a time-frequency matrix, then the noise power spectral density function is estimated, and the noise residue is dynamically suppressed by using an asymmetric window function, the rising edge time of the window function is 5 ms, and the falling edge time is 20 ms; the enhanced audio signal is subjected to automatic segmentation of the cough segment, and the segmentation is based on the energy envelope threshold method, the threshold is set to 3 times the standard deviation of the local mean energy, the segmentation window length is 200 ms, and the overlap rate is 50%; the mel frequency cepstral coefficients of the segmented cough audio are extracted, the order is 13 dimensions, the first eight autocorrelation coefficients of the linear prediction residual spectrum are calculated, and the harmonic energy ratio parameter is superimposed, the parameter is defined as the ratio of the total energy in the fundamental frequency and its second and third harmonic frequency bands to the energy of the full frequency band, and the transient zero-crossing rate, i.e. the value of the number of signal zero-crossings per unit time after normalization processing, is calculated; the above four types of parameters are spliced into a two-dimensional feature matrix along the time axis, and input into a pathological feature extractor composed of three parallel convolutional neural network branches, the first branch adopts a one-dimensional convolution kernel size of three to extract local time sequence patterns, the second branch adopts a structure of a hole convolution kernel size of 5 and an expansion rate of two to capture medium-range dependencies, and the third branch adopts a structure of a max pooling layer followed by a fully connected layer to extract global statistical features; the feature vectors output by the three branches are spliced and input into a disease-specific classification module, the module includes three independent fully connected layers corresponding to bronchospasm feature detection, lung parenchyma exudation feature recognition and upper respiratory tract obstruction feature discrimination, respectively, the number of neurons in each layer is 128, 96 and 64 respectively, and the activation function adopts a rectified linear unit; end-to-end training is carried out through a cross-entropy loss function and a disease label, a label smoothing strategy is introduced in the training process, the smoothing coefficient is set to 0.1, and a feature decoupling regularization term is applied to force each classifier to respond only to its corresponding pathological feature dimension; in the inference stage, the input is a cough audio to be tested, and the output is the probability values of three disease categories and the corresponding feature activation heat map, the horizontal axis of the heat map is the time frame number, the vertical axis is the feature dimension index, and the color intensity represents the contribution weight of the dimension to the classification decision.

[0007] According to another aspect of the present application, a system for diagnosing respiratory diseases by analyzing cough sounds with disease characteristics is provided, which comprises: a multi-channel acoustic sensing array module for deploying a ring array structure composed of six omnidirectional microphones, each microphone being uniformly distributed at an interval of 60° angle on a rigid circular ring support with a diameter of 30 cm, the sampling frequency being set to 44100 Hz, and the quantization bit number being 16 bits; a sound source positioning and event triggering module for calculating the time difference of sound source arrival at each microphone through a beamforming algorithm, constructing a three-dimensional sound source positioning function, and triggering a cough event capture mechanism when the positioning result converges within a preset human oral cavity space coordinate range and the duration is greater than 0.5 s; an adaptive noise suppression module for starting an adaptive spectral subtraction and Wiener filtering cascade processing procedure, first performing short-time Fourier transform on the original multi-channel signal to obtain a time-frequency matrix, then estimating the noise power spectral density function, and dynamically suppressing noise residues using an asymmetric window function with a rising edge time of 5 ms and a falling edge time of 20 ms; a cough segment automatic segmentation module for performing cough segment automatic segmentation on the enhanced audio signal, the segmentation being based on an energy envelope threshold method with a threshold set to 3 times the standard deviation of the local mean energy, a segmentation window length of 200 ms, and an overlap rate of 50%; a pathological acoustic feature extraction module for extracting the Mel-frequency cepstral coefficients of the segmented cough audio, taking 13 dimensions as the order, simultaneously calculating the first eight-order autocorrelation coefficients of the linear prediction residual spectrum, and superimposing the harmonic energy ratio parameter defined as the ratio of the total energy in the fundamental frequency and its second and third harmonic frequency bands to the full-band energy, and then calculating the transient zero-crossing rate, i.e. the value of the number of signal zero-crossings per unit time after normalization processing; a deep time sequence feature encoding module for concatenating the above four types of parameters along the time axis into a two-dimensional feature matrix, inputting the feature matrix into a pathological feature extractor composed of three parallel convolutional neural network branches, the first branch using a one-dimensional convolution kernel size of three to extract local time sequence patterns, the second branch using a hollow convolution kernel size of 5 and an expansion rate of two to capture medium-range dependencies, and the third branch using a max-pooling layer followed by a fully connected layer structure to extract global statistical features; a disease-specific classification module for inputting the concatenated feature vectors output by the three branches into the disease-specific classification module, which comprises three independent fully connected layers corresponding to bronchospasm feature detection, lung parenchyma exudation feature recognition, and upper respiratory tract obstruction feature discrimination, respectively, the number of neurons in each layer being 128, 96, and 64 respectively, and the activation function being a rectified linear unit; a model training and reasoning module for performing end-to-end training with the disease labels through a cross-entropy loss function, introducing a label smoothing strategy during the training process, and setting the smoothing coefficient to 0.1. Simultaneously, a feature decoupling regularization term is applied, forcing each classifier to respond only to its corresponding pathological feature dimension. During the inference phase, the input is the cough audio to be tested, and the output is the probability values ​​of three disease categories and the corresponding feature activation heatmap. The horizontal axis of the heatmap represents the time frame number, the vertical axis represents the feature dimension index, and the color intensity represents the contribution weight of that dimension to the classification decision.

[0008] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0009] 1. By constructing a dual guarantee mechanism of physical space constraints and signal processing algorithms, it can still stably capture high signal-to-noise ratio cough audio in high-noise environments such as homes, solving the feature pollution problem caused by sound source aliasing in traditional single-microphone solutions;

[0010] 2. A time-frequency domain joint enhancement strategy is adopted to effectively suppress non-stationary interference such as television sound, conversation sound, and traffic noise, while preserving the unique transient impact and harmonic structure of cough sound, providing a clean input for subsequent feature extraction;

[0011] 3. Establish a pathophysiology-guided deep feature extraction architecture that organically integrates four types of acoustic parameters: Mel cepstral spectrum, linear prediction residual, harmonic energy ratio, and transient zero-crossing rate, breaking through the limitations of traditional methods that rely on only a single feature dimension.

[0012] 4. Design a disease-specific multi-branch classifier. Through independent modeling and joint optimization of three pathological features—bronchospasm, pulmonary parenchymal exudation, and upper respiratory tract obstruction—accurate mapping of cough sounds to respiratory disease subtypes can be achieved, avoiding misjudgment caused by feature confusion in general classifiers.

[0013] 5. Introducing feature decoupling regularization terms and activation heatmap visualization mechanisms makes the model decision-making process clinically interpretable, helps doctors understand the basis of algorithm judgment, and improves diagnostic credibility and practicality. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall technical solution architecture of a method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics, as proposed in this invention.

[0015] Figure 2 This is a schematic diagram of the core principle framework of deep temporal feature extraction and disease-specific classification based on pathophysiological priors in this invention;

[0016] Figure 3 This is a logical flowchart of the multi-channel acoustic sensor array and adaptive noise suppression collaborative front-end acquisition in this invention;

[0017] Figure 4 This is a flowchart illustrating the logical flow of the joint enhancement of cough audio in the time and frequency domains and the construction of pathological acoustic feature vectors in this invention.

[0018] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the parallel decoupled classification of three disease characteristics: bronchospasm, pulmonary parenchymal exudation, and upper respiratory tract obstruction in this invention.

[0019] Figure 6 This is a logical flowchart of the feature activation heatmap generation and clinical interpretability decision output in this invention; Detailed Implementation

[0020] Please refer to Figures 1 to 6 This invention provides a method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics. Its core lies in constructing a front-end acquisition system that works in conjunction with a multi-channel acoustic sensor array and an adaptive noise suppression model to achieve precise spatial localization and physical isolation of the cough sound source in a complex acoustic environment. Then, a time-frequency domain joint enhancement algorithm is used to perform nonlinear filtering and spectrum reconstruction on the original audio signal to eliminate steady-state and non-steady-state interference components and obtain a high-fidelity cough audio segment.

[0021] Building upon this foundation, a deep temporal feature extraction network guided by prior knowledge of pathophysiology is established. This network integrates four types of acoustic parameters—Melch cepstral coefficients, linear prediction residual spectrum, harmonic energy ratio, and transient zero-crossing rate—to construct a pathological feature vector space for cough sounds. Furthermore, a respiratory disease-specific classifier architecture is introduced. This architecture consists of a bronchospasm feature detection subnetwork, a lung parenchymal exudation feature recognition subnetwork, and an upper airway obstruction feature discrimination subnetwork. Each subnetwork corresponds to three typical respiratory diseases: asthma, pneumonia, and laryngitis. A multi-task joint training mechanism achieves disease-oriented decoupling and discrimination boundary optimization of the feature space, ultimately outputting a clinically interpretable disease probability distribution and confidence assessment.

[0022] The method includes the following steps:

[0023] S1. Deploy a ring array structure consisting of six omnidirectional microphones. Each microphone is evenly distributed at a 60° interval on a rigid circular support with a diameter of 30cm. The sampling frequency is set to 44100Hz and the quantization bit depth is 16bit.

[0024] S2. Calculate the time difference of the sound source arriving at each microphone through the beamforming algorithm, construct a three-dimensional sound source localization function, and trigger the cough event capture mechanism when the localization result converges within the preset human oral cavity spatial coordinate range and the duration is greater than 0.5s.

[0025] S3. Start the adaptive spectral subtraction and Wiener filtering cascade processing flow. First, perform short-time Fourier transform on the original multi-channel signal to obtain the time-frequency matrix. Then, estimate the noise power spectral density function and use an asymmetric window function to dynamically suppress noise residue. The rise time of the window function is 5ms and the fall time is 20ms.

[0026] S4. Perform automatic cough segmentation on the enhanced audio signal. The segmentation is based on the energy envelope threshold method. The threshold is set to 3 times the standard deviation of the local mean energy. The segmentation window length is 200ms and the overlap rate is 50%.

[0027] S5. Extract the Mel-Cepstral Coefficients of the segmented cough audio, with an order of 13. Simultaneously calculate the first eight autocorrelation coefficients of the linear prediction residual spectrum and superimpose the harmonic energy ratio parameter, which is defined as the ratio of the total energy in the fundamental frequency and its second and third harmonic frequency bands to the total energy of the entire frequency band. Then calculate the transient zero-crossing rate, which is the normalized value of the number of times the signal crosses zero points per unit time.

[0028] S6. The above four types of parameters are concatenated into a two-dimensional feature matrix along the time axis and input into a pathological feature extractor consisting of three parallel convolutional neural network branches. The first branch uses a one-dimensional convolutional kernel with a size of three to extract local temporal patterns. The second branch uses a dilated convolutional kernel with a size of 5 and an inflation rate of two to capture mid-range dependencies. The third branch uses a max pooling layer followed by a fully connected layer to extract global statistical features.

[0029] S7. The three branch output feature vectors are concatenated and input into the disease-specific classification module. This module contains three independent fully connected layers, which correspond to bronchospasm feature detection, lung parenchymal exudation feature recognition, and upper airway obstruction feature discrimination, respectively. The number of neurons in each layer is 128, 96, and 64, respectively, and the activation function is a modified linear unit.

[0030] S8. End-to-end training is performed using the cross-entropy loss function and disease labels. During the training process, a label smoothing strategy is introduced with a smoothing coefficient of 0.1. At the same time, a feature decoupling regularization term is applied to force each classifier to respond only to its corresponding pathological feature dimension.

[0031] S9. In the inference stage, input the cough audio to be tested, and output the probability values ​​of three disease categories and the corresponding feature activation heatmaps. The horizontal axis of the heatmap is the time frame number, the vertical axis is the feature dimension index, and the color intensity represents the contribution weight of the dimension to the classification decision.

[0032] In S1, a ring array structure consisting of six omnidirectional microphones is deployed. Each microphone is evenly distributed at a 60° interval on a rigid ring support with a diameter of 30cm. The sampling frequency is set to 44100Hz and the quantization bit depth is 16bit.

[0033] The physical layout of the array structure ensures omnidirectional coverage of the sound source direction in three-dimensional space. Its geometric symmetry eliminates directional blind spots, allowing at least three microphones to capture the cough sound source simultaneously, regardless of whether it is located in front of, to the side of, or behind the array, thus meeting the basic requirements of beamforming algorithms for spatial redundancy.

[0034] Each microphone unit is equipped with an independent preamplifier circuit and an anti-aliasing filter. The preamplifier circuit gain is set to 40x, and the anti-aliasing filter cutoff frequency is 22050Hz to ensure no spectral aliasing within the Nyquist frequency range.

[0035] The sampling frequency of 44,100 Hz was chosen based on the physiological characteristic that the main energy of a cough is concentrated between 50 Hz and 8 kHz. This sampling rate is sufficient to cover the entire effective frequency band of a cough and preserve its high-frequency transient components.

[0036] The 16-bit quantization provides a dynamic range of 96 dB, which is sufficient to handle sound pressure level differences of up to 60 dB to 80 dB between coughing sounds and background noise in a home environment, preventing weak coughing signals from being drowned out by quantization noise.

[0037] Synchronous sampling of the six microphones is driven by the same crystal oscillator, with clock jitter less than 0.1μs, ensuring time consistency between channels and providing a physical basis for subsequent time difference calculations. The array support is made of rigid aluminum alloy and covered with a sound-absorbing foam layer to suppress the transmission of structural noise caused by the support's own vibration.

[0038] The microphone mounting holes are precision machined, with a positional error of less than 0.5mm and an angle error of less than 0.5°, ensuring the spatial resolution of the beamforming algorithm.

[0039] The entire array system is connected to the central processing unit via a universal serial bus interface. Data transmission adopts an isochronous transmission mode to ensure the real-time performance and integrity of the audio stream.

[0040] In S2, the time difference between the arrival of the sound source at each microphone is calculated using a beamforming algorithm, and a three-dimensional sound source localization function is constructed. When the localization result converges within the preset human oral cavity spatial coordinate range and the duration is greater than 0.5s, the cough event capture mechanism is triggered.

[0041] The beamforming algorithm uses the generalized cross-correlation phase transform method to calculate the time delay estimate between any two microphone pairs. A total of 15 cross-correlation functions are generated for the six-channel signals. Each function is calculated by fast Fourier transform, with 1024 transformation points.

[0042] The time delay estimate is smoothed by a median filter with a window length of 5 frames to eliminate transient noise interference.

[0043] The three-dimensional sound source localization function is constructed based on the principle of spherical intersection. A right-handed Cartesian coordinate system is established with the geometric center of the array as the origin. The sound source position coordinates are solved by the least squares method, and the iterative convergence threshold is set to 0.01m.

[0044] The preset spatial coordinate range of the human oral cavity is defined as a spherical space with a radius of 0.8m centered at the origin. This range covers the possible positions of the user's mouth relative to the array in typical sitting or standing postures. The positioning result must be within this space for 10 consecutive frames, and the positioning error of each frame must be less than 0.05m, in order to be determined as a valid cough sound source.

[0045] The 0.5s duration setting is based on the physiological characteristics of cough. The duration of a single cough is usually between 0.1 and 0.3s, while the duration of a continuous cough sequence is mostly between 0.5s and 1.5s. This threshold can effectively distinguish between a single cough and a continuous cough sequence, while excluding interference from transient environmental noise.

[0046] Once the trigger mechanism is activated, the system locks the current sound source direction, initiates subsequent noise suppression and signal enhancement processes, and records the sound source spatial coordinates for subsequent multi-channel signal alignment.

[0047] In S3, the adaptive spectral subtraction and Wiener filtering cascaded processing flow is initiated. First, a short-time Fourier transform is performed on the original multi-channel signal to obtain the time-frequency matrix. Then, the noise power spectral density function is estimated, and an asymmetric window function is used to dynamically suppress noise residue. The rise time of the window function is 5ms and the fall time is 20ms.

[0048] The short-time Fourier transform uses a Hamming window with a length of 25 ms, a frame shift of 10 ms, and 1024 transform points, generating a time-frequency matrix with a dimension of the number of frames multiplied by 513. The noise power spectral density function is estimated using the minimum statistics method, with a statistical window length of 3 s and an update step size of 0.5 s. The initial noise spectrum is calculated from the background noise collected during the system's silent period. The adaptive spectral subtraction formula is as follows:

[0049] ;

[0050] in, The enhanced power spectrum at the k-th frequency point;

[0051] The power spectrum includes noise.

[0052] Noise power spectrum;

[0053] The over-subtraction factor has a value of 4.

[0054] This is the lower limit factor of the spectrum, with a value of 0.01.

[0055] The transfer function of the Wiener filter is calculated by the following equation:

[0056] ;

[0057] The asymmetric window function uses a combination of exponential rising and linear falling forms. The expression for the rising segment is:

[0058] ,

[0059] in It takes 5ms;

[0060] The expression for the descent segment is:

[0061] ,

[0062] in It takes 20ms.

[0063] This window function quickly builds up gain at the onset of a cough to prevent transient impacts from being suppressed, and slowly reduces gain during the decay phase to prevent the speech tail from being truncated.

[0064] The cascaded signal is reconstructed into a time-domain waveform via inverse short-time Fourier transform and synthesized using an overlapping addition method to ensure phase continuity. The signal-to-noise ratio of the enhanced signal is improved by more than 15 dB, and the speech distortion is less than 5%, meeting the signal fidelity requirements for subsequent feature extraction.

[0065] In S4, the enhanced audio signal is automatically segmented for cough segments. The segmentation is based on the energy envelope threshold method, with the threshold set to three times the standard deviation of the local mean energy. The segmentation window length is 200ms and the overlap rate is 50%.

[0066] The energy envelope is obtained by calculating the sum of squares of the signal in each frame, with a frame length of 25ms and a frame shift of 10ms. The local mean energy is calculated using a sliding window with a window length of 1.5s and a step size of 0.1s, and the standard deviation is updated synchronously.

[0067] A dynamic threshold adjustment mechanism ensures effective detection of cough events under varying background noise levels. A 200ms segmentation window covers the duration of most individual coughs, and a 50% overlap rate guarantees feature integrity at segmentation boundaries.

[0068] The segmentation process employs a dual-threshold detection strategy, with an initial threshold of 3 times the standard deviation and an ending threshold of 1.5 times the standard deviation, to avoid premature truncation due to energy fluctuations.

[0069] After segmentation, the fragments must meet the minimum duration of 50ms and the maximum duration of 500ms. Fragments exceeding these ranges are considered invalid and are discarded.

[0070] Valid cough fragments are stored in chronological order, with timestamps and spatial coordinate labels of the sound source attached for subsequent batch processing.

[0071] In S5, the Mel-Cepstral Coefficients of the segmented cough audio are extracted, with an order of 13. At the same time, the first eight autocorrelation coefficients of the linear prediction residual spectrum are calculated, and the harmonic energy ratio parameter is superimposed. This parameter is defined as the ratio of the total energy of the fundamental frequency and its second and third harmonic frequency bands to the total energy of the entire frequency band. Then, the transient zero-crossing rate is calculated, which is the normalized value of the number of times the signal crosses zero points per unit time.

[0072] The Mel cepstral coefficient extraction process includes pre-emphasis, framing, windowing, Fourier transform, Mel filter bank mapping, logarithmic compression, and discrete cosine transform. The pre-emphasis coefficient is 0.97, the framing parameters are the same as described above, and the number of Mel filter banks is 40, covering a frequency range of 50Hz to 8kHz.

[0073] The linear prediction residual spectrum was obtained through tenth-order linear prediction analysis. The prediction coefficients were solved using the autocorrelation method. The first eight autocorrelation coefficients of the residual signal were normalized and used as the eigenvectors.

[0074] The calculation of the harmonic energy ratio parameter requires first estimating the fundamental frequency using the autocorrelation method, with a search range of 70Hz to 1200Hz. The delay corresponding to the maximum autocorrelation peak is then used as the fundamental frequency period.

[0075] The fundamental frequency and its second and third harmonic frequency bands are defined as rectangular windows with a bandwidth of 20Hz centered on each harmonic frequency, and energy calculations are performed in the power spectral domain.

[0076] Before calculating the transient zero-crossing rate, the signal needs to be normalized to zero mean. Zero-crossing detection uses the sign change method, which counts the number of zero-crossings in each frame and divides it by the frame length to obtain the normalized rate.

[0077] The four types of parameters are aligned according to time frames to form a two-dimensional feature matrix with a dimension of 25 times the number of frames. The matrix includes a 13-dimensional Mel cepstrum, an 8-dimensional linear prediction residual autocorrelation, a 1-dimensional harmonic energy ratio, a 1-dimensional transient zero-crossing rate, and two additional dimensions for first-order and second-order difference coefficients to enhance the dynamic characteristics of the time series.

[0078] In S6, the above four types of parameters are concatenated into a two-dimensional feature matrix along the time axis and input into a pathological feature extractor consisting of three parallel convolutional neural network branches. The first branch uses a one-dimensional convolutional kernel with a size of three to extract local temporal patterns. The second branch uses a dilated convolutional kernel with a size of 5 and an inflation rate of two to capture mid-range dependencies. The third branch uses a max pooling layer followed by a fully connected layer to extract global statistical features.

[0079] The first branch contains three 1D convolutional layers with 32, 64, and 128 kernels respectively. The activation function is a modified linear unit (MRU), and each layer is followed by batch normalization and random deactivation with a probability of 0.2. The second branch contains two dilated convolutional layers with 64 kernels in the first layer and 128 kernels in the second layer. The activation function and regularization configuration are the same as in the first branch.

[0080] The third branch first performs max pooling on the input feature matrix along the time axis, with the pooling window covering all time frames, and outputs a 25-dimensional static feature vector. This vector is then passed through two fully connected layers with 256 and 128 neurons, respectively. The activation function is a modified linear unit with a random deactivation probability of 0.3.

[0081] The three branches output feature vectors with dimensions of 128, 128, and 128 respectively, which are concatenated to form a 384-dimensional joint feature vector. This design ensures the complementary fusion of local transient features, mid-range contextual features, and global statistical features, providing multi-scale representation for subsequent disease classification.

[0082] In S7, the feature vectors output from the three branches are concatenated and input into the disease-specific classification module. This module contains three independent fully connected layers, corresponding to bronchospasm feature detection, lung parenchymal exudation feature recognition, and upper airway obstruction feature discrimination, respectively. The number of neurons in each layer is 128, 96, and 64, respectively, and the activation function is a modified linear unit.

[0083] The bronchospasm feature detection layer receives 384-dimensional input and outputs 128-dimensional features. After activation by a modified linear unit, the features are input to the second layer of 96-dimensional neurons, then activated again and input to the third layer of 64-dimensional neurons. Finally, the asthma probability is output through a soft maximum function.

[0084] The lung parenchymal exudation feature recognition layer and the upper respiratory tract obstruction feature discrimination layer have the same structure, and output the probabilities of pneumonia and laryngitis respectively.

[0085] The three classifiers share the parameters of the previous feature extractor, but a feature decoupling regularization term is applied during training, which is mathematically expressed as:

[0086] ;

[0087] in, Let be the weight matrix of the i-th classifier;

[0088] It is the Frobenius norm.

[0089] This regularization term forces the weight vectors of different disease classifiers to be orthogonal, reducing feature response overlap and improving specificity. The cross-entropy loss function, combined with a label smoothing strategy, has a smoothing coefficient of 0.1 to avoid overfitting the model to the training labels. Training uses a stochastic gradient descent optimizer with an initial learning rate of 0.001, a batch size of 32, and 50 training epochs, with the learning rate decaying to 1 / 10 of its original value every ten epochs. During the inference phase, the input is the audio of the cough to be tested. After processing through the aforementioned steps, the output is the probability values ​​for three disease categories. The disease with the highest probability is taken as the diagnosis result, and a feature activation heatmap is generated simultaneously.

[0090] The heatmap data is obtained by the dot product of the input feature vector and the weight matrix of the last fully connected layer of each classifier. After soft maximum normalization, it is mapped to color intensity. The horizontal axis is the time frame number and the vertical axis is the feature dimension index, which intuitively shows which time points and which feature dimensions contribute the most to the current diagnosis, providing a basis for decision-making for clinicians.

[0091] This invention, through the aforementioned steps, achieves high-fidelity acquisition and accurate disease classification of cough audio in high-noise environments such as homes. Multi-channel array and beamforming technologies solve the sound source aliasing problem; the adaptive noise suppression algorithm effectively separates cough sounds from background interference; multi-parameter fusion feature extraction overcomes the limitations of single feature dimensions; the disease-specific classification architecture avoids the misjudgment tendency of general models; and feature decoupling and heatmap mechanisms enhance the model's interpretability.

[0092] The entire method was tested in a real-world home environment, demonstrating improved average classification accuracy and reduced misdiagnosis rate for three types of diseases: asthma, pneumonia, and laryngitis, meeting the accuracy requirements for clinical auxiliary diagnosis. The system has low response latency, supports real-time continuous monitoring, and is suitable for various application scenarios such as home health monitoring, remote medical consultation, and early screening for epidemics.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing respiratory diseases by analyzing cough sounds based on disease characteristics, characterized in that, include: Deploy a ring array structure consisting of six omnidirectional microphones; The time difference between the arrival of the sound source at each microphone is calculated by beamforming algorithm, and a three-dimensional sound source localization function is constructed. When the localization result converges within the preset human oral cavity spatial coordinate range and the duration is greater than 0.5s, the cough event capture mechanism is triggered. The adaptive spectral subtraction and Wiener filtering cascade processing flow is initiated. First, a short-time Fourier transform is performed on the original multi-channel signal to obtain the time-frequency matrix. Then, the noise power spectral density function is estimated, and an asymmetric window function is used to dynamically suppress the noise residue. Automatic cough segmentation is performed on the enhanced audio signal. The segmentation is based on the energy envelope threshold method, and the threshold is set to 3 times the standard deviation of the local mean energy. Mel-frequency cepstral coefficients of the segmented cough audio are extracted, with an order of 13. The first eight autocorrelation coefficients of the linear prediction residual spectrum are calculated, and the harmonic energy ratio parameter is superimposed. This parameter is defined as the ratio of the total energy of the fundamental frequency and its second and third harmonic frequency bands to the total energy of the entire frequency band. Then, the transient zero-crossing rate is calculated, which is the normalized value of the number of times the signal crosses zero points per unit time. The above four types of parameters are concatenated into a two-dimensional feature matrix along the time axis and input into a pathological feature extractor consisting of three parallel convolutional neural network branches; End-to-end training is performed using the cross-entropy loss function and disease labels. During training, a label smoothing strategy is introduced with a smoothing coefficient of 0.

1. At the same time, a feature decoupling regularization term is applied to force each classifier to respond only to its corresponding pathological feature dimension. During the inference phase, the input is the audio of the cough to be tested, and the output is the probability values ​​of three disease categories and the corresponding feature activation heatmap. The horizontal axis of the heatmap is the time frame number, the vertical axis is the feature dimension index, and the color intensity represents the contribution weight of the dimension to the classification decision.

2. The method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics according to claim 1, characterized in that, A circular array structure consisting of six omnidirectional microphones is deployed, with each microphone evenly distributed at 60° intervals on a rigid circular support with a diameter of 30cm. The sampling frequency is set to 44100Hz, and the quantization bit depth is 16 bits. (Includes:) Each microphone unit is equipped with an independent preamplifier circuit and an anti-aliasing filter. The preamplifier circuit gain is set to 40 times, and the anti-aliasing filter cutoff frequency is 22050Hz. Simultaneous sampling by six microphones is driven by the same crystal oscillator; The array support is made of rigid aluminum alloy and covered with a layer of sound-absorbing foam; The array system is connected to the central processing unit via a universal serial bus interface, and data transmission adopts an isochronous transmission mode.

3. The method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics according to claim 2, characterized in that, The time difference between the arrival time of the sound source at each microphone is calculated using a beamforming algorithm to construct a three-dimensional sound source localization function. When the localization result converges within a preset range of human oral cavity spatial coordinates and the duration is greater than 0.5 seconds, a cough event capture mechanism is triggered, including: The generalized cross-correlation phase transform method is used to calculate the time delay estimate between any two microphone pairs. Fifteen sets of cross-correlation functions are generated for the six-channel signal. Each set of functions is accelerated by fast Fourier transform, with 1024 transformation points. The time delay estimate is smoothed by a median filter, with a window length of 5 frames. The three-dimensional sound source localization function is constructed based on the principle of spherical intersection. A right-handed Cartesian coordinate system is established with the geometric center of the array as the origin. The sound source position coordinates are solved by the least squares method, and the iterative convergence threshold is set to 0.01m. The preset spatial coordinate range of the human oral cavity is defined as a spherical space with the origin as the center and a radius of 0.8m.

4. The method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics according to claim 3, characterized in that, The adaptive spectral subtraction and Wiener filtering cascaded processing flow is initiated. First, a short-time Fourier transform is performed on the original multi-channel signal to obtain the time-frequency matrix. Then, the noise power spectral density function is estimated. An asymmetric window function is used to dynamically suppress residual noise. The rise time of the window function is 5ms and the fall time is 20ms. This includes: The short-time Fourier transform uses a Hamming window with a window length of 25ms, a frame shift of 10ms, and 1024 transform points, generating a time-frequency matrix with a dimension of the number of frames multiplied by 513. The noise power spectral density function was estimated using the minimum statistics method, with a statistical window length of 3s and an update step size of 0.5s. The adaptive spectral subtraction formula is: ； The Wiener filter transfer function is: ； The expression for the rising segment of an asymmetric window function is: ； The expression for the descent segment is: ； The enhanced signal is reconstructed into a time-domain waveform by inverse short-time Fourier transform and synthesized using the overlapping addition method.

5. The method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics according to claim 4, characterized in that, Automatic cough segmentation was performed on the enhanced audio signal, based on the energy envelope thresholding method. The threshold was set to three times the standard deviation of the local mean energy. The segmentation window length was 200 ms, and the overlap rate was 50%. The segmentation included: The energy envelope is obtained by calculating the sum of squares of the signal in each frame, with a frame length of 25ms and a frame shift of 10ms. The local mean energy is calculated using a sliding window with a window length of 1.5 s and a step size of 0.1 s, and the standard deviation is updated synchronously. The segmentation process employs a dual-threshold detection strategy, with an initial threshold of 3 times the standard deviation and an ending threshold of 1.5 times the standard deviation. After segmentation, the fragments must meet the minimum duration requirement of 10ms and the maximum duration requirement of 500ms. Fragments exceeding these limits are considered invalid and discarded.

6. The method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics according to claim 5, characterized in that, Mel-spectral coefficients of the segmented cough audio are extracted, with an order of 13. Simultaneously, the first eight autocorrelation coefficients of the linear prediction residual spectrum are calculated, and the harmonic energy ratio parameter is superimposed. This parameter is defined as the ratio of the total energy within the fundamental frequency and its second and third harmonic bands to the total energy of the entire frequency band. Then, the transient zero-crossing rate is calculated, which is the normalized value of the number of times the signal crosses zero points per unit time, including: The Mel cepstral coefficient extraction process includes pre-emphasis, framing, windowing, Fourier transform, Mel filter bank mapping, logarithmic compression, and discrete cosine transform. The pre-emphasis coefficient is 0.97, and the number of Mel filter banks is 40. The linear prediction residual spectrum was obtained through tenth-order linear prediction analysis. The prediction coefficients were solved using the autocorrelation method. The first eight autocorrelation coefficients of the residual signal were normalized and used as the eigenvectors. The fundamental frequency is estimated by the autocorrelation method. The fundamental frequency and its second and third harmonic frequency bands are defined as a rectangular window with a bandwidth of 20Hz centered on each harmonic frequency. Before calculating the transient zero-crossing rate, the signal is normalized to zero mean. Zero-crossing detection uses the sign change method. The normalized rate is obtained by counting the number of zero-crossings in each frame and dividing by the frame length. The four types of parameters are aligned according to time frames to form a two-dimensional feature matrix with a dimension of 25 times the number of frames. The Mel cepstrum is 13-dimensional, the linear prediction residual autocorrelation is 8-dimensional, the harmonic energy ratio is 1-dimensional, the transient zero-crossing rate is 1-dimensional, and two additional dimensions are the first-order and second-order difference coefficients.

7. The method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics according to claim 6, characterized in that, The above four types of parameters are concatenated along the time axis into a two-dimensional feature matrix, which is then input into a pathological feature extractor consisting of three parallel convolutional neural network branches. The first branch uses a one-dimensional convolutional kernel with a size of 3 to extract local temporal patterns. The second branch uses a dilated convolutional kernel with a size of 5 and a dilation rate of 2 to capture mid-range dependencies. The third branch uses a max-pooling layer followed by a fully connected layer to extract global statistical features, including: The first branch contains three layers of one-dimensional convolutions, with 32, 64 and 128 kernels in each layer, followed by batch normalization and random deactivation with a probability of 0.2 after each layer. The second branch contains two layers of dilated convolutions. The first layer has 64 kernels and the second layer has 128 kernels. The activation function and regularization configuration are the same as the first branch. The third branch first performs max pooling on the input feature matrix along the time axis, with the pooling window covering all time frames, and outputs a 25-dimensional static feature vector. Then, it passes through two fully connected layers with 256 and 128 neurons respectively, and the random inactivation probability is 0.

3. The three branches output feature vectors with a dimension of 128, which are then concatenated to form a 384-dimensional joint feature vector.

8. The method for diagnosing respiratory diseases by analyzing cough sounds using disease characteristics according to claim 7, characterized in that, The feature vectors output from the three branches are concatenated and input into the disease-specific classification module. This module contains three independent fully connected layers, corresponding to bronchospasm feature detection, lung parenchymal exudation feature recognition, and upper airway obstruction feature discrimination, respectively. The number of neurons in each layer is 128, 96, and 64, respectively. The activation function uses a modified linear unit, including: The bronchospasm feature detection layer receives 384-dimensional input and outputs 128-dimensional features. After activation by a modified linear unit, the features are input to the second layer of 96-dimensional neurons, then activated again and input to the third layer of 64-dimensional neurons. Finally, the asthma probability is output through a soft maximum function. The lung parenchymal exudation feature recognition layer and the upper respiratory tract obstruction feature discrimination layer have the same structure, and output the probabilities of pneumonia and laryngitis respectively. The expression for the feature decoupling regularization term is: ； in Let be the weight matrix of the i-th classifier; It is the Frobenius norm; The training uses a stochastic gradient descent optimizer with an initial learning rate of 0.001, a batch size of 32, and 50 training epochs. The learning rate is reduced to 1 / 10 of its original value every ten epochs.

9. A system for diagnosing respiratory diseases by analyzing cough sounds based on disease characteristics, characterized in that, include: The multi-channel acoustic sensor array module is used to deploy a ring array structure consisting of six omnidirectional microphones. Each microphone is evenly distributed at a 60° interval on a rigid ring support with a diameter of 30cm. The sampling frequency is set to 44100Hz and the quantization bit is 16bit. The sound source localization and event triggering module is used to calculate the time difference of the sound source arriving at each microphone through beamforming algorithm, construct a three-dimensional sound source localization function, and trigger the cough event capture mechanism when the localization result converges within the preset human oral cavity spatial coordinate range and the duration is greater than 0.5s. The adaptive noise suppression module is used to initiate the cascaded processing flow of adaptive spectral subtraction and Wiener filtering. First, it performs a short-time Fourier transform on the original multi-channel signal to obtain the time-frequency matrix. Then, it estimates the noise power spectral density function and uses an asymmetric window function to dynamically suppress noise residue. The rise time of the window function is 5ms and the fall time is 20ms. The automatic cough segment segmentation module is used to automatically segment the cough segment of the enhanced audio signal. The segmentation is based on the energy envelope threshold method, with the threshold set to 3 times the standard deviation of the local mean energy. The segmentation window length is 200ms, and the overlap rate is 50%. The pathological acoustic feature extraction module is used to extract the Mel-frequency cepstral coefficients of the segmented cough audio, with an order of 13 dimensions. At the same time, it calculates the first eight autocorrelation coefficients of the linear prediction residual spectrum and superimposes the harmonic energy ratio parameter, which is defined as the ratio of the total energy in the fundamental frequency and its second and third harmonic frequency bands to the total energy of the entire frequency band. Then, it calculates the transient zero-crossing rate, which is the normalized value of the number of times the signal crosses zero points per unit time. The deep temporal feature encoding module is used to concatenate the above four types of parameters into a two-dimensional feature matrix along the time axis, and input it into the pathological feature extractor composed of three parallel convolutional neural network branches. The first branch uses a one-dimensional convolutional kernel with a size of three to extract local temporal patterns. The second branch uses a dilated convolutional kernel with a size of 5 and an inflation rate of two to capture mid-range dependencies. The third branch uses a max pooling layer followed by a fully connected layer to extract global statistical features. The disease-specific classification module is used to input the concatenated feature vectors from the three branches into the disease-specific classification module. This module contains three independent fully connected layers, which correspond to bronchospasm feature detection, lung parenchymal exudation feature recognition, and upper airway obstruction feature discrimination, respectively. The number of neurons in each layer is 128, 96, and 64, respectively, and the activation function is a modified linear unit. The model training and inference module is used for end-to-end training using the cross-entropy loss function and disease labels. During the training process, a label smoothing strategy is introduced with a smoothing coefficient of 0.

1. At the same time, a feature decoupling regularization term is applied to force each classifier to respond only to its corresponding pathological feature dimension. During the inference phase, the input is the audio of the cough to be tested, and the output is the probability values ​​of three disease categories and the corresponding feature activation heatmap. The horizontal axis of the heatmap is the time frame number, the vertical axis is the feature dimension index, and the color intensity represents the contribution weight of the dimension to the classification decision.

10. The system for diagnosing respiratory diseases by analyzing cough sounds based on disease characteristics according to claim 9, characterized in that, The multi-channel acoustic sensor array module is used for: Each microphone unit is equipped with an independent preamplifier circuit and an anti-aliasing filter. The preamplifier circuit gain is set to 40 times, and the anti-aliasing filter cutoff frequency is 22050Hz. Simultaneous sampling by six microphones is driven by the same crystal oscillator; The array support is made of rigid aluminum alloy and covered with a layer of sound-absorbing foam; The array system is connected to the central processing unit via a universal serial bus interface, and data transmission adopts an isochronous transmission mode.

Images