A lung disease acoustic recognition method and device based on artificial intelligence

Abstract

The application provides a lung disease acoustic recognition method and device based on artificial intelligence. The lung disease acoustic recognition method based on artificial intelligence provided by the application comprises the following steps: acquiring a lung disease acoustic signal, framing the acoustic signal to obtain a plurality of acoustic signal segments; screening effective acoustic segments from the plurality of acoustic signal segments based on an adaptive zero-crossing rate algorithm; and identifying the health risk level corresponding to the effective acoustic segments based on an artificial intelligence model. The lung disease acoustic recognition method and device based on artificial intelligence provided by the application not only assist doctors in clinical diagnosis, but also realize automatic analysis and intelligent recognition of lung disease acoustic signals of different groups of people, and have high efficiency and universality. In addition, the recognition method is non-contact, non-invasive, low-cost, simple to operate, and supports detection and health management anytime and anywhere, realizes early screening and early warning of potential patients, reduces the rate of missed diagnosis, improves the detection rate of diseases, and realizes early diagnosis and prevention of lung diseases.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based method and apparatus for acoustic recognition of lung diseases. Background Technology

[0002] Currently, chronic obstructive pulmonary disease (COPD) has become one of the chronic respiratory diseases with high morbidity and mortality rates worldwide. This disease is characterized by its long course, irreversibility, and high relapse rate, making accurate diagnosis and assessment crucial for its prevention and management. Existing methods for assessing COPD involve measuring the ratio of forced expiratory volume in one second (FEV1) to forced vital capacity (FVC). This requires patients to visit a hospital or specialized medical examination facility, use a dedicated pulmonary function testing machine, and complete breathing exercises under the guidance of a professional technician. Chest imaging and blood gas analysis are often used as auxiliary diagnostic tools. However, existing technologies suffer from poor usability and high costs, requiring expensive specialized equipment and professional personnel, making them difficult to implement in grassroots and remote areas and unsuitable for large-scale screening. Furthermore, the testing methods are inconvenient, leading to low compliance. The testing procedures pose a risk of inconvenience to the elderly, frail individuals, or those with impaired lung function, and improper operation can affect the accuracy of the results. At the same time, there are also problems such as delayed detection and missing the best intervention period. It is difficult to seek medical attention when early symptoms appear. Existing technologies are not sensitive enough to the subtle pathophysiological changes in early COPD, are highly subjective, and are prone to missing the best time for early intervention, which affects subsequent treatment. Summary of the Invention

[0003] In view of this, this application provides an artificial intelligence-based acoustic recognition method and device for lung diseases, which can meet the needs of pre-screening and auxiliary diagnosis that is simple, fast, low-cost, and suitable for a wide range of people.

[0004] Specifically, this application is implemented through the following technical solution:

[0005] The first aspect of this application provides an artificial intelligence-based acoustic recognition method for lung diseases, the method comprising:

[0006] Acoustic signals of lung disease are acquired, and the acoustic signals are framed to obtain multiple acoustic signal segments.

[0007] The effective acoustic segments among the multiple acoustic signal segments are selected based on the adaptive zero-crossing rate algorithm;

[0008] The health risk level corresponding to the effective acoustic segment is identified based on an artificial intelligence model.

[0009] A second aspect of this application provides an artificial intelligence-based acoustic recognition device for lung diseases, comprising a processing module and a generation module; wherein,

[0010] The processing module is used to acquire acoustic signals of lung diseases, and to segment the acoustic signals into frames to obtain multiple acoustic signal segments;

[0011] The processing module is used to filter the effective acoustic segments from the plurality of acoustic signal segments based on an adaptive zero-crossing rate algorithm;

[0012] The generation module is used to identify the health risk level corresponding to the effective acoustic segment based on an artificial intelligence model.

[0013] The AI-based acoustic recognition method and apparatus for lung diseases provided in this application significantly improves the accuracy of AI model recognition results by accurately locating relevant features in the acoustic scene of lung diseases through two levels. On the one hand, an adaptive dynamic threshold zero-crossing rate algorithm is adopted. The baseline for zero crossing and the threshold for zero crossing rate are determined based on the time-domain and frequency-domain variation trends of the pitch and timbre of the acoustic signal of lung diseases. This extends from features at a single time point to features over a period of time, accurately filtering out the corresponding effective signals. On the other hand, during model recognition, a feature sound detection module is added to the AI ​​model. Based on the detection of feature sounds, the key features of lung disease acoustics are located, and the attention weight is dynamically adjusted according to the detection object. This achieves controllable and adaptive attention weights based on the features of the input signal. Once the focus needs to be adjusted, only the detection object of the feature sound detection module needs to be adjusted, without adjusting other model structures and parameters. This further improves the attention to effective features and enhances the accuracy of model recognition. By combining an adaptive zero-crossing rate algorithm with an artificial intelligence model, this method not only automatically filters out high-quality, effective acoustic segments from raw lung disease acoustic signals, avoiding the influence of invalid or interfering signals on subsequent analysis, but also utilizes the artificial intelligence model to extract features and classify these effective acoustic segments, achieving accurate identification of health risk levels. This method reduces reliance on human experience, improves the automation and accuracy of identification, and can be widely applied to the early screening and auxiliary diagnosis of lung diseases. It has the advantages of strong real-time performance, high robustness, and significant clinical application value. Attached Figure Description

[0014] Figure 1 A flowchart of an embodiment of the artificial intelligence-based acoustic recognition method for lung diseases provided in this application;

[0015] Figure 2 A flowchart of Embodiment 2 of the AI-based acoustic recognition method for lung diseases provided in this application;

[0016] Figure 3 A flowchart of Embodiment 3 of the AI-based acoustic recognition method for lung diseases provided in this application;

[0017] Figure 4A flowchart of Embodiment 4 of the AI-based acoustic recognition method for lung diseases provided in this application;

[0018] Figure 5 This is a schematic diagram of the structure of an embodiment of the artificial intelligence-based acoustic recognition device for lung diseases provided in this application. Detailed Implementation

[0019] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0020] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0021] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein can be interpreted as "when," "when," or "in response to a determination." Specific embodiments are given below to describe the technical solutions of this application in detail.

[0022] Figure 1 This is a flowchart of an embodiment of the artificial intelligence-based acoustic recognition method for lung diseases provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0023] S101. Acquire acoustic signals of lung disease, and divide the acoustic signals into frames to obtain multiple acoustic signal segments.

[0024] Specifically, pulmonary acoustic signals are audio signals related to lung diseases collected by specific devices (such as pulmonary function analyzers, stethoscopes, or other medical testing equipment). Examples include low-frequency noise in coughing sounds, and persistent wheezing and crackling sounds in breathing. These signals can reflect the pathological state of the lungs. Framing involves dividing the signal into multiple short segments for subsequent feature extraction and classification. Acoustic signal segments are sub-parts of the aforementioned pulmonary acoustic signals.

[0025] In this specific implementation, COPD patients exhibit specific acoustic characteristics in their coughing, breathing, and speaking acoustic signals due to pathophysiological changes such as airway narrowing, inflammation, and increased mucus secretion. The user's acoustic signals are collected using a high-sensitivity acquisition device. To ensure data quality, the acquisition device can be a smartphone with a built-in microphone or a dedicated acquisition unit, capable of covering the respiratory frequency range of 20 Hz to 2 kHz, ensuring the capture of subtle abnormal acoustic features during coughing, breathing, and speaking in COPD patients. For example, the user completes a series of vocalization tasks, such as forceful coughing, deep breathing, wheezing, and reading aloud, using an application on a smartphone or dedicated acquisition device.

[0026] Furthermore, after acquiring the acoustic signals, preprocessing is performed. This preprocessing includes noise reduction, silent segment removal, normalization, and framing of the lung disease acoustic signals to obtain multiple acoustic signal segments. Preprocessing is used to eliminate environmental noise and individual vocal differences, making the effective segments more prominent and facilitating subsequent model recognition.

[0027] In practice, for example, the acquired audio signal is securely uploaded to a cloud server. The server performs preprocessing on the audio, such as noise reduction, silence removal, and signal enhancement, to eliminate environmental interference and segment out effective acoustic segments (such as a single cough sound).

[0028] Furthermore, the acoustic signal of lung disease exhibits non-stationary characteristics overall, but it can be approximated as a stationary process within a short time range. Therefore, dividing the long-term acoustic signal into multiple acoustic signal segments by framing can effectively preserve local statistical features.

[0029] Understandably, framing not only improves the accuracy of signal analysis but also reduces computational complexity, making subsequent feature extraction and model training more efficient, while enhancing the ability to monitor lung diseases.

[0030] Specifically, the steps for dividing the acoustic signal into frames to obtain multiple acoustic signal segments are as follows:

[0031] Step 1: Identify the volume change trend of the acoustic signal of the lung disease.

[0032] Specifically, the volume change trend is the overall change pattern of the acoustic signal volume over time. By calculating the volume change in each time period, the fluctuation trend of the volume change can be identified, that is, the increase or decrease of the volume, so as to avoid interference from instantaneous noise and provide an analytical basis for subsequent identification of abrupt change points.

[0033] Furthermore, by calculating short-time energy from the acoustic signal in frames, and performing envelope extraction and smoothing, the volume variation trend of the signal over time can be obtained. Acoustic signals from lung diseases are acquired using equipment, and bandpass filtering and amplitude normalization are applied to eliminate interference from environmental noise and equipment differences. Frames are created with fixed frame lengths and frame shifts, and windowing is used to maintain short-time stationarity. Logarithmic compression and smoothing filtering are performed on the volume sequence, and the envelope curve of the signal is extracted to obtain the volume variation trend of the acoustic signals from lung diseases.

[0034] Step 2: Mark abrupt change points according to the volume change trend, where the difference in acoustic signal volume before and after the abrupt change point is greater than a preset threshold.

[0035] Specifically, abrupt change points refer to moments when the signal volume changes drastically, reflecting certain important pathological characteristics in the acoustic signal (such as wheezing and crackling sounds). These abrupt change points can be identified by analyzing the changes in volume and comparing them with preset thresholds.

[0036] Furthermore, an acceleration threshold is determined based on the upper and lower limits of the volume of the lung disease acoustic signal. The lung disease acoustic signal is segmented using the duration interval of the lung disease characteristic signal as a unit. The rate of change of each segment is calculated. The mutation interval is determined based on the rate of change and the acceleration threshold. The point with the largest rate of change within the mutation interval is taken as the mutation point.

[0037] It should be noted that this step analyzes the rate of change of volume in the acoustic signals of lung diseases and, in conjunction with an acceleration threshold, identifies abrupt change intervals and points within the acoustic signals to achieve precise localization of pathological acoustic features. The acceleration threshold limits the sensitivity to volume changes, thus distinguishing between stable states and abrupt changes. The acoustic signals of lung diseases are segmented into time intervals, dividing long-duration audio into shorter segments. This reduces noise interference and facilitates precise analysis of local abnormal events while maintaining processing efficiency. When the rate exceeds the threshold, an abrupt change interval is identified within that segment, corresponding to a sharp change in signal energy. When the rate exceeds the acceleration threshold, the abrupt change interval is automatically identified without manual annotation. Within the abrupt change interval, the time point with the highest rate of change is selected as the abrupt change point, which serves as the time boundary for subsequent acoustic signal segmentation and feature extraction, ensuring accurate localization and analysis of pathological acoustic features.

[0038] Specifically, a large number of acoustic signals from lung diseases in target lung disease patients are collected. The maximum value is used as the upper limit of the acceleration threshold, and the minimum value as the lower limit, to obtain the volume range. Further, the rate of change of each segment is calculated. The volume difference between two consecutive time points is selected and divided by the time difference between the two time points to obtain the volume change per unit time. The formula for calculating the rate of change is as follows:

[0039]

[0040] in, Indicates the rate of change; and This indicates the volume at two different points in time. and This indicates two points in time.

[0041] Furthermore, the mutation interval is determined based on the rate of change and the acceleration threshold. When the rate of change is greater than the acceleration threshold, the point is marked as a candidate mutation point. When there are consecutive candidate mutation points, they are merged into a mutation interval. Among them, the point with the largest rate of change in the mutation interval is taken as the mutation point.

[0042] In a specific implementation, for example, on the volume change trend curve, if the volume measured at the front window of 100 ms is 10 dB and the volume measured at the back window of 100 ms is 3 dB, then the volume difference is 7 dB. If the preset threshold is 5 dB, then the volume difference is greater than the preset threshold, and the start time of the back window is the abrupt change point.

[0043] Understandably, setting a preset threshold to prevent the selection of non-mutation points is beneficial to improving the accuracy of subsequent feature extraction and model training, reducing the interference of irrelevant noise on model learning, and thus improving the reliability and accuracy of disease identification.

[0044] Step 3: Calculate the frequency change trend of the acoustic signal of the lung disease, align the frequency change trend with the volume change trend, and determine the target point corresponding to the mutation point on the frequency change trend.

[0045] Specifically, the frequency change trend is the pattern of how the acoustic signal frequency changes over time. Aligning the frequency change pattern with the volume change trend ensures that the volume and frequency correspond at the same point in time, allowing us to find the target point on the frequency change trend corresponding to the corresponding abrupt change in volume.

[0046] It should be noted that the time of the abrupt change point determined in the time domain is mapped onto the time axis of the frequency change trend, thus aligning the time of the volume change point with the time of the frequency change trend. Within the frequency change trend, the frequency value corresponding to the mapped time point is taken as the target point.

[0047] In practice, for example, when the acoustic signal of lung disease shows a sudden change at 1.2 seconds, the frequency value corresponding to the same time point will be found in the frequency change trend and marked as the target point, thereby determining the target point for subsequent risk assessment.

[0048] Specifically, acoustic signals from lung diseases not only exhibit abrupt changes in volume, but also often show significant frequency shifts or energy concentrations in frequency distribution. For example, wheezing typically manifests as a narrow band of sustained high-frequency components, while crackles are characterized by transient, broadband high-frequency components. Therefore, by performing frequency analysis on acoustic signals from lung diseases to obtain frequency change trends and aligning them with pre-marked volume change trends, the corresponding positions of abrupt change points in the frequency dimension can be effectively verified and located.

[0049] Understandably, linking frequency change trends with volume change trends breaks through the limitations of a single feature and provides a basis for subsequent refinement of breakthrough points.

[0050] Step 4: Correct the abrupt change point based on the frequency change state of the target point.

[0051] Specifically, correcting abrupt change points involves eliminating those that do not meet the criteria based on frequency changes. Abrupt change points in the acoustic signals of lung diseases may exhibit temporal shifts in both volume and frequency. For example, some popping sounds may show an instantaneous peak in volume, but a subsequent energy expansion in frequency distribution. Relying solely on volume trends to label abrupt change points can easily lead to localization errors. In practical implementation, for instance, the volume difference in the acoustic signal may exceed a preset threshold, but the frequency change trend may not meet pathological significance.

[0052] Furthermore, the time interval of the lung disease characteristic signal is used as the unit, and the target point is used as the center to divide it to obtain the correction interval. It is determined whether the rate of change of the target point in the correction interval is the largest in the correction interval. If the determination is true, the point is used as the target point for correction. If the determination is false, the next target point is taken until the target point with the largest rate of change in the interval is taken.

[0053] It should be noted that, around the initially determined target point, a local time window is defined, and segments within this time window are extracted as the correction interval. The rate of change of all points within the correction interval is iterated, and the maximum value is found, which is then retained as the target point.

[0054] Understandably, combining the pathological specificity of frequency can filter out meaningless mutation points, avoid misdiagnosis, and significantly improve the clinical reliability of mutation points.

[0055] Step 5: Use the corrected mutation point as a reference to frame the acoustic signal of lung disease.

[0056] Specifically, framing involves dividing a continuous acoustic signal into multiple acoustic signal segments over time. Breaking down complex, long signals into shorter segments reduces the difficulty of subsequent analysis. By using corrected abrupt change points as a benchmark for framing, it ensures that each frame covers the start and end range of abrupt change points in the acoustic signals associated with lung diseases. Simultaneously, framing enables structured classification, providing standardized input for automated diagnosis.

[0057] Understandably, efficiently extracting acoustic signal fragments from acoustic signals of lung diseases not only reduces the complexity of subsequent pathological feature analysis, but also provides standardized input for subsequent lung disease diagnosis and analysis.

[0058] S102. Adjust the multiple acoustic signal segments based on the adaptive zero-crossing rate algorithm to obtain multiple effective acoustic segments.

[0059] Specifically, the adaptive zero-crossing rate algorithm determines the zero-crossing baseline of the zero-crossing rate based on the detection environment acoustic signal of the lung disease acoustic signal, and determines the zero-crossing rate threshold corresponding to the acoustic signal segment based on the fluctuation change of the acoustic signal segment. Each valid acoustic segment includes a complete acoustic event.

[0060] Furthermore, the adaptive zero-crossing rate algorithm dynamically adjusts the zero-crossing rate threshold based on the actual characteristics of the acoustic signal segments, filtering out effective acoustic segments related to lung pathology. In lung disease acoustic signals, coughing, wheezing, or other abnormal breathing sounds are often accompanied by high-frequency vibrations, leading to a significant increase in the zero-crossing rate. By analyzing the zero-crossing rate of each acoustic signal segment, effective acoustic segments are filtered out. The adaptive zero-crossing rate algorithm further dynamically adjusts the threshold based on segment characteristics, avoiding misjudging non-pathological segments as effective signals due to noise or slight fluctuations. It should be noted that the adaptive zero-crossing rate algorithm can dynamically adjust the standard according to the actual situation, avoiding both missed detection of weak pathological signal segments due to an excessively high fixed threshold and misjudgment due to an excessively low fixed threshold, ensuring the selection of effective segments and providing a reliable data foundation for subsequent pathological analysis.

[0061] Step 1: Obtain the detection environment acoustic signal of the lung disease acoustic signal, and determine the zero-crossing reference line of the zero-crossing rate based on the amplitude of the environment acoustic signal.

[0062] Specifically, the zero-crossing baseline is a reference value determined based on the statistical results of detected environmental acoustic signals. It is used to correct the shortcomings of traditional zero-crossing rate calculations that use absolute zero as the reference. This step determines the zero-crossing baseline by detecting the amplitude characteristics of the environmental acoustic signal, thus adjusting the zero-crossing judgment benchmark from absolute zero to the typical amplitude level of environmental noise.

[0063] In practice, for example, a 5-10 second pure detection ambient acoustic signal is collected, the collected ambient acoustic signal is preprocessed, and the amplitude statistical characteristics of the ambient acoustic signal are used as the core basis to set a zero-crossing baseline.

[0064] Step 2: Determine the zero-crossing rate threshold corresponding to the acoustic signal segment based on the fluctuation changes of a preset number of adjacent acoustic signal segments and the zero-crossing baseline;

[0065] Determining the zero-crossing rate threshold corresponding to an acoustic signal segment based on the fluctuation changes of a preset number of adjacent acoustic signal segments and the zero-crossing reference line includes: determining a first acoustic signal segment preceding the acoustic signal segment and a second acoustic signal segment following the acoustic signal segment; determining the amplitude change curves of the first acoustic signal segment and the second acoustic signal segment; predicting the change curve of the current acoustic signal segment based on the difference between the amplitude change curves; and adjusting the zero-crossing reference line based on the difference between the change curve of the current acoustic signal segment and the actual change curve.

[0066] Specifically, fluctuation changes can be obtained by calculating the instantaneous amplitude, energy change, or volume difference between adjacent segments. By statistically analyzing fluctuation changes, a zero-crossing rate threshold can be used to distinguish between normal and abnormal fluctuations, thus determining whether an acoustic signal segment contains a valid acoustic segment.

[0067] Specifically, the zero-crossing rate threshold, used to determine whether an acoustic signal segment is valid, needs to be dynamically calculated in conjunction with the acoustic segment to adapt to different scenarios under various conditions. Compared to fixed threshold methods, adaptive thresholds can better adapt to the amplitude and frequency variations of acoustic signals in different segments, thus improving recognition accuracy.

[0068] Understandably, by referencing the fluctuations of adjacent segments, the zero-crossing rate threshold can be dynamically adjusted to avoid noise misjudgment and missed detection of abnormal events.

[0069] Step 3: For any acoustic signal segment, calculate the zero-crossing rate of the arbitrary acoustic signal segment.

[0070] Specifically, the zero-crossing rate is the number of times an acoustic signal waveform changes from a positive amplitude to a negative amplitude, or from a negative amplitude to a positive amplitude, within a single signal frame. It reflects the frequency characteristics and noise level of the signal. For example, high-frequency signals typically have a higher zero-crossing rate, while the zero-crossing rate of stationary noise is relatively stable, and the zero-crossing rate of meaningless static signals is extremely low or close to zero.

[0071] In practical implementation, for example, if zero crosses 15 times within a 10ms segment, the zero crossing rate is 1.5 times / ms (15 times / 10ms = 1.5 times / ms).

[0072] Step 4: Filter valid acoustic segments based on the zero-crossing rate threshold and the numerical relationship between the zero-crossing rates.

[0073] Specifically, by comparing the zero-crossing rate with a threshold value, valid acoustic segments are selected, and invalid segments with excessively high or low zero-crossing rates are removed. For each acoustic signal segment, the zero-crossing rate of that segment is compared with a threshold. If the zero-crossing rate is greater than the threshold, the segment is marked as a valid acoustic segment. If the zero-crossing rate is less than or equal to the threshold, it is marked as an invalid segment.

[0074] In practice, for example, if the zero-crossing rate of the current segment is 1.6 times / ms (16 times / 10ms = 1.6 times / ms) and the zero-crossing rate threshold is 1.45 times / ms (14.5 times / 10ms = 1.45 times / ms), the two are compared. If the zero-crossing rate is greater than the threshold, it is determined to be a valid acoustic segment; if the zero-crossing rate is 0.9 times / ms (9 times / 10ms = 0.9 times / ms) and the threshold remains unchanged, it is determined to be an invalid segment.

[0075] In a specific implementation, in one possible approach, to eliminate the interference of ambient background noise on the volume analysis of lung disease acoustic signals, an analysis of the ambient background noise will be performed before acquiring the lung disease acoustic signals. The ambient sound signal (e.g., indoor noise when no patient is speaking, equipment background noise) is collected for a continuous duration of no less than 2 seconds. The average volume value of the collected ambient sound signal is calculated. The volume value at each sampling point is corrected to obtain the corrected volume value. ( Following the steps above, substitute the corrected volume value into the calculation to obtain the corrected calculation method.

[0076] Understandably, this can quickly remove invalid segments, significantly reduce redundant calculations in subsequent feature extraction, disease detection, and other processes, and improve the overall efficiency of the workflow.

[0077] S103. Identify the health risk level corresponding to the effective acoustic segment based on an artificial intelligence model.

[0078] Specifically, the artificial intelligence model detects acoustic events associated with lung diseases and adjusts the attention weights of corresponding effective acoustic segments based on these associated acoustic events.

[0079] Furthermore, the artificial intelligence model is a deep learning model that correlates acoustic signals with health risk levels. The health level is a quantitative classification of the severity of potential health problems based on acoustic signal characteristics. It can output risk levels in real time, allowing time for subsequent intervention and facilitating early disease detection and treatment. Attention weights are importance coefficients assigned to different segments of the acoustic signal. Through the mapping relationship between effective acoustic segments and health risk levels, the health risks corresponding to each segment can be automatically identified.

[0080] Furthermore, the artificial intelligence model achieves the mapping of effective acoustic segments to health risk levels through multi-level processing.

[0081] (1) Convolutional layer: The time spectrum of the effective acoustic segment is used as input. The convolutional neural network is used to perform convolution operation on the time spectrum to extract the local acoustic features corresponding to each effective acoustic segment.

[0082] Specifically, a time-frequency spectrogram is an image that corresponds to time and frequency, reflecting the process of frequency change over time. A convolutional neural network (CNN) is a deep learning model for processing two-dimensional images, composed of multiple convolutional layers, pooling layers, etc., used to automatically extract local features. Convolution is the core computation of a CNN; by sliding the convolution kernel across the Mel spectrogram, it performs an inner product operation with the pixel values ​​of the corresponding region to obtain local acoustic features. The short-time Fourier transform (SFT) divides the continuous acoustic signal into a series of overlapping short time windows, performs a Fourier transform on each window individually, and finally arranges the frequency domain results of all windows in time to form the Mel spectrogram.

[0083] Specifically, the time-spectrum map combines the temporal and frequency information of the acoustic signal, enabling the model to capture instantaneous acoustic events and frequency patterns. The Mel spectrogram generated by the short-time Fourier transform is used as input to construct an input tensor. This input tensor is fed into a convolutional layer, where the convolutional kernel slides along both the time and frequency dimensions, performing a weighted summation operation on local regions to extract local acoustic features. The feature map output by the convolutional layer retains information about local acoustic patterns and can be used for subsequent pooling, activation, and classification operations. Repeated convolutional operations are used to extract higher-level feature representations, from low-level spectral primitives to more complex combinations of acoustic structures.

[0084] In practice, for example, it is trained using supervised learning based on acoustic data from tens of thousands of patients with chronic obstructive pulmonary disease diagnosed by pulmonary function testing and healthy controls. The convolutional layer uses 32 convolutional kernels of size 3×3. Each kernel slides on the Mel spectrogram, performs convolution operations on local regions, processes them with the ReLU activation function, and then compresses the dimensions using 2×2 max pooling, ultimately outputting a 32-channel local acoustic feature map.

[0085] Understandably, the feature extraction capabilities of convolutional neural networks overcome the limitations of manually defined features. They can capture local areas that are easily overlooked by traditional methods through convolution operations, thus laying a solid data and feature foundation for the entire artificial intelligence model to accurately identify health risk levels from the source.

[0086] (2) Attention weight layer: detect lung disease-related acoustic events, take the lung disease-related acoustic events as input, locate the target effective acoustic segments corresponding to the related acoustic events, calculate the similarity between the query vector and the key vector, convert the obtained similarity value into an attention value, and the attention value of the target effective acoustic segment is greater than that of other effective acoustic segments.

[0087] Specifically, the attention weight layer is used to focus the model on the more important parts of the task and give them higher attention, while reducing the influence of irrelevant information. The query vector is the target template that needs to be focused on, a reference vector used to filter key information. The key vector is the label of all information to be evaluated, a remapping of local acoustic features, consistent with the query vector, to facilitate the calculation of the matching degree. The similarity is a quantified value of the degree of matching between the query vector and the key vector, calculated through mathematical operations. The attention value is the value obtained after normalizing the similarity, used to assign high weights to highly relevant features and low weights to low-relevance features.

[0088] In this step, local acoustic features are used as input and linearly mapped to generate query vectors and key vectors, resulting in attention values. High attention values ​​correspond to pathological acoustic features, while low attention values ​​correspond to irrelevant features, thus optimizing the model. Specifically, for example, suppose the AI ​​model takes three input local acoustic features, corresponding to key vectors K1, K2, and K3, respectively. Simultaneously, a query vector Q is generated. The dot product similarity between the query vector Q and each key vector is calculated, yielding similarities of 0.2, 0.6, and 1.0, respectively. Subsequently, these scores are normalized using the softmax function, resulting in corresponding attention values ​​of 0.12, 0.26, and 0.62.

[0089] Specifically, the core purpose of the attention weighting layer is to enable the model to automatically focus on the local acoustic features most important to the task, thereby enhancing the expressive power of key acoustic patterns. By calculating the similarity between the query vector and the key vector, an importance score, i.e., an attention value, is obtained for each local feature. This attention value is then applied to the corresponding feature to amplify key acoustic signals and suppress noise or non-key features, achieving feature-weighted optimization.

[0090] In practice, multiple local acoustic feature segments F1, F2, ..., F are extracted. n And the corresponding attention values ​​α1, α2, ..., α are calculated through the attention mechanism. n Next, element-wise weighting is performed on each local acoustic feature segment to obtain the weighted result, which is then summed.

[0091] Understandably, by assigning high weights to key pathological features such as wheezing and crackling sounds based on similarity, while suppressing background noise, the discriminative power of local acoustic features is enhanced.

[0092] (3) Recurrent layer, which uses the local acoustic features fused by the attention mechanism as input to the long short-term memory network to obtain temporal dynamic features.

[0093] Specifically, the attention mechanism is a weighted feature selection method used to focus the model on disease-related local acoustic features. Long Short-Term Memory (LSTM) networks, through gating mechanisms including input, forget, and output gates, are used to capture the dynamic evolution of local acoustic features over time. Local acoustic features are extracted from Mel spectrograms, including, for example, the decay pattern of a cough or the wheezing cycle in breathing sounds. Temporal dynamic features reflect the evolution trend of acoustic signals over time, including the occurrence time, duration, interval patterns, intensity, and frequency trends of abnormal sounds.

[0094] In practice, for example, the forward long short-term memory network scans the sequence from frame 1 to frame 215 to capture the temporal evolution from early to late stages; the reverse long short-term memory network scans from frame 215 to frame 1 to capture the dependencies from late to early stages. Each layer of the long short-term memory network processes the data through a gating mechanism: the forget gate filters out redundant features of normal breathing sounds from frames 1 to 50; the input gate stores the first wheezing feature (enhanced by attention) from frames 51 to 60 into the cell state; when processing the second wheezing from frames 101 to 110, the cell state retains the time stamp of the first wheezing, and the output gate calculates the interval between the two as 41 frames (101-60 = 41 frames).

[0095] Understandably, compared to traditional feature extraction methods, Long Short-Term Memory (LSTM) networks can retain richer evolutionary details, significantly improving the ability of artificial intelligence models to perceive dynamic changes in diseases.

[0096] (4) Fully connected layer and output layer, taking temporal dynamic features as input, predict the classification probability of the temporal dynamic features output.

[0097] Specifically, a fully connected layer is a network structure composed of multiple layers of neurons, with each neuron fully connected to all neurons in the layer above it. The output layer is the final output unit of the model, receiving the fused features from the fully connected layers and converting them into probability distributions for various health risk levels through an activation function.

[0098] Furthermore, the output temporal dynamic features are input into a fully connected layer, which performs a linear transformation on them and uses an activation function, such as the softmax activation function, to convert the result into classification probabilities. Based on the category corresponding to the classification probability, the health risk level of the valid acoustic segment is determined. For example, when the probability value of a popping sound in an acoustic segment is the highest and exceeds a preset threshold, the segment is determined to correspond to a higher risk level of lung disease.

[0099] Understandably, by introducing convolutional neural networks, attention mechanisms, and long short-term memory networks into the artificial intelligence model, it is possible not only to fully extract the local features and global temporal dynamic information of acoustic signals, but also to automatically focus on pathological acoustic abnormalities under the action of the attention mechanism, thereby significantly improving the model's accuracy and robustness in identifying abnormal acoustic segments.

[0100] Specifically, in this embodiment, identifying the health risk level corresponding to the effective acoustic segment based on an artificial intelligence model includes:

[0101] (1) Receive the classification probability output by the artificial intelligence model.

[0102] Specifically, the classification probability reflects the model's confidence in a feature belonging to a pathological sound. By setting a pathological label matching threshold (e.g., 0.5), the model's probability output can be transformed into a definite classification result. When the probability exceeds the threshold, it is determined to be an abnormal sound, thus mapping the continuous probability space to a discrete pathological label space.

[0103] Specifically, the corresponding classification probability will be output through an activation function. In a concrete implementation, for example, using the Softmax activation function, for an input vector z=[z1,z2,...,z...]... K The softmax output vector σ(z) = [σ1, σ2, ..., σz] K Each element of ] is defined as:

[0104]

[0105] in, It is the i-th element of the input vector; the output is... Indicates category The predicted probability; .

[0106] Understandably, by using classification probabilities and preset thresholds to make judgments, the automatic identification of abnormal sounds is achieved. This method can map continuous probabilities to clear pathological judgments, reducing the risk of missed and false detections, while making the detection results controllable and reliable, facilitating subsequent analysis and processing.

[0107] (2) Determine the risk level based on the classification probability and the preset threshold.

[0108] Specifically, by combining the classification probabilities output by the artificial intelligence model with preset thresholds, continuous probability values ​​can be converted into specific health risk levels. This determines whether an acoustic signal meets a certain pathological risk standard, thereby achieving quantitative risk assessment. In practice, for example, if the threshold is set to 0.6 and the classification probability is 0.65, then the risk level corresponding to the acoustic signal is determined to be unhealthy, because its probability 0.65 ≥ 0.6.

[0109] (3) Generate the screening report and send the screening report to the user terminal for display.

[0110] Specifically, the system receives the classification probabilities and determined risk levels output by the artificial intelligence model. It then standardizes the description of the risk levels and sends the generated screening report to user terminals (such as mobile apps, PCs, and hospital management systems) via a secure communication channel. This visualizes and makes the model analysis results more understandable, improving users' awareness and management efficiency regarding health risks.

[0111] Understandably, this involves converting continuous probability values ​​into specific health risk levels to achieve quantitative risk assessment. Simultaneously, the assessment results are generated into a visual screening report and sent to the user's terminal, enabling remote real-time monitoring and improving the user's understanding and management of their health status.

[0112] The AI-based acoustic recognition method for lung diseases provided in this embodiment accurately locates relevant features in the acoustic scene of lung diseases at two levels, greatly improving the accuracy of the AI ​​model's recognition results. On the one hand, an adaptive dynamic threshold zero-crossing rate algorithm is adopted. The baseline for zero crossing and the threshold for zero crossing rate are determined based on the time-domain and frequency-domain variation trends of the pitch and timbre of the lung disease acoustic signal. This extends from features at a single time point to features over a period of time, accurately filtering out the corresponding effective signals. On the other hand, during model recognition, a feature sound detection module is added to the AI ​​model. Based on the detection of feature sounds, the key features of lung disease acoustics are located, and then the attention weight is dynamically adjusted according to the detection object. This achieves controllable and adaptive attention weights based on the features of the input signal. Once the focus needs to be adjusted, only the detection object of the feature sound detection module needs to be adjusted, without adjusting other model structures and parameters. This further improves the attention to effective features and enhances the accuracy of model recognition. By combining an adaptive zero-crossing rate algorithm with an artificial intelligence model, this method not only automatically filters out high-quality, effective acoustic segments from raw lung disease acoustic signals, avoiding the influence of invalid or interfering signals on subsequent analysis, but also utilizes the artificial intelligence model to extract features and classify these effective acoustic segments, achieving accurate identification of health risk levels. This method reduces reliance on human experience, improves the automation and accuracy of identification, and can be widely applied to the early screening and auxiliary diagnosis of lung diseases. It has the advantages of strong real-time performance, high robustness, and significant clinical application value.

[0113] Figure 2 This is a flowchart of Embodiment 2 of the AI-based acoustic recognition method for lung diseases provided in this application. Please refer to... Figure 2 The lung disease acoustic recognition method based on artificial intelligence provided in this embodiment includes:

[0114] S201. Use the relevant statistical features extracted from the time-spectrum graph by the convolutional neural network as the query vector, and use the relevant statistical features as the key vector.

[0115] Understandably, this design can make full use of the local acoustic feature information extracted by the convolutional neural network, so that the model considers its own features and maintains global consistency when calculating the attention value, thereby improving the sensitivity and recognition accuracy of pathological features, while reducing interference from irrelevant features, improving the overall detection performance and model robustness.

[0116] The specific implementation principle and process of this step are described in detail in the above embodiments, and will not be repeated here.

[0117] S202. Calculate the similarity between the query vector and the key vector, and convert the similarity into an attention value.

[0118] Specifically, the similarity calculation includes dot product operations and normalization, ensuring that when local acoustic features contain pathological acoustic features such as wheezing or crackling sounds, their similarity to the query vector is high, resulting in a larger attention value, while the attention value for non-disease-specific features is smaller. This mechanism enables the model to automatically focus on key pathological information in acoustic data, improving recognition accuracy and robustness.

[0119] Specifically, the implementation principle and process of this step are detailed in the above embodiments and will not be repeated here.

[0120] Understandably, the model can automatically learn and focus on pathological features in acoustic signals that are truly diagnostically valuable without human intervention. By constructing an attention matrix, the model assigns higher weights to features containing key abnormal information such as wheezing or crackles, thereby effectively improving the discriminative power and expressive ability of pathological signals.

[0121] Figure 3 This is a flowchart of Embodiment 3 of the AI-based acoustic recognition method for lung diseases provided in this application. Please refer to... Figure 3 The lung disease acoustic recognition method based on artificial intelligence provided in this embodiment includes:

[0122] S301. Determine multiple associated acoustic events of lung disease and their corresponding temporal relationships, with each acoustic event having different acoustic characteristics.

[0123] Specifically, associated acoustic events are acoustic events directly related to pathological features. Temporal relationships refer to the chronological order of these events.

[0124] Furthermore, based on the acoustic signals of the lung disease, multiple associated acoustic events are identified, each with different acoustic characteristics, such as wheezing, crackling, or coughing. Different pathological events manifest as specific frequency, amplitude, or zero-crossing rate patterns in the acoustic signals. By analyzing these characteristics, event types can be distinguished, and the time sequence of event occurrence can be determined.

[0125] Understandably, correlating multiple related acoustic events with temporal relationships can fully capture the key differences between the temporal variation patterns of acoustic signals and pathological characteristics, improve the accuracy and robustness of abnormal sound detection, and enable the model to more effectively distinguish between disease-related signals such as wheezing and crackling sounds and normal acoustic fluctuations.

[0126] S302. Based on the temporal relationship, locate the first target valid acoustic segment corresponding to the first associated acoustic event.

[0127] Specifically, based on temporal relationships, the first target segment of the first associated acoustic event is determined as the starting point of the time series, and the first valid target acoustic segment is filtered by traversing the time series until the first segment that meets the conditions is found.

[0128] It should be noted that the specific implementation principle and process of selecting the first target valid acoustic segment can be found in the above content, and will not be repeated here.

[0129] S303. Based on the first target effective acoustic segment and the timing relationship, locate the subsequent target effective acoustic segment.

[0130] Specifically, after identifying the first effective acoustic segment, the system continues to traverse sequentially according to the temporal relationship to locate subsequent effective acoustic segments. Through temporal localization, the first effective segment and subsequent effective segments are concatenated into a continuous signal sequence, avoiding feature loss due to isolated segments.

[0131] It should be noted that the specific implementation principles and processes for screening effective acoustic segments can be found in the above content, and will not be repeated here.

[0132] Understandably, by determining the temporal relationship of multiple acoustic events, it is possible to ensure that events with different pathological characteristics are covered, and to avoid missing key acoustic segments due to signal complexity.

[0133] Figure 4 The flowchart for Embodiment 4 of the AI-based acoustic recognition method for lung diseases provided in this application is shown below. Please refer to... Figure 4 The lung disease acoustic recognition method based on artificial intelligence provided in this embodiment includes:

[0134] S401. Determine the first associated acoustic event among multiple associated acoustic events of lung disease based on temporal relationships.

[0135] Specifically, based on the temporal relationship of the acoustic signals related to lung disease, the first event occurring among multiple associated acoustic events of lung disease is determined. The first associated acoustic event corresponds to the earliest weak signal that is directly related to COPD pathology, such as the first pathological cough occurring in an early-stage patient within 0.5 seconds.

[0136] Furthermore, by focusing on the earliest weak signal, abnormalities can be missed due to focusing on subsequent strong signals, thereby improving the ability to identify pathological events.

[0137] S402. Calculate the maximum time threshold between each subsequent associated acoustic event and the previous associated acoustic event.

[0138] Specifically, the maximum time threshold is the largest time interval between two adjacent associated acoustic events, used to distinguish between continuous pathological events and discontinuous events. If the interval between two associated events exceeds this threshold, it can be determined as a discontinuous pathological event (such as two coughs with too long an interval, one of which may be a pathological signal and the other a normal throat clearing), and should be excluded from subsequent time intervals.

[0139] In practice, for example, all statistically significant adjacent event intervals are sorted from smallest to largest, and the interval value corresponding to the 95th percentile after sorting is taken as the longest time threshold. Extreme values ​​(such as intervals > 30 seconds) are excluded and determined to be non-continuous pathological events, and are not included in the statistics.

[0140] S403. Locate the first target effective acoustic segment, and determine the candidate effective acoustic segments by taking the first target effective acoustic segment as the starting point and the sum of all the longest time thresholds as the time interval.

[0141] Specifically, by using the first target segment as the starting point for statistical thresholding, a time interval can be formed, which can then be used to filter subsequent valid segments.

[0142] In practical implementation, for example, all acoustic segments corresponding to the first identified associated acoustic event (e.g., timestamps 1.2–1.22 seconds, 1.21–1.23 seconds, ... 1.7–1.72 seconds) are selected. The segment with the smallest starting timestamp (e.g., 1.2–1.22 seconds) is chosen as the first target valid acoustic segment. The time interval is defined by taking the starting timestamp of the first target valid acoustic segment as the starting point. As the endpoint, among which This is the threshold for the longest time.

[0143] S405. Select other valid target valid acoustic segments from the candidate valid acoustic segments.

[0144] Specifically, among the candidate valid acoustic segments, the first target valid acoustic segment is removed, and only candidate segments with timestamps greater than the first target segment are retained. This avoids redundant filtering and ensures that valid acoustic segments of subsequent related events can be extracted accurately and sequentially.

[0145] Understandably, by calculating the longest time threshold between adjacent events and forming candidate time intervals starting from the first effective acoustic segment of the target, it is possible to effectively distinguish between continuous and discontinuous pathological events, eliminate irrelevant signals with excessively long intervals, and improve the accuracy of subsequent event extraction.

[0146] Figure 5 This is a schematic diagram of the structure of an embodiment of the artificial intelligence-based acoustic recognition device for lung diseases provided in this application. Please refer to... Figure 5 The apparatus provided in this embodiment includes a processing module 510 and a generation module 520; wherein,

[0147] The processing module 510 is used to acquire acoustic signals of lung diseases, and to segment the acoustic signals into frames to obtain multiple acoustic signal segments;

[0148] The processing module 510 is used to adjust the plurality of acoustic signal segments based on an adaptive zero-crossing rate algorithm to obtain a plurality of effective acoustic segments; wherein, the adaptive zero-crossing rate algorithm determines the zero-crossing baseline of the zero-crossing rate based on the detection environment acoustic signal of the lung disease acoustic signal, and determines the zero-crossing rate threshold corresponding to the acoustic signal segment based on the fluctuation change of the acoustic signal segment, and each effective acoustic segment includes a complete acoustic event;

[0149] The generation module 520 is used to identify the health risk level corresponding to the effective acoustic segment based on an artificial intelligence model, wherein the artificial intelligence model detects lung disease-related acoustic events and adjusts the attention weight of the corresponding effective acoustic segment based on the related acoustic events.

[0150] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0151] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0152] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An artificial intelligence-based acoustic recognition method for lung diseases, characterized in that, The method includes: Acquire acoustic signals of lung disease, and divide the acoustic signals into frames to obtain multiple acoustic signal segments; The multiple acoustic signal segments are adjusted based on an adaptive zero-crossing rate algorithm to obtain multiple effective acoustic segments; wherein, the adaptive zero-crossing rate algorithm determines the zero-crossing baseline of the zero-crossing rate based on the detection environment acoustic signal of the lung disease acoustic signal, and determines the zero-crossing rate threshold corresponding to the acoustic signal segment based on the fluctuation change of the acoustic signal segment, and each effective acoustic segment includes a complete acoustic event. The model identifies the health risk level corresponding to the effective acoustic segment based on an artificial intelligence model, wherein the artificial intelligence model detects lung disease-related acoustic events and adjusts the attention weight of the corresponding effective acoustic segment based on the related acoustic events. Adjusting the effective acoustic segments among the multiple acoustic signal segments based on an adaptive zero-crossing rate algorithm includes: Acquire the detection environment acoustic signal of the lung disease acoustic signal, and determine the zero-crossing reference line of the zero-crossing rate based on the amplitude of the environment acoustic signal; The zero-crossing rate threshold corresponding to the acoustic signal segment is determined based on the fluctuation changes of a preset number of adjacent acoustic signal segments and the zero-crossing baseline; wherein, the zero-crossing rate threshold for statistical fluctuation changes is used to distinguish between normal fluctuations and abnormal fluctuations. For any acoustic signal segment, calculate the zero-crossing rate of the arbitrary acoustic signal segment; Valid acoustic segments are selected based on the zero-crossing rate threshold and the numerical relationship between the zero-crossing rates.

2. The method according to claim 1, characterized in that, Determining the zero-crossing rate threshold corresponding to an acoustic signal segment based on the fluctuation changes of a preset number of adjacent acoustic signal segments and the zero-crossing baseline includes: Determine the first acoustic signal segment preceding the second acoustic signal segment; Determine the amplitude variation curves of the first acoustic signal segment and the second acoustic signal segment; Predict the change curve of the current acoustic signal segment based on the difference between the amplitude change curves; The zero-crossing baseline is adjusted based on the difference between the change curve of the current acoustic signal segment and the actual change curve.

3. The method according to claim 1, characterized in that, Acoustic signals of lung disease are acquired, and the acoustic signals are framed to obtain multiple acoustic signal segments, including: Identify the volume variation trend of the acoustic signals of the lung disease; The abrupt change point is marked according to the volume change trend, and the difference in acoustic signal volume before and after the abrupt change point is greater than a preset threshold. Calculate the frequency change trend of the acoustic signal of the lung disease, align the frequency change trend with the volume change trend, and determine the target point corresponding to the mutation point on the frequency change trend; The abrupt change point is corrected based on the frequency change state of the target point; The acoustic signal of lung disease is framed based on the corrected mutation point.

4. The method according to claim 1, characterized in that, The artificial intelligence model includes: The convolutional layer takes the time-spectrum map of the effective acoustic segment as input, and performs convolution operation on the time-spectrum map through the convolutional neural network to extract the local acoustic features corresponding to each effective acoustic segment; An attention weight layer detects lung disease-related acoustic events. Using the lung disease-related acoustic events as input, it locates the target effective acoustic segments corresponding to the associated acoustic events, calculates the similarity between the query vector and the key vector, and converts the obtained similarity value into an attention value. The attention value of the target effective acoustic segment is greater than that of other effective acoustic segments. The recurrent layer takes the local acoustic features fused with the attention mechanism as input to the long short-term memory network to obtain temporal dynamic features; The fully connected layer and the output layer take temporal dynamic features as input and predict the classification probability of the temporal dynamic features.

5. The method according to claim 4, characterized in that, The identification of the health risk level corresponding to the effective acoustic segment based on the artificial intelligence model includes: The relevant statistical features extracted from the time-spectrum graph by the convolutional neural network are used as the query vector, and the relevant statistical features are also used as the key vector. The similarity between the query vector and the key vector is calculated, and the similarity is converted into an attention value. When the local acoustic features include pathological acoustic features, the attention value between the local acoustic features and the query vector is greater than the attention value of non-disease-specific features.

6. The method according to claim 1, characterized in that, Based on an artificial intelligence model, the detection of lung disease-related acoustic events is performed. Using these lung disease-related acoustic events as input, the target effective acoustic segment corresponding to each related acoustic event is located, including: Multiple associated acoustic events in lung disease and their corresponding temporal relationships were identified, with each acoustic event having different acoustic characteristics. Based on the temporal relationship, locate the first target valid acoustic segment corresponding to the first associated acoustic event; Based on the first target effective acoustic segment and the timing relationship, the subsequent target effective acoustic segment is located.

7. The method according to claim 6, characterized in that, After identifying the lung disease-related acoustic events based on an artificial intelligence model, and locating the target effective acoustic segment corresponding to the lung disease-related acoustic events as input, the method further includes: Based on temporal relationships, the first associated acoustic event among multiple associated acoustic events in lung disease was identified; Calculate the maximum time threshold between each subsequent associated acoustic event and the previous associated acoustic event; Locate the first target effective acoustic segment, and using the first target effective acoustic segment as the starting point, determine the candidate effective acoustic segments by summing all the longest time thresholds as the time interval; Other valid target acoustic segments are selected from the candidate valid acoustic segments.

8. The method according to claim 1, characterized in that, The health risk level corresponding to the effective acoustic segment is identified based on an artificial intelligence model, including: Receive the classification probability output by the artificial intelligence model; The risk level is determined based on the classification probability and a preset threshold. A screening report is generated and sent to the user terminal for display.

9. An acoustic recognition device for lung diseases based on artificial intelligence, characterized in that, The device includes a processing module and a generation module; wherein... The processing module is used to acquire acoustic signals of lung diseases, and to segment the acoustic signals into frames to obtain multiple acoustic signal segments; The processing module is used to adjust the plurality of acoustic signal segments based on an adaptive zero-crossing rate algorithm to obtain a plurality of valid acoustic segments. The adaptive zero-crossing rate algorithm determines a zero-crossing baseline for the zero-crossing rate based on the ambient acoustic signal used to detect the lung disease acoustic signal, and determines a zero-crossing rate threshold corresponding to the acoustic signal segment based on the fluctuation changes of the acoustic signal segments. Each valid acoustic segment includes a complete acoustic event. The processing module is also used to acquire the ambient acoustic signal used to detect the lung disease acoustic signal, determine the zero-crossing baseline for the zero-crossing rate based on the amplitude of the ambient acoustic signal, and determine the zero-crossing rate threshold corresponding to the acoustic signal segment based on the fluctuation changes of a preset number of adjacent acoustic signal segments and the zero-crossing baseline. Specifically, it statistically analyzes the fluctuation changes and distinguishes between normal and abnormal fluctuations. For any acoustic signal segment, it calculates the zero-crossing rate of that segment. Valid acoustic segments are then selected based on the zero-crossing rate threshold and the numerical relationship between the zero-crossing rates. The generation module is used to identify the health risk level corresponding to the effective acoustic segment based on an artificial intelligence model, wherein the artificial intelligence model detects lung disease-related acoustic events and adjusts the attention weight of the corresponding effective acoustic segment based on the related acoustic events.

Images