COPD identification system based on audio representation of radar respiration signal on body surface

Abstract

The application discloses a COPD identification system based on audio representation of a body surface radar respiratory signal, and comprises the following modules: a signal acquisition module, which is used for acquiring echo data of a detection object based on a body surface radar and performing distance dimension fast Fourier transform, extracting a phase sequence of a stable target to obtain a respiratory micro-motion signal; a respiratory audio generation module, which is used for encoding and normalized mapping of the respiratory micro-motion signal to obtain respiratory representation audio; a COPD pathological phase perception module, which is used for overlapping and dividing the respiratory representation audio, and generating a mel time-frequency sequence based on a COPD pathological characteristic frequency band; and a deep learning classification module, which is used for taking the mel time-frequency sequence as input data, learning a pathological mode in a time-frequency feature by using a global attention mechanism, and outputting a COPD prediction result. The application realizes completely non-invasive respiratory monitoring by the cooperation of a millimeter wave radar and a series of modules, and is particularly suitable for early identification and diagnosis of COPD.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence medical diagnostic equipment technology, and more specifically to a COPD identification system based on the audio representation of respiratory signals from body surface radar. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a common, preventable, and treatable chronic airway disease characterized by persistent airflow limitation and corresponding respiratory symptoms. Early diagnosis and intervention are crucial for slowing disease progression. Traditional COPD diagnosis relies primarily on pulmonary function tests (such as vital capacity measurement). While accurate, this method requires specialized equipment and medical personnel, and demands high levels of cooperation from some elderly or frail patients. The procedures are also time-consuming, making large-scale early screening and in-hospital and out-of-hospital monitoring difficult.

[0003] In recent years, with the development of non-contact sensing technology and artificial intelligence, using radar or microphones to collect respiratory audio signals for COPD-assisted diagnosis has become a research hotspot. Respiratory audio contains rich physiological and pathological information, especially features such as prolonged expiratory phase, wheezing, and decreased respiratory energy, which are important auscultatory indicators for COPD.

[0004] However, existing respiratory signal analysis methods based on audio or radar still have the following shortcomings:

[0005] Limitations of feature extraction: Traditional signal processing methods, such as short-time Fourier transform (STFT), can convert time-domain signals into time-frequency spectra, but when processing complex respiratory audio, they often struggle to adaptively capture subtle pathological features and are sensitive to noise.

[0006] Bottlenecks in model architecture: Traditional machine learning or shallow neural network models have difficulty effectively modeling long-distance dependencies when processing high-dimensional time series data, resulting in insufficient overall understanding of breathing patterns and affecting diagnostic accuracy.

[0007] Low data utilization efficiency: Before inputting audio signals into the model, there is a lack of effective preprocessing and framing strategies for pathological features, which leads to the loss or submersion of key pathological information such as the start and end points of breathing and wheezing segments during the processing.

[0008] Therefore, there is an urgent need for a COPD detection device that can efficiently extract pathological features from respiratory audio and use advanced artificial intelligence models for accurate identification, suitable for early screening and long-term monitoring of COPD. Summary of the Invention

[0009] In view of the above problems, the present invention is proposed to provide a COPD identification system based on the audio representation of respiratory signals from a body surface radar to overcome or at least partially solve the above problems. The system acquires respiratory micro-motion signals measured by a body surface radar without contact, realizes the acoustic representation of respiratory data through dynamic standardized modeling and feature audio processing, and performs intelligent COPD identification by combining a deep learning model. It is suitable for early screening and long-term health monitoring systems.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a COPD identification system based on the audio representation of respiratory signals from a surface radar, comprising the following modules: Signal acquisition module: used to acquire echo data of the detection object based on the surface radar and perform range-dimensional fast Fourier transform, count the range unit with the highest frequency in the whole sequence as the stable target, extract the phase sequence of the stable target and perform phase unwrapping and correction to obtain the respiratory micro-motion signal; Breathing audio generation module: used to perform resting-state fast encoding or differential enhancement encoding on the breathing micro-motion signal, and normalize and map it into a standard audio signal to obtain breathing characterization audio; COPD pathological phase sensing module: used to overlap and segment the respiratory characterization audio using a sliding window, and generate a Mel time-frequency sequence based on COPD pathological feature frequency bands; Deep learning classification module: It is used to take the Mel time-frequency sequence as input data, use the global attention mechanism to learn the pathological patterns in the time-frequency features, and output COPD prediction results.

[0011] Furthermore, the signal acquisition module specifically includes: Radar data rearrangement unit: used to acquire echo data of the detected object based on the surface radar; the echo data is an integer sequence interleaved with in-phase components and quadrature components; it is also used to restore the interleaved integer sequence to the independent components of each antenna and rearrange them into a complex radar data matrix by using an equal-interval sampling method. Static clutter filtering unit: used to perform a range-dimensional fast Fourier transform on each rearranged frame of data, and subtract the mean spectrum of all frames from the resulting range spectrum to obtain a new range spectrum after filtering out static clutter; Stable target selection unit: used to select the globally stable energy maximum distance unit through mode statistics, extract its phase and perform unwrapping and jump correction to obtain a continuous and stable phase sequence; Respiratory frequency band optimization unit: The corrected phase sequence is processed by a 7-point weighted smooth differential operator to obtain the phase differential sequence value as the respiratory micro-motion signal.

[0012] Furthermore, the 7-point weighted smoothing differential operator processes the corrected phase sequence, as expressed by the formula:

[0013] in, Indicates the first n The phase differential sequence value output at the frame; This represents a continuous phase sequence after jump correction.

[0014] Furthermore, the breathing audio generation module is used to directly perform normalization and audio mapping on the resting phase differential sequence of the detected object before movement to obtain breathing characterization audio, depending on the application scenario; or to estimate the optimal alignment offset using the cross-correlation function and perform differential processing based on the phase differential sequence of the detected object before and after movement, and then perform normalization and audio mapping to obtain breathing characterization audio.

[0015] Furthermore, the estimation of the optimal alignment offset using the cross-correlation function and the subsequent differential processing are expressed by the following formula:

[0016]

[0017] in, This indicates a pathological enhancement differential sequence. i Indicates the discrete-time sampling point index. This represents the phase differential sequence before motion. This represents the sequence value after the phase differential sequence has been aligned by the optimal alignment offset. This represents the optimal alignment offset corresponding to the maximum cross-correlation result. express The total length of the effective sampling points used in the calculation The sliding of the phase differential sequence after motion The subsequent sequence value.

[0018] Furthermore, the COPD pathological phase sensing module includes: Overlapping segmentation unit: used to segment the respiratory characterization audio into a series of ordered overlapping segments based on the average respiratory cycle of the detected object; Mel time-frequency sequence generation unit: used to perform frame processing on each segment and convert it into a two-dimensional Mel spectrogram; wherein, the two-dimensional Mel spectrogram is a Mel time-frequency sequence covering the 200-800Hz frequency band filtered by 128 triangular filters, used to capture COPD-related pathological features and respiratory energy decay.

[0019] Furthermore, the overlapping segmentation unit segments the breathing characterization audio into a series of ordered overlapping segments with the following segmentation parameters: audio segmentation window length is set to 2.0s; overlap rate is 0.4; and step size is 1.2s. The obtained number m The segment is:

[0020] in, Indicates the segmented first m A segment, This represents the audio sequence representing respiration after normalization and mapping. Indicates step size, Indicates the length of the audio segmentation window. α Indicates the overlap rate.

[0021] Furthermore, the deep learning classification module is built based on the visual Transformer model. It takes the Mel time-frequency sequence as input data, and after positional encoding, multi-head self-attention mechanism, and MLP encoding, outputs the anomaly probability of each segment through the classification head:

[0022] in, Indicates the prediction of the first m The probability of an anomaly in a time-frequency segment. Indicates the first m The global classification label feature vector extracted after encoding a Mel spectrogram; and These represent the learnable weight matrix and bias vector of the fully connected layer in the classification head, respectively. The anomaly probabilities of all segments are aggregated and compared with a preset threshold to obtain the COPD prediction result.

[0023] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a COPD identification system based on the audio representation of respiratory signals from body surface radar, which has the following beneficial effects: This invention's system requires no sensors worn by the patient and no professional medical personnel to operate. It can capture and intelligently analyze respiratory signals using surface radar in complex environments, providing an efficient, unobtrusive, and accurate technological means for home screening and clinical auxiliary diagnosis of COPD. It achieves end-to-end identification from raw radar signals to COPD risk probability.

[0024] This invention constructs a signal acquisition module, establishing a unified signal processing standard; it also constructs a respiratory audio generation module, converting weak respiratory radar signals into "audio"; and a COPD pathological phase sensing module, which, through overlapping segmentation and Mel-spectrum sequence generation, solves the problem of pathological feature truncation caused by frame division in traditional short-time Fourier transform when processing respiratory signals. This enables the system to more stably capture the core pathological features unique to COPD, such as "prolonged expiratory phase" and "energy attenuation in the 200-800Hz frequency band," significantly improving diagnostic specificity. Finally, the feature maps input to the deep learning classification model can accurately highlight pathological regions, effectively suppressing noise interference in irrelevant frequency bands, laying a solid data foundation for subsequent high-precision identification. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 This is a framework diagram of a COPD identification system based on the audio representation of respiratory signals from a surface radar, provided in an embodiment of the present invention. Figure 2 This is a structural framework diagram of the signal acquisition module provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention discloses a COPD identification system based on the audio representation of respiratory signals from a body surface radar, referring to... Figure 1 As shown, it includes the following modules: Signal acquisition module: used to acquire echo data of the detected object based on the surface radar and perform range-dimensional fast Fourier transform, count the range cell with the highest frequency in the whole sequence as the stable target, extract the phase sequence of the stable target and perform phase unwrapping and correction to obtain the respiratory micro-motion signal; Respiratory audio generation module: used to perform resting-state fast encoding or differential enhancement encoding on respiratory micro-movement signals, and normalize and map them into standard audio signals to obtain respiratory characterization audio; COPD pathology phase sensing module: used to overlap and segment respiratory characterization audio using a sliding window and generate Mel time-frequency sequences based on COPD pathology feature frequency bands; Deep learning classification module: It takes Mel time-frequency series as input data, uses a global attention mechanism to learn pathological patterns in time-frequency features, and outputs COPD prediction results.

[0029] This embodiment is applied to a non-intrusive early COPD screening scenario in a hospital or home environment. In this scenario, an intelligent identification device integrating a dynamic standard surface radar and a COPD recognition system is placed in a pre-screening triage area or next to the examination bed in the examination room, or even in a fixed location within the home. Subjects (such as middle-aged and elderly people who have smoked for a long time or people with respiratory symptoms) do not need to wear any physiological sensors. They only need to maintain a natural and calm breathing state within the radar detection range (e.g., on a bed or sofa) while minimizing large-scale physical activities.

[0030] After the system is activated, the dynamic standard surface radar emits non-ionizing electromagnetic waves and receives echo signals reflected from the human chest and abdomen. The COPD identification system captures these micro-Doppler signals in real time and converts them into time-series data reflecting respiratory movements. Subsequently, the system performs overlapping framing of continuous respiratory signals to ensure the capture of the complete inspiratory-expiratory cycle, especially the key pathological features of the expiratory phase.

[0031] Next, the system converts each frame of the time-domain respiratory signal into a Mel spectrogram. This process utilizes a customized 128-channel filter bank to enhance the feature representation in the 200-800Hz frequency band, highlighting pathological acoustic features specific to COPD patients, such as wheezing or decreased breath sounds. The generated two-dimensional spectrogram is then fed into the Visual Transformer (ViT) model for inference. The ViT model uses its multi-head self-attention mechanism to analyze the spatiotemporal features in the spectrogram and calculate the probability that the subject's respiratory signal belongs to the "COPD abnormality" category. Finally, the system outputs visualized diagnostic suggestions (such as "low risk," "suspected abnormality, please re-examine," or "high risk") on the display screen and generates an electronic report containing respiratory waveforms, spectral features, and risk scores for doctors' reference or archiving.

[0032] The following describes in detail the modules included in the COPD identification system in this embodiment: 1. Signal acquisition module.

[0033] The signal acquisition module in this embodiment refers to... Figure 2 As shown, it includes: radar data rearrangement unit, static clutter filtering unit, stable target selection unit, and breathing band optimization unit.

[0034] This embodiment uses a TI IWR1443 millimeter-wave radar in conjunction with a DCA1000 data acquisition card. The original (.bin) file is in I / Q interleaved int16 format. Let Rx0 and Rx1 represent the first and second physical receiving antennas on the radar hardware, respectively; I0 and I1 represent the in-phase components of the baseband echo signals received by receiving antennas Rx0 and Rx1, respectively; and Q0 and Q1 represent the quadrature components of the baseband echo signals received by receiving antennas Rx0 and Rx1, respectively. The data is stored in cyclic order as follows: .

[0035] 1.1 The radar data rearrangement unit, based on the radar hardware configuration parameters—number of frames, number of chirs, number of receiving antennas, and number of sampling points—reorganizes the I / Q interleaved integer sequence into a complex radar data cube according to the data organization specifications specified by the device. The original radar (.bin) file is an I / Q interleaved sample and needs to be rearranged into a complex form. First, an int16 array is read from the binary stream, rearranged frame by frame, and then I / Q complex recovery is performed on each frame.

[0036] This embodiment uses equal-interval sampling to restore the interleaved one-dimensional sequence into independent components of each antenna; the interval step size in this embodiment is selected as 4.

[0037] Extraction of in-phase and quadrature components of antenna Rx0: I 0 = raw[0::4], Q 0 = raw[2::4]; Extraction of in-phase and quadrature components of antenna Rx0: I 1 = raw[1::4] ，Q 1 = raw[3::4]; Wherein, raw represents a one-dimensional 16-bit integer data sequence read directly from a binary file (.bin); [0::4], [1::4], [2::4] and [3::4] respectively represent taking a value every 4 elements starting from indices 0, 1, 2 and 3.

[0038] This process of forming and rearranging complex sequences results in a final shape of ( N frame ，N chirp ，N Rx ，N ADC Complex radar data matrix, N frame 、N chirp 、N Rx and N ADCThese represent the total number of frames acquired by the radar, the number of frequency-modulated continuous waves (Chirps) contained in each frame, the total number of receiving antennas used, and the number of analog-to-digital converter (ADC) sampling points for each Chirp, respectively.

[0039] 1.2 Static Clutter Removal Unit: This unit eliminates static clutter through inter-frame averaging; specifically, it includes: For each frame acquired by the radar, perform an FFT along the ADC dimension to obtain the original range spectrum. X r [ n,k ],in, k For distance cell index, n This is the sampling frame index. The inter-frame mean is then subtracted to eliminate static background and obtain the new range spectrum. for:

[0040] in, j For summation index.

[0041] The stable target selection unit is to select the distance unit corresponding to the energy peak in each frame, and the unit with the highest frequency in the whole sequence is selected as the stable target.

[0042] 1.3. Stable target selection unit: Select the distance unit corresponding to the energy peak in each frame, and count the unit that appears most frequently in the whole sequence as the stable target.

[0043] Because the raw I / Q data contains strong static clutter such as wall reflections, directly selecting the energy-maximum range cell within a single frame is susceptible to noise interference, leading to target position jumps. Simultaneously, motion artifacts may be mixed into the phase sequence, resulting in a low signal-to-noise ratio for micro-motion features. To improve target stability, this embodiment uses the following method to extract the energy-maximum index frame-by-frame from the range FFT results, expressed by the formula:

[0044]

[0045] in, k n Indicates the first n The index of the distance cell with the highest energy in the frame; Perform mode statistics on the peak index sequence of all frames to obtain globally stable target range cells, and select the value with the highest frequency:

[0046] To avoid near-field interference (such as antenna coupling, direct waves), this embodiment shields elements with a distance index less than 3 (k < 3), and ultimately determines the stable target for subsequent processing as k. .

[0047] Extract target cell k Phase of the complex sequence and its expansion:

[0048] in, This represents the original unwrapped phase at the nth frame; This represents the phase unwrapping function, used to eliminate phase defects generated when crossing boundaries. or Phase winding and jump phenomena; This represents a mathematical operator for extracting the phase angle of a complex number; In the new distance spectrum, the first... n Frame in stable target range cell The complex distance spectrum value of the static background at the location.

[0049] Subsequently, to further eliminate local phase abrupt changes caused by random impulse noise or non-physiological large-scale body movements, this embodiment calculates the phase increment:

[0050] in, Indicates the first n Phase increment at frame.

[0051] This embodiment makes T The preset threshold for determining phase abrupt change; if Then let The continuous phase after jump correction is obtained by incremental accumulation reconstruction as follows:

[0052] in, Indicates the first n The final continuous phase value after jump correction at the frame; This indicates the initial reference phase of the phase sequence; l This is the frame sequence number traversal index during the summation process.

[0053] 1.4. Breathing frequency band optimization unit: To highlight the rate of change of micro-motion and suppress high-frequency noise, differential processing is performed to optimize the breathing frequency band (0.1–2Hz).

[0054] In this embodiment, a 7-point weighted smoothing differential operator is used to calculate the corrected phase sequence. The calculation formula is as follows:

[0055] in, Indicates the first n The phase differential sequence value output at the frame is the respiratory micro-motion signal; This represents the continuous phase sequence output by the stable target selection unit after jump correction; coefficients 1, 4, 5 and the normalization factor. These are the filter weighting coefficients, forming a combined differential kernel that integrates low-pass smoothing and center difference.

[0056] The kernel function exhibits gain fluctuations of less than 0.5 dB in the 0.1–2 Hz (adult respiratory frequency range) and attenuation of ≥15 dB in the >5 Hz high-frequency range, making it suitable for respiratory micro-movements.

[0057] 2. Breathing audio generation module: Based on different application scenarios, it performs corresponding signal processing.

[0058] One approach involves using resting-state fast encoding on the differential sequence of the subject's resting phase before movement, followed by direct normalization and audio mapping to obtain the respiratory characterization audio. Specifically, this includes: The resting phase differential sequence before exercise output by the respiratory band optimization unit The sequence already contains temporal dynamics of airflow limitation, such as respiratory waveform distortion and phase jitter; subsequent modules can effectively extract these airflow limitation features. Execute directly Normalization and audio mapping. This mode eliminates the need for subjects to undergo exercise stress testing, making it particularly suitable for elderly or mobility-impaired patients.

[0059] Another approach involves using differential coding based on the phase differential sequence before and after the object's motion. This is done by estimating the optimal alignment offset using a cross-correlation function, performing differential processing, and then executing the code. Normalization and audio mapping yield the breathing representation audio.

[0060] In this embodiment, after acquiring the pre-exercise phase differential sequence, the subject is guided to undergo a short-term exercise load (such as stationary cycling), and the post-exercise phase differential sequence is acquired. The pathological enhancement differential sequence is calculated using cross-correlation alignment.

[0061] in, This indicates a pathological enhancement differential sequence. This represents the phase differential sequence before motion. This represents the optimal alignment offset of the phase differential sequence after motion. Aligned sequence values ​​after translation on the time axis i For discrete-time sampling point index, This is the optimal alignment offset corresponding to the maximum cross-correlation result, used to shift the sequence after translation on the time axis to find the best matching point.

[0062] In this embodiment, the optimal alignment offset is estimated using the cross-correlation function as follows:

[0063] in, This represents the total number of micro-motion velocity sequences before the exercise intervention; This represents the optimal alignment offset corresponding to the maximum cross-correlation result. Its physical meaning is the time frame difference required for two breathing sequences to achieve maximum overlap in the breathing cycle (such as from the start of inspiration to the start of inspiration). The sliding of the phase differential sequence after motion The subsequent sequence value.

[0064] This embodiment achieves the optimal alignment offset. Then, the pathological enhancement differential sequence is calculated using the following formula:

[0065] Because of static anatomical differences in chest cavity size and absolute distance from the radar among different subjects, direct comparison of their respiratory signals is insufficient to highlight subtle pathological features of COPD. This embodiment first strictly aligns the respiratory waveforms before and after exercise in time phase using cross-correlation alignment offset, and then performs point-by-point subtraction. This operation physically completely cancels out the common baseline of static anatomical differences and normal breathing, thereby greatly enhancing the abnormal dynamic changes such as airflow limitation induced by exercise load; for example, in COPD patients after exercise, the expiratory phase exhibits abnormally prolonged resistance, a subtle change that is highlighted in the differential sequence. The middle will be significantly amplified.

[0066] Next, this embodiment applies the difference sequence implement Normalization yields the normalized results of the pathological enhancement-type differential sequences:

[0067] Then, audio mapping is performed. In this embodiment, the relative change ratio of micro-motions is preserved to ensure signal comparability between different subjects. The resulting audio mapping is:

[0068] in, This represents the respiratory characterization audio sequence obtained after normalization and mapping; in this embodiment, it is a 16-bit discrete audio sequence. This represents the normalization result of the pathological enhancement differential sequence; `round` represents the floor function; 32767 is the largest positive value of a 16-bit signed integer, used to stretch the normalized amplitude of [-1, 1] to the audio dynamic range. Then, the above audio mapping sequence... Duration standardization is performed by extending the audio to a fixed duration T through loop splicing or truncation.fix =10s (covering ≥4 complete respiratory cycles) to obtain the final standardized respiratory characterization audio. .

[0069] 3. COPD pathological phase sensing module, specifically including an overlap segmentation unit and a Mel time-frequency sequence generation unit, performs overlap segmentation and time-frequency focusing on respiratory characterization audio.

[0070] Overlapping segmentation unit: Sets the length of the audio segmentation window. The duration was 2.0 seconds (matching the average adult respiratory cycle of 2.5 seconds); overlap rate α The value is 0.4, and the step size is... It is 1.2s; the obtained first time is 1.2s. m The segment is:

[0071] Ensure that key pathological phases (such as the expiratory initiation point) are fully covered by at least two segments.

[0072] Mel time-frequency sequence generation unit: converts each 2.0s audio segment (sampling rate 16kHz) into a time-frequency two-dimensional feature representation with temporal structure, i.e., Mel spectrum sequence.

[0073] This embodiment first performs frame segmentation on the audio segment, using a Hamming window (window length 25ms, frame shift 10ms), and obtains the amplitude spectrum through short-time Fourier transform. Then, 128 triangular Mel filters are designed to cover the pathological characteristic frequency band of 80–800Hz. This frequency band accurately covers COPD-related wheezing harmonics (200–500Hz) and expiratory phase energy decay characteristics. The amplitude spectrum is mapped through the Mel filter bank to obtain the Mel spectrum, whose dimensions are: T m ∈R Nt×F in, T m Indicates the first m The two-dimensional Mel spectrum matrix corresponding to each audio segment Nt The number of time frames is determined by the audio length and frame shift. F =128 is the number of Mel bands.

[0074] This embodiment generates T m It is a two-dimensional matrix, with the horizontal axis representing time frames and the vertical axis representing Mel frequency bands. Each element represents the energy intensity of the corresponding time and frequency band. This representation is essentially a time-frequency plot, which can be used as input to a visual Transformer, enabling it to learn the temporal dependencies and frequency domain patterns in breathing audio through an attention mechanism.

[0075] 4. Deep learning classification module.

[0076] The deep learning classification module in this embodiment is built based on the visual Transformer model, specifically the standard ViT-Base model. To achieve accurate recognition of complex breathing patterns, the specific implementation and data processing of this module are as follows: 1) Time-frequency graph segmentation and serialization embedding: First, the two-dimensional Mel time-frequency sequence (i.e., Mel spectrogram) output by the COPD pathology phase sensing module is divided into non-overlapping fixed-size local image patches. Each image patch is flattened and mapped to a one-dimensional hidden layer feature vector through a linear projection matrix (in this embodiment, the hidden dimension of each layer is configured to be 768). Subsequently, a learnable global classification token ([CLS] Token) feature vector is concatenated to the beginning of the feature sequence, and a one-dimensional learnable positional encoding is superimposed element-wise on each feature vector in the sequence, thereby obtaining an input feature matrix containing rich spatial time-frequency sequence information.

[0077] 2) Multi-layer feature self-attention encoding and pathological feature capture: The constructed input feature matrix is ​​then fed sequentially into a 12-layer Transformer encoder for deep feature extraction. Each layer contains a multi-head self-attention (MSA) unit and an MLP encoding unit.

[0078] In this process, the multi-head self-attention mechanism dynamically captures long-range pathological dependencies between different frequency bands and time periods in the Mel time-frequency sequence by calculating the attention weights of global features. For example, the model can adaptively associate abnormal inspiratory-to-expiratory ratio features across multiple respiratory cycles and capture the temporal dependencies between intermittent wheezing harmonics distributed in the 200-800Hz frequency band. Subsequently, nonlinear mapping of features is performed through MLP coding units. After 12 layers of stacked extraction, the first feature vector (i.e., the vector corresponding to the global classification label) in the output sequence of the last layer is extracted and used as the high-dimensional global fusion pathological feature vector of the current Mel spectral segment.

[0079] 3) Fragment-level probability classification and result aggregation: Finally, the extracted global classification feature vector is input into the classification head, and the anomaly probability of the current segment is output through the Softmax normalized exponential function:

[0080] in, Indicates the prediction of the first m The probability of an anomaly in a time-frequency segment; Indicates the first m The global classification label feature vector extracted after encoding a Mel spectrogram; and These represent the learnable weight matrix and bias vector of the fully connected layer in the classification head, respectively.

[0081] Finally, this embodiment aggregates the anomaly probabilities of all segments, compares them with a preset threshold, and outputs a COPD prediction result of "abnormal" or "normal".

[0082] This embodiment constructs a highly efficient processing system from raw radar signals to pathological diagnosis. Through modules such as phase correction, adaptive audio encoding, overlapping segmentation, and time-frequency feature extraction, the respiratory micro-motion signals acquired by millimeter-wave radar are converted into audio representations with pathological expression capabilities. A deep learning classification module is then constructed based on a visual Transformer model to perform end-to-end classification.

[0083] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0084] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A COPD identification system based on the audio representation of respiratory signals from surface radar, characterized in that, Includes the following modules: Signal acquisition module: used to acquire echo data of the detection object based on the surface radar and perform range-dimensional fast Fourier transform, count the range unit with the highest frequency in the whole sequence as the stable target, extract the phase sequence of the stable target and perform phase unwrapping and correction, perform differential processing on the corrected phase sequence to obtain the phase differential sequence value as the respiratory micro-motion signal; The breathing audio generation module is used to perform resting-state fast encoding or differential enhancement encoding on the breathing micro-motion signal, and normalize and map it into a standard audio signal to obtain breathing characterization audio. Specifically, depending on the application scenario, it can directly perform normalization and audio mapping on the resting phase differential sequence of the detected object before movement to obtain breathing characterization audio; or based on the phase differential sequence of the detected object before and after movement, it can use the cross-correlation function to estimate the optimal alignment offset and perform differential processing, and then perform normalization and audio mapping to obtain breathing characterization audio. The estimation of the optimal alignment offset using the cross-correlation function and the subsequent difference processing are expressed by the following formula: in, This indicates a pathological enhancement differential sequence. i Indicates the discrete-time sampling point index. This represents the phase differential sequence before motion. This represents the sequence value after the phase differential sequence has been aligned by the optimal alignment offset. This represents the optimal alignment offset corresponding to the maximum cross-correlation result. express The total length of the effective sampling points used in the calculation The sliding of the phase differential sequence after motion The subsequent sequence value; COPD pathology phase sensing module: used to overlap and segment the respiratory characterization audio using a sliding window, and generate a Mel time-frequency sequence based on COPD pathology feature frequency bands; specifically including: Overlapping segmentation unit: used to segment the respiratory characterization audio into a series of ordered overlapping segments based on the average respiratory cycle of the detected object; Mel time-frequency sequence generation unit: used to perform frame processing on each segment and convert it into a two-dimensional Mel spectrogram; wherein, the two-dimensional Mel spectrogram is a Mel time-frequency sequence covering the 200-800Hz frequency band filtered by 128 triangular filters, used to capture COPD-related pathological features and respiratory energy decay; Deep learning classification module: It takes the Mel time-frequency sequence as input, uses a global attention mechanism to learn the pathological patterns in the time-frequency features, and outputs COPD prediction results.

2. The system as described in claim 1, characterized in that, The signal acquisition module specifically includes: Radar data rearrangement unit: used to acquire echo data of the detected object based on the surface radar; the echo data is an integer sequence interleaved with in-phase components and quadrature components; it is also used to restore the interleaved integer sequence to the independent components of each antenna and rearrange them into a complex radar data matrix by using an equal-interval sampling method. Static clutter filtering unit: used to perform a range-dimensional fast Fourier transform on each rearranged frame of data, and subtract the mean spectrum of all frames from the resulting range spectrum to obtain a new range spectrum after filtering out static clutter; Stable target selection unit: used to select the globally stable energy maximum distance unit through mode statistics, extract its phase and perform unwrapping and jump correction to obtain a continuous and stable phase sequence; Respiratory frequency band optimization unit: The corrected phase sequence is processed by a 7-point weighted smooth differential operator to obtain the phase differential sequence value as the respiratory micro-motion signal.

3. The system as described in claim 2, characterized in that, The 7-point weighted smoothing differential operator processes the corrected phase sequence, as expressed by the formula: in, Indicates the first n The phase differential sequence value output at the frame; This represents a continuous phase sequence after jump correction.

4. The system as described in claim 1, characterized in that, The overlapping segmentation unit segments the breathing characterization audio into a series of ordered overlapping segments with the following segmentation parameters: audio segmentation window length is set to 2.0s; overlap rate is 0.4; and step size is 1.2s. The obtained number m The segment is: in, Indicates the segmented first m A segment, This represents the audio sequence representing respiration after normalization and mapping. Indicates step size, Indicates the length of the audio segmentation window. α Indicates the overlap rate.

5. The system as described in claim 1, characterized in that, The deep learning classification module is built based on the visual Transformer model. It takes the Mel time-frequency sequence as input data, processes it through position encoding, multi-head self-attention mechanism, and MLP encoding, and then outputs the anomaly probability of each segment through the classification head. in, Indicates the prediction of the first m The probability of an anomaly in a time-frequency segment. Indicates the first m The global classification label feature vector extracted after encoding a Mel spectrogram; and These represent the learnable weight matrix and bias vector of the fully connected layer in the classification head, respectively. The anomaly probabilities of all segments are aggregated and compared with a preset threshold to obtain the COPD prediction result.

Images