Breathing state distinguishing method based on data fusion

Abstract

The invention relates to the technical field of vital sign monitoring, in particular to a breathing state distinguishing method based on data fusion, which comprises the following steps: acquiring multichannel physiological signals such as electrocardio, photoelectric volume pulse wave, breathing movement and the like to construct a group tensor; the method comprises the following steps: performing constraint non-negative tensor decomposition to construct a general physiological collaborative primitive library and generate a standard state template, then collecting a to-be-detected individual signal, calling the library to perform coupling matrix decomposition, calculating an individual mode expression weight vector, and judging an individual respiration state through a minimum Euclidean distance between the vector and the standard template. According to the method, collaborative analysis is constructed by fusing multi-source information such as electrocardio and photoelectric volume pulse waves, interference of a single sensor is reduced, a common physiological mode is extracted from a group to form a standard template, individual data is matched with the template through coupling decomposition and weight calculation, and the breathing state of an individual is accurately judged according to the principle of minimizing reconstruction errors. And the reliability and adaptability are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of vital sign monitoring technology, and in particular to a method for determining respiratory status based on data fusion. Background Technology

[0002] The field of vital sign monitoring technology mainly involves the collection, analysis, and discrimination of basic human physiological parameters. Its core aspects include the perception of vital sign signals such as respiration, heart rate, body temperature, and blood pressure; the design of signal acquisition equipment; signal feature extraction; and the identification and judgment of various vital sign states. It encompasses a complete technical system from sensor signal acquisition, data fusion, parameter analysis to state recognition. Traditional respiratory state discrimination methods involve directly detecting human respiratory-related signals using devices such as airflow sensors, resistance bands, pressure sensors, or optical sensors, and then extracting and discriminating feature parameters from the signals through fixed threshold comparisons or statistical feature analysis to achieve classification and identification of respiratory states.

[0003] Existing technologies heavily rely on signals from direct-contact sensors such as airflow or pressure sensors. In practical applications, when a person moves, speaks, or the sensor is not worn properly, signal distortion or interruption can easily occur, severely affecting data quality. These technologies primarily employ fixed threshold comparisons or simple statistical feature analysis. This type of judgment logic cannot adapt to physiological differences between individuals or changes in the state of the same individual in different situations, lacking adaptability and flexibility. For example, a slight change in sleeping posture may cause fluctuations in pressure sensor readings, which may then be incorrectly judged as abnormal breathing by a fixed threshold, generating unnecessary false alarms and interfering with the accuracy of continuous monitoring. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a respiratory state discrimination method based on data fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a respiratory state discrimination method based on data fusion, comprising the following steps:

[0006] S1: Acquire ECG, photoplethysmography pulse wave, and respiratory motion signals to construct multi-channel physiological signals, analyze the time-spectrum diagram of the multi-channel physiological signals and calculate the arithmetic mean, and generate a population physiological signal tensor;

[0007] S2: Perform constrained nonnegative tensor decomposition on the population physiological signal tensor to obtain and combine the physiological channel contribution pattern matrix, time evolution pattern matrix, and frequency activation pattern matrix to construct a general physiological coordination primitive library, and calibrate the general physiological coordination primitive library to generate a standard state template;

[0008] S3: Collect multi-channel physiological signals of the individual to be tested and construct an individual physiological data matrix. Call the general physiological collaborative primitive library to perform coupled matrix decomposition on the individual physiological data matrix. Calculate the individual pattern expression weight vector with the constraint of minimizing reconstruction error and numerical difference.

[0009] S4: Calculate the Euclidean distance between the individual pattern expression weight vector and the standard state template, and determine the breathing state corresponding to the smallest Euclidean distance as the individual breathing state discrimination result.

[0010] As a further aspect of the present invention, the group physiological signal tensor specifically refers to the group-level time-frequency feature representation of multi-channel physiological signals, the general physiological collaborative primitive library includes the physiological channel contribution pattern matrix, the time evolution pattern matrix and the frequency activation pattern matrix, the standard state template is the collaborative primitive parameter configuration of multiple respiratory states, the individual pattern expression weight vector is the projection weight describing individual physiological data on the general physiological collaborative primitive library, and the individual respiratory state discrimination result specifically refers to the respiratory state category with the smallest distance from the standard state template.

[0011] As a further aspect of the present invention, the step of obtaining the group physiological signal tensor specifically includes:

[0012] S101: Acquire electrocardiogram (ECG) signals, photocapacitance-volume pulse wave (PVP) signals, and respiratory motion signals; perform synchronous time window segmentation on the ECG signals, PPV signals, and respiratory motion signals to obtain signal segments; construct data frames by using multiple signal segments within the same time window as column vectors; stack all the data frames of the time windows to generate a multi-channel physiological signal matrix.

[0013] S102: Apply Hanning window function preprocessing to each column of signal vector in the multi-channel physiological signal matrix, perform short-time Fourier transform to calculate energy spectral density, and stitch the energy spectral density along the time axis to generate a two-dimensional time spectrum for each signal channel. Aggregate the two-dimensional time spectrums of all the signal channels to obtain a single-user time spectrum set.

[0014] S103: Based on multiple single-user time-spectrum sets, each single-user time-spectrum set is stacked as a two-dimensional slice along the new user dimension to establish a three-dimensional data tensor. At the same time, the population averaging operator is called to calculate the arithmetic mean of the corresponding elements in the three-dimensional data tensor along the new user dimension and perform dimensionality reduction to generate a population physiological signal tensor.

[0015] As a further aspect of the present invention, the step of obtaining the standard state template specifically includes:

[0016] S201: Perform constrained nonnegative tensor decomposition on the population physiological signal tensor, decompose the population physiological signal tensor into mutually independent contribution mode factors, evolution mode factors and activation mode factors, and reconstruct the contribution mode factors, the evolution mode factors and the activation mode factors into physiological channel contribution mode matrix, time evolution mode matrix and frequency activation mode matrix respectively, and establish a multidimensional physiological mode matrix group.

[0017] S202: Call the multidimensional physiological pattern matrix group, obtain the signal strength weight and time synchronization weight defined by the preset coordination rule, and perform weighted combination on the elements of the physiological channel contribution pattern matrix, the time evolution pattern matrix and the frequency activation pattern matrix according to the signal strength weight and the time synchronization weight to construct a general physiological coordination primitive library;

[0018] S203: Based on the general physiological coordination primitive library, calculate the activation frequency and duration of multiple primitives and multiply them to generate coordination mode stability, set a coordination mode activation benchmark value, screen all physiological coordination primitives whose coordination mode stability exceeds the coordination mode activation benchmark value, perform numerical calibration, and generate a standard state template.

[0019] As a further aspect of the present invention, the step of calculating the individual pattern representation weight vector specifically comprises:

[0020] S301: Collect multi-channel physiological signals of the individual under resting state, align the multi-channel physiological signals with timestamps according to the preset sampling frequency and channel index, and aggregate the aligned signal data stream into a two-dimensional data structure, wherein the row vector represents the time series and the column vector represents the signal channel, and establish an individual physiological data matrix;

[0021] S302: Call the general physiological coordinating primitive library, and according to the preset physiological primitives in the library, perform non-negative coupled matrix decomposition on the individual physiological data matrix to split the individual physiological data matrix into a primitive pattern matrix and a weight coefficient matrix. Extract the primitive pattern matrix and concatenate it horizontally to generate a decomposed coordinating pattern matrix.

[0022] S303: Based on the decomposed cooperative pattern matrix, a randomly initialized undetermined weight vector is set. The decomposed cooperative pattern matrix and the undetermined weight vector are multiplied by matrix to obtain a reconstructed physiological matrix. The root mean square error between the reconstructed physiological matrix and the individual physiological data matrix is ​​calculated as the reconstruction error, and the Manhattan distance is calculated as the numerical difference. The element values ​​in the undetermined weight vector are iteratively adjusted by gradient descent until the weighted sum of the reconstruction error and the numerical difference is lower than the convergence threshold, thus obtaining the individual pattern expression weight vector.

[0023] As a further aspect of the present invention, the step of obtaining the individual respiratory state discrimination result specifically includes:

[0024] S401: Based on the multiple standard vectors and associated breathing state labels stored in the standard state template, the individual pattern expression weight vector is called, and the Euclidean distance is calculated by performing the sum of squares and square roots of the differences between the individual and all the standard vectors in the standard state template. Each generated distance value is paired with the breathing state label associated with the standard vector to form a state distance entry. All the state distance entries are aggregated to generate a state distance vector group.

[0025] S402: Traverse all the state distance entries in the state distance vector group, extract the distance value encapsulated in each state distance entry, retrieve the minimum value by loop comparison, lock the state distance entry with the minimum value, obtain the position identifier in the state distance vector group, and establish a minimum distance index;

[0026] S403: Invoke the minimum distance index and the state distance vector group, and based on the position information indicated by the minimum distance index, perform direct addressing within the state distance vector group to lock the corresponding state distance entry, access the state distance entry and parse out the encapsulated breathing state label, confirm the breathing state label as the discrimination conclusion, and obtain the individual breathing state discrimination result.

[0027] As a further aspect of the present invention, in the construction step of the universal physiological synergistic primitive library, the weighted combination specifically comprises:

[0028] Based on the aforementioned multidimensional physiological pattern matrix, using the formula:

[0029] ;

[0030] Calculate the co-factor components;

[0031] in, This represents a single coordinating primitive component used to generate universal physiological coordinating primitives. The signal strength weights are calibrated based on experimental data. This refers to the normalized energy amplitude of the corresponding physiological channel under the target cooperative mode, extracted from the physiological channel contribution mode matrix. This is a synchronization quantification obtained by calculating the peak point cross-correlation function of the differentiated time pattern sequences in the time evolution pattern matrix;

[0032] The calculated coordinating primitive components are called, and the original elements in the physiological channel contribution pattern matrix, the time evolution pattern matrix and the frequency activation pattern matrix are multiplied and weighted. The weighted results are then integrated to obtain the general physiological coordinating primitive library.

[0033] As a further aspect of the present invention, in the step of calculating the individual pattern expression weight vector, the determination process of the weighted sum of the reconstruction error and the numerical difference being lower than the convergence threshold is specifically as follows:

[0034] Through the formula:

[0035] ;

[0036] Calculate the portfolio loss value;

[0037] in, This represents the combined loss value generated during a single iteration of optimization. These are preset hyperparameters used to balance reconstruction accuracy and numerical sparsity. The reconstruction error is obtained by calculating the square root of the mean of the sum of squares of the differences between all corresponding elements in the reconstructed physiological matrix and the individual physiological data matrix. The numerical difference is obtained by calculating the sum of the absolute values ​​of the differences between all corresponding elements between the reconstructed physiological matrix and the individual physiological data matrix;

[0038] The convergence threshold is determined in the following way:

[0039] The first combined loss value calculated using the randomly initialized undetermined weight vector is used as the initial combined loss value, and the initial combined loss value is multiplied by a preset decay coefficient to obtain the convergence threshold.

[0040] As a further aspect of the present invention, the step of dividing the synchronization time window includes:

[0041] Set a fixed-length time window and an overlap rate between windows. The length of the time window is set according to empirical statistical values ​​to cover the duration of at least two complete respiratory cycles, and the overlap rate is set within the range of 50% to 75% according to the signal stability requirements.

[0042] Using the continuously detected R wave peaks in the electrocardiogram signal as the synchronization reference timestamp, the time interval between adjacent R waves is calculated and the time axis of other signal channels is calibrated. The time axis of the photocapacitance-volume pulse wave signal and the respiratory motion signal are aligned to ensure that all physiological signals are strictly synchronized at the beginning of each time window.

[0043] Based on the aligned multiple signal channels, starting from the synchronization start point, data segments are obtained by sliding and truncating according to the set time window length and the overlap rate, and the signal segment is established.

[0044] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0045] In this invention, a multi-channel physiological signal collaborative analysis framework is constructed by integrating multi-source information such as electrocardiogram, photoplethysmography (PPG), and respiratory motion. This eliminates reliance on the stability of a single sensor, effectively reducing interference caused by individual activity or poor sensor contact. Furthermore, common physiological collaborative patterns are extracted from the time-spectrum diagram of the group to form a standard state template. This template profoundly reveals the intrinsic correlation of multi-dimensional physiological signals under different respiratory states. The physiological data of the individual under test are coupled, decomposed, and weighted with this template. Based on the principle of minimizing reconstruction error, the standard state that best matches the individual pattern is found, achieving accurate discrimination of the individual's respiratory state and greatly improving the reliability and adaptability of the results. Attached Figure Description

[0046] Figure 1 This is a flowchart of the respiratory state discrimination method based on data fusion of the present invention;

[0047] Figure 2 This is a flowchart of the population physiological signal tensor generation process of the present invention;

[0048] Figure 3 This is a flowchart illustrating the construction process of the universal physiological synergistic primitive library and standard state template of this invention.

[0049] Figure 4 This is a flowchart of the calculation process for the individual pattern representation weight vector in this invention;

[0050] Figure 5 This is a flowchart illustrating the individual respiratory status determination process of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0052] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0053] Please see Figure 1 and Figure 2 This invention provides a technical solution: a respiratory state discrimination method based on data fusion, comprising the following steps:

[0054] S1: Acquire ECG, photoplethysmography pulse wave, and respiratory motion signals to construct multi-channel physiological signals, analyze the time-spectrum diagram of the multi-channel physiological signals and calculate the arithmetic mean, and generate a population physiological signal tensor;

[0055] The population physiological signal tensor specifically refers to the group-level time-frequency characteristic representation of multi-channel physiological signals.

[0056] The specific steps for obtaining the group physiological signal tensor are as follows:

[0057] S101: Acquires ECG signals, photocapacitance-volume pulse wave signals and respiratory motion signals, performs synchronous time window segmentation on ECG signals, photocapacitance-volume pulse wave signals and respiratory motion signals to obtain signal segments, constructs data frames by using multiple signal segments in the same time window as column vectors, stacks all time window data frames, and generates a multi-channel physiological signal matrix.

[0058] The steps for segmenting the synchronization time window include: setting a fixed-length time window and an overlap rate between windows. The length of the time window is set according to empirical statistical values ​​to cover the duration of at least two complete respiratory cycles, and the overlap rate is set within the range of 50% to 75% according to the signal stability requirements.

[0059] Using the continuously detected R wave peaks in the electrocardiogram signal as the synchronization reference timestamp, the time interval between adjacent R waves is calculated and the time axis of other signal channels is calibrated. The time axis of the photocapacitance-volume pulse wave signal and the respiratory motion signal are aligned to ensure that all physiological signals are strictly synchronized at the beginning of each time window.

[0060] Based on the aligned multiple signal channels, starting from the synchronization start point, data segments are obtained by sliding and truncating according to the set time window length and overlap rate to establish signal segments.

[0061] S102: Apply Hanning window function to preprocess each column of signal vector in the multi-channel physiological signal matrix, perform short-time Fourier transform to calculate energy spectral density, stitch the energy spectral density along the time axis to generate a two-dimensional time spectrum for each signal channel, and aggregate the two-dimensional time spectrums of all signal channels to obtain a single-user time spectrum set;

[0062] S103: Based on multiple single-user time-spectrum sets, each single-user time-spectrum set is stacked as a two-dimensional slice along the new user dimension to establish a three-dimensional data tensor. At the same time, the population averaging operator is called to calculate the arithmetic mean of the corresponding elements in the three-dimensional data tensor along the new user dimension and perform dimensionality reduction to generate a population physiological signal tensor.

[0063] In a specific implementation scenario, taking an adult male subject in a sleep monitoring environment as an example, physiological signals were collected for 10 consecutive minutes in a resting state. A three-lead electrocardiogram (ECG) recorder was used to acquire ECG signals at a sampling frequency of 500 Hz. A finger clip-on photoelectric sensor was used to acquire photoplethysmography (PPG) signals at a sampling frequency of 100 Hz, and a piezoelectric respiratory sensor strapped to the chest was used to acquire respiratory movement (RESP) signals at a sampling frequency of 50 Hz. First, the PPG and RESP signals were upsampled, unifying their sampling frequency to 500 Hz, consistent with the ECG signal, to achieve initial time resolution alignment.

[0064] Next, the segmentation parameters of the synchronization time window were set. To determine the time window length, a preliminary validation experiment was conducted: 30 healthy volunteers were recruited, and physiological signals were collected for 5 minutes each under three states: resting, deep breathing, and shallow breathing. Their respiratory rates were analyzed; under resting conditions, the respiratory rate ranged from 12 to 20 breaths / minute, i.e., each respiratory cycle lasted 3 to 5 seconds. Under deep breathing conditions, the cycle length increased to 6 to 8 seconds, while under shallow breathing conditions, the cycle length shortened to 2 to 3 seconds. To cover at least two of the longest complete respiratory cycles, a time window length of 16 seconds was selected. To determine the overlap rate, the effects of overlap rates of 50%, 60%, 70%, and 75% on the spectral leakage of the subsequent short-time Fourier transform results were tested. The results showed that, within a 16-second window, a 75% overlap rate provided the highest time resolution while ensuring signal stability and was the best among the tested overlap rates. Therefore, the final time window length was determined to be 16 seconds, and the overlap rate to be 75%.

[0065] Subsequently, the R-wave peak in the ECG signal was used as the synchronization reference. In the first 30 seconds of acquired ECG signal, the R-wave peak was identified using a detection algorithm based on thresholds and waveform morphology, and its occurrence timestamp was recorded as follows. For example, the timestamps for detecting the first four R-wave peaks are 1.20 seconds, 2.05 seconds, 2.95 seconds, and 3.80 seconds, respectively. The time intervals between adjacent R-waves, i.e., the RR intervals (RRI), are calculated to be 0.85 seconds, 0.90 seconds, and 0.85 seconds, respectively. Based on these RRI values, the time axes of the PPG and RESP signals are calibrated. Specifically, the main peak in the PPG signal is identified, its timestamp is compared with the most recent ECGR wave timestamp, and the average delay is calculated (e.g., the average lag of the PPG peak behind the R-wave peak is 0.25 seconds). The time axis of the entire PPG signal is shifted forward by 0.25 seconds. For the RESP signal, the start point of its inspiratory phase is aligned with the R-wave timestamp, and its time axis is fine-tuned using interpolation or resampling techniques to ensure that the phase relationship of the three signals remains consistent at each R-wave timestamp.

[0066] After timeline alignment, starting from the synchronization start point (e.g., the calibrated time point t=0), a data segment from 0 to 16 seconds is extracted. This segment contains three channels, each with 16 seconds × 500 Hz = 8000 data points. This 8000x3 data block constitutes the first signal segment. Due to the 75% overlap, the starting point of the next window is 16 seconds × (1-75%) = 4 seconds. Therefore, the second signal segment starts at t=4 seconds, extracting data from 4 to 20 seconds, also forming an 8000x3 data block. This process of continuously extracting data over 10 minutes (600 seconds) generates a total of (600-16) / 4+1 = 147 signal segments. The three column vectors (ECG, PPG, RESP) in each signal segment (8000x3) are used as columns of a data frame. These 147 data frames are then stacked chronologically to generate a multi-channel physiological signal matrix.

[0067] Continuing the previous steps, for any one of the 147 generated signal segments (e.g., the first one, i.e., data from t=0 to 16 seconds), this segment is an 8000x3 matrix containing 8000 time points and 3 physiological channels. For each column of the signal vector in this matrix—the ECG signal vector (8000x1), PPG signal vector (8000x1), and RESP signal vector (8000x1)—a Hanning window function is applied for preprocessing. The value of the Hanning window function smoothly approaches zero at both ends of the window. It is applied to the signal vector by multiplying each data point in the signal vector by the value of the window function at the corresponding position. Taking the ECG signal vector as an example, the vector consisting of its 8000 data points is multiplied element-wise by a Hanning window vector of length 8000 to obtain the windowed ECG signal vector.

[0068] Next, a Short-Time Fourier Transform (STFT) is performed on each windowed signal vector. The STFT parameters are set as follows: the number of FFT points (NFFT) is set to 1024, the window function (used within the STFT) is a Hanning window with a length of 512 sampling points, and the overlap length between windows is 256 sampling points. For an 8000-point ECG signal vector, the STFT divides it into multiple overlapping segments of 512 points each. For each segment, a 1024-point Fourier transform is calculated to obtain the spectrum for that time period. The square of the amplitude of each spectral component is calculated to obtain the energy spectral density. The energy spectral density results for all time periods are arranged along the time axis to generate a two-dimensional time-spectrum graph. The vertical axis of this graph represents frequency (0 Hz to 250 Hz, corresponding to half of the 1024-point FFT), and the horizontal axis represents time (from 0 seconds to 16 seconds). For example, this process can generate a 513x32 matrix, where 513 is the number of frequency bins and 32 is the number of time windows. Performing the same operation on the PPG and RESP signal vectors yields a 513x32 time-spectrum matrix for each. Now, for this user's data within the 0-16 second timeframe, three time-spectrum maps corresponding to ECG, PPG, and RESP are obtained. These three time-spectrum maps (three 513x32 matrices) are aggregated to form a single-user time-spectrum set within a single time window. Applying this process to all 147 signal segments yields the complete single-user time-spectrum set for the test subject.

[0069] To construct the population physiological signal tensor, multi-user data needs to be collected. Following the procedures in S101 and S102, data was collected and processed from a population of 50 subjects, with each subject's data being processed into a single-user time-spectrum dataset. Assuming we focus on a specific frequency band of respiratory rate (e.g., 0.1 Hz to 0.8 Hz) and a specific temporal evolution pattern, a representative sub-region is selected from each 513x32 time-spectrum plot, for example, a 64x32 sub-matrix (64 frequency bins, 32 time bins). Therefore, for each signal channel (ECG, PPG, RESP) of each subject, there is a corresponding 64x32 two-dimensional time-spectrum plot.

[0070] Next, the time-spectral data of these 50 subjects are stacked. Taking the ECG channel as an example, the ECG time-spectral data of the 50 subjects (all 64x32 matrices) are treated as two-dimensional slices and stacked along a new "user" dimension. This operation generates a three-dimensional data tensor with dimensions of 64x32x50 (frequency x time x user). Similarly, this operation is performed on the PPG and RESP channels, resulting in two more 64x32x50 three-dimensional data tensors. These three tensors are concatenated along the channel dimensions to form a four-dimensional data tensor with dimensions of 64x32x3x50, which is the initial set of three-dimensional data tensors (for ease of understanding, the first three dimensions are considered as a single three-dimensional data tensor).

[0071] Then, the group averaging operator is invoked to calculate the arithmetic mean along the new user dimension. Specifically, for each fixed (frequency, time, channel) coordinate point in the tensor, there are 50 values ​​from different subjects. For example, for the coordinate point (frequency index = 10, time index = 5, channel index = 1), its corresponding 50 energy spectral density values ​​might be [0.85, 0.91, 0.88, ..., 0.93]. The arithmetic mean of these 50 values ​​is calculated as: (0.85 + 0.91 + 0.88 + ... + 0.93) / 50 = 0.89. This average is used as the new value of the tensor at that coordinate point after dimensionality reduction. This averaging operation is performed on all 64x32x3 coordinate points in the four-dimensional tensor. This process compresses the user dimension (dimension 50), ultimately generating a 64x32x3 three-dimensional tensor. This three-dimensional tensor no longer contains individual-specific information, but reflects the average time-frequency characteristics of the entire population under a specific physiological state. This is what the present invention refers to as the population physiological signal tensor.

[0072] Please see Figure 1 and Figure 3 S2: Perform constrained nonnegative tensor decomposition on the population physiological signal tensor, obtain and combine the physiological channel contribution pattern matrix, time evolution pattern matrix, and frequency activation pattern matrix to construct a general physiological coordination primitive library, and calibrate the general physiological coordination primitive library to generate standard state templates;

[0073] The universal physiological synergistic primitive library includes a physiological channel contribution pattern matrix, a temporal evolution pattern matrix, and a frequency activation pattern matrix.

[0074] The standard state template is for configuring the collaborative primitive parameters for multiple respiratory states.

[0075] The specific steps for obtaining the standard state template are as follows:

[0076] S201: Perform constrained nonnegative tensor decomposition on the population physiological signal tensor, decompose the population physiological signal tensor into mutually independent contribution mode factors, evolution mode factors and activation mode factors, and reconstruct the contribution mode factors, evolution mode factors and activation mode factors into physiological channel contribution mode matrix, time evolution mode matrix and frequency activation mode matrix respectively, and establish a multidimensional physiological mode matrix group.

[0077] S202: Call the multidimensional physiological pattern matrix group, obtain the signal strength weight and time synchronization weight defined by the preset coordination rules, and perform weighted combination on the elements of the physiological channel contribution pattern matrix, time evolution pattern matrix and frequency activation pattern matrix according to the signal strength weight and time synchronization weight to construct a general physiological coordination primitive library;

[0078] In the construction steps of the universal physiological synergistic primitive library, the weighted combination is specifically as follows:

[0079] Based on a multidimensional physiological pattern matrix, using the formula:

[0080] ;

[0081] Calculate the co-factor components;

[0082] in, This represents a single coordinating primitive component used to generate universal physiological coordinating primitives. The signal strength weights are determined based on experimental data. This refers to the normalized energy amplitude of the corresponding physiological channel under the target cooperative mode, extracted from the physiological channel contribution mode matrix. The synchronization metric is obtained by calculating the peak point cross-correlation function of the differentiated time pattern sequences in the time evolution pattern matrix;

[0083] The calculated coordinating primitive components are called, and the original elements in the physiological channel contribution pattern matrix, time evolution pattern matrix and frequency activation pattern matrix are multiplied and weighted. The weighted results are then integrated to obtain a general physiological coordinating primitive library.

[0084] S203: Based on a general physiological coordination primitive library, the activation frequency and duration of multiple primitives are calculated and their products are obtained to generate coordination mode stability. A coordination mode activation benchmark value is set, all physiological coordination primitives whose coordination mode stability exceeds the coordination mode activation benchmark value are screened, and numerical calibration is performed to generate a standard state template.

[0085] A weighted combination of the physiological channel contribution pattern matrix, the temporal evolution pattern matrix, and the frequency activation pattern matrix is ​​used to construct a universal physiological coordinating primitive library. The core of this process lies in the coordinating primitive components. The calculation.

[0086] formula The detailed explanation is as follows: This formula is used to calculate a coordinating primitive component. The logic is that an effective physiological coordination pattern depends not only on the signal strength of each physiological signal channel but also on their temporal synchronization. The formula uses an adjustable weight... This combines these two characteristics. This represents the final calculated value of a single coordinating primitive component, which integrates energy and synchronization information; It is the signal strength weight, a scalar between 0 and 1, used to balance the importance of energy contribution and time synchronization; It is the normalized energy amplitude of the corresponding physiological channel under a specific cooperative mode, extracted from the physiological channel contribution mode matrix, reflecting the signal strength of a single channel; It is a synchronization quantification value obtained by calculating the cross-correlation of peak points of different time pattern sequences in the time evolution pattern matrix, reflecting the degree of temporal coupling between channels. The calculation logic of the formula is: using... Weight the energy amplitude, using The synchronization degree is weighted and then the two values ​​are added together to obtain a comprehensive evaluation value.

[0087] Parameter acquisition and calculation examples: 1. Signal strength weighting The weights were set through a cross-validation experiment. A validation dataset containing 100 labeled respiratory states (normal breathing, rapid breathing, slow breathing, and apnea) was used. The weights were then... Traverse from 0.1 to 0.9, with a step size of 0.1. For each... For each value, a corresponding general physiological coordinating primitive library is generated, and this library is used to determine the respiratory state of the validation dataset and calculate its accuracy.

[0088] Weight value ( ) Discrimination accuracy (%) 0.1 88.5 0.2 91.2 0.3 93.6 0.4 95.1 0.5 96.8 0.6 97.5 0.7 98.1 0.8 97.2 0.9 96.4

[0089] Table 1. Experimental Data for Signal Strength Weight Calibration

[0090] As shown in Table 1, when When set to 0.7, the system achieved its highest discrimination accuracy of 98.1%. Therefore, in this embodiment, the selected... .

[0091] Normalized energy amplitude Suppose that from the physiological channel contribution pattern matrix, the k-th primitive related to the "normal breathing" pattern is extracted, with corresponding channel contribution values ​​of ECG=1.2, PPG=1.5, and RESP=3.8. To calculate the RESP channel... First, the contribution vector [1.2, 1.5, 3.8] of this primitive is normalized using the L2 norm. The norm is... The normalized vector is [1.2 / 4.26, 1.5 / 4.26, 3.8 / 4.26]. [0.28, 0.35, 0.89]. Therefore, the normalized energy amplitude of the RESP channel in this mode. It is 0.89.

[0092] Synchronized metric values From the time evolution pattern matrix, extract the time evolution pattern sequences corresponding to the PPG and RESP channels, respectively, related to the aforementioned "normal breathing" pattern. Assume the peak times (in seconds) of these two sequences over a given period are: Sequence A (PPG): [2.1, 5.5, 8.9, 12.3]; Sequence B (RESP): [2.0, 5.4, 8.8, 12.2]. Calculate the Pearson cross-correlation function of these two sequences at a delay of 0. Since the peak times of the two sequences are very close, their cross-correlation coefficient is calculated to be close to 1.0. In this example, the calculated... .

[0093] Example calculation:

[0094] Substitute the above parameter values ​​into the formula:

[0095] ;

[0096] This result indicates that the overall contribution of the RESP channel in the "normal breathing" mode is 0.917. The underlying principle is based on introducing signal strength weights. For normalized energy amplitude and synchronization metric By employing weighted combinations, the construction of synergistic primitives not only considers the energy contribution intensity of each physiological channel but also incorporates their temporal synergistic dynamics, thereby more comprehensively characterizing the coupling relationships within the physiological system.

[0097] S203: After constructing the general physiological coordination primitive library, it needs to be screened and calibrated to generate standard state templates. This process first calculates the coordination pattern stability of each physiological coordination primitive. Taking a coordination primitive representing "shallow and rapid breathing" as an example, an analysis was conducted on a 300-second population average data set. It was observed that this primitive was activated (its activation coefficient exceeded a specific activation threshold) 25 times, i.e., the activation frequency was 25 times / 300 seconds = 0.0833 Hz, or 5 times / minute. Simultaneously, the duration of each activation was recorded, resulting in a duration sequence of [3.1, 3.5, 3.2, 3.8, ..., 3.3] seconds. The arithmetic mean of this sequence was calculated to be an average duration of 3.4 seconds. The coordination pattern stability of this coordination primitive was calculated as the product of the activation frequency and the average duration: 5 times / minute × 3.4 seconds = 17.0 (the unit can be considered as times / second / minute).

[0098] Next, a baseline value for co-mode activation was established. This baseline value was based on the analysis of a reference dataset with known states. This dataset contained physiological data from 50 subjects under four defined states: normal breathing, deep and slow breathing, shallow and rapid breathing, and simulated apnea. A primitive library was applied to this data to calculate the co-mode stability of all primitives under each state. Statistical analysis revealed that when a primitive correctly corresponded to its physiological state (e.g., the "deep and slow breathing" primitive was activated when the subject was performing deep and slow breathing), its stability value ranged from 12.0 to 45.0; while when a primitive did not match the physiological state (e.g., the "deep and slow breathing" primitive was accidentally activated when the subject was performing shallow and rapid breathing), its stability value was below 8.0. To improve the accuracy and reliability of primitive selection, the baseline value for co-mode activation was set within the gap between these two distributions. Specifically, the average of the maximum stability value of 8.0 under mismatched states and the minimum stability value of 12.0 under matched states was selected, i.e., (8.0 + 12.0) / 2 = 10.0. Therefore, the baseline value for activating the collaborative mode is set to 10.0.

[0099] Next, a screening process is performed. The stability of the "shallow and rapid breathing" synergistic motif, calculated above, is 17.0, exceeding the set baseline of 10.0; therefore, this physiological synergistic motif is retained. Conversely, if the stability of another "noise pattern" motif is calculated to be 4.5, which is below 10.0, then that motif is filtered out. This process iterates through all motifs, retaining only those with a stability exceeding 10.0.

[0100] Finally, all selected physiological coordinating primitives are numerically calibrated to generate standard state templates. For each respiratory state (e.g., "normal breathing"), there may be multiple coordinating primitives learned from different subsets of subject groups or different time periods, all of which have passed stability screening. For example, after screening, three highly similar "normal breathing" primitives remain. The numerical calibration process involves averaging the corresponding parameters of these three primitives (weights from the physiological channel contribution pattern, time evolution pattern, and frequency activation pattern matrix). For example, the channel contribution vector of primitive 1 is [0.3, 0.4, 0.9], that of primitive 2 is [0.28, 0.42, 0.88], and that of primitive 3 is [0.31, 0.39, 0.91]. The channel contribution vector of the calibrated standard "normal breathing" template is the average of these three parameters: [(0.3+0.28+0.31) / 3,(0.4+0.42+0.39) / 3,(0.9+0.88+0.91) / 3]=[0.297,0.403,0.897]. After averaging all parameters, a unique standard vector representing the "normal breathing" state is obtained, and a "normal breathing" label is associated with it. The same operation is performed on other selected states (such as "shallow and rapid breathing", "deep and slow breathing", etc.), ultimately forming a set containing multiple standard vectors and corresponding state labels, which is the standard state template.

[0101] Please see Figure 1 and Figure 4 S3: Collect multi-channel physiological signals of the individual to be tested and construct the individual physiological data matrix. Call the general physiological collaborative primitive library to perform coupled matrix decomposition on the individual physiological data matrix. Calculate the individual pattern expression weight vector with the constraint of minimizing reconstruction error and numerical difference.

[0102] The individual pattern representation weight vector is the projection weight of individual physiological data onto the general physiological coordinating primitive library.

[0103] The specific steps for calculating the individual pattern representation weight vector are as follows:

[0104] S301: Collect multi-channel physiological signals of the individual under resting state, align the multi-channel physiological signals with timestamps according to the preset sampling frequency and channel index, and aggregate the aligned signal data stream into a two-dimensional data structure, where row vectors represent time series and column vectors represent signal channels, and establish an individual physiological data matrix;

[0105] S302: Call the general physiological coordinating primitive library, perform non-negative coupled matrix decomposition on the individual physiological data matrix according to the preset physiological primitives in the library, split the individual physiological data matrix into primitive pattern matrix and weight coefficient matrix, extract the primitive pattern matrix and concatenate it horizontally to generate the decomposed coordinating pattern matrix.

[0106] S303: Based on the decomposition of the cooperative pattern matrix, a randomly initialized undetermined weight vector is set. The decomposition of the cooperative pattern matrix and the undetermined weight vector are multiplied by matrix to obtain the reconstructed physiological matrix. The root mean square error between the reconstructed physiological matrix and the individual physiological data matrix is ​​calculated as the reconstruction error, and the Manhattan distance is calculated as the numerical difference. The element values ​​in the undetermined weight vector are iteratively adjusted by the gradient descent method until the weighted sum of the reconstruction error and the numerical difference is lower than the convergence threshold, thus obtaining the individual pattern expression weight vector.

[0107] In the step of calculating the individual pattern representation weight vector, the specific process for determining whether the weighted sum of reconstruction error and numerical difference is below the convergence threshold is as follows:

[0108] Through the formula:

[0109] ;

[0110] Calculate the portfolio loss value;

[0111] in, This represents the combined loss value generated during a single iteration of optimization. These are preset hyperparameters used to balance reconstruction accuracy and numerical sparsity. The reconstruction error is calculated by taking the square root of the mean of the sum of squares of the differences between all corresponding elements in the reconstructed physiological matrix and the individual physiological data matrix. The numerical difference is obtained by calculating the sum of the absolute values ​​of the differences between all corresponding elements between the reconstructed physiological matrix and the individual physiological data matrix;

[0112] The convergence threshold is determined in the following way:

[0113] The first combined loss value calculated using the randomly initialized undetermined weight vector is used as the initial combined loss value, and the initial combined loss value is multiplied by a preset decay coefficient to obtain the convergence threshold.

[0114] After obtaining the decomposed collaborative pattern matrix, iterative optimization is needed to calculate the individual pattern representation weight vectors. The core of this process is minimizing a combined loss value. .

[0115] formula The detailed explanation is as follows: This formula defines the objective function in the optimization process, namely the combined loss value generated in a single iteration. The logic is that the optimal individual pattern representation weight vector should enable the physiological matrix reconstructed using this weight and the general primitive library to approximate the original individual physiological data matrix in two aspects: numerical accuracy and pattern sparsity. In the formula, It is the total loss; It is a hyperparameter between 0 and 1, used to balance the importance of reconstruction accuracy and numerical sparsity; It is the reconstruction error, which is obtained by calculating the square root of the mean of the sum of squares of the differences between all corresponding elements between the reconstructed physiological matrix and the individual physiological data matrix. It measures the accuracy of the reconstruction. The Manhattan distance (RMSE) is a numerical difference calculated by summing the absolute values ​​of the differences between all corresponding elements of two matrices. It acts as a proxy for sparsity; a smaller RMSE value indicates a more concentrated distribution of the differences between the two matrices, which helps generate a sparse weight vector. The formula's logic is to weight and sum the RMSE and Manhattan distance to obtain a comprehensive loss metric. The goal of gradient descent is to adjust the undetermined weight vector to achieve this... Minimize the value.

[0116] Parameter acquisition and examples: 1. Hyperparameters The determination of this parameter also depends on experimental validation. A separate validation dataset will be used to test different... The impact of the value on the final judgment result.

[0117] hyperparameter values ​​( ) Mean Reconstruction Error (RMSE) Average weight sparsity Discrimination accuracy (%) 0.2 0.08 0.92 94.5 0.4 0.12 0.85 96.7 0.6 0.15 0.78 98.2 0.8 0.25 0.65 97.1

[0118] Table 2 Hyperparameters Select Experiment Data Table

[0119] As shown in Table 2 (where weight sparsity is the proportion of zero elements in the weight vector), when... At that time, although the reconstruction error increased slightly, the discrimination accuracy reached its peak. This indicates that a moderate sparsity constraint (by...) The introduction of [this method] helps the model focus on the most critical physiological patterns, thereby improving discrimination performance. Therefore, this embodiment selects [this method]. .

[0120] Reconstruction error Numerical differences Assume that both the individual physiological data matrix and the reconstructed physiological matrix are simplified to a small 2x2 matrix. Individual physiological data matrix In a certain iteration, the reconstructed physiological matrix is ​​generated from the undetermined weight vector. Calculate the difference between the two.

[0121] Determining the convergence threshold: Before the iteration begins, the first combined loss value is calculated using a randomly initialized undetermined weight vector (e.g., [0.25, 0.25, 0.25, 0.25] for the four standard states), denoted as . Assuming the calculation yields... The attenuation coefficient is empirically set to 0.0001, a commonly used value in similar optimization problems to ensure convergence accuracy. Convergence threshold = .

[0122] Example calculation:

[0123] Substitute the above parameters into the loss function formula:

[0124]

[0125] .

[0126] Please see Figure 1 and Figure 5 S4: Calculate the Euclidean distance between the individual pattern expression weight vector and the standard state template, and determine the breathing state corresponding to the minimum Euclidean distance as the individual breathing state discrimination result.

[0127] The individual respiratory state discrimination result specifically refers to the respiratory state category that is closest to the standard state template.

[0128] The specific steps for obtaining the individual respiratory status assessment results are as follows:

[0129] S401: Based on the storage of multiple standard vectors and associated respiratory state labels in the standard state template, call the individual pattern expression weight vector, and calculate the Euclidean distance by performing the sum of squares and square roots of the differences between it and all standard vectors in the standard state template. Pair each generated distance value with the respiratory state label associated with the standard vector to form a state distance entry. Aggregate all state distance entries to generate a state distance vector group.

[0130] S402: Traverse all state distance entries in the state distance vector group, extract the distance value encapsulated in each state distance entry, retrieve the minimum value through loop comparison, lock the state distance entry with the minimum value, obtain the position identifier in the state distance vector group, and establish the minimum distance index;

[0131] S403: Call the minimum distance index and the state distance vector group. At the same time, based on the location information indicated by the minimum distance index, perform direct addressing within the state distance vector group to lock the corresponding state distance entry. Access the state distance entry and parse out the encapsulated breathing state label. Confirm the breathing state label as the discrimination conclusion and obtain the individual breathing state discrimination result.

[0132] After obtaining the pattern representation weight vector of the individual to be tested, the final discrimination stage begins. Assume the standard state template contains four standard breathing states and their corresponding standard vectors (i.e., templates), with a vector dimension of 10: - Template 1 (normal breathing): Template 2 (Deep, slow breathing): Template 3 (shallow, rapid breathing): Template 4 (Sleep Apnea):

[0133] The calculated individual pattern representation weight vector is: Next, the vector W is compared with each standard vector V_i in the standard state template by performing the square root of the sum of the squared differences, i.e., calculating the Euclidean distance.

[0134] Calculate the Euclidean distance D1 between W and V1 (normal breathing): Assuming that the sum of squares of the differences is calculated to be 0.0081, then Pair the distance value 0.09 with the associated breathing status label “normal breathing” to form the first status distance entry: (0.09, “normal breathing”).

[0135] Calculate the Euclidean distance D2 between W and V2 (deep, slow breathing): Because the corresponding elements of W and V2 differ significantly, for example, the square of the difference in the first term is... This is much larger than the terms in D1. Assuming the calculated total sum of squares is 3.88, then... This constitutes the second state distance entry: (1.97, "Deep and slow breathing").

[0136] Similarly, calculate the Euclidean distance D3 between W and V3 (shallow, rapid breathing), assuming the result is... This constitutes the third state distance entry: (1.24, "shallow and rapid breathing"). Calculate the Euclidean distance D4 between W and V4 (apnea), assuming the result is... This constitutes the fourth state distance entry: (4.51, "apnea").

[0137] These four state distance entries are aggregated together to generate a state distance vector group: [(0.09, "Normal breathing"), (1.97, "Deep and slow breathing"), (1.24, "Shallow and rapid breathing"), (4.51, "Pause in breathing")].

[0138] By traversing the state distance vector group and comparing the distance values ​​of each entry, the entry with the smallest value and its corresponding position index are found. In this example, the minimum value is 0.09, and its position index is 1, thus establishing the minimum distance index.

[0139] The minimum distance index (with a value of 1) established in the previous step and the state distance vector group are invoked. Based on the index value of 1, a direct addressing operation is performed within the state distance vector group. In programming implementation, this is equivalent to accessing the first element of an array or list (assuming the index starts from 1). The corresponding locked state distance entry is (0.09, "normal breathing"). Next, the internal structure of this entry is accessed, parsed, and the breathing state label encapsulated within is extracted. In this example, the extracted label is the string "normal breathing". This breathing state label is confirmed as the final discrimination conclusion. Thus, the breathing state discrimination result of the tested individual is obtained as "normal breathing".

[0140] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the scope of protection defined by the claims of the present invention.

Claims

1. A respiratory state discrimination method based on data fusion, characterized in that, Includes the following steps: S1: Acquire ECG, photoplethysmography pulse wave, and respiratory motion signals to construct multi-channel physiological signals, analyze the time-spectrum diagram of the multi-channel physiological signals and calculate the arithmetic mean, and generate a population physiological signal tensor; S2: Perform constrained nonnegative tensor decomposition on the population physiological signal tensor to obtain and combine the physiological channel contribution pattern matrix, time evolution pattern matrix, and frequency activation pattern matrix to construct a general physiological coordination primitive library, and calibrate the general physiological coordination primitive library to generate a standard state template; S3: Collect multi-channel physiological signals of the individual to be tested and construct an individual physiological data matrix. Call the general physiological collaborative primitive library to perform coupled matrix decomposition on the individual physiological data matrix. Calculate the individual pattern expression weight vector with the constraint of minimizing reconstruction error and numerical difference. S4: Calculate the Euclidean distance between the individual pattern expression weight vector and the standard state template, and determine the breathing state corresponding to the smallest Euclidean distance as the individual breathing state discrimination result.

2. The respiratory state discrimination method based on data fusion according to claim 1, characterized in that, The group physiological signal tensor specifically refers to the group-level time-frequency feature representation of multi-channel physiological signals. The general physiological collaborative primitive library includes the physiological channel contribution pattern matrix, the time evolution pattern matrix, and the frequency activation pattern matrix. The standard state template is the collaborative primitive parameter configuration for multiple respiratory states. The individual pattern expression weight vector is the projection weight describing the individual physiological data on the general physiological collaborative primitive library. The individual respiratory state discrimination result specifically refers to the respiratory state category with the smallest distance from the standard state template.

3. The respiratory state discrimination method based on data fusion according to claim 2, characterized in that, The specific steps for obtaining the group physiological signal tensor are as follows: S101: Acquire electrocardiogram (ECG) signals, photocapacitance-volume pulse wave (PVP) signals, and respiratory motion signals; perform synchronous time window segmentation on the ECG signals, PPV signals, and respiratory motion signals to obtain signal segments; construct data frames by using multiple signal segments within the same time window as column vectors; stack all the data frames of the time windows to generate a multi-channel physiological signal matrix. S102: Apply Hanning window function preprocessing to each column of signal vector in the multi-channel physiological signal matrix, perform short-time Fourier transform to calculate energy spectral density, and stitch the energy spectral density along the time axis to generate a two-dimensional time spectrum for each signal channel. Aggregate the two-dimensional time spectrums of all the signal channels to obtain a single-user time spectrum set. S103: Based on multiple single-user time-spectrum sets, each single-user time-spectrum set is stacked as a two-dimensional slice along the new user dimension to establish a three-dimensional data tensor. At the same time, the population averaging operator is called to calculate the arithmetic mean of the corresponding elements in the three-dimensional data tensor along the new user dimension and perform dimensionality reduction to generate a population physiological signal tensor.

4. The respiratory state discrimination method based on data fusion according to claim 3, characterized in that, The specific steps for obtaining the standard state template are as follows: S201: Perform constrained nonnegative tensor decomposition on the population physiological signal tensor, decompose the population physiological signal tensor into mutually independent contribution mode factors, evolution mode factors and activation mode factors, and reconstruct the contribution mode factors, the evolution mode factors and the activation mode factors into physiological channel contribution mode matrix, time evolution mode matrix and frequency activation mode matrix respectively, and establish a multidimensional physiological mode matrix group. S202: Call the multidimensional physiological pattern matrix group, obtain the signal strength weight and time synchronization weight defined by the preset coordination rule, and perform weighted combination on the elements of the physiological channel contribution pattern matrix, the time evolution pattern matrix and the frequency activation pattern matrix according to the signal strength weight and the time synchronization weight to construct a general physiological coordination primitive library; S203: Based on the general physiological coordination primitive library, calculate the activation frequency and duration of multiple primitives and multiply them to generate coordination mode stability, set a coordination mode activation benchmark value, screen all physiological coordination primitives whose coordination mode stability exceeds the coordination mode activation benchmark value, perform numerical calibration, and generate a standard state template.

5. The respiratory state discrimination method based on data fusion according to claim 4, characterized in that, The specific steps for calculating the individual pattern representation weight vector are as follows: S301: Collect multi-channel physiological signals of the individual under resting state, align the multi-channel physiological signals with timestamps according to the preset sampling frequency and channel index, and aggregate the aligned signal data stream into a two-dimensional data structure, wherein the row vector represents the time series and the column vector represents the signal channel, and establish an individual physiological data matrix; S302: Call the general physiological coordinating primitive library, and according to the preset physiological primitives in the library, perform non-negative coupled matrix decomposition on the individual physiological data matrix to split the individual physiological data matrix into a primitive pattern matrix and a weight coefficient matrix. Extract the primitive pattern matrix and concatenate it horizontally to generate a decomposed coordinating pattern matrix. S303: Based on the decomposed cooperative pattern matrix, a randomly initialized undetermined weight vector is set. The decomposed cooperative pattern matrix and the undetermined weight vector are multiplied by matrix to obtain a reconstructed physiological matrix. The root mean square error between the reconstructed physiological matrix and the individual physiological data matrix is ​​calculated as the reconstruction error, and the Manhattan distance is calculated as the numerical difference. The element values ​​in the undetermined weight vector are iteratively adjusted by gradient descent until the weighted sum of the reconstruction error and the numerical difference is lower than the convergence threshold, thus obtaining the individual pattern expression weight vector.

6. The respiratory state discrimination method based on data fusion according to claim 5, characterized in that, The specific steps for obtaining the individual respiratory status discrimination result are as follows: S401: Based on the multiple standard vectors and associated breathing state labels stored in the standard state template, the individual pattern expression weight vector is called, and the Euclidean distance is calculated by performing the sum of squares and square roots of the differences between the individual and all the standard vectors in the standard state template. Each generated distance value is paired with the breathing state label associated with the standard vector to form a state distance entry. All the state distance entries are aggregated to generate a state distance vector group. S402: Traverse all the state distance entries in the state distance vector group, extract the distance value encapsulated in each state distance entry, retrieve the minimum value by loop comparison, lock the state distance entry with the minimum value, obtain the position identifier in the state distance vector group, and establish a minimum distance index; S403: Invoke the minimum distance index and the state distance vector group, and based on the position information indicated by the minimum distance index, perform direct addressing within the state distance vector group to lock the corresponding state distance entry, access the state distance entry and parse out the encapsulated breathing state label, confirm the breathing state label as the discrimination conclusion, and obtain the individual breathing state discrimination result.

7. The respiratory state discrimination method based on data fusion according to claim 4, characterized in that, In the construction steps of the general physiological synergistic primitive library, the weighted combination specifically includes: Based on the aforementioned multidimensional physiological pattern matrix, using the formula: ； Calculate the co-factor components; in, This represents a single coordinating primitive component used to generate universal physiological coordinating primitives. The signal strength weights are calibrated based on experimental data. This refers to the normalized energy amplitude of the corresponding physiological channel under the target cooperative mode, extracted from the physiological channel contribution mode matrix. This is a synchronization quantification obtained by calculating the peak point cross-correlation function of the differentiated time pattern sequences in the time evolution pattern matrix; The calculated coordinating primitive components are called, and the original elements in the physiological channel contribution pattern matrix, the time evolution pattern matrix and the frequency activation pattern matrix are multiplied and weighted. The weighted results are then integrated to obtain the general physiological coordinating primitive library.

8. The respiratory state discrimination method based on data fusion according to claim 5, characterized in that, In the step of calculating the individual pattern representation weight vector, the process for determining whether the weighted sum of the reconstruction error and the numerical difference is lower than the convergence threshold is as follows: Through the formula: ； Calculate the portfolio loss value; in, This represents the combined loss value generated during a single iteration of optimization. These are preset hyperparameters used to balance reconstruction accuracy and numerical sparsity. The reconstruction error is obtained by calculating the square root of the mean of the sum of squares of the differences between all corresponding elements in the reconstructed physiological matrix and the individual physiological data matrix. The numerical difference is obtained by calculating the sum of the absolute values ​​of the differences between all corresponding elements between the reconstructed physiological matrix and the individual physiological data matrix; The convergence threshold is determined in the following way: The first combined loss value calculated using the randomly initialized undetermined weight vector is used as the initial combined loss value, and the initial combined loss value is multiplied by a preset decay coefficient to obtain the convergence threshold.

9. The respiratory state discrimination method based on data fusion according to claim 3, characterized in that, The steps for dividing the synchronization time window include: Set a fixed-length time window and an overlap rate between windows. The length of the time window is set according to empirical statistical values ​​to cover the duration of at least two complete respiratory cycles, and the overlap rate is set within the range of 50% to 75% according to the signal stability requirements. Using the continuously detected R wave peaks in the electrocardiogram signal as the synchronization reference timestamp, the time interval between adjacent R waves is calculated and the time axis of other signal channels is calibrated. The time axis of the photocapacitance-volume pulse wave signal and the respiratory motion signal are aligned to ensure that all physiological signals are strictly synchronized at the beginning of each time window. Based on the aligned multiple signal channels, starting from the synchronization start point, data segments are obtained by sliding and truncating according to the set time window length and the overlap rate, and the signal segment is established.

Images