Multi-parameter non-invasive health monitoring method and device, system and storage medium based on FPGA-MCU heterogeneous architecture

Abstract

This invention discloses a multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture, comprising the following steps: S1, multi-frequency signal excitation and data acquisition; S2, FPGA parallel real-time impedance calculation, data transmission, and fusion modeling; S3, multi-physiological parameter fusion prediction; and S4, risk warning and health assessment. This invention extends the excitation frequency to 2MHz, utilizing the physical characteristic that the capacitance effect of the stratum corneum of the skin tends to zero under high-frequency conditions. Combined with low-squares linear fitting in the high-frequency band, an equivalent impedance model of deep human tissue is established, achieving systematic elimination of interference from the skin surface. The extracted equivalent pure resistance parameters of deep tissue are highly correlated with deep physiological indicators such as blood volume and tissue fluid content, significantly overcoming the fundamental limitation of existing bioelectrical impedance technology, which can only obtain surface impedance. Furthermore, the deep human impedance spectrum can more effectively detect physiological parameters such as blood glucose and blood pressure.

Description

Technical Field

[0001] This invention belongs to the field of medical and health monitoring technology, specifically relating to a multi-parameter non-invasive health monitoring method, device, system, and storage medium based on an FPGA-MCU heterogeneous architecture. Background Technology

[0002] With the accelerating aging of the population and changes in lifestyle, cardiovascular disease, hypertension, and diabetes have become major chronic diseases threatening human health. Their early asymptomatic and insidious development pose significant challenges to clinical screening and prevention.

[0003] In the field of impedance monitoring, existing technologies typically employ integrated impedance conversion chips such as the AD5933 for single-frequency or low-frequency sweeps. These chips have low sampling rates and are easily affected by the capacitance effect of the epidermal stratum corneum, making it difficult to effectively penetrate deep tissues to extract pure resistance information. In terms of cardiac output assessment, the traditional Kubicek impedance cardiac recording (ICG) technique relies on the assumption of a static baseline, and its accuracy is easily reduced when faced with complex motion interference. In terms of non-invasive blood pressure monitoring, existing techniques based on photoplethysmography (PPG) and pulse wave conduction time (PTT) are highly susceptible to changes in vascular smooth muscle tension, requiring frequent cuff calibration.

[0004] Existing technologies are mostly limited to low-frequency sampling and superficial tissue impedance acquisition, which are easily affected by the capacitance effect of the skin surface. They are difficult to accurately obtain the content of deep tissue fluid and hemodynamic information. Moreover, they lack the ability to acquire multi-frequency and multi-channel data in a coordinated manner and the collaborative monitoring mechanism that integrates multi-frequency, deep and multi-modal features, resulting in insufficient accuracy in assessing cardiovascular function and metabolic status. Summary of the Invention

[0005] (1) Technical problems to be solved To address the shortcomings of existing technologies, the present invention aims to provide a multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture, which seeks to solve the problems of superficial measurement limitations, single-parameter monitoring bottlenecks, and poor patient compliance caused by invasiveness in existing bioelectrical impedance health monitoring technologies.

[0006] (2) Technical solution To address the aforementioned technical problems, this invention provides a multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture, comprising the following steps: S1. Multi-frequency signal excitation and data acquisition: Under the scheduling of the control unit, multi-frequency sinusoidal excitation signals are generated and injected into the human body through the electrode array. Then, the synchronous high-frequency digital sampling of the human body voltage and current signals is completed to obtain a high-speed parallel digital sampling sequence. Step S2: FPGA Parallel Real-Time Impedance Calculation, Data Transmission, and Fusion Modeling: The digital sampling sequence is input into the FPGA processing unit, and the real and imaginary parts of the signal at each excitation frequency are extracted through parallel computation. The real and imaginary parts are serially transmitted to the MCU processing unit via the SPI protocol for accurate floating-point calculation of complex impedance to obtain the complex impedance data at each frequency. The calculated complex impedance data is then encapsulated and transmitted to the host computer client via Wi-Fi. The host computer client uses EEMD to adaptively decompose and reduce noise in the complex impedance data. The complex impedance data at each frequency is then fused to construct an equivalent impedance model of deep human tissue. Step S3, Multi-physiological parameter fusion prediction: Extract multi-dimensional fusion features from the denoised deep human tissue impedance equivalent model to form a 42-dimensional multi-parameter fusion feature vector, input it into a customized state space neural network, and output multiple core physiological parameters such as heart rate, respiratory rate, blood pressure, and blood glucose. Step S4, Risk Warning and Health Assessment: Input multiple core physiological parameters into the pre-trained inference model and output real-time health status assessment and multi-level risk warning signals.

[0007] Preferably, in step S1, a multi-frequency sinusoidal excitation signal in the range of 1kHz to 2MHz is generated. After the excitation signal is processed by the analog front end, it is injected into the human body through the electrode array. The amplitude of the excitation current injected into the human body is strictly limited to a safety threshold below the milliampere level.

[0008] Furthermore, in step S1, a dual-channel high-speed ADC with a sampling rate of 65 MSPS and a resolution of 14 bits is used to complete the synchronous high-frequency digital sampling of voltage and current signals to obtain a high-speed parallel digital sampling sequence at a sampling rate far exceeding that of the Nyquist theorem.

[0009] Furthermore, in step S2, the parallel computation adopts an iterative algorithm based on recursive equations, and in step S2, the C language is used to perform the accurate floating-point solution of the final complex impedance.

[0010] Furthermore, in step S2, the host computer client first performs dynamic baseline calibration on the calculated complex impedance data and then uses EEMD to adaptively decompose and reduce noise in the complex impedance data.

[0011] A multi-parameter non-invasive health monitoring device based on an FPGA-MCU heterogeneous architecture includes: The hardware acquisition module is used to synthesize multi-frequency excitation signals and inject them into the human body through an electrode array, and to condition and digitally sample the response signals. The FPGA digital processing module is used to perform real-time parallel computation on the sampled data and extract impedance characteristic information at each frequency point; The MCU processing and communication module is used to complete complex impedance calculation, inference model operation, and wired or wireless data transmission. The host computer signal processing module is used to perform EEMD adaptive signal decomposition, dynamic baseline calibration, and high-frequency deep impedance modeling. The feature extraction and SSNN inference module is used to perform multi-parameter fusion feature extraction and end-to-end integrated inference of seven types of physiological parameters. The risk assessment and early warning module is used to perform multi-parameter joint anomaly detection and output three-level risk warning signals (high-risk, medium-risk, and low-risk) based on the classification threshold. The power management module is used to provide a stable power supply for the above modules and supports at least one external power supply method; A multi-electrode array is arranged on the surface of the object under test, and the excitation current amplitude is within the safe threshold range.

[0012] A multi-parameter non-invasive health monitoring system based on an FPGA-MCU heterogeneous architecture includes: The above-mentioned multi-parameter non-invasive health monitoring device based on FPGA-MCU heterogeneous architecture; Cloud-based health management platform: Supports real-time synchronization and sharing of data from multiple devices, enabling multi-level data sharing among medical institutions, communities, and families, long-term trend analysis, and early warning information push for abnormal events.

[0013] Beneficial effects Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extends the excitation frequency to 2MHz and utilizes the physical property that the capacitance effect of the stratum corneum of the skin tends to zero under high-frequency conditions. It combines low-squares linear fitting in the high-frequency band to establish an equivalent impedance model of deep human tissue, thus achieving systematic elimination of interference from the skin surface. The extracted equivalent pure resistance parameters of deep tissue are highly correlated with deep physiological indicators such as blood volume and tissue fluid content, which significantly breaks through the fundamental limitation of existing bioelectrical impedance technology that can only obtain surface impedance. At the same time, the deep human impedance spectrum can more effectively detect physiological parameters such as blood glucose and blood pressure.

[0014] This invention deploys the Goertzel algorithm in a fully parallel manner within an FPGA, fully leveraging the hardware advantages of the FPGA's high-speed parallel throughput to achieve real-time preprocessing of massive ADC sampling data. Simultaneously, the complex impedance calculation task involving intensive floating-point operations is offloaded to an MCU that supports C language, forming a heterogeneous processing architecture that combines hardware and software. This solves the problems of excessive resource consumption, insufficient real-time performance, and wasted computing power in traditional DFT / FFT schemes on embedded hardware, with the system's end-to-end processing latency strictly controlled within 0.1 seconds.

[0015] This invention introduces Integrated Empirical Mode Decomposition (EEMD) to adaptively decompose impedance time-series signals, which can effectively separate the fast-wave cardiac component and the slow-wave respiratory component at different time scales and suppress random noise. The extracted cardiac impedance waveform and respiratory waveform have high purity, providing a high-quality signal foundation for subsequent multi-parameter feature extraction and inference. Compared with traditional bandpass filtering methods, it has an essential advantage in non-stationary physiological signal processing.

[0016] This invention innovatively integrates bioelectrical impedance analysis (BIA) technology with SSNN to achieve end-to-end integrated reasoning of seven core physiological parameters: heart rate, respiratory rate, stroke volume, cardiac output, systolic blood pressure, diastolic blood pressure, and blood glucose. It fills the technological gap in non-invasive multi-parameter joint monitoring and represents a qualitative leap in monitoring dimensions and clinical application value compared to existing single-disease risk assessment or single-parameter blood glucose monitoring solutions.

[0017] This invention creatively adopts a Z-axis spatial isolation strategy, sinking the entire power management network to the bottom layer of the PCB and using the complete inner ground plane as an electromagnetic shielding layer. This physically cuts off the interference of high-frequency switching radiation on the analog front end, achieving the high signal-to-noise ratio required for medical-grade signal acquisition in a highly integrated portable device. It effectively solves the electromagnetic compatibility problem of the long-term coexistence of high-speed digital signals and weak analog signals in portable medical devices.

[0018] This invention designs a data caching and breakpoint resume mechanism based on a circular buffer at the MCU firmware layer, which can ensure zero loss of impedance data when the wireless link is interrupted momentarily. Combined with the low power consumption design of the entire link, the system can continuously support more than 24 hours of uninterrupted high-frequency multi-lead impedance monitoring in a completely offline portable state, taking into account both medical-grade data reliability and the need for non-intrusive daily wear. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of the multi-parameter non-invasive health monitoring method according to an embodiment of the present invention. Figure 2 This is a hardware system architecture block diagram according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the top and bottom layers of the PCB. Figure 4 This is a block diagram of the internal logic of the FPGA. Figure 5 The FPGA implementation flow of the Goertzel algorithm and the verification diagram of fixed-point error; Figure 6 A schematic diagram of the equivalent circuit model of deep human body impedance and the linear fitting curve in the high-frequency band. Figure 7 This is a schematic diagram of EEMD signal decomposition. Figure 8This is a schematic diagram illustrating the formation of multi-parameter fused feature vectors. Figure 9 A schematic diagram of the customized SSNN model architecture; Figure 10 Examples of the host computer client display interface and training interface; Figure 11 This is a flowchart for multi-level risk warning. Detailed Implementation

[0020] 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.

[0021] It should be noted that the use of terms such as "an embodiment," "embodiment," and "exemplary embodiment" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when describing a specific feature, structure, or characteristic in conjunction with embodiments, implementing such a feature, structure, or characteristic in conjunction with other embodiments should be within the knowledge of those skilled in the art.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] This embodiment provides a multi-parameter non-invasive health monitoring method based on FPGA-MCU heterogeneous architecture and bioimpedance spectroscopy, including the following steps: Step S1, Multi-frequency signal excitation and data acquisition: Under the scheduling of the control unit, a multi-frequency sinusoidal excitation signal is generated and injected into the human body through an electrode array. Then, the synchronous high-frequency digital sampling of the human body voltage and current signals is completed to obtain a high-speed parallel digital sampling sequence.

[0024] The specific process for this step is as follows: like Figure 2 and Figure 3 As shown, the hardware acquisition terminal of this invention uses a field-programmable gate array as the underlying real-time signal processing core, and is equipped with an ESP32 microprocessor as the global master control and wireless routing hub.

[0025] The ESP32 sends frequency control words to the high-speed direct digital frequency synthesizer via the SPI bus, and synthesizes multi-frequency sinusoidal excitation signals in the range of 1kHz to 2MHz according to the preset frequency sweep strategy. After the excitation signal is processed by the analog front end, it is injected into the human body through the electrode array. The amplitude of the excitation current injected into the human body is strictly limited to a safety threshold below the milliampere level to ensure medical-grade electrical safety.

[0026] The weak voltage and current response signals returned by the human body, containing physiological impedance information, are amplified and matched at the primary level by an ultra-low noise instrumentation amplifier. Then, they undergo deep purification via an active low-pass filter to ensure an extremely high signal-to-noise ratio for the analog signals input to subsequent stages. The purified signals are then fed into an AD9248 dual-channel high-speed ADC with a sampling rate of 65 MSPS and a resolution of 14 bits. This ADC performs synchronous high-frequency digital sampling of the voltage and current signals at a sampling rate far exceeding that of the Nyquist theorem, obtaining a high-speed parallel digital sampling sequence. This high-speed parallel digital sampling sequence is then directly fed into the FPGA.

[0027] At the PCB design level, the system adopts a four-layer high-density stack-up structure. The top layer is analog + digital, the inner layer 1 is a ground plane layer, the inner layer 2 is a power plane layer, and the bottom layer is a power management layer. The DC-DC buck-boost core and LDO power management network are all placed at the bottom layer. A solid electromagnetic shielding layer is constructed using the complete inner ground plane, i.e., the Z-axis spatial isolation strategy. This physically cuts off the contamination of the top-layer analog front-end signal by high-frequency switching radiation, ensuring the extremely high signal-to-noise ratio required for the acquisition of weak signals of deep biological impedance. Through the coordinated cooperation of the above hardware architecture, this embodiment achieves stable output and high-speed synchronous sampling of medical-grade multi-band excitation signals, providing a high-quality raw data foundation for subsequent impedance extraction and physiological parameter inference.

[0028] Step S2: FPGA Parallel Real-Time Impedance Calculation, Data Transmission, and Fusion Modeling: The digital sampling sequence is input into the FPGA processing unit, and the real and imaginary parts of the signal at each excitation frequency are extracted through parallel computation. The real and imaginary parts are serially transmitted to the MCU processing unit via the SPI protocol. The complex impedance at each frequency is obtained by precise floating-point calculation using C language. The calculated data is then encapsulated and transmitted to the host computer client via Wi-Fi. The host computer client first performs dynamic baseline calibration and then uses EEMD to adaptively decompose and reduce noise in the complex impedance data. Finally, the complex impedance data at each frequency is fused to construct an equivalent model of deep human tissue impedance.

[0029] In this step, the specific process of FPGA parallel real-time impedance calculation and data transmission is as follows: like Figure 4 and Figure 5As shown, this invention implements a parallel Goertzel algorithm for accurate extraction of impedance at multiple frequency points within an FPGA. Compared with traditional DFT / FFT schemes, the Goertzel algorithm only calculates amplitude and phase information for a single or a few specified frequency points, perfectly adapting to the application scenario of single-excitation frequency point detection in this invention, with extremely low resource overhead.

[0030] The Goertzel algorithm iteratively executes the following second-order difference equation (n from 0 to N-1): Where x[n] is the ADC sampled value, k is the target frequency point number, and N is the number of sampling points calculated each time (N=1024 in this embodiment).

[0031] After iteration, the real and imaginary parts of the signal at the target frequency are calculated using the Xcal module using the following formula, where, , , State variables output by the Goertzel iterator: The FPGA hardware implementation uses four DSP48A1 multipliers to construct a 35-bit multiplier, performing operations using signed fixed-point numbers in Q17.17 format (allowing a range of approximately [-131072, 131072]), which is much larger than the actual variation range of s[n] in 1024 iterations (the absolute values ​​are all within 1024), effectively avoiding precision loss. The 70-bit Q35.34 format data output from the multiplication is then processed by a 70-bit adder constructed using a LUT. MATLAB simulations of 500 iterations verify that the error rate of the Q17.17 fixed-point state variable s[n] compared to double-precision floating-point calculation does not exceed 0.1% throughout the entire iteration process, indicating that fixed-point calculation does not introduce significant errors.

[0032] In the top-level logic, the GoertzelIter and Xcal modules are dual-instantiated to process the voltage and current sampling channels in parallel, respectively, achieving fully parallel computation of dual-channel data. The output results are serially transmitted to the ESP32 MCU via the SPI protocol. After receiving the four channels of real and imaginary data, the ESP32 uses C language to perform precise floating-point calculation of the final complex impedance, and encapsulates the calculated spectrum data before transmitting it to the host computer client via Wi-Fi (TCP / IP protocol), with the end-to-end latency strictly controlled within 0.1 seconds. The precise floating-point calculation of the final complex impedance is as follows: Amplitude: Phase: To prevent data loss due to momentary interruption of the Wi-Fi link, the MCU's underlying firmware implements a data caching and breakpoint resumption mechanism based on a circular buffer: when the link is disconnected, impedance data is suspended and temporarily stored in the on-chip SRAM, and timing retransmission is automatically performed after the link handshake is restored, fundamentally ensuring the absolute continuity and integrity of pathological data.

[0033] In this step, the specific process of fusion modeling is as follows: like Figure 6 As shown, this invention comprehensively considers and develops a method for modeling the equivalent impedance of deep human tissue based on linear fitting of high-frequency impedance spectra. Based on a linear equivalent circuit model of skin impedance, the model uses a linear resistor R_s to simulate the impedance of deep skin tissue, and a parallel combination of a linear capacitor C_p and a resistor R_p to simulate the impedance of the stratum corneum. Its complex impedance expression is as follows: As the frequency increases, the capacitive effect of the skin surface gradually weakens, and the impedance tends to be in a pure resistance state. Therefore, fitting high-frequency data can effectively remove the interference of the stratum corneum and extract the equivalent pure resistance parameters of deep tissues.

[0034] For the high-frequency impedance data of each channel, the following linear fit was performed using the least squares method: The formula for finding the optimal parameters is: Where N is the total number of high-frequency data points across multiple frequency bands. This is the measured value of the imaginary part of the impedance at the i-th frequency point. Let be the measured real part of the impedance at the i-th frequency point. The high-frequency extrapolated intercept obtained from the fitting is... The equivalent pure resistance, which approximates the deep tissues of the human body, is closely related to physiological states such as pleural blood volume and deep tissue fluid conductivity; the fitting slope k reflects the impedance dispersion characteristics and can be used to assess the magnitude of the capacitance effect at the skin-tissue interface.

[0035] like Figure 7 As shown, after completing the multi-band impedance data acquisition, the host computer client first performs dynamic baseline calibration: extracting the low-frequency components of the impedance signals of each band to estimate the slowly changing baseline, and subtracting them from the original signal to eliminate baseline drift caused by changes in electrode contact and fluctuations in skin conductance. Subsequently, Integrated Empirical Mode Decomposition (EEMD) is introduced to adaptively decompose the impedance time-series signal; EEMD effectively overcomes the mode aliasing problem of EMD by repeatedly adding auxiliary white noise to the signal and performing EMD decomposition and then taking the mean, adaptively separating components at different time scales, as follows: (1) The random noise components corresponding to the first few high-frequency IMFs are discarded; (2) The intermediate IMF represents the rapidly changing components caused by the heartbeat (frequency approximately 0.8 Hz to 3 Hz), which is retained as the noise-reduced heartbeat impedance waveform; (3) The last low-frequency IMF represents the slowly changing component caused by respiration (frequency approximately 0.1 Hz to 0.5 Hz), used to calculate respiratory rate.

[0036] Through the modeling and decomposition process of the above-mentioned deep human tissue impedance equivalent model, step S2 achieves accurate extraction of deep tissue impedance and effective separation of heartbeat and respiratory components, providing a high-quality signal foundation for subsequent multi-parameter feature extraction.

[0037] Step S3, Multi-physiological parameter fusion prediction: Extract multi-dimensional fusion features from the denoised deep human tissue impedance equivalent model to form a 42-dimensional multi-parameter fusion feature vector, input it into a customized state space neural network, and output multiple core physiological parameters such as heart rate, respiratory rate, blood pressure, and blood glucose.

[0038] The multidimensional fusion feature extraction process is as follows: like Figure 8 As shown, this invention extracts multi-dimensional fusion features from the denoised impedance waveform and the deep impedance model to form a 42-dimensional multi-parameter fusion feature vector, which mainly includes the following three categories: Basic characteristics: Frequency domain characteristics include impedance spectrum slope k and high-frequency extrapolation intercept. It reflects changes in the content of deep tissue fluid; the time-domain characteristics include the peak amplitude of the heartbeat, the slope of the rising edge, and the heartbeat cycle, which are related to stroke volume (SV) and myocardial contractility.

[0039] Blood pressure-related characteristics: impedance wave propagation velocity The impedance wave propagation time from the anterior-posterior leads is positively correlated with vascular elasticity; the diastolic impedance change rate... The slope of the rising edge of impedance during diastole reflects peripheral vascular resistance; impedance amplitude fluctuation coefficient. The ratio of the peak-to-trough difference in impedance during a cardiac cycle to the baseline impedance is related to the blood flow pulse pressure.

[0040] Blood glucose-related characteristics: Mean of the real part of mid-frequency impedance The average value of the real part of the impedance within a sliding window in the frequency band from 10kHz to 100kHz is negatively correlated with blood conductivity; the impedance temperature coefficient The effect of glucose metabolism and heat production on impedance was corrected by the rate of change of impedance signal with ambient temperature; multi-lead impedance difference. The impedance difference between the left chest and right chest and between the front chest and back is taken to reflect the uniformity of blood distribution throughout the body.

[0041] By fusing the above three types of features in multiple dimensions, this embodiment systematically quantifies deep physiological information from multiple dimensions, including frequency domain, time domain, and spatial leads, providing high-dimensional and highly discriminative feature inputs for subsequent SSNN multi-parameter inference.

[0042] Furthermore, the multi-parameter inference process of a custom state-space neural network (SSNN) is as follows: like Figure 9 As shown, this invention designs a custom state-space neural network (SSNN), drawing on the core idea of ​​structured state-space models (multi-parameter fusion features), and adopts an integrated design of "three-level core architecture and multi-parameter prediction branches". The overall network structure satisfies: The three-level core architecture of a custom state-space neural network includes the following three modules: (1) Spectrum-Time Domain Fusion Embedding Module (STE-Conv1D): integrates frequency domain features (k, The vector is mapped to a low-dimensional embedding vector through a fully connected layer and then concatenated and fused with the temporal features to achieve complementary features between the two domains. (2) Bidirectional state space modeling unit (B-SSMU): It adopts continuous convolution kernels to construct state transfer and gating mechanism, and simultaneously captures the rapid fluctuation of heartbeat and the long-term trend of hemodynamics in impedance signal, respectively treating them as short-term and long-term dependencies, adapting to the multi-scale characteristics of physiological signals, and effectively solving the problems of low parallel efficiency and unstable gradient propagation of traditional RNN. (3) Sparse abnormal event enhancement layer (SEE-FFN): Introducing L1 regularization constraints, the contribution of abnormal signals such as weak heartbeat, sudden change in blood pressure, and abnormal blood sugar to the output is enhanced by weight sparsification, thereby improving the model’s sensitivity to hidden functional abnormalities.

[0043] The multi-parameter prediction branch consists of the following four components: (1) Multi-parameter adaptation layer: The feature weights are dynamically allocated through the attention mechanism. For blood pressure prediction, the relevant features such as impedance wave propagation velocity are enhanced, and for blood glucose prediction, the core features such as the real part of mid-frequency impedance are enhanced. (2) Blood Pressure Prediction Branch (SBP / DBP): Using feature refinement and piecewise regression strategies, systolic blood pressure (SBP) and diastolic blood pressure (DBP) are fitted separately using a 2-layer MLP, and a vascular elasticity calibration coefficient is introduced. The formula for correcting the effects of age-related arteriosclerosis is as follows: Where a, b, c, d, e, and f are regression coefficients obtained from model training. The MAE of SBP is controlled within ±5 mmHg, and the MAE of DBP is controlled within ±3 mmHg. (3) Blood glucose prediction branch (FPG / 2hPG): Integrates a time period recognition module, automatically distinguishes between fasting and postprandial states based on the diurnal rhythm of impedance signals and postprandial impedance fluctuation patterns, constructs a personalized mapping model by combining individual vital signs parameters, and introduces a blood glucose calibration coefficient. To correct for individual metabolic differences; (4) Joint optimization layer: Unify the inference loss of all parameters to ensure the balance between mutual constraints and accuracy during multi-parameter synchronous inference.

[0044] The anomaly detection module constructs a clinical anomaly threshold library covering all parameters. It integrates the inference results of all parameters through a weighted fusion algorithm, and outputs the single-parameter anomaly probability and the multi-parameter joint anomaly risk value, providing a quantitative basis for subsequent risk assessment and graded early warning. The clinical anomaly threshold library includes multiple aspects such as cardiovascular parameter abnormalities (e.g., sudden drop in SV); blood pressure abnormalities (SBP≥140mmHg or≤90mmHg, DBP≥90mmHg or≤60mmHg); blood glucose abnormalities (FPG≥7.0mmol / L or≤3.9mmol / L, 2hPG≥11.1mmol / L).

[0045] Step S4, Risk Warning and Health Assessment: Input multiple core physiological parameters into the pre-trained inference model and output real-time health status assessment and multi-level risk warning signals.

[0046] The specific optimization of the inference model training and deployment in this step is as follows: The training data was based on clinical data provided by the PLA Navy General Hospital, covering healthy volunteers, patients with abnormal cardiac function, patients with hypertension, and patients with type 2 diabetes. BIA signals were correlated one-to-one with measurements from standard equipment. SBP / DBP were measured using an electronic blood pressure monitor, and FPG / 2hPG were measured using venous blood biochemistry. The data covered individuals aged 18 to 75 years with a body mass index (BMI) of 18.5 to 35 kg / m². 2 Different groups of people are considered to ensure that the model has good generalization ability.

[0047] The loss function employs a multi-parameter weighted loss sum, specifically including: MSE loss for cardiovascular parameters (HR, SV, etc.) (weight 0.3), MSE loss for SBP / DBP (weight 0.3), MSE loss for FPG / 2hPG (weight 0.3), and cross-entropy loss for anomaly detection (weight 0.1). Based on the PyTorch framework, early stopping regularization and cross-validation are used to suppress overfitting, and the Adam optimizer dynamically adjusts the learning rate to accelerate convergence. The deployment mode continues the "lightweight front-end and high-performance back-end computing" architecture: FPGA and ESP32 work together to complete data acquisition and transmission, and model inference is deployed entirely on the host computer client, avoiding the limitation of hardware computing power on model accuracy.

[0048] The host computer client and visualization system of the present invention, such as Figure 10 and Figure 11 As shown, the host computer client is developed using the Python PyQt framework, providing a multi-tab graphical user interface. Its main functional modules are as follows: Data receiving and parsing module: Real-time reception of multi-channel, multi-frequency raw impedance data transmitted by the MCU, dynamic rendering of impedance waveforms, support for reliable TCP transmission, and the ability to switch to UDP mode to reduce transmission latency.

[0049] Real-time inference and results display module: Users can manually or automatically input impedance data and auxiliary vital signs parameters (body temperature, BMI, etc.). After the system calls the inference engine at the underlying level, it outputs the predicted values ​​of seven types of physiological parameters and their corresponding risk levels in a highly readable format in real time.

[0050] Model training monitoring module: Supports users to import local CSV format data files, prevents interface lag through multi-threading mechanism, prints training logs, validation set loss and early stopping status for each round in real time, and automatically plots and saves learning curves.

[0051] Multi-level risk warning module: Constructs a three-level dynamic warning mechanism of high-risk, medium-risk, and low-risk, dynamically displaying risk heat maps and historical trend comparison charts; high-risk alarms automatically send alarm notifications to emergency contacts and medical service providers through the cloud platform, with a warning trigger delay of less than 100 milliseconds, fully ensuring timely response to sudden cardiovascular events. Through the coordinated operation of the above modules, the host computer client realizes closed-loop management of the entire process from receiving raw impedance data and real-time inference to visualized warnings.

[0052] This invention also provides a multi-parameter non-invasive health monitoring device based on an FPGA-MCU heterogeneous architecture, used to implement the multi-parameter non-invasive health monitoring method based on the FPGA-MCU heterogeneous architecture, comprising: The hardware acquisition module is used to synthesize multi-frequency excitation signals and inject them into the human body through an electrode array, and to condition and digitally sample the response signals. The FPGA digital processing module is used to perform real-time parallel computation on the sampled data and extract impedance characteristic information at each frequency point; The MCU processing and communication module is used to complete complex impedance calculation, inference model operation, and wired or wireless data transmission. The host computer signal processing module is used to perform EEMD adaptive signal decomposition, dynamic baseline calibration, and high-frequency deep impedance modeling. The feature extraction and SSNN inference module is used to perform multi-parameter fusion feature extraction and end-to-end integrated inference of seven types of physiological parameters. The risk assessment and early warning module is used to perform multi-parameter joint anomaly detection and output three-level risk warning signals (high-risk, medium-risk, and low-risk) based on the classification threshold. The power management module is used to provide a stable power supply for the above modules and supports at least one external power supply method, such as power supply using Z-axis isolation design; A multi-electrode array is arranged on the surface of the object under test, and the excitation current amplitude is within the safe threshold range.

[0053] The present invention also provides a multi-parameter non-invasive health monitoring system, including a memory and a processor. The memory stores a computer program executed by the processor, which, when run by the processor, performs a multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture.

[0054] The processor includes: The above-mentioned multi-parameter non-invasive health monitoring device based on FPGA-MCU heterogeneous architecture; Cloud-based health management platform: Supports real-time synchronization and sharing of data from multiple devices, enabling multi-level data sharing among medical institutions, communities, and families, long-term trend analysis, and early warning information push for abnormal events; Storage medium: stores a computer program that can execute the above steps, and implements the multi-parameter non-invasive health monitoring method based on FPGA-MCU heterogeneous architecture of the present invention when run by a processor.

[0055] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture, characterized in that, Includes the following steps: S1. Multi-frequency signal excitation and data acquisition: Under the scheduling of the control unit, multi-frequency sinusoidal excitation signals are generated and injected into the human body through the electrode array. Then, the synchronous high-frequency digital sampling of the human body voltage and current signals is completed to obtain a high-speed parallel digital sampling sequence. Step S2, FPGA Parallel Real-Time Impedance Calculation, Data Transmission and Fusion Modeling: The digital sampling sequence is input into the FPGA processing unit, and the real and imaginary parts of the signal at each excitation frequency are extracted through parallel calculation; The real and imaginary data are serially transmitted to the MCU processing unit via the SPI protocol to perform accurate floating-point calculation of the complex impedance to obtain the complex impedance data at each frequency point. The calculated complex impedance data is encapsulated and transmitted to the host computer client via Wi-Fi. The host computer client uses EEMD to adaptively decompose and reduce noise in the complex impedance data. It also integrates the complex impedance data at each frequency point to construct an equivalent model of deep human tissue impedance. Step S3, Multi-physiological parameter fusion prediction: Extract multi-dimensional fusion features from the denoised deep human tissue impedance equivalent model to form a 42-dimensional multi-parameter fusion feature vector, input it into a customized state space neural network, and output multiple core physiological parameters such as heart rate, respiratory rate, blood pressure, and blood glucose. Step S4, Risk Warning and Health Assessment: Input multiple core physiological parameters into the pre-trained inference model and output real-time health status assessment and multi-level risk warning signals.

2. The multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture according to claim 1, characterized in that, In step S1, a multi-frequency sinusoidal excitation signal in the range of 1kHz to 2MHz is generated. After the excitation signal is processed by the analog front end, it is injected into the human body through the electrode array. The amplitude of the excitation current injected into the human body is strictly limited to a safety threshold below the milliampere level.

3. The multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture according to claim 1, characterized in that, In step S1, a dual-channel high-speed ADC with a sampling rate of 65 MSPS and a resolution of 14 bits is used to complete the synchronous high-frequency digital sampling of voltage and current signals at a sampling rate far exceeding that of the Nyquist theorem, thereby obtaining a high-speed parallel digital sampling sequence.

4. The multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture according to claim 1, characterized in that, In step S2, parallel computation employs an iterative algorithm based on recursive equations, and in step S2, C language is used to perform accurate floating-point calculation of the final complex impedance.

5. The multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture according to claim 1, characterized in that, In step S2, the host computer client first performs dynamic baseline calibration on the calculated complex impedance data, and then uses EEMD to adaptively decompose and reduce noise in the complex impedance data.

6. A multi-parameter non-invasive health monitoring device based on an FPGA-MCU heterogeneous architecture, used to implement the multi-parameter non-invasive health monitoring method based on an FPGA-MCU heterogeneous architecture as described in any one of claims 1 to 5, characterized in that, include: The hardware acquisition module is used to synthesize multi-frequency excitation signals and inject them into the human body through an electrode array, and to condition and digitally sample the response signals. The FPGA digital processing module is used to perform real-time parallel computation on the sampled data and extract impedance characteristic information at each frequency point; The MCU processing and communication module is used to complete complex impedance calculation, inference model operation, and wired or wireless data transmission. The host computer signal processing module is used to perform EEMD adaptive signal decomposition, dynamic baseline calibration, and high-frequency deep impedance modeling. The feature extraction and SSNN inference module is used to perform multi-parameter fusion feature extraction and end-to-end integrated inference of seven types of physiological parameters. The risk assessment and early warning module is used to perform multi-parameter joint anomaly detection and output three-level risk warning signals (high-risk, medium-risk, and low-risk) based on the classification threshold. The power management module is used to provide a stable power supply for the above modules and supports at least one external power supply method; A multi-electrode array is arranged on the surface of the object under test, and the excitation current amplitude is within the safe threshold range.

7. A multi-parameter non-invasive health monitoring system based on an FPGA-MCU heterogeneous architecture, characterized in that, include: The multi-parameter non-invasive health monitoring device based on an FPGA-MCU heterogeneous architecture as described in claim 6; Cloud-based health management platform: Supports real-time synchronization and sharing of data from multiple devices, enabling multi-level data sharing among medical institutions, communities, and families, long-term trend analysis, and early warning information push for abnormal events.

8. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the multi-parameter non-invasive health monitoring method based on FPGA-MCU heterogeneous architecture and bioimpedance spectrum as described in any one of claims 1 to 5.

Images