Dynamic blood glucose monitoring and intelligent pump speed adjusting system and method

Abstract

The invention discloses a dynamic blood glucose monitoring and intelligent pump speed adjusting system and method, and relates to the technical field of biomedical engineering.The dynamic blood glucose monitoring and intelligent pump speed adjusting method comprises the steps that precise monitoring is conducted through a multi-mode sensing module, an enzyme electrode sensor adopts an anti-interference coating, a Raman spectrum dynamic correction algorithm is combined, and motion artifacts and temperature drift are restrained; the metabolic feature fusion calibration module is integrated with a time convolution network, time sequence modeling is carried out on diet, exercise and historical blood glucose data, an individualized metabolic parameter database is constructed, and self-adaptive signal drift compensation is achieved; the self-adaptive pump control module innovatively designs a double-layer fuzzy PID controller, a base layer generates an infusion rate according to a real-time blood glucose state and insulin activity, an optimization layer introduces an LSTM metabolism prediction model, and the slope of an infusion curve is dynamically adjusted in combination with future behavior pre-judgment; in the edge-cloud collaborative architecture, a lightweight TCN model is locally deployed to guarantee real-time control, and a cloud end optimizes a global digital twinborn model through federal learning and generates a multi-dimensional metabolic risk early warning report.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical engineering, and in particular to a dynamic blood glucose monitoring and intelligent pump speed regulation system and method. Background Art

[0002] Globally, over 500 million people suffer from diabetes. Traditional blood glucose monitoring, such as fingertip blood tests, is invasive, time-consuming, and cumbersome. Dynamic monitoring can improve treatment accuracy. Insulin pump therapy requires real-time adjustment of the infusion rate based on blood glucose fluctuations. Traditional fixed-rate methods are prone to hypoglycemia or hyperglycemia. Minimally invasive glucose sensors offer improved accuracy, enabling continuous blood glucose monitoring.

[0003] A deep learning-based blood glucose prediction model can provide early warning of hypoglycemia and hyperglycemia trends, and combined with reinforcement learning, it enables personalized pump speed adjustment. By simulating the human pancreas' "perception-decision-execution" mechanism, a fully automated closed-loop system of "blood glucose monitoring-algorithm decision-making-pump speed adjustment" is constructed. An aging society is driving the development of home-based and intelligent chronic disease management. Dynamic monitoring and intelligent adjustment systems can reduce the frequency of medical visits for patients and improve their quality of life.

[0004] However, existing technologies still have problems that need to be solved: signal drift and calibration defects. Enzyme electrode sensors are easily affected by temperature and motion artifacts, and the drift error can reach 18% after 72 hours. The existing calibration algorithm relies on finger blood input at fixed time points and does not integrate the user's diet / metabolic characteristics. Summary of the Invention

[0005] In order to solve the above technical problems, a dynamic blood glucose monitoring and intelligent pump speed adjustment system and method are provided. This technical solution solves the above problems of signal drift and calibration defects.

[0006] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0007] A dynamic blood glucose monitoring and intelligent pump speed adjustment system, comprising:

[0008] Multimodal sensing module: Integrates an enzyme electrode sensor, a temperature compensation sensor, and a motion acceleration sensor. The enzyme electrode sensor is coated with a nanoscale anti-interference layer for simultaneous detection of subcutaneous interstitial fluid glucose concentration, ambient temperature, and user motion status. The non-invasive Raman spectroscopy detection unit compensates for enzyme electrode signal drift through a dynamic skin transmittance correction algorithm.

[0009] Metabolic feature fusion calibration module: User metabolic parameter database, storing personalized glycated hemoglobin, insulin sensitivity factor, and carbohydrate ratio data; dynamic calibration algorithm, based on time convolution network, performs time series modeling on the enzyme electrode raw signal, and input parameters include real-time diet records, exercise energy consumption, and historical blood sugar fluctuation patterns;

[0010] Adaptive pump control module: A two-layer fuzzy PID controller. The first layer calculates the basal pump rate based on the current blood glucose level, blood glucose change rate, and insulin activity in the body. The second layer introduces a metabolic prediction model to dynamically adjust the slope of the insulin infusion curve.

[0011] Edge-cloud collaborative architecture: The local edge computing unit deploys lightweight AI modules to perform real-time signal processing and control decisions; the cloud-based digital twin platform aggregates multi-user data through federated learning to optimize the global model and generate personalized metabolic risk warning reports.

[0012] Preferably, the multimodal sensing module specifically includes:

[0013] Anti-drying unit: The nanoscale anti-interference layer consists of a polyurethane substrate doped with the conductive polymer PEDOT:PSS and uniformly dispersed with cerium oxide nanoparticles. A porous network structure is formed by electrochemical deposition, and the catalytic activity of CeO2 accelerates the decomposition of H2O2, inhibiting baseline drift caused by the accumulation of enzyme reaction byproducts. A zwitterionic polymer is grafted onto the surface. Dynamic Raman spectroscopy correction is performed using a low-power laser to penetrate the skin surface and collect Raman spectra of subcutaneous tissue. A mapping relationship between glucose concentration and spectral signals is established through the characteristic peak intensity ratio.

[0014] Cross-modal data fusion unit: inputs temperature sensor data, accelerometer data and enzyme electrode signals into the extended Kalman filter to construct the state equation;

[0015] Calibration trigger unit: Triggered by a dietary event, the camera captures the plate image, uses the ResNet-18 model to identify the food type, and combines the CR value to calculate the expected blood sugar fluctuation threshold. Calibration is initiated when the enzyme electrode reading deviates from the expected value by more than a predetermined ratio; triggered by exercise status, the accelerometer automatically shortens the Raman spectrum detection interval after detecting medium to high intensity exercise; abnormal drift judgment: if the Raman spectrum correction coefficient is greater than the predetermined threshold for multiple consecutive times, the user is forced to perform a finger blood calibration.

[0016] Preferably, the multimodal sensing module specifically includes:

[0017] Miniaturized packaging unit: A three-dimensional stacked structure with a flexible polyimide substrate, an enzyme electrode array integrated on the upper layer, Raman lasers and photodetectors embedded in the middle layer, and temperature / acceleration sensors arranged on the bottom layer. Electrical connections between components are achieved through gold wire bonding.

[0018] Motion artifact suppression unit: Acceleration signal classification, using support vector machines to distinguish movement types, and dynamically adjusting filtering parameters for different activity levels; high-frequency noise filtering, designing a fourth-order Butterworth low-pass filter; skin contact uses a bionic microneedle array to penetrate the stratum corneum and directly contact the interstitial fluid of the dermis; the needle surface is coated with a hyaluronic acid sustained-release layer to maintain the hydration state of the detection interface.

[0019] Preferably, the metabolic signature fusion calibration module specifically includes:

[0020] Personalized metabolic parameter database: including glycated hemoglobin, insulin sensitivity factor, and carbohydrate ratio; metabolic feature vector generation, extracting features based on historical data, including blood glucose fluctuation patterns, and identifying user-specific patterns through K-means clustering; insulin activity decay curve, establishing a bi-exponential model; metabolic rate calibration, estimating basal metabolic rate by combining heart rate variability and respiratory entropy;

[0021] Temporal convolutional network time series modeling: The input layer includes raw enzyme electrode signals, dietary records, exercise energy consumption, and historical blood sugar fluctuations. The network structure is an eight-layer causal dilated convolution, and the output of each layer is processed by a gated activation unit. Skip connections fuse shallow features, and the output layer is a bidirectional LSTM to predict the future short-term blood sugar drift compensation coefficient.

[0022] Multimodal data fusion unit: Priority weight allocation, Raman spectroscopy data is preferred within the golden standard period after meals, and enzyme electrode signals are relied upon during sleep; conflict resolution mechanism, when the difference in data from different sensors is greater than a predetermined ratio, metabolic model arbitration is initiated; regular evaluation of calibration effects, online optimization of TCN network parameters through gradient descent method; introduction of adversarial training to generate extreme metabolic scenario data.

[0023] Preferably, the metabolic signature fusion calibration module specifically includes:

[0024] Event-driven calibration trigger unit: Metabolic event detection. When postprandial hyperglycemia does not reach the expected level, carbohydrate intake is greater than the preset value through image recognition, activating TCN short-term prediction calibration. When abnormal hypoglycemia after exercise is detected by the accelerometer and exceeds the set threshold, the calibration interval is shortened. When the hourly change in the insulin sensitivity mutation ISF value exceeds the set ratio, the fingerstick blood verification calibration is forced to start.

[0025] Cross-modal calibration execution process: The Raman spectroscopy unit collects baseline values ​​and corrects the subcutaneous glucose concentration using a skin transmittance compensation algorithm; the enzyme electrode signal is preprocessed by temperature compensation and motion artifact filtering; the TCN network outputs a drift compensation coefficient, which is weighted and fused with the metabolic model prediction value; if the error after calibration is still greater than the established ratio, a hardware-level reset is triggered, the sensor power supply circuit is restarted, and the enzyme electrode working potential is reinitialized.

[0026] Preferably, the adaptive pump control module specifically includes:

[0027] The first layer: Calculate the basic pump rate. The input parameters include the current blood glucose level, blood glucose change rate, and insulin activity in vivo. Fuzzy subsets are divided using trapezoidal function, Gaussian function, and trigonometric function, and a fuzzy rule base is established. The center of gravity method is used to calculate the basic infusion rate. The maximum infusion volume is set according to the IOB value.

[0028] The second layer: Dynamic parameter adjustment driven by metabolic prediction, using the LSTM metabolic prediction model. The input features include time series data and event data. The network structure is a bidirectional LSTM layer and a self-attention mechanism, which outputs a blood glucose prediction curve for the short term in the future. The model deviation is corrected in real time through the Kalman filter. The key features of the prediction curve are extracted, including peak blood glucose, time to peak, and area under the curve, and fuzzy variables are constructed for dynamic adjustment.

[0029] Preferably, the adaptive pump control module specifically includes:

[0030] The MEMS piezoelectric pump uses a 3D-printed titanium alloy pump chamber and piezoelectric ceramic actuator; an integrated high-precision flow sensor; the inner wall of the drug reservoir is coated with a heparin-like coating; a pulse reverse cleaning protocol is set to automatically execute reverse pressure pulses at regular intervals; a real-time control loop, and timing synchronization includes triggering the ADC to read sensor data acquisition through hardware interrupts, implementing fuzzy PID calculations through dedicated FPGA logic circuits, and directly controlling the pump body execution through the piezoelectric driver PWM pulse.

[0031] Preferably, the edge-cloud collaborative architecture specifically includes:

[0032] Lightweight AI model compression: Through knowledge distillation, the cloud-trained LSTM metabolic prediction model is compressed into a TCN network. Eight-bit fixed-point quantization uses the TensorFlow Lite Micro framework, with dynamic pruning to dynamically shut down redundant neurons based on real-time data streams. Real-time inference integrates edge TPUs, supports parallel processing of multimodal sensor data, and uses a circular buffer for memory management. Local control decisions and priority task scheduling utilize an offline emergency mechanism to store core control logic and maintain short-term basic operation during network outages. An LRU cache strategy is used to retain the most recent day's metabolic signature data for local model invocation.

[0033] Cloud-based digital twin platform: Federated learning for global optimization. Through a secure aggregation protocol and Paillier homomorphic encryption, edge nodes upload model gradients, which are aggregated in the cloud to generate global model updates. Differential privacy noise is introduced. Each user has their own personalized model branch, and the cloud quickly adapts to new users through meta-learning. Branches share the underlying feature extraction layer, but have exclusive access to the top-level decision-making parameters.

[0034] Digital twin metabolic modeling, multi-scale physiological simulation, organ-level modeling, and simulation of pancreatic insulin secretion dynamics based on the finite element method; molecular-level modeling, simulating the insulin receptor binding process through molecular dynamics; risk warning logic, short-term warning, combining real-time blood sugar trends and weather data; long-term warning, predicting the risk of diabetes complications through digital twin deduction;

[0035] Report generation uses GPT-4Turbo to generate personalized recommendation text; embedded explainable AI displays key decision-making basis; three-dimensional metabolic map, WebGL renders the user's whole-body glucose distribution heat map, and marks insulin-sensitive areas; VR mode supports doctors to immersively observe historical metabolic event chains.

[0036] Preferably, the edge-cloud collaborative architecture specifically includes:

[0037] Edge-cloud collaboration: The hierarchical transmission strategy includes abnormal alarms as the highest level, with instant transmission and no compression; metabolic feature vectors as the high level, with hourly transmission and LZ4 compression; raw sensor data as the low level, with daily transmission and Huffman encoding; the protocol stack design includes CoAP over DTLS as the transport layer, which supports reliable communication for low-power devices; the application layer is a custom binary protocol; model hot updates, incremental push, the cloud generates model difference packages every day, and the edge merges and updates them through the BSDiff algorithm; update verification, using EdDSA digital signature verification; A / B test diversion, a predetermined proportion of users as the gray release group, first testing the new control strategy; feedback data is collected through edge embedding points, and the full data is pushed after meeting the standards.

[0038] Furthermore, a method for dynamic blood glucose monitoring and intelligent pump speed adjustment is provided for implementing the dynamic blood glucose monitoring and intelligent pump speed adjustment system described above, comprising:

[0039] An enzyme electrode sensor coated with a nanoscale anti-interference layer continuously acquires subcutaneous interstitial fluid glucose concentration signals. It also simultaneously collects ambient temperature, three-dimensional motion acceleration, and Raman spectroscopy data from the skin surface. This data is then cross-modally fused using an extended Kalman filter to suppress motion artifacts and temperature drift interference.

[0040] A time-series convolutional network calibration model is constructed based on the user's individual metabolic parameters. When dietary events or moderate to high-intensity exercise are detected, a Raman spectroscopy dynamic compensation algorithm is triggered, correcting skin transmittance deviations using a Monte Carlo light transmission model. Only when the error exceeds a certain percentage after a predetermined number of consecutive Raman compensations is the Bluetooth-connected portable blood glucose meter activated for mandatory fingerstick calibration.

[0041] Two-layer fuzzy PID pump speed regulation: The first layer generates a basal infusion rate based on the current blood glucose level, blood glucose change rate, and insulin activity in the body through a fuzzy rule base. The second layer introduces an LSTM metabolic prediction model, inputting the expected short-term dietary calories and exercise plan to dynamically adjust the slope of the infusion curve.

[0042] The local edge computing unit deploys a lightweight TCN model to perform real-time signal processing and control instruction generation; the cloud aggregates multi-user data through federated learning, updates the global digital twin model and generates personalized metabolic risk warning reports.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] This invention utilizes the dynamic synergy of a nano-anti-interference coating and Raman spectroscopy to construct an optical-electrochemical dual-mode calibration system, significantly improving signal stability. When dietary or exercise events are detected, a Monte Carlo optical transmission model corrects for deviations in the optical properties of subcutaneous tissue in real time, achieving non-invasive and precise compensation. This overcomes the limitations of traditional systems that rely on frequent fingerstick blood calibration. This adaptive compensation logic minimizes the need for manual intervention and demonstrates enhanced environmental robustness in complex life scenarios.

[0045] By organically integrating lightweight model compression, a hierarchical data transmission strategy, and a federated learning framework, a hybrid intelligent system with both real-time and evolutionary capabilities is constructed. A miniaturized TCN model deployed at the edge ensures millisecond-level response to critical control commands, while a cloud-based digital twin platform continuously optimizes metabolic models through multi-scale physiological simulation. Both platforms leverage privacy-preserving computing technologies to ensure the secure transfer of data value. This architecture avoids the control latency risks associated with complete reliance on the cloud while overcoming the limited generalization capabilities of a single local model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is an internal framework diagram of a dynamic blood glucose monitoring and intelligent pump speed adjustment system;

[0047] Figure 2 The figure is a flow chart of a method for dynamic blood glucose monitoring and intelligent pump speed adjustment. DETAILED DESCRIPTION

[0048] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0049] Reference Figure 1 As shown, a dynamic blood glucose monitoring and intelligent pump speed adjustment system includes:

[0050] Multimodal sensing module: Integrates an enzyme electrode sensor, a temperature compensation sensor, and a motion acceleration sensor. The enzyme electrode sensor is coated with a nanoscale anti-interference layer for simultaneous detection of subcutaneous interstitial fluid glucose concentration, ambient temperature, and user motion status. The non-invasive Raman spectroscopy detection unit compensates for enzyme electrode signal drift through a dynamic skin transmittance correction algorithm.

[0051] Metabolic feature fusion calibration module: User metabolic parameter database, storing personalized glycated hemoglobin, insulin sensitivity factor, and carbohydrate ratio data; dynamic calibration algorithm, based on time convolution network, performs time series modeling on the enzyme electrode raw signal, and input parameters include real-time diet records, exercise energy consumption, and historical blood sugar fluctuation patterns;

[0052] Adaptive pump control module: A two-layer fuzzy PID controller. The first layer calculates the basal pump rate based on the current blood glucose level, blood glucose change rate, and insulin activity in the body. The second layer introduces a metabolic prediction model to dynamically adjust the slope of the insulin infusion curve.

[0053] Edge-cloud collaborative architecture: The local edge computing unit deploys lightweight AI modules to perform real-time signal processing and control decisions; the cloud-based digital twin platform aggregates multi-user data through federated learning to optimize the global model and generate personalized metabolic risk warning reports.

[0054] It should be noted that the anti-interference sensing matrix, nano-coating technology, and polyurethane-PEDET:PSS / CeO2 composite coating (thickness <50nm) on the surface of the enzyme electrode can significantly improve signal stability against lactic acid interference caused by sweating during exercise (lactate dehydrogenase adsorption rate is reduced by 87%) and drug interference (such as acetaminophen cross-reaction is reduced to 0.5%).

[0055] Multi-source data verification, Raman spectrum (785nm laser) and enzyme electrode signals are collected synchronously, and the skin transmittance compensation coefficient (ΔT) is calculated in real time through the Monte Carlo light transmission model. When the deviation between the two is greater than 15%, the third-level calibration protocol (including Bluetooth finger blood verification) is triggered.

[0056] Dynamic temperature compensation, integrating MEMS thermopile and polynomial fitting algorithm (order = 5), eliminates the nonlinear effect of ambient temperature fluctuation (-10℃~45℃) on enzyme activity (compensation efficiency > 92%).

[0057] An adaptive modeling framework for metabolic characteristics, dynamic updating of individualized parameters, and incremental updates of HbA1c values ​​every 30 days are based on the time-weighted average of blood glucose (TIR) ​​and the red blood cell lifespan model; ISF parameters are optimized hourly, and the reinforcement learning model input includes the standard deviation of blood glucose fluctuations and the post-exercise insulin sensitivity enhancement factor β1 = 1.12.

[0058] Innovative time series modeling: the TCN network uses dilated causal convolution (dilation coefficients 1, 2, 4, ..., 128) to capture long-range dependencies.

[0059] Collaborative decision logic of two-layer PID:

[0060] The basic control layer outputs the basic pump speed in real time according to the fuzzy rule base (147 clinical rules), supports micro-dose adjustment of 0.025U / h, and limits the maximum rate to 80% of the daytime mode.

[0061] In the prediction optimization layer, the LSTM model (128 hidden units) inputs the next 30-minute diet plan (carbohydrate accuracy of image recognition ±5g) and exercise schedule (data synchronized with Apple HealthKit) and dynamically adjusts the slope of the insulin infusion curve (regulation factor α∈[0.5,2.0]).

[0062] Emergency response strategy for metabolic mutations, dynamic priority scheduling:

[0063] Event Level Trigger Conditions Response Action Emergency (L1) BG<3.9mmol / L and ROC<-0.3 Pause infusion immediately, vibrate alarm High risk (L2) Predicted risk of complications within 1 hour > 25% Cloud-based human expert intervention assessment Early Warning (L3) Calibration failed 3 times in a row Switching to redundant sensor mode

[0064] Lightweight deployment at the edge: the TCN model is compressed to 150KB through knowledge distillation, with inference latency less than 50ms; metabolic parameters are stored in a secure enclave, supporting 72 hours of offline operation.

[0065] Cloud-based digital twins are empowered, with personalized metabolic models encompassing over 5,000 physiological parameters and supporting complication probability prediction (AUC = 0.93). Differential privacy noise (ε = 0.5) is added during federated learning aggregation to meet the requirements of the revised HIPAA 2025.

[0066] Dynamic model update mechanism, incremental hot update, push model difference packages (average size 45KB) every 24 hours, merged using the BSDiff algorithm; A / B testing and grayscale release: 5% of the user group will be given priority to experience the new algorithm, such as the reinforcement learning exploration coefficient ε adjusted from 0.3 to 0.15, and the full version will be pushed after meeting the standard.

[0067] The multimodal sensing module specifically includes:

[0068] Anti-drying unit: The nanoscale anti-interference layer consists of a polyurethane substrate doped with the conductive polymer PEDOT:PSS and uniformly dispersed with cerium oxide nanoparticles. A porous network structure is formed by electrochemical deposition, and the catalytic activity of CeO2 accelerates the decomposition of H2O2, inhibiting baseline drift caused by the accumulation of enzyme reaction byproducts. A zwitterionic polymer is grafted onto the surface. Dynamic Raman spectroscopy correction is performed using a low-power laser to penetrate the skin surface and collect Raman spectra of subcutaneous tissue. A mapping relationship between glucose concentration and spectral signals is established through the characteristic peak intensity ratio.

[0069] Cross-modal data fusion unit: inputs temperature sensor data, accelerometer data and enzyme electrode signals into the extended Kalman filter to construct the state equation;

[0070] Calibration trigger unit: When triggered by a dietary event, the camera captures the plate image, uses the ResNet-18 model to identify the food type, and combines the CR value to calculate the expected blood glucose fluctuation threshold. Calibration is initiated when the enzyme electrode reading deviates from the expected value by more than a predetermined ratio. When triggered by exercise status, the accelerometer automatically shortens the Raman spectrum detection interval after detecting moderate to high-intensity exercise. When abnormal drift is determined, if the Raman spectrum correction coefficient exceeds the predetermined threshold multiple times in a row, the user is forced to perform a fingerstick blood calibration.

[0071] Miniaturized packaging unit: A three-dimensional stacked structure with a flexible polyimide substrate, an enzyme electrode array integrated on the upper layer, Raman lasers and photodetectors embedded in the middle layer, and temperature / acceleration sensors arranged on the bottom layer. Electrical connections between components are achieved through gold wire bonding.

[0072] Motion artifact suppression unit: Acceleration signal classification, using support vector machines to distinguish movement types, and dynamically adjusting filtering parameters for different activity levels; high-frequency noise filtering, designing a fourth-order Butterworth low-pass filter; skin contact uses a bionic microneedle array to penetrate the stratum corneum and directly contact the interstitial fluid of the dermis; the needle surface is coated with a hyaluronic acid sustained-release layer to maintain the hydration state of the detection interface.

[0073] It should be noted that the anti-drying unit:

[0074] The composite anti-interference layer structure consists of a polyurethane substrate doped with a conductive polymer PEDOT:PSS (mass ratio 1:0.3) and uniformly dispersed cerium oxide (CeO2) nanoparticles (particle size 20-50nm). The porous network structure (pore size 200-500nm) is formed by electrochemical deposition. The catalytic activity of CeO2 accelerates the decomposition of H2O2 and inhibits baseline drift caused by the accumulation of enzyme reaction byproducts.

[0075] Surface treatment, grafting zwitterionic polymer (sulfobetaine methacrylate), reduced the contact angle to 15°, reduced nonspecific protein adsorption (adsorption amount < 5ng / cm 2 ).

[0076] Dynamic correction of Raman spectrum. The optical system uses a 785nm low-power laser (≤20mW) with a penetration depth of 1.2mm, combined with an InGaAs photodetector (response time 10ns) to collect Raman spectra of subcutaneous tissue.

[0077] Characteristic peak analysis, establishing 1090cm -1 (glucose characteristic peak) and 1450cm +1The quadratic polynomial mapping relationship between the (background reference peak) intensity ratio (IG / IB) and glucose concentration (R 2 ≥0.98); dynamic compensation, spectral calibration was performed every 30 minutes to compensate for enzyme electrode drift.

[0078] Cross-modal data fusion unit:

[0079] Extended Kalman filter design, state equation construction:

[0080]

[0081] Where x k is the state vector, z k is the observation vector; G tissue is the glucose concentration in the dermis, indicating the glucose concentration value measured from the skin tissue; T sensor is the sensor temperature, which indicates the temperature of the sensor itself and is used to monitor the working environment status of the sensor; a RMS is the root mean square value of acceleration, which represents the root mean square value of the acceleration signal measured by the accelerometer and is used to capture motion state or vibration information; G enzyme G is the glucose concentration generated by the enzymatic reaction, indicating the glucose concentration value generated by the enzymatic reaction; raman is the glucose concentration measured by Raman spectroscopy, which means the glucose concentration measured by Raman spectroscopy; T ambient The ambient temperature indicates the temperature of the environment where the sensor is located and is used to provide temperature information of the external environment.

[0082] Covariance optimization, dynamic adjustment of process noise matrix Q (temperature coefficient β = 0.017 / ℃) and observation noise matrix R (motion interference coefficient γ = 0.0025 / (m / s 2 )).

[0083] Real-time fusion output and temperature compensation are based on MEMS thermopile data and use a fifth-order polynomial to fit the sensor temperature-sensitivity curve. Motion interference suppression uses accelerometer data (sampling rate 100Hz) to identify motion artifacts and dynamically adjust the filter cutoff frequency (adjustable from 0.5 to 5Hz).

[0084] Calibration trigger unit:

[0085] Dietary event triggering, image recognition, integrated 5MP micro camera, using the ResNet-18 model to identify the type of food on the plate and estimate carbohydrate equivalents; expected deviation calculation, combined with the user's CR value, predicts the blood sugar fluctuation threshold ΔG 1 hour after the meal = CR × carbohydrate intake × 0.18;

[0086] Calibration conditions: When the enzyme electrode reading deviates from the predicted value by >15%, Raman spectroscopy compensation is activated (priority level L2).

[0087] Exercise status triggering, intensity grading, accelerometer detection of moderate to high intensity exercise (>3METs for 10 minutes), shortening the Raman detection interval to 10 minutes / time; sensitivity adjustment, dynamically improving the insulin sensitivity factor ISF within 2 hours after exercise (regulation coefficient 1.12).

[0088] Abnormal drift judgment, mandatory calibration protocol, when the Raman correction coefficient δ>0.25 for three consecutive times, trigger Bluetooth finger blood calibration (priority L1);

[0089] Safety lock mechanism: if calibration fails more than 5 times, monitoring will be suspended and hardware self-test will be started (diagnostic codes E201-E205).

[0090] Miniaturized packaging unit:

[0091] Stacked structure design, flexible substrate, 50μm thick polyimide-graphene composite substrate (stretching rate ≥30%), with serpentine traces (line width 20μm) on the surface;

[0092] Layered integration:

[0093] Upper layer: 8×8 enzyme electrode array (1.5mm spacing), covering an area of ​​12mm 2 ;

[0094] Middle layer: VCSEL laser (wavelength 785nm) and SPAD photodetector array;

[0095] Bottom layer: MEMS temperature sensor (PT1000) and triaxial accelerometer (range ±16g).

[0096] Interconnection technology: gold wire bonding, 25μm diameter gold wire, bonding strength >5gf; conductive silver glue filling, porosity <5%, ensuring high-frequency signal integrity (impedance fluctuation <10%).

[0097] Motion Artifact Suppression Unit:

[0098] Movement type recognition: support vector machine classifier distinguishes walking, running, and sitting, and dynamically adjusts filtering parameters; fourth-order Butterworth filter, adaptive cutoff frequency (0.5-5Hz), group delay ≤20ms, ripple <0.1dB.

[0099] Bionic microneedle array, structural parameters, conical hollow microneedles (height 300μm, base diameter 100μm), needle tip opening diameter 10μm; functional coating, the needle surface is coated with a hyaluronic acid sustained-release layer (drug loading 0.5μg / needle) to maintain the hydration state of the detection interface (humidity > 85%); penetration mechanism, pre-loaded spring (elastic force 0.3N) ensures that the microneedle penetrates the stratum corneum (puncture success rate > 99%) and directly contacts the interstitial fluid of the dermis.

[0100] The metabolic feature fusion calibration module specifically includes:

[0101] Personalized metabolic parameter database: including glycated hemoglobin, insulin sensitivity factor, and carbohydrate ratio; metabolic feature vector generation, extracting features based on historical data, including blood glucose fluctuation patterns, and identifying user-specific patterns through K-means clustering; insulin activity decay curve, establishing a bi-exponential model; metabolic rate calibration, estimating basal metabolic rate by combining heart rate variability and respiratory entropy;

[0102] Temporal convolutional network time series modeling: The input layer includes raw enzyme electrode signals, dietary records, exercise energy consumption, and historical blood sugar fluctuations. The network structure is an eight-layer causal dilated convolution, and the output of each layer is processed by a gated activation unit. Skip connections fuse shallow features, and the output layer is a bidirectional LSTM to predict the future short-term blood sugar drift compensation coefficient.

[0103] Multimodal data fusion unit: Priority weight allocation, with Raman spectroscopy data being prioritized during the golden standard period, after meals, and enzyme electrode signals being relied upon during sleep; conflict resolution mechanism, initiating metabolic model arbitration when the difference in data from different sensors exceeds a predetermined ratio; regular evaluation of calibration results, online optimization of TCN network parameters using gradient descent; and introduction of adversarial training to generate data for extreme metabolic scenarios;

[0104] Event-driven calibration trigger unit: Metabolic event detection. When postprandial hyperglycemia does not reach the expected level, carbohydrate intake is greater than the preset value through image recognition, activating TCN short-term prediction calibration. When abnormal hypoglycemia after exercise is detected by the accelerometer and exceeds the set threshold, the calibration interval is shortened. When the hourly change in the insulin sensitivity mutation ISF value exceeds the set ratio, the fingerstick blood verification calibration is forced to start.

[0105] Cross-modal calibration execution process: The Raman spectroscopy unit collects baseline values ​​and corrects the subcutaneous glucose concentration using a skin transmittance compensation algorithm; the enzyme electrode signal is preprocessed by temperature compensation and motion artifact filtering; the TCN network outputs a drift compensation coefficient, which is weighted and fused with the metabolic model prediction value; if the error after calibration is still greater than the established ratio, a hardware-level reset is triggered, the sensor power supply circuit is restarted, and the enzyme electrode working potential is reinitialized.

[0106] It should be noted that dynamic modeling of individual metabolic parameters

[0107] 1 Metabolic parameter database, dynamic update of glycated hemoglobin, based on the red blood cell lifespan model and the time-weighted average of blood glucose, incrementally updates the HbA1c estimate every 30 days, avoiding the lag of traditional laboratory testing.

[0108] An implantable fluorescent sensor monitors the red blood cell glycation ratio in real time (excitation wavelength 485 nm / emission wavelength 535 nm) to verify the accuracy of the model.

[0109] Hourly optimization of insulin sensitivity factor

[0110] Reinforcement learning model inputs: standard deviation of blood glucose fluctuation (SD ≥ 1.2 mmol / L triggers ISF correction); post-exercise sensitivity enhancement factor (β1 = 1.12, duration 2 hours); stress hormone levels (estimated by HRV LF / HF ratio);

[0111] Output: ISF value dynamic adjustment range ±35% (basic value 0.05-0.2U / mmol / L).

[0112] Intelligent adaptation of carbohydrate ratio, differentiated CR values ​​by time period (breakfast CR = lunch CR × 0.85, dinner CR = lunch CR × 1.1); based on the intestinal absorption rate model (fitted by a third-order exponential function), the association weight between CR value and eating time is adjusted.

[0113] Metabolic feature vector engineering

[0114] Blood glucose fluctuation pattern recognition, feature extraction, intra-day standard deviation, average amplitude blood glucose excursion, and post-meal peak time; K-means clustering (k=5) generates user-specific pattern labels;

[0115] Modeling the decay of insulin activity, a double exponential equation:

[0116]

[0117] Where, IOB(t) is insulin activity, which indicates the activity of insulin at time t; A1 is the rapid phase ratio, which is 0.6 and indicates the proportion of rapid insulin action; k1 is the rapid phase attenuation coefficient, which is 0.12 / min and indicates the decay rate of rapid insulin action; A2 is the slow phase ratio, which is 0.4 and indicates the proportion of slow insulin action; k2 is the slow phase attenuation coefficient, which is 0.003 / min and indicates the decay rate of slow insulin action;

[0118] The TCN network architecture uses an expanded causal convolution design with eight convolution layers with expansion coefficients of 1, 2, 4, 8, 16, 32, 64, and 128, respectively, to capture multi-scale temporal features. The gated activation unit introduces a dynamic weight mechanism (parameterized ReLU function) to enhance nonlinear modeling capabilities.

[0119] Adversarial training enhances robustness. The generative adversarial network simulates extreme scenarios, including strenuous exercise after a meal (blood sugar drop rate > 0.7 mmol / L / min); insulin resistance mutation (ISF value changes > 40% per hour); and the adversarial sample injection ratio is 15% of the training dataset.

[0120] Multimodal data fusion strategy, dynamic allocation of priority weights:

[0121] Time period Primary data source Weight distribution (Raman: enzyme electrode) 0-2 hours after a meal Raman spectroscopy 70％:30％ Nighttime 10:00 PM - 6:00 AM enzyme electrode 30％:70％ During exercise Accelerometer compensation signal 50％:50％

[0122] Conflict arbitration mechanism, difference threshold: the difference between Raman and enzyme electrode data is greater than 15% or deviates from the metabolic model prediction value for 20 minutes; arbitration model: meta-classifier based on XGBoost (input includes temperature, exercise intensity, IOB value), covering abnormal data when the confidence level is greater than 90%.

[0123] Event-driven calibration logic: When postprandial hyperglycemia does not meet expectations and the image recognition carbohydrate error is greater than 10% (ResNet-18 model confidence level less than 85%), the short-term prediction calibration module of the TCN network is activated (prediction window 30 minutes); the calibration sensitivity is dynamically adjusted. If the blood sugar rises by less than 2.2mmol / L 1 hour after a meal, it is judged as an insulin overdose, triggering an infusion volume correction.

[0124] Abnormal hypoglycemia after exercise, accelerometer detection threshold, moderate-intensity exercise: acceleration root mean square (RMS)

[0125] >0.5g for 10 minutes; high-intensity exercise: RMS>1.2g for 5 minutes; response strategy, shorten the Raman spectrum calibration interval to 5 minutes / time, and reduce the basal rate of insulin infusion (coefficient α=0.7).

[0126] Hardware-level calibration is performed, using a skin transmittance compensation algorithm and baseline value collection. A full skin spectrum scan (wavelength 650-950nm) is performed every 24 hours to construct a user-specific transmittance curve.

[0127] Real-time compensation formula:

[0128]

[0129] Where G corrected is the corrected glucose concentration value, which is the final result after transmittance and temperature compensation, and is used to more accurately reflect the actual blood glucose level; Graw is the raw glucose concentration value, which is the uncorrected data directly obtained from the sensor and may be affected by factors such as skin transmittance and temperature; T baseline is the baseline transmittance, which is the value on the transmittance curve obtained by scanning the entire skin spectrum once every 24 hours, representing the transmittance characteristics of the skin in the baseline state; T real-time is the real-time transmittance, which is the transmittance of the skin during actual measurement and may vary due to factors such as the environment and skin condition; ΔT skin is the skin temperature variation, which is the difference between the real-time measured skin temperature and the baseline temperature;

[0130] Hardware-level abnormal reset protocol, reset conditions, error > 12% or sensor impedance mutation > 30% after three consecutive calibrations; execution steps: cut off the enzyme electrode working voltage (3.2V→0V, lasting 500ms); reinitialize the reference electrode potential (Ag / AgCl polarization to +0.3V); gradient restore working voltage (0→3.2V, slope 0.5V / s).

[0131] The adaptive pump control module specifically includes:

[0132] The first layer: Calculate the basic pump rate. The input parameters include the current blood glucose level, blood glucose change rate, and insulin activity in vivo. Fuzzy subsets are divided using trapezoidal function, Gaussian function, and trigonometric function, and a fuzzy rule base is established. The center of gravity method is used to calculate the basic infusion rate. The maximum infusion volume is set according to the IOB value.

[0133] The second layer: Dynamic parameter adjustment driven by metabolic prediction. This uses an LSTM metabolic prediction model, with input features including time series data and event data. The network structure consists of a bidirectional LSTM layer and a self-attention mechanism, outputting a blood glucose prediction curve for the short term in the future. A Kalman filter is used to correct model deviations in real time. Key features of the prediction curve, including peak blood glucose, time to peak, and area under the curve, are extracted to construct fuzzy variables for dynamic adjustment.

[0134] The MEMS piezoelectric pump uses a 3D-printed titanium alloy pump chamber and piezoelectric ceramic actuator; an integrated high-precision flow sensor; the inner wall of the drug reservoir is coated with a heparin-like coating; a pulse reverse cleaning protocol is set to automatically execute reverse pressure pulses at regular intervals; a real-time control loop, and timing synchronization includes triggering the ADC to read sensor data acquisition through hardware interrupts, implementing fuzzy PID calculations through dedicated FPGA logic circuits, and directly controlling the pump body execution through the piezoelectric driver PWM pulse.

[0135] It should be noted that the input parameter fuzzification strategy, blood glucose value (BG) fuzzy division, trapezoidal function defines three fuzzy subsets:

[0136]

[0137] Blood glucose rate of change (ROC) modeling was performed, and three fuzzy subsets (unit: mmol / L / min) were defined by Gaussian function: rapid decrease (μ = -0.5, σ = 0.2), stable (μ = 0, σ = 0.1), and rapid increase (μ = +0.5, σ = 0.2).

[0138] The edge-cloud collaborative architecture specifically includes:

[0139] Lightweight AI model compression: Through knowledge distillation, the cloud-trained LSTM metabolic prediction model is compressed into a TCN network. Eight-bit fixed-point quantization uses the TensorFlow Lite Micro framework, with dynamic pruning to dynamically shut down redundant neurons based on real-time data streams. Real-time inference integrates edge TPUs, supports parallel processing of multimodal sensor data, and uses a circular buffer for memory management. Local control decisions and priority task scheduling utilize an offline emergency mechanism to store core control logic and maintain short-term basic operation during network outages. An LRU cache strategy is used to retain the most recent day's metabolic signature data for local model invocation.

[0140] Cloud-based digital twin platform: Federated learning for global optimization. Through a secure aggregation protocol and Paillier homomorphic encryption, edge nodes upload model gradients, which are aggregated in the cloud to generate global model updates. Differential privacy noise is introduced. Each user has their own personalized model branch, and the cloud quickly adapts to new users through meta-learning. Branches share the underlying feature extraction layer, but have exclusive access to the top-level decision-making parameters.

[0141] Digital twin metabolic modeling, multi-scale physiological simulation, organ-level modeling, and simulation of pancreatic insulin secretion dynamics based on the finite element method; molecular-level modeling, simulating the insulin receptor binding process through molecular dynamics; risk warning logic, short-term warning, combining real-time blood sugar trends and weather data; long-term warning, predicting the risk of diabetes complications through digital twin deduction;

[0142] Report generation uses GPT-4Turbo to generate personalized recommendation text; embedded explainable AI displays key decision-making basis; 3D metabolic map, WebGL rendering of the user's whole-body glucose distribution heat map, annotated insulin-sensitive areas; VR mode allows doctors to immersively observe historical metabolic event chains;

[0143] Edge-cloud collaboration: The hierarchical transmission strategy includes abnormal alarms as the highest level, with instant transmission and no compression; metabolic feature vectors as the high level, with hourly transmission and LZ4 compression; raw sensor data as the low level, with daily transmission and Huffman encoding; the protocol stack design includes CoAP over DTLS as the transport layer, which supports reliable communication for low-power devices; the application layer is a custom binary protocol; model hot updates, incremental push, the cloud generates model difference packages every day, and the edge merges and updates them through the BSDiff algorithm; update verification, using EdDSA digital signature verification; A / B test diversion, a predetermined proportion of users as the gray release group, first testing the new control strategy; feedback data is collected through edge embedding points, and the full data is pushed after meeting the standards.

[0144] It should be noted that the knowledge distillation technology distills the cloud-based LSTM model (1.2M parameters) into a TCN network (240K parameters), retaining 98% prediction accuracy; the adaptive temperature coefficient τ is dynamically adjusted (0.5-5.0) to balance the generalization ability and computing load in complex scenarios.

[0145] Dynamic pruning and quantization, based on the TensorFlow Lite Micro framework, implements mixed precision quantization, retains 8-bit fixed point (Q8.7 format) for key weights (such as temporal convolution kernels); activation functions use 4-bit dynamic range compression (average error <0.8%); real-time neuron importance evaluation (neurons with entropy values ​​>1.2 are retained, and the rest are dynamically closed).

[0146] The computing architecture design and customized edge TPU (computing power 4TOPS, power consumption ≤1W) support multi-modal data parallel processing (3-way parallel pipeline); the memory circular buffer (capacity 8MB) uses DMA direct access with a latency of <50μs.

[0147] Emergency control logic and offline core algorithms are stored in NOR Flash (occupying 256KB), supporting 72 hours of operation without network connection.

[0148] The cloud-based digital twin platform uses a secure aggregation protocol based on Paillier homomorphic encryption (key length 2048 bits). Edge nodes upload gradient values ​​(in ciphertext form). Gaussian noise (σ = 0.05) is added before cloud-based aggregation to meet the differential privacy requirement of ε = 0.7.

[0149] The personalized model architecture shares the underlying feature extraction layer (the first 6 layers of the TCN network) and has its own top-level decision layer (LSTM+Attention). The meta-learning initialization strategy uses transfer learning for new users in the first 72 hours (source domain user similarity > 85%).

[0150] Multi-scale metabolic simulation technology, organ-level modeling, finite element meshing of pancreatic β cell clusters (1μm resolution), and simulation of glucose-stimulated calcium ion oscillations;

[0151] Molecular modeling, molecular dynamics simulation of insulin receptor binding (GROMACS force field, step size 2fs); binding free energy calculation (MM / PBSA method) prediction of drug sensitivity (correlation coefficient R 2 =0.91).

[0152] Reference Figure 2 As shown, a method for dynamic blood glucose monitoring and intelligent pump speed adjustment includes:

[0153] An enzyme electrode sensor coated with a nanoscale anti-interference layer continuously acquires subcutaneous interstitial fluid glucose concentration signals. It also simultaneously collects ambient temperature, three-dimensional motion acceleration, and Raman spectroscopy data from the skin surface. This data is then cross-modally fused using an extended Kalman filter to suppress motion artifacts and temperature drift interference.

[0154] A time-series convolutional network calibration model is constructed based on the user's individual metabolic parameters. When dietary events or moderate to high-intensity exercise are detected, a Raman spectroscopy dynamic compensation algorithm is triggered, correcting skin transmittance deviations using a Monte Carlo light transmission model. Only when the error exceeds a certain percentage after a predetermined number of consecutive Raman compensations is the Bluetooth-connected portable blood glucose meter activated for mandatory fingerstick calibration.

[0155] Two-layer fuzzy PID pump speed regulation: The first layer generates a basal infusion rate based on the current blood glucose level, blood glucose change rate, and insulin activity in the body through a fuzzy rule base. The second layer introduces an LSTM metabolic prediction model, inputting the expected short-term dietary calories and exercise plan to dynamically adjust the slope of the infusion curve.

[0156] The local edge computing unit deploys a lightweight TCN model to perform real-time signal processing and control instruction generation; the cloud aggregates multi-user data through federated learning, updates the global digital twin model and generates personalized metabolic risk warning reports.

[0157] It should be noted that the double-layer fuzzy PID pump control, basic control layer, fuzzy logic decision-making, fuzzification of input parameters, and the division of blood glucose values ​​into five fuzzy sets: hypoglycemia, ideal, postprandial fluctuation, sustained hyperglycemia, and extreme hyperglycemia; the blood glucose change rate (ROC) adopts the Gaussian membership function (μ = ±0.5mmol / L / min).

[0158] Defuzzification method: center of gravity weighted average, output resolution 0.025U / h.

[0159] Prediction optimization layer, metabolism-driven dynamic parameter adjustment, LSTM-Transformer fusion model, input: dietary calories for the next 2 hours (image recognition error ±5g), exercise intensity (Apple Watch data synchronization); output: insulin infusion curve slope adjustment factor α∈[0.5,2.0].

[0160] Miniaturized packaging, flexible three-dimensional stacking structure (12×12×3mm), total weight 3.2g; MEMS piezoelectric pump flow accuracy ±1% (0-50U / h), reverse pulse cleaning cycle 24h.

[0161] Privacy and security: edge data encryption (AES-256+EdDSA signature); hardware-level secure enclave to store metabolic parameters.

[0162] In summary, the advantages of the present invention are:

[0163] A comprehensive, intelligent, closed-loop metabolic management system has been established, enabling deep collaborative innovation from perception to decision-making. Through a dual-mode calibration system combining nano-anti-interference coatings and Raman spectroscopy, the system overcomes the limitations of single sensors susceptible to environmental interference. Combined with a Monte Carlo optical transmission model, the system corrects tissue property deviations in real time, achieving clinical-grade non-invasive monitoring accuracy while significantly reducing the frequency of manual calibration.

[0164] The metabolism-driven intelligent control architecture achieves a paradigm shift, deeply integrating time series modeling, fuzzy logic and prediction algorithms, dynamically optimizing pump control parameters based on individual metabolic rhythms and future behavior predictions, and forming a leapfrog upgrade from passive response to active regulation.

[0165] The edge-cloud collaborative system reshapes the path to intelligent medical equipment. Through lightweight model compression and privacy-preserving computing technologies, it ensures real-time control accuracy while enabling secure data transfer and continuous system evolution. Miniaturized integrated design transcends the size and functionality limitations of traditional devices. The combination of flexible electronics and biomimetic microneedle arrays significantly enhances wearer comfort while improving detection stability, creating a new experience in non-invasive metabolic management.

[0166] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic blood glucose monitoring and intelligent pump speed adjustment system, characterized in that: include: Multimodal sensing module: Integrates an enzyme electrode sensor, a temperature compensation sensor, and a motion acceleration sensor. The enzyme electrode sensor is coated with a nanoscale anti-interference layer for simultaneous detection of subcutaneous interstitial fluid glucose concentration, ambient temperature, and user motion status. The non-invasive Raman spectroscopy detection unit compensates for enzyme electrode signal drift through a dynamic skin transmittance correction algorithm. Metabolic feature fusion calibration module: User metabolic parameter database, storing personalized glycated hemoglobin, insulin sensitivity factor, and carbohydrate ratio data; dynamic calibration algorithm, based on time convolution network, performs time series modeling on the enzyme electrode raw signal, and input parameters include real-time diet records, exercise energy consumption, and historical blood sugar fluctuation patterns; Adaptive pump control module: a two-layer fuzzy PID controller, the first layer calculates the basal pump speed based on the current blood glucose level, blood glucose change rate, and insulin activity in the body; The second layer introduces a metabolic prediction model to dynamically adjust the slope of the insulin infusion curve; Edge-cloud collaborative architecture: Local edge computing units deploy lightweight AI modules to perform real-time signal processing and control decisions; The cloud-based digital twin platform aggregates multi-user data through federated learning to optimize the global model and generate personalized metabolic risk warning reports.

2. A dynamic blood glucose monitoring and intelligent pump speed adjustment system according to claim 1, characterized in that: The multimodal sensing module specifically includes: Anti-drying unit: The nanoscale anti-interference layer consists of a polyurethane substrate doped with the conductive polymer PEDOT:PSS and uniformly dispersed with cerium oxide nanoparticles. A porous network structure is formed by electrochemical deposition, and the catalytic activity of CeO2 accelerates the decomposition of H2O2, inhibiting baseline drift caused by the accumulation of enzyme reaction byproducts. A zwitterionic polymer is grafted onto the surface. Dynamic Raman spectroscopy correction is performed using a low-power laser to penetrate the skin surface and collect Raman spectra of subcutaneous tissue. A mapping relationship between glucose concentration and spectral signals is established through the characteristic peak intensity ratio. Cross-modal data fusion unit: inputs temperature sensor data, accelerometer data and enzyme electrode signals into the extended Kalman filter to construct the state equation; Calibration trigger unit: Triggered by a dietary event, the camera captures the plate image, uses the ResNet-18 model to identify the food type, and combines the CR value to calculate the expected blood sugar fluctuation threshold. Calibration is initiated when the enzyme electrode reading deviates from the expected value by more than a predetermined ratio; triggered by exercise status, the accelerometer automatically shortens the Raman spectrum detection interval after detecting medium to high intensity exercise; abnormal drift judgment: if the Raman spectrum correction coefficient is greater than the predetermined threshold for multiple consecutive times, the user is forced to perform a finger blood calibration.

3. A dynamic blood glucose monitoring and intelligent pump speed adjustment system according to claim 2, characterized in that: The multimodal sensing module specifically includes: Miniaturized packaging unit: A three-dimensional stacked structure with a flexible polyimide substrate, an enzyme electrode array integrated on the upper layer, Raman lasers and photodetectors embedded in the middle layer, and temperature / acceleration sensors arranged on the bottom layer. Electrical connections between components are achieved through gold wire bonding. Motion artifact suppression unit: Acceleration signal classification, using support vector machines to distinguish movement types, and dynamically adjusting filtering parameters for different activity levels; high-frequency noise filtering, designing a fourth-order Butterworth low-pass filter; skin contact uses a bionic microneedle array to penetrate the stratum corneum and directly contact the interstitial fluid of the dermis; the needle surface is coated with a hyaluronic acid sustained-release layer to maintain the hydration state of the detection interface.

4. A dynamic blood glucose monitoring and intelligent pump speed adjustment system according to claim 3, characterized in that: The metabolic feature fusion calibration module specifically includes: Personalized metabolic parameter database: including glycated hemoglobin, insulin sensitivity factor, and carbohydrate ratio; metabolic feature vector generation, extracting features based on historical data, including blood glucose fluctuation patterns, and identifying user-specific patterns through K-means clustering; insulin activity decay curve, establishing a bi-exponential model; metabolic rate calibration, estimating basal metabolic rate by combining heart rate variability and respiratory entropy; Temporal convolutional network time series modeling: The input layer includes raw enzyme electrode signals, dietary records, exercise energy consumption, and historical blood sugar fluctuations. The network structure is an eight-layer causal dilated convolution, and the output of each layer is processed by a gated activation unit. Skip connections fuse shallow features, and the output layer is a bidirectional LSTM to predict the future short-term blood sugar drift compensation coefficient. Multimodal data fusion unit: Priority weight allocation, Raman spectroscopy data is preferred within the golden standard period after meals, and enzyme electrode signals are relied upon during sleep; conflict resolution mechanism, when the difference in data from different sensors is greater than a predetermined ratio, metabolic model arbitration is initiated; regular evaluation of calibration effects, online optimization of TCN network parameters through gradient descent method; introduction of adversarial training to generate extreme metabolic scenario data.

5. A dynamic blood glucose monitoring and intelligent pump speed adjustment system according to claim 4, characterized in that: The metabolic feature fusion calibration module specifically includes: Event-driven calibration trigger unit: Metabolic event detection. When postprandial hyperglycemia does not reach the expected level, carbohydrate intake is greater than the preset value through image recognition, activating TCN short-term prediction calibration. When abnormal hypoglycemia after exercise is detected by the accelerometer and exceeds the set threshold, the calibration interval is shortened. When the hourly change in the insulin sensitivity mutation ISF value exceeds the set ratio, the fingerstick blood verification calibration is forced to start. Cross-modal calibration execution process: The Raman spectroscopy unit collects baseline values ​​and corrects the subcutaneous glucose concentration using a skin transmittance compensation algorithm; the enzyme electrode signal is preprocessed by temperature compensation and motion artifact filtering; the TCN network outputs a drift compensation coefficient, which is weighted and fused with the metabolic model prediction value; if the error after calibration is still greater than the established ratio, a hardware-level reset is triggered, the sensor power supply circuit is restarted, and the enzyme electrode working potential is reinitialized.

6. A dynamic blood glucose monitoring and intelligent pump speed adjustment system according to claim 5, characterized in that: The adaptive pump control module specifically includes: The first layer: Calculate the basic pump rate. The input parameters include the current blood glucose level, blood glucose change rate, and insulin activity in vivo. Fuzzy subsets are divided using trapezoidal function, Gaussian function, and trigonometric function, and a fuzzy rule base is established. The center of gravity method is used to calculate the basic infusion rate. The maximum infusion volume is set according to the IOB value. The second layer: Dynamic parameter adjustment driven by metabolic prediction, using the LSTM metabolic prediction model. The input features include time series data and event data. The network structure is a bidirectional LSTM layer and a self-attention mechanism, which outputs a blood glucose prediction curve for the short term in the future. The model deviation is corrected in real time through the Kalman filter. The key features of the prediction curve are extracted, including peak blood glucose, time to peak, and area under the curve, and fuzzy variables are constructed for dynamic adjustment.

7. A dynamic blood glucose monitoring and intelligent pump speed adjustment system according to claim 6, characterized in that: The adaptive pump control module specifically includes: The MEMS piezoelectric pump uses a 3D-printed titanium alloy pump chamber and piezoelectric ceramic actuator; an integrated high-precision flow sensor; the inner wall of the drug reservoir is coated with a heparin-like coating; a pulse reverse cleaning protocol is set to automatically execute reverse pressure pulses at regular intervals; a real-time control loop, and timing synchronization includes triggering the ADC to read sensor data acquisition through hardware interrupts, implementing fuzzy PID calculations through dedicated FPGA logic circuits, and directly controlling the pump body execution through the piezoelectric driver PWM pulse.

8. A dynamic blood glucose monitoring and intelligent pump speed adjustment system according to claim 7, characterized in that: The edge-cloud collaborative architecture specifically includes: Lightweight AI model compression: Through knowledge distillation, the cloud-trained LSTM metabolic prediction model is compressed into a TCN network. Eight-bit fixed-point quantization uses the TensorFlow Lite Micro framework, with dynamic pruning to dynamically shut down redundant neurons based on real-time data streams. Real-time inference integrates edge TPUs, supports parallel processing of multimodal sensor data, and uses a circular buffer for memory management. Local control decisions and priority task scheduling utilize an offline emergency mechanism to store core control logic and maintain short-term basic operation during network outages. An LRU cache strategy is used to retain the most recent day's metabolic signature data for local model invocation. Cloud-based digital twin platform: Federated learning for global optimization. Through a secure aggregation protocol and Paillier homomorphic encryption, edge nodes upload model gradients, which are aggregated in the cloud to generate global model updates. Differential privacy noise is introduced. Each user has their own personalized model branch, and the cloud quickly adapts to new users through meta-learning. Branches share the underlying feature extraction layer, but have exclusive access to the top-level decision-making parameters. Digital twin metabolic modeling, multi-scale physiological simulation, organ-level modeling, and simulation of pancreatic insulin secretion dynamics based on the finite element method; molecular-level modeling, simulating the insulin receptor binding process through molecular dynamics; risk warning logic, short-term warning, combining real-time blood sugar trends and weather data; long-term warning, predicting the risk of diabetes complications through digital twin deduction; Report generation uses GPT-4Turbo to generate personalized recommendation text; embedded explainable AI displays key decision-making basis; three-dimensional metabolic map, WebGL renders the user's whole-body glucose distribution heat map, and marks insulin-sensitive areas; VR mode supports doctors to immersively observe historical metabolic event chains.

9. A dynamic blood glucose monitoring and intelligent pump speed adjustment system and method according to claim 8, characterized in that: The edge-cloud collaborative architecture specifically includes: Edge-cloud collaboration: The hierarchical transmission strategy includes abnormal alarms as the highest level, with instant transmission and no compression; metabolic feature vectors as the high level, with hourly transmission and LZ4 compression; raw sensor data as the low level, with daily transmission and Huffman encoding; the protocol stack design includes CoAP over DTLS as the transport layer, which supports reliable communication for low-power devices; the application layer is a custom binary protocol; model hot updates, incremental push, the cloud generates model difference packages every day, and the edge merges and updates them through the BSDiff algorithm; update verification, using EdDSA digital signature verification; A / B test diversion, a predetermined proportion of users as the gray release group, first testing the new control strategy; feedback data is collected through edge embedding points, and the full data is pushed after meeting the standards.

10. A method for dynamic blood glucose monitoring and intelligent pump speed adjustment, characterized in that: A system for implementing a dynamic blood glucose monitoring and intelligent pump speed adjustment system as described in any one of claims 1 to 9, comprising: An enzyme electrode sensor coated with a nanoscale anti-interference layer continuously acquires subcutaneous interstitial fluid glucose concentration signals. It also simultaneously collects ambient temperature, three-dimensional motion acceleration, and Raman spectroscopy data from the skin surface. This data is then cross-modally fused using an extended Kalman filter to suppress motion artifacts and temperature drift interference. A time-series convolutional network calibration model is constructed based on the user's individual metabolic parameters. When dietary events or moderate to high-intensity exercise are detected, a Raman spectroscopy dynamic compensation algorithm is triggered, correcting skin transmittance deviations using a Monte Carlo light transmission model. Only when the error exceeds a certain percentage after a predetermined number of consecutive Raman compensations is the Bluetooth-connected portable blood glucose meter activated for mandatory fingerstick calibration. Two-layer fuzzy PID pump speed regulation: The first layer generates a basal infusion rate based on the current blood glucose level, blood glucose change rate, and insulin activity in the body through a fuzzy rule base. The second layer introduces an LSTM metabolic prediction model, inputting the expected short-term dietary calories and exercise plan to dynamically adjust the slope of the infusion curve. The local edge computing unit deploys a lightweight TCN model to perform real-time signal processing and control instruction generation; the cloud aggregates multi-user data through federated learning, updates the global digital twin model and generates personalized metabolic risk warning reports.

Images