Ring type blood glucose noninvasive detection device based on multi-modal sensing fusion
By integrating multimodal fusion technology of optical, electrochemical and mechanical sensors, combined with self-supervised learning and Bayesian network models, the accuracy and stability issues of non-invasive blood glucose detection have been solved, enabling personalized blood glucose monitoring and health management.
Patent Information
- Application Number
- CN202511599492.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-01-27
AI Technical Summary
Existing non-invasive blood glucose testing technologies suffer from insufficient accuracy, poor stability, and a lack of personalized adaptation, making it difficult to meet the needs of high-frequency monitoring. In particular, false alarms or missed alarms occur frequently due to individual differences and environmental interference.
Employing multimodal sensing fusion technology, integrating optical, electrochemical, and mechanical sensors, and combining self-supervised learning algorithms to construct a personalized digital twin baseline, cross-modal verification is achieved through variational mode decomposition and dynamic weight allocation using a Bayesian network model, outputting accurate blood glucose values and generating personalized health intervention suggestions.
It significantly improves the accuracy and clinical applicability of non-invasive blood glucose testing, achieves dynamic adaptation to individual physiological changes, reduces the risk of false alarms and missed alarms, and provides personalized health management solutions.
Smart Images
Figure CN121400822A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical testing equipment technology, specifically to a ring-type non-invasive blood glucose testing device based on multimodal sensor fusion. Background Technology
[0002] With the continued rise in the prevalence of chronic metabolic diseases, accurate and convenient blood glucose monitoring has become a core requirement for global health management. Traditional blood glucose testing relies on invasive blood sampling, which presents problems such as pain, infection risks, and operational complexity, making it particularly unsuitable for high-frequency monitoring scenarios. Non-invasive blood glucose testing technology, due to its non-invasiveness and user-friendliness, has become a research hotspot in the field of biosensing in recent years. Existing non-invasive technologies mainly include optical, electrochemical, and mechanical methods; however, single-modal sensors are susceptible to individual differences, environmental interference, and physiological fluctuations, resulting in insufficient detection accuracy and stability. Multimodal sensor fusion technology, by integrating complementary information from different physical fields, has become a key direction for overcoming the bottlenecks of non-invasive testing.
[0003] Traditional single-modal non-invasive blood glucose monitoring technologies have significant limitations. While optical methods can reflect tissue metabolic characteristics, they are easily affected by skin pigmentation, moisture content, and light scattering, resulting in poor signal stability. Electrochemical methods are sensitive to the concentration of metabolites in the microenvironment, but sweat secretion or temperature changes can cause baseline drift, affecting the reliability of long-term monitoring. Mechanical methods indirectly infer blood glucose through skin deformation, but individual skin elasticity and pressure distribution vary greatly, resulting in weak model generalization ability. In addition, existing technologies mostly rely on static threshold judgments and lack dynamic adaptation to user behavior and physiological rhythms, leading to insufficient personalization. For example, the blood glucose baseline of the same user at different times may vary significantly due to changes in metabolic state, but traditional methods cannot adjust the judgment criteria in real time, which easily leads to false alarms or missed alarms. Although multimodal fusion technology has been proposed, most studies simply superimpose data without solving the problems of decoupling interference between modalities and dynamic weight allocation, making it difficult to meet the requirements of clinical-grade accuracy. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a ring-type non-invasive blood glucose monitoring device based on multimodal sensor fusion. This device integrates optical, electrochemical, and mechanical sensors to simultaneously collect and fuse multimodal data from the fingertip. It utilizes a self-supervised learning algorithm to construct a personalized digital twin baseline, combines variational mode decomposition and dynamic weight allocation to reduce interference, and relies on a Bayesian network model to achieve cross-modal validation. Finally, it outputs accurate blood glucose values and generates personalized health intervention suggestions. This device overcomes the limitations of traditional single-modal detection, significantly improving the accuracy and clinical applicability of non-invasive blood glucose monitoring, and providing an innovative solution for chronic disease management.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a ring-type non-invasive blood glucose detection device based on multimodal sensor fusion, the device comprising: Data acquisition module: integrates optical sensors, electrochemical sensors and mechanical sensors to simultaneously acquire optical metabolic data, microenvironment electrochemical data and skin mechanical characteristic data of the fingertip. The acquired data is initially fused through a multimodal data fusion formula. Baseline Construction Module: Based on the initial data from the data acquisition module, combined with the user's entered diet logs, exercise intensity, and circadian rhythm behavior data, a user-specific blood glucose-physiological-behavioral digital twin baseline is constructed using a self-supervised learning algorithm. The baseline is dynamically updated through a baseline update formula. Interference decoupling module: The real-time data acquired by the data acquisition module is decomposed into multiple intrinsic mode functions through variational mode decomposition. Then, based on the dual indicators of mechanical signal stability and electrochemical microenvironment consistency, dynamic weights are assigned to each mode data using the mode deviation calculation formula and dynamic weight allocation formula. Cross-motif validation module: Construct blood glucose prediction sub-networks corresponding to optical, electrochemical, and mechanical modes respectively, forming a Bayesian network model. Calculate the final blood glucose value through a blood glucose value fusion formula. When the signal of one mode is disturbed, the predicted value is automatically corrected based on the consistency results of the other two modes. Results Feedback Module: Displays the final blood glucose test results from the cross-momentum validation module in real time, and combines them with the individual physiological baseline from the baseline construction module to generate personalized health intervention suggestions for diet and exercise through the intervention index calculation formula.
[0006] Furthermore, the multimodal data fusion calculation formula in the data acquisition module is as follows: ,in, To fuse feature values; Normalized data collected by optical sensors; Normalized data collected by electrochemical sensors; Normalized data collected by the mechanical sensor; , , These are the initial weights.
[0007] Furthermore, the baseline update calculation formula in the baseline construction module is as follows: ,in, This represents the baseline drift for the day. This represents the baseline drift from the previous day. This represents the mean of the fusion features for that day. This represents the initial baseline characteristics.
[0008] Furthermore, the modal deviation calculation formula in the interference decoupling module is as follows: ,in, For the first Modal deviation; For the first Real-time data of modalities; For the first Baseline data for the modality.
[0009] Furthermore, the dynamic weight allocation calculation formula in the interference decoupling module is as follows: ,in, The weights are adjusted for the i-th mode; , , These are deviations in optical, electrochemical, and mechanical modes, respectively. For the first Modal deviation.
[0010] Furthermore, the blood glucose value fusion calculation formula in the cross-monotone verification module is as follows: ,in, This is the final predicted blood glucose value; The predicted blood glucose value for the optical mode; The predicted blood glucose value for the electrochemical mode; The predicted blood glucose value for the mechanical mode; , , The credibility weights for each modality.
[0011] Furthermore, the formula for calculating the intervention index in the result feedback module is as follows: ,in, The intervention intensity index; This is the final predicted blood glucose value; Baseline blood glucose level; These are behavioral influencing factors.
[0012] Furthermore, the self-supervised learning algorithm in the baseline construction module employs a contrastive learning framework. It constructs positive and negative sample pairs and optimizes the feature extraction network based on a contrastive loss function, the formula of which is: ,in, To compare the loss function values, used to quantify the deviation of the optimization target of the feature extraction network; This is the feature vector of the currently processed user multimodal data after preliminary feature extraction; To and Positive sample feature vectors derived from multimodal data of the same user at different time periods; To and Negative sample feature vectors derived from concurrent multimodal data from different users; (a,b) is the vector similarity calculation function, specifically the cosine similarity between vector a and vector b, which is used to characterize the degree of similarity between the features of two samples. The temperature adjustment parameter is used to control the steepness of the distribution of similarity calculation results; the data source for constructing the positive and negative sample pairs is multimodal data of the same user at different time periods and multimodal data of different users at the same time.
[0013] Furthermore, the variational mode decomposition in the interference decoupling module is constrained by signal sparsity and reconstruction accuracy, and the number of intrinsic mode functions is adaptively determined through iterative optimization. The objective function during the decomposition process is to minimize the mode energy difference.
[0014] Furthermore, the Bayesian network model in the cross-modal verification module contains three input nodes and one output node. The input nodes correspond to the predicted values of optical, electrochemical, and mechanical modes, respectively, and the output node is the final blood glucose prediction value. The conditional probability distribution between each node is obtained based on the statistical analysis of historical multimodal blood glucose data and is updated periodically according to newly collected data.
[0015] Compared with existing technologies, this ring-type non-invasive blood glucose monitoring device based on multimodal sensor fusion has the following advantages: I. This invention achieves improved accuracy and anti-interference capability of non-invasive blood glucose detection through multimodal data fusion technology integrating optical, electrochemical, and mechanical sensors. The optical sensor captures the metabolic optical characteristics of the fingertip, the electrochemical sensor monitors the concentration of metabolites in the microenvironment, and the mechanical sensor analyzes skin deformation and pressure distribution. The three modal data are decoupled from environmental interference in real time through a dynamic weighted fusion formula, ensuring that even when a single modal signal is abnormal, other modal data can still be cross-validated through a Bayesian network model. Finally, a reliable result is output through the blood glucose value fusion formula. This design overcomes the limitations of traditional single-modal detection, which is easily affected by sweat, temperature, etc., and improves the clinical applicability of non-invasive detection.
[0016] Second, this invention achieves intelligent dynamic health management by constructing a personalized digital twin baseline through a self-supervised learning algorithm. The system combines user diet logs, exercise intensity, and circadian rhythm data, extracts individual metabolic characteristics using a contrastive learning framework, and dynamically adjusts baseline parameters through a baseline update formula. When blood glucose deviation from the baseline is detected, the intervention index calculation formula can integrate blood glucose value, baseline difference, and behavioral influencing factors to generate customized diet and exercise recommendations. This closed-loop management model not only improves the continuity of blood glucose monitoring but also reduces the risk of complications through behavioral intervention, providing an innovative solution for chronic disease management.
[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0019] Figure 1 A flowchart of a ring-type non-invasive blood glucose testing device based on multimodal sensor fusion; Figure 2 This is a framework diagram of a ring-type non-invasive blood glucose detection device based on multimodal sensor fusion. Detailed Implementation
[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0021] Example 1: Daily blood glucose monitoring at home for diabetic patients Scene background A 50-year-old patient with type 2 diabetes for 5 years requires multiple daily blood glucose monitoring sessions to adjust his diet and medication. He regularly exercises in the morning, and his blood glucose fluctuates significantly when his diet is irregular. He uses this ring-type non-invasive blood glucose monitoring device for home monitoring.
[0022] The specific implementation steps are as follows: Figure 1 As shown: Data acquisition module operation: The patient wears the ring-shaped device on the left index finger, and the device's data acquisition module immediately activates. The optical sensor begins collecting optical metabolic data from the fingertip, the electrochemical sensor simultaneously acquires electrochemical data of the fingertip microenvironment, and the mechanical sensor captures skin mechanical characteristic data. After acquisition, the module calls the multimodal data fusion formula to initially fuse these three types of data, obtaining the fused initial data. The multimodal data fusion calculation formula is as follows: ,in, To fuse feature values; Normalized data collected by optical sensors; Normalized data collected by electrochemical sensors; Normalized data collected by the mechanical sensor; , , These are the initial weights.
[0023] The baseline construction module works by receiving initial data from the data acquisition module. Simultaneously, the patient uses the accompanying mobile app to record their previous day's dietary log, including details such as two steamed buns and a cup of soy milk for breakfast, a bowl of rice, stir-fried vegetables, and a small amount of braised pork for lunch, and millet porridge and cucumber salad for dinner. They also record their exercise intensity (30 minutes of brisk walking in the morning) and circadian rhythm data (going to bed at 10 PM and waking up at 6 AM). The module utilizes a self-supervised learning algorithm, combining this data with the initial data, to construct a patient-specific digital twin baseline for blood glucose, physiological function, and behavior. The self-supervised learning algorithm employs a contrastive learning framework, constructing positive and negative sample pairs and optimizing the feature extraction network based on a contrastive loss function. The formula for the contrastive loss function is: ,in, To compare the loss function values; This is the feature vector of the currently processed user multimodal data after preliminary feature extraction; To and Positive sample feature vectors derived from multimodal data of the same user at different time periods; To and Negative sample feature vectors derived from concurrent multimodal data from different users; (a,b) is the vector similarity calculation function; This is the temperature adjustment parameter. Every day thereafter, the module will call the baseline update formula to dynamically update the baseline based on the patient's newly entered information and newly collected initial data, ensuring that the baseline accurately reflects the patient's current physical condition. The baseline update calculation formula is: ,in, This represents the baseline drift for the day. This represents the baseline drift from the previous day. This represents the mean of the fusion features for that day. This represents the initial baseline characteristics.
[0024] Interference decoupling module operation: At 10:00 AM one day, a patient wears a device to perform a blood glucose test in an office. At this time, the office air conditioner blows directly on the patient's hand, which may interfere with the test data. The interference decoupling module first receives the real-time data acquired by the data acquisition module at that moment, and then decomposes the mixed signal in the real-time data into multiple intrinsic mode functions through variational mode decomposition. Next, the module calculates the deviation of each mode data based on two indicators: mechanical signal stability and electrochemical microenvironment consistency. This process calls the mode deviation calculation formula, which is: ,in, For the first Modal deviation; For the first Real-time data of modalities; For the first Baseline data for each modality. Then, based on the calculated deviation, the module calls the dynamic weighting formula to assign dynamic weights to the optical, electrochemical, and mechanical modal data. This reduces the weight of modal data more susceptible to air conditioning blast interference and increases the weight of relatively stable modal data, thereby minimizing the impact of interference on subsequent blood glucose calculations. The dynamic weighting formula is as follows: ,in, The weights are adjusted for the i-th mode; , , These are deviations in optical, electrochemical, and mechanical modes, respectively. For the first Modal deviation.
[0025] The cross-motif validation module works as follows: It receives the modal data and corresponding dynamic weights processed by the interference decoupling module. First, the module constructs sub-networks for blood glucose prediction corresponding to the optical, electrochemical, and mechanical modes, respectively. These three sub-networks together form a Bayesian network model. Each sub-network calculates its respective modal prediction value based on the modal data and weights. Then, the module calls the blood glucose value fusion formula, combining the confidence weights of each modality, to fuse the blood glucose prediction values of the three modalities, obtaining the final blood glucose prediction value. The blood glucose value fusion calculation formula is as follows: ,in, This is the final predicted blood glucose value; The predicted blood glucose value for the optical mode; The predicted blood glucose value for the electrochemical mode; The predicted blood glucose value for the mechanical mode; , , The confidence weights for each modality are defined. Assuming the optical modality signal is slightly disturbed by dry hand skin, the module automatically corrects the final blood glucose prediction value based on the consistency of the electrochemical and mechanical modality prediction results, ensuring accuracy.
[0026] The results feedback module receives the final blood glucose test results from the cross-phasic validation module and displays them in real time on the small display screen of the ring device, which the patient can view directly. Simultaneously, the module combines the latest individual physiological baseline of the patient from the baseline construction module, calls the intervention index calculation formula, and generates personalized health intervention suggestions. The intervention index calculation formula is as follows: ,in, The intervention intensity index; This is the final predicted blood glucose value; Baseline blood glucose level; The behavioral influencing factor is as follows: If the final blood glucose level is slightly higher than the baseline blood glucose level, and the behavioral influencing factor shows that the patient has not exercised that day, the suggestion is "The current blood glucose is slightly high. You can take a 20-minute slow walk around 3 pm. Reduce the intake of staple food at dinner and increase the intake of vegetables by 1 serving"; if the blood glucose level is within the normal range, the suggestion is "The current blood glucose is normal. Maintain the current diet and exercise habits and pay attention to not overeating at dinner".
[0027] In summary, in the scenario of daily home blood glucose monitoring for diabetic patients, the device of this invention achieves accurate and non-invasive detection through multi-module collaboration. The data acquisition module integrates three types of sensors to simultaneously collect multi-dimensional data from the fingertip, which is initially fused using a multimodal data fusion formula; the baseline construction module combines user data such as diet and exercise, constructs a dedicated digital twin baseline using a self-supervised learning algorithm, and dynamically optimizes it using a baseline update formula; the interference decoupling module processes real-time data using variational mode decomposition, and reduces the impact of interference by relying on dual indicators, modal bias, and dynamic weight allocation formulas; the cross-motif validation module uses a Bayesian network model as its core, calculates the final blood glucose value using a blood glucose value fusion formula, and corrects interference bias; the result feedback module displays the results in real time, calls the intervention index calculation formula to generate personalized suggestions, and meets the needs of patients for home monitoring and health management.
[0028] Example 2: Community healthcare blood glucose screening scenario Scene background A community hospital is conducting a free diabetes screening program, targeting residents aged 40 and above in the community. The program uses this ring-type non-invasive blood glucose testing device to conduct rapid blood glucose screening in order to preliminarily determine whether residents have abnormal blood glucose levels, so as to facilitate further examination and intervention.
[0029] The specific implementation steps are as follows: Figure 2 As shown: Data Acquisition Module Operation: A 45-year-old resident, under the guidance of community hospital staff, wears a ring-shaped device on the middle finger of their right hand. After the data acquisition module is activated, the optical sensor collects optical metabolic data from the resident's fingertip, the electrochemical sensor collects microenvironment electrochemical data, and the mechanical sensor collects skin mechanical characteristic data; all three are performed simultaneously. After acquisition, the module calls the multimodal data fusion formula to initially fuse the three types of data, obtaining initial fused data, which is then transmitted to subsequent modules.
[0030] Baseline Construction Module Operation: Since this was the resident's first time using the device, the baseline construction module first acquired initial fusion data from the data acquisition module. Staff assisted the resident in entering their basic dietary information through the hospital's computer system, such as a typical breakfast of porridge and eggs, lunch and dinner mainly consisting of rice with occasional noodles, and a preference for sweets; their exercise history was almost nonexistent; and their circadian rhythm data showed they usually went to sleep after 11 PM and woke up around 7:30 AM. Using a self-supervised learning algorithm, the module combined this entered information with the initial fusion data to construct a personalized digital twin baseline for the resident's blood glucose, physiological, and behavioral patterns. Considering the specific nature of community screening, if the resident participates in screenings again, the module will call the baseline update formula to dynamically update the baseline based on the new entered information and newly collected data.
[0031] Interference decoupling module operation: The screening site was crowded and noisy. The resident's hand trembled slightly during the test, potentially affecting data accuracy. After receiving real-time data from the data acquisition module, the interference decoupling module decomposes the mixed signal into multiple intrinsic mode functions (EMFs) using variational mode decomposition. Then, based on mechanical signal stability and electrochemical microenvironment consistency indices, it calculates the deviation of each mode data using the mode deviation calculation formula. Finally, it uses a dynamic weight allocation formula to assign dynamic weights to each mode data, reducing the weight of modes significantly affected by hand tremors and increasing the weight of other stable modes, thus eliminating some interference factors.
[0032] The cross-motif validation module works as follows: This module receives modal data and dynamic weights processed by the interference decoupling module, constructs blood glucose prediction sub-networks corresponding to optical, electrochemical, and mechanical modes, forming a Bayesian network model. Each sub-network calculates its own predicted blood glucose value based on the corresponding modal data. The module then calls the blood glucose value fusion formula, combines it with the confidence weights of each modality, and fuses them to obtain the final predicted blood glucose value. If a certain modality is interfered with by changes in ambient light, the module automatically corrects the final blood glucose value based on the consistency of the prediction results of the other two modalities, ensuring the reliability of the screening results.
[0033] The results feedback module displays the final blood glucose test results in real time on the computer screen connected to the ring device, allowing staff to view them immediately. Simultaneously, the module combines the resident's individual physiological baseline from the baseline construction module, calls the intervention index calculation formula, and generates health intervention recommendations. If the resident's blood glucose level is high, the recommendation is: "Initial screening shows high blood glucose; it is recommended to go to the hospital for a venous blood glucose test for further confirmation; reduce sweet food intake in daily life, and engage in moderate-intensity exercise at least three times a week for at least 30 minutes each time, such as brisk walking or jogging." If the blood glucose level is normal, the recommendation is: "This screening shows normal blood glucose; continue to maintain good dietary and sleep habits, and have regular health checkups."
[0034] In summary, the device of this invention demonstrates efficient and convenient screening capabilities in community healthcare blood glucose screening scenarios. The data acquisition module rapidly and synchronously collects multimodal data from residents' fingertips, completing initial fusion through a multimodal data fusion formula. The baseline construction module, with staff assistance, combines residents' physiological behavior data and constructs a dedicated baseline using a self-supervised learning algorithm, which is then dynamically updated using a baseline update formula. The interference decoupling module reduces interference in the complex screening environment using variational mode decomposition and related formulas. The cross-modal validation module relies on a Bayesian network model and blood glucose value fusion formula to ensure reliable blood glucose test results. The result feedback module presents the results in real time, calls the intervention index calculation formula to generate targeted suggestions, and helps communities quickly complete blood glucose screening, providing a basis for subsequent medical interventions.
[0035] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A ring-type non-invasive blood glucose monitoring device based on multimodal sensor fusion, characterized in that, The device includes: Data acquisition module: integrates optical sensors, electrochemical sensors and mechanical sensors to simultaneously acquire optical metabolic data, microenvironment electrochemical data and skin mechanical characteristic data of the fingertip. The acquired data is initially fused through a multimodal data fusion formula. Baseline Construction Module: Based on the initial data from the data acquisition module, combined with the user's entered diet logs, exercise intensity, and circadian rhythm behavior data, a user-specific blood glucose-physiological-behavioral digital twin baseline is constructed using a self-supervised learning algorithm. The baseline is dynamically updated through a baseline update formula. Interference decoupling module: The real-time data acquired by the data acquisition module is decomposed into multiple intrinsic mode functions through variational mode decomposition. Then, based on the dual indicators of mechanical signal stability and electrochemical microenvironment consistency, dynamic weights are assigned to each mode data using the mode deviation calculation formula and dynamic weight allocation formula. Cross-motif validation module: Construct blood glucose prediction sub-networks corresponding to optical, electrochemical, and mechanical modes respectively, forming a Bayesian network model. Calculate the final blood glucose value through a blood glucose value fusion formula. When the signal of one mode is disturbed, the predicted value is automatically corrected based on the consistency results of the other two modes. Results Feedback Module: Displays the final blood glucose test results from the cross-momentum validation module in real time, and combines them with the individual physiological baseline from the baseline construction module to generate personalized health intervention suggestions for diet and exercise through the intervention index calculation formula.
2. The ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The multimodal data fusion calculation formula in the data acquisition module is as follows: ,in, To fuse feature values; Normalized data collected by optical sensors; Normalized data collected by electrochemical sensors; Normalized data collected by mechanical sensors; , , These are the initial weights.
3. The ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The baseline update calculation formula in the baseline construction module is as follows: ,in, This represents the baseline drift for the day. This represents the baseline drift from the previous day. This represents the mean of the fusion features for that day. This represents the initial baseline characteristics.
4. The ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The modal deviation calculation formula in the interference decoupling module is as follows: ,in, For the first Modal deviation; For the first Real-time data of modalities; For the first Baseline data for the modality.
5. A ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The dynamic weight allocation calculation formula in the interference decoupling module is as follows: ,in, The weights are adjusted for the i-th mode; , , These are deviations in optical, electrochemical, and mechanical modes, respectively. For the first Modal deviation.
6. A ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The formula for calculating blood glucose fusion in the cross-momentum verification module is as follows: ,in, This is the final predicted blood glucose value; The predicted blood glucose value for the optical mode; The predicted blood glucose value for the electrochemical mode; The predicted blood glucose value for the mechanical mode; , , The credibility weights for each modality.
7. A ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The formula for calculating the intervention index in the result feedback module is as follows: ,in, The intervention intensity index; This is the final predicted blood glucose value; Baseline blood glucose level; These are behavioral influencing factors.
8. A ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The self-supervised learning algorithm in the baseline construction module adopts a contrastive learning framework. It constructs positive and negative sample pairs and optimizes the feature extraction network based on a contrastive loss function, the formula of which is: ,in, To compare the loss function values; This is the feature vector of the currently processed user multimodal data after preliminary feature extraction; To and Positive sample feature vectors derived from multimodal data of the same user at different time periods; To and Negative sample feature vectors derived from concurrent multimodal data from different users; (a,b) is the vector similarity calculation function; These are temperature control parameters.
9. A ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The variational mode decomposition in the interference decoupling module is constrained by signal sparsity and reconstruction accuracy. The number of intrinsic mode functions is determined adaptively through iterative optimization. The objective function during the decomposition process is to minimize the mode energy difference.
10. A ring-type non-invasive blood glucose detection device based on multimodal sensor fusion according to claim 1, characterized in that, The Bayesian network model in the cross-modal validation module contains three input nodes and one output node. The input nodes correspond to the predicted values of optical, electrochemical, and mechanical modes, respectively, and the output node is the final blood glucose prediction value. The conditional probability distribution between each node is obtained based on the statistical analysis of historical multimodal blood glucose data and is updated periodically according to newly collected data.