Smart watch health monitoring management and early warning method based on multi-mode sensor

Abstract

The invention discloses a smart watch health monitoring management and early warning method based on a multi-modal sensor, and the method comprises the following steps: S1, collecting and preprocessing multi-modal physiological data, and generating a time sequence health data set; s2, constructing a health state memory network, and outputting individualized health state characterization; s3, performing cross attention fusion on the layered time sequence features to form a global health feature vector; s4, performing health anomaly detection and scoring by adopting a firefly algorithm; s5, optimizing a firefly algorithm search process by using a genetic algorithm; s6, dynamically adjusting an early warning threshold value, and generating personalized health early warning information; and S7, early warning is triggered during anomaly detection, and a health alarm is sent through wireless communication. Through multi-modal data acquisition, deep health state modeling and an intelligent optimization algorithm, real-time accurate detection and personalized early warning of health abnormity by the smart watch are realized, so that an efficient and reliable health management solution is provided for a user.

Description

Technical Field

[0001] The present invention relates to the technical field of wearable health monitoring, and particularly to an intelligent watch health monitoring management and early warning method based on multi-modal sensors. Background Art

[0002] With the rapid development of Internet of Things, big data and artificial intelligence technologies, wearable devices have gradually penetrated into people's daily lives and are particularly remarkable in the field of health monitoring. In recent years, intelligent watches have gradually become an important tool for health management due to their advantages such as portability, real-time data collection, and multi-functional integration. Currently, the intelligent watches on the market mainly rely on photoplethysmogram, electrocardiogram signals, accelerometers, and other single or a small number of sensors to obtain users' health data, and monitor indicators such as heart rate and blood oxygen through simple signal processing and statistical analysis. However, these methods have many deficiencies. In the prior art, most intelligent watches have problems such as high noise, serious loss, and unstable signals during data collection, and lack an effective mechanism for the fusion processing of sensor data, which cannot fully reflect the complex physiological state of users. Especially in sports, strenuous activities, or extreme environments, single-sensor data is difficult to meet the requirements of precise monitoring, resulting in easy omission or false alarm of abnormal states.

[0003] In terms of health state modeling, traditional methods mainly rely on a single time series model, such as long short-term memory network (LSTM) or Transformer structure, to capture short-term changes and long-term trends respectively. However, a single model often has difficulty in balancing local details and global trends. In the prior art, the fusion processing of multi-modal sensor data generally adopts simple cascading or averaging methods, etc., and fails to fully utilize the cross-correlation between modalities, resulting in inaccurate extraction of the time series characteristics of health data, and the abnormal detection results are easily affected by noise, reducing the accuracy and robustness of the early warning system. At the same time, traditional health early warning strategies usually adopt fixed threshold methods and cannot be adjusted in real time according to the dynamic changes of individual health states, which makes the early warning system less flexible when facing individual differences and the early warning effect is greatly limited.

[0004] In addition, some current studies attempt to introduce meta-heuristic algorithms (such as traditional firefly algorithm) or genetic algorithms for parameter optimization to improve the accuracy and self-adaptability of abnormal detection. However, single algorithms have their own limitations. The traditional firefly algorithm is easily affected by initial parameters and is prone to falling into local optima, while the genetic algorithm often ignores the excavation of local details during the global search process. The respective deficiencies of the two lead to limited application effects in actual health monitoring. Facing such a complex health data environment, how to achieve efficient fusion of multi-modal data, accurate health state modeling, and personalized early warning based on intelligent optimization algorithms has become a key problem to be solved urgently.

[0005] Therefore, how to provide an intelligent watch health monitoring management and warning method based on multi-modal sensors is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to propose an intelligent watch health monitoring management and warning method based on multi-modal sensors. The present invention makes full use of multi-modal sensor data acquisition, deep learning technology, health state memory network, hierarchical time series feature cross-attention mechanism, and intelligent optimization technology combining improved firefly algorithm and genetic algorithm, and details how to implement comprehensive preprocessing and fusion of multi-modal physiological data collected by intelligent watches. By constructing short-term and long-term health state modeling networks to achieve individualized health state time series representation, combined with an improved optimization algorithm to accurately detect health abnormalities, and using an adaptive gradient correction method to dynamically adjust personalized warning thresholds, thereby generating real-time and accurate health warning information. The present invention has the advantages of comprehensive data acquisition, accurate health state modeling, accurate anomaly detection, sensitive warning, and personalized customization, significantly improving the performance and practicality of intelligent watches in health monitoring and warning.

[0007] According to an embodiment of the present invention, an intelligent watch health monitoring management and warning method based on multi-modal sensors includes the following steps:

[0008] S1. Collect multi-modal physiological data of the intelligent watch, preprocess the multi-modal physiological data, and generate a time series health data set;

[0009] S2. Based on the time series health data set, construct a health state memory network, including a short-term health state module and a long-term health trend module, and combine the attention mechanism to output an individualized health state time series representation;

[0010] S3. Use a hierarchical time series feature cross-attention mechanism to perform feature fusion on the time series health data set and the individualized health state time series representation to form a global health feature vector;

[0011] S4. Based on the firefly algorithm, perform health anomaly detection on the global health feature vector, calculate individual health anomaly scores, and by introducing a dynamic adaptive entropy-driven mechanism, dynamically correct the light intensity calculation and step size adjustment strategies using the information entropy of health data;

[0012] S5. Use the genetic algorithm to optimize the search process of the firefly algorithm, generate new individuals by gene crossover, and adjust the mutation rate by adaptive entropy mutation, and select firefly individuals with higher fitness to replace individuals with lower fitness in each iteration;

[0013] S6. Based on the temporal representation of the individualized health status and the individual health anomaly score, dynamically adjust the personalized health warning threshold, and use the adaptive gradient correction method to iteratively adjust the warning parameters to generate personalized health warning information;

[0014] S7. When an individual's health anomaly is detected, trigger a user prompt according to the generated personalized health warning information, and in the case of a serious anomaly, send a health alert message to a preset emergency contact or a medical service platform through the wireless communication module.

[0015] Optionally, the said S2 specifically includes:

[0016] S21. Extract health data from the temporal health dataset to generate a standardized temporal health data sequence X = [x1, x2, …, x T , where x T represents the multimodal health data collected at the t-th time step, and T represents the total number of time steps;

[0017] S22. Construct a short-term health status module, and use a long short-term memory network to model the data within a short time window to generate a short-term health status representation H STM :

[0018] H STM = {h t = σ(W x x t + W h h t-1 + b) | t = T - T s + 1, …, T};

[0019] Wherein, T s is the short-term window length, σ is the activation function, W x and W h are the input and hidden state weight matrices respectively, b is the bias, h t is the hidden state, and h t-1 represents the hidden state at the previous time step;

[0020] S23. Construct a long-term health trend module, and use the Transformer structure to globally model the data within a long time window to generate a long-term health trend representation H LTM :

[0021]

[0022] Wherein, X long is the data within the long time window, T l is the long-term window length, d is the scaling factor, LayerNorm represents the layer normalization operation, and softmax represents the normalization;

[0023] S24. Use the attention mechanism to perform feature fusion on the short-term health status representation H STM and the long-term health trend representation H LTM to generate an individualized health status time-series representation H status :

[0024]

[0025] where W a is the attention weight matrix, H E is the entropy of health data, λ is the entropy-driven adjustment coefficient, μ is the short-term feature weight adjustment parameter, and d is the scaling factor;

[0026] S25. Output the individualized health status time-series representation H status .

[0027] Optionally, the specific steps of S3 are as follows:

[0028] S31. The feature matrix output by the time-series health data set is X, and the individualized health status time-series representation is H status , and an input feature matrix is generated through a concatenation operation:

[0029] F input = Concat(X, H status );

[0030] where F input represents the input feature matrix, and Concat represents the concatenation operation;

[0031] S32. Perform local cross-attention calculation on the input feature matrix F input to generate a low-level fusion feature F low ;

[0032] S33. Perform global cross-attention calculation on the low-level fusion feature F low to generate a high-level fusion feature F high ;

[0033] S34. Use a hierarchical feature fusion mechanism to perform cross-fusion on the low-level fusion feature F low and the high-level fusion feature F high to generate a fusion feature F fused :

[0034] F fused = LayerNorm(σ(κH E )) ⊙ (αF low )) + (1 - σ(κH E )) ⊙ ((1 - α)F high ));

[0035] Among them, σ() represents the activation function, and H E is the entropy of health data information, κ is the entropy adjustment factor, α is the low-level feature weight parameter, ⊙ represents element-wise multiplication, and LayerNorm represents the layer normalization operation;

[0036] S35. Generate the global health feature vector H fused by using the fused feature F input and the input feature F global through the residual connection with entropy modulation:

[0037] H global = LayerNorm(F input + γF fused + δ(F fused ⊙ σ(κH E )));

[0038] Among them, γ and δ are learnable weight coefficients.

[0039] Optionally, the specific steps of S4 include:

[0040] S41. Use the output global health feature vector H global as the input of the improved firefly algorithm;

[0041] S42. Initialize the firefly population in the firefly algorithm, and the position of the firefly individual is where i = 1, 2,..., N, R is the set of real numbers, d ′ is the dimension, and N is the total number of the firefly population. The initial position is randomly sampled from the global health feature vector;

[0042] S43. In each iteration, update the light intensity value I E between any two fireflies X i and X j by introducing the drive of the health data information entropy H ij :

[0043]

[0044] Among them, represents the light intensity value of firefly X i to firefly X j in the t-th generation. exp() represents the exponential function, γ1 is the light absorption coefficient, λ1 is the entropy adjustment coefficient, μ is the gradient modulation coefficient, represents the local gradient of the position of firefly X i in the t-th generation, represents the local gradient of the position of firefly X j in the t-th generation, Denote the updated firefly X in the (t + 1)-th generation i For the firefly X j The light intensity value;

[0045] S44. Comprehensively utilize the information entropy of health data to globally modulate the step size, and introduce a local gradient ratio modulation term to dynamically update the firefly movement step size α:

[0046]

[0047] where α (t) is the step size in the t-th generation, α (t+1) is the step size in the (t + 1)-th generation, β is the step size adjustment factor, θ is the local gradient modulation coefficient, denotes the norm of the position gradient of the firefly individual in the t-th generation, ||X (t) || denotes the norm of the position vector of the firefly individual in the t-th generation, ∈ is a small positive number to prevent division by zero;

[0048] S45. After reaching the predetermined maximum number of iterations T max calculate the health anomaly score S of each firefly individual according to the final light intensity values among the fireflies : i :

[0049]

[0050] where φ() is a function that maps the light intensity value to the anomaly score, N is the total number of the firefly population, Var(X i ) denotes the position variance of the firefly X i during the iteration process, ν is the stability modulation coefficient, and γ2 is the variance scaling factor.

[0051] Optionally, the S5 specifically includes:

[0052] S51. Extract the positions of each individual in the firefly population and the corresponding health anomaly scores where i = 1, 2, …, N, t represents the iteration generation, and N is the population size;

[0053] S52. Adopt the gene crossover technology and introduce an entropy modulation term to generate the position of a new individual, and calculate the position of the newly generated firefly individual:

[0054]

[0055] where, is the position of the newly generated firefly individual, and are respectively the positions of two randomly selected firefly individuals in the t-th generation, and η (t)is the gene crossover coefficient of the t-th generation, ξ is the crossover modulation coefficient, and H E is the entropy of health data, and σ() represents the activation function;

[0056] S53. Based on the mutation mechanism, combined with the entropy weight modulation and individual difference normalization techniques, calculate the mutation probability P of the current iteration m :

[0057]

[0058] where P m0 is the base mutation rate, λ m is the mutation regulation coefficient, τ is the inter-individual difference modulation coefficient, |||| represents the vector norm, ∈ is a small positive number to prevent division by zero, and exp() represents the exponential function;

[0059] S54. Adopt a selection and update mechanism based on feedback entropy modulation to select and update the firefly population, and update the positions of firefly individuals:

[0060]

[0061] where ε() is a function that maps the position of a firefly individual to a fitness value, reflecting the health anomaly score, is the feedback entropy factor, ψ is the feedback modulation coefficient, ∈ is a small positive number to prevent division by zero, represents the updated position of firefly individual i in the (t + 1)-th generation, is the candidate position of the new firefly individual i, represents the current position information of firefly individual i in the t-th generation;

[0062] S55. Comprehensively utilize the variance information of the population positions to adaptively update the gene crossover coefficient:

[0063]

[0064] where η (t+1) is the updated gene crossover coefficient in the (t + 1)-th generation, δ1 is the diversity modulation coefficient, represents the variance of the firefly positions in the t-th generation, is the mean norm of the firefly position vectors in the t-th generation.

[0065] Optionally, the specific content of S6 includes:

[0066] S61. From the temporal characterization H of the individual health status status and the individual health anomaly score vector S = [S1, S2,..., S N T ​Extract data related to health warnings, where a feature vector representing the individual's health change trend is extracted from the time-series representation of the individual's health status, and the abnormal score value corresponding to each firefly individual is extracted from the individual health abnormal score vector;

[0067] S62. Set the personalized health warning threshold T (t) with an initial value of T (0) , where t represents the iteration generation;

[0068] S63. Use the adaptive gradient correction method to iteratively adjust the warning threshold T:

[0069]

[0070] where α1 is the learning rate, α2 is the entropy modulation coefficient, H E is the health data information entropy, is the gradient of the warning loss function L with respect to the threshold T, T (t) represents the personalized health warning threshold at the t-th iteration, T (t+1) represents the personalized health warning threshold updated after the (t + 1)-th iteration;

[0071] S64. After reaching the predetermined maximum number of iterations T max , determine the stable threshold

[0072] S65. Use the dynamic feedback fusion mechanism to generate personalized health warning information Q:

[0073] Q = σ(μ1(S - T * ) + μ2ln(1 + H E ));

[0074] where μ1 is the score difference scaling coefficient, μ2 is the entropy feedback adjustment coefficient, ln() represents the natural logarithm function, and σ() represents the activation function.

[0075] The beneficial effects of the present invention are:

[0076] The present invention adopts multi-modal sensor data acquisition, deep learning and intelligent optimization algorithms to achieve comprehensive preprocessing, deep fusion and accurate analysis of smartwatch health data. By constructing a health status memory network and a hierarchical time-series feature cross-attention mechanism, the present invention can simultaneously capture the individual's short-term physiological fluctuations and long-term health trends, making the time-series representation of the health status more comprehensive and accurate. At the same time, an intelligent optimization technology combining an improved firefly algorithm and a genetic algorithm is used to dynamically adjust the search strategy and warning threshold, effectively reducing the noise interference and local optimum problems in the anomaly detection process, and ensuring the high accuracy and real-time performance of the anomaly score and health warning information.

[0077] Through the fine preprocessing and fusion of multimodal physiological data, the present invention significantly improves the effective utilization rate and anti-interference ability of the data, and then realizes the accurate modeling of the health status. Based on this, the system can timely capture the subtle changes in the user's health status and quickly trigger an alarm when the health data is abnormal, so as to provide personalized and dynamic health management suggestions for the user. This method not only improves the monitoring accuracy, but also enhances the system's adaptability to complex health environments, making the smart watch more reliable and secure in real-time health monitoring and early warning.

[0078] In addition, the present invention introduces the health data information entropy and local gradient modulation mechanism in the optimization search process, realizes the organic combination of global search and local optimization, and further improves the accuracy of health anomaly detection. By using the dynamic adaptive gradient correction method to adjust the warning threshold in real time, the present invention can automatically adjust the warning parameters according to the changes in the user's health status, so as to achieve personalized health warning. Overall, the present invention has significant advantages in improving the health monitoring accuracy of smart watches, reducing the risk of false alarms and missed reports, and realizing personalized health management, providing a more secure, accurate and efficient health management solution for users. Description of the Drawings

[0079] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0080] Figure 1 is a flowchart of the smart watch health monitoring management and warning method based on multimodal sensors proposed by the present invention;

[0081] Figure 2 is a schematic diagram of the personalized health warning generation module of the smart watch health monitoring management and warning method based on multimodal sensors proposed by the present invention. Detailed Embodiments

[0082] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0083] Refer to Figure 1 and Figure 2 , the smart watch health monitoring management and warning method based on multimodal sensors includes the following steps:

[0084] S1. Collect the multimodal physiological data of the smart watch, preprocess the multimodal physiological data, and generate a time-series health data set;

[0085] S2. Based on the time-series health dataset, construct a health state memory network, including a short-term health state module and a long-term health trend module, and combine the attention mechanism to output an individualized health state time-series representation;

[0086] S3. Adopt a hierarchical time-series feature cross-attention mechanism to fuse the features of the time-series health dataset and the individualized health state time-series representation to form a global health feature vector;

[0087] S4. Based on the firefly algorithm, perform health anomaly detection on the global health feature vector, calculate the individual health anomaly score, and by introducing a dynamic adaptive entropy-driven mechanism, dynamically correct the light intensity calculation and step size adjustment strategy using the information entropy of the health data;

[0088] S5. Use the genetic algorithm to optimize the search process of the firefly algorithm, generate new individuals by gene crossover, and adjust the mutation rate through adaptive entropy mutation. In each iteration, select the firefly individual with higher fitness to replace the individual with lower fitness;

[0089] S6. Based on the individualized health state time-series representation and the individual health anomaly score, dynamically adjust the personalized health warning threshold, and use the adaptive gradient correction method to iteratively adjust the warning parameters to generate personalized health warning information;

[0090] S7. When an individual health anomaly is detected, trigger a user prompt according to the generated personalized health warning information, and in the case of a serious anomaly, send a health alert message to a preset emergency contact or a medical service platform through the wireless communication module.

[0091] In this embodiment, the S2 specifically includes:

[0092] S21. Extract health data from the time-series health dataset to generate a standardized time-series health data sequence X = [x1, x2, …, x T , where x T represents the multimodal health data collected at the t-th time step, and T represents the total number of time steps;

[0093] S22. Construct a short-term health state module, and use a long short-term memory network to model the data within a short time window to generate a short-term health state representation H STM :

[0094] H STM = {h t = σ(W x x t + W h h t-1 + b) | t = T - T s + 1, …, T};

[0095] Among them, T s is the short-term window length, σ is the activation function, W x and W h are the input and hidden state weight matrices respectively, b is the bias, h t is the hidden state, and h t-1 represents the hidden state at the previous time step;

[0096] S23. Construct a long-term health trend module, use the Transformer structure to globally model the data within a long time window, and generate a long-term health trend representation H LTM :

[0097]

[0098] Among them, X long is the data within the long time window, T l is the long-term window length, d is the scaling factor, LayerNorm represents the layer normalization operation, and softmax represents the normalization;

[0099] S24. Use the attention mechanism to perform feature fusion on the short-term health state representation H STM and the long-term health trend representation H LTM to generate an individualized health state time series representation H status :

[0100]

[0101] Among them, W a is the attention weight matrix, H E is the information entropy of health data, λ is the entropy-driven adjustment coefficient, μ is the short-term feature weight adjustment parameter, and d is the scaling factor;

[0102] S25. Output the individualized health state time series representation H status .

[0103] In this embodiment, the specific content of S3 includes:

[0104] S31. The feature matrix output by the time series health data set is X and the individualized health state time series representation is H status , and an input feature matrix is generated through a concatenation operation:

[0105] F input = Concat(X, H status );

[0106] Among them, F input represents the input feature matrix, and Concat represents the concatenation operation;

[0107] S32. Calculate the local cross-attention on the input feature matrix F input to generate the low-level fusion feature F low ;

[0108] S33. Calculate the global cross-attention on the low-level fusion feature F low to generate the high-level fusion feature F high ;

[0109] S34. Use the hierarchical feature fusion mechanism to cross-fuse the low-level fusion feature F low and the high-level fusion feature F high to generate the fusion feature F fused :

[0110] F fused = LayerNorm(σ(κH E )⊙(αF low )+(1 - σ(κH E ))⊙((1 - α)F high ));

[0111] where σ() represents the activation function, H E is the entropy of health data, κ is the entropy adjustment factor, α is the low-level feature weight parameter, ⊙ represents element-wise multiplication, and LayerNorm represents the layer normalization operation;

[0112] S35. Generate the global health feature vector H fused by the residual connection with entropy modulation between the fusion feature F input and the input feature F global :

[0113] H global = LayerNorm(F input + γF fused + δ(F fused ⊙σ(κH E )));

[0114] where γ and δ are learnable weight coefficients.

[0115] In this embodiment, the S4 specifically includes:

[0116] S41. Use the output global health feature vector H global as the input of the improved firefly algorithm;

[0117] S42. Initialize the firefly population in the firefly algorithm, and the position of the firefly individual is where i = 1, 2, …, N, R is the set of real numbers, d ′For dimension, N is the total number of the firefly population, and the initial position is randomly sampled from the global health feature vector;

[0118] S43. In each iteration, by introducing the health data information entropy H E Drive and modulate the local gradient difference to update the light intensity value I i between any two fireflies X j and X ij :

[0119]

[0120] Among them, represents that in the t-th generation, the light intensity value of firefly X i to firefly X j , exp() represents the exponential function, γ1 is the light absorption coefficient, λ1 is the entropy adjustment coefficient, μ is the gradient modulation coefficient, represents the local gradient of the position of firefly X i in the t-th generation, represents the local gradient of the position of firefly X j in the t-th generation, represents the light intensity value of the updated firefly X i to firefly X j in the (t + 1)-th generation;

[0121] S44. Comprehensively utilize the health data information entropy to globally modulate the step size, and at the same time introduce the local gradient ratio modulation term to dynamically update the firefly movement step size α:

[0122]

[0123] Among them, α (t) is the step size in the t-th generation, α (t+1) is the step size in the (t + 1)-th generation, β is the step size adjustment factor, θ is the local gradient modulation coefficient, represents the norm of the position gradient of the firefly individual in the t-th generation, ||X (t) || represents the norm of the position vector of the firefly individual in the t-th generation, ∈ is a small positive number to prevent division by zero;

[0124] S45. After reaching the predetermined maximum number of iterations T max , calculate the health anomaly score S of each firefly individual according to the final light intensity value i between fireflies:

[0125]

[0126] Among them, φ() is a function that maps the light intensity value to the anomaly score, N is the total number of the firefly population, Var(Xi ) represents Firefly X i The position variance during the iteration process, ν is the stability modulation coefficient, and γ2 is the variance scaling factor.

[0127] In this embodiment, the S5 specifically includes:

[0128] S51. Extract the positions of each individual in the firefly population and the corresponding health abnormality scores where i = 1, 2, …, N, t represents the iteration generation, and N is the population size;

[0129] S52. Adopt the gene crossover technique and introduce the entropy modulation term to generate the positions of new individuals, and calculate the positions of the newly generated firefly individuals:

[0130]

[0131] where, is the position of the newly generated firefly individual, and are the positions of two randomly selected firefly individuals in the t-th generation, η (t) is the gene crossover coefficient in the t-th generation, ξ is the crossover modulation coefficient, H E is the entropy of health data information, and σ() represents the activation function;

[0132] S53. On the basis of the mutation mechanism, combine the entropy weight modulation and the individual difference normalization technique to calculate the mutation probability P of the current iteration m :

[0133]

[0134] where, P m0 is the basic mutation rate, λ m is the mutation adjustment coefficient, τ is the inter-individual difference modulation coefficient, |||| represents the vector norm, ∈ is a small positive number to prevent division by zero, and exp() represents the exponential function;

[0135] S54. Adopt the selection and update mechanism based on feedback entropy modulation to perform selection and update on the firefly population, and update the positions of firefly individuals:

[0136]

[0137] where, ε() is a function that maps the positions of firefly individuals to fitness values, reflecting the health abnormality scores, is the feedback entropy factor, ψ is the feedback modulation coefficient, ∈ is a small positive number to prevent division by zero, represents the updated position of firefly individual i in the (t + 1)-th generation, is the candidate position of the new firefly individual i, Denote the current position information of the i-th firefly individual in the t-th generation;

[0138] S55. Comprehensively utilize the variance information of the population position to adaptively update the gene crossover coefficient:

[0139]

[0140] where η (t+1) is the updated gene crossover coefficient in the (t + 1)-th generation, δ1 is the diversity modulation coefficient, represents the variance of the firefly positions in the t-th generation, is the mean norm of the firefly position vectors in the t-th generation.

[0141] In this embodiment, the specific steps of S6 are as follows:

[0142] S61. Extract data related to health warning from the time series representation H status of the individual health status and the individual health abnormality score vector S = [S1, S2,..., S N T Among them, extract the feature vector representing the individual health change trend from the time series representation of the individual health status, and extract the abnormality score value corresponding to each firefly individual from the individual health abnormality score vector;

[0143] S62. Set the initial value of the personalized health warning threshold T (t) as T (0) , where t represents the iteration generation;

[0144] S63. Use the adaptive gradient correction method to iteratively adjust the warning threshold T:

[0145]

[0146] where α1 is the learning rate, α2 is the entropy modulation coefficient, H E is the health data information entropy, is the gradient of the warning loss function L with respect to the threshold T, T (t) represents the personalized health warning threshold at the t-th iteration, and T (t+1) represents the personalized health warning threshold updated at the (t + 1)-th iteration;

[0147] S64. After reaching the predetermined maximum iteration number T max , determine the stable threshold

[0148] S65. Use the dynamic feedback fusion mechanism to generate the personalized health warning information Q:

[0149] Q = σ(μ1(S - T​* ) + μ2ln(1 + H E ));

[0150] Among them, μ1 is the scoring difference scaling coefficient, μ2 is the entropy feedback adjustment coefficient, ln() represents the natural logarithm function, and σ() represents the activation function.

[0151] Example 1:

[0152] To verify the feasibility of the present invention in implementation, the present invention is applied to the cardiovascular department of a certain tertiary hospital. 300 volunteers are selected for a health monitoring experiment. The volunteers cover patients with cardiovascular diseases, diabetes patients, and healthy individuals, with an age distribution between 20 - 75 years old. They all wear smartwatches supporting multimodal sensors for long-term health data collection and real-time monitoring. The experimental period is 90 days, covering multiple typical scenarios in daily life, including resting state, exercise state, mental stress state, abnormal health state, etc., to evaluate the health monitoring and early warning capabilities of the present invention.

[0153] In the experiment, the multimodal physiological data collected by the smartwatch includes signals such as heart rate, blood oxygen saturation, skin temperature, skin conductance, acceleration, etc. The system uses a health state memory network to model these data in the short term and long term, and extracts individualized health state time series representations through a hierarchical time series feature cross-attention mechanism. Based on this representation, the present invention uses an improved firefly-genetic algorithm to optimize the health anomaly detection process, and combines an adaptive gradient correction method to dynamically adjust the personalized health early warning threshold to ensure the accuracy and personalization of the early warning.

[0154] In the resting state (such as sleep monitoring) scenario, the system analyzed the overnight health data of 300 volunteers and detected nocturnal arrhythmia conditions. Traditional methods mainly detect anomalies based on fixed heart rate thresholds, with a high false alarm rate and poor adaptability to individual differences. The method of the present invention combines multimodal data analysis and individual health trend modeling, and can accurately identify nocturnal abnormal heart rhythm events. The detection rate is increased by 32.1% compared with traditional methods, and the false alarm rate is reduced by 27.5%.

[0155] In the exercise state, the system monitors physiological parameters such as heart rate, step frequency, and blood oxygen of volunteers during activities such as morning running, fast walking, and fitness. Traditional methods have a high false alarm rate due to the over-limit fluctuation of a single index, especially during strenuous exercise. The method of the present invention significantly reduces the false alarm rate through individualized health state modeling and dynamic adjustment of the health early warning threshold. The false alarm rate in the fast walking scenario is reduced from 18.4% to 7.2%, and the false alarm rate in the fitness scenario is reduced from 22.3% to 8.5%.

[0156] Under psychological stress, volunteers participated in stress scenarios such as high-intensity meetings and exams to test the system's monitoring ability of the impact of psychological stress on health. The experiment found that traditional methods are prone to false alarms of health abnormalities under high-pressure conditions, while the method of the present invention can comprehensively evaluate the actual impact of stress on individual health based on data such as skin conductance and heart rate variability, and reduce the false alarm rate through adaptive warning threshold adjustment. Experimental data shows that the false alarm rate has dropped from 21.5% to 9.3%.

[0157] In the detection scenario of abnormal health events, the system successfully detected and early warned 19 hypoglycemia events and 14 hypertension events. Compared with traditional methods, the present invention can detect hypoglycemia events 5.8 minutes earlier and hypertension emergencies 4.9 minutes earlier, significantly improving the response time to sudden health abnormalities. In addition, the results of the user satisfaction survey show that more than 90% of users recognize the health monitoring and early warning capabilities of the method of the present invention, especially in terms of the reduction of false alarm rate and the improvement of early detection ability.

[0158] Table 1 Comparison table of the health monitoring and early warning performance of smart watches

[0159]

[0160] The smart watch health monitoring and early warning method of the present invention shows significant advantages in all indicators. Experimental data shows that in the detection of nocturnal arrhythmia, the detection rate of the present invention reaches 89.5%, which is 32.1% higher than 67.4% of traditional methods, while the false alarm rate drops from 25.1% to 9.6%, and the reduction rate reaches 27.5%. This effect is mainly attributed to the health state memory network and hierarchical temporal feature cross-attention mechanism adopted by the present invention, enabling the system to capture both short-term fluctuations and long-term trends of individuals, thus more accurately reflecting the true physiological state.

[0161] In the detection of sudden hypoglycemia events, the detection rate of the present invention is 94.2%, higher than 70.3% of traditional methods. At the same time, the false alarm rate drops from 20.8% to 7.5%, greatly improving the accuracy of abnormal detection. The optimized strategy combining the improved firefly algorithm and genetic algorithm enables the system to accurately identify health abnormalities in complex multi-modal data and further reduce false alarms through dynamic warning threshold adjustment. Similarly, in the detection of sudden hypertension events, the detection rate of the present invention reaches 92.5%, and the false alarm rate drops from 19.2% of the traditional method to 8.3%. This shows that the method of the present invention has good robustness and accuracy in the detection of different health events.

[0162] Tests under motion states also prove the superiority of the present invention. In the scenario of fast walking, the false alarm rate of the traditional method is as high as 18.4%, while that of the present invention is only 7.2%. This fully shows that through multi-modal data fusion and personalized health state modeling, the present invention can effectively distinguish physiological fluctuations caused by exercise from real health abnormalities. Similarly, in the state of fitness exercise, the effects of improved detection rate and reduced false alarm rate are also very obvious, all showing the stable performance of the present invention in dealing with dynamic and complex health data.

[0163] In addition, the present invention also performs excellently in the aspect of early detection ability. In the detection of hypoglycemia events, the present invention detects abnormalities 5.8 minutes earlier on average, while for hypertension events, it is 4.9 minutes earlier. This greatly improves the response time in case of sudden health conditions and provides users with more opportunities for prevention and intervention. Generally speaking, the present invention not only has a significant improvement in the detection rate and false alarm rate compared with the traditional method, but also has great advantages in the early detection of health abnormal events, and can effectively improve the real-time performance and reliability of health monitoring.

[0164] To sum up, the present invention realizes accurate and real-time health abnormality detection and personalized early warning by deeply fusing multi-modal sensor data, constructing a health state memory network and a hierarchical temporal feature cross-attention mechanism, and combining an improved firefly-genetic algorithm optimization strategy. Experimental data fully prove that the present invention is superior to the existing methods in terms of health abnormality detection rate, false alarm rate control and early detection ability, greatly improving the application effect of smart watches in health management and user satisfaction. This achievement provides solid technical support for the popularization and application of smart watches in the field of health monitoring and lays a foundation for the development of future wearable health monitoring systems.

[0165] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered by the protection scope of the present invention.

Claims

1. An intelligent watch health monitoring, management and warning method based on multi-modal sensors, characterized in that, It includes the following steps: S1. Collect multi-modal physiological data of the smartwatch, preprocess the multi-modal physiological data, and generate a time-series health dataset; S2. Based on the time-series health dataset, construct a health state memory network, including a short-term health state module and a long-term health trend module, and combine the attention mechanism to output an individualized health state time-series representation; S3. Adopt a hierarchical time-series feature cross-attention mechanism to perform feature fusion on the time-series health dataset and the individualized health state time-series representation to form a global health feature vector; S4. Based on the firefly algorithm, perform health anomaly detection on the global health feature vector, calculate the individual health anomaly score, and by introducing a dynamic adaptive entropy-driven mechanism, dynamically correct the light intensity calculation and step adjustment strategy using the information entropy of the health data; S5. Use the genetic algorithm to optimize the search process of the firefly algorithm, generate new individuals through gene crossover, and adjust the mutation rate through adaptive entropy mutation. In each iteration, select the firefly individual with higher fitness to replace the individual with lower fitness; S6. Based on the individualized health state time-series representation and the individual health anomaly score, dynamically adjust the personalized health warning threshold, and use the adaptive gradient correction method to iteratively adjust the warning parameters to generate personalized health warning information; S7. When an individual health anomaly is detected, trigger a user prompt according to the generated personalized health warning information, and in the case of a serious anomaly, send a health alert message to a preset emergency contact or a medical service platform through the wireless communication module.

2. The intelligent watch health monitoring management and warning method based on a multi-modal sensor according to claim 1, wherein The specific content of S2 includes: S21. Extract health data from the time-series health dataset to generate a standardized time-series health data sequence X = [x1, x2, …, x T , where x T represents the multimodal health data collected at the t-th time step, and T represents the total number of time steps; S22. Construct a short-term health status module, use a long short-term memory network to model the data within a short time window, and generate a short-term health status representation H STM : H STM = {h t = σ(W x x t + W h h t-1 + b) | t = T - T s + 1, …, T}; Among them, T s is the short-term window length, σ is the activation function, W x and W h are the input and hidden state weight matrices respectively, b is the bias, h t is the hidden state, and h t-1 represents the hidden state at the previous time step; S23. Construct a long-term health trend module, use the Transformer structure to globally model the data within a long time window, and generate a long-term health trend representation H LTM : Among them, X long is the data within a long time window, T l is the length of the long time window, d is the scaling factor, LayerNorm represents the layer normalization operation, and softmax represents normalization; S24. Use the attention mechanism to perform feature fusion on the short-term health state representation H STM and the long-term health trend representation H LTM to generate an individualized health state time series representation H status : Among them, W a is the attention weight matrix, H E is the entropy of health data, λ is the entropy-driven adjustment coefficient, μ is the short-term feature weight adjustment parameter, and d is the scaling factor; S25. Output the time series representation H of the individualized health status status for output.

3. The intelligent watch health monitoring management and warning method based on a multi-modal sensor according to claim 1, characterized in that The specific content of S3 includes: S31. The feature matrix output by the time series health data set is X, and the time series representation of the individual health state is H status , and an input feature matrix is generated through a cascading operation: F input = Concat(X, H status ); Among them, F input represents the input feature matrix, and Concat represents the concatenation operation; S32. Apply local cross-attention calculation to the input feature matrix F input to generate a low-level fusion feature F low ; S33. Apply global cross-attention calculation to the low-level fusion feature F low to generate the high-level fusion feature F high ; S34. Use a hierarchical feature fusion mechanism to cross-fuse the low-level fused feature F low with the high-level fused feature F high to generate the fused feature F fused : F fused = LayerNorm(σ(κH E )) ⊙ (αF low )) + (1 - σ(κH E )) ⊙ ((1 - α)F high )); Among them, σ() represents the activation function, and H E is the entropy of health data, κ is the entropy adjustment factor, α is the low-level feature weight parameter, ⊙ represents element-wise multiplication, and LayerNorm represents the layer normalization operation; S35. Merge the fused feature F fused with the input feature F input to generate a global health feature vector H through a residual connection with entropy modulation global : H global = LayerNorm(F input + γF fused + δ(F fused ⊙ σ(κH E ))); Among them, γ and δ are learnable weight coefficients.

4. The intelligent watch health monitoring management and warning method based on multimodal sensors according to claim 1, characterized in that, The specific content of S4 includes: S41. Use the output global health feature vector H global as the input for improving the firefly algorithm; S42. Initialize the firefly population in the firefly algorithm, where the position of a firefly individual is where \(i = 1, 2, \ldots, N\), \(R\) is the set of real numbers, \(d\) ′ is the dimension, \(N\) is the total number of the firefly population, and the initial position is randomly sampled from the global health feature vector; S43. In each iteration, by introducing the entropy H of health data E Drive and modulate the local gradient difference to update the light intensity value I between any two fireflies X i and X j : ij : Among them, represents firefly X in the t-th generation i for firefly X j the light intensity value, exp() represents the exponential function, γ1 is the light absorption coefficient, λ1 is the entropy adjustment coefficient, and μ is the gradient modulation coefficient. represents the local gradient of the position of firefly X i in the t-th generation, represents the local gradient of the position of firefly X j in the t-th generation, represents the updated firefly X in the (t + 1)-th generation i for firefly X j the light intensity value; S44. Comprehensively use the information entropy of the health data to globally modulate the step size, and at the same time introduce a local gradient ratio modulation term to dynamically update the firefly movement step size α: where α (t) is the step size of the t-th generation, and α (t+1) is the step size of the (t + 1)-th generation, β is the step size adjustment factor, θ is the local gradient modulation coefficient, represents the norm of the position gradient of the firefly individual in the t-th generation, ‖X (t) ‖ represents the norm of the position vector of the firefly individual in the t-th generation, and ∈ is a small positive number to prevent division by zero; S45. After reaching the predetermined maximum number of iterations T max calculate the health anomaly score S of each firefly individual according to the final light intensity value among the fireflies i :​ Among them, φ() is a function that maps the light intensity value to an anomaly score, N is the total number of firefly populations, and Var(X i ) represents the position variance of firefly X i during the iterative process, ν is the stability modulation coefficient, and γ2 is the variance scaling factor.

5. The intelligent watch health monitoring management and warning method based on multi-modal sensors according to claim 1, characterized in that The specific content of S5 includes: S51. Extract the positions of individual fireflies in the firefly population and the corresponding health anomaly scores where i = 1, 2, …, N, t represents the iteration generation, and N is the population size; S52. Adopt gene crossover technology, and introduce an entropy modulation term to generate the position of a new individual, and calculate the position of the newly generated firefly individual: Among them, is the position of the newly generated firefly individual, and are the positions of two randomly selected firefly individuals in the t-th generation, respectively. η (t) is the gene crossover coefficient in the t-th generation, ξ is the crossover modulation coefficient, and H E is the entropy of health data information, and σ() represents the activation function; S53. Based on the mutation mechanism, the mutation probability P of the current iteration is calculated by combining entropy weight modulation and individual difference normalization techniques m : Among them, P m0 is the basic mutation rate, λ m is the mutation adjustment coefficient, τ is the individual difference modulation coefficient, ‖‖ represents the vector norm, ∈ is a small positive number to prevent division by zero, and exp() represents the exponential function; S54. Adopt a selection update mechanism based on feedback entropy modulation to perform selection update on the firefly population, and update the position of the firefly individual: where ε() is a function that maps the position of a firefly individual to a fitness value, reflecting the health anomaly score, is the feedback entropy factor, ψ is the feedback modulation coefficient, and ∈ is a small positive number to prevent division by zero, represents the updated position of firefly individual i in the (t + 1)-th generation, is the candidate position of the new firefly individual i, represents the current position information of firefly individual i in the t-th generation; S55. Comprehensively use the variance information of the population position to adaptively update the gene crossover coefficient: Among them, η (t+1) is the gene crossover coefficient updated in the (t + 1)-th generation, δ1 is the diversity modulation coefficient, represents the variance of the firefly positions in the t-th generation, is the mean norm of the firefly position vectors in the t-th generation.

6. The intelligent watch health monitoring management and warning method based on multi-modal sensors according to claim 1, characterized in that, The specific content of S6 includes: S61. Extract data related to health warnings from the time-series representation H of the individualized health status status and the individual health abnormality score vector S = [S1, S2, …, S N T Among them, extract the feature vector representing the individual health change trend from the time-series representation of the individualized health status, and extract the abnormality score value corresponding to each firefly individual from the individual health abnormality score vector;​ S62. Set the personalized health warning threshold T (t) with an initial value of T (0) , where t represents the iteration generation; S63. Use the adaptive gradient correction method to iteratively adjust the warning threshold T: Among them, α1 is the learning rate, α2 is the entropy modulation coefficient, and H E is the entropy of health data information, is the gradient of the warning loss function L with respect to the threshold T, and T (t) represents the personalized health warning threshold at the t-th iteration, and T (t+1) represents the personalized health warning threshold updated after the (t + 1)-th iteration; S64. After reaching the predetermined maximum number of iterations T max determine the stability threshold T * = T (Tmax) ; S65. Adopt a dynamic feedback fusion mechanism to generate personalized health warning information Q: Q = σ(μ1(S - T * ) + μ2ln(1 + H E )); Among them, μ1 is the scoring difference scaling coefficient, μ2 is the entropy feedback adjustment coefficient, ln() represents the natural logarithm function, and σ() represents the activation function.

Images