Blood, urine and body fluid collecting, detecting, analyzing and uploading system based on multi-sensor integration

Abstract

The invention discloses a blood, urine and body fluid collecting, detecting, analyzing and uploading system based on multi-sensor integration, which comprises the following modules: a blood collecting module for collecting a blood sample; the urine collection module is used for detecting urine; the body fluid collecting module is used for detecting body fluid; the data processing module is used for preprocessing data; the feature extraction and classification module is used for constructing a multi-modal feature matrix based on a neural structure search network, a support vector machine and a random forest algorithm; the anomaly detection module is used for generating a weighted anomaly score in combination with an isolated forest and a depth anomaly detection model; the time sequence analysis module is used for analyzing the health trend by using a variational auto-encoder-generative adversarial network; the health portrait and risk assessment module is used for constructing a health map by adopting a graph convolutional network; and the encryption and uploading management module is used for encrypting by using AES and RSA and managing the data access authority based on the intelligent contract. According to the invention, a multi-modal sensor and an intelligent algorithm are integrated, and collection, detection, analysis, encryption and uploading of blood, urine and body fluid are realized.

Description

Technical Field

[0001] The present invention relates to the technical fields of health monitoring and data processing, and particularly to a blood, urine and body fluid collection, detection, analysis and uploading system based on multi-sensor integration. Background Art

[0002] In the current fields of health monitoring and data management, with the improvement of people's health awareness, personalized and real-time health management systems have received extensive attention. Most existing health monitoring systems focus on the collection and analysis of single body fluid samples. For example, blood tests usually focus on indicators such as blood sugar and cholesterol, urine tests focus on the monitoring of protein and pH, while body fluid (such as sweat, saliva) monitoring is usually used for the detection of electrolytes and metabolites. These systems usually only monitor single body fluid data sources and cannot achieve comprehensive analysis of multiple body fluid indicators, resulting in great limitations in the assessment of an individual's health status. At the same time, traditional sensors or combinations of a small number of sensors are generally used in existing systems, making it difficult to simultaneously monitor multiple physiological indicators in different types of body fluids. In addition, the data collected by sensors has not undergone systematic multi-level processing, and there are still deficiencies in data noise removal, anomaly detection, and feature extraction. Especially when dealing with dynamic, time-series biological signals, the prior art cannot fully capture key health trends and potential risks.

[0003] In terms of data processing and anomaly detection, the deficiencies of the prior art are even more obvious. Most systems lack comprehensive preprocessing of the collected data. For example, there are obvious deficiencies in denoising, data smoothing, and missing value filling. Conventional data processing methods may lead to information loss or incomplete data. In addition, for the detection of outliers, most systems rely on traditional threshold setting and cannot dynamically and comprehensively detect and identify abnormal health signals. In the prior art, basic machine learning models such as isolation forest and support vector machine are usually used for anomaly detection. However, due to the complexity of biological signals, traditional machine learning methods are difficult to handle the complex characteristics and dynamic change trends of multi-modal body fluid data. In addition, when performing anomaly detection on multiple data sources, the prior art fails to fully consider the correlation between various biological signals, and often cannot form an accurate hierarchical warning mechanism, easily resulting in the omission or false alarm of high-risk health problems.

[0004] Existing health monitoring systems often lack effective time series analysis methods, especially systematic analysis means for short-term health trend prediction. Traditional time series models, such as autoregressive, moving average models, etc., often perform poorly in the dynamic analysis of health data and cannot fully utilize the long-term and short-term correlation features of biological signals. In addition, existing time series analysis methods based on generative adversarial networks (GANs) and autoencoders are relatively simple and cannot adapt to the complex dynamic features of multimodal health data. For short-term health trend prediction, existing technologies often struggle to balance accuracy and real-time performance, making it difficult for prediction models to be practically applied in health monitoring scenarios. Due to the lack of dynamic analysis of body fluid data, existing systems are difficult to provide users with real-time health status prediction and personalized health guidance, limiting their practical application value.

[0005] In terms of health profiling and risk assessment, existing health assessment models are mostly based on single biological signals and usually use simple methods such as logistic regression and decision trees, lacking the correlation analysis between multimodal data and unable to form a complete profile that truly reflects the user's health status. Since health data usually has complex internal correlations, for example, blood and urine indicators may reflect the same health risk, existing technologies are difficult to effectively construct a health correlation network of different biological signals and cannot conduct a unified risk assessment of multimodal health indicators. Algorithms such as Bayesian networks and graph convolutional networks have been applied to health management in recent years, but their use in the analysis of multimodal body fluid data is not yet mature. Existing risk assessment systems often can only achieve basic health risk scoring, lacking the ability of personalized dynamic adjustment and unable to provide users with a comprehensive and accurate health risk level.

[0006] In addition, the protection measures of existing health monitoring systems in data encryption and secure uploading are relatively single. Usually, simple symmetric encryption (such as AES) or asymmetric encryption (such as RSA) is used to process data, but it fails to combine multiple encryption methods to deal with data of different sensitivity levels, lacking flexible permission management and hierarchical data sharing functions. At the same time, due to the application of blockchain and smart contract technologies in health data management being still in its initial stage, existing technologies have not been able to achieve access control based on smart contracts and dynamic allocation of encryption policies. Existing encrypted uploading schemes also lack strategies for fragmenting and hierarchical management of data, resulting in insufficient security of highly sensitive data during transmission and storage. In addition, for high-frequency data transmission, existing technologies have not effectively optimized the efficiency of data transmission. Especially for the transmission requirements of highly sensitive data, conventional uploading methods are prone to transmission delays or insufficient priority allocation of data transmission, affecting the user experience.

[0007] Therefore, how to provide a blood and urine body fluid collection, detection, analysis and uploading system based on multi-sensor integration is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a blood, urine and body fluid collection, detection, analysis and upload system based on multi-sensor integration. The present invention integrates multi-modal sensors and advanced intelligent algorithms to realize the synchronous collection, analysis and dynamic management of multiple body fluids such as blood, urine and body fluid. Through technologies such as neural architecture search, anomaly detection models and time series analysis, high-risk health characteristics are accurately identified and real-time warnings are provided. In addition, personalized health portraits are generated through graph convolutional networks and Bayesian networks to ensure a comprehensive and accurate assessment of the health status. The system uses AES and RSA hierarchical encryption combined with smart contracts to control permissions, and realizes efficient and secure upload and sharing of data with a sharding and hierarchical upload strategy, significantly improving data security and privacy protection effects.

[0009] The blood, urine and body fluid collection, detection, analysis and upload system based on multi-sensor integration according to an embodiment of the present invention includes the following modules:

[0010] The blood collection module is used to collect blood samples through micro needles or micro blood collection devices, detect blood glucose, cholesterol and lactate indicators through electrochemical sensors, and perform blood component analysis in combination with near-infrared spectroscopy sensors, and transmit the data to the microcontroller unit and mark the time stamp and sample type;

[0011] The urine collection module is used to collect urine samples through a portable urine collector, detect urine pH, protein content and sugar indicators through a pH sensor, conductivity sensor, urine protein sensor and urine sugar sensor, and transmit the data to the microcontroller unit and mark the time stamp and sample type;

[0012] The body fluid collection module is used to collect body fluid samples through wearable sweat sensors or cotton swab saliva collection devices, detect sodium, potassium and chloride ion concentrations through electrolyte sensors, and analyze metabolites in body fluids through cortisol and glucose sensors, and transmit the data to the microcontroller unit and mark the time stamp and sample type;

[0013] The data processing module is used to smooth the collected data by the sliding window method, perform denoising processing using Kalman filtering, and complete the missing data items by the K-nearest neighbor algorithm;

[0014] The feature extraction and classification module is used to extract metabolic features of blood samples based on the neural architecture search network, and classify key indicators in urine and body fluid samples in combination with support vector machine and random forest algorithms, identify high-risk features, and form a multi-modal feature matrix;

[0015] The anomaly detection module is used to perform anomaly detection on the multi-modal feature matrix through the isolation forest model and the deep anomaly detection model, generate weighted anomaly scores, establish a hierarchical warning mechanism and real-time warning of high-risk data;

[0016] A time series analysis module for performing time series analysis through a variational autoencoder - generative adversarial network, predicting short - term health trends, and generating health status predictions;

[0017] A health portrait and risk assessment module for constructing a user's health portrait using a graph convolutional network, associating health indicators in blood, urine, and body fluids to form a health map, and scoring risk factors for multimodal health indicators based on a Bayesian network to generate a personalized health risk level;

[0018] An encryption and upload management module for encrypting sensing data and user - sensitive information using AES and RSA, managing data access rights through a smart contract, and splitting highly sensitive data into several small data fragments and uploading them in layers using a sharding strategy.

[0019] A method for blood, urine, and body fluid collection, detection, analysis, and upload based on multi - sensor integration according to an embodiment of the present invention includes the following steps:

[0020] S1. The blood collection module uses a microneedle or a micro - blood collection device to collect a blood sample, detects blood glucose, cholesterol, and lactate indicators through an electrochemical sensor, combines with a near - infrared spectroscopy sensor for blood component analysis, and transmits the data to a micro - control unit to mark the time stamp and sample type;

[0021] S2. The urine collection module uses a portable urine collector. The user injects the urine sample into the detection module, and pH sensors, conductivity sensors, urine protein, and urine glucose sensors detect urine acidity, protein content, and sugar indicators, and transmit the data to the micro - control unit to mark the time stamp and sample type;

[0022] S3. The body fluid collection module collects samples through a wearable sweat sensor or a cotton swab - type saliva collection device. Electrolyte sensors detect sodium, potassium, and chloride ion concentrations, and cortisol and glucose sensors analyze metabolites in the body fluid, and transmit the data to the micro - control unit to mark the time stamp and sample type;

[0023] S4. The micro - control unit smooths the data through a sliding window method, performs denoising processing using a Kalman filter, and fills in missing data items through a K - nearest neighbor algorithm;

[0024] S5. Extracts metabolic characteristics of the blood sample based on a neural architecture search network, classifies key indicators in urine and body fluids using a support vector machine and a random forest algorithm, identifies high - risk characteristics, and forms a multimodal feature matrix;

[0025] S6. Performs anomaly detection through an isolation forest model and a deep anomaly detection model, identifies abnormal changes in indicators, establishes a hierarchical early - warning mechanism, and real - time warns of high - risk data;

[0026] S7. Use variational autoencoder - generative adversarial network for time - series analysis, predict short - term health trends, and generate health status predictions;

[0027] S8. Use graph convolutional network to construct user health portraits, associate various health indicators of blood, urine, and body fluids to form a health atlas, and perform risk factor scoring on multi - modal health indicators based on Bayesian network to generate personalized health risk levels;

[0028] S9. Encrypt sensing data and user - sensitive information using AES and RSA, manage data access rights based on smart contracts, and achieve secure data upload and sharing through sharding and hierarchical upload strategies.

[0029] Optionally, the specific steps of S5 are as follows:

[0030] S51. Extract features from blood samples based on neural architecture search network, define the neural architecture search space, including different numbers of convolutional layers, numbers of neurons, and types of activation functions, optimize the model structure through a combination of greedy search and evolutionary algorithm, and automatically select the optimal architecture to adapt to the non - linear relationships of different metabolic features; the optimized model inputs blood glucose, cholesterol, and lactic acid metabolic feature data to generate a blood metabolic feature matrix X = [x1, x2, …, x n , where x i represents the feature vector of the i - th metabolite, and n represents the number of metabolic features;

[0031] S52. Through a dynamic weight allocation mechanism, adaptively adjust the hierarchical weights of the neural architecture search network, and perform weighted processing based on the differences and importance of each metabolic feature:

[0032]

[0033] where, x′ i represents the feature vector of the i - th metabolite after weighting, w i represents the weight dynamically allocated according to feature importance, and α, β, λ, δ, γ, η, κ, and θ represent adjustment parameters, and exp represents the exponential function;

[0034] S53. Classify the key risk features of urine samples based on support vector machine, select the non - linear kernel function K(x i , x j ) = <φ(x′ i ), φ(x′ j )>, where x j represents the feature vector of the j - th metabolite after weighting, φ represents the mapping function, and use the hyperplane of the support vector machine to construct a high - risk feature classification model;

[0035] S54. Use an improved random forest algorithm to perform multi-layer classification analysis on the electrolyte and metabolite characteristics of body fluid samples. By introducing feature sampling weights, high-risk features are preferentially introduced into the decision tree construction process to form a multi-level feature classification model, and the final classification voting is carried out through a forest composed of T weighted decision trees;

[0036] S55. After the extraction of blood, urine, and body fluid characteristics is completed, a multi-modal feature matrix is generated, and the multi-modal feature matrix is standardized and dimension-reduced in combination with data labels to form a multi-modal feature matrix.

[0037] Optionally, the specific content of S6 is as follows:

[0038] S61. Use an isolation forest model to perform preliminary anomaly detection on the multi-modal feature matrix of blood, urine, and body fluids. Construct isolation trees containing multi-dimensional features, calculate the anomaly scores of each sample point by randomly selecting split points multiple times to generate isolation path lengths, and the isolation forest model judges the degree of anomaly based on the path lengths of the sample points. Sample points with shorter paths indicate a higher possibility of anomaly;

[0039] S62. Further detect feature relationships based on a deep anomaly detection model. The deep anomaly detection model performs distribution feature modeling and reconstruction on the key indicators in the multi-modal feature matrix through hierarchical encoding of a multi-layer neural network, generates hidden layer features from the input multi-modal feature matrix through an encoder, and calculates the input and reconstruction errors as anomaly scores;

[0040] S63. Combine the anomaly scores of the isolation forest model and the reconstruction errors of the deep anomaly detection model to generate a weighted anomaly score for each sample point z i Calculate the final anomaly score:

[0041]

[0042] where S(z i ) represents the weighted anomaly score of sample point z i , Score(z i ) represents the anomaly score of sample point z i by the isolation forest model, Anomaly_Score(z i ) represents the anomaly score of sample z i by the deep anomaly detection model, α1 represents the weighting coefficient, β1 represents the smoothing parameter, δ1 represents the adjustment parameter, θ1 represents the smoothing adjustment coefficient, and γ1 represents the attenuation coefficient;

[0043] S64. For the weighted anomaly scores of each sample point, set a hierarchical warning strategy within different time duration ranges, judge the severity of the anomaly, and adjust the risk level according to the duration:

[0044]

[0045] Among them, θ L represents the low threshold, and θ H represents the high threshold, T 短期 represents the short-term time threshold, and T 突发 represents the burst time threshold, δ t , μ, η t , ω, γ t , ξ and τ represent adjustment parameters;

[0046] S65. When mild, moderate or severe anomalies are detected, trigger the corresponding feedback mechanism according to the warning level, notify the user of the abnormal changes in high-risk features in real time, and mark them as high priority when severe and emergency warning levels are detected;

[0047] S66. Through the dynamic adaptive mechanism, automatically update the weighted coefficients and warning thresholds of the isolation forest and the deep anomaly detection model according to the user's health feedback and new data.

[0048] Optionally, the S7 specifically includes:

[0049] S71. Convert the time series data of the multimodal feature matrix into a latent variable U through a variational autoencoder, retaining key time series information and potential health features;

[0050] S72. Reconstruct the latent variable U through the decoder of the variational autoencoder to generate reconstructed data R that is consistent with the distribution of the original time series data, and calculate the reconstruction loss:

[0051]

[0052] Among them, Loss recon represents the reconstruction loss, m represents the number of data points of the original time series data, q j represents the j-th original data point, r j represents the j-th reconstructed data point, α j represents the adaptive regularization coefficient, β j represents the nonlinear amplification coefficient, ξ j represents the exponential decay coefficient, λ j represents the weight for controlling abnormal features, γ s represents the smoothing coefficient, δ s represents the adjustment parameter, θ s represents the logarithmic adjustment coefficient;

[0053] S73. Input the latent variable U into the generator of the generative adversarial network, generate synthetic time series data V through the generator, and optimize using random noise input and adversarial training;

[0054] S74. The discriminator of the generative adversarial network discriminates between the generated time series data V and the original time series data. By learning the distribution difference between the real data and the generated data, the discriminator optimizes the output of the generator.

[0055] S75. Introduce a time dependence mechanism, construct the data generated by the variational autoencoder - generative adversarial network into a time series model, and use a long short - term memory network to predict the short - term health trend of key health indicators to generate a health status prediction; the time series model adopts multi - scale analysis to construct multi - level feature representations with different time windows for predicting the changing trend of health status over time.

[0056] Optionally, the specific steps of S8 are as follows:

[0057] S81. Construct a multi - modal health atlas through a graph convolutional network. Consider each health indicator in blood, urine, and body fluids as a node, and establish edges between nodes based on the correlation or causal relationship between health indicators to form a graph structure G=(V, E), where V represents the set of health indicator nodes and E represents the set of associated edges between nodes.

[0058] S82. In the health atlas, perform multi - layer graph convolutional operations on each health indicator node, and update the features of each node layer by layer by aggregating the information of neighboring nodes:

[0059]

[0060] where, H (l+1) represents the node feature matrix of the (l + 1)-th layer, σ represents the activation function, N(v) represents the set of neighboring nodes of node v, H (l) represents the node feature matrix of the l - th layer of node v, represents the node feature matrix of the l - th layer of neighboring node u, W (l) represents the weight matrix of the l - th layer, b (l) represents the bias value of the l - th layer, α h and β h represent adjustment coefficients, δ h represents the adjustment parameter of the exponential decay term;

[0061] S83. Integrate the node features after graph convolution into the health portrait model to form a multi - modal health status atlas. The atlas reflects the dynamic relationship of the user's health indicators through the recursive update and aggregation of the features of each node.

[0062] S84. Based on the health status map, use the Bayesian network to score the risk factors of health indicator nodes, associate each health node with the risk factor nodes in the Bayesian network, and evaluate the risk level of each health node by calculating the conditional probability to generate the conditional probability distribution of multimodal health indicators;

[0063] S85. Generate the risk factor scores of the multimodal health map through the comprehensive analysis of the Bayesian network:

[0064]

[0065] Among them, Risk_Factor represents the risk factor score, n represents the total number of health indicator nodes, v i represents the i-th health indicator node, Pa(v i ) represents the set of parent nodes of the i-th health indicator node, P(v i | Pa(v i )) represents the conditional probability, represents the value of the k-th risk factor node, K represents the total number of risk factor nodes, γ k , ρ, ξ k and η k represent the adjustment parameters, v j represents the j-th health indicator node, represents the expected health indicator value of the j-th health indicator node, ∈ represents the smoothing coefficient, ζ r represents the parameter for adjusting the logarithmic term;

[0066] S86. Based on the calculated risk factor scores, construct a personalized health risk level assessment matrix, generate the user's personalized health risk level, and form a dynamic health management plan.

[0067] Optionally, the specific content of S9 includes:

[0068] S91. Encrypt the sensing data and user sensitive information according to the sensitivity level. The low-sensitivity data is encrypted using AES symmetric encryption to generate the ciphertext C AES , and the high-sensitivity data is encrypted using RSA asymmetric encryption to generate the ciphertext C RSA , and the encrypted data is grouped according to the type and sensitivity level;

[0069] S92. Split the encrypted high-sensitivity data into several small data segments according to the sharding strategy, encrypt each data segment using a separate AES key, and mark a unique segment ID;

[0070] S93. Configure hierarchical upload priorities for the sharded data segments. The low-sensitivity data segments are uploaded to the cloud storage in batches according to the normal priority, and the high-sensitivity data segments are set to a high priority for upload according to real-time requirements and are preferentially processed in the transmission channel. The upload priority of each data segment is controlled by the upload delay threshold ΔT.

[0071] S94. Use smart contracts to manage the access rights of sharded data. The smart contract records the permission levels, access frequencies, and access rights of each user, and generates a dynamic permission control table P = [p ij , where p ij represents the access right of user i to shard j. The smart contract verifies the user's rights during data access and performs shard decryption and access control according to the hierarchical level of the data.

[0072] S95. During the data upload process, adopt a hierarchical transmission strategy for high-priority data segments, set the shard sequence parameter Ω i , sort the segments according to the urgency and importance of the data, and preferentially upload the more important segments in the transmission channel, while the low-priority segments will be delayed for processing.

[0073] S96. Implement the storage and sharing of sharded data through a distributed storage system in the cloud. After all the sharded encrypted data is uploaded, the access can be dynamically controlled by the smart contract. When a user accesses the data, the smart contract assigns the decryption key according to the permission level and gradually decrypts the sharded data to provide personalized data sharing services in a sharded and hierarchical manner.

[0074] The beneficial effects of the present invention are as follows:

[0075] First of all, through the multi-sensor integrated blood, urine, and body fluid collection, detection, analysis, and upload system of the present invention, the comprehensive, accurate, and dynamic monitoring and evaluation of users' health data are realized, overcoming the deficiencies of the prior art in multi-body fluid comprehensive analysis, real-time warning, personalized health management, and data security. By integrating a variety of high-precision sensors in the system, the synchronous collection of multi-body fluid data such as blood, urine, and body fluid is realized, providing a reliable data basis for the comprehensive analysis of users' health status. Combining advanced algorithms such as neural architecture search network, support vector machine, and random forest, the system can effectively extract the metabolic characteristics in blood, urine, and body fluid, form a multi-modal feature matrix, and accurately classify high-risk features. Compared with the traditional single data source analysis method, the introduction of the multi-modal feature matrix enables the system to have higher health index recognition accuracy and anomaly detection sensitivity, thus improving the accuracy and diversity of health data.

[0076] In addition, the present invention adopts an isolation forest and a deep anomaly detection model, which can dynamically detect outliers in multi-body fluid data and provide real-time warnings in combination with a hierarchical warning mechanism, enabling high-risk health states to be discovered and processed in a timely manner. Compared with traditional static threshold anomaly detection methods, the dynamic anomaly detection in the present invention can be adaptively adjusted based on the weights of different features, ensuring the accuracy of the detection results and reducing the possibility of false alarms or missed alarms. In terms of health trend prediction, the system innovatively uses a variational autoencoder-generative adversarial network for time series analysis, effectively capturing the short-term dynamic change trends of health states and generating personalized health state predictions. This method makes up for the deficiencies of time series prediction models in the prior art, enabling users to foresee potential health risks and intervene in advance.

[0077] The present invention also uses a graph convolutional network to construct a user health portrait, associates various health indicators by constructing a health map, and scores risk factors for multi-modal data based on a Bayesian network to generate a personalized health risk level. Different from existing single-feature risk scoring systems, this health portrait model uses multi-modal body fluid data and multi-dimensional correlation analysis, making the health risk assessment results more accurate and comprehensive, and being able to better reflect the user's health status and potential risks. Through the system's hierarchical risk assessment, users can obtain personalized health management suggestions, improving the pertinence and effectiveness of health management.

[0078] In terms of data security, the present invention implements a hierarchical encryption strategy combining AES and RSA to ensure that the encryption intensity of data with different sensitivities adapts to the user privacy protection requirements. The system finely manages data access permissions through smart contracts and combines a sharding and hierarchical uploading strategy to solve the deficiencies in data uploading security and transmission efficiency in the prior art. The sharding strategy splits highly sensitive data into multiple encrypted segments, effectively preventing the risk of data leakage, and the hierarchical uploading ensures the efficiency and security of high-priority data during transmission, significantly improving the security and real-time performance of the system during data transmission. BRIEF 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. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0080] Figure 1 is a schematic structural diagram of a blood, urine and body fluid collection, detection, analysis and upload system based on multi-sensor integration proposed by the present invention;

[0081] Figure 2 is the overall flowchart of a method for blood, urine and body fluid collection, detection, analysis and upload based on multi-sensor integration proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

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

[0083] Reference Figure 1 , a blood, urine, and body fluid collection, detection, analysis, and upload system based on multi-sensor integration, including the following modules:

[0084] A blood collection module for collecting a blood sample through a microneedle or a micro blood collection device, detecting blood glucose, cholesterol, and lactate indicators through an electrochemical sensor, and performing blood component analysis in combination with a near-infrared spectroscopy sensor, transmitting the data to a microcontroller unit and marking the timestamp and sample type;

[0085] A urine collection module for collecting a urine sample through a portable urine collector, detecting urine acidity, protein content, and sugar indicators through a pH sensor, a conductivity sensor, a urinary protein sensor, and a urine sugar sensor, and transmitting the data to a microcontroller unit and marking the timestamp and sample type;

[0086] A body fluid collection module for collecting a body fluid sample through a wearable sweat sensor or a cotton swab saliva collection device, detecting the concentrations of sodium, potassium, and chloride ions through an electrolyte sensor, and analyzing metabolites in the body fluid through a cortisol and glucose sensor, transmitting the data to a microcontroller unit and marking the timestamp and sample type;

[0087] A data processing module for smoothing the collected data through a sliding window method, performing denoising processing using a Kalman filter, and filling in missing data items through a K-nearest neighbor algorithm;

[0088] A feature extraction and classification module for extracting metabolic features of a blood sample based on a neural architecture search network, and classifying key indicators in urine and body fluid samples in combination with a support vector machine and a random forest algorithm, identifying high-risk features, and forming a multi-modal feature matrix;

[0089] An anomaly detection module for performing anomaly detection on the multi-modal feature matrix through an isolation forest model and a deep anomaly detection model, generating a weighted anomaly score, establishing a hierarchical early warning mechanism, and real-time warning of high-risk data;

[0090] A time series analysis module for performing time series analysis through a variational autoencoder-generative adversarial network, predicting short-term health trends, and generating a health status prediction;

[0091] Health Image and Risk Assessment Module, which is used to construct a user health image by using a graph convolutional network, associate health indicators in blood, urine and body fluids to form a health map, and score risk factors for multimodal health indicators based on a Bayesian network to generate a personalized health risk level;

[0092] Encryption and Upload Management Module, which is used to encrypt sensing data and user sensitive information with AES and RSA, manage data access rights through smart contracts, and split highly sensitive data into several small data segments and upload them in layers by using a sharding strategy.

[0093] Reference Figure 2 , A method for collecting, detecting, analyzing and uploading blood, urine and body fluids based on multi-sensor integration, including the following steps:

[0094] S1. The blood collection module uses a microneedle or a micro blood collection device to collect a blood sample, detects blood glucose, cholesterol and lactate indicators through an electrochemical sensor, combines a near-infrared spectroscopy sensor to analyze blood components, and transmits the data to the microcontroller unit to mark the timestamp and sample type;

[0095] S2. The urine collection module uses a portable urine collector. The user injects the urine sample into the detection module. The pH sensor, conductivity sensor, urinary protein and urinary sugar sensors detect the urine acidity, protein content and sugar indicators, and transmit the data to the microcontroller unit to mark the timestamp and sample type;

[0096] S3. The body fluid collection module collects samples through a wearable sweat sensor or a cotton swab type saliva collection device. The electrolyte sensor detects the concentrations of sodium, potassium and chloride ions, and the cortisol and glucose sensors analyze the metabolites in the body fluid, and transmit the data to the microcontroller unit to mark the timestamp and sample type;

[0097] S4. The microcontroller unit smooths the data by using the sliding window method, performs denoising processing by using the Kalman filter, and completes the missing data items by using the K-nearest neighbor algorithm;

[0098] S5. Extract the metabolic characteristics of the blood sample based on the neural architecture search network, combine the support vector machine and the random forest algorithm to classify the key indicators in urine and body fluids, identify high-risk characteristics, and form a multimodal feature matrix;

[0099] S6. Perform anomaly detection through the isolation forest model and the deep anomaly detection model, identify abnormal changes in indicators, establish a hierarchical early warning mechanism, and real-time warn of high-risk data;

[0100] S7. Use the variational autoencoder-generative adversarial network for time series analysis, predict short-term health trends, and generate health status predictions;

[0101] S8. Construct a user health profile using a graph convolutional network, associate various health indicators of blood, urine, and body fluids to form a health atlas, and perform risk factor scoring on multimodal health indicators based on a Bayesian network to generate a personalized health risk level;

[0102] S9. Encrypt the sensor data and user sensitive information using AES and RSA, manage data access permissions based on smart contracts, and achieve secure data upload and sharing through sharding and hierarchical upload strategies.

[0103] In this embodiment, the S5 specifically includes:

[0104] S51. Extract features from blood samples based on a neural architecture search network, define a neural architecture search space including different numbers of convolutional layers, neurons, and activation function types, optimize the model structure through a combination of greedy search and evolutionary algorithms, and automatically select the optimal architecture to adapt to the non-linear relationships of different metabolic features; the optimized model inputs blood glucose, cholesterol, and lactate metabolic feature data to generate a blood metabolic feature matrix X = [x1, x2,..., x n , where x i represents the feature vector of the i-th metabolite, and n represents the number of metabolic features;

[0105] S52. Through a dynamic weight allocation mechanism, adaptively adjust the hierarchical weights of the neural architecture search network, and perform weighted processing based on the differences and importance of each metabolic feature:

[0106]

[0107] where x′ i represents the feature vector of the i-th metabolite after weighting, w i represents the weight dynamically allocated according to feature importance, and α, β, λ, δ, γ, η, κ, and θ represent adjustment parameters, and exp represents the exponential function;

[0108] S53. Classify the key risk features of urine samples based on a support vector machine, select a non-linear kernel function K(x i , x j ) = <φ(x′ i ), φ(x′ j )>, where x j represents the feature vector of the j-th metabolite after weighting, φ represents the mapping function, and use the hyperplane of the support vector machine to construct a high-risk feature classification model;

[0109] S54. Use an improved random forest algorithm to perform multi-level classification analysis on the electrolyte and metabolite characteristics of body fluid samples. By introducing feature sampling weights, high-risk features are preferentially introduced into the decision tree construction process, thereby forming a multi-level feature classification model, and the final classification voting is carried out through a forest composed of T weighted decision trees;

[0110] S55. After the extraction of blood, urine and body fluid characteristics is completed, a multi-modal feature matrix is generated, and the multi-modal feature matrix is standardized and dimension-reduced in combination with data labels to form a multi-modal feature matrix.

[0111] In this embodiment, the specific content of S6 is as follows:

[0112] S61. Use an isolation forest model to perform preliminary anomaly detection on the multi-modal feature matrix of blood, urine and body fluids, construct isolation trees containing multi-dimensional features, generate isolation path lengths by randomly selecting split points multiple times to calculate the anomaly scores of each sample point, and the isolation forest model judges the degree of anomaly based on the path lengths of the sample points. Sample points with shorter paths indicate a higher possibility of anomaly;

[0113] S62. Further detect the feature relationship based on a deep anomaly detection model. The deep anomaly detection model performs distribution feature modeling and reconstruction on the key indicators in the multi-modal feature matrix through hierarchical encoding of a multi-layer neural network, generates hidden layer features from the input multi-modal feature matrix through an encoder, and calculates the input and reconstruction error as the anomaly score;

[0114] S63. Combine the anomaly score of the isolation forest model and the reconstruction error of the deep anomaly detection model to generate a weighted anomaly score for each sample point z i Calculate the final anomaly score:

[0115]

[0116] where S(z i ) represents the weighted anomaly score of sample point z i , Score(z i ) represents the anomaly score of sample point z i obtained by the isolation forest model, Anomaly_Score(z i ) represents the anomaly score of sample z i obtained by the deep anomaly detection model, α1 represents the weighting coefficient, β1 represents the smoothing parameter, δ1 represents the adjustment parameter, θ1 represents the smoothing adjustment coefficient, and γ1 represents the attenuation coefficient;

[0117] S64. For the weighted anomaly scores of each sample point, set a hierarchical warning strategy within different time duration ranges, judge the severity of the anomaly and adjust the risk level according to the duration:

[0118]

[0119] Among them, θ L represents the low threshold, and θ H represents the high threshold, T 短期 represents the short-term time threshold, and T 突发 represents the burst time threshold, δ t , μ, η t , ω, γ t , ξ and τ represent adjustment parameters;

[0120] S65. When mild, moderate or severe anomalies are detected, trigger the corresponding feedback mechanism according to the warning level, notify the user of the abnormal changes in high-risk features in real time, and mark them as high priority when severe and emergency warning levels are detected;

[0121] S66. Through the dynamic adaptive mechanism, automatically update the weighted coefficients and warning thresholds of the isolation forest and the deep anomaly detection model according to the user's health feedback and new data.

[0122] In this embodiment, the specific steps of S7 include:

[0123] S71. Convert the time series data of the multimodal feature matrix into a latent variable U through a variational autoencoder, retaining the key time series information and potential health features;

[0124] S72. Reconstruct the latent variable U through the decoder of the variational autoencoder to generate reconstructed data R with a distribution consistent with the original time series data, and calculate the reconstruction loss:

[0125]

[0126] Among them, Loss recon represents the reconstruction loss, m represents the number of data points of the original time series data, q j represents the j-th original data point, r j represents the j-th reconstructed data point, α j represents the adaptive regularization coefficient, β j represents the nonlinear amplification coefficient, ξ j represents the exponential decay coefficient, λ j represents the weight controlling abnormal features, γ s represents the smoothing coefficient, δ s represents the adjustment parameter, θ s represents the logarithmic adjustment coefficient;

[0127] S73. Input the latent variable U into the generator of the generative adversarial network, generate synthetic time series data V through the generator, and optimize it using random noise input and adversarial training;

[0128] S74. Use the discriminator of the generative adversarial network to discriminate between the generated time series data V and the original time series data. The discriminator optimizes the output of the generator by learning the distribution difference between real data and generated data;

[0129] S75. Introduce a time-dependence mechanism, construct the data generated by the variational autoencoder - generative adversarial network into a time series model, and use a long short-term memory network to predict the short-term health trend of key health indicators to generate a health status prediction; the time series model adopts multi-scale analysis to construct multi-level feature representations with different time windows for predicting the changing trend of health status over time.

[0130] In this embodiment, the specific steps of S8 are as follows:

[0131] S81. Construct a multi-modal health atlas through a graph convolutional network, regard each health indicator in blood, urine, and body fluids as a node, and establish edges between nodes based on the correlation or causal relationship between health indicators to form a graph structure G=(V, E), where V represents the set of health indicator nodes and E represents the set of associated edges between nodes;

[0132] S82. In the health atlas, perform multi-layer graph convolutional operations on each health indicator node, and update the features of each node layer by layer by aggregating neighbor node information:

[0133]

[0134] where, H (l+1) represents the node feature matrix of the (l + 1)-th layer, σ represents the activation function, N(v) represents the set of neighbor nodes of node v, H (l) represents the node feature matrix of the l-th layer of node v, represents the node feature matrix of the l-th layer of neighbor node u, W (l) represents the weight matrix of the l-th layer, b (l) represents the bias value of the l-th layer, α h and β h represent adjustment coefficients, δ h represents the adjustment parameter of the exponential decay term;

[0135] S83. Integrate the node features after graph convolution into the health portrait model to form a multi-modal health status atlas, and the atlas reflects the dynamic relationship of the user's health indicators through the recursive update and aggregation of each node feature;

[0136] S84. Based on the health status map, use the Bayesian network to score the risk factors of health indicator nodes, associate each health node with the risk factor nodes in the Bayesian network, and evaluate the risk level of each health node by calculating the conditional probability to generate the conditional probability distribution of multi-modal health indicators;

[0137] S85. Generate the risk factor score of the multi-modal health map through the comprehensive analysis of the Bayesian network:

[0138]

[0139] Among them, Risk_Factor represents the risk factor score, n represents the total number of health indicator nodes, v i represents the i-th health indicator node, Pa(v i ) represents the set of parent nodes of the i-th health indicator node, P(v i |Pa(v i )) represents the conditional probability, represents the value of the k-th risk factor node, K represents the total number of risk factor nodes, γ k , ρ, ξ k and η k represent adjustment parameters, v j represents the j-th health indicator node, represents the expected health indicator value of the j-th health indicator node, ∈ represents the smoothing coefficient, ζ r represents the parameter for adjusting the logarithmic term;

[0140] S86. Based on the calculated risk factor score, construct a personalized health risk level assessment matrix, generate the user's personalized health risk level, and form a dynamic health management plan.

[0141] In this embodiment, the specific steps of S9 are as follows:

[0142] S91. Encrypt the sensing data and user sensitive information according to the sensitivity level. The low-sensitivity data is encrypted using AES symmetric encryption to generate the ciphertext C AES , and the high-sensitivity data is encrypted using RSA asymmetric encryption to generate the ciphertext C RSA . The encrypted data is grouped according to the type and sensitivity level;

[0143] S92. Split the encrypted high-sensitivity data into several small data segments according to the sharding strategy. Each data segment is encrypted using a separate AES key and marked with a unique segment ID;

[0144] S93. Configure a hierarchical upload priority for the sharded data segments. The low-sensitivity data segments are uploaded to the cloud storage in batches according to the normal priority, and the high-sensitivity data segments are set to a high priority for upload according to real-time requirements and are preferentially processed in the transmission channel. The upload priority of each data segment is controlled by the upload delay threshold ΔT;

[0145] S94. Use a smart contract to manage the access rights of the sharded data. The smart contract records the permission level, access frequency, and access rights of each user, and generates a dynamic permission control table P = [p ij , where p ij represents the access right of user i to shard j. The smart contract verifies the user's permission during data access and performs shard decryption and access control according to the hierarchical level of the data;

[0146] S95. During the data upload process, adopt a hierarchical transmission strategy for the high-priority data segments, set the shard order parameter Ω i , sort the segments according to the urgency and importance of the data, and preferentially upload the more important segments in the transmission channel, while the low-priority segments will be delayed for processing;

[0147] S96. Implement the storage and sharing of the sharded data through a distributed storage system in the cloud. After all the sharded encrypted data is uploaded, the access can be dynamically controlled by the smart contract. When a user accesses the data, the smart contract allocates the decryption key according to the permission level and gradually decrypts the sharded data to provide personalized data sharing services in a sharded and hierarchical manner.

[0148] Example 1:

[0149] To verify the feasibility of the present invention in practice, the present invention is applied to the health monitoring of community elderly people. Through simple and convenient device operations, users or medical staff can collect blood, urine, and body fluid samples. For example, a blood sample can be collected through a microneedle blood collection device, and multiple detections such as blood glucose, cholesterol, and lactic acid can be performed. Combining with a near-infrared spectroscopy sensor to analyze the blood components, the data will be transmitted to the microcontroller unit in real time and marked with a timestamp and sample type. Further, urine is collected through a portable urine collector to detect the pH value, protein content, and sugar content; body fluids such as sweat are collected through a wearable sensor to analyze metabolic components such as sodium, potassium, and cortisol. The synchronous collection and transmission of these multi-body fluid indicators enable the multi-dimensional comprehensive analysis of health data.

[0150] In the data processing stage, the system first smooths the collected data to remove noise and outliers. At the same time, the K-nearest neighbor algorithm is used to complete some missing data items to ensure data integrity. This process effectively solves the problem that a single collection device cannot fully collect all health data. After data smoothing, the system inputs the metabolic characteristics of blood, urine, and body fluids into the feature extraction module, and automatically extracts key high-risk features through a neural architecture search network and a support vector machine algorithm. These high-risk features are classified and integrated to finally form a multi-modal feature matrix.

[0151] In terms of anomaly detection, this system combines the Isolation Forest and the Deep Anomaly Detection Model to automatically analyze the multi-modal feature matrix, identify outliers, and generate anomaly scores. The Isolation Forest model detects outliers in the multi-dimensional feature matrix, while the Deep Anomaly Detection Model further analyzes the correlation relationships between features to ensure more accurate judgment of outliers. This multi-model joint detection method solves the problem of insufficient sensitivity of a single detection model to the complex correlation relationships of multi-body fluid data. In the application scenario, if the system detects that the outliers of blood glucose and cortisol are close to the high-risk threshold, it will send a warning reminder to the user or management personnel, indicating that health intervention is needed.

[0152] To meet the needs of chronic disease patients for health trend monitoring, the system innovatively uses a Variational Autoencoder-Generative Adversarial Network for time series analysis of multi-modal data. Through time series analysis, the system can predict the short-term health trends of users, helping users or medical staff to identify potential health risks in a timely manner and carry out early intervention. In the actual scenario, the system can perform time series prediction for indicators such as high blood glucose and urine protein. If the prediction results show that the indicators will gradually rise to the risk threshold, it will automatically prompt the user to intervene to avoid the occurrence of a high-risk state.

[0153] In addition, the Health Portrait and Risk Assessment Module in the system can construct a health map based on the health indicators of blood, urine, and body fluids. The Graph Convolutional Network is used to analyze the correlation between multi-body fluid data, and the Bayesian Network constructs a multi-modal health risk factor score based on this to generate a precise personalized health risk level assessment for users. This method not only provides a more comprehensive assessment of an individual's health status but also identifies high-risk health characteristics of specific populations, helping medical staff to conduct hierarchical management of users according to the health risk level.

[0154] This system also adopts a multi-layer encryption mechanism that combines AES and RSA encryption to ensure data security. Before data upload, the system performs hierarchical encryption on sensing data and users' sensitive information, and manages data access rights through smart contracts during the upload process. The smart contract is responsible for managing the access rights of different users and medical staff. Highly sensitive data is uploaded through sharding and layering strategies to ensure privacy and security during data transmission, while improving the efficiency of data transmission. This encryption and sharded upload strategy solves the problems of transmission security and privacy protection for high-frequency data upload, providing reliable technical support for data security in home health monitoring.

[0155] This system can be widely applied in families, communities, and medical institutions, providing an integrated solution for the collection, anomaly detection, health prediction, and data security management of multi-body fluid health data, and realizing more comprehensive, accurate, and secure health management services. Through the practical application of this system, remarkable results have been achieved in the monitoring and management of chronic disease-related indicators such as blood glucose and urinary protein, effectively making up for the deficiencies of existing health monitoring systems.

[0156] Table 1 Experimental comparison data table of the existing technology and the present invention of the health monitoring system

[0157]

[0158] The above Table 1 compares the specific performance data of the health monitoring system under the existing technology and the present invention, showing the significant advantages of the present invention in various health monitoring indicators. First, in terms of collection efficiency, the present invention improves the hourly sample collection volume from 15 in the existing technology to 21 through multi-sensor integration, with an increase of 40%, effectively meeting the high-frequency monitoring requirements. Secondly, the coverage rate of health index analysis has increased from 60% to 95%, enabling users to obtain more comprehensive health information, especially the multi-index coverage ability in multi-modal body fluid detection, laying a foundation for the comprehensive analysis of health status.

[0159] In terms of data processing, the present invention uses technologies such as the sliding window method, Kalman filter, and K-nearest neighbor algorithm to improve data integrity to 98.5%, while there are significant defects in data processing in the existing technology, and the proportion of missing values is as high as 8%. This significant improvement reduces analysis errors, improves data reliability, and helps to obtain more accurate health assessment results. In terms of feature extraction and classification accuracy, the present invention adopts advanced models such as the neural architecture search network and random forest algorithm, increasing the feature extraction accuracy from 70% to 89%, reducing the interference of data noise, and greatly improving the reliability of detection results.

[0160] In terms of anomaly detection, the present invention combines the Isolation Forest and the deep anomaly detection model, and the sensitivity is significantly increased to 90%, an increase of 25% compared with 65% of the prior art, effectively solving the deficiency of the single anomaly detection model in sensitivity and reducing the possibility of false alarms and missed detections. Further, in terms of health trend prediction, the Variational Autoencoder - Generative Adversarial Network model of the present invention increases the prediction accuracy from 50% to 85%, enabling users to better grasp short - term health changes and achieve preventive intervention.

[0161] In terms of risk assessment, through multi - modal data analysis of the Graph Convolutional Network and the Bayesian Network, the present invention increases the number of indicators covered by risk assessment from 3 to 10. This expansion makes the health portrait more comprehensive, the risk assessment more personalized, and facilitates users to understand health risks. In terms of data transmission, the encryption and sharded upload strategy of the present invention reduces the data transmission delay from 3.2 seconds to 1.0 second, improves the real - time performance of data, and increases the data security score to 9, ensuring the effective protection of user privacy.

[0162] The above - mentioned 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, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A blood, urine and body fluid collection, detection, analysis and upload system based on multi-sensor integration, characterized in that, It includes the following modules: A blood collection module, which is used to collect blood samples through microneedles or micro blood collection devices, detect blood glucose, cholesterol, and lactate indicators through electrochemical sensors, and perform blood component analysis in combination with near-infrared spectroscopy sensors, and transmit the data to the microcontroller unit and mark the timestamp and sample type; A urine collection module, which is used to collect urine samples through a portable urine collector, detect urine pH, protein content, and sugar indicators through a pH sensor, conductivity sensor, urine protein sensor, and urine sugar sensor, and transmit the data to the microcontroller unit and mark the timestamp and sample type; A body fluid collection module, which is used to collect body fluid samples through wearable sweat sensors or cotton swab saliva collection devices, detect the concentrations of sodium, potassium, and chloride ions through electrolyte sensors, and analyze metabolites in body fluids through cortisol and glucose sensors, and transmit the data to the microcontroller unit and mark the timestamp and sample type; A data processing module, which is used to smooth the collected data by the sliding window method, perform denoising processing using Kalman filtering, and complete the missing data items through the K-nearest neighbor algorithm; A feature extraction and classification module, which is used to extract metabolic features of blood samples based on a neural architecture search network, and classify key indicators in urine and body fluid samples in combination with support vector machine and random forest algorithms, identify high-risk features, and form a multi-modal feature matrix; An anomaly detection module, which is used to perform anomaly detection on the multi-modal feature matrix through an isolation forest model and a deep anomaly detection model, generate weighted anomaly scores, establish a hierarchical early warning mechanism and real-time early warning of high-risk data; A time series analysis module, which is used to perform time series analysis through a variational autoencoder-generative adversarial network, predict short-term health trends, and generate health status predictions; A health portrait and risk assessment module, which is used to construct a user health portrait using a graph convolutional network, associate health indicators in blood, urine, and body fluids to form a health map, and perform risk factor scoring on multi-modal health indicators based on a Bayesian network to generate a personalized health risk level; An encryption and upload management module, which is used to encrypt sensing data and user sensitive information using AES and RSA, manage data access rights through smart contracts, and use a sharding strategy to split highly sensitive data into several small data segments and upload them in layers.

2. The blood, urine and body fluid collection, detection, analysis and upload system based on multi-sensor integration according to claim 1 performs a blood, urine and body fluid collection, detection, analysis and upload method based on multi-sensor integration, characterized in that, It includes the following steps: S1. The blood collection module uses a microneedle or a micro blood collection device to collect a blood sample, detects blood glucose, cholesterol, and lactate indicators through an electrochemical sensor, performs blood component analysis in combination with a near-infrared spectroscopy sensor, and transmits the data to the microcontroller unit to mark the timestamp and sample type; S2. The urine collection module uses a portable urine collector. The user injects the urine sample into the detection module. The pH sensor, conductivity sensor, urine protein, and urine sugar sensors detect urine pH, protein content, and sugar indicators, and transmit the data to the microcontroller unit to mark the timestamp and sample type; S3. The body fluid collection module collects samples through a wearable sweat sensor or a cotton swab saliva collection device. The electrolyte sensor detects the concentrations of sodium, potassium, and chloride ions, and the cortisol and glucose sensors analyze the metabolites in the body fluid and transmit the data to the microcontroller unit to mark the timestamp and sample type; S4. The microcontroller unit smooths the data through the sliding window method, performs denoising processing using Kalman filtering, and completes the missing data items through the K-nearest neighbor algorithm; S5. Based on the neural architecture search network, extract the metabolic characteristics of the blood sample, combine the support vector machine and the random forest algorithm to classify the key indicators in the urine and body fluid, identify the high-risk characteristics, and form a multi-modal feature matrix; S6. Perform anomaly detection through the isolation forest model and the deep anomaly detection model, identify abnormal changes in the indicators, establish a hierarchical early warning mechanism, and give real-time early warning of high-risk data; S7. Use the variational autoencoder-generative adversarial network for time series analysis, predict the short-term health trend, and generate a health status prediction; S8. Use the graph convolutional network to construct a user health profile, associate the health indicators of blood, urine, and body fluid to form a health map, and perform risk factor scoring on the multi-modal health indicators based on the Bayesian network to generate a personalized health risk level; S9. Encrypt the sensing data and user sensitive information using AES and RSA, manage the data access rights based on the smart contract, and achieve secure data upload and sharing through the sharding and hierarchical upload strategy.

3. The method for collecting, detecting, analyzing and uploading blood, urine and body fluids based on multi-sensor integration according to claim 2, wherein, The specific content of S5 includes: S51. Feature extraction is performed on blood samples based on a neural architecture search network. A neural architecture search space is defined, including different numbers of convolutional layers, numbers of neurons, and types of activation functions. The model structure is optimized through a combination of greedy search and evolutionary algorithms, and the optimal architecture is automatically selected to adapt to the non-linear relationships of different metabolic features. The optimized model inputs blood glucose, cholesterol, and lactic acid metabolic feature data to generate a blood metabolic feature matrix X = [x1, x2, …, x n , where x i represents the feature vector of the i-th metabolite, and n represents the number of metabolic features; S52. Through the dynamic weight allocation mechanism, adaptively adjust the hierarchical weights of the neural architecture search network, and perform weighted processing based on the differences and importance of each metabolic characteristic: where, x′ i represents the eigenvector of the i-th metabolite after weighting, w i represents the weight dynamically assigned according to the feature importance, α, β, λ, δ, γ, η, κ, and θ represent the adjustment parameters, and exp represents the exponential function; S53. Classify the key risk features of urine samples based on support vector machines, and select the non-linear kernel function K(x i , x j ) = <φ(x′ i ), φ(x′ j )>, where x j represents the feature vector of the j-th metabolite after weighting, φ represents the mapping function, and construct a high-risk feature classification model using the hyperplane of the support vector machine; S54. Use the improved random forest algorithm to perform multi-layer classification analysis on the electrolyte and metabolite characteristics of the body fluid samples. By introducing the feature sampling weight, the high-risk characteristics are preferentially introduced into the decision tree construction process, thereby forming a multi-level feature classification model, and the final classification voting is performed through the forest composed of T weighted decision trees; S55. After the characteristics of blood, urine, and body fluid are extracted, a multi-modal feature matrix is generated, and the multi-modal feature matrix is standardized and dimension-reduced in combination with the data label to form a multi-modal feature matrix.

4. The method for collecting, detecting, analyzing and uploading blood, urine and body fluids based on multi-sensor integration according to claim 2, wherein The specific content of S6 includes: S61. Use the isolation forest model to perform preliminary anomaly detection on the multi-modal feature matrix of blood, urine, and body fluid, construct isolation trees containing multi-dimensional features, calculate the anomaly scores of each sample point by randomly selecting split points multiple times to generate the isolation path length, and the isolation forest model judges the degree of anomaly based on the path length of the sample point. The sample points with shorter paths indicate a higher possibility of anomaly; S62. Further detect the feature relationship based on the deep anomaly detection model. The deep anomaly detection model performs distribution feature modeling and reconstruction on the key indicators in the multi-modal feature matrix through the hierarchical coding of the multi-layer neural network, generates the hidden layer features from the input multi-modal feature matrix through the encoder, and calculates the input and reconstruction error as the anomaly score; S63. Generate a weighted anomaly score by combining the anomaly score of the Isolation Forest model and the reconstruction error of the Deep Anomaly Detection model for each sample point z i Calculate the final anomaly score: Among them, S(z i ) represents the weighted anomaly score of the sample point z i , Score(z i ) represents the anomaly score of the sample point z i by the isolation forest model, Anomaly_Score(z i ) represents the anomaly score of the sample z i by the deep anomaly detection model, α1 represents the weighting coefficient, β1 represents the smoothing parameter, δ1 represents the adjustment parameter, θ1 represents the smoothing adjustment coefficient, and γ1 represents the attenuation coefficient; S64. For the weighted anomaly scores of each sample point, set a hierarchical warning strategy within different time durations, judge the severity of the anomaly, and adjust the risk level according to the duration: where, θ L represents a low threshold, θ H represents a high threshold, T 短期 represents a short-term time threshold, T 突发 represents a burst time threshold, δ t , μ, η t , ω, γ t , ξ and τ represent adjustment parameters; S65. When mild, moderate, or severe anomalies are detected, trigger the corresponding feedback mechanism according to the warning level, notify the user in real time of the abnormal changes in high-risk features, and mark them as high-priority when severe and emergency warning levels are detected; S66. Through a dynamic adaptive mechanism, automatically update the weighted coefficients and warning thresholds of the isolation forest and deep anomaly detection models according to the user's health feedback and new data.

5. The method for collecting, detecting, analyzing and uploading blood, urine and body fluid based on multi-sensor integration according to claim 2, wherein The specific content of S7 includes: S71. Convert the time series data of the multi-modal feature matrix into a latent variable U through a variational autoencoder, retaining the key time series information and potential health features; S72. Reconstruct the latent variable U through the decoder of the variational autoencoder to generate reconstructed data R that is consistent with the distribution of the original time series data, and calculate the reconstruction loss; Among them, Loss recon represents the reconstruction loss, m represents the number of data points of the original time series data, q j represents the j-th original data point, r j represents the j-th reconstructed data point, α j represents the adaptive regularization coefficient, β j represents the nonlinear amplification coefficient, ξ j represents the exponential decay coefficient, λ j represents the weight for controlling abnormal features, γ s represents the smoothing coefficient, δ s represents the adjustment parameter, θ s represents the logarithmic adjustment coefficient; S73. Input the latent variable U into the generator of the generative adversarial network, generate synthetic time series data V through the generator, and optimize it using random noise input and adversarial training; S74. Discriminate the generated time series data V and the original time series data through the discriminator of the generative adversarial network. The discriminator optimizes the output of the generator by learning the distribution difference between the real data and the generated data; S75. Introduce a time dependence mechanism, construct the data generated by the variational autoencoder-generative adversarial network into a time series model, and use a long short-term memory network to predict the short-term health trend of key health indicators to generate a health status prediction; the time series model uses multi-scale analysis to construct multi-level feature representations with different time windows for predicting the change trend of the health status over time.

6. The method for collecting, detecting, analyzing and uploading blood, urine and body fluid based on multi-sensor integration according to claim 2, wherein, The specific content of S8 includes: S81. Construct a multi-modal health map through a graph convolutional network, regard each health indicator in blood, urine, and body fluids as a node, and establish edges between nodes based on the correlation or causal relationship between health indicators to form a graph structure G=(V, E), where V represents the set of health indicator nodes and E represents the set of associated edges between nodes; S82. In the health map, perform multi-layer graph convolutional operations on each health indicator node, and update the features of each node layer by layer by aggregating the information of neighboring nodes; Among them, H (l+1) represents the node feature matrix of the (l + 1)-th layer, σ represents the activation function, N(v) represents the set of neighbor nodes of node v, and H (l) represents the node feature matrix of node v at the l-th layer, represents the node feature matrix of neighbor node u at the l-th layer, W (l) represents the weight matrix of the l-th layer, b (l) represents the bias value of the l-th layer, α h and β h represent the adjustment coefficients, and δ h represents the adjustment parameter of the exponential decay term; S83. Integrate the node features after graph convolution into the health portrait model to form a multi-modal health status map, and the map reflects the dynamic relationship of the user's health indicators through the recursive update and aggregation of the features of each node; S84. Based on the health status map, use a Bayesian network to score the risk factors of health indicator nodes, associate each health node with the risk factor nodes in the Bayesian network, and evaluate the risk level of each health node by calculating the conditional probability to generate the conditional probability distribution of multi-modal health indicators; S85. Generate the risk factor score of the multi-modal health map through the comprehensive analysis of the Bayesian network; Among them, Risk_Factor represents the risk factor score, n represents the total number of health index nodes, v i represents the i-th health index node, Pa(v i ) represents the set of parent nodes of the i-th health index node, P(v i |Pa(v i )) represents the conditional probability, represents the value of the k-th risk factor node, K represents the total number of risk factor nodes, γ k , ρ, ξ k and η k represent the adjustment parameters, v j represents the j-th health index node, represents the expected health index value of the j-th health index node, ∈ represents the smoothing coefficient, ζ r represents the parameter for adjusting the logarithmic term; S86. Based on the calculated risk factor score, construct a personalized health risk level assessment matrix, generate the user's personalized health risk level, and form a dynamic health management plan.

7. The method for collecting, detecting, analyzing and uploading blood, urine and body fluid based on multi-sensor integration according to claim 2, characterized in that, The specific steps of S9 are as follows: S91. Encrypt the sensing data and user sensitive information according to the sensitivity level. For low-sensitivity data, use AES symmetric encryption to generate ciphertext C AES , and for high-sensitivity data, use RSA asymmetric encryption to generate ciphertext C RSA . After encryption, group the data by type and sensitivity level; S92: Split the encrypted high-sensitivity data into several small data segments according to the sharding strategy. Each data segment is encrypted with a separate AES key and marked with a unique segment ID. S93: Configure a hierarchical upload priority for the sharded data segments. The low-sensitivity data segments are uploaded to the cloud storage in batches according to the normal priority. The high-sensitivity data segments are set to a high priority for upload according to real-time requirements and are preferentially processed in the transmission channel. The upload priority of each data segment is controlled by the upload delay threshold ΔT. S94. Use a smart contract to manage the access rights of sharded data. The smart contract records the permission level, access frequency, and access rights of each user, and generates a dynamic permission control table P = [p ij , where p ij represents the access right of user i to shard j. The smart contract verifies the user's permission during data access and performs shard decryption and access control according to the hierarchical level of the data; S95. During the data upload process, a hierarchical transmission strategy is adopted for high-priority data segments, and a sharding order parameter Ω is set. i Segments are sorted according to the urgency and importance of the data, and the more important segments are preferentially uploaded in the transmission channel, while the low-priority segments will be delayed for processing. S96: Implement the storage and sharing of sharded data through a distributed storage system in the cloud. After all the sharded encrypted data is uploaded, it can be dynamically accessed by a smart contract. When a user accesses the data, the smart contract allocates a decryption key according to the permission level and decrypts the sharded data step by step, providing a personalized data sharing service in a sharded and hierarchical manner.

Images