CN-122025161-A - Urinary data management method combined with deep learning

CN122025161ACN 122025161 ACN122025161 ACN 122025161ACN-122025161-A

Abstract

The invention provides a urinary data management method combining deep learning in the technical field of intersection of artificial intelligence and medical Internet of things, which comprises the steps of S1, inputting collected real-time urinary data into a data cleaning model by a user terminal for cleaning, then desensitizing, compressing to obtain a compressed data packet, encrypting and uploading the compressed data packet to a server, S2, storing the encrypted data packet to a user-specific data container by the server, periodically reading the urinary desensitization data, inputting a data drift detection model and a collection strategy generation model, sending the obtained drift parameters and the collection parameters to the user terminal, compensating and correcting the collected real-time sensor monitoring data, and adjusting the data collection strategy of a sensor array, and S3, accessing the server by the user terminal through a small program for online management of the user-specific data container. The invention has the advantages of improving the data transmission efficiency, the data quality and the overall intelligent level of the system.

Inventors

BAI PEIDE
CHEN BIN
WANG TAO
GONG HE
XIE YIJIE
ZHENG QI
GUAN BING
SHI ZHIYUAN

Assignees

厦门大学附属第一医院(厦门市第一医院、厦门市红十字医院、厦门市糖尿病研究所)

Dates

Publication Date: 20260512
Application Date: 20251219

Claims (10)

1. A urinary data management method combined with deep learning is characterized by comprising the following steps: step S1, a user terminal sends a registration request to a server through a small program so as to execute registration operation comprising registration information input and sensor array binding, after registration is completed, the server negotiates a dynamic session key with the user terminal, and the server distributes a user exclusive data container based on the registration information; Step S2, the user terminal collects real-time sensor monitoring data through a bound sensor array, collects input real-time life logs through a small program, and inputs real-time urinary data comprising the real-time sensor monitoring data and the real-time life logs into a pre-deployed data cleaning model to obtain real-time optimized urinary data; Step S3, the user terminal performs data desensitization based on differential privacy on the real-time optimized urinary data to obtain urinary desensitization data, and performs data compression on the urinary desensitization data to obtain a compressed data packet; S4, the user terminal encrypts the compressed data packet based on the dynamic session key to obtain an encrypted data packet, and uploads the encrypted data packet to a server; s5, the server stores the received encrypted data packet in a user exclusive data container in real time, and records an access log of the user exclusive data container for security audit and abnormal access detection in real time; step S6, the server periodically reads an encrypted data packet from the user exclusive data container, decrypts and decompresses the encrypted data packet to obtain urinary desensitization data, and respectively inputs the urinary desensitization data into a pre-deployed data drift detection model and an acquisition strategy generation model to respectively obtain drift parameters and acquisition parameters; Step S7, the server encrypts and transmits the drift parameters and the acquisition parameters to the user terminal based on the dynamic session key so as to compensate and correct the subsequently acquired real-time sensor monitoring data and adjust the data acquisition strategy of the sensor array; and S8, the user terminal accesses the server through the small program, and performs online management on the exclusive data container of the user by combining the access log.
2. The urinary data management method of claim 1, wherein said step S1 is specifically: After the user terminal accesses the applet based on the two-dimension code, the program discovery portal or the program address, the user interface of the applet acquires the recorded registration information, and then guides the user to call the wireless communication module of the user terminal to establish connection with the sensor array, reads the equipment information of the sensor array, binds the registration information with the equipment information to obtain registration data, encrypts the registration data into registration ciphertext data, generates a registration request based on the registration ciphertext data, and sends the registration request to the server based on a TLS protocol, wherein the registration information at least comprises account name, password, name, gender, age, ethnicity, residence city, birth date, mobile phone number, height, weight and disease information; The server analyzes the received registration request to obtain registration ciphertext data, decrypts and verifies the registration ciphertext data to obtain registration data, generates a unique registration serial number after performing registration verification on the registration data, and stores the registration serial number and the registration data in a registration user management table of the hardware security module in an associated manner so as to execute registration operation; after registration is completed, the server negotiates dynamic session keys with the user terminal based on ECDH algorithm, and the server allocates a user-specific data container based on the registration information.
3. The urinary data management method in combination with deep learning of claim 1, wherein in the step S1, the server negotiates a dynamic session key with the user terminal specifically: the server creates a pair of private keys S_priv and public keys S_pub based on an ECDH algorithm to obtain a local device serial number SN_S, calculates a check code JY_S of the public keys S_pub and the device serial number SN_S through a BLAKE3 algorithm, and creates a pair of private keys U_priv and public keys U_pub based on the ECDH algorithm to obtain a local device serial number SN_U, calculates a check code JY_U of the public keys U_pub and the device serial number SN_U through the BLAKE3 algorithm, and exchanges the public keys S_pub, the device serial number SN_S, the check code JY_S, the public keys U_pub, the device serial number SN_U and the check code JY_U with the user terminal based on a TLS protocol; The server checks the public key U_pub and the equipment serial number SN_U based on the received check code JY_U, and the user terminal checks the public key S_pub and the equipment serial number SN_S based on the received check code JY_S; After verification, the server calculates a shared key through elliptic curve point multiplication operation based on the private key S_priv and the public key U_pub, and stores the shared key into a hardware security module of the server; the user terminal calculates a shared secret key through elliptic curve point multiplication operation based on the private key U_priv and the public key S_pub, and stores the shared secret key into a hardware security module of the user terminal; The server and the user terminal acquire the current token time stamp, round the token time stamp by taking half an hour as key time to obtain a time token, and derive a dynamic session key based on the shared key, the equipment serial number SN_S, the equipment serial number SN_U and the time token through HKDF functions; and the server and the user terminal execute time synchronization operation based on a preset time synchronization period through an NTP time synchronization protocol.
4. The urinary data management method combined with deep learning as claimed in claim 1, wherein in the step S2, the data cleaning model is constructed based on a noise filtering module, a missing value complementing module, an abnormality detecting and correcting module, a consistency verifying and optimizing module, a multi-source data fusion module and an output optimizing module which are sequentially connected, and the data cleaning model is used for sequentially carrying out noise reduction, missing value processing, outlier processing, inconsistent data correction, time synchronization, data alignment and preprocessing of feature extraction on input initial urinary data and outputting structured optimized urinary data; The data cleaning model is trained through marked historical urinary data in advance, is compressed by combining a quantization technology and a knowledge distillation technology before deployment, and performs performance verification on the compressed data cleaning model so as to adapt to the computing resources of a user terminal.
5. The urinary data management method of claim 1, wherein said step S3 is specifically: The user terminal desensitizes the real-time optimized urinary data by adopting a localized differential privacy technology based on a preset privacy budget epsilon and a relaxation parameter delta, wherein noise sampled from Laplacian distribution is added to a numerical field in the data, and the classified data is disturbed by adopting a random response technology so as to ensure that the output urinary desensitized data meets (epsilon, delta) -differential privacy constraint; And then, according to the characteristic of discrete and continuous mixing of the urinary desensitization data, adopting a mixed compression strategy of fusion dictionary coding, entropy coding and predictive coding to carry out high-efficiency compression to generate a compressed data packet.
6. The urinary data management method of claim 1, wherein said step S4 is specifically: The user terminal calculates a hash value of the compressed data packet based on an SHA256 algorithm, divides the compressed data packet into a plurality of data blocks based on a preset block size, and sequentially constructs a data block sequence based on each data block; Calculating an MAC value based on a preset character string tag and the dynamic session key through an HMAC algorithm, taking the first 128 bits of the MAC value as a seed, generating a replacement sequence based on the seed through a pseudo-random number generator, and rearranging a data block sequence based on the replacement sequence to obtain a scrambling sequence; And calling the dynamic session key to encrypt the disordered sequence by adopting an AES-CTR mode to obtain an encrypted sequence, packaging the encrypted sequence and the hash value into an encrypted data packet, and uploading the encrypted data packet to a server based on a TLS protocol.
7. The urinary data management method of claim 1, wherein said step S5 is specifically: The server stores the received encrypted data packet into a user exclusive data container in real time, synchronously records the storage time of the encrypted data packet in the user exclusive data container, records access logs of the user exclusive data container for security audit and abnormal access detection in real time, calculates log fingerprints of the access logs based on an SHA-3 algorithm when the access logs are updated, and stores the log fingerprints into a blockchain; The access log comprises basic access information, access source and context information, operation object and range details, operation results and system response; the basic access information comprises an access time stamp, a visitor identity identifier and an access operation type; The access source and context information comprises a source IP address and geographic location, user terminal information, a network protocol and authentication credentials; The operation object and range details comprise a target data container ID and a data range related to operation; the operation result and system response comprise operation result state codes, data change abstracts and system triggering behaviors.
8. The urinary data management method in combination with deep learning of claim 1, wherein in step S6, the data drift detection model is constructed based on a time-series coding module, a drift feature extraction module, a drift detection module and a parameter generation module; The time sequence coding module is constructed based on an input standardization unit, a sequence embedding unit and a time convolution unit; the input normalization unit normalizes the input urinary desensitization data by adopting Z-score normalization to obtain normalized data, the sequence embedding unit embeds the normalized data into a high-dimensional vector sequence by using a Transformer encoder, and the time convolution unit carries out convolution operation on the high-dimensional vector sequence by adopting a time sequence convolution network to output a time sequence feature diagram; The drift feature extraction module is constructed based on a multi-scale feature extraction unit, an attention mechanism unit and a feature dimension reduction unit; the multi-scale feature extraction unit performs multi-resolution analysis on the time sequence feature map by using a wavelet transformation network in combination with a convolutional neural network to extract features of different time scales and output a multi-scale feature map, the attention mechanism unit applies attention weights to the multi-scale feature map by using a multi-head self-attention network, focuses on a time point related to drift and outputs a weighted feature vector; The drift detection module is constructed based on a drift classification unit, a drift regression unit and an countermeasure verification unit, wherein the drift classification unit adopts a support vector data description network to carry out drift detection on the dimension reduction feature vector and output a drift probability fraction; The parameter generation module is constructed based on a parameter optimization unit and a parameter formatting unit, wherein the parameter optimization unit uses a depth Q network, optimizes the preliminary drift parameter through a reward mechanism based on the drift mark and the preliminary drift parameter, and outputs the optimized drift parameter; the acquisition strategy generation model is constructed based on a time sequence feature coding module, a context aggregation module, a strategy generation module and a parameter optimization module; The system comprises a time sequence feature coding module, a context aggregation module, a strategy generation module and a parameter optimization module, wherein the time sequence feature coding module is used for extracting multi-scale time sequence features from urinary desensitization data through a transducer encoder and outputting feature sequences, the context aggregation module is used for carrying out attention pooling and global average pooling on the feature sequences and generating global context vectors after fusion, the strategy generation module is used for generating preliminary acquisition parameter vectors through multi-layer perceptron regression based on the global context vectors, and the parameter optimization module is used for applying constraint optimization on the preliminary acquisition parameter vectors and outputting final acquisition parameters.
9. The urinary data management method of claim 1, wherein said step S7 is specifically: The server equally divides the dynamic session key into a first sub-key and a second sub-key, encrypts the drift parameter through the first sub-key to obtain a first encryption parameter, encrypts the acquisition parameter through the second sub-key to obtain a second encryption parameter, encrypts the first encryption parameter and the second encryption parameter through the dynamic session key to obtain a second encryption parameter, maps each character of the second encryption parameter based on a preset mapping rule to obtain parameter ciphertext data, and transmits the parameter ciphertext data to a user terminal through a TLS protocol; and the user terminal decrypts the received parameter ciphertext data based on the dynamic session key to obtain the drift parameter and the acquisition parameter, compensates and corrects the real-time sensor monitoring data acquired subsequently based on the drift parameter, and adjusts the data acquisition strategy of the sensor array based on the acquisition parameter.
10. The urinary data management method of claim 1, wherein said step S8 is specifically: The user terminal inputs an account name and a password through a user interface displayed by the applet so as to perform one-time authentication, and then accesses the server; the user terminal encrypts the second random number based on the latest dynamic session key to obtain a first encrypted random number, the first encrypted random number is sent to the server, the server encrypts the first random number based on the latest dynamic session key to obtain a second encrypted random number, the second encrypted random number is sent to the user terminal, and secondary authentication is carried out based on the first encrypted random number and the second encrypted random number; and querying an access log through the user interface, performing integrity check on the access log based on the log fingerprint of the blockchain storage card, and performing online management comprising adding, deleting, modifying, querying and exporting on the data stored in the user-specific data container based on the authorization of the server.

Description

Urinary data management method combined with deep learning Technical Field The invention relates to the technical field of intersection of artificial intelligence and medical internet of things, in particular to a urinary data management method combining deep learning. Background The urinary system of the human body plays an important physiological role in excreting metabolic waste and regulating the balance of body fluids. The continuous monitoring and analysis of various physiological parameters reflecting the functional status (hereinafter referred to as "urinary data") is helpful for assessing physiological status, studying the adaptive changes of the organism under different conditions, and providing information support for health management. Traditional urinary data acquisition relies on single point, discrete tests in hospital settings, such as urine dipstick tests, laboratory urine routine and blood biochemical analysis. Although the accuracy is high, the method has the inherent limitations of long sampling interval, isolated data and the like, is difficult to capture continuous dynamic changes of parameters in daily life, and is also difficult to correlate and analyze with other health data. With the development of the sensor and the Internet of things technology, a technical scheme capable of realizing continuous monitoring outside a hospital appears. The technical scheme is generally based on a wearable/implantable sensor, a portable device and an intelligent terminal, can measure parameters such as urine flow rate, urine volume, urine ion concentration and the like, is combined with user log information, and is transmitted to a cloud for storage and analysis in a wireless manner. Although the prior art has realized continuous data collection, the following significant drawbacks still exist in the aspect of the overall architecture, especially the front-end data processing and end-to-end security system, and the reliability, instantaneity and large-scale application security of the system are restricted: (1) Data transmission redundancy and timeliness bottlenecks: In the prior art, the sensor node or the terminal side generally lacks effective data preprocessing capability, the original data stream contains a large amount of noise, invalid fragments and systematic errors, and the data stream is fully uploaded to the cloud without screening and compression. The cloud end not only occupies a large amount of communication bandwidth and increases transmission delay and packet loss risk, but also causes that a large amount of computation resources are consumed by the cloud end to clean data, so that the overall link delay from data generation to an available result is too long, and near real-time response to rapid parameter change is difficult to realize. (2) Data security and privacy protection architecture is vulnerable: the resulting multidimensional physiological and behavioral data is continuously monitored for highly sensitive personal health information. The current scheme generally adopts a static encryption key, lacks a dynamic key updating mechanism, and is easy to bring leakage and cracking risks after long-term use. In addition, the data is usually stored in the cloud in a plaintext or decryptable form after being uploaded, and once the server is attacked or internal leakage occurs, complete sensitive information of the user is easily exposed. (3) Sensor data quality challenge: The sensor is susceptible to matrix effects, pollution, environmental interference and the like, has problems of accuracy and stability, and generally lacks an effective online calibration mechanism, and has poor long-term data comparability. In addition, in order to realize continuous monitoring, the requirement on the compliance of users is high, and errors can be introduced into improper operation or habit change, so that the data quality is affected. (4) System stiffness, lack of personalized adaptive capabilities: the existing system mostly adopts a fixed acquisition strategy, a processing pipeline and an alarm threshold value, and cannot be dynamically adjusted according to individual differences of users, real-time states or data characteristics. Such "one-touch" designs are difficult to optimize the efficiency of resource (e.g., power, computing power, storage) utilization, and may result in missed detection of critical events or unnecessary resource consumption, affecting the sustainability of long-term monitoring. Therefore, how to provide a urinary data management method combining deep learning, which realizes high-efficiency real-time processing, dynamic safety protection and personalized self-adaptive analysis on continuously monitored urinary data, so that on the premise of guaranteeing data safety and privacy, the improvement of data transmission efficiency, data quality and overall intelligent level of a system becomes a technical problem to be solved urgently. Disclosure of Invention The invention a