CN117147544B

CN117147544B - Urine detection and analysis system of intelligent closestool

Info

Publication number: CN117147544B
Application number: CN202311103892.3A
Authority: CN
Inventors: 彭伟杰; 罗少雄
Original assignee: Shenzhen Guobang Biotechnology Co ltd
Current assignee: Shenzhen Guobang Biotechnology Co ltd
Priority date: 2023-08-30
Filing date: 2023-08-30
Publication date: 2024-07-12
Anticipated expiration: 2043-08-30
Also published as: CN117147544A

Abstract

The invention relates to a urine detection and analysis system of an intelligent closestool. A real-time multi-index analysis module is adopted, and the module carries out real-time quantification on various biological markers in urine through a miniaturized biological sensor and is combined with a self-adaptive sampling technology, and the sampling technology can automatically adjust the urine sampling frequency according to the physiological state and behavior of a user; the system further integrates a non-contact detection module, adopts a non-contact mode of optics or electromagnetic field to collect urine samples, and carries out micro-processing and detection of the samples through an integrated micro-fluidic chip; the system also includes an AI diagnostic engine that uses machine learning algorithms to diagnose and predict urine analysis data in real-time and can generate personalized health reports. The system is also provided with a multi-mode user interaction interface and a data security and privacy protection mechanism, wherein the mechanism adopts end-to-end encryption and blockchain technology to ensure the security and privacy of user data.

Description

Urine detection and analysis system of intelligent closestool

Technical Field

The invention relates to a urine detection and analysis system of an intelligent closestool.

Background

The existing intelligent closestool urine detection and analysis system has certain convenience and practicability, but has the following problems: for example, the accuracy problem leads to inaccuracy of urine sample analysis, such as errors in detecting indicators of uric acid, protein and the like, thereby affecting medical diagnosis and health assessment. The response time delay affects the user's timely feedback, resulting in the user possibly missing the best treatment or changing the lifestyle opportunities. High energy consumption lacks optimization, which means that the device needs to replace batteries or connect to a power source more frequently, increasing use costs and inconvenience. In conditions such as poor environmental adaptability, e.g., in conditions of humidity, large temperature variation, or light, the detection results are easily affected. The user experience is poor, for example, the interface is not intuitive, the operation is complex, or the feedback mechanism such as sound and visual prompt is not friendly. The data fusion and interpretation capabilities are inadequate, i.e., the system is not able to effectively combine multi-source data for accurate and useful health assessment. Such as lack of adaptive sampling or dynamic thresholding functionality, makes sample collection inflexible under different conditions. Insufficient security and privacy protection of data, such as unencrypted transmissions or storage, results in user data that is easily stolen or misused. The cost is high, and the cost is increased due to the application of a high-precision sensor and a complex algorithm, so that the method is not suitable for large-scale or low-cost application. In the absence of an effective self-diagnosis and fault recovery mechanism, for example, when a system fails or fails, automatic diagnosis and recovery are not possible, requiring manual intervention. The complexity of the system results in difficult maintenance, requiring specialized technicians to perform periodic checks and updates. Interference with a particular drug or foodstuff affects the accuracy of the test, e.g., certain drugs may affect the chemical composition of urine, and thus the test results. Storage and handling problems such as urine samples can affect the results, such as exposure of the sample to air or improper temperatures, which can lead to chemical composition changes. Lack of multi-user support and personalized settings does not meet the specific needs of different users or family members. Poor durability, such as corrosion or physical damage to sensors and other components, affects their life and performance. In the absence of integration with other health monitoring devices, for example, data synchronization and analysis with smart watches, health trackers, etc. are not possible. Untimely or unstable system software updates may result in new medical research and diagnostic techniques not being incorporated in time. The measurement range is limited, for example, only a few indexes in urine can be detected, and the health condition of a user cannot be comprehensively estimated. In cases such as high operation complexity, a certain medical or technical knowledge is required for the user to effectively use. The medical accuracy of reporting and interpretation is to be improved, such as using popular or ambiguous languages, and the detection results and their medical significance cannot be accurately interpreted.

Disclosure of Invention

The invention aims to provide a urine detection and analysis system of an intelligent closestool, so that part of defects and shortcomings pointed out in the background art are overcome.

The invention solves the technical problems as follows: a real-time multi-index analysis module is adopted, and the module carries out real-time quantification on various biological markers in urine through a miniaturized biological sensor and is combined with a self-adaptive sampling technology, and the sampling technology can automatically adjust the urine sampling frequency according to the physiological state and behavior of a user; the system further integrates a non-contact detection module, adopts a non-contact mode of optics or electromagnetic field to collect urine samples, and carries out micro-processing and detection of the samples through an integrated micro-fluidic chip; the system also includes an AI diagnostic engine that uses machine learning algorithms to diagnose and predict urine analysis data in real-time and can generate personalized health reports.

Further, the personalized health report is generated according to the historical data and health conditions of each user;

The system is also provided with a multi-mode user interaction interface, comprising a voice, a touch screen and a mobile phone APP, and is provided with a data security and privacy protection mechanism, wherein the mechanism adopts an end-to-end encryption and blockchain technology to ensure the security and privacy of user data; the system further has a remote medical treatment integration function, and urine analysis data are safely transmitted to a remote medical expert for further analysis and diagnosis; the system also uses a disposable sensor made of biodegradable material.

Further, the self-adaptive sampling technology is driven by a dynamic frequency adjustment algorithm based on reinforcement learning, the algorithm adjusts sampling frequency in real time and is combined with a multi-source data fusion module, and the fusion module fuses the data of the environmental sensor with urine detection data by using a Kalman filtering algorithm; the system further adopts a time sequence analysis and sampling decision threshold adjustment mechanism to dynamically adjust the threshold of the sampling decision according to the output result of the AI diagnosis engine and the predicted urine index trend; the system comprises a context sensing mechanism, wherein the context sensing mechanism uses an embedded camera or an infrared sensor to detect whether a user is in a state suitable for sampling;

The system also comprises a self-adaptive sample storage mechanism, a pre-sampling data quality detection module and a sample data preprocessing and compressing module, wherein the modules interact with the integrated micro-fluidic chip and the non-contact detection module to realize automatic collection, storage and micro-processing of the urine sample; the system provides real-time feedback and user guidance through a multi-mode user interaction interface, the interface is combined with a data security and privacy protection mechanism, and the security and privacy of user data are ensured by adopting an end-to-end encryption and blockchain technology.

Further, the integrated non-contact detection module is driven by an optical calibration subsystem with a bimodal optical sensor, uses two lasers with different wavelengths to respectively detect protein and sugar in urine, adopts a high-frequency electric field pulse shape optimized by a machine learning algorithm, and is used for improving the measurement accuracy of conductivity and pH value; the module further comprises an instant data checking mechanism, wherein the instant data checking mechanism is used for carrying out real-time data comparison through the redundant sensor, and a local data cache is used for quick local analysis; the non-contact detection module is also provided with a sensor thermal management subsystem, the optimal working temperature of the sensor is maintained by using a liquid cooling or air cooling mode, the position and the angle of the sensor are adjusted in real time through a micro-electro-mechanical system (MEMS) accelerometer and a gyroscope, and the sensor is changed based on the physical state of a user;

The non-contact detection module interacts with the self-adaptive sampling technology module, and dynamically adjusts sensor parameters by using a machine learning algorithm so as to adapt to sampling frequency and time points; interaction with the data preprocessing and compression module to determine the data to be processed or transmitted preferentially; the data of the non-contact detection module is input to the multi-source data fusion module, is fused with environment data, and is provided with different QoS priorities so as to improve the overall data analysis and user experience performance of the system.

Further, the AI diagnosis engine firstly acquires biochemical data of the urine sample through the non-contact detection module and performs pretreatment, and then performs feature extraction by using the multilayer CNN, in particular to key feature identification aiming at high dimension; meanwhile, an LSTM module is embedded and used for processing time series analysis of urine samples, in particular for progress tracking of chronic diseases; model updating is carried out on the premise of ensuring user privacy by using a federal learning mechanism;

The AI diagnosis engine also applies a real-time back propagation algorithm to perform model optimization, and evaluates the uncertainty of the quantized result through a confidence interval; the system dynamically adjusts model parameters according to feedback obtained from the adaptive sampling module through a dynamic weight distribution strategy; the AI diagnostic engine also utilizes a second order optimization algorithm: the method comprises the steps of carrying out rapid model convergence by Newton method, and simultaneously automatically triggering an input data fault detection and self-repairing subroutine when inconsistent or error data are received; further, the AI diagnostic engine processes the multimodal data from the multimodal data fusion module and performs a more accurate diagnosis based thereon; to speed up large-scale data processing and model reasoning, the AI diagnostic engine uses a general purpose graphics processing unit for parallel computation.

Further, the AI diagnostic engine firstly performs feature extraction through a multi-layer Convolutional Neural Network (CNN); this process is driven by the formula f (x) =relu (W x+b) for identifying key features from biochemical data of urine samples; where f (x) is the feature map, W is the convolution kernel, x represents the convolution operation, b is the bias term, and ReLU represents the activation function;

next, the engine processes the time series data using a long short time memory network (LSTM),

The formula ct=f _t⊙ct_t-1+i_t⊙tanh(W·[h_t-1,x_t ] +b) is used to track the progression of chronic disease;

wherein f _t,i_t is a forget gate and an input gate, while, if is element-by-element multiplication, ct is the current cell state;

the engine further applies a real-time back-propagation algorithm for model optimization, wherein the gradient is defined by Calculating;

where J (θ) is the loss function, Is the gradient with respect to the parameter θ;

In order to realize rapid model convergence, the engine adopts a second-order optimization algorithm Newton method,

Its update rule is represented by formulaDefinition;

Where H (f (θ)) is a hessian matrix for representing the second derivative of the loss function.

Further, the multi-source data fusion module utilizes a Kalman filtering algorithm to fuse the data received from the environment sensor and the urine detection module in real time; wherein the Kalman filtering algorithm dynamically changes the covariance matrix through adaptive covariance adjustment; specifically, covariance matrixWhere α is an adjustable factor, var (x _t) is the variance of the last N data points, thereby enabling adaptive processing of data uncertainty;

at the same time, the system introduces a sensor calibration mechanism to correct sensor bias, using the formula Wherein a and b are calibration parameters;

The fusion module further comprises a data anomaly detection mechanism, and a Markov distance formula is calculated: d ²＝(x-μ)^T·Σ^-1 (x- μ) to label potential outlier data points; when D ² > threshold, the data point will be marked as abnormal;

the system also adopts a self-adaptive weight distribution mechanism to dynamically adjust the Kalman gain according to the reliability of urine detection and environmental data, and the specific formula is as follows

The fusion module also includes an advanced noise model to more accurately describe the system noise, specifically a Gaussian Mixture Model (GMM): the system also includes an AI diagnostic engine that receives the fused data, data confidence and anomaly tags from the fusion module for more accurate diagnosis; the system comprises an adaptive sampling module, a sampling module and a sampling module, wherein the adaptive sampling module adjusts sampling frequency of urine and environmental data according to an anomaly tag triggered by a data anomaly detection mechanism; the system realizes high-precision and low-energy consumption comprehensive analysis and diagnosis of urine detection and environmental factors through interaction of the modules and the technical points and related calculation formulas.

A method for urine detection analysis system of intelligent closestool, this method is through a embedded database to carry on data cache and inquiry, and use the characteristic to withdraw the characteristic extraction algorithm to carry on the extraction of the key characteristic to the original urine sample data, then adopt STL algorithm to decompose and strengthen the accuracy of the time series analysis model with the seasonality and trend;

The time sequence model is trained and optimized through an automatic super-parameter optimization algorithm by using Bayesian optimization; after model training, real-time interpretability scoring is performed by using an AI diagnosis engine and adopting a SHAP method; based on the output result of the AI diagnosis engine and the confidence interval of the model, calculating to obtain a new sampling decision threshold value, optimizing the threshold value through a dynamic learning rate and state feedback control logic, and carrying out sliding average or exponential smoothing post-processing; the newly calculated sampling decision threshold is finally applied to the adaptive sampling module to achieve more accurate and personalized urine detection.

Further, the automatic super-parameter tuning algorithm comprises a data preprocessing module, wherein the data preprocessing module is used for collecting and storing urine samples and environmental sensor data in a fixed time interval and simultaneously carrying out data cleaning, abnormal value detection and normalization processing; initializing a basic time sequence model, and dynamically optimizing model super-parameters through a search range and initial seed points of a Bayesian optimization algorithm; after model training and verification, updating a time sequence model by applying an optimal super-parameter combination, and combining model output and a predicted confidence interval;

Dynamically adjusting the sampling decision threshold by kalman filtering or state feedback logic; the system further includes an AI diagnostic engine for receiving the output of the time series model for further analysis and interpretation using the advanced machine learning model to generate an interpretability score; finally, the results of the AI diagnostic engine are used to further optimize the time series model and bayesian optimization algorithm, wherein new labels and diagnostic results are used for model retraining, enabling a tight association of the system with user needs and health conditions; through the series of interaction and feedback mechanisms, the system ensures continuous optimization of the model and the threshold value, improves the accuracy of urine detection and increases the individuation and interpretation capability of the system.

Further, the dynamic learning rate and state feedback control logic comprises a specific learning rate scheduling algorithm and a gradient shearing technology, wherein the dynamic adjustment of the learning rate is realized in an optimizer, and meanwhile, the momentum parameter is used for guiding optimization;

The system also comprises a state feedback control logic module, kp, ki and Kd parameters in the PID controller are automatically adjusted by using a genetic algorithm or a simulated annealing optimization algorithm, a second-order low-pass filter is added before state feedback to reduce noise, and a design algorithm compensates state feedback delay caused by communication or calculation delay;

The system is further provided with a data synchronization mechanism, timestamp synchronization is established between the data preprocessing module and the time sequence model so as to ensure the consistency of the data, and a priority queue is used between the AI diagnosis engine and the time sequence model so as to ensure that more urgent or important data can be processed preferentially; the system also includes a fast model weight online update mechanism that allows updating the time series model and the AI diagnostic engine without downtime, thereby meeting the need for flexible and reliable operation while maintaining high accuracy and response speed.

The invention has the beneficial effects that:

1. accuracy is improved: urine samples can be analyzed more accurately by high-precision sensors and advanced data processing algorithms, providing a more reliable health assessment.

2. Real-time feedback: the system can provide detection results in a short time, provide timely health feedback for users, and is beneficial to early diagnosis and treatment of potential health problems.

3. Energy efficiency: by means of the optimized energy management system, power consumption can be reduced, and operation cost can be reduced.

4. Environmental suitability: has good environment adaptability and can keep high performance under various environmental conditions.

5. User experience: the easy-to-use interface and various feedback mechanisms (e.g., visual, audible cues) enhance the user experience.

6. Data fusion: the multisource data can be effectively fused through the advanced algorithm, and comprehensive health assessment is provided for the user.

7. Dynamic adjustment: the system has the functions of self-adaptive sampling and dynamic threshold adjustment, and can provide high-quality service under different use scenes and conditions.

8. Data security: privacy and security of user data is ensured through encryption and other security mechanisms.

9. Cost effectiveness: although a high-precision sensor and a complex algorithm are adopted, the balance between cost and performance is realized through the optimal design, so that the system has a wide application prospect.

10. Self-diagnosis and recovery: the system has excellent self-diagnosis and fault recovery capabilities, and can reduce maintenance cost and improve system reliability.

11. And (3) personalized setting: the multi-user and personalized setting is supported, and the specific requirements of different users can be met.

12. Integration: the system can synchronize and analyze data with other intelligent health equipment (such as a smart watch, a health tracker and the like) to provide more comprehensive health management service.

Drawings

FIG. 1 is a flow chart of a urine detection and analysis system of an intelligent toilet according to the present invention.

Detailed Description

The following describes the embodiments of the present invention in detail with reference to the drawings.

Examples: the invention relates to an integrated intelligent closestool urine detection and analysis system; the following is an explanation of some key points:

The real-time multi-index analysis module: this module uses a micro-biosensor that can detect multiple biomarkers (e.g., sugar, proteins, ions, etc.) in urine in real time, providing a comprehensive health assessment. The use scenario of the module is: assuming a diabetic patient uses the intelligent closestool, the sensor can detect sugar, ketone bodies, trace proteins and the like in urine in real time. In this way, once abnormal data is detected, the system can prompt the user immediately, possibly avoiding acute or chronic complications due to diabetes.

Adaptive sampling technique: this technique can automatically adjust the frequency of urine sampling based on the physiological state and behavior of the user (e.g., amount of exercise, eating habits, etc.). By the aid of the method, analysis accuracy is improved, and energy use is optimized. The use scenario of the module is: a hypertensive patient often needs to monitor sodium and potassium levels. The system can adjust the frequency of urine sampling according to the salt intake and the weight of the user. For example, if the user has recently increased salt intake, the system will increase the frequency of sodium and potassium detection.

A contactless detection module: the module collects urine samples in an optical or electromagnetic field mode, so that a user can use the module more conveniently and sanitarily. The use scenario of the module is: in the influenza season, people are particularly concerned about cross-infection. The intelligent closestool collects urine samples by using the optical sensor, and a user does not need to contact any surface, so that the possibility of cross infection is greatly reduced.

Integrated microfluidic chip: the micro-fluidic chip can perform micro-processing and detection on the sample, so that the speed and accuracy of detection are improved, and the cost is reduced. The use scenario of the module is: consider a pregnant woman using the intelligent toilet. The micro-fluidic chip can be used for rapidly and accurately detecting hCG (human chorionic gonadotrophin) of a trace urine sample, so that a pregnant woman can know whether the pregnant woman is pregnant or not in the earliest stage, and the pregnancy management can be started earlier.

AI diagnostic engine: this is the decision core of the system, which uses machine learning algorithms to parse the collected data, enabling the generation of real-time diagnostics and prognostics, as well as personalized health reports. The use scenario of the module is: a user who suffers from urinary tract infection for a long time uses the intelligent closestool. The AI diagnostic engine performs real-time analysis based on the data of leukocytes, erythrocytes, bacteria, etc. in urine, and then generates a personalized health report. This report will not only indicate whether there is evidence of urinary tract infection, but will also predict when the user is most likely to reoccur urinary tract infection based on past data and give preventive advice.

Specific example 1: when the user gets a smart toilet in Zhang Zoujin home, the facial recognition technology immediately recognizes his identity. The touch screen of the toilet is activated, a personalized user interface is displayed, and simultaneously a voice assistant Alice greets the sheetlet and inquires: "do you good, small, need to do today's urine tests? "based on data from the past month, the system records show 95% user satisfaction, indicating that the multimodal user interface is very popular.

The small tab clicks "yes" and then uses the intelligent toilet to perform urine detection. After the test is finished, all data are encrypted by an end-to-end encryption technology, so that information security is ensured. The encrypted data is stored in a secure database using blockchain technology, which ensures the integrity and non-tamper ability of the data. Statistics data of the past 12 months show that the data security and privacy protection mechanism of the system is extremely reliable, and no event of data leakage or tampering occurs.

These encrypted urine test data are then sent to the telemedicine expert Zhao Yisheng via a secure channel. According to statistics, 80% of users are willing to share their health data, which is a very high rate. After receiving the data, the Zhao doctor starts to perform further analysis and diagnosis. After analyzing the urine sample of the small sheet, the Zhao doctor makes an online consultation with the small sheet through the built-in video communication function of the system, and gives out some valuable medical advice.

Finally, the disposable sensor for detection is automatically discarded into a dedicated biodegradation processing unit. These sensors are made of biodegradable PLA material, which is both environmentally friendly and efficient. This practice has been shown to reduce the plastic waste yield by about 60%.

The series of processes not only show the advantages of the intelligent closestool urine detection analysis system in the aspects of technical realization, user acceptance and environmental protection, but also fully prove the high integration and cooperative work capability among all the modules of the system. This provides a comprehensive and careful solution to modern intelligent health management.

Specific example 2: week health monitoring of the sheetlet:

The small sheet is a patient with hypertension and diabetes mellitus, and recently, an intelligent closestool urine detection and analysis system is used for closely monitoring the health condition of the patient.

The first day: adaptive sampling and environment fusion

In the morning, the small pieces used the intelligent toilet for the first time. A self-adaptive sampling module driven by a reinforcement learning algorithm (Q-learning) of the system identifies the sheetlet and dynamically adjusts the urine sampling frequency. Meanwhile, the Kalman filtering algorithm fuses environmental data (such as indoor temperature and humidity) and urine data.

Wherein, the Q-learning algorithm of reinforcement learning:

State definition: values for each biomarker after each urine sample.

Action definition: whether to take the next sample and when.

Bonus function: positive rewards are obtained by accurately detecting abnormal values, and negative rewards are obtained by false detection or omission detection.

Q value update: updates are made using the bellman equation.

Whereas the kalman filter algorithm:

The equation of state: the current state is predicted using the state at the previous time and the control input.

Observation equation: the predicted state is compared with the actual observations.

Updating the equation: the state equation is updated based on the error of the observation equation.

Third day: dynamic threshold adjustment and prediction

Based on the data of the past few days of the sheetlet, the AI diagnostic engine of the system uses the ARIMA time series model for preliminary analysis. The sodium content of the sheetlet was found to be increasing, which may be a signal for hypertension. The system dynamically adjusts the threshold using the Sigmoid function, detecting this indicator more frequently.

Wherein, ARIMA time series model:

Autoregressive (AR): the current value is predicted using the first few observations of the urine biomarker.

Moving Average (MA): the error between the observed value and the predicted value is modeled.

Difference (I): the data is subjected to a first order or higher order difference to achieve data smoothing.

The dynamic threshold is smoothly adjusted by using the Sigmoid function, so that sampling errors caused by abrupt change of the threshold can be avoided.

Fifth day: data quality and preprocessing

The One-Class SVM algorithm of the system performs a pre-sampling data quality check. Then, the sample data is subjected to PCA reduction and compression, and micro-processing is performed by the micro-fluidic chip.

The One-Class SVM algorithm described above maps the data points of the normal samples to a high dimensional space to find an optimal hyperplane by maximizing the distance between the normal data points and the origin. The PCA dimension reduction and compression is to find the main components of the original data by using a eigenvalue decomposition method, and then select the first few main components to perform the dimension reduction of the data.

Seventh day: data security and telemedicine integration

After one week, the system uses the RSA algorithm to generate an encrypted health report and uses the SHA-256 hash function to record data on the blockchain, ensuring its security and integrity. And then transmitted to a remote medical expert granddoctor through a secure channel for further analysis and advice. This week of embodiment shows the specific application of the various technical modules and core operations in detail, thus constituting a comprehensive, safe and personalized health monitoring scheme.

The RSA algorithm uses public key for encryption and private key for decryption. And the SHA-256 hash function is for data security and integrity verification of the blockchain.

Specific example 3: the small money is a middle-aged person with a mild bladder problem, and the intelligent closestool with the advanced contactless detection module starts to be used. When he is urinating, the bimodal optical sensor activates and emits two lasers of different wavelengths, one for determining the protein content in the urine and the other for sugar. The high frequency electric field pulse shape (optimized machine learning algorithm) is also used to measure the conductivity and pH of urine over a specific time window, e.g., 1 second.

Actual data: the optical sensor measured protein content at 0.03g/dL and sugar content at 0.2g/dL at wavelengths of 400nm and 600nm, while the electric field pulse measured conductivity at 1.5S/m and pH at 6.5.

All the data are checked in real time and compared with the data of the other group of redundant sensors, so that errors are reduced. The data is buffered in a local buffer, which can be analyzed more quickly.

Actual data: the redundant sensor measured a protein content of 0.031g/dL, a sugar content of 0.21g/dL, a conductivity of 1.49S/m and a pH of 6.51, which is highly consistent with the original data.

The sensor thermal management subsystem maintains the sensor at an optimal working temperature, which is assumed to be 25 ℃, in a liquid cooling mode, and simultaneously the MEMS accelerometer and the gyroscope adjust the position and the angle of the sensor in real time.

These data then interact with the adaptive sampling technique module. Assuming the system finds that the urine sugar and protein content of the small money does not change much from 10 pm to 4 am, the machine learning algorithm automatically reduces the sampling frequency for this period.

These data are then passed to a data preprocessing and compression module, wherein certain data are marked for preferential processing or transmission according to predefined rules (such as abrupt pH changes or abnormally high protein content).

Finally, all of these data are input to a multi-source data fusion module for fusion with environmental data (e.g., indoor temperature and humidity). In addition, the system also sets different QoS priorities to ensure quick response to the key health indicators, thereby improving the overall user experience. Through the series of steps, the small money can obtain more accurate and timely health feedback, and medical specialists can more efficiently carry out remote diagnosis and treatment suggestion.

Specific embodiment 4 is a scene continuation continuing with embodiment 3: the money starts his day and uses the high-end intelligent toilet as usual. First, the non-contact detection module measures biochemical data in the urine of little money, such as protein content of 0.03g/dL, sugar content of 0.2g/dL, conductivity of 1.5S/m and pH value of 6.5.

1. Preprocessing and feature extraction

These data are first received by the AI diagnostic engine and pre-processed, such as normalized and missing value interpolation. Subsequently, a multi-layer Convolutional Neural Network (CNN) performs feature extraction on these pre-processed data to identify patterns of key biomarkers.

2. Time series analysis

The system also recorded urine sample data for the past week of the small money. Long-term memory networks (LSTM) are used for time series analysis, particularly useful for tracking chronic disease progression.

Actual data: for example, the sugar content of small money fluctuates from 0.18g/dL to 0.21g/dL in the past seven days, and LSTM recognizes an ascending trend.

3. Federal learning and model updating

The model is updated and optimized on the premise of not revealing user data through a federal learning mechanism.

4. Model optimization and quantization uncertainty

A real-time back propagation algorithm is used for model optimization. Meanwhile, uncertainty of model output is quantified through calculation of confidence intervals.

Actual data: for example, the model predicts that a small amount is at 40% risk for diabetes in the next year with a confidence interval of + -5%.

5. Dynamic weight distribution and second order optimization algorithm

Based on feedback from the adaptive sampling module, the AI diagnostic engine adjusts model parameters via a dynamic weight distribution strategy. Meanwhile, a second order optimization algorithm such as newton's method is used to accelerate model convergence.

6. Multi-modal data fusion and fault detection

These urine data are fused with environmental data obtained from other sensors, such as indoor temperature and humidity sensors. When inconsistent or erroneous data is received, the fault detection and self-repair subroutine is automatically triggered.

7. Parallel computing

To process large-scale data and achieve fast model reasoning, AI diagnostic engines use a general-purpose Graphics Processing Unit (GPU) for parallel computation.

Through this series of highly advanced calculations and analyses, the AI diagnostic engine not only provides his health report to the small money, but also predicts his future health risk for one year. All this is done while protecting the data privacy of the small money thanks to federal learning and end-to-end encryption techniques.

Specific example 5: one week health monitoring of xiao Zhao:

1. Feature extraction (CNN)

A smart toilet was used for Zhao Mei days and the data of urine samples were collected by a contactless detection module. Suppose that on a certain day, the biochemical data obtained includes a protein content of 0.1g/dL and a sugar content of 0.2g/dL.

The calculation process comprises the following steps:

performing feature extraction by using a Convolutional Neural Network (CNN), wherein a specific calculation formula is as follows;

f (x) =relu (w\ astx +b). In this hypothetical example, let w= [0.1,0.2], b=0.1.

f(x)＝ReLU(0.1*0.1+0.2*0.2+0.1)

f(x)＝ReLU(0.01+0.04+0.1)

f(x)＝ReLU(0.15)

f(x)＝0.15

2. Time series analysis (LSTM)

Xiao Zhao urine sample data was recorded over the past week to track the progression of chronic disease.

The calculation process comprises the following steps:

Using long-short-term memory networks (LSTM), a specific updated formula is ct=f _t⊙ct_t1+i_t⊙tanh(W·[h_t1,x_t ] +b.

Let ct _t1＝0.5,f_t＝0.6,i_t＝0.7,W＝[0.1,0.2],h_t1＝0.4,x_t =0.1 and b=0.2 at the last time point.

ct＝0.6⊙0.5+0.7⊙tanh(0.1·0.4+0.2·0.1+0.2)

ct＝0.3+0.7⊙tanh(0.04+0.02+0.2)

ct＝0.3+0.7⊙tanh(0.26)

ct＝0.3+0.7⊙0.255

ct＝0.3+0.1785

ct＝0.4785

3. Model optimization (counter-propagation and Newton method)

Let the loss function J (θ) = (f (x) -y ², where y is the actual value. Assume that the initial θ value is 0.5.

Back propagation:

Newton method:

The update rule is

Let the inverse of the hessian matrix H (f (\theta)) be 1.

(θ_{"{new}}＝0.5-1*2(0.15

Specific example 6: the king is in an intelligent home environment and uses intelligent toilets:

1. and (3) data acquisition: urine sample data (e.g., urine protein content 0.1g/dL, sugar content 0.2 g/dL) were obtained by the smart toilet, while environmental data (e.g., room temperature 24 ℃ C., humidity 50%) were obtained by the environmental sensor.

2. Multi-source data fusion (Kalman filter)

Covariance adjustment:

Initial covariance matrix Α=0.5, variance Var (x _t) =0.05 of the last N urine data points.

Sensor calibration:

Using a calibration formula Assuming a=0.9, b=0.1, the urine protein content after calibration is 0.9×0.1+0.1=0.19 g/dL

Anomaly detection (mahalanobis distance):

let mean μ= [0.15,0.18] and covariance matrix

D²＝(x-μ)^T·Σ¹·(x-μ)

Assuming D ² = 1.5 and a threshold of 1, this data point would be marked as abnormal.

Self-adaptive weight distribution:

The confidence levels of urine data and environmental data were assumed to be 0.9 and 0.7, respectively.

Noise model (gaussian mixture model):

Assuming that the system noise consists of two gaussian distributions, weights 0.3 and 0.7, respectively, mean and variance μ ₁ =0, And mu ₂ = 0.1,

AI diagnostic engine

The AI diagnostic engine receives post-fusion data (urine data and environmental data), data credibility, and anomaly tags from the fusion module. Based on these, it is possible to diagnose that the king is at slight risk for diabetes and suggest further examination.

4. Self-adaptive sampling module

Because there is a distinct data point, this module may decide to increase the sampling frequency of urine and environmental data to obtain more accurate information.

Specific example 7: personalized urine detection of chapter women:

1. Data caching and querying (embedded database):

Suppose that a chapter woman's smart toilet stores daily urine sample data such as protein content, sugar, electrolyte, etc. for the past year. These data are stored in a high-speed embedded database.

2. Feature extraction:

Raw urine sample data (now 0.2g/dL protein content, 0.1g/dL sugar) is subjected to dimension reduction and key feature extraction using feature extraction algorithms such as PCA or t-SNE.

3. Time series analysis (STL algorithm):

the STL algorithm is used to decompose seasonal and trending components. It is assumed that in the past year, the protein content of chapter women has increased in summer, but generally has a decreasing trend. The STL algorithm is able to identify these patterns and perform a more accurate analysis accordingly.

4. Automatic super-parametric tuning (bayesian optimization):

In the model training stage, bayesian optimization is used for automatically adjusting super parameters of the model, such as learning rate, regularization parameters and the like, so as to realize better prediction accuracy.

5. Real-time interpretability score (SHAP method):

SHAP (SHAPLEY ADDITIVE Explanations) is used to calculate the interpretability of model predictions. For example, it may be pointed out that protein content is the major factor that leads to a model predictive chapter women at risk of kidney problems.

6. Calculating a sampling decision threshold:

Based on the output of the AI diagnostic engine and the confidence interval of the model (e.g., the probability of predicting kidney problems for chapter women is 60%), a new sampling decision threshold is calculated. Assuming the original threshold is 0.5, the new threshold is optimized to 0.55 by the dynamic learning rate and state feedback logic.

7. Post-processing (sliding average or exponential smoothing):

The newly calculated sampling decision threshold is smoothed using a running average or exponential smoothing.

8. And the self-adaptive sampling module is used for:

A new sampling decision threshold (e.g., 0.55) is applied to the adaptive sampling module to achieve more accurate and personalized urine detection. Because the new threshold is higher than the original threshold (0.5), the system may increase the sampling frequency of the urine sample to more closely monitor the health status of the chapter woman.

Specific example 8: health monitoring of chapter women, assuming chapter women is an intelligent toilet user, worry about the risk of diabetes.

S1, a data preprocessing module:

In the chapter woman example, the intelligent toilet automatically collects urine samples and environmental data every morning. Suppose that data collected on a day includes a protein content of 200mg/dL for urine, and an indoor temperature of 25℃and a humidity of 60%. During the data cleansing phase, the system automatically removes unreasonable data. For example, if the normal protein content range is 10-30mg/dL, then a 200mg/dL reading is clearly an outlier and therefore would be rejected by the system.

Then, in the outlier detection phase, the data is further screened. For example, if the protein content of urine suddenly rises to 45mg/dL, and the previous data is typically in the range of 10-30mg/dL, this 45mg/dL reading will be marked as abnormal and further examination or exclusion will be performed.

Finally, in the normalization stage, all data are normalized to a range of 0-1. Taking the protein content of urine as an example, it is assumed that the remaining data range after outliers are removed is 12-20mg/dL. A specific protein content reading, such as 18mg/dL, will be processed using the simplest min-max normalization method. The calculation process is as follows:

With such normalization, the protein content of 18mg/dL was converted to 0.75, which further facilitates subsequent advanced analysis and model training.

S2, initializing a basic time sequence model:

Initializing a basic time series model is typically the first step in modeling historical data. In the case of chapter women, the present case can consider the use of ARIMA (autoregressive integrated moving average model) as the initial model. The ARIMA model is applicable to non-seasonal, and other types of time series data.

Since the ARIMA model has three main parameters: p (number of autoregressive terms), d (number of differences), q (number of moving average terms). In the initialization phase, a basic combination of parameters can be set, for example p=1, d=1, q=1.

Suppose that women in the present case have urine protein content data over the past 30 days, which have been washed and normalized by the preceding pretreatment steps.

Using the 30 day data, a basic ARIMA (1, 1) model can be fitted.

After the model fitting is completed, model evaluation is required. This is typically done by calculating metrics such as AIC (red pool information criterion), BIC (bayesian information criterion), or RMSE (root mean square error).

Suppose that the case has urine protein content data (normalized) for the following chapter women over 30 days:

[0.2,0.22,0.19,0.21,0.25,0.23,0.2,0.18,0.19,0.21,…]

1) These data and parameters (p=1, d=1, q=1) were used to fit the ARIMA model.

2) Assuming that the AIC score of the model is 120, this provides a benchmark model for this case, which can be further optimized by subsequent steps (e.g., bayesian optimization).

In this way, the scheme successfully initializes a basic time series model and prepares for more advanced analysis and optimization. This step provides the basis for subsequent model optimization and personalized health advice.

S3, bayesian optimization algorithm

After the basic ARIMA model initialization and data preprocessing is completed, the next task is to further optimize the performance of the model. Here, the scheme adopts a bayesian optimization algorithm to perform automatic super-parameter tuning.

First, a search range is set. Taking learning rate and regularization parameters as examples, search ranges of 0.001 to 0.1 and 0.1 to 1 are set, respectively. These ranges were selected based on previous research and experimental experience.

Then, an initial seed point is selected. Assume that three sets of initial hyper-parameter values are selected for testing:

1. learning rate=0.005, regularization parameter=0.5

2. Learning rate=0.05, regularization parameter=0.8

3. Learning rate=0.01, regularization parameter=0.2

These initial seed points are used in the present case to train the ARIMA model and record the AIC scores for each set of parameters. For example, a first set of parameters may yield aic=115, a second set yields aic=110, and a third set yields aic=112.

The results of these preliminary experiments are then used by a bayesian optimization algorithm to build a probabilistic model (typically a gaussian process model). The algorithm will then select a new combination of parameters to maximize a certain objective function (e.g., the amount of improvement desired), and then perform a new round of model training and evaluation.

It is assumed that in this round the algorithm selects a learning rate=0.025 and regularization parameter=0.35 and gets aic=108, which score is better than any initial seed point.

S4, model training and verification:

After multiple iterations using the bayesian optimization algorithm, it is assumed that the scheme finds a set of best performing hyper-parameters: the learning rate was 0.05 and the regularization parameter was 0.5. The set of parameters is determined from the lowest AIC value (e.g., aic=108). In the following, the present case will explain in detail how this set of hyper-parameters is used to retrain and validate the model.

And using the found optimal super-parameter combination, retraining the ARIMA model. In this example, the learning rate is set to 0.05 and the regularization parameter is set to 0.5. Assume that 180 days of urine samples and environmental data are present. Using these data and the optimal superparameters, the model training resulted in lower training errors, e.g., from 0.2 to 0.15.

To verify the generalization ability of the model, cross-validation can be used in this case. Assume that the scheme uses 5-fold cross-validation. In 5-fold cross validation, the validation error for each fold is 0.16,0.14,0.15,0.17,0.16, respectively, and the average validation error is 0.156.

The quality of the model is determined using an evaluation index (e.g., AIC, BIC, RMSE, etc.). In this example, the AIC value of the new model is reduced to 108, which is better than any previous round of Bayesian optimization. The old model has an AIC value of 115 and the new model has an AIC value of 108, indicating a better fit of the new model.

Finally, the present case also requires evaluating the model on a separate test set.

Assuming that the test set contains 30 days of urine and environmental data, the error of the model on the test set is 0.14, which is similar to the cross-validation result, and the model has good generalization capability.

S5, dynamically adjusting sampling decision threshold

The sampling decision threshold is dynamically adjusted to make the urine detection of the intelligent toilet more accurate and personalized. The following are examples of how the decision threshold may be dynamically updated using Kalman filtering or state feedback logic:

Kalman filtering method:

initially, an initial threshold is set, such as 0.5.

Assume that in the first week of urine sample detection, the proportion of abnormal data detected by the system is 0.45, which is below the current threshold of 0.5.

The system continues to collect new urine sample data and analyze it with a trained model.

At the second week, the proportion of abnormal data detected by the system increased to 0.55.

Kalman filtering is used to estimate the optimal decision threshold for the next point in time.

The kalman filter algorithm estimates that the new optimal threshold should be 0.52 based on historical data and current observations.

The new optimal threshold is compared with the current threshold and adjustments are made if necessary.

Since the new optimal threshold value of 0.52 is close to the current threshold value of 0.5, the system decides to fine tune, setting the threshold value to 0.52.

State feedback logic:

and calculating the error between the current threshold value and the actual detection result.

At a threshold of 0.5, the proportion of abnormal samples actually detected is 0.6.

Based on the calculated error, a state feedback logic is used to adjust the threshold.

The error is 0.1 (0.6-0.5) and the state feedback logic suggests that the threshold be raised by the same amount, i.e., the new threshold is 0.6.

In both cases, the threshold is effectively and dynamically adjusted, so that the health state and the needs of the user can be reflected more accurately. This procedure has not only theoretical support, but also demonstrated its feasibility by practical data examples.

S6, AI diagnostic engine:

On the basis of a simple time series model, a more advanced model such as XGBoost or a neural network is selected. It is assumed that in a simple time series model, the average detection accuracy of proteins in urine samples is 85%. After XGBoost is used, the precision is improved to 95%. Features for advanced model training are selected and constructed based on urine and environmental data.

In addition to basic proteins, sugars, etc., interactive terms and polynomial features, such as protein interactions with temperature, are added. Training is performed using the selected features and the advanced model. Training was performed using 80% of the data, the remaining 20% being used for validation. The model reached 97% accuracy on the validation set.

Generating an interpretability score:

SHAP (SHAPLEY ADDITIVE exPlanations) or the like is selected to interpret the prediction results of the advanced model. SHAP values show that protein and sugar are the most important features affecting urine detection results. An interpretability score is generated for each prediction result.

In the case where one urine sample is predicted to be abnormal, the SHAP value shows a protein effect of +0.5, a sugar effect of +0.3, and a temperature effect of-0.1. These interpretations are then used to further optimize the advanced model. If the SHAP value frequently shows that a feature has less impact on the predicted outcome, the feature may be de-weighted or deleted in the next round of model training.

S7, continuously optimizing a model and an algorithm:

the output of the AI diagnostic engine is used to retrain the time series model and bayesian optimization algorithm. New labels (e.g., chapter women are diagnosed with a high risk of diabetes) and diagnostic results are also used to further optimize the model.

Continuous optimization of models and algorithms:

The output of the AI diagnostic engine is used to add a new tag to the urine sample. For example, if a chapter woman is diagnosed with a high risk of diabetes, this information is added as a new label. The original model has only two labels of normal and abnormal. More finely divided tags for "high risk of diabetes", "proteinuria" and the like are now added. Combining the new label with the original urine and environmental data, normalizing and cleaning.

For example, there are 1000 new samples, 20 of which are marked as "high risk of diabetes". The time series model is retrained using the new tags and data.

After the new tag was introduced, the F1 score of the model was raised from 0.8 to 0.85.

Optimization of a Bayesian optimization algorithm:

Depending on the new labels and diagnostic results, it may be necessary to adjust the search range of the bayesian optimization algorithm or add new constraints. The search range of the original learning rate is 0.001 to 0.1, and is now adjusted to 0.005 to 0.08. Bayesian optimization is re-performed using the new search scope and constraints.

Within the new search range, a new optimal solution with a learning rate of 0.06 and a regularization parameter of 0.4 is found. The found optimal parameters are used for retraining of the time series model. The F1 score of the model was further improved to 0.9 using the new optimal parameters. Through the steps, not only is the time sequence model optimized, but also a Bayesian optimization algorithm is more suitable for newly-appearing labels and diagnosis results.

Feasibility and actual data:

Suppose that the following urine sample data (glucose content) have been collected for several days:

Day1:0.1

Day2:0.12

day3:0.18 (marked as abnormal)

Day4:0.11

The optimal model prediction is obtained through pretreatment and optimization, and then the health condition of the chapter and woman is further analyzed. Since Day 3's data is marked as anomalous, the sampling decision threshold of the system may be adjusted to a more sensitive level, thereby more closely monitoring the health of the chapter women.

Specific example 9: advanced optimization of intelligent toilet based urine detection and analysis system:

dynamic learning rate and state feedback control logic:

S1, dynamic learning rate adjustment, wherein the system adopts a preset learning rate scheduling algorithm, specifically an exponential decay algorithm, so as to gradually optimize the performance of the model. In the initial phase of system operation, the model uses a relatively high initial learning rate of 0.1 for urine data analysis. This high learning rate helps the model converge quickly and captures the main features of the urine sample in a short time.

After the first 5 urine sample analyses, the accuracy of the model has reached a satisfactory 90%. But to avoid overcommitted and further optimize the model, the learning rate scheduling algorithm of the system automatically reduces the learning rate from 0.1 to 0.05. This adjustment is based on the behavior of the model on the first 5 samples and is made according to a preset decay factor.

This reduced learning rate not only helps the model capture subtle patterns in urine samples more finely, but also reduces the risk of the model producing overfitting on future samples. At the same time, the system also sets an early-stop mechanism to monitor the performance of the model on the validation set. If the accuracy of the model is not significantly improved in 10 consecutive iterations, the system will automatically stop training the model to preserve the optimal model weights.

S2, gradient shearing and momentum parameters, namely, besides dynamic learning rate adjustment, are skillfully combined in an optimizer so as to ensure the stability and high-efficiency performance of the model.

In order to prevent the problem of gradient explosions, the system is provided with a gradient threshold of 5. This means that during each training iteration, if the calculated gradient value exceeds this threshold value, it will be clipped to 5. For example, in a particular training iteration, if the model calculates a gradient value of 8 when analyzing urine data, then the gradient value is clipped to 5 to prevent model weight updates from being too severe, resulting in an unstable training process.

Meanwhile, in order to make the model weight update smoother, the system also introduces a momentum parameter, which is set to be 0.9. Momentum considers not only the current gradient but also the previous gradient, thereby smoothly adjusting the model weights. This momentum effect helps the model converge faster and reduces oscillations during training. These two techniques: the combined action of gradient shearing and momentum parameters effectively prevents the problems of gradient explosion and weight vibration possibly occurring in the urine data analysis process.

S3, the state feedback control logic plays a key role, particularly in dynamic adjustment of data sampling frequency and other parameters. This module employs simulated annealing or genetic algorithms to automatically optimize three parameters of a PID (proportional-integral-derivative) controller: kp (proportionality constant), ki (integration constant), and Kd (differentiation constant).

Taking a simulated annealing algorithm as an example:

the initial Kp, ki, kd is set to 1,0.1,0.01. This is based on experience or some heuristic approach.

This function is typically based on a system performance indicator, such as in a smart toilet, which may be "complete urine detection in the shortest time as accurately as possible".

At each step, the algorithm will try to change the values of Kp, ki, kd. These new values will be evaluated according to the objective function. If the new values are better (or with some probability better) than the old values, they will replace the old values. This process is repeated until the system reaches a certain stop condition (e.g., number of iterations or temperature threshold). The resulting Kp, ki, kd will be used in the PID controller.

Examples: it is assumed that after 50 iterations of the simulated annealing algorithm, the values of Kp, ki, kd become 1.2,0.12,0.015. The new set of parameters improves the objective function (e.g. accuracy and response speed, etc.) compared to the initial setting.

Therefore, by automatically optimizing the values of Kp, ki and Kd, the state feedback control logic module not only improves the accuracy of the intelligent closestool urine detection system, but also improves the operation efficiency of the intelligent closestool urine detection system. This proves to be very feasible for state feedback control logic optimization using simulated annealing or genetic algorithms.

S4, noise reduction and delay compensation, which are critical, particularly when the system relies on real-time or near real-time data to make decisions.

Prior to the state feedback logic, the system reduces noise by a second order low pass filter. This is done to reduce data errors due to environmental factors such as temperature fluctuations or electromagnetic interference.

The design and parameters of the filter will be set depending on the desired noise reduction level. Once the design is complete, all data entering the state feedback logic passes through the filter.

The performance of the filter is quantified by comparison with the raw data.

Practical application: it is assumed that the original state feedback data contains about 10% noise. After processing by the second-order low-pass filter, the noise level is reduced to 2%. This reduction may have a significant performance improvement in subsequent urine analysis.

Delay compensation:

Delay compensation algorithms are designed to solve problems caused by communication or computation delays. This is typically accomplished by adding a time stamp within the state feedback logic or using other advanced time synchronization methods. Once the design is complete, the algorithm is integrated into the state feedback logic to compensate for the delay in real time. Similarly, it is also desirable to quantify the impact of this compensation measure on system performance. Through the noise reduction and delay compensation measures, the urine detection analysis system of the intelligent closestool not only can perform data analysis more accurately, but also can cope with various running environments and conditions more flexibly.

Synchronization and update mechanism of data and model:

1. data synchronization is a key ring to ensure accuracy and reliability of models. The following is how data synchronization is achieved, and its feasibility in practical applications.

When collecting urine samples and environmental data (e.g., temperature, humidity), a time stamp is attached to each data point. All data passes through a preprocessing module before entering the time series model. The module will check the time stamps of the urine sample and environmental data to ensure that they are collected at the same or close points in time.

If the time stamps of the two sets of data differ too much (e.g., more than 0.5 seconds), the preprocessing module may discard the data or mark them as "needed review" to prevent errors in the model due to unsynchronized data.

Once the data synchronization is complete, they are fed into a time series model for further analysis and diagnosis.

Examples: the following two sets of data are assumed:

Urine sample data: timestamp 08:30:00.500

Environmental data (temperature, humidity): timestamp 08:30:00.200

Since the difference between the two timestamps is only 0.3 seconds, below the preset 0.5 second threshold, the two sets of data are considered to be synchronized.

The preprocessing module may further examine other criteria and then send the synchronized data to the time series model for analysis. If the difference between the time stamps exceeds 0.5 seconds, for example:

Urine sample data: timestamp 08:35:00.100

Environmental data (temperature, humidity): timestamp 08:35:01.000

Because of the 0.9 second difference, exceeding the 0.5 second threshold, the preprocessing module will either flag the two sets of data or choose to discard them to prevent errors in subsequent analysis. By the method, the urine detection and analysis system of the intelligent closestool ensures high synchronization of data, so that accuracy and reliability of a model are improved.

2. Priority queue handling-a method of efficiently handling tasks of different priorities is to use priority queues. This ensures that more urgent or important data samples can be processed preferentially.

Priority queue processing step:

As urine sample data passes through the preprocessing module, they are classified and labeled according to certain predefined rules or indicators. For example, if the sugar exceeds a certain threshold, the sample may be marked as "high priority". The data samples enter the corresponding queues according to their priorities. Normal samples enter the normal queue, while high priority samples enter the priority queue.

The AI diagnostic engine will first check the priority queue. Only when the priority queue is empty will data be fetched from the regular queue for analysis. For high priority tasks, the system will complete the analysis as short as possible and feed back the results to the user or medical personnel quickly.

Examples: suppose we have the following two urine samples:

Sample a: the sugar is normal, the priority is "low", and the processing time is expected to be 10 seconds.

Sample B: the sugar is high, the priority is "high", and the processing time is expected to be 5 seconds.

First, both sample a and sample B enter the preprocessing module for data cleansing and time stamp synchronization. They are then sent to the corresponding queues.

Sample a enters the regular queue.

Sample B is marked as "high priority" and enters the priority queue due to the high sugar.

The AI diagnostic engine first examines the priority queue to find sample B therein and therefore performs the analysis preferentially. Even if sample a arrives earlier than sample B, it will be processed earlier because of the higher priority of B. In practice, the priority queue ensures that samples requiring urgent processing can be analyzed quickly, which is particularly important in health diagnostics and emergency medical events.

3. On-line model update, which is an important function, ensures that the system can be adjusted in time according to new data, thereby improving diagnosis accuracy and efficiency. The following are detailed steps and one example of implementing online model updates.

Updating an online model:

The system will continue to acquire new data from the urine detection analysis of the intelligent toilet. Such data may come from predetermined tests or actual use. A background process monitors the performance metrics of the model, such as accuracy, recall, etc., in real time. When these indicators are below a certain threshold or a preset update condition is reached (e.g., every 1000 new samples collected), a model update is triggered. The new data is preprocessed and feature extracted as the old data.

Fine tuning of the model is performed using the new data. This typically involves iterative optimization based on existing model weights. New model weights will replace old weights. This step is automated and the system does not need to be shut down.

The new model is evaluated using a validation set to ensure performance at least comparable to or better than the old model. All update procedures are recorded and monitored for subsequent analysis and improvement.

The example assumes that the system has been running for a period of time and 1000 new urine samples have been collected. The current accuracy of the model is 90% and the trigger condition we set is one update every 1000 new samples.

1. Triggering model updating: the system automatically triggers the update procedure as a sufficient number of new samples have been collected.

2. Data preprocessing: the new 1000 samples were subjected to data cleaning and feature extraction, ready for model fine tuning.

3. Fine tuning of the model: based on the current model weight, new 1000 samples are used for iterative optimization.

4. And (5) weight updating: new weights are calculated after fine tuning and used to replace old model weights.

5. And (3) verification: on a separate validation set, the accuracy of the new model was estimated to be 92%.

6. And (3) line feeding: because the new model has better performance, the system can automatically adopt new weight, and the whole process is completed within 3 seconds without stopping the machine. The online model updating mechanism is not only feasible, but also highly automated and optimized, and can ensure that the system always operates in an optimal state.

The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims

1. The urine detection analysis system of the intelligent closestool is characterized by adopting a real-time multi-index analysis module, wherein the module carries out real-time quantification on various biological markers in urine through a miniaturized biological sensor and is combined with an adaptive sampling technology, and the sampling technology can automatically adjust the urine sampling frequency according to the physiological state and behavior of a user; the system further integrates a non-contact detection module, adopts a non-contact mode of optics or electromagnetic field to collect urine samples, and carries out micro-processing and detection of the samples through an integrated micro-fluidic chip; the system also includes an AI diagnostic engine that uses a machine learning algorithm to diagnose and predict urine analysis data in real time and can generate personalized health reports;

The self-adaptive sampling technology is driven by a dynamic frequency adjustment algorithm based on reinforcement learning, the algorithm adjusts sampling frequency in real time and is combined with a multi-source data fusion module, and the fusion module fuses the data of the environmental sensor with urine detection data by using a Kalman filtering algorithm; the system further adopts a time sequence analysis and sampling decision threshold adjustment mechanism to dynamically adjust the threshold of the sampling decision according to the output result of the AI diagnosis engine and the predicted urine index trend; the system includes a context-aware mechanism that uses an embedded camera or infrared sensor to detect if a user is in a state suitable for sampling.

2. The urine detection and analysis system for intelligent toilets according to claim 1, wherein said personalized health report is generated based on historical data and health status of each user;

The system is also provided with a multi-mode user interaction interface, comprising a voice, a touch screen and a mobile phone APP, and is provided with a data security and privacy protection mechanism, wherein the mechanism adopts an end-to-end encryption and blockchain technology to ensure the security and privacy of user data; the system further has a remote medical treatment integration function, and urine analysis data are safely transmitted to a remote medical expert for further analysis and diagnosis; the system also uses a disposable sensor made of biodegradable material;

3. The urine detection and analysis system of the intelligent closestool according to claim 1, wherein the integrated non-contact detection module is driven by an optical calibration subsystem with a bimodal optical sensor, two lasers with different wavelengths are used for respectively detecting protein and sugar in urine, and a high-frequency electric field pulse shape optimized by a machine learning algorithm is adopted for improving the measurement accuracy of conductivity and pH value; the module further comprises an instant data checking mechanism, wherein the instant data checking mechanism is used for carrying out real-time data comparison through the redundant sensor, and a local data cache is used for quick local analysis; the non-contact detection module is also provided with a sensor thermal management subsystem, the optimal working temperature of the sensor is maintained by using a liquid cooling or air cooling mode, the position and the angle of the sensor are adjusted in real time through a micro-electro-mechanical system (MEMS) accelerometer and a gyroscope, and the sensor is changed based on the physical state of a user;

4. The urine detection and analysis system of the intelligent closestool according to claim 1, wherein the AI diagnosis engine firstly acquires biochemical data of a urine sample through a non-contact detection module and performs pretreatment, and then performs feature extraction by using a multi-layer CNN, in particular to key feature identification of high-dimensional data; meanwhile, an LSTM module is embedded and used for processing time series analysis of urine samples, in particular for progress tracking of chronic diseases; model updating is carried out on the premise of ensuring user privacy by using a federal learning mechanism;

5. The urine detection and analysis system of claim 4, wherein the AI diagnosis engine performs feature extraction by a multi-layer Convolutional Neural Network (CNN); this process is driven by the formula f (x) =relu (W x+b) for identifying key features from biochemical data of urine samples; where f (x) is the feature map, W is the convolution kernel, x represents the convolution operation, b is the bias term, and ReLU represents the activation function;

The formula ct=f _t⊙ct^t-1+i_t⊙tanh(W·[h_t-1,x_t ] +b) is used to track the progression of chronic disease;

Its update rule is represented by formulaDefinition;

6. The urine detection and analysis system of claim 3, wherein the multi-source data fusion module utilizes a kalman filter algorithm to fuse the data received from the environmental sensor and the urine detection module in real time; wherein the Kalman filtering algorithm dynamically changes the covariance matrix through adaptive covariance adjustment; specifically, covariance matrixWhere α is an adjustable factor, var (x _t) is the variance of the last N data points, thereby enabling adaptive processing of data uncertainty;

the system also adopts a self-adaptive weight distribution mechanism to dynamically adjust the Kalman gain according to the reliability of urine detection and environmental data, and the specific formula is as follows:

7. A method for the urine detection analysis system of the intelligent toilet of any one of claims 1-6, the method comprising the steps of performing high-speed data caching and query through an embedded database, extracting key features from original urine sample data by using a feature extraction algorithm, and then enhancing the accuracy of a time series analysis model by seasonal and trend decomposition by using an STL algorithm;

The sequence analysis model is trained and optimized through an automatic super-parameter optimization algorithm by using Bayesian optimization; after model training, real-time interpretability scoring is performed by using an AI diagnosis engine and adopting a SHAP method; based on the output result of the AI diagnosis engine and the confidence interval of the model, calculating to obtain a new sampling decision threshold value, optimizing the threshold value through a dynamic learning rate and state feedback control logic, and carrying out sliding average or exponential smoothing post-processing; the newly calculated sampling decision threshold is finally applied to the adaptive sampling module to achieve more accurate and personalized urine detection.

8. The method of claim 7, wherein the automatic super-parameter tuning algorithm comprises a data preprocessing module for collecting and storing urine samples and environmental sensor data at regular time intervals, and performing data cleaning, outlier detection and normalization processing simultaneously; initializing a basic time sequence model, and dynamically optimizing model super-parameters through a search range and initial seed points of a Bayesian optimization algorithm; after model training and verification, updating a time sequence model by applying an optimal super-parameter combination, and combining model output and a predicted confidence interval;

9. The method of claim 7, wherein the dynamic learning rate and state feedback control logic comprises implementing dynamic adjustment of learning rate in an optimizer using momentum parameters for directional optimization using specific learning rate scheduling algorithms and gradient clipping techniques;