Data collaboration system and method supporting individualized baseline modeling algorithm

Abstract

The invention discloses a data collaboration system and method for supporting an individualized baseline modeling algorithm, and the system comprises a multi-scene baseline management module which is used for storing and switching baseline parameters of different use environments; the group collaborative calibration module is used for summarizing the aging data of the multiple devices and generating compensation parameters; the cold start acceleration module is used for quickly establishing an initial baseline through transfer learning by utilizing an existing data source; the cross-device baseline unification module is used for fusing baseline parameters of multiple devices to realize judgment consistency; the data cooperative communication module is used for realizing data transmission required by the modules; and the privacy protection module is used for encryption and differential privacy processing. According to the method, the technical bottleneck of a single-machine individualized baseline modeling algorithm in a multi-scene, long-period use and multi-equipment scene is solved, the cold start time is shortened, and the equipment aging misjudgment rate is reduced. The method can be applied to individualized health monitoring equipment such as intelligent water cups, water purifiers and glucometers.

Description

A data collaboration system and method supporting individualized baseline modeling algorithms Technical Field

[0001] This invention relates to the field of data management and sensor calibration for intelligent testing equipment, and particularly to a data collaboration system and method that supports individualized baseline modeling algorithms. Background Technology

[0002] In recent years, with the increasing health awareness of the public and the development of Internet of Things (IoT) technology, personalized health monitoring devices have been widely used. The applicant's earlier patent application proposed a personalized baseline modeling method that eliminates the need for a large pre-established material database. It can identify unknown pollutants simply by causing spectral characteristics to deviate from the baseline, and can automatically adapt to the user's drinking habits and regional water quality differences.

[0003] However, existing personalized baseline modeling methods face the following technical bottlenecks in practical applications: When users change their usage scenarios, the original baseline becomes invalid, requiring users to re-collect samples in the new environment to establish a new baseline. During the initial 3-5 days in the new environment, the system cannot provide reliable judgments, and some users abandon the device during this waiting period. When a new device is used for the first time, it requires collecting N≥20 samples to establish reliable baseline parameters, requiring a 3-5 day cold start period. During this period, the system cannot provide reliable anomaly judgments, leading to long waiting times and user skepticism about the product's value, resulting in a new user churn rate as high as 35%. The detection equipment ages during use, leading to a decrease in detection accuracy. Furthermore, with the popularization of smart homes, users often own multiple similar devices: a primary water cup at home, a spare water cup in the office, and a portable model for travel. Each device establishes an independent baseline, resulting in inconsistent judgments for the same cup of water on different devices, severely impacting user trust. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems of difficult baseline migration, slow cold start, lack of equipment aging compensation, and inconsistent cross-device judgment in existing individualized baseline modeling methods. It provides a data collaboration system and method that supports individualized baseline modeling algorithms. Through minimum necessary data transmission and privacy protection mechanisms, it achieves automatic baseline switching across multiple scenarios, enabling seamless baseline migration when users change environments. It employs cold start acceleration, migrating the initial baseline from the cloud when a new device is used, requiring only 5-10 sample collections before it is ready for use. It provides real-time compensation for equipment aging through group collaborative calibration, eliminating the need for factory calibration. When a user has multiple devices, it unifies the baseline across devices, allowing multiple devices to share a unified baseline.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: Firstly, the present invention provides a data collaboration system supporting an individualized baseline modeling algorithm. The individualized baseline modeling algorithm establishes an individualized baseline model based on the low-dimensional discrete feature vector of the target object. The low-dimensional discrete feature vector is composed of the ratio and difference features of a set band pair. It includes one or more of the following modules: a multi-scenario baseline management module, used to identify user scenarios and store baseline parameters for different usage environments. When a scenario change is detected, it switches to the baseline parameters of the corresponding scenario and uses a weighted fusion mechanism to gradually adjust the weights of the old and new scenario baseline parameters within a preset sampling period. If a new scenario is identified... The system includes: a cold start acceleration module; a group collaborative calibration ...

[0006] As a further implementation, the data collaborative communication module executes desensitization processing logic when transmitting data. The terminal device only uploads statistical features that have been locally aggregated to the cloud. These statistical features include the baseline center vector, covariance matrix, or anomaly deviation. Uploading unprocessed raw spectral response data is prohibited to achieve data privacy protection and regulatory compliance.

[0007] As a further implementation, scene recognition in the multi-scenario baseline management module is based on location information, specifically including: the device acquiring the current GPS coordinates; converting the GPS coordinates into grid numbers with a preset precision; and matching the grid numbers with scene IDs stored on the cloud server.

[0008] As a further implementation method, the initial baseline acquisition and individualized fine-tuning in the cold start acceleration module includes: querying anonymized baseline datasets of the same region and water source type from the cloud server based on the current location information of the device and the water source type; performing statistical analysis on the baseline datasets to calculate the median or weighted average to obtain the initial baseline parameters; collecting M current environmental samples to calculate the local baseline parameters; and fusing the initial baseline with the local baseline through a weighted fusion algorithm.

[0009] As a further implementation method, the aging compensation model establishment method of the group collaborative calibration module includes: each device periodically uploads aging data packets, which include: device serial number, sensor batch number, usage time, wavelength drift, and baseline center change; the cloud server groups the data according to the sensor batch number; regression analysis is performed on the data of each batch to fit the aging compensation model, the covariance matrix correction is calculated, and the aging compensation parameters are sent to all devices in the corresponding batch.

[0010] As a further implementation method, the fusion calculation method of the cross-device baseline unification module includes: weighted averaging of the baseline centers of each device. ;in Let be the sample size of the i-th device; the larger the sample size, the higher the weight. The baseline center of the i-th device is used; the covariance matrices of each device are pooled and fused: .

[0011] As a further implementation, the encryption methods of the privacy protection module include: encrypting the baseline parameters using the AES-256 algorithm, with the encryption key derived from the user's biometrics or the device's unique identifier; adding Laplace noise to the uploaded aging data; and downgrading the location information by reducing the precise GPS coordinates to grid numbers before uploading.

[0012] Secondly, this invention also provides a data collaboration method supporting an individualized baseline modeling algorithm. The individualized baseline modeling algorithm establishes an individualized baseline model based on a low-dimensional discrete feature vector of the target object. The low-dimensional discrete feature vector is composed of ratio and difference features of a set band pair. The method includes one or more of the following steps: a multi-scene baseline management step: when the device detects that the deviation of the current sample exceeds the scene switching threshold, a scene recognition process is triggered. Based on the scene recognition result, the baseline parameters of the corresponding scene are obtained from the data storage location and the currently used baseline is replaced. If a new scene is identified, a cold start acceleration step is triggered. A cold start acceleration step: when the device is used for the first time or enters a new scene, initial baseline parameters are obtained from an existing data source; a small number of current environment samples are collected, with the sample size M being less than the sample size N required by traditional methods; the initial baseline parameters are then... The baseline parameters are fused with the small number of samples to generate baseline parameters adapted to the current user; Group collaborative calibration steps: Multiple devices periodically upload aging-related data to a centralized data coordination node, which aggregates the aging-related data and establishes an aging compensation model; aging compensation parameters for each device are calculated based on the aging compensation model; the aging compensation parameters are distributed to each device, and each device uses the compensation parameters to correct its local baseline parameters; Cross-device baseline unification steps: When a user has multiple devices, their respective baseline parameters are obtained from each device; the multiple baseline parameters are fused and calculated to generate unified baseline parameters; the unified baseline parameters are synchronized to each device; During the execution of the above steps, the transmitted data is encrypted and differentially private; ensuring that it is impossible to infer the original information of a single user from the transmitted data.

[0013] Thirdly, the present invention also provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the data collaboration system supporting the individualized baseline modeling algorithm.

[0014] Fourthly, the present invention also provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute the data collaboration system that supports the individualized baseline modeling algorithm.

[0015] The beneficial effects of the above invention are as follows: Addressing the difficulty of baseline migration, this invention proposes a multi-scenario baseline management module. This module stores baseline parameters from multiple user scenarios. When an abnormally high deviation of the current sample is detected, the scenario recognition process is automatically triggered. Based on the recognition result, the corresponding scenario's baseline is retrieved from the storage location and replaced with the local baseline. No manual user operation is required; the system automatically recognizes scenario changes. When first entering another scenario, a baseline is established in just one day, and subsequent entry into the same scenario achieves zero-second switching, improving efficiency.

[0016] This invention addresses the slow cold start problem by proposing a cold start acceleration module. Although new devices lack historical user data, they can utilize anonymized data from other users as prior knowledge. Based on the new user's location information and water source type, the module queries the cloud for the baseline distribution of other users in the region, calculates the median as the initial baseline, and collects a small number of samples. Through Bayesian updates or weighted fusion, the initial baseline is fine-tuned to suit the new device. This enables rapid activation of new devices, reducing the activation time from 5 days to 1 day.

[0017] This invention addresses the issue of inadequate compensation for equipment aging by proposing a collaborative calibration module. This module aggregates data from multiple devices, fits an aging compensation model through regression analysis, and distributes compensation parameters to each device. The devices then correct the wavelength center in real time during testing. Cloud-based automatic calibration reduces factory repair costs, and compensation parameters are updated weekly to maintain continuous testing accuracy.

[0018] This invention addresses the issue of inconsistent judgments across devices by proposing a cross-device baseline unification module. Multiple devices of a user upload their respective baseline parameters, which are then weighted and fused in the cloud to generate unified baseline parameters, which are then synchronized to all devices. Subsequently, the judgment results for the same glass of water are consistent across all three devices. This eliminates the potential for inconsistent judgments for the same glass of water, improving the user experience. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative examples and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention.

[0020] Figure 1 is an overall flowchart of the data collaboration system according to some specific implementations of the present invention; Figure 2 is a schematic diagram of the multi-scenario baseline management module according to some specific implementations of the present invention; Figure 3 is a schematic diagram of the group collaborative calibration module according to some specific implementations of the present invention.

[0021] Figure 4 is a layered architecture diagram of the privacy protection module according to some specific implementations of the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0023] This embodiment discloses a data collaboration system supporting an individualized baseline modeling algorithm. The individualized baseline modeling algorithm establishes an individualized baseline model based on the low-dimensional discrete feature vector of the target object. The low-dimensional discrete feature vector is composed of ratio and difference features of a set band pair. As shown in Figure 1, it includes one or more of the following modules: a multi-scenario baseline management module, used to identify user scenarios and store baseline parameters for different usage environments. When a scenario change is detected, it switches to the baseline parameters of the corresponding scenario and uses a weighted fusion mechanism to gradually adjust the weights of the old and new scenario baseline parameters within a preset sampling period. If a new scenario is identified, a cold start acceleration module is triggered. The system comprises the following modules: a cold start acceleration module, used to quickly establish an initial baseline through transfer learning when the device is used for the first time or enters a new scenario, fine-tuning the initial baseline with a small number of samples, and generating baseline parameters adapted to the current user through a weighted fusion algorithm; a group collaborative calibration module, used to aggregate aging data from multiple devices and generate compensation parameters to correct local baseline parameters; a cross-device baseline unification module, used to fuse baseline parameters from various devices to generate unified baseline parameters and synchronize the unified baseline parameters to each device; a data collaborative communication module, which enables data transmission required by the above modules; and a privacy protection module, which includes encrypting the baseline parameters during transmission.

[0024] At least two of the following modules described in this disclosure—the multi-scenario baseline management module, the group collaborative calibration module, the cold start acceleration module, and the cross-device baseline unification module—work collaboratively.

[0025] S1: Multi-scenario baseline management module, used to identify user scenarios and store baseline parameters for different usage environments. When a scenario change is detected, it switches to the baseline parameters of the corresponding scenario and uses a weighted fusion mechanism to gradually adjust the weights of the old and new scenario baseline parameters within a preset sampling period. If a new scenario is identified, the cold start acceleration module is triggered.

[0026] In one embodiment, the user is Mr. Zhang, male, 35 years old, residing in Haidian District, Beijing. He has been using smart water cup device A at home to test drinking water for 6 months, and a stable individualized baseline model of "Beijing-Haidian-wtw3bk" has been established. ,in, This indicates the baseline center of the individualized baseline model. This represents the covariance matrix of the individualized baseline model. This indicates the number of times the baseline center of the individualized baseline model was sampled. One day, Mr. Zhang traveled to Huangpu District, Shanghai on business and checked into a hotel.

[0027] S11: Zhang arrives at the Shanghai hotel and turns on the tap in his room to get water. As shown in Figure 2, he uses smart water cup device A to perform a test: Smart water cup device A collects the spectral data of the current liquid and obtains a low-dimensional discrete feature vector. Read the baseline center from the local baseline, i.e., the individualized baseline model of "Beijing-Haidian-wtw3bk": ; Calculate the Mahalanobis distance between the current liquid's low-dimensional discrete eigenvector and the baseline center in the individualized baseline model: =8.2; Determined based on Mahalanobis distance: , The scene switching threshold is triggered, and the multi-scene baseline management module is activated. The smart water cup device A prompts the user on the display screen: "The current liquid characteristics are significantly different from those of the usual scenes. It may be in a new environment. It is automatically adapting..."

[0028] The scene switching threshold It is dynamically determined based on a comprehensive volatility index derived from an individualized baseline model, for example... Where k is an empirical coefficient, ranging from 3 to 5, to ensure that scene switching is not triggered erroneously within the normal fluctuation range. This is the overall volatility index for the individualized baseline model.

[0029] S12: Smart water bottle device A executes a scene recognition mechanism, prioritizing GPS positioning: Smart water bottle device A obtains its current coordinates via GPS module; converts the coordinates to Geohash encoding: wtw3yz; queries the local cache to see if a baseline with a scene ID containing "wtw3yz" exists. If not, it is determined to be a new scene; scene recognition can also choose to perform a cloud query. Smart water bottle device A connects to the cloud server via 4G / Wi-Fi; uploads a request to query the cloud database: the historical scene list of user_123456 includes: Scene 1: "Beijing-Haidian-wtw3bk" baseline; Scene 2: "Beijing-Chaoyang-wtw3cd" baseline; query result: no matching Shanghai scene, so it is determined to be a new scene; scene recognition can also choose to allow manual user input confirmation. Smart water bottle device A pops up a prompt: "Are you in the following scene? A. Home B. Office C. Other"; the user selects "C. Other"; smart water bottle device A further asks: "Is this a new scene?"; the user selects "Yes".

[0030] Since the scenario was identified as new, the system triggered the cold start acceleration module. It performed transfer learning via the cloud, querying the corresponding anonymized baseline dataset in the database, filtering and calculating the dataset to obtain the initial baseline. The initial baseline was then downloaded to the device and labeled "Shanghai-Huangpu-Temporary Baseline," with the baseline center of the temporary baseline defined. .

[0031] S13: Smart water cup device A used a "Shanghai-Huangpu-Temporary Baseline" and made minor adjustments. During one day, Mr. Zhang drank water normally at the hotel. Smart water cup device A collected 5 samples using the temporary baseline: V1=[0.44, 0.31, 0.18, 0.63, 0.28, 0.51, 0.39, 0.41]; V2=[0.43, 0.29, 0.20, 0.62, 0.27, 0.53, 0.37, 0.40]; V3=[0.45, 0.32, 0.19, 0.61, 0.28, 0.52, 0.38, 0.42]; V4=[0.42, 0.30, 0.18, 0.60, 0.26, 0.51, 0.39, 0.39]; V5=[0.44, 0.31, 0.19, 0.62, 0.27, 0.52, 0.38, 0.41].

[0032] Calculate the temporary baseline center: V 0_local =(1 / 5)ΣV i =[0.436, 0.306, 0.188, 0.616, 0.272, 0.518, 0.382, 0.406], and covariance matrix Σ local .

[0033] Using a weighted fusion algorithm, such as Bayesian update α=0.7, baseline parameters adapted to the current user are generated: baseline center V. 0_new = 0.7×V 0_init + 0.3×V 0_local = 0.7×[0.42,...] + 0.3×[0.436,...]= [0.4248,0.3018, 0.1894, 0.6118, 0.2716, 0.5154, 0.3814, 0.4018];covariance matrix Σ new = 0.8×Σ init + 0.2×Σ local .

[0034] S14: One day later, the new individualized baseline model is activated, and the smart water cup device A begins to use the activated individualized baseline model to make anomaly judgments. At the same time, the smart water cup device A uploads the activated individualized baseline model to the cloud storage via encrypted transmission and saves it to the scene 3: "Shanghai-Huangpu-wtw3yz" individualized baseline model.

[0035] When the user returns to their home in Beijing, and the system detects again that the Mahalanobis distance between the feature vector and the individualized baseline of Scene 1 "Beijing-Haidian-wtw3bk" is below the threshold, it will automatically and seamlessly switch back to the original baseline and treat this switching behavior as positive feedback to enhance the accuracy of scene recognition.

[0036] S2: Cold start acceleration module, used to quickly establish an initial baseline through transfer learning when the device is used for the first time or enters a new scenario, fine-tuning the initial baseline with a small number of samples, and generating baseline parameters adapted to the current user through a weighted fusion algorithm; In one embodiment, user Wang, female, 28 years old, residing in Wuhou District, Chengdu, purchased a new device B to test tap water.

[0037] S21: When Wang first turned on the device, he connected to the mobile app via Bluetooth and completed the following settings: User registration: Create an account; Location authorization: The app obtains GPS coordinates and gets the Geohash code: wm6n; Water source type: Select tap water.

[0038] S22: The device uploads a cold start request, the cloud server performs transfer learning, and queries the database: queries the Chengdu Wuhou District, the water supply baseline model within an area of ​​about 10km×10km, matches the historical baseline data of 200 users, and performs high-quality baseline filtering to obtain the 80 users who are currently using the water stably.

[0039] The initial baseline was calculated, and the baseline centers and median of the covariance matrix of 80 users were obtained. These were used as the baseline centers V of the initial baseline. 0_init The covariance matrix Σ init Based on the sample size and CV calculation, the confidence level is calculated to be 0.85, which indicates the reliability of the initial baseline.

[0040] S23: The cloud downloads the initial baseline to the new device B. The device displays a prompt: "The initial baseline has been prepared for you. Please drink water 5-10 times to complete the personalized adaptation."

[0041] S24: Perform local fine-tuning. Within one day, Mr. Wang used the new device B for normal drinking water. The new device B collected 5 samples. Calculate the local baseline center: V 0_local =(1 / 5)ΣV i =[0.436, 0.306, 0.188, 0.616, 0.272, 0.518, 0.382, 0.406]; and covariance matrix Σ local .

[0042] Using a weighted fusion algorithm, such as Bayesian update, baseline parameters adapted to the current user are generated. The weight coefficient α is dynamically adjusted based on the confidence level: α = 0.5 + 0.3 × confidence = 0.5 + 0.3 × 0.85 = 0.755; V 0_adapted =α×V 0_init + (1-α)×V 0_local ;Σ adapted =0.8×Σ init + 0.2×Σ local .

[0043] S25: After 5 samplings, the system determines that the baseline is usable, provided that N ≥ 5 and the standard deviation of the local samples < the threshold; New device B uses V 0_adapted As a formal baseline, anomaly detection is performed. At the same time, the new device B uploads the baseline model to the cloud storage via encrypted transmission, saving the scenario 1: "Chengdu-Wuhou-wm6n" baseline.

[0044] A comparative experiment was conducted between the cold start method of transfer learning disclosed in this disclosure and the traditional method. N=200 new users were randomly divided into two groups. The technical effects are shown in Table 1 below. The cold start time of the new device disclosed in this disclosure decreased by 74%; the new user retention rate increased by 42%; and the misjudgment rate in the first week was slightly improved. This is because the transfer baseline disclosed in this disclosure provides good prior knowledge.

[0045] Table 1. Comparison of Cold Start Acceleration Effects | Group | Cold Start Method | Activation Time | First Week False Positive Rate | New User 7-Day Retention Rate | |---|---|---|---|---| | Control Group | Traditional Method (20 samples collected) | 4.2 days | 22% | 65% | | Experimental Group | This Invention Transfer Learning (5 samples collected) | 1.1 days | 18% | 92% Table S3: Group collaborative calibration module, used to summarize aging data of multiple devices and generate compensation parameters to correct local baseline parameters; specifically including: based on the summarized aging-related data of multiple devices, establishing a device aging compensation model, generating aging compensation parameters according to the aging compensation model, and sending the aging compensation parameters to the corresponding terminal devices through the data collaborative communication module. After receiving the aging compensation parameters, each device uses them to correct its local baseline parameters and compensate for baseline distortion caused by sensor aging.

[0046] In one embodiment, as shown in Figure 3, a batch of MEMS-FPI filters, batch number "202601-A", was produced simultaneously. A total of 5000 chips were manufactured and integrated into 5000 smart water bottles for sale. After six months of use, some users reported through the app that they frequently received false alarms, indicating abnormalities even when the water was normal. The app's judgment was not limited to human feedback but also relied on sensor self-test signals as weighting factors, such as drastic changes in ambient temperature and humidity, and abnormal light source drive current. If the physical self-test signal was normal, it was more likely to be judged as a deviation from the actual state rather than aging. After collecting data, the customer service department found that the false alarm rate of this batch of devices had increased from an initial 15% to 28%.

[0047] S31: Each device performs an aging self-test once a week. The aging self-test is only performed when the device meets preset conditions, such as when the device is idle, has sufficient power (>30%), and more than 168 hours have passed since the last self-test, in order to avoid interfering with normal user use and to ensure data validity.

[0048] In one embodiment, the self-test time is during a period when the user does not usually use the device, such as 2:00 AM every Sunday; the self-test method is as follows: test a standard sample by drawing a sealed standard water bag built into the device to the test window using a micro pump; measure the spectrum of the standard water bag and compare it with the factory calibration value to calculate the wavelength drift and the change in the baseline center.

[0049] S32: Each device uploads aging data to the cloud. If there is no network connection, it is cached locally and uploaded in batches the next time a network connection is available.

[0050] Before measuring the standard water tank, the system first measures the air spectrum for calibration and confirms that the light source intensity is within the normal range to ensure that the calculation of aging data can exclude other interfering factors and enhance the reliability of the aging data.

[0051] S33: Cloud server aggregates data by batch number: Batch: "202601-A"; Number of valid devices: 4850, of which 150 did not upload data due to user not being connected to the network / device failure; Data cleaning: outlier data and short-term usage data with no obvious aging effect were removed; Remaining valid data: 4600 records.

[0052] S34: Fit the aging compensation model and use regression analysis to fit the model to obtain the expected wavelength drift: Δλ=a·T+b·T², where T is the number of days used; a=0.0102 nm / day; b = -0.000018 nm / day²; the model's coefficient of determination is R²=0.87.

[0053] Among them, a·T reflects the gradual contamination of the MEMS mirror surface, such as dust and water vapor, leading to a monotonic drift of the wavelength center; b·T² reflects the self-cleaning effect after the contamination reaches a certain level, such as the surface tension starting to take effect, resulting in a slight decrease in the drift rate. The coefficient of determination index is used to measure the goodness of fit of the model to the data, with a value ranging from 0 to 1. The closer the value is to 1, the better the fit. In this embodiment, R²=0.87 indicates that the quadratic polynomial model can explain 87% of the wavelength drift change, indicating a good fit, and can be used for subsequent compensation parameter calculations.

[0054] S35: Besides wavelength drift, MEMS aging also affects the covariance matrix Σ of baseline parameters. Principal component analysis was performed in the cloud: the cloud input consisted of ΔV0 vectors from 4600 devices; the first principal component (PC1) contributed 82% of the variance, corresponding to the fourth feature dimension, i.e., the 1450nm band; the second principal component (PC2) contributed 12% of the variance, corresponding to the second feature dimension, i.e., the 1200nm band; the remaining principal components contributed <6% of the variance. Aging mainly affects the 1450nm and 1200nm bands, which correspond to the absorption peaks of water molecules, and mirror contamination has the greatest impact on them.

[0055] Calculate the covariance matrix correction: ΔΣ = principal component reconstruction (changes in covariance of PC1 and PC2); specific values ​​(8×8 matrix, diagonal element example): ΔΣ[4,4] = 0.012 (variance increases in the 1450nm band); ΔΣ[2,2] = 0.005 (variance increases in the 1200nm band); other elements < 0.002 (no significant change).

[0056] S36: The cloud server pushes aging compensation parameters to all devices in the "202601-A" batch.

[0057] S37: After receiving the compensation parameters, the device stores them in local Flash memory. During each subsequent detection: wavelength correction is performed: current usage time T = 180 days; expected drift is calculated: Δλ = a·T + b·T² = 1.836 - 0.583 = 1.253 nm; real-time correction is performed when acquiring spectra: λ_corrected = λ_raw - Δλ, where λ_raw represents the real-time acquired spectral data; and covariance matrix correction is performed: the offset ΔΣ of the covariance matrix is ​​obtained, and the local baseline covariance Σ_old is read; update: Σ_new = Σ_old + ΔΣ; Σ_new is used when calculating Mahalanobis distance subsequently.

[0058] A comparative experiment was conducted using the group collaborative calibration method of this disclosure for aging compensation and the traditional method. N=100 devices were used for 6 months. The technical effects are shown in Table 2 below. The effect of the solution disclosed in this disclosure is comparable to that of hardware calibration at the factory, but the cost is reduced from 50 yuan / time to 0 yuan. In addition, users do not need to send the equipment, resulting in a better user experience.

[0059] Table 2. Comparison of Aging Compensation Effects | Group | Compensation Method | Misjudgment Rate | Missed Detection Rate | User Complaint Rate | Control Group 1 | No Compensation | 28% | 5% | 35% | Control Group 2 | Hardware Calibration After Return to Factory | 12% | 2% | 8% | Experimental Group | Software Compensation Based on This Invention | 12% | 2% | 7% Table S4: Cross-device baseline unification module, used to fuse baseline parameters of various devices, generate unified baseline parameters, and synchronize the unified baseline parameters to each device; specifically, it includes: when a user has multiple devices, fusing and calculating the baseline parameters independently established by each device to generate unified baseline parameters, synchronizing the unified baseline parameters to each device, and replacing the local baseline parameters of each device; in one embodiment, user Li has 3 devices: Device A: Family cup, used for 6 months, N=150; Device B: Office cup, used for 3 months, N=120; Device C: Travel cup, just purchased, N=0.

[0060] S41: Li binds multiple devices in the App and performs device management operations: Device A is already bound to account "user_345678"; Add device B: Scan the QR code of device B and bind it to the same account; Add device C: Perform the same operation. The App displays: "You have bound 3 devices. Do you want to enable the cross-device baseline unification function?", explaining: "After enabling, multiple devices will share the same set of judgment standards to ensure consistent judgment results." Li selects "Enable".

[0061] S42: Each device uploads its local baseline to the cloud: Device A's individualized baseline model: {V0_A, Σ_A, N_A=150}; Device B's individualized baseline model: {V0_B, Σ_B, N_B=120}; Device C's individualized baseline model: N_C=0 (no baseline data).

[0062] S43: The cloud executes a weighted fusion algorithm to perform a weighted average of the baseline centers of each device. ;in Let be the sample size of the i-th device; the larger the sample size, the higher the weight. The baseline center of the i-th device is used; the covariance matrices of each device are pooled and fused: Covariance fusion considers not only the covariance matrix of each device itself, but also the contribution of the deviation between the baseline center of each device and the unified baseline center to the overall variance.

[0063] S44: The cloud will and Push to all devices: Device A: Receives update notification, downloads unified baseline, and replaces local baseline; Device B: Performs the same operation; Device C: Since there are no samples, it directly uses unified baseline as initial baseline, which is equivalent to cold start acceleration.

[0064] S45: One day later, Li filled a glass with tap water at home and tested it using three different devices: After using a unified baseline, device A: D m =1.8, judged as "normal"; Device B: D m =2.1, judged as "normal"; Device C:D m =2.5, judged as "normal". If not standardized, the following may occur: for the same cup of water, equipment A:D m =1.2, judged as "normal"; Device B: D m =4.8, judged as "minor abnormality"; Device C: cannot be judged.

[0065] The use of a cross-device baseline unified module for judgment improves the consistency of detection and is immediately available for new devices C without waiting for a cold start.

[0066] S5: Privacy protection module, including: encrypting baseline parameters during transmission.

[0067] In one implementation, as shown in Figure 4, the encryption method of the privacy protection module consists of four layers. Layer 1, the device local layer, uses plaintext storage, and the stored data includes baseline parameters. It includes local detection history, user personal settings, and scene tag ID; security features include physical isolation, local fast read / write, and device lock screen protection, used to upload local data to the transport layer.

[0068] Layer 2 is the transport layer: It uses AES-256 encryption to build an encrypted channel, encrypts the data uploaded from the device's local layer, and forms encrypted data packets to ensure the security of the data during transmission.

[0069] The third layer is the cloud storage layer: it adopts a distributed encrypted storage method, and the stored data includes anonymized scenario models, global optimization parameters {Ω}, user group statistics and encrypted backup data; security features include data desensitization, key fragmentation management and access control auditing, and it also synchronizes data with the application service layer.

[0070] Layer 4 is the application service layer: It provides application functions through interactive methods and secure API interfaces, including privacy settings interface, data usage report, model update recommendation and security alarm notification; security features include identity authentication and authorization, the principle of least privilege and data lifecycle management, which are used to realize data synchronization with the cloud storage layer and user-facing interactive services.

[0071] As a more advanced privacy protection method, the group collaborative calibration module can adopt a federated learning architecture. The cloud only distributes the initial aging model to each device, and each device uses local aging data to train the model locally. Only the encrypted model gradient is uploaded to the cloud for aggregation. The original aging data never leaves the device, further protecting user privacy.

[0072] In one embodiment, for the technical solution of using a local SD card to store baselines for multiple scenarios: the device is equipped with an 8GB SD card to store baseline data for up to 100 scenarios; the user manually selects the scenario through the App: home, office, travel; all data is stored locally and not uploaded to the cloud. Its drawbacks are: (1) It cannot achieve group collaborative calibration: a single device cannot obtain aging data from other devices; after 6 months of use, the misjudgment rate still increases from 15% to 28%; the user must return the device to the factory for calibration, costing 50 yuan per time.

[0073] (2) Scene recognition depends on manual user input: The actual probability of users forgetting to switch scenes is 35%; incorrect scene selection leads to misjudgment and poor user experience.

[0074] (3) High risk of data loss: The failure rate of SD cards is about 3% per year. Once the SD card is damaged, all scenario baselines will be lost, and users will need to recollect all scenario data. In addition, users of this technical solution must accept the additional cost of regular factory calibration.

[0075] In one embodiment, data between devices is transmitted directly between devices via Bluetooth P2P without cloud relay. That is, device A and device B are paired via Bluetooth, which requires manual operation by the user; the baseline data of device A is transmitted to device B via Bluetooth; no cloud server is involved. Its drawbacks are: (1) it cannot achieve group collaborative calibration, as Bluetooth is a one-to-one transmission and cannot aggregate data from N>2 devices; and its aging compensation problem remains unresolved.

[0076] (2) Cold start still requires 5 days: The new device cannot obtain an initial baseline from other users’ anonymized data; 20 samples still need to be collected for the first use.

[0077] (3) High operational complexity: Bluetooth transmission distance is short, and users need to bring the two devices close together; the Bluetooth pairing process is relatively cumbersome, and the user abandonment rate is as high as 60% in actual testing. In addition, this technical solution only partially solves the problem of cross-device unification, and other problems remain unsolved, making user operation complicated.

[0078] In one embodiment, the device only performs cloud backup and does not perform collaborative calibration. The baseline data is uploaded to the cloud and used only for backup and recovery. It does not perform group data aggregation and aging compensation model training. The cloud is only used for passive storage. Its defects are: (1) The aging problem is not solved. After 6 months of use, the misjudgment rate increases from 15% to 28%. Users still need to return the device to the factory for calibration, and the device is unavailable during the return period, resulting in decreased user satisfaction.

[0079] (2) The cold start still takes 5 days, and it is impossible to use other user data for transfer learning, resulting in a high churn rate of new users.

[0080] A comparative experiment was conducted using the technical solution disclosed herein, the traditional method, and the above-described embodiments. The technical effects are shown in Table 3 below: This disclosure, through the organic combination of four modules, outperforms existing technologies and competing solutions in all key indicators. Through the automatic calibration of this disclosure, the user-intervention calibration cycle is extended from 3-6 months to over 24 months. In particular, it reduces cold start time and aging misjudgment rate, which is something that those skilled in the art could not have foreseen or achieved without implementing this invention.

[0081] Table 3. Comparison of Overall Effects of Various Technical Solutions Solution | Cross-Device Consistency | Cold Start Time | Misjudgment Rate After 6 Months of Aging | Maintenance Cost | User Operation Complexity | Traditional Standalone Solution | Poor (D value difference > 3) | 5 days | 28% | 0 yuan (but limited functionality) | Low | Local SD Card | Medium (D value difference ~ 1.5) | 5 days | 28% | 50 yuan / half a year | Medium (requires manual scene selection) | Bluetooth P2P | Excellent (D value difference < 1) | 5 days | 28% | 50 yuan / half a year | High (cumbersome pairing) | Cloud Backup Only | Excellent (D value difference < 1) | 5 days | 28% | 50 yuan / half a year | Low | This Invention Solution | Excellent (D value difference < 0.5) | 1 day | 12% | 0 yuan | Low (automated) This report has been piloted in the products of a smart water bottle manufacturer. In a 6-month trial in a certain city, with 2000 users, the results showed: the 7-day retention rate of new users increased from 65% to 92%; after 6 months, the device misjudgment rate remained at 12%, compared to 28% in the control group; user satisfaction ratings improved from 3.2 / 5 to 4.5 / 5; the return-to-factory calibration rate decreased from 18% / year to <1% / year, with most users resolving issues through software calibration; and the cloud service cost was approximately 0.5 yuan / user / year, including server, bandwidth, and storage, far lower than the 50 yuan / time cost of return-to-factory calibration.

[0082] Example 2 This example provides a data collaboration method supporting an individualized baseline modeling algorithm. The individualized baseline modeling algorithm establishes an individualized baseline model based on the low-dimensional discrete feature vector of the target object. The low-dimensional discrete feature vector is composed of the ratio feature and difference feature of a set band pair. It includes one or more of the following steps: Multi-scene baseline management step: When the device detects that the deviation of the current sample exceeds the scene switching threshold, a scene recognition process is triggered. According to the scene recognition result, the baseline parameters of the corresponding scene are obtained from the data storage location and the currently used baseline is replaced. If a new scene is identified, a cold start acceleration step is triggered. Cold start acceleration step: When the device is used for the first time or enters a new scene, the initial baseline parameters are obtained from the existing data source. A small number of current environment samples are collected, with the sample size M being less than the sample size N required by the traditional method. The initial baseline is then used to... The baseline parameters are fused with the small number of samples to generate baseline parameters adapted to the current user; Group collaborative calibration steps: Multiple devices periodically upload aging-related data to a centralized data coordination node, which aggregates the aging-related data and establishes an aging compensation model; aging compensation parameters for each device are calculated based on the aging compensation model; the aging compensation parameters are distributed to each device, and each device uses the compensation parameters to correct its local baseline parameters; Cross-device baseline unification steps: When a user has multiple devices, their respective baseline parameters are obtained from each device; the multiple baseline parameters are fused and calculated to generate unified baseline parameters; the unified baseline parameters are synchronized to each device; During the execution of the above steps, the transmitted data is encrypted and differentially private; ensuring that it is impossible to infer the original information of a single user from the transmitted data.

[0083] Example 3: A non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the data collaboration system supporting the individualized baseline modeling algorithm.

[0084] Example 4: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute the data collaboration system that supports the individualized baseline modeling algorithm.

[0085] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0087] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A data collaboration system for supporting individualized baseline modeling algorithms, characterized in that, The individualized baseline modeling algorithm establishes an individualized baseline model based on the low-dimensional discrete feature vector of the target object. The low-dimensional discrete feature vector is composed of the ratio feature and difference feature of a set band pair. Includes one or more of the following modules: Multi-scenario baseline management module, used to identify user scenarios and store baseline parameters for different usage environments. When a scenario change is detected, it switches to the baseline parameters of the corresponding scenario and uses a weighted fusion mechanism to gradually adjust the weights of the old and new scenario baseline parameters within a preset sampling period. If a new scenario is identified, the cold start acceleration module is triggered. The cold start acceleration module is used to quickly establish an initial baseline through transfer learning when the device is used for the first time or enters a new scenario. It uses a small number of samples to fine-tune the initial baseline and generates baseline parameters adapted to the current user through a weighted fusion algorithm. The group collaborative calibration module is used to aggregate aging data from multiple devices and generate compensation parameters to correct local baseline parameters; the cross-device baseline unification module is used to fuse baseline parameters from various devices, generate unified baseline parameters, and synchronize the unified baseline parameters to each device. Data collaborative communication module: Enables the data transmission required by the above modules; The privacy protection module includes: encrypting baseline parameters during transmission.

2. A data collaboration system for supporting individualized baseline modeling algorithms according to claim 1, characterized in that, In the data collaborative communication module, an anonymization processing logic is executed when transmitting data. The terminal device only uploads statistical features that have been locally aggregated to the cloud. These statistical features include the baseline center vector, covariance matrix, or anomaly deviation. Uploading unprocessed raw spectral response data is prohibited to achieve data privacy protection.

3. A data collaboration system for supporting individualized baseline modeling algorithms according to claim 1, characterized in that, Scene recognition in the multi-scene baseline management module is based on location information, specifically including: the device acquiring the current GPS coordinates; converting the GPS coordinates into grid numbers with a preset precision; and matching the grid numbers with scene IDs stored on the cloud server.

4. A data collaboration system for supporting individualized baseline modeling algorithms according to claim 1, characterized in that, The initial baseline acquisition and individualized fine-tuning in the cold start acceleration module includes: querying anonymized baseline datasets of the same region and water source type from the cloud server based on the device's current location information and water source type; performing statistical analysis on the baseline datasets to calculate the median or weighted average to obtain the initial baseline parameters; collecting M current environmental samples to calculate the local baseline parameters; and fusing the initial baseline with the local baseline using a weighted fusion algorithm.

5. A data collaboration system for supporting individualized baseline modeling algorithms according to claim 1, characterized in that, The method for establishing the aging compensation model in the group collaborative calibration module includes: each device periodically uploading an aging data package, which includes: device serial number, sensor batch number, usage time, wavelength drift, and baseline center change; the cloud server groups the data according to the sensor batch number; regression analysis is performed on the data of each batch to fit the aging compensation model, the covariance matrix correction is calculated, and the aging compensation parameters are sent to all devices in the corresponding batch.

6. A data collaboration system for supporting individualized baseline modeling algorithms according to claim 1, characterized in that, The fusion calculation method in the cross-device baseline unification module includes: weighted averaging of the baseline centers of each device. ;in Let be the sample size of the i-th device; the larger the sample size, the higher the weight. The baseline center of the i-th device is used; the covariance matrices of each device are pooled and fused: 。 7. A data collaboration system for supporting individualized baseline modeling algorithms according to claim 1, characterized in that, The encryption methods in the privacy protection module include: encrypting baseline parameters using the AES-256 algorithm, with the encryption key derived from the user's biometrics or the device's unique identifier; adding Laplace noise to the uploaded aging data; and downgrading location information by reducing precise GPS coordinates to grid numbers before uploading.

8. A data collaboration method supporting individualized baseline modeling algorithms, characterized in that, The individualized baseline modeling algorithm establishes an individualized baseline model based on the low-dimensional discrete feature vector of the target object. The low-dimensional discrete feature vector is composed of the ratio feature and difference feature of a set band pair. The process includes one or more of the following steps: Multi-scene baseline management step: When the device detects that the deviation of the current sample exceeds the scene switching threshold, a scene recognition process is triggered. Based on the scene recognition result, the baseline parameters of the corresponding scene are obtained from the data storage location and the currently used baseline is replaced. If a new scene is identified, a cold start acceleration step is triggered. Cold start acceleration step: When the device is used for the first time or enters a new scene, initial baseline parameters are obtained from an existing data source. A small number of current environment samples are collected, with the sample size M being less than the sample size N required by traditional methods. The initial baseline parameters are fused with the small number of samples to generate baseline parameters adapted to the current user. Group collaborative calibration steps: Multiple devices periodically upload aging-related data to a centralized data coordination node. The centralized data coordination node aggregates the aging-related data and establishes an aging compensation model. The aging compensation parameters for each device are calculated based on the aging compensation model. The aging compensation parameters are then distributed to each device, and each device uses the compensation parameters to correct its local baseline parameters. Cross-device baseline unification steps: When a user has multiple devices, obtain the baseline parameters from each device; perform fusion calculation on the multiple baseline parameters to generate unified baseline parameters; synchronize the unified baseline parameters to each device; During the execution of the above steps, the transmitted data is encrypted and differential privacy processed; Ensure that it is impossible to infer the original information of a single user from the transmitted data.

9. A non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement a data collaboration system supporting an individualized baseline modeling algorithm as described in any one of claims 1-7.

10. An electronic device, comprising: A processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute a data collaboration system that implements the individualized baseline modeling algorithm as described in any one of claims 1-7.

Images