A multi-pet cross-device pet identity attribution and health warning method and system

Abstract

This invention discloses a method and system for pet identification and health early warning across multiple pets and devices. The method specifically includes: using high-confidence event segments with identification confidence levels exceeding a preset confidence threshold as valid samples for training and dynamically updating the individual health baseline model of the corresponding pet; comparing real-time monitoring data with the output parameters of the individual health baseline model of the corresponding pet, calculating the individual anomaly degree of each feature, and obtaining a comprehensive anomaly score through weighted fusion and multi-source cross-validation; and outputting tiered health early warning information based on the relationship between the comprehensive anomaly score and a preset tiered threshold. This invention achieves automatic differentiation, independent filing, and accurate early warning of health data for multiple pets in the same household, thereby improving the accuracy, stability, and practicality of pet health monitoring in multi-pet households.

Description

Technical Field

[0001] This invention relates to the field of pet health monitoring and intelligent data processing technology, and in particular to a method and system for pet identification and health early warning across multiple pets and devices. Background Technology

[0002] With the continued growth of the pet economy, smart home, Internet of Things and artificial intelligence technologies are becoming increasingly popular in home settings. The field of pet health monitoring is undergoing profound changes, gradually expanding from simple feeding management to multiple directions such as activity recognition, vital sign collection, abnormal behavior early warning and long-term health assessment.

[0003] Currently, typical pet smart monitoring systems comprise a variety of devices, including smart feeders, water dispensers, litter boxes, smart collars, cameras, positioning modules, and mobile or cloud-based management platforms. However, while the industry has achieved some capabilities in device networking, remote control, and partial data collection, technologies truly focused on continuous pet health monitoring, comprehensive analysis, and intelligent early warning are still in a rapidly developing but not yet fully mature stage. Furthermore, although existing device forms are becoming increasingly diverse, many products are still designed around single functions; multi-functional collaboration and unified data management are becoming clear development trends.

[0004] Research progress shows that pet monitoring technology is evolving towards multimodal, non-contact, and intelligent analysis. For example, a 2024 study (Hu R, Gao Y, Peng G, Yang H and Zhang J (2024) A novel approach for contactless heart rate monitoring from pet facial videos. Front.Vet. Sci. 11:1495109. doi: 10.3389 / fvets.2024.1495109) proposed a non-contact heart rate monitoring method based on pet facial videos and imaging photoplethysmography (iPPG). Under both natural and artificial light conditions, the mean absolute error of heart rate monitoring in dogs and cats was low, demonstrating the feasibility of video-based non-contact heart rate monitoring in pet scenarios. The 2024 RayPet study applied FMCW millimeter-wave radar to pet activity and posture recognition (Ehsan Sadeghi, Abel van Raalte, Alessandro Chiumento, Paul Havinga. (2024). "RayPet: Unveiling Challenges and Solutions for Activity and Posture Recognition in Pets Using FMCW") The study "Mm-Wave Radar." arXiv:2404.15340, Apr 2024. https: / / arxiv.org / abs / 2404.15340, after preprocessing to address issues such as random movement of small animals, noise, and sparse point clouds, achieved an overall recognition accuracy of 89%. Another 2024 study (Mustafa M. Matalgah; Mohammed AliAlqodah, "Appendix A," in Real-Time Ground-Based Flight Data and CockpitVoice Recorder: Implementation Scenarios and Feasibility Analysis , IEEE,2024, pp.145-151, doi: 10.1002 / 9781119984894.app1.) used three types of sensors—accelerometer, gyroscope, and magnetometer—to build a cat activity recognition model, achieving high training and validation accuracy.Meanwhile, the 2025 study on wearable animal design (Marta Siguín, Roberto Casas, Oscar Casas, Teresa Blanco, Towards Effective Wearable Design: 20 Key Factors for Monitoring Physiological Health in Animals, Results in Engineering, Volume 27, 2025, 106001, ISSN 2590-1230, https: / / doi.org / 10.1016 / j.rineng.2025.106001.) proposed a "20-factor framework" for continuous physiological monitoring devices, clearly pointing out that insufficient adaptation to individual animal characteristics, limited measurement methods, and lack of standardization in usage conditions directly affect measurement accuracy and animal acceptance of wearing the device. These studies indicate that while existing technologies have made significant progress in perception accuracy and recognition algorithms, different technological approaches remain independent, and a unified technological system for multi-pet family scenarios has not yet been formed.

[0005] However, existing pet monitoring technologies suffer from several prominent system architecture issues. First, most existing systems are based on single-device, single-modality, or single-pet scenarios. While they can achieve localized monitoring, their support for unified cross-device access, unified time alignment, and multi-source data correlation analysis is extremely limited. Related reviews indicate that current research on veterinary remote vital sign monitoring rarely combines different technologies into complete monitoring systems; most studies remain at the stage of validating single sensing methods, and the scale and cost-effectiveness of clinical validation are insufficiently assessed. Second, existing products tend to focus on "recording data" rather than "building individual profiles and providing continuous early warnings," resulting in a large volume of data that is difficult to reliably translate into continuous judgments that provide guidance on disease risk. Finally, the data heterogeneity among existing devices is significant. Different terminals differ in sampling frequency, timestamp accuracy, spatial location, and data format, making direct fusion of multi-source data difficult.

[0006] In multi-pet households, the shortcomings of existing technologies are further amplified. In such environments, different pets may take turns accessing the same feeder, litter box, or sleeping area, or frequently move between multiple monitoring devices. Therefore, the system must not only collect data but also address the crucial question of "which pet does this data belong to?" Animal re-identification research in 2025 (Pshanth C. Ravoor, Sudarshan TSB, Deep Learning Methods for Multi-Species Animal Re-identification and Tracking–a Survey, ComputerScience Review, Volume 38, 2020, 100289, ISSN 1574-0137, https: / / doi.org / 10.1016 / j.cosrev.2020.100289.), (Cigdem Beyan, Anil Osman Tur, Ehsan Karimi, Fromspecies-specific models to universal re-ID: a survey of animal re-identification,Information Fusion,Volume 133,2026,104323,ISSN References 1566-2535 (https: / / doi.org / 10.1016 / j.inffus.2026.104323) point out that animal re-identification is an emerging field. Compared with human re-identification, it presents unique complexities such as species diversity, significant environmental changes, strong posture differences, and limited data volume. Many methods effective in human scenarios cannot be directly generalized to animal scenarios. This means that in multi-pet households, relying solely on a single camera, wearable device, or general identification model often fails to reliably achieve individual-level data attribution. Once an attribution error occurs, subsequent individual health records and baselines will be contaminated by erroneous samples, ultimately leading to false positives, false negatives, or distorted warning levels.

[0007] In summary, the core deficiency of existing technologies lies not simply in the insufficient number of sensors or the inadequacy of algorithm complexity, but in the lack of a complete technological chain for scenarios involving multiple pets, multi-device linkage, and parallel heterogeneous data collection. Existing solutions often only address one specific aspect of "collection," "identification," or "early warning," lacking a systematic design for the entire process of "unified cross-device data access—candidate event slicing—individual identity attribution—attribution confidence assessment—individual health baseline update—anomaly scoring and tiered early warning." In other words, existing technologies have not fundamentally solved problems such as data attribution confusion, individual baseline contamination, and unstable cross-device integrated early warning in multi-pet environments. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for pet identification and health early warning across multiple devices. By unifying the access, time-series alignment, identification, individual health baseline modeling and updating, and anomaly assessment and graded early warning of monitoring data from multiple devices, the invention achieves automatic differentiation, independent filing, and accurate early warning of health data of multiple pets in the same household. This improves the accuracy, stability, and practicality of pet health monitoring in multi-pet household scenarios, thereby solving at least one of the aforementioned problems in the prior art.

[0009] In a first aspect, the present invention provides a method for pet identification and health alerts across multiple pets and devices, the method specifically comprising: The system integrates and aligns the raw monitoring data collected from various pet monitoring terminals in a unified manner, and segments the continuous data into candidate behavioral event segments according to preset event rules. Based on candidate behavioral event fragments, event feature vectors are extracted and matched with pre-established individual reference feature vectors of each pet under the same account using multi-feature weighted matching. The affiliation score of each pet is calculated, and the target affiliation object is determined based on the affiliation score, and the affiliation confidence is calculated. High-confidence event segments with attribution confidence levels higher than a preset confidence threshold are used as valid samples for training and dynamic updating of the corresponding pet's individual health baseline model. The real-time monitoring data is compared with the output parameters of the corresponding pet's individual health baseline model to calculate the individual anomaly degree of each feature. A comprehensive anomaly score is obtained through weighted fusion and multi-source cross-validation. Based on the relationship between the comprehensive anomaly score and the preset grading threshold, graded health warning information is output.

[0010] Secondly, the present invention provides a multi-pet cross-device pet identification and health early warning system, the system specifically comprising: The data processing module is used to uniformly access and time-series align the raw monitoring data collected by various pet monitoring terminals, and to segment continuous data into candidate behavioral event segments according to preset event rules. The rating calculation module is used to extract event feature vectors based on candidate behavioral event fragments, and perform multi-feature weighted matching with the individual reference feature vectors of each pet pre-established under the same account to calculate the affiliation score of each pet, determine the target affiliation object based on the affiliation score, and calculate the affiliation confidence. The model training module is used to take high-confidence event fragments with an attribution confidence level higher than a preset confidence threshold as valid samples for training and dynamic updating of the individual health baseline model of the corresponding pet. The graded early warning module compares real-time monitoring data with the output parameters of the corresponding pet's individual health baseline model, calculates the individual anomaly degree of each feature, and obtains a comprehensive anomaly score through weighted fusion and multi-source cross-validation. Based on the relationship between the comprehensive anomaly score and the preset graded threshold, it outputs graded health early warning information.

[0011] Thirdly, the present invention provides a computer device, including: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the multi-pet cross-device pet identity attribution and health warning method as described in any of the above methods.

[0012] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multi-pet cross-device pet identity attribution and health warning method as described in any of the above methods.

[0013] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention achieves automatic differentiation, independent filing, and accurate early warning of health data of multiple pets in the same family by unifying access, time-series alignment, identity attribution, individual health baseline modeling and updating, as well as anomaly assessment and hierarchical early warning of monitoring data from multiple devices, thereby improving the accuracy, stability, and practicality of pet health monitoring in multi-pet family scenarios. 2. This invention solves the problems of difficult cross-device unified access and time sequence alignment of raw monitoring data collected from multiple types of pet monitoring terminals, and divides continuous data into candidate behavioral event segments according to preset event rules. This solves the problems of inconsistent time of multi-source data and achieves effective integration and preliminary processing of data from different devices, providing a unified and accurate data foundation for subsequent analysis. 3. This invention extracts event feature vectors from candidate behavioral event fragments and performs multi-feature weighted matching with pre-established individual reference feature vectors for each pet under the same account to calculate the attribution score for each pet, thereby determining the target attribution object and calculating the attribution confidence. This method comprehensively considers multiple dimensions such as the pet's static features, dynamic behavioral features, spatial location features, and multi-device consistency features, greatly improving the accuracy of pet identity attribution and effectively solving the problem of data attribution confusion in multi-pet environments. 4. This invention uses high-confidence event segments with an attribution confidence level higher than a preset confidence threshold as valid samples for training and dynamically updating the individual health baseline model for the corresponding pet. In this way, the individual health baseline model can be continuously optimized based on the pet's actual behavior and health data, ensuring that the model always accurately reflects the pet's health status and avoiding inaccurate individual baselines caused by erroneous sample contamination. 5. This invention compares real-time monitoring data with the output parameters of the corresponding pet's individual health baseline model, calculates the individual anomaly degree of each feature, and obtains a comprehensive anomaly score through weighted fusion and multi-source cross-validation. Based on the relationship between the comprehensive anomaly score and a preset grading threshold, it outputs tiered health warning information. This multi-dimensional, comprehensive anomaly assessment method, combined with a multi-source cross-validation mechanism, can effectively improve the accuracy and reliability of health warnings, reduce false alarms and missed alarms, and provide more timely and effective protection for pet health. 6. This invention is designed for multi-pet households and can automatically distinguish and independently file monitoring data of multiple pets under the same account, reducing the problem of data confusion between different pets; 7. By establishing a long-term health baseline for each pet and identifying abnormalities based on this baseline, this invention avoids the volatility problem caused by relying solely on single monitoring results for judgment, thereby improving the pertinence and stability of health assessment. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating a method for pet identification and health early warning across multiple devices for multiple pets, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the algorithm framework for a multi-pet cross-device pet identity attribution and health early warning method provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the multi-device data timing alignment and identity attribution module provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the multi-pet identity feature modeling and data attribution matching algorithm provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the network structure of an individual health baseline model provided in an embodiment of the present invention; Figure 6 This is a flowchart illustrating the health baseline update and graded early warning module provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a multi-pet cross-device pet identity attribution and health early warning system provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0016] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0017] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0018] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0019] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0020] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0022] The core innovation of this invention lies in its innovative cross-device pet identification mechanism for multi-pet households. At the data attribution level, it unifies the access, time-series alignment, and event slicing of multi-source data from smart homes, feeders, water dispensers, litter boxes, and visual terminals. Combined with multi-dimensional features such as pet weight ranges, activity rhythms, and dwelling preferences, it achieves automatic differentiation and accurate attribution of monitoring data for multiple pets under the same account. At the model level, it constructs a two-layer algorithm architecture of "attribution confidence control + individual health baseline modeling," using only high-confidence attribution data for updating the corresponding pet's health record and health baseline, thus avoiding contamination of the individual model by erroneous attribution data. At the early warning level, it further combines behavioral sequences, environmental data, and owner-uploaded image data for cross-validation, forming a complete algorithmic closed loop from multi-device data collection, identity determination, baseline learning to anomaly identification and three-level early warning output. This upgrades pet health monitoring in multi-pet households from "data recording" to "precise attribution + individualized early warning."

[0023] This invention addresses three core issues in pet health monitoring in multi-pet households: First, data is not differentiated by pet. Addressing the data confusion issues common in multi-pet environments with existing solutions, this invention achieves automatic attribution of monitoring data for multiple pets under the same account through unified cross-device data processing and individual identity feature matching, significantly improving data differentiation capabilities in multi-pet scenarios. Second, health baselines are prone to distortion. This invention establishes an independent digital profile for each pet and forms a personalized health baseline after approximately a week of learning. Combined with attribution confidence control for model updates, this effectively avoids erroneous samples entering individual health models, improving the stability and scientific rigor of subsequent anomaly assessments. Third, early warning results are not reliable enough. This invention achieves a three-tiered early warning output—reminders, observation suggestions, and emergency alerts—through a cross-validation mechanism using behavioral sequences, environmental data, and image data.

[0024] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating a multi-pet cross-device pet identification and health alert method according to an embodiment of the present invention is shown below: S101, performs unified access and time-series alignment of raw monitoring data collected from multiple types of pet monitoring terminals, and divides continuous data into candidate behavioral event segments according to preset event rules; S102, Based on candidate behavioral event fragments, extract event feature vectors and perform multi-feature weighted matching with the individual reference feature vectors of each pet pre-established under the same account to calculate the affiliation score of each pet, determine the target affiliation object based on the affiliation score and calculate the affiliation confidence. S103, high-confidence event segments with attribution confidence higher than the preset confidence threshold are used as valid samples for training and dynamic updating of the individual health baseline model of the corresponding pet; S104 compares the real-time monitoring data with the output parameters of the corresponding pet's individual health baseline model, calculates the individual anomaly degree of each feature, and obtains a comprehensive anomaly score through weighted fusion and multi-source cross-validation. Based on the relationship between the comprehensive anomaly score and the preset grading threshold, it outputs graded health warning information.

[0025] In this embodiment, the raw monitoring data collected by various pet monitoring terminals, including smart feeders, water dispensers, litter boxes, smart collars, cameras, and positioning modules, are first uniformly accessed and time-series aligned. Since the time base of data collected by different terminals may differ, these raw data need to be aligned according to a unified time standard to ensure the accuracy of subsequent analysis. For example, feeding times recorded by smart feeders and pet activity times captured by cameras must be unified to the same time coordinate system. After time-series alignment, continuous data is segmented into candidate behavioral event segments according to preset event rules. These preset event rules can be set based on common pet behavior patterns, such as eating behavior, activity behavior, and resting behavior. Taking eating behavior as an example, when a smart feeder detects food being consumed, the data collected by all relevant monitoring terminals within this time period is segmented into a candidate behavioral event segment. This segment contains various information about the pet's eating process, such as eating time, eating duration, and activity status during eating.

[0026] Based on the candidate behavioral event fragments obtained from the above segmentation, event feature vectors are extracted. Event feature vectors can be extracted from multiple dimensions, such as pet appearance features (e.g., coat color, body shape, patterns) and posture features (e.g., standing, lying down, walking) from image data collected by cameras; and motion features (e.g., speed, trajectory, frequency) from motion data collected by smart collars. After extracting the event feature vectors, they are matched with pre-established individual reference feature vectors for each pet under the same account using a multi-feature weighted matching process. Individual reference feature vectors are obtained by collecting and analyzing a large amount of data from the pet when it first uses the system; they represent the unique characteristics of that pet. Multi-feature weighted matching assigns a weight to each feature based on its importance in determining identity attribution, and then calculates the similarity between the feature vector of the candidate behavioral event fragment and the individual reference feature vectors of each pet. Based on the calculated similarity, an attribution score is calculated for each pet; the higher the attribution score, the greater the likelihood that the candidate behavioral event fragment belongs to that pet. The target attribution is determined based on the attribution score, i.e., the pet with the highest attribution score. At the same time, the attribution confidence score is calculated. The attribution confidence score reflects the accuracy of the judgment of the target attribution object. Its value range is usually between 0 and 1. The closer it is to 1, the more accurate the judgment is.

[0027] High-confidence event segments with attribution confidence exceeding a pre-set confidence threshold are considered valid samples. This threshold is set based on the actual application scenario and system requirements; only when the attribution confidence exceeds this threshold is the candidate behavioral event segment considered to accurately reflect the corresponding pet's behavioral characteristics. These valid samples are used to train and dynamically update the individual health baseline model for the corresponding pet. The individual health baseline model describes various characteristic parameters of a pet in a normal healthy state. It can be trained on valid samples using machine learning algorithms (such as decision trees and neural networks). As the pet continuously generates new high-confidence event segments, these new data are continuously added to the training set, dynamically updating the individual health baseline model to adapt to changes in the pet's health status over time. For example, as a pet ages, its exercise ability and eating habits may change; dynamically updating the individual health baseline model can more accurately reflect the pet's current normal health state.

[0028] The real-time monitoring data is compared with the output parameters of the corresponding pet's individual health baseline model. Real-time monitoring data is collected by the pet monitoring terminal at the current moment, while the output parameters of the individual health baseline model represent the expected characteristic values ​​of the pet in a normal healthy state. By comparing the two, the individual anomaly degree of each feature is calculated. The individual anomaly degree reflects the degree of deviation of the real-time monitoring data from the normal healthy state in a certain feature dimension. For example, if the pet's real-time movement speed is much lower than the normal movement speed output by the individual health baseline model, then the individual anomaly degree of that movement speed feature will be high. After calculating the individual anomaly degrees of each feature, a comprehensive anomaly score is obtained through weighted fusion and multi-source cross-validation. Weighted fusion assigns corresponding weights to the individual anomaly degree of each feature based on the degree of influence of different features on the pet's health status, and then performs a comprehensive calculation. Multi-source cross-validation considers data collected by multiple monitoring terminals to ensure the accuracy of the comprehensive anomaly score. Finally, a graded health warning information is output based on the relationship between the comprehensive anomaly score and a preset grading threshold. The preset grading threshold divides the pet's health status into different levels, such as normal, mildly abnormal, moderately abnormal, and severely abnormal. When the overall anomaly score exceeds a certain threshold, the system outputs a corresponding level of health warning information to remind pet owners to pay close attention to their pets' health. For example, when the overall anomaly score exceeds the severe anomaly threshold, the system outputs a severe anomaly warning information, recommending that the pet owner take their pet to the vet for examination immediately.

[0029] like Figure 2 As shown, the entire process of this invention is divided into five main stages: data input, data fusion and feature extraction, individual identification and attribution, health baseline modeling and analysis, and health early warning and output.

[0030] During the data input phase, data is accessed through various multi-source monitoring devices, including but not limited to smart feeders, water dispensers, and litter boxes. The types of data collected cover movement data, appetite data, sleep data, videos / pictures, and behavior logs, providing comprehensive raw data support for subsequent analysis.

[0031] In the data fusion and feature extraction stage, heterogeneous data from different devices are synchronized, and noise removal and dimensionality reduction are performed to ensure data quality and usability, enabling data analysis within a unified framework. Key information such as motion features, behavioral patterns, and physiological indicators are extracted from the processed data, providing specific feature parameters for individual identification and health analysis.

[0032] In the individual identification and attribution stage, recognition probability is calculated based on multimodal data (such as video re-id, RF signals, behavioral phase lines, etc.). Through automatic data attribution and archiving, data is assigned to corresponding pet accounts based on identity tags, and pet profiles are created or updated, including pet name, description, and other information. While determining the target attribution object, attribution confidence is calculated to quantify the accuracy of identification and provide a reliable basis for subsequent data processing.

[0033] In the health baseline modeling and analysis phase, time series analysis methods, combined with nonparametric baseline models and machine learning techniques, are used to establish a personalized health baseline for each pet. The individual health baseline model is dynamically adjusted based on newly collected data to ensure that the baseline can reflect changes in the pet's health status in a timely manner.

[0034] In the health early warning and output phase, real-time monitoring data is compared with the output parameters of the individual health baseline model to calculate the individual anomaly degree of each feature. A comprehensive anomaly score is obtained through weighted fusion and multi-source cross-validation. Based on the relationship between the score and preset grading thresholds, tiered health early warning information is output, including disease risk alerts, anomaly levels, and care recommendations. Once an anomaly is detected, the system will promptly push intelligent early warning information, enabling pet owners to understand their pet's health status and take appropriate measures.

[0035] The entire process, through unified cross-device data access, candidate event slicing, individual identity attribution, attribution confidence assessment, individual health baseline update, and anomaly scoring and graded early warning, forms a complete technical chain, effectively solving problems such as data attribution confusion, individual baseline contamination, and instability of cross-device comprehensive early warning in multi-pet environments.

[0036] In some embodiments, step S101 above, which involves uniformly accessing and aligning the raw monitoring data collected by multiple types of pet monitoring terminals, and segmenting continuous data into candidate behavioral event segments according to preset event rules, specifically includes: The raw monitoring data from multiple types of pet monitoring terminals are uniformly accessed, and each data point is appended with a device identifier, account identifier, collection timestamp, and spatial location identifier to form multi-source data. The pet monitoring terminals include smart nests, smart feeders, smart water dispensers, litter box monitoring devices, environmental sensors, and visual acquisition terminals. Multi-source data is preprocessed by outlier removal, missing value correction, clock correction, and sampling frequency unification to transform data from scattered sources, heterogeneous formats, and different time granularities to the same time axis, thus obtaining time-aligned continuous data. Based on a preset time window or behavioral state change rules, the continuous data after time alignment is sliced ​​into event segments to generate candidate behavioral event fragments. The candidate behavioral event fragments include eating events, drinking events, sleeping events, staying events, activity events, and toileting events.

[0037] In this embodiment, data collected from multiple pet monitoring terminals in a home environment are uniformly accessed, and each data entry is appended with a device identifier, account identifier, collection timestamp, and spatial location identifier. Its working principle lies in the fact that a data processing server performs outlier removal, missing value correction, clock correction, unified sampling frequency, and unified timeline mapping on data uploaded from different devices. This transforms data that was originally scattered in origin, heterogeneous in format, and had different time granularities into a unified analytical framework. Then, based on preset time windows or behavioral state change rules, continuous data is segmented into candidate behavioral event segments, such as eating events, drinking events, sleeping events, staying events, activity events, or toileting events. This technology fundamentally solves the problem of difficulty in directly comparing and correlating data from different devices, providing a unified and traceable data foundation for subsequent identification and health analysis.

[0038] like Figure 3 As shown, pet-related data is acquired through various devices: smart collars collect data on the pet's movement, heart rate, and body temperature; smart feeders record food and water intake; smart litter boxes provide data on excretion and weight; and indoor cameras perform visual analysis and track movement trajectories. This data from different devices provides rich raw material for subsequent processing.

[0039] Specifically, the process begins with heterogeneous data parsing of various data collected from the multi-device raw data access layer, converting them into a unified standard framework format for subsequent processing. Bluetooth / RFID explicit identification and biometric implicit matching technologies are then utilized. Bluetooth / RFID explicit identification determines the approximate identity range of a pet through relevant signals; biometric implicit matching combines visual and weighted data with stored network sequence (possibly network sequence and other related feature information), while also performing behavioral pattern consistency analysis—comparing real-time behavior with dynamic indicators (possibly preset dynamic feature indicators) and dynamic fundamental characteristics (dynamic basic properties)—to provide more evidence for identification. Clock synchronization is then performed to ensure data collected from different devices remains consistent over time. A sliding window aggregation method is used to divide and aggregate continuous data streams according to specific time windows, preparing for subsequent processing. Multi-source evidence fusion is then performed based on the previously processed data, comprehensively considering various features and information. Next, confidence scores are calculated, and through multi-faceted analysis and judgment, the degree of correlation between the data and the individual pet, as well as the probability of their identification, are determined. Based on the calculated confidence scores and other information, the data is separated and routed to the corresponding individual pet profiles. For example, pet A, pet B, and pet C each have their own independent files, and the data will be accurately allocated to the corresponding files, providing individualized data support for subsequent health baseline modeling and early warning.

[0040] After data separation and routing are completed, a health baseline for each pet is established based on data from its individual file. Through continuous monitoring and analysis, when real-time data deviates from the health baseline, anomaly alerts are issued. Relevant alert information is generated based on the independently created file data, enabling real-time monitoring and anomaly warnings for pets' health status.

[0041] The entire process involves steps such as multi-device data access, data parsing and unification, time-series alignment, probabilistic correlation decision-making, and data separation and routing. It enables cross-device data processing, identity attribution, and health warning functions for multiple pets, forming a complete technical logic system.

[0042] In some embodiments, step S102 above, which involves extracting event feature vectors based on candidate behavioral event fragments and performing multi-feature weighted matching with pre-established individual reference feature vectors for each pet under the same account to calculate the ownership score of each pet, specifically includes: For each candidate behavioral event fragment, multi-dimensional features for identity recognition are extracted, and a corresponding event feature vector is constructed. The event feature vector includes static features for characterizing the inherent attributes of an individual, dynamic behavioral features for reflecting the individual's living habits, spatial location features for indicating the location where the event occurred, and multi-device consistency features for reflecting the collaborative relationship of multiple devices. Retrieve the individual reference feature vectors pre-established for each pet under the same account. The individual reference feature vectors are generated based on the historical identity files of the corresponding pets and contain feature items of the same dimension as the event feature vectors. Calculate the static feature matching degree, dynamic behavior similarity, spatial location matching degree, and multi-device consistency verification value between the event feature vector of each candidate behavior event segment and the individual reference feature vector of each pet. According to the preset weight allocation rules, the calculated static feature matching degree, dynamic behavior similarity, spatial location matching degree and multi-device consistency verification value are weighted and fused to generate the ownership score of each pet corresponding to each candidate behavior event fragment.

[0043] In this embodiment, an independent identity profile is established for each pet under the same account, and identity features to distinguish different individuals are extracted from long-term historical data. Its working principle is that for each candidate behavioral event segment, the system first extracts the corresponding identity recognition features and constructs an event feature vector. Let the candidate behavioral event segment be at time... The event feature vector is: ; in, This represents the individual feature values ​​in the event feature vector, which includes at least weight features, body shape contour features, pressure distribution features, activity intensity features, event occurrence time features, spatial location features, and behavioral rhythm features. Subsequently, Let each feature value be denoted as a feature value in the individual reference feature vector. The system retrieves the first feature value from the same account. The individual reference feature vector in a pet's historical identity file is denoted as: ; A multi-feature weighted matching algorithm was used to calculate the matching score between candidate behavioral event fragments and each pet.

[0044] Specifically, firstly, static feature similarity, dynamic behavior similarity, spatial matching degree, and multi-device consistency matching value are calculated respectively. Let the... The static feature similarity for each pet is The dynamic behavior similarity is Spatial matching degree is Multi-device consistency matching value Then they can be expressed as: ; in, Used to characterize the degree of matching between static identity features such as weight, body shape, and pressure distribution; It is used to characterize the degree of similarity among dynamic behavioral features such as activity rhythms, sleep patterns, food time distribution, water drinking frequency, and toilet cycle; Used to characterize the degree of spatial matching between the location of an event and historical area stay preferences; This is used to characterize the co-occurrence consistency or mutual exclusion constraint value among data collected by different devices within the same time window. Where, Represents the similarity function. This represents a spatial location matching function. represents the multi-device consistency coefficient, p represents the number of features in the event feature vector when calculating static feature similarity, and q represents the number of dynamic behavior features when calculating dynamic behavior similarity. Represents the event feature vector of the candidate behavioral event segment at time step [time]. The k-th eigenvalue in This represents the individual reference feature vector in the historical identity profile of the i-th pet under the same account. The k-th eigenvalue in This represents the m-th feature value in the event feature vector of the candidate behavioral event fragment that is related to dynamic behavior. This represents the m-th feature value related to dynamic behavior in the individual reference feature vector of the i-th pet's historical identity file under the same account. This indicates the location information where the candidate behavioral event occurred. This represents the historical area stay preference information of the i-th pet. This represents the weight of the k-th static identity feature. This represents the weight of the m-th dynamic behavior feature. The weights representing spatial matching degree are used to measure the importance of spatial matching degree in the overall matching degree calculation. The weight represents the consistency across multiple devices.

[0045] After obtaining the above-mentioned item matching results, the system performs weighted fusion of each matching result according to preset weights to obtain the first result. The pet's corresponding ownership rating : ; in, , , , The preset weighting coefficients satisfy the following conditions: ; The identification characteristics can include relatively stable static features such as weight range, body shape, and pressure distribution, as well as dynamic features reflecting individual lifestyle habits, such as activity rhythm, sleep patterns, feeding time distribution, water drinking frequency, toileting cycle, and regional stay preferences. When calculating the attribution result, not only the similarity of a single feature is considered, but also the consistency between the event occurrence time and the pet's historical behavioral rhythm, the degree of agreement between the event occurrence location and the pet's spatial preferences, and the mutual exclusion or co-occurrence relationships between data from different devices within the same time window, thus outputting the final attribution result. This embodiment achieves a crucial leap from "collecting data" to "determining which pet the data belongs to," which is a key inventive point that distinguishes this invention from existing single-pet, single-device monitoring solutions.

[0046] Furthermore, the step of calculating the static feature matching degree, dynamic behavior similarity, spatial location matching degree, and multi-device consistency verification value between the event feature vector of each candidate behavioral event segment and the individual reference feature vector of each pet specifically includes: By comparing the weight measurement value, body shape contour parameters and pressure distribution data in the event feature vector with the corresponding historical static archive data in the individual reference feature vector, the static feature matching degree is calculated. The static feature matching degree is used to measure the degree of conformity between the current event and the target pet in terms of inherent physiological attributes. By analyzing the periodic match between the activity occurrence time, activity duration, and behavior frequency distribution in the event feature vector and the corresponding historical behavior rhythm pattern in the individual reference feature vector, dynamic behavior similarity is calculated. The dynamic behavior similarity is used to measure the degree of matching between the current event and the target pet in terms of time series behavior habits. By comparing the spatial location identifier carried in the event feature vector with the historical area stay preference recorded in the individual reference feature vector, the spatial location matching degree is calculated. The spatial location matching degree is used to measure whether the current event location is within the target pet's daily activity range. By acquiring associated data collected by multiple monitoring devices within the same time window, the coexistence relationship or mutual exclusion constraint between the current event reflected by the event feature vector and the detection results of other devices is analyzed, and a multi-device consistency verification value is calculated. The multi-device consistency verification value is used to verify the logical self-consistency of the current event from the perspective of multi-source monitoring.

[0047] In this embodiment, static feature information such as weight measurement, body contour parameters, and stress distribution data are extracted from the event feature vector of candidate behavioral event fragments. These static features reflect the inherent physiological attributes of pets; for example, a pet's weight is relatively stable over a certain period, and body contour and stress distribution are also unique to each individual. Simultaneously, corresponding historical static profile data is obtained from the pre-established reference feature vectors of each pet under the same account. Then, by precisely comparing the static feature data in the current event feature vector with the historical static profile data, the static feature matching degree is calculated. For example, if the pet's weight measurement in the current event is very close to the weight record in a pet's historical profile, and the body contour parameters and stress distribution data are also highly consistent, then the pet's static feature matching degree in the current event will be high. The static feature matching degree is mainly used to measure the degree of conformity between the current event and the target pet in terms of inherent physiological attributes, providing an important basis for subsequent attribution scoring.

[0048] To calculate dynamic behavior similarity, dynamic behavioral features such as the time of occurrence, duration, and frequency distribution of the activity are first extracted from the event feature vector. These features reflect the pet's behavioral habits at different times; for example, pets are usually more active during certain periods, and the duration of each activity follows a certain pattern. Simultaneously, the corresponding historical behavioral rhythm pattern is obtained from the individual reference feature vector. Next, the periodic match between the dynamic behavioral features of the current event and the historical behavioral rhythm pattern is analyzed. If the time of occurrence, duration, and frequency of the pet's activity in the current event are highly consistent with a pet's historical behavioral rhythm, then the dynamic behavior similarity will be high. Dynamic behavior similarity is mainly used to measure the degree of matching between the current event's time-series behavioral habits and the target pet, helping to determine whether the event is more likely to belong to a specific pet.

[0049] When calculating spatial location fit, the spatial location identifier carried in the event feature vector is first obtained, indicating the specific location where the current event occurred. Simultaneously, historical area stay preferences are extracted from the individual reference feature vector, reflecting the pet's activity range and frequently visited areas in daily life. Then, the spatial location identifier of the current event is compared with the historical area stay preferences in terms of spatial topology. For example, if the location of the current event is within the area where a pet frequently roams, the spatial location fit will be high. Spatial location fit is mainly used to measure whether the location of the current event matches the target pet's daily activity range, further assisting in determining the attribution of the event.

[0050] To calculate the multi-device consistency check value, it is necessary to obtain correlated data collected by multiple monitoring devices within the same time window. In multi-pet households, various monitoring devices are typically deployed, such as smart feeders, cameras, and positioning modules. The analysis focuses on the coexistence relationship or mutual exclusion constraint between the current event reflected in the event feature vector and the detection results of other devices. For example, if a camera detects a pet active in a certain area, and the positioning module also shows that the pet is in the same area, then this coexistence relationship indicates data consistency. Conversely, if the detection results of different devices are contradictory, it may affect the multi-device consistency check value. The multi-device consistency check value obtained in this way is used to verify the logical self-consistency of the current event from the perspective of multi-source monitoring, ensuring the reliability and accuracy of the data, and providing assurance for the accurate attribution of pets.

[0051] By calculating the static feature matching degree, dynamic behavior similarity, spatial location matching degree, and multi-device consistency verification value, we can comprehensively and accurately evaluate the correlation between candidate behavioral event fragments and the reference feature vectors of each pet individual, laying a solid foundation for subsequent calculation of attribution scores and determination of target attribution objects.

[0052] In some embodiments, step S102 above, which involves determining the target belonging object based on the belonging score and calculating the belonging confidence level, specifically includes: Compare the ownership scores of the same candidate behavior event segment with the pets under the same account, and determine the pet with the highest ownership score as the target ownership object of the current candidate behavior event segment; Extract the score difference between the highest and second-highest attribution scores of the target object, calculate the stability index of the attribution score of the target object in the historical matching process, and obtain the consistency index between multiple source devices in the current time window. The attribution confidence of the current candidate behavioral event fragment is generated by comprehensively calculating the score difference, stability index, and consistency index.

[0053] In this embodiment, the affiliation scores of each pet are compared, and the pet with the highest affiliation score is selected as the target affiliation object for the current candidate behavioral event segment, that is: ; This indicates the ownership rating among all pets. The system reaches the index of the largest pet. Furthermore, to improve the reliability of ownership determination, the system calculates ownership confidence by combining the difference between the highest and second-highest ownership scores, the stability of multiple historical matches, and the consistency results of multi-source devices. Let the highest ownership score be... The second highest attribution score was Then the confidence level of attribution It can be represented as: ; in, , , Calculate the weighting coefficients for the confidence level, and satisfy: ; This indicates the stability index of multiple historical matches. This indicates the consistency index of multi-source devices within the current time window. To prevent extremely small positive numbers with a denominator of zero.

[0054] In some embodiments, step S103 above, which involves using high-confidence event segments with attribution confidence levels higher than a preset confidence threshold as valid samples for training and dynamically updating the individual health baseline model of the corresponding pet, specifically includes: When the attribution confidence is greater than or equal to the preset confidence threshold, the current candidate behavioral event segment is determined to be a high-confidence event segment and is used as a valid sample; otherwise, the current candidate behavioral event segment is marked as a low-confidence event to be confirmed. The valid samples are preprocessed to generate a multidimensional health monitoring sequence, which includes the amount of activity, resting time, sleep duration, number of meals, number of water drinks, number of toilet visits, respiratory rate, heart rate, weight and corresponding environmental parameters within a unit time window. Based on a time-series modeling network for continuous health monitoring data of pets, an independent individual health baseline model is constructed for each pet. The multidimensional health monitoring sequences corresponding to multiple valid samples belonging to the corresponding pets are used as training data, and the stable monitoring results of multiple consecutive days are used as baseline labels to initially train the individual health baseline model and generate individual health baseline parameters. Based on the confidence level of the new valid samples, corresponding sample update weights are assigned, and the model parameters at the current time are corrected according to the sample update weights.

[0055] In this embodiment, when the attribution confidence satisfies: ; in, When a pre-set confidence threshold is set, the system determines that the candidate behavioral event fragment belongs to the first... Only pets, and complete automatic ownership; when If the event is not confirmed, it is marked as a low-confidence event awaiting confirmation. This enables automatic matching and data attribution based on multi-dimensional identity features in multi-pet scenarios.

[0056] After data attribution is completed, instead of directly incorporating all monitoring data into the corresponding pet's health record, an attribution confidence control mechanism is introduced. Its working principle is that the system only uses monitoring data with an attribution confidence level higher than a preset threshold as valid samples to establish or update the corresponding pet's individual health baseline model; monitoring data with a low attribution confidence level is only cached, reviewed, or used for auxiliary display, and does not directly enter the baseline update process. Let the... The confidence level of the sample at time step is The preset threshold is Then the sample validity determination function can be expressed as: ; in, This indicates that the monitoring data is used as a valid sample in the individual health baseline model training or update process. This indicates that the monitoring data is only used for caching, verification, or auxiliary display.

[0057] like Figure 4 As shown, the data comes from various pet monitoring terminals, including smart collars, smart food bowls, smart water fountains, smart litter boxes, and environmental sensors. These devices collect various types of information about the pets; for example, smart collars can acquire pet movement data, and smart food bowls record eating-related data, providing comprehensive raw data support for subsequent processing. The data input from multiple devices is divided into time windows according to preset time rules to prepare for subsequent behavioral event detection. Pet behavioral events, such as eating and activity, are detected using specific algorithms or rules, and candidate behavioral event segments are extracted from the data segmented by time windows, with each segment corresponding to a possible behavioral event.

[0058] For the generated candidate behavioral event fragments, multiple features are extracted: weight features: recording the pet's weight information at the time of the event; body shape features: describing the pet's body shape and appearance during the event; stress distribution features: reflecting the distribution of body stress under specific behaviors; activity intensity features: reflecting the intensity of the pet's activity during the event; time features: containing information related to the time of the event; spatial location features: recording the spatial location of the event; and behavioral rhythm features: reflecting the rhythm and pattern of the pet's behavior. These features together constitute the event feature vector, used for subsequent identity recognition.

[0059] The similarity between static features (such as weight and body shape) in the event feature vector and static identity features in each pet's historical identity file is calculated to obtain the static feature matching degree. Dynamic behavioral features (such as activity intensity and behavioral rhythm) in the event are compared with dynamic behavioral habits in the pet's historical file to calculate the dynamic behavior similarity. The spatial location features of the event are compared with the pet's historical area preferences to calculate the spatial location matching degree. Data collected by multiple monitoring devices within the same time window is analyzed to verify the logical consistency of the current event under the detection results of other devices, and a multi-device consistency check value is calculated. According to a preset weight allocation rule, the static feature matching degree, dynamic behavior similarity, spatial location matching degree, and multi-device consistency check value are weighted and fused to generate an affiliation score for each pet corresponding to each candidate behavioral event segment. The affiliation scores of different pets corresponding to the same candidate behavioral event segment are compared, and the pet with the highest score is selected as the target affiliation object.

[0060] By comparing the difference between the highest and second-highest attribution scores of the target pet, statistically analyzing the stability indicators of the target pet in historical matching, and obtaining consistency indicators among multiple source devices within the current time window, a comprehensive calculation of attribution confidence is performed to assess the reliability of identity attribution. Various historical characteristic information for each pet is stored, including static identity characteristics, dynamic behavioral habits, regional stay preferences, and historical association records across multiple devices, providing foundational data for multi-dimensional feature matching calculations. After determining the target pet and calculating the attribution confidence, further processing, such as fusion, is performed on the attribution scores to provide accurate data attribution information for subsequent health baseline modeling.

[0061] Furthermore, the process of using the multidimensional health monitoring sequences corresponding to multiple sets of valid samples belonging to the corresponding pet as training data, and using stable monitoring results over multiple consecutive days as baseline labels, to initially train the individual health baseline model and generate individual health baseline parameters specifically includes: A training sample sequence is constructed based on the multidimensional health monitoring sequences corresponding to multiple sets of valid samples belonging to the corresponding pets; Based on the stable monitoring results over several consecutive days, health monitoring data within the corresponding time window are extracted as baseline labels. These baseline labels are used to characterize the normal physiological fluctuation range and behavioral patterns of pets in a healthy state. The training sample sequence is input into a preset temporal modeling network for model forward propagation. The temporal modeling network is used to receive multidimensional health monitoring sequences through the input layer, perform preliminary feature extraction through the feature encoding layer, and then send them to the temporal feature extraction layer to capture the behavioral rhythms and physiological changes in the continuous time dimension. The output layer generates predicted individual health baseline parameters, which include activity baseline, sleep pattern baseline, food and water intake baseline, toilet cycle baseline, respiratory rate baseline, heart rate baseline, and weight fluctuation baseline.

[0062] In this embodiment, the individual health baseline model is preferably constructed and trained using a time-series modeling network based on continuous pet health monitoring data, and an independent individual health baseline model is established for each pet. The model input is a multidimensional health monitoring sequence arranged in chronological order, let the th... The input feature vector corresponding to each time window is ; in, Each element in the input feature vector represents a different health monitoring parameter within a unit time window. This parameter includes at least the activity level, resting duration, sleep duration, number of meals, number of water drinks, number of toilet visits, respiratory rate, heart rate, weight, and at least one auxiliary parameter from the corresponding timestamp, diurnal time period, and ambient temperature and humidity. To ensure that data from different sources can be uniformly incorporated into the model, the system first performs time alignment, missing value imputation, outlier removal, normalization, and sliding time window slicing on the original monitoring data to generate a training sample sequence. ; in, ... Represents the training sample sequence Each input feature vector in the, This represents the length of the sliding time window. The model network structure includes at least an input layer, a feature encoding layer, a temporal feature extraction layer, and an output layer. The temporal feature extraction layer preferably uses any one of LSTM, GRU, TCN, or Transformer to extract the behavioral rhythms and physiological changes of the pet over continuous time. Let the hidden representation obtained after the temporal feature extraction layer be: ; in, Represents a time-series modeling network. Represents network parameters. The output layer is based on the hidden representation. Generate the corresponding individual health baseline parameters for the pet, denoted as: ; in, This represents the individual's baseline health parameters, i.e., the output parameters. This represents each element in the output parameters, which at least include activity baseline, sleep pattern baseline, food and water intake baseline, toilet cycle baseline, respiratory rate baseline, heart rate baseline, weight fluctuation baseline, and their normal fluctuation range. During training, the system selects historical monitoring data with an attribution confidence level higher than a preset threshold as valid training samples, and uses stable monitoring results from multiple consecutive days or weeks as baseline labels to complete the initial model training. Let the baseline label be... The model training objective function can then be expressed as: ; in, Indicates the number of training samples. Let represent the L2 norm. Therefore, only samples that meet the confidence criteria participate in model parameter optimization. During the model update phase, the system assigns different sample weights based on the confidence level of the new assigned data, so that high-confidence samples have a greater impact on model parameter updates and baseline correction. Let the ... The sample update weights at time step are Then it can be expressed as: ; Furthermore, the dynamic update process of the individual health baseline model can be represented as: ; in, This represents the baseline health parameters at the previous moment. This represents the baseline parameters predicted based on the current high-confidence samples. The dynamic update weights represent a positive correlation with the attribution confidence level. Therefore, the higher the attribution confidence level, the stronger the correction effect on the individual health baseline model; conversely, the lower the attribution confidence level, the weaker the impact on model updates or the less it participates in the updates. Through this design, the model can continuously reflect the stable physiological and behavioral characteristics of different pets, providing a basis for subsequent anomaly detection, health assessment, and early warning. This technique effectively avoids the problems of individual profile contamination, health baseline distortion, and inaccurate subsequent anomaly judgments caused by incorrect data attribution in existing technologies, enabling the model to truly reflect the individualized health status of each pet.

[0063] like Figure 5 As shown, the training process and dynamic updates of the individual health baseline model are demonstrated. The overall process starts from the input layer, goes through multiple processing layers, and finally reaches the health baseline output layer, realizing the determination of pet ownership and health warning.

[0064] The input layer is responsible for collecting multi-source data, including activity sequences, sleep sequences, food / water intake sequences, toilet cycle sequences, respiratory rate sequences, heart rate sequences, weight sequences, as well as timestamps, diurnal time periods, and environmental temperature and humidity information. This data provides comprehensive raw information for subsequent processing. The raw input data undergoes operations such as time alignment, missing value imputation, outlier removal, and normalization. Time alignment ensures consistency in the time dimension for data collected from different devices; missing value imputation and outlier removal guarantee data integrity and accuracy; normalization makes the data comparable, facilitating subsequent processing. Simultaneously, by using sliding time window slicing, the continuous data stream is divided into data segments within multiple time windows, and environmental auxiliary feature encoding is used to incorporate environmental information into the data processing.

[0065] The multidimensional feature encoding layer encodes features from the preprocessed data, including physiological features and temporal location, transforming the data into feature vectors suitable for model processing. It integrates various feature information to construct a multidimensional temporal sample tensor, preparing for subsequent temporal feature extraction.

[0066] The backbone network for temporal feature extraction includes the following structure: (1) Bi-LSTM / GRU unit: using bidirectional long short-term memory network or gated recurrent unit to extract long-term behavioral rhythm features of data and capture the long-term dependence of pet behavior and physiological features in time series. (2) Temporal Convolution Network (TCN) module: further modeling short-term fluctuations and long-term dependencies through temporal convolution network to extract local and global features of data in the time dimension. (3) Self-Attention / Transformer Encoder module: using self-attention mechanism or Transformer encoder to capture cross-modal temporal correlations and mine the intrinsic connection between different features.

[0067] The feature fusion layer combines features from different modalities, such as combining activity features with physiological features, to form a more comprehensive feature representation. The fused features are then further processed and transformed by a fully connected layer to prepare for subsequent weighted calculation of attribution confidence.

[0068] Next, the feature representation of the current sample is determined, and different weights are assigned to the samples based on their attribution confidence levels. Higher attribution confidence results in greater sample weights and a greater role in model updates. By adjusting sample weights, the influence of high-confidence samples on the model is enhanced, while the interference from low-confidence samples is suppressed, improving the model's accuracy and stability. The weighted samples are then used to dynamically update the pet's health baseline, ensuring that the baseline promptly reflects the pet's latest health status.

[0069] Multiple health baselines are established, including activity level baseline, sleep pattern baseline, food and water intake baseline, toileting cycle baseline, respiratory rate baseline, heart rate baseline, and weight fluctuation baseline. These baselines are determined based on the pet's historical health data and normal physiological fluctuation range, and are used to compare with real-time monitoring data to determine the pet's health status.

[0070] The individual health baseline model has the following beneficial effects: (1) It establishes a unique health baseline model for each pet, fully considering individual differences. (2) It uses data with high attribution confidence to train and update the model, improving the accuracy and reliability of the model. (3) The model can output individualized behavioral and physiological baseline parameters for each pet, providing accurate reference standards for health assessment. (4) By comparing real-time data with the health baseline, abnormalities can be detected in a timely manner, providing a scientific basis for health early warning.

[0071] In some embodiments, step S104 above, which involves comparing the real-time monitoring data with the output parameters of the corresponding pet's individual health baseline model, calculating the individual anomaly degree of each feature, and obtaining a comprehensive anomaly score through weighted fusion and multi-source cross-validation, specifically includes: For real-time monitoring data of pets that have been successfully assigned to the target pet, extract the multi-dimensional monitoring feature vector at the current moment; The real-time monitoring feature value of each dimension in the multidimensional monitoring feature vector is compared with the benchmark value of the corresponding dimension in the individual health baseline parameters. Based on the relative magnitude of the deviation of the real-time monitoring feature value from the benchmark value, and combined with the normal fluctuation range of the real-time monitoring feature in the health baseline, the single abnormality degree corresponding to each real-time monitoring feature is calculated. According to the preset dimensional feature weight allocation rules, the individual anomaly degrees of each dimension's real-time monitoring features are weighted and fused to generate a preliminary comprehensive anomaly score. The multi-source cross-validation correction factor is obtained based on the multi-source cross-validation mechanism. The preliminary comprehensive anomaly score is then fused with the multi-source cross-validation correction factor to generate the target comprehensive anomaly score.

[0072] In this embodiment, by comparing the deviation between real-time data and an individual's health baseline, and combining multi-dimensional signals such as changes in activity levels, sleep interruption frequency, abnormal water intake frequency, abnormal eating behavior, abnormal toileting patterns, and changes in environmental factors, a weighted anomaly scoring algorithm is used to generate an anomaly score. Specifically, let the first... The multidimensional monitoring feature vector at time step is The corresponding individual health baseline vector is Then, the individual outlier of each feature dimension can be expressed as: ; in, For the first Anomaly degree of dimensional features This is the current real-time monitoring value. To correspond to the health baseline value, This represents the normal range or standard deviation of this characteristic within a healthy baseline. To prevent extremely small positive numbers with a denominator of zero, the system then weights and fuses the anomalies of each dimension according to preset weights to obtain a comprehensive anomaly score. : ; in, For the first The weight coefficients corresponding to the dimensional features, and satisfying 'n' represents the number of features, which include at least one or more of the following: activity level, sleep interruption frequency, water intake frequency, food intake frequency, toilet visits frequency, respiratory rate, heart rate, and environmental parameters. Furthermore, to reduce false alarms from a single device, the system also introduces a multi-source cross-validation correction factor. The final anomaly score is obtained. : ; Specifically, when multiple devices detect abnormalities that deviate from the healthy baseline within the same time window, When the anomaly originates from a single instantaneous fluctuation and lacks collaborative verification from other devices, The system then uses the final anomaly score as a basis. The relationship between the health warning level and the preset grading thresholds outputs the corresponding health warning level; for example, when When it is judged as a normal state, when Timely output reminders to pay attention, when Observe the output suggestions when The system outputs emergency alerts in real time. Through the aforementioned anomaly scoring algorithm, the system can achieve quantitative health risk assessment based on multi-source continuous monitoring data. It outputs tiered early warning results, such as reminders, observation suggestions, and emergency alerts, based on the scoring results. Furthermore, to reduce false alarm rates, the system can perform multi-source cross-validation on anomalies triggered by a single device. For example, if activity levels are continuously decreasing, but water intake frequency increases, sleep quality declines, and ambient temperature and humidity are within normal ranges, the anomaly score can be increased, and a risk warning can be issued earlier. Conversely, if the anomaly originates from a single instantaneous data fluctuation and lacks collaborative verification from other devices, the warning level can be reduced or the output delayed. This embodiment realizes a shift from shallow data recording to in-depth health trend analysis and early risk warning, improving the system's ability to identify long-term health problems and its practical value in early warning.

[0073] Furthermore, the step of obtaining the multi-source cross-validation correction factor based on the multi-source cross-validation mechanism specifically includes: Acquire correlated data collected by multiple monitoring devices within the same time window, and analyze whether the current anomaly has been verified by other devices. A positive correction factor is generated when multiple devices detect abnormalities that deviate from the health baseline. A negative correction factor is generated when the anomaly originates from a single instantaneous fluctuation and lacks verification from other devices.

[0074] In this embodiment, a specific time window is set, the size of which can be adjusted reasonably according to actual conditions, for example, it can be set to a few minutes or tens of minutes. Then, correlated data is collected from multiple different monitoring devices within the same time window. These monitoring devices include, but are not limited to, smart collars, cameras, smart feeders, water dispensers, litter box monitoring devices, etc. Each device monitors the pet's behavior and physiological characteristics from different angles and records relevant data. For example, smart collars can record the pet's movement data and heart rate information, cameras can capture the pet's activity, and smart feeders can record the pet's eating status, etc. After acquiring this correlated data, it is analyzed, focusing on how data that has been initially identified as abnormal is reflected in the data of other devices, that is, analyzing whether the current anomaly is verified by the collaborative efforts of other devices.

[0075] During the analysis of correlated data, if multiple devices detect anomalies deviating from the health baseline, this indicates a high degree of confidence in the anomaly. For example, if a smart collar detects an abnormally high heart rate in a pet, a camera captures abnormal restlessness, and a smart feeder shows the pet's food intake is significantly off-range, the monitoring results from multiple devices corroborate each other, suggesting the pet may indeed have a health problem. In this case, a positive correction factor is generated. The purpose of the positive correction factor is to positively adjust the initial comprehensive anomaly score, increasing its value to more accurately reflect the severity of the pet's health abnormality, making subsequent warnings more timely and accurate.

[0076] When analyzing correlated data, if an anomaly is found to originate from a single, instantaneous fluctuation and lacks verification from other devices, the reliability of such an anomaly is relatively low. For example, a smart collar might record a brief anomaly in a pet's heart rate at a particular moment, but the camera shows the pet is in a normal resting state, and other devices do not detect any related anomalies. This is likely due to a single measurement error or a momentary fluctuation caused by the pet's accidental behavior. In such cases, a negative correction factor is generated. This negative correction factor adjusts the initial comprehensive anomaly score negatively, reducing the anomaly score value and avoiding false alarms caused by inaccurate data from a single instance, thus improving the accuracy and reliability of health alerts.

[0077] By obtaining the multi-source cross-validation correction factor through the above steps and integrating it with the preliminary comprehensive anomaly score, a more accurate target comprehensive anomaly score can be generated. This provides a reliable basis for outputting graded health warning information based on the relationship between the score and the preset grading threshold, thereby better protecting the health of pets.

[0078] In some embodiments, step S104 above, which involves outputting graded health warning information based on the relationship between the comprehensive anomaly score and a preset grading threshold, specifically includes: Multiple grading thresholds are set to classify health risk levels. These thresholds are used to divide the abnormal score range into multiple warning level ranges from low to high. The warning level ranges include normal state range, alert range, recommended observation range, and emergency alarm range. The comprehensive anomaly score is compared with each grading threshold to determine the warning level range into which the comprehensive anomaly score falls. Based on the warning level range that the comprehensive anomaly score falls into, the corresponding warning level output rules are matched to generate corresponding graded health warning information. The graded health warning information includes the anomaly type, the degree of anomaly, the recommended intervention measures, and the corresponding target pet identification.

[0079] In this embodiment, multiple grading thresholds are set to classify health risk levels. These thresholds are determined comprehensively based on a large amount of experimental data, the experience of pet health experts, and the needs of real-world scenarios. For example, by analyzing a large amount of pet health data, the comprehensive abnormal score range for different health states can be identified. The abnormal score range is divided into multiple warning level ranges from low to high, typically including a normal state range, a warning interval, a recommended observation range, and an emergency alert range. The normal state range indicates that the pet's health is good, and all indicators are within the normal range; the warning interval means that the pet may have experienced some minor changes, requiring the pet owner to continue monitoring; the recommended observation range indicates that the pet may have potential health problems, and the owner is advised to further observe the pet's behavior and condition; the emergency alert range indicates that the pet may have experienced a more serious health abnormality, requiring immediate action.

[0080] After obtaining the comprehensive anomaly score of the pet, it is compared with each set grading threshold. This process is similar to determining the position interval of a point on the number line. For example, assuming the set grading thresholds are T1, T2, and T3 respectively, the comprehensive anomaly score S is compared with these thresholds in sequence. If S < T1, it falls into the normal state interval; if T1 ≤ S < T2, it falls into the attention reminder interval; if T2 ≤ S < T3, it falls into the recommended observation interval; if S ≥ T3, it falls into the emergency alarm interval. In this way, the early warning level interval into which the comprehensive anomaly score falls is accurately determined.

[0081] According to the early warning level interval into which the comprehensive anomaly score falls, the corresponding early warning level output rules are matched. Each early warning level interval has pre-set early warning information output rules, which specify in detail the information content to be included at different levels. For example, when the comprehensive anomaly score falls into the normal state interval, the generated graded health early warning information mainly indicates that the pet's current health status is good and no special intervention is required. When it falls into the attention reminder interval, in addition to indicating that the pet is in the attention reminder state, the early warning information will also prompt the pet owner to pay attention to changes in the pet's daily behaviors such as diet and exercise. If it falls into the recommended observation interval, the early warning information will clearly identify the type of anomaly, such as possible abnormal activity level or abnormal heart rate, etc., give the degree of anomaly as medium, and recommend that the owner observe the pet's symptoms in the next period of time and record relevant information. When it falls into the emergency alarm interval, the early warning information will elaborate on the type of anomaly and the severe degree of the anomaly, give urgent recommended intervention measures, such as sending the pet to the doctor immediately, etc., and clearly identify the corresponding target pet identity label to ensure that the pet owner can clearly know which pet has an urgent health problem.

[0082] Through the above steps, it is possible to accurately output graded health early warning information based on the relationship between the comprehensive anomaly score and the preset grading threshold, helping pet owners understand the pet's health status in a timely manner and take corresponding measures to ensure the pet's health.

[0083] As Figure 6 shown, through multiple devices, various data of the pet are collected in real time, covering aspects such as weight, body temperature, heart rate, respiration, sleep, and exercise. The collected real-time monitoring data is compared and analyzed with the pre-established individual reference characteristics of the pet, and the attribution confidence level is calculated. Then, a judgment is made according to the preset attribution confidence level threshold: If the threshold is reached (high confidence level): it indicates that the current data can be more accurately attributed to a certain pet. This data is used as a high-confidence event segment and enters the individual health baseline update link.

[0084] If the threshold (low confidence) is not reached, it means that the accuracy of the data attribution is questionable. It is marked as a low confidence event fragment, stored in the pending confirmation cache, and will not be used for health baseline updates for the time being. Further analysis or manual intervention may be required.

[0085] Individualized features of pets are extracted from high-confidence event fragments, reflecting their unique physiological and behavioral characteristics. By combining historical data with newly extracted features, the long-term health baseline of pets is dynamically adjusted to better adapt to changes in their health status. Different weights are assigned to different data points based on their quality and reliability, ensuring that high-quality data plays a greater role in the baseline update process. The updated health baseline will more accurately reflect the normal health status of pets.

[0086] Low-confidence event fragments are stored in an unconfirmed cache, and this data requires further processing and verification. Possible methods include manual verification and analysis in conjunction with more monitoring data to determine their correct attribution before deciding whether to use them for health baseline updates.

[0087] Real-time monitoring data is compared with the updated individual health baseline to calculate the individual anomaly degree of each feature. A comprehensive anomaly score is then obtained through weighted fusion and multi-source cross-validation. Multi-source cross-validation analyzes correlated data collected by multiple monitoring devices within the same time window to verify the authenticity and severity of the anomalies, generating correction factors to adjust the comprehensive anomaly score.

[0088] Based on the relationship between the comprehensive anomaly score and the preset grading threshold, a tiered health warning system is output. This system is divided into three levels: Emergency Alert (high risk, immediate notification), Recommended Observation (medium risk, manual confirmation required), and Attention Reminder (low risk, continuous monitoring), allowing pet owners or relevant personnel to take appropriate measures according to the different warning levels.

[0089] The entire process forms a complete closed loop, from data collection to identification and ownership determination, and then to health baseline updates and abnormal warnings, enabling real-time and accurate monitoring and early warning of pet health status, effectively solving the problems of data ownership and health monitoring in multi-pet environments.

[0090] In summary, this invention, through a complete technical chain encompassing "unified access across multiple devices - candidate event slicing - individual identity attribution - attribution confidence control - individual health baseline modeling - multi-source trend analysis and hierarchical early warning," forms a cross-device pet health management method suitable for multi-pet households. Compared with existing technologies, this invention not only solves the problem of inaccurate data allocation in multi-pet environments but also significantly improves the reliability of individual health records and the accuracy of abnormal early warnings, providing technical support for upgrading pet health monitoring systems from single recording tools to intelligent, personalized, and continuous health management platforms.

[0091] Reference Figure 7 An embodiment of the present invention provides a multi-pet cross-device pet identity attribution and health early warning system 7, the system 7 specifically including: The data processing module 701 is used to uniformly access and time-series align the raw monitoring data collected by various pet monitoring terminals, and to divide the continuous data into candidate behavioral event segments according to preset event rules. The scoring calculation module 702 is used to extract event feature vectors based on candidate behavioral event fragments, and perform multi-feature weighted matching with the individual reference feature vectors of each pet pre-established under the same account to calculate the affiliation score of each pet, determine the target affiliation object based on the affiliation score, and calculate the affiliation confidence. The model training module 703 is used to take high-confidence event fragments with an attribution confidence higher than a preset confidence threshold as valid samples for training and dynamic updating of the individual health baseline model of the corresponding pet. The graded early warning module 704 is used to compare real-time monitoring data with the output parameters of the corresponding pet's individual health baseline model, calculate the single anomaly degree of each dimension feature, and obtain a comprehensive anomaly score through weighted fusion and multi-source cross-validation. Based on the relationship between the comprehensive anomaly score and the preset graded threshold, it outputs graded health early warning information.

[0092] It is understandable that, such as Figure 1 The content of the multi-pet cross-device pet identity attribution and health warning method embodiments shown herein is applicable to the multi-pet cross-device pet identity attribution and health warning system embodiments. The specific functions implemented by the multi-pet cross-device pet identity attribution and health warning system embodiments are as follows: Figure 1 The illustrated multi-pet cross-device pet identification and health alert method is the same as the one shown, and achieves the same beneficial effects. Figure 1 The beneficial effects achieved by the multi-pet cross-device pet identity attribution and health early warning method embodiment shown are also the same.

[0093] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0095] Reference Figure 8 The present invention also provides a computer device 8, including: a memory 802 and a processor 801, and a computer program 803 stored on the memory 802. When the computer program 803 is executed on the processor 801, it implements the multi-pet cross-device pet identity attribution and health warning method as described in any of the above methods.

[0096] The computer device 8 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 8 may include, but is not limited to, a processor 801 and a memory 802. Those skilled in the art will understand that... Figure 8 The computer device 8 is merely an example and does not constitute a limitation on the computer device 8. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0097] The processor 801 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0098] In some embodiments, the memory 802 may be an internal storage unit of the computer device 8, such as a hard disk or memory of the computer device 8. In other embodiments, the memory 802 may be an external storage device of the computer device 8, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 8. Furthermore, the memory 802 may include both internal and external storage units of the computer device 8. The memory 802 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 802 can also be used to temporarily store data that has been output or will be output.

[0099] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-pet cross-device pet identity attribution and health warning method as described in any of the above methods.

[0100] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0101] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0102] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0103] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0104] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for pet identification and health early warning across multiple devices, characterized in that, The method specifically includes: The system integrates and aligns the raw monitoring data collected from various pet monitoring terminals in a unified manner, and segments the continuous data into candidate behavioral event segments according to preset event rules. Based on candidate behavioral event fragments, event feature vectors are extracted and matched with pre-established individual reference feature vectors of each pet under the same account using multi-feature weighted matching. The affiliation score of each pet is calculated, and the target affiliation object is determined based on the affiliation score, and the affiliation confidence is calculated. High-confidence event segments with attribution confidence levels higher than a preset confidence threshold are used as valid samples for training and dynamic updating of the corresponding pet's individual health baseline model. The real-time monitoring data is compared with the output parameters of the corresponding pet's individual health baseline model to calculate the individual anomaly degree of each feature. A comprehensive anomaly score is obtained through weighted fusion and multi-source cross-validation. Based on the relationship between the comprehensive anomaly score and the preset grading threshold, graded health warning information is output.

2. The method according to claim 1, characterized in that, The process involves unified access and time-series alignment of raw monitoring data collected from multiple pet monitoring terminals, and segmenting continuous data into candidate behavioral event fragments according to preset event rules. Specifically, this includes: The raw monitoring data from multiple types of pet monitoring terminals are uniformly accessed, and each data point is appended with a device identifier, account identifier, collection timestamp, and spatial location identifier to form multi-source data. The pet monitoring terminals include smart nests, smart feeders, smart water dispensers, litter box monitoring devices, environmental sensors, and visual acquisition terminals. Multi-source data is preprocessed by outlier removal, missing value correction, clock correction, and sampling frequency unification to transform data from scattered sources, heterogeneous formats, and different time granularities to the same time axis, thus obtaining time-aligned continuous data. Based on a preset time window or behavioral state change rules, the continuous data after time alignment is sliced ​​into event segments to generate candidate behavioral event fragments. The candidate behavioral event fragments include eating events, drinking events, sleeping events, staying events, activity events, and toileting events.

3. The method according to claim 1, characterized in that, The process of extracting event feature vectors based on candidate behavioral event fragments and performing multi-feature weighted matching with pre-established individual reference feature vectors for each pet under the same account to calculate the ownership score of each pet includes: For each candidate behavioral event fragment, multi-dimensional features for identity recognition are extracted, and a corresponding event feature vector is constructed. The event feature vector includes static features for characterizing the inherent attributes of an individual, dynamic behavioral features for reflecting the individual's living habits, spatial location features for indicating the location where the event occurred, and multi-device consistency features for reflecting the collaborative relationship of multiple devices. Retrieve the individual reference feature vectors pre-established for each pet under the same account. The individual reference feature vectors are generated based on the historical identity files of the corresponding pets and contain feature items of the same dimension as the event feature vectors. Calculate the static feature matching degree, dynamic behavior similarity, spatial location matching degree, and multi-device consistency verification value between the event feature vector of each candidate behavior event segment and the individual reference feature vector of each pet. According to the preset weight allocation rules, the calculated static feature matching degree, dynamic behavior similarity, spatial location matching degree and multi-device consistency verification value are weighted and fused to generate the ownership score of each pet corresponding to each candidate behavior event fragment.

4. The method according to claim 3, characterized in that, The calculation of the static feature matching degree, dynamic behavior similarity, spatial location matching degree, and multi-device consistency verification value between the event feature vector of each candidate behavioral event segment and the individual reference feature vector of each pet specifically includes: By comparing the weight measurement value, body shape contour parameters and pressure distribution data in the event feature vector with the corresponding historical static archive data in the individual reference feature vector, the static feature matching degree is calculated. The static feature matching degree is used to measure the degree of conformity between the current event and the target pet in terms of inherent physiological attributes. By analyzing the periodic match between the activity occurrence time, activity duration, and behavior frequency distribution in the event feature vector and the corresponding historical behavior rhythm pattern in the individual reference feature vector, dynamic behavior similarity is calculated. The dynamic behavior similarity is used to measure the degree of matching between the current event and the target pet in terms of time series behavior habits. By comparing the spatial location identifier carried in the event feature vector with the historical area stay preference recorded in the individual reference feature vector, the spatial location matching degree is calculated. The spatial location matching degree is used to measure whether the current event location is within the target pet's daily activity range. By acquiring associated data collected by multiple monitoring devices within the same time window, the coexistence relationship or mutual exclusion constraint between the current event reflected by the event feature vector and the detection results of other devices is analyzed, and a multi-device consistency verification value is calculated. The multi-device consistency verification value is used to verify the logical self-consistency of the current event from the perspective of multi-source monitoring.

5. The method according to claim 1, characterized in that, The process of determining the target affiliation object based on the affiliation score and calculating the affiliation confidence score specifically includes: Compare the ownership scores of the same candidate behavior event segment with the pets under the same account, and determine the pet with the highest ownership score as the target ownership object of the current candidate behavior event segment; Extract the score difference between the highest and second-highest attribution scores of the target object, calculate the stability index of the attribution score of the target object in the historical matching process, and obtain the consistency index between multiple source devices in the current time window. The attribution confidence of the current candidate behavioral event fragment is generated by comprehensively calculating the score difference, stability index, and consistency index.

6. The method according to claim 5, characterized in that, The step of using high-confidence event segments with attribution confidence levels exceeding a preset confidence threshold as valid samples for training and dynamic updating of the corresponding pet's individual health baseline model specifically includes: When the attribution confidence is greater than or equal to the preset confidence threshold, the current candidate behavioral event segment is determined to be a high-confidence event segment and is used as a valid sample; otherwise, the current candidate behavioral event segment is marked as a low-confidence event to be confirmed. The valid samples are preprocessed to generate a multidimensional health monitoring sequence, which includes the amount of activity, resting time, sleep duration, number of meals, number of water drinks, number of toilet visits, respiratory rate, heart rate, weight and corresponding environmental parameters within a unit time window. Based on a time-series modeling network for continuous health monitoring data of pets, an independent individual health baseline model is constructed for each pet. The multidimensional health monitoring sequences corresponding to multiple valid samples belonging to the corresponding pets are used as training data, and the stable monitoring results of multiple consecutive days are used as baseline labels to initially train the individual health baseline model and generate individual health baseline parameters. Based on the confidence level of the new valid samples, corresponding sample update weights are assigned, and the model parameters at the current time are corrected according to the sample update weights.

7. The method according to claim 6, characterized in that, The process involves using multidimensional health monitoring sequences corresponding to multiple valid samples belonging to the corresponding pet as training data, and using stable monitoring results from multiple consecutive days as baseline labels to initially train the individual health baseline model, generating individual health baseline parameters, specifically including: A training sample sequence is constructed based on the multidimensional health monitoring sequences corresponding to multiple sets of valid samples belonging to the corresponding pets; Based on the stable monitoring results over several consecutive days, health monitoring data within the corresponding time window are extracted as baseline labels. These baseline labels are used to characterize the normal physiological fluctuation range and behavioral patterns of pets in a healthy state. The training sample sequence is input into a preset temporal modeling network for model forward propagation. The temporal modeling network is used to receive multidimensional health monitoring sequences through the input layer, perform preliminary feature extraction through the feature encoding layer, and then send them to the temporal feature extraction layer to capture the behavioral rhythms and physiological changes in the continuous time dimension. The output layer generates predicted individual health baseline parameters, which include activity baseline, sleep pattern baseline, food and water intake baseline, toilet cycle baseline, respiratory rate baseline, heart rate baseline, and weight fluctuation baseline.

8. The method according to claim 1, characterized in that, The process involves comparing real-time monitoring data with the output parameters of the corresponding pet's individual health baseline model, calculating the individual anomaly degree of each feature, and obtaining a comprehensive anomaly score through weighted fusion and multi-source cross-validation. Specifically, this includes: For real-time monitoring data of pets that have been successfully assigned to the target pet, extract the multi-dimensional monitoring feature vector at the current moment; The real-time monitoring feature value of each dimension in the multidimensional monitoring feature vector is compared with the benchmark value of the corresponding dimension in the individual health baseline parameters. Based on the relative magnitude of the deviation of the real-time monitoring feature value from the benchmark value, and combined with the normal fluctuation range of the real-time monitoring feature in the health baseline, the single abnormality degree corresponding to each real-time monitoring feature is calculated. According to the preset dimensional feature weight allocation rules, the individual anomaly degrees of each dimension's real-time monitoring features are weighted and fused to generate a preliminary comprehensive anomaly score. The multi-source cross-validation correction factor is obtained based on the multi-source cross-validation mechanism. The preliminary comprehensive anomaly score is then fused with the multi-source cross-validation correction factor to generate the target comprehensive anomaly score.

9. The method according to claim 8, characterized in that, The process of obtaining the multi-source cross-validation correction factor based on the multi-source cross-validation mechanism specifically includes: Acquire correlated data collected by multiple monitoring devices within the same time window, and analyze whether the current anomaly has been verified by other devices. A positive correction factor is generated when multiple devices detect abnormalities that deviate from the health baseline. A negative correction factor is generated when the anomaly originates from a single instantaneous fluctuation and lacks verification from other devices.

10. A multi-pet cross-device pet identification and health early warning system, characterized in that, The system specifically includes: The data processing module is used to uniformly access and time-series align the raw monitoring data collected by various pet monitoring terminals, and to segment continuous data into candidate behavioral event segments according to preset event rules. The rating calculation module is used to extract event feature vectors based on candidate behavioral event fragments, and perform multi-feature weighted matching with the individual reference feature vectors of each pet pre-established under the same account to calculate the affiliation score of each pet, determine the target affiliation object based on the affiliation score, and calculate the affiliation confidence. The model training module is used to take high-confidence event fragments with an attribution confidence level higher than a preset confidence threshold as valid samples for training and dynamic updating of the individual health baseline model of the corresponding pet. The graded early warning module compares real-time monitoring data with the output parameters of the corresponding pet's individual health baseline model, calculates the individual anomaly degree of each feature, and obtains a comprehensive anomaly score through weighted fusion and multi-source cross-validation. Based on the relationship between the comprehensive anomaly score and the preset graded threshold, it outputs graded health early warning information.

Images