A pet health report intelligent generation method and system based on multi-factor weight

Abstract

This invention discloses an intelligent pet health report generation method and system based on multi-factor weighting. The method specifically includes: using a hierarchical multi-factor weighting model to calculate the adjustment coefficients of each factor in a weight factor library in parallel, synthesizing the original weights through element-wise multiplication, and outputting the dynamic weights of each health indicator after normalization; based on the dynamic weights, weighted summing of the standardized scores of each health indicator, identifying abnormal changes in the indicators using a trend analysis algorithm, and determining the pet's health risk level based on the weighted score and the number of abnormal indicators; based on the health risk level, combining dynamic weights and trend analysis, generating a health report including data summaries, anomaly interpretations, trend analysis, and personalized suggestions using natural language generation technology. This invention achieves context-aware weighting of data through a dynamic weight analysis engine, and automatically outputs structured, personalized, and interpretable health reports by combining natural language generation technology.

Description

Technical Field

[0001] This invention relates to the field of intelligent health monitoring and natural language generation technology, and in particular to a method and system for intelligently generating pet health reports based on multi-factor weights. Background Technology

[0002] With the booming development of the pet economy and the widespread adoption of smart home devices, pet health monitoring has become a research hotspot for the application of artificial intelligence in vertical industries.

[0003] In academic research, studies on methods for generating pet health reports are constantly deepening both domestically and internationally. The domestic Fun-ASR team (2026) utilizes a high-precision speech recognition system (Fun-ASR) and AI technology to directly convert spoken content into standardized veterinary electronic medical records (Keyu An etc. (2025). "Fun-ASR Technical Report." arXiv:2509.12508, Sep 2025. https: / / doi.org / 10.48550 / arXiv.2509.12508); the patent development team for the intelligent pet health monitoring and assessment system (2025) uses deep learning algorithms to effectively solve the problem of high false alarm rates from single data sources (Publication No.: CN120241007A). The CVS Group technical team abroad (2025) is able to generate predictive health risk reports (CVS Group. Quality Improvement and Research Report 2025 [R / OL]. (2025). https: / / www.cvsukltd.co.uk / news / cvs-publishes-quality-improvement-and-research-report / .) and (Simpson C. CVS launches parasite risk assessment initiative [N / OL]. Vet Times, (2025). https: / / www.vettimes.com / news / vets / small-animal-vets / cvs-launches-parasite-risk-assessment-initiative.); the PetPartner R&D team (see application link: https: / / petpartner.app / ) (2025-2026) generates comprehensive solutions covering multiple aspects such as medical care, nutrition, and behavior.

[0004] While current intelligent monitoring reports have made some progress in data collection, fundamental limitations remain in the core areas of intelligent decision-making and accurate assessment. First, most existing systems output relatively simple data, merely listing raw data without in-depth analysis and interpretation, making it difficult for users to extract valuable health information. Second, the generated reports are often generic, lacking personalized content tailored to the individual characteristics and health conditions of different pets, failing to meet pet owners' needs for accurate health assessments. Third, they rely heavily on static thresholds and independent analysis of single-modal data, employing a "data stacking + simple comparison" model, lacking dynamic adjustment and flexible decision-making capabilities, and struggling to cope with complex and changing household scenarios. Furthermore, when assessing pet health, they fail to allocate weights reasonably according to the importance of different factors, resulting in an inaccurate representation of the degree of influence of each indicator on the final health assessment result, affecting the accuracy of the assessment. In addition, the system-generated reports often only provide the health assessment results, lacking detailed explanations of the assessment process and basis, making it difficult for users to understand the health implications behind the data, reducing user trust in the system. Finally, the system design failed to fully consider the actual needs of users and could not continuously optimize its performance based on user feedback and subsequent diagnostic results, resulting in stagnant system performance and an inability to adapt to ever-changing real-world application scenarios.

[0005] The root causes of the above problems mainly lie in the following four aspects: First, a biased technical positioning: Most existing systems position themselves as "data collection tools," focusing too much on data collection while neglecting data interpretation and analysis. This results in outputs that merely display raw data, failing to provide users with valuable health information. Second, limitations in algorithm design: Most systems remain at the stage of single-modal analysis or simple data splicing, lacking deep fusion mechanisms. They cannot fully explore the correlation characteristics between different modalities of data, making it difficult to effectively integrate and utilize multi-source data. Third, a lack of context awareness: Dynamic factors such as environmental factors, behavioral states, and data quality are not incorporated into the decision-making model, leading to rigid judgment logic. This makes it impossible to adjust the evaluation strategy in real time according to the actual situation, easily resulting in misjudgments and false alarms. Fourth, a lack of feedback loops: The system design is disconnected from actual user needs. Without an effective feedback mechanism, the system cannot be continuously optimized and improved based on usage effects, resulting in stagnant system performance and an inability to meet the ever-growing needs of users. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for intelligent generation of pet health reports based on multi-factor weighting. This method integrates pet physiological data, behavioral data, environmental data, and historical records. It achieves context-aware weighting of data through a dynamic weighting analysis engine and combines natural language generation technology to automatically output structured, personalized, and interpretable health reports, thereby solving at least one of the aforementioned problems in the prior art.

[0007] In a first aspect, the present invention provides a method for intelligently generating pet health reports based on multi-factor weights, the method specifically comprising: Collect pet physiological data, behavioral data, environmental data, and historical data, and perform data cleaning, filtering, and feature extraction to generate multi-dimensional quantitative features; Based on multi-dimensional quantitative characteristics, a weight factor library is constructed that includes environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors. A hierarchical multi-factor weighted model is used to calculate the adjustment coefficients of each factor in the weight factor library in parallel, and the original weights are synthesized by element-wise multiplication. After normalization, the dynamic weights of each health indicator are output. Based on dynamic weights, the standardized scores of each health indicator are weighted and summed, and trend analysis algorithms are used to identify abnormal changes in the indicators. The health risk level of the pet is determined based on the weighted score and the number of abnormal indicators. Based on health risk levels, combined with dynamic weights and trend analysis, a health report is generated using natural language generation technology, which includes data summaries, anomaly interpretations, trend analysis, and personalized suggestions. The weight allocation and adjustment reasons in the decision-making process are also embedded into the health report. The health report is output to the user in a structured format, and the weight strategy is optimized online using a reinforcement learning algorithm based on user interaction feedback and subsequent diagnostic results.

[0008] Secondly, the present invention provides a pet health report intelligent generation system based on multi-factor weighting, the system specifically comprising: The data acquisition module is used to collect pet physiological data, behavioral data, environmental data, and historical data, and to perform data cleaning, filtering, and feature extraction to generate multi-dimensional quantitative features. The weighting factor module is used to build a weighting factor library based on multi-dimensional quantitative features, including environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors. The weight adjustment module is used to calculate the adjustment coefficients of each factor in the weight factor library in parallel using a hierarchical multi-factor weighted model, synthesize the original weights through element-wise multiplication, and output the dynamic weights of each health indicator after normalization. The health assessment module is used to perform weighted summation of standardized scores for each health indicator based on dynamic weights, combine trend analysis algorithms to identify abnormal changes in the indicators, and determine the pet's health risk level based on the weighted score and the number of abnormal indicators. The report generation module is used to generate health reports based on health risk levels, combined with dynamic weights and trend analysis, using natural language generation technology. The reports include data summaries, anomaly interpretations, trend analysis, and personalized suggestions, and embed the weight allocation and adjustment reasons from the decision-making process into the health reports. The strategy optimization module is used to output health reports to users in a structured format and optimize the weight strategy online using reinforcement learning algorithms based on user interaction feedback and subsequent diagnostic results.

[0009] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention integrates pet physiological data, behavioral data, environmental data, and historical records. It achieves context-aware weighting of data through a dynamic weight analysis engine and combines natural language generation technology to automatically output structured, personalized, and interpretable health reports. It is suitable for algorithm implementation in scenarios such as intelligent monitoring of family pets, auxiliary diagnosis of pet medical care, and pet insurance underwriting. 2. This invention collects multi-source data and generates multi-dimensional quantitative features, constructs a hierarchical multi-factor weighted model, realizes dynamic weight calculation of health indicators, combines trend analysis to generate intelligent health reports and optimize weight strategies online, providing accurate, dynamic and personalized support for pet health management; 3. This invention extracts multiple factors from multi-dimensional quantitative features to construct a weight factor library, comprehensively considering various factors that affect the weight of pet health indicators, making the weight assessment more scientific and reasonable; 4. This invention uses a hierarchical multi-factor weighted model to calculate and process dynamic weights in parallel, which can adjust the weights of each health indicator in real time according to the individual pet's condition and various factors, thereby improving the accuracy of health assessment. 5. This invention uses appropriate methods to calculate adjustment coefficient vectors for different factors, ensuring that the adjustment of the weights of each factor on health indicators is scientific and reasonable, and improving the accuracy of dynamic weight calculation. 6. This invention calculates a comprehensive health score based on dynamic weights, combines trend analysis to identify abnormal changes in indicators, and reasonably determines the level of health risk, providing a comprehensive and accurate basis for assessing the health status of pets. 7. This invention generates a health report with rich content based on multiple information sources and embeds relevant factors of the decision-making process into the report, so that pet owners can clearly understand their pet's health status and the basis for recommendations; 8. Based on user interaction feedback and diagnostic results, this invention uses reinforcement learning algorithms to optimize weight strategies online, so that the weight allocation continuously adapts to the actual situation, thereby improving the accuracy and practicality of health reports; 9. This invention constructs an individual-specific weight prediction model to achieve a personalized optimal weight allocation scheme for each pet, further improving the accuracy and relevance of pet health assessment and report generation. Attached Figure Description

[0010] 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.

[0011] Figure 1 This is a flowchart illustrating a method for intelligently generating pet health reports based on multi-factor weights, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a pet health report intelligent generation system based on multi-factor weighting, provided in an embodiment of the present invention. Detailed Implementation

[0012] 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.

[0013] The following is an explanation of algorithm-related terminology: GBDT (Gradient Boosting Decision Tree) is an ensemble learning algorithm that iteratively trains a decision tree and accumulates predictions, making it suitable for tasks involving learning historical weights. Q-learning: Q-learning is a model-free reinforcement learning algorithm that maintains a Q-table to record state-action values ​​and uses it to optimize weights. NLG: Natural Language Generation, a technique for converting structured data into natural language text; Min-Max Scaler: Min-Max normalization is a linear transformation method that maps data to the interval [0,1]. Sliding window filtering: a signal processing method that smooths data by using the mean or median within a sliding window to eliminate noise; Health baseline: The normal range of various indicators for a pet in a healthy state, dynamically updated based on historical data.

[0014] 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 an embodiment of the intelligent pet health report generation method based on multi-factor weighting disclosed in this invention is shown below: S101 collects physiological, behavioral, environmental, and historical data of pets, and performs data cleaning, filtering, and feature extraction to generate multi-dimensional quantitative features.

[0015] In this embodiment, environmental sensors deployed in the pet activity area collect data such as room temperature, humidity, light intensity, and noise. Specifically, physiological data processing algorithms are used to dynamically update the health reference range for each pet by collecting physiological signals such as body temperature, heart rate, and respiratory rate. Behavioral data processing algorithms are used to collect behavioral data such as activity level, vocalization frequency, food intake, and water intake to extract behavioral pattern characteristics such as the pet's active periods and resting patterns. Environmental data processing algorithms are used to collect environmental data such as room temperature, humidity, light intensity, and noise level, converting continuous values ​​into discrete states to facilitate subsequent weight calculations. A historical record management algorithm is used to construct a pet health record database, storing static record data such as breed, age, sex, medical history, medication records, and vaccination information.

[0016] After completing the multi-source data collection, the data needs to be systematically organized, classified, and labeled to construct a data sample set for algorithm training and testing. The labeled data is divided into training set, validation set, and test set in a 7:2:1 ratio to support algorithm training, parameter tuning, and performance evaluation, respectively.

[0017] In physiological data processing, a sliding window filtering algorithm was used to denoise the raw body temperature and heart rate data, removing sensor noise and transient outliers. After denoising, NumPy and Pandas libraries were used to extract quantitative features such as mean, variance, maximum, minimum, and fluctuation coefficient.

[0018] In behavioral data processing, anomaly detection algorithms are used to remove outliers caused by equipment failures, and then trend features such as month-on-month change rate, year-on-year change rate, and trend slope are calculated.

[0019] In environmental data processing, a moving average algorithm is used for smoothing to eliminate instantaneous fluctuations, while converting continuous values ​​into discrete state codes.

[0020] In the historical record processing, a pet health record database is built to store static record data and dynamically calculate individual health baselines based on historical monitoring data.

[0021] S102, based on multi-dimensional quantitative characteristics, constructs a weight factor library that includes environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors.

[0022] In this embodiment, the MinMaxScaler function from the Scikit-learn library is used to normalize various features, mapping features of different scales to the [0,1] interval. After normalization, a weight factor library is constructed, including five categories: environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors, as shown in the table below:

[0023] S103 uses a hierarchical multi-factor weighted model to calculate the adjustment coefficients of each factor in the weight factor library in parallel, and synthesizes the original weights by element-wise multiplication. After normalization, it outputs the dynamic weights of each health indicator.

[0024] In this embodiment, a hierarchical dynamic weighted model is adopted. First, a basic weight vector is obtained based on the pet's static profile (breed, age, medical history). Subsequently, the adjustment coefficient vectors for each indicator based on five categories of factors are calculated in parallel: environmental factors are generated through rule mapping; individual factors are dynamically adjusted based on physiological baselines; behavioral factors utilize the Isolation Forest algorithm to automatically identify the deviation between current behavior and individual historical patterns and generate adjustment coefficients; data quality factors also employ the Isolation Forest algorithm to detect sensor signal anomalies and reduce weight accordingly; and historical pattern factors calculate the deviation by comparing the current value with historical data from the same period. Finally, the base weights and the five adjustment coefficient vectors are multiplied element-wise to synthesize the original weights. After normalization, the original weights are converted into a probability distribution to ensure that the sum of all weights is 1, thus obtaining the final dynamic weights. This algorithm ensures that each indicator receives independent contextual adjustments through element-wise multiplication, and introduces isolated forests to achieve behavioral anomaly identification and data quality detection in unsupervised scenarios, enabling the system to possess true individualized contextual awareness capabilities.

[0025] Each factor outputs an adjustment coefficient vector with the same dimension as the base weights, where: Environmental factors Generated through mapping using predefined rules. For example, at a room temperature of 32℃, the body temperature adjustment coefficient is 1.21, respiration is 1.175, and activity level is maintained at 1.0. Individual factors Calculated based on the deviation of real-time indicators from the individual baseline. When the heart rate deviates from the baseline by 1.5 times the standard deviation, the adjustment factor = 1 + 0.2 × deviation. Behavioral factors This paper introduces an isolated forest model to identify behavioral anomalies. A dedicated model is trained based on 8-dimensional behavioral features to obtain behavioral anomaly scores. Generate adjustment coefficient (First activity level coefficient) =0.5, first heart rate coefficient =0.3, First body temperature coefficient =0.1); Data quality factor An isolated forest approach is used to detect signal anomalies. A global model is trained based on quality features to obtain quality anomaly classification. Generate adjustment coefficient (Second heart rate coefficient) =0.8, first respiratory coefficient =0.7, First body temperature coefficient =0.5, Second Activity Coefficient =0.3); Historical pattern factors Calculate the deviation by comparing the current value with the historical value for the same period. Generate adjustment coefficient (Second respiratory coefficient) =0.3, third heart rate coefficient =0.2).

[0026] Based on individual historical data, a personalized weight optimization model is constructed. First, historical monitoring data and corresponding user feedback / veterinary diagnosis results are collected for each pet to build a training sample set. Then, the gradient boosting tree (GBDT) algorithm is used to train an individual-specific weight prediction model. Finally, the trained model is deployed to a real-time system to predict the optimal weight allocation scheme based on the current context characteristics.

[0027] Specifically, the Gradient Boosting Tree (GBDT) algorithm is used to train a dedicated weight prediction model for each pet. The training process is as follows: First, historical monitoring data (including environmental, behavioral, and physiological features) and corresponding health outcomes (such as veterinary diagnoses or user feedback) for the pet over the past 3-6 months are collected. Then, an inverse optimization algorithm is introduced to generate training labels—for each historical data point, the "known health outcome" is used as the objective, and the "weight allocation that best explains the outcome" is solved in reverse. That is, under the premise of satisfying medical prior constraints (such as body temperature weight not being too low and short-nosed cat breathing weight having a minimum requirement), a set of weights is found to maximize the match between the weighted health score and the diagnosis result by solving a constraint optimization problem. This transforms the abstract diagnosis into a specific weight vector, which serves as the learning objective of the GBDT model. Next, the contextual features of the historical moment are used as input X, and the optimal weights obtained through inverse optimization are used as output Y to construct a training sample set. The model structure employs a Lightweight Gradient Boosting Tree (LightGBM). The input layer consists of 50-80 dimensional contextual feature vectors (e.g., room temperature 28.5℃, short-nosed cat breed, post-exercise state, etc.), and the output layer is an M-dimensional weight vector (e.g., weights for various indicators such as body temperature 0.28, heart rate 0.32, respiration 0.25, etc.). Iterative training is performed using 200 decision trees of depth 6, with each tree fitting the prediction residual of the previous tree. Ultimately, a non-linear mapping relationship is learned: "Given the current context, how much weight should be assigned to each indicator?" After training, when new real-time data arrives, the model directly outputs a personalized optimal weight allocation based on the current contextual features. The model is retrained monthly using the latest accumulated data, allowing the weight strategy to continuously adapt to changes in the pet's individual characteristics (e.g., age, changes in medical history, seasonal patterns, etc.), thus achieving a leap from "general medical guidelines" to "personalized models."

[0028] S104, based on dynamic weights, performs a weighted summation of the standardized scores of each health indicator, combines a trend analysis algorithm to identify abnormal changes in the indicators, and determines the pet's health risk level based on the weighted score and the number of abnormal indicators.

[0029] In this embodiment, during the health status assessment process, a dynamic weight matrix is ​​first applied to calculate the weighted average of each indicator to obtain a comprehensive health score. The formula is: Health Score = Σ( × ),in For dynamic weights, Standardized scores are calculated for each indicator based on individual reference ranges. Building upon this, and combining time-series data, a trend analysis algorithm is used to calculate the month-on-month change rate, year-on-year change rate, and trend slope for each indicator. These are compared with the past 7-day average, 30-day average, and historical baseline to identify significant trends. Finally, based on the weighted health score and the number of abnormal indicators, a risk level determination algorithm is used to classify the pet's health status into four levels: "Normal" if all indicators are normal and show no significant trend change; "Attention" if one or two indicators are slightly abnormal or show a significant downward trend; "Warning" if two or three indicators are significantly abnormal or multiple indicators show a continuous downward trend; and "Seek Medical Attention" if key indicators are severely abnormal or meet preset emergency rules. This assessment process achieves a systematic judgment from data weighting and trend identification to risk classification, providing a scientific basis for the subsequent generation of personalized reports.

[0030] In health status assessment, a dynamic weight matrix is ​​first applied to weight each indicator to obtain a comprehensive health score. Specifically, standardized scores for each indicator are calculated based on an individual's health baseline, then multiplied by their corresponding weights and summed.

[0031] Trend analysis employs time series analysis to calculate the rate of change of each indicator compared to its 7-day and 30-day averages, identifying significant trends. Indicators showing a continuous decline over several days are marked as key areas of focus.

[0032] Risk level determination is based on weighted health scores and the number of abnormal indicators. Decision tree rules are used to classify pet health status into four levels: normal, attention level, warning level, and veterinary level.

[0033] S105, based on health risk levels, combines dynamic weights and trend analysis, and uses natural language generation technology to generate a health report that includes data summaries, anomaly interpretations, trend analysis, and personalized suggestions. It also embeds the weight allocation and adjustment reasons from the decision-making process into the health report.

[0034] In this embodiment, report text generation employs a combination of template filling and dynamic generation. First, a library of various text templates is constructed, including data summary templates, anomaly interpretation templates, trend analysis templates, and suggestion templates. Then, based on the health assessment results and dynamic weights, an appropriate template is selected for filling. For content requiring personalized generation, a pre-trained language model is invoked for dynamic generation, improving the text's naturalness and readability. The generated text includes data summaries, anomaly interpretations, trend analyses, and personalized suggestions. Weight transparency is a key feature of this invention; the report text embeds information such as the weight values ​​of each indicator, the reasons for weight adjustments, and their contribution to the conclusions, enabling users to understand the decision-making process.

[0035] S106 outputs the health report to the user in a structured format and optimizes the weight strategy online using a reinforcement learning algorithm based on user interaction feedback and subsequent diagnostic results.

[0036] In this embodiment, to provide users with intuitive and easy-to-understand monitoring results, the algorithm outputs the pet's health status in the form of a structured report. The report integrates data charts, tables, textual interpretations, and suggestion lists, and is generated in PDF / HTML / JSON formats. It supports automatic generation on a daily, weekly, or monthly basis, and also supports triggering instant reports when an anomaly is detected. The generated report is pushed to the user's mobile app, WeChat mini-program, or email via API, allowing users to view their pet's health status and historical trends at any time.

[0037] Based on user interaction feedback, a reinforcement learning algorithm is used to optimize the weight strategy online. The system records user feedback on reports (adopting suggestions / ignoring suggestions / manual correction) and subsequent veterinary diagnostic results to construct reward signals; the Q-learning algorithm is used to update the weight strategy model, with positive feedback reinforcing the current weight strategy and negative feedback triggering weight adjustments; through continuous iteration, the system's weight allocation strategy continuously approaches the optimal value.

[0038] Specifically, a Q-learning reinforcement learning algorithm is adopted to update the weight strategy online based on user behavior feedback. In the implementation, the state space consists of current environmental characteristics (temperature, humidity, noise, etc.), behavioral characteristics (activity state, trends, etc.), deviations in physiological indicators, and historical user feedback tendencies, forming approximately 100,000 possible combinations through discretization encoding. The action space is defined as the adjustment operations on the weights of various health indicators, including three directions: increase, decrease, and maintain, with three amplitudes: small, medium, and large, totaling approximately 60 selectable actions. The reward function is designed by integrating three types of signals: instant feedback reward (user adoption +1.0, ignore -0.3, manual correction). The system employs a weighted approach, which balances exploration and utilization (initial ε=0.3, gradually decreasing to 0.05). It provides rewards for delayed diagnosis (+2.0 for confirmed cases, -2.0 for missed cases, -1.0 for false positives) and soft metrics (read completion rate, next-day retention, etc.). Each time user feedback is received, the Q-table is updated according to the formula Q(s,a)←Q(s,a)+α[R+γmaxQ(s′,a′)-Q(s,a)], causing the system to gradually favor weight adjustments based on historically positive feedback in similar situations. For example, when the system repeatedly detects that reports of "increased heart rate after exercise" are ignored by users, the Q-value decreases. Subsequently, in similar situations, the system automatically reduces the heart rate weight and focuses on other metrics, thus achieving continuous personalized optimization of the weighting strategy.

[0039] In some embodiments, step S101 above, which involves collecting the pet's physiological data, behavioral data, environmental data, and historical data, and performing data cleaning, filtering, and feature extraction to generate multi-dimensional quantitative features, specifically includes: Physiological signals such as body temperature, heart rate and respiratory rate of pets were collected. After noise reduction using a sliding window filtering algorithm, the mean, variance and fluctuation coefficient were extracted as physiological features. Collect behavioral data on pets' activity levels, vocalization frequency, food intake, and water intake. After removing outliers caused by equipment malfunctions through anomaly detection, calculate the month-on-month change rate, year-on-year change rate, and trend slope as behavioral trend characteristics. Data on room temperature, humidity, light intensity, and noise levels in the environment are collected, and instantaneous fluctuations are smoothed using a moving average algorithm, while continuous values ​​are converted into discrete state codes as environmental features. A historical archive database containing information on breed, age, gender, medical history, and medication records was constructed, and an individual's health baseline was dynamically calculated based on historical monitoring data as an archive feature.

[0040] In this embodiment, wearable sensors or non-contact monitoring devices are used to collect three core physiological signals of the pet in real time: body temperature, heart rate, and respiratory rate. During the acquisition process, a sliding window filtering algorithm is used for dynamic noise reduction to address device noise, motion artifacts, and environmental interference present in the original signals. Specifically, the signal is segmented into fixed time windows (e.g., 5 seconds), and the mean of the data within the window is calculated as the filtered output. The sliding window gradually covers the entire signal sequence to achieve noise suppression. Subsequently, feature extraction is performed on the filtered physiological signals: the mean of each signal within the monitoring period is calculated to reflect the overall level, the variance is calculated to quantify the fluctuation amplitude, and the fluctuation coefficient is calculated by summing the absolute values ​​of the differences between adjacent sampling points. Finally, a physiological feature vector containing the mean, variance, and fluctuation coefficient is generated.

[0041] Behavioral data collection is achieved through an IoT device integrating an accelerometer, microphone, smart feeder, and water dispenser. Specifically, this includes recording pet activity levels using a triaxial accelerometer, collecting vocalization frequencies via microphone, and obtaining food and water intake data from the smart feeder and water dispenser. For outliers caused by equipment malfunctions or environmental interference, a 3σ-based anomaly detection algorithm is used for data cleaning: the mean and standard deviation of each behavioral indicator over the historical monitoring period are calculated, and data points exceeding the mean ± 3 times the standard deviation are marked as outliers and removed. Subsequently, three types of trend characteristics are calculated for the cleaned behavioral data: month-on-month change rate reflects the relative change of the current data compared to the previous period (e.g., the previous day); year-on-year change rate reflects the relative change of the current data compared to the same period last year (e.g., seasonal behavioral patterns); and trend slope is quantified by fitting the slope of the data sequence using linear regression.

[0042] Environmental data was collected using temperature and humidity sensors, light sensors, and noise sensors deployed in the pet activity area. Specifically, the data included four categories of indicators: room temperature, humidity, light intensity, and noise level. To address the instantaneous fluctuations in environmental data, a moving average algorithm was used for smoothing: the average of continuously collected environmental data was calculated over a fixed time window, and the mean of the data within that window was output as the smoothed environmental value. Subsequently, the continuous numerical environmental data was converted into discrete state codes to reduce data dimensionality and improve model robustness. The final generated environmental feature vector is a combination of discrete state codes for each indicator; for example, "room temperature - suitable, humidity - comfortable, light intensity - bright, noise - quiet" will serve as the input for the environmental dimensions.

[0043] The construction of the pet health record features is based on the long-term accumulation of historical health data, specifically including five basic categories of information: breed, age, sex, medical history, and medication records, as well as an individual health baseline dynamically calculated through historical monitoring data. First, static information such as breed, age, and sex, along with medical history information such as past illnesses, surgical history, and allergies, are obtained from the pet hospital's electronic medical record system or through manual user input, while also recording current medication and its duration. Then, the individual health baseline is calculated based on historical physiological monitoring data: historical data for each physiological indicator are clustered over time to identify the data distribution interval under normal conditions, and the median of the distribution interval is extracted as the individual baseline value for that indicator. The final generated record feature vector contains both static information encoding and dynamic baseline values, providing an individualized reference for health assessment.

[0044] In some embodiments, step S102 above, which involves constructing a weight factor library based on multi-dimensional quantitative features, including environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors, specifically includes: Based on multi-dimensional quantitative features, environmental factors are constructed by extracting parameters such as room temperature, light intensity, noise level, and time period. These environmental factors are used to reflect the adjustment needs of the external environment on the weight of the indicators. Based on multi-dimensional quantitative features, individual factors are constructed by extracting parameters such as variety, age, gender, medical history, and health baseline. These individual factors are used to reflect the benchmark impact of individual differences on the weight of indicators. Based on multi-dimensional quantitative features, behavioral factors are constructed by extracting parameters of current activity status and recent trend changes. These behavioral factors are used to capture the basis for real-time correction of indicator weights by behavioral patterns. Based on multi-dimensional quantitative features, data quality factors are constructed by extracting parameters such as sensor signal strength, occlusion degree, acquisition duration, and data integrity. These data quality factors are used to evaluate the degree to which data credibility inhibits the weight of indicators. Based on multi-dimensional quantitative features, historical response patterns and seasonality parameters of individuals are extracted to construct historical pattern factors. These historical pattern factors are used to identify the need for enhanced indicator weights due to historical deviations. Environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors are integrated to construct a weighted factor library.

[0045] In this embodiment, the construction of environmental factors aims to quantify the impact of the external environment on the weighting of pet health indicators. First, three continuously monitored parameters—room temperature, light intensity, and noise level—are extracted from environmental data. These are then combined with time-of-day parameters (e.g., day / night, weekday / restday) for dynamic adjustment. For example, room temperature data is acquired using temperature and humidity sensors deployed in the pet's activity area. If the current room temperature exceeds the pet's comfort range (e.g., the suitable temperature for dogs is 18-28℃), the weighting coefficient of environmental factors on physiological indicators (e.g., heart rate) will dynamically increase. The light intensity parameter is set with a threshold based on the pet's visual sensitivity; when the light intensity is below the threshold, the weight of behavioral activity indicators (e.g., activity level) will be suppressed. The noise level parameter is converted into a weighting coefficient through a hierarchical mapping; for example, when the noise exceeds 60dB, the weight of stress response-related indicators (e.g., respiratory rate) increases by 20%. The time-of-day parameter is determined through historical behavioral pattern analysis. If the pet's activity frequency is significantly higher at night than during the day, the overall weight of nighttime data is reduced by 15% to reduce nighttime misjudgments. The final environmental factors are generated by weighted summation of the parameters. For example, in a high-noise environment at noon in summer, a cat's overall environmental factors show an increase of 12% in the weight of physiological indicators and an inhibition of 8% in the weight of behavioral indicators.

[0046] Individual factors are constructed based on pet profile data and dynamic health baselines to reflect the impact of individual differences on the weighting benchmark. First, three static characteristics—breed, age, and sex—are extracted from the profile data. For example, Golden Retrievers, due to their larger size, have a 10% higher weight for joint load indicators compared to Poodles. Age parameters are mapped using piecewise functions; for instance, metabolic indicators for 1-3 year old puppies have a 15% increased weight, while immune indicators for pets over 7 years old have a 20% increased weight. Sex characteristics are adjusted differentially based on specific disease risks; for example, mammary gland disease-related indicators for unspayed female cats have a 25% increased weight. Medical history parameters are processed using association analysis techniques; if a pet has a history of arthritis, the weight of exercise-related indicators is reduced by 18%. Health baseline parameters are dynamically calculated based on historical physiological data. For example, if a dog's long-term heart rate baseline is 90 beats / minute, and the current monitoring value deviates from the baseline by 10%, the corresponding indicator weight is dynamically adjusted by 5%. Finally, individual factors are generated by weighted fusion of static features and dynamic baselines. For example, a 5-year-old unspayed male Golden Retriever with a history of arthritis will have a 13% reduced weight for joint load and a 22% increased weight for immune indicators.

[0047] Behavioral factors capture dynamic correction criteria through real-time behavioral data and historical trend analysis. Current activity status parameters are extracted from behavioral data, including activity intensity (e.g., steps), eating speed, and interaction frequency. For example, when activity intensity suddenly decreases by 30%, the weight of related physiological indicators increases by 12%; when eating speed abnormally increases by 25%, the weight of digestive indicators increases by 10%. Recent trend change parameters are calculated through time series analysis. If a behavioral indicator (e.g., nighttime activity) shows an upward trend for three consecutive days with a slope exceeding a threshold, the weight of the corresponding health indicator (e.g., fatigue) decreases by 8. Behavioral factors also incorporate behavioral pattern mutation detection. For example, if a pet suddenly changes from an active state to a still state for more than 2 hours, the weight of the associated indicator is dynamically adjusted by 15%. The final behavioral factor is generated by weighting the current state correction value and the trend change coefficient. For example, if a cat's nighttime activity decreases but its daytime activity has been continuously increasing recently, its behavioral factor will show a 5% decrease in the weight of metabolic indicators and a 10% increase in the weight of sleep quality.

[0048] The data quality factor is used to assess the degree to which the reliability of sensor data suppresses the weighting. First, the sensor signal strength parameter is extracted. When the signal strength is below a threshold (e.g., accelerometer signal <0.2g), the corresponding behavioral indicator weight is reduced by 20%. The occlusion parameter is determined by analyzing data volatility; if the standard deviation of data in a certain period exceeds 30% of the historical mean, the weight of data collected in that period is suppressed by 15%. The collection duration parameter is set according to data integrity requirements; indicators with a single collection duration of less than 30 seconds have their weight reduced by 10%. The data integrity parameter is determined by calculating the proportion of missing values; if the missing rate of a physiological indicator exceeds 5%, its weight is suppressed by 25%. The final data quality factor is generated by the weighted sum of the suppression coefficients of each parameter. For example, if a dog's step count data has 12% missing data due to signal occlusion, its data quality factor will show a suppression of the behavioral indicator weight by 18.

[0049] The historical pattern factor identifies the need for weight enhancement by mining individual historical response patterns. Individual historical response pattern parameters extract historical sensitivity indicators of pets to environmental changes and treatment interventions. For example, if a dog's heart rate increases by an average of 15% during high summer temperatures, the weight of this indicator increases by 12 during summer. Seasonal pattern parameters are based on time series analysis; if a behavioral indicator shows cyclical changes in winter and is associated with health deterioration, the weight of this indicator increases by 10 during winter. The historical pattern factor also introduces a deviation warning mechanism: when an indicator deviates from the historical baseline by more than 20% for three consecutive monitoring values, the weight of that indicator dynamically increases by 8 in subsequent assessments. The final historical pattern factor is generated by weighting historical response intensity and seasonal patterns. For example, if a cat has experienced decreased appetite in the past three summers and has been diagnosed with heat stress, its historical pattern factor shows a 20-fold increase in the weight of the summer appetite indicator.

[0050] The five factors mentioned above are dynamically fused using a multilayer perceptron. The input layer receives the original parameters of each factor, hidden layer 1 extracts nonlinear features using the ReLU activation function, hidden layer 2 introduces a Dropout mechanism to prevent overfitting, and the output layer uses the Softmax function to normalize and generate the final weight adjustment coefficients. For example, if the weight of room temperature in the environmental factor is 0.3 and the weight of age in the individual factor is 0.25, then the final weight of this physiological indicator is 0.3 × environmental adjustment coefficient + 0.25 × individual baseline coefficient. During the training phase, a reinforcement learning algorithm is used to dynamically adjust the network parameters based on user feedback and diagnostic results. For example, when a user repeatedly questions the unreasonable weight allocation of a certain indicator, the influence weight of that factor in subsequent calculations is reduced. The final weight factor library enables dynamic and personalized allocation of weights for each health indicator.

[0051] In some embodiments, step S103 above, which involves using a hierarchical multi-factor weighted model to calculate the adjustment coefficients of each factor in the weight factor library in parallel, synthesizing the original weights through element-wise multiplication, and outputting the dynamic weights of each health indicator after normalization, specifically includes: The built-in mapping table is queried based on the pet's breed, age, and medical history to obtain the basic weight vector of each health indicator; Calculate the adjustment coefficient vector of each factor in the weight factor library for each indicator, and multiply the basic weight vector of each health indicator with the adjustment coefficient vector of each factor for each indicator element by element to synthesize the original weight vector. The original weight vector is normalized so that the sum of the weights of all health indicators is one, and the dynamic weights are output.

[0052] In this embodiment, the basic weight vectors of each health indicator are obtained by querying a built-in mapping table based on the pet's breed, age, and medical history. This built-in mapping table is pre-constructed through extensive experiments and data analysis, and it records in detail the basic weights of various health indicators for pets of different breeds, ages, and medical histories. For example, for a certain breed of canine in its puppyhood, the basic weights of heart-related health indicators may be relatively low, while the basic weights of bone development-related health indicators may be relatively high. If the pet has a history of heart disease, the basic weights of heart-related health indicators will be further adjusted to better reflect its actual health condition. By querying this built-in mapping table, the basic weight vectors of each health indicator of the current pet can be quickly and accurately obtained, providing foundational data for subsequent calculations.

[0053] The adjustment coefficient vector for each indicator is calculated for each factor in the weighted factor library. The weighted factor library includes environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors. For environmental factors, factors such as temperature, humidity, and air quality in the pet's environment are considered, and the degree of influence of these factors on each health indicator is analyzed to determine the adjustment coefficient for each indicator. For example, in a high-temperature environment, the burden on a pet's heart may increase, so the adjustment coefficient of the environmental factor for heart-related health indicators will increase accordingly. Individual factors mainly consider individual characteristics such as breed, size, and sex of the pet. Different individual characteristics have different effects on health indicators, thus determining the adjustment coefficient of the individual factor for each indicator. For example, large dogs experience relatively greater joint pressure, so the adjustment coefficient of the individual factor for joint-related health indicators may be higher than that for small dogs. Behavioral factors analyze the pet's daily behaviors, such as exercise volume, sleep duration, and diet, to determine the impact of these behaviors on health indicators and determine the adjustment coefficient of the behavioral factor for each indicator. For example, the adjustment coefficient for muscle-related health indicators may decrease for pets with insufficient exercise. The data quality factor assesses the accuracy and completeness of collected pet physiological and behavioral data, determining the adjustment coefficient for each indicator based on data quality. Poor data quality may reduce the weight of the corresponding health indicator. The historical pattern factor refers to the pet's past health data changes, analyzing their impact on current health indicators and determining the adjustment coefficient for each indicator. By calculating the adjustment coefficient vector for each indicator using these factors separately, the influence of various factors on health indicators can be comprehensively considered.

[0054] Then, the basic weight vector of each health indicator is multiplied element-wise with the adjustment coefficient vector of each factor for each indicator to synthesize the original weight vector. In other words, for each health indicator, its basic weight is multiplied with the adjustment coefficient of each factor to obtain the original weight of the indicator after considering the influence of each factor.

[0055] The original weight vector is normalized so that the sum of the weights of all health indicators is one, resulting in dynamic weights. Normalization adjusts the weights of each health indicator to a reasonable range, facilitating subsequent health assessments and report generation. Specifically, each element in the original weight vector is divided by the sum of all elements, ensuring that the sum of the weights of all health indicators equals 1. The dynamic weights output after normalization more accurately reflect the relative importance of each health indicator in the current situation, providing a reliable basis for generating pet health reports.

[0056] Furthermore, the step of calculating the adjustment coefficient vector of each factor in the weight factor library for each indicator specifically includes: Based on environmental factors, an environmental adjustment coefficient vector is generated through predefined rule mapping. Based on individual factors, an individual adjustment coefficient vector is generated according to the degree of deviation between real-time indicators and the individual's health baseline; Based on behavioral factors, the isolated forest algorithm is introduced to identify the degree of abnormality between the current behavior and the individual's historical patterns, and to generate a behavioral adjustment coefficient vector. Based on the data quality factor, the isolated forest algorithm is introduced to detect the degree of anomaly in sensor signals and generate a data quality adjustment coefficient vector; Based on historical pattern factors, a historical pattern adjustment coefficient vector is generated by comparing the deviation of the current indicator value from the historical data of the same period.

[0057] In this embodiment, the pet's environment is comprehensively monitored, acquiring various environmental data including temperature, humidity, air quality, and light intensity. Then, based on a rule mapping table derived from extensive experiments and expert experience, the table details the correspondence and degree of influence between different environmental data ranges and various health indicators. For example, regarding the pet's respiratory health indicators, when the ambient air quality is poor (e.g., dust concentration exceeds a certain threshold), the rule mapping table determines that this environmental factor has a negative impact on respiratory health indicators and assigns a corresponding adjustment coefficient value. The worse the air quality, the larger the adjustment coefficient value, indicating a more severe impact on respiratory health indicators. By performing this rule matching and adjustment coefficient determination on all environmental data and various health indicators, an environmental adjustment coefficient vector is ultimately generated. Each element in this vector corresponds to the degree to which a health indicator is affected by environmental factors.

[0058] In the initial stage of the pet health monitoring system, a detailed individual health record will be established for each pet, including normal baseline ranges for various health indicators. These baseline ranges are determined based on information such as the pet's breed, age, sex, and past medical history. During monitoring, real-time data on various physiological indicators of the pet, such as heart rate, blood pressure, and body temperature, will be collected. The real-time collected indicator data will be compared with the baseline ranges in the individual health record to calculate the degree of deviation of each indicator from the baseline. Based on the magnitude of the deviation, a corresponding adjustment coefficient will be assigned according to pre-set rules. The greater the deviation, the larger the adjustment coefficient value, indicating that the indicator is more significantly affected by individual factors. By performing this calculation on all health indicators, an individual adjustment coefficient vector will be generated, reflecting the influence of individual factors on each health indicator.

[0059] We continuously collect various behavioral data about pets in their daily lives, such as activity levels, movement patterns, rest periods, and feeding frequency, and establish a database of individual pets' historical behavioral patterns. During each new monitoring cycle, we collect pet behavior data in real time. We then use the Isolation Forest algorithm to analyze the real-time behavioral data and individual historical behavioral patterns. The Isolation Forest algorithm can quickly and effectively identify outliers in the data, allowing us to determine the degree of difference between the current pet's behavior and its historical normal behavior patterns. For example, if a pet's daily activity level is usually relatively stable, but suddenly decreases significantly within a certain period, the Isolation Forest algorithm will identify this anomaly. Based on the degree of anomaly, we assign corresponding behavioral adjustment coefficients. The higher the degree of anomaly, the larger the adjustment coefficient value, meaning the current behavior may have a greater impact on the pet's health indicators. By processing all behavior-related health indicators in this way, we generate a behavioral adjustment coefficient vector, reflecting the impact of behavioral factors on each health indicator.

[0060] In pet health monitoring systems, various sensors are used to collect physiological, behavioral, and environmental data from pets. However, sensors may be affected by various factors, leading to anomalies in the collected signals, such as excessive signal noise, missing data, and data mutations. To assess data quality, an isolated forest algorithm is introduced to analyze the signals collected by the sensors. This algorithm can detect abnormal components in the signal and determine the quality of the data. Based on the degree of data quality anomaly, a corresponding data quality adjustment coefficient is assigned. The worse the data quality, the smaller the adjustment coefficient value, because low-quality data has a greater impact on the reliability of health indicator assessments, and its influence on the final result should be appropriately reduced. By performing this analysis on the data collected by all sensors, a data quality adjustment coefficient vector is generated, reflecting the impact of data quality factors on various health indicators.

[0061] The pet health monitoring system's database stores a large amount of health indicator data for pets across various historical time periods. In a new monitoring cycle, the system acquires the current values ​​of each pet's health indicators. These current values ​​are compared with historical data from the same period (e.g., the same season, the same month), calculating the degree of deviation for each indicator. For example, if a pet's average heart rate is relatively stable each summer, but its heart rate on a particular day this summer is significantly higher than the historical average, this indicates a deviation from the historical pattern. Based on the magnitude of the deviation, a corresponding historical pattern adjustment coefficient is assigned according to pre-defined rules. The greater the deviation, the larger the adjustment coefficient, indicating a more significant influence of historical pattern factors on the current indicator. By performing this calculation on all health indicators, a historical pattern adjustment coefficient vector is generated, reflecting the impact of historical pattern factors on each health indicator.

[0062] By calculating adjustment coefficient vectors for environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors, the impact of various factors on pet health indicators can be comprehensively and accurately considered, providing an important basis for generating scientific and reasonable pet health reports.

[0063] Furthermore, the generation of an environmental adjustment coefficient vector based on environmental factors through predefined rule mapping specifically includes: The room temperature parameter, light intensity parameter, noise level parameter, and time period parameter contained in the environmental factors are extracted from the weight factor library as environmental input parameters. Based on a pre-set medical prior rule base, the state of each environmental input parameter is determined, mapping the room temperature parameter to high temperature, normal temperature or low temperature, the light intensity parameter to strong light, normal or weak light, the noise level parameter to noisy, normal or quiet, and the time period parameter to day and night label and season label. Based on the status determination results of each environmental input parameter, query the predefined adjustment coefficient mapping table to obtain the preliminary adjustment range of each health indicator that matches the current environmental status; The initial adjustment ranges corresponding to each environmental input parameter are combined and superimposed to generate an environmental adjustment coefficient vector with the same dimension as the basic weight vector. Each element in the environmental adjustment coefficient vector corresponds to the weight adjustment coefficient of a health indicator under the current environmental conditions.

[0064] In this embodiment, various parameters closely related to environmental factors are extracted from a weighted factor library as basic environmental input parameters for subsequent analysis. These parameters include room temperature parameters, which reflect the temperature of the space where the pet lives; different temperatures have different effects on the pet's physiological functions; light intensity parameters, which reflect the lighting conditions of the pet's living environment; excessive or insufficient light may affect the pet's behavior and health; noise level parameters, which measure the noise level of the pet's surrounding environment; excessive noise may cause stress reactions in the pet; and time period parameters, which include day / night and seasonal labels, as the pet's activity patterns and health needs differ at different times of day and in different seasons. By comprehensively extracting these parameters, a complete picture of the pet's environmental conditions can be drawn.

[0065] Based on a pre-defined medical rule base, the extracted environmental input parameters are meticulously assessed for their respective states. For room temperature, based on medical research and pet physiology, it is mapped to high, normal, or low temperatures. For example, if the room temperature exceeds the upper limit of the pet's suitable survival temperature by a certain range, it is considered a high temperature; within the suitable temperature range, it is considered a normal temperature; and below the lower limit of the suitable temperature, it is considered a low temperature. Similarly, light intensity is mapped to strong, normal, or weak light states according to medical knowledge. Light intensity exceeding a certain value indicates strong light, within a suitable range, it is considered normal, and below a certain value, it is considered weak. Noise level is mapped to noisy, normal, or quiet states based on the degree of noise's impact on the pet. For example, noise exceeding a certain decibel threshold indicates a noisy state, within a reasonable range, it is considered normal, and below a certain decibel level, it is considered quiet. Time period parameters are clearly labeled with day / night and seasonal tags, such as daytime, nighttime, spring, and summer. This state assessment allows for a clear understanding of the specific status of each environmental parameter.

[0066] Based on the status determination results of each environmental input parameter, a predefined adjustment coefficient mapping table is queried. This mapping table, derived from extensive experiments and expert experience, records in detail the initial adjustment range for each health indicator under different environmental conditions. For example, when the room temperature is high, the mapping table indicates that the initial adjustment range for the pet's heart rate is an increase of a certain percentage; when the light intensity is strong, the mapping table provides a corresponding initial adjustment range for the pet's eye health indicators. By querying this mapping table, the initial adjustment range for each health indicator matching the current environmental condition can be quickly obtained, providing a basis for subsequent weight adjustments.

[0067] The initial adjustment magnitudes corresponding to each environmental input parameter are comprehensively superimposed. Since the effects of each environmental parameter on different health indicators are interrelated and synergistic, it is necessary to comprehensively consider the initial adjustment magnitudes of all environmental parameters. For example, high room temperature increases the adjustment magnitude of the heart rate indicator, while strong light may also have some impact on the heart rate indicator; these two effects are analyzed and superimposed. In this way, an environmental adjustment coefficient vector with the same dimension as the basic weight vector is generated. Each element in this vector corresponds to the weight adjustment coefficient of a health indicator under the current environmental conditions, accurately reflecting the degree of influence of environmental factors on the weights of each health indicator, providing important support for the subsequent generation of scientifically sound pet health reports.

[0068] Furthermore, the generation of an individual adjustment coefficient vector based on individual factors and the degree of deviation between real-time indicators and the individual's health baseline specifically includes: Extract the individual factors, including the variety parameters, age parameters, gender parameters, medical history parameters, and health baseline parameters, from the weighted factor library; Based on the variety parameters, we identify special physiological structural characteristics; based on the medical history parameters, we identify indicators related to past diseases; based on the age parameters, we identify life cycle stage characteristics; and we determine the health indicators that are strongly correlated with individual characteristics as the key targets for regulation. The values ​​of each health indicator are obtained in real time and compared with the corresponding individual health baseline range in the health baseline parameters. The deviation of each health indicator from the health baseline is calculated. The deviation of each health indicator is matched with the key adjustment objects associated with individual characteristics. The indicators that are successfully matched are assigned enhanced adjustment weights, and the indicators that are not matched are assigned conventional adjustment weights. By combining the deviation of each indicator with the allocation of adjustment weights, an individual adjustment coefficient vector with the same dimension as the basic weight vector is generated. Each element in the individual adjustment coefficient vector corresponds to a weight adjustment coefficient of a health indicator based on individual characteristics and real-time deviation.

[0069] In this embodiment, various parameters closely related to individual factors are extracted from a weighted factor library. These parameters cover breed parameters, as different breeds of pets have unique physiological structures and genetic characteristics; for example, some breeds may be more susceptible to specific diseases. Age parameters are also included, as pets at different life stages exhibit significant differences in physical function and health needs, with different health concerns for puppies and senior pets. Gender parameters are also included, as gender may affect certain physiological indicators and disease susceptibility. Medical history parameters record past illnesses of the pet, which may have a lasting impact on its current health status. Finally, health baseline parameters reflect the normal range of various indicators for the pet in a healthy state and serve as an important reference for assessing the pet's current health status. By comprehensively extracting these parameters, a complete picture of the pet's individual characteristics can be drawn.

[0070] Based on the extracted breed parameters, a thorough analysis of the unique physiological characteristics of each breed is conducted. For example, some brachynose breeds may have specific respiratory problems, so respiratory-related health indicators require close monitoring. Based on medical history parameters, health indicators associated with past illnesses are carefully identified. If a pet has a history of heart disease, indicators related to cardiac function, such as heart rate and blood pressure, should be prioritized for adjustment. Simultaneously, age parameters are used to accurately identify the pet's life cycle stage. Different stages of a pet's development and metabolic rate differ, leading to variations in key health indicators. For instance, growth and development indicators may be more important for puppies, while organ function indicators are more crucial for older pets. Through this analysis, health indicators strongly correlated with individual pet characteristics are identified and prioritized for subsequent weighting adjustments.

[0071] The system acquires real-time health indicator values ​​for pets via intelligent monitoring devices. These values ​​reflect the pet's current health status. The acquired real-time health indicator values ​​are then compared index by index with the corresponding individual health baseline range in the health baseline parameters. This comparison precisely calculates the degree of deviation of each health indicator from the health baseline. The degree of deviation can be measured by the difference between the actual value and the upper or lower limit of the baseline range, providing a clear understanding of the magnitude of the difference between the current health indicator and the normal state.

[0072] The calculated deviations of each health indicator are matched with the previously identified key moderating targets related to individual characteristics. Indicators that match successfully—those that are strongly correlated with the pet's individual characteristics but also exhibit significant deviations—are assigned enhanced moderating weights. This means these indicators will be more significantly affected in subsequent weight adjustments because they may play a more crucial role in the pet's health. Unmatched indicators—those with weak correlations to individual characteristics or small deviations—are assigned regular moderating weights and processed according to general adjustment rules. This differentiated allocation of moderating weights more accurately reflects the importance of different health indicators to the individual pet's health.

[0073] By considering the deviation levels of each indicator and the allocation of adjustment weights, an individual adjustment coefficient vector with the same dimension as the basic weight vector is generated. During the generation process, the magnitude of deviation and the type of adjustment weight assigned to each health indicator are fully considered. In the final generated individual adjustment coefficient vector, each element corresponds to a weight adjustment coefficient for a health indicator based on individual characteristics and real-time deviation. This accurately reflects the impact of the pet's individual characteristics and real-time health status on the weights of each health indicator, providing an important basis for generating scientifically sound pet health reports.

[0074] Furthermore, based on behavioral factors, the isolated forest algorithm is introduced to identify the degree of anomalousness between the current behavior and the individual's historical patterns, generating a behavior adjustment coefficient vector, specifically including: The current activity status parameters, vocalization frequency parameters, food intake parameters, water intake parameters, and recent trend change parameters contained in the weighted factor library are extracted as behavioral features. Collect behavioral data sequences of pets in their historical health states to construct a historical behavioral sample library that reflects the normal behavioral patterns of individuals. Based on a historical behavior sample database, an isolated forest algorithm is used to train a unique behavior anomaly detection model for each pet. The behavior anomaly detection model is used to construct multiple isolated trees by randomly cutting the feature space, gathering normal behavior points that are difficult to isolate near the root of the tree, and segmenting abnormal behavior points that are easy to isolate near the top of the tree. Input the behavioral features of the current moment into the trained Isolation Forest model, calculate the isolation score of the current behavioral pattern in the distribution of the individual's historical behavior, and obtain the behavioral abnormality score. Based on the score of the degree of behavioral abnormality, and combined with the correlation strength between each health indicator and the behavioral state, the health indicators closely related to behavior are assigned an enhanced moderation coefficient, and the health indicators with a weaker correlation to behavior are assigned a basic moderation coefficient. The correlation strength between the comprehensive behavioral abnormality score and each health indicator is assigned to generate a behavioral adjustment coefficient vector with the same dimension as the basic weight vector. Each element in the behavioral adjustment coefficient vector corresponds to a weight adjustment coefficient of a health indicator based on the current behavioral abnormality.

[0075] In this embodiment, various parameters included in the behavioral factors are extracted from the weighted factor library to serve as behavioral characteristics. These parameters include: current activity state parameters, which reflect whether the pet is currently active, resting, or in another specific activity state; vocalization frequency parameters, as different vocalization frequencies may indicate different emotions or physical conditions in the pet; food intake parameters, recording the amount of food the pet is currently eating, which is closely related to its digestive system health; water intake parameters, reflecting the pet's water intake, which has an important impact on the health of organs such as the kidneys; and recent trend change parameters, which reflect the dynamic changes in the pet's behavior over a recent period, such as whether activity levels are gradually increasing or decreasing. By comprehensively extracting these parameters, a complete description of the pet's current behavioral characteristics can be obtained.

[0076] To accurately identify the degree of abnormality in a pet's current behavior, it is necessary to collect behavioral data sequences of the pet in its historical healthy state. These data sequences should cover various behavioral information of the pet over different time periods, including activity level, vocalization frequency, food intake, and water intake. This collected historical behavioral data is organized and stored to construct a historical behavioral sample library specifically reflecting an individual's normal behavioral patterns. Based on this constructed historical behavioral sample library, an isolated forest algorithm is used to train a dedicated behavioral anomaly detection model for each pet. The principle of the isolated forest algorithm is to construct multiple isolated trees by randomly segmenting the feature space. During the construction process, normal behavioral points, due to their universality and regularity, are usually difficult to isolate and tend to cluster near the root of the tree; while abnormal behavioral points, due to their specificity and randomness, are relatively easy to isolate and tend to cluster near the treetop. In this way, the model can learn the feature distribution of the pet's normal behavioral patterns, thus gaining the ability to identify abnormal behavior. Each pet has its unique behavioral habits and patterns, therefore, training a dedicated model for each pet can more accurately detect abnormal behaviors.

[0077] The behavioral features extracted at the current moment are input into the already trained Isolation Forest model. The model calculates the isolation score of the current behavior pattern within the individual's historical behavior distribution based on the learned distribution of normal behavior patterns. This score is the behavioral anomalousness score. A higher score indicates a greater difference between the current behavior and the pet's historical normal behavior patterns, and thus a higher degree of anomalousness; a lower score indicates that the current behavior is closer to the pet's historical normal behavior patterns, and a lower degree of anomalousness. In this way, the anomalousness of the pet's current behavior can be quantitatively assessed.

[0078] Based on the calculated scores for behavioral abnormality, adjustment coefficients are assigned according to the correlation strength between various health indicators and behavioral states. For health indicators closely related to behavior—for example, if a pet's activity level is closely related to heart health, a high score for behavioral abnormality indicates a possible abnormality in the activity level, which may affect heart health—an enhanced adjustment coefficient is assigned to this heart-related health indicator, making it more significantly affected in subsequent weight adjustments. Conversely, for health indicators with weaker correlations to behavior—such as the relatively small correlation between a pet's coat quality and its current behavioral state—a basic adjustment coefficient is assigned, and adjustments are made according to general rules. This differentiated allocation of adjustment coefficients more accurately reflects the degree to which different health indicators are affected by behavioral abnormalities.

[0079] By analyzing the correlation strength between the comprehensive behavioral abnormality score and various health indicators, a behavior adjustment coefficient vector with the same dimension as the base weight vector is generated. During generation, the correlation between each health indicator and behavioral abnormality, as well as the magnitude of the behavioral abnormality score, are fully considered. For example, health indicators closely related to behavior and with high behavioral abnormality scores will have relatively larger behavior adjustment coefficients; conversely, health indicators with weaker correlation to behavior and low behavioral abnormality scores will have relatively smaller behavior adjustment coefficients. In the final generated behavior adjustment coefficient vector, each element corresponds to a weight adjustment coefficient for a health indicator based on the current degree of behavioral abnormality. This accurately reflects the impact of pet behavioral abnormalities on the weights of various health indicators, providing an important basis for generating scientifically sound pet health reports.

[0080] Furthermore, the step of introducing the isolated forest algorithm based on the data quality factor to detect the anomaly degree of sensor signals and generating a data quality adjustment coefficient vector specifically includes: The sensor signal strength parameter, occlusion parameter, acquisition duration parameter, and data integrity parameter included in the weight factor library are extracted as data quality features. Collect historical quality characteristic data of each sensor under normal operating conditions, and construct a data quality sample library that reflects the distribution of normal signal quality. Based on the data quality sample library, a global data quality detection model is trained using the isolated forest algorithm. The global data quality detection model is used to construct multiple isolated trees by randomly cutting the feature space, gathering high-quality data points with stable signals, less occlusion, and complete collection near the tree roots, and segmenting low-quality data points with intermittent signals, severe occlusion, and missing collection near the tree tops. Input the data quality features of each sensor at the current moment into the trained isolated forest model, calculate the isolation score of the current signal quality in the global quality distribution, and obtain the data quality anomaly score. Based on the data quality anomaly scores of each sensor, and combined with the differences in the sensitivity of different health indicators to data quality, an inhibitory adjustment coefficient is assigned to health indicators that rely on low-quality sensors, and a maintenance or enhancement adjustment coefficient is assigned to health indicators that rely on high-quality sensors. By combining the data quality anomaly scores of each sensor with the sensitivity distribution of each health indicator, a data quality adjustment coefficient vector with the same dimension as the basic weight vector is generated. Each element in the data quality adjustment coefficient vector corresponds to a weight adjustment coefficient of a health indicator based on the current data acquisition quality.

[0081] In this embodiment, various parameters included in the data quality factors are extracted from the weighting factor library and used as data quality features. These parameters include: sensor signal strength parameters, which reflect the strength of the signal received by the sensor (weak signal strength may indicate interference or malfunction); occlusion parameters, which measure whether the sensor is obstructed by objects during operation and the degree of obstruction (occlusion can severely affect the accuracy of sensor data); acquisition duration parameters, which record the time spent by the sensor to acquire data (too short an acquisition duration may result in incomplete and valid data); and data integrity parameters, which reflect whether there are any missing data points (incomplete data can affect the comprehensive assessment of the pet's health status). By comprehensively extracting these parameters, the quality characteristics of the pet health data acquisition process can be fully described.

[0082] To accurately detect the degree of anomaly in current sensor signals, it is necessary to collect historical quality characteristic data of each sensor under normal operating conditions. This data should cover various quality characteristic information of sensors operating normally under different time periods and environmental conditions. The collected historical quality characteristic data is then organized and stored to construct a data quality sample library specifically reflecting the distribution of normal signal quality. Based on this constructed data quality sample library, a global data quality detection model is trained using the Isolation Forest algorithm. The core principle of the Isolation Forest algorithm is to construct multiple isolated trees by randomly cutting the feature space. During the construction process, high-quality data points with stable signals, minimal obstruction, and complete acquisition are usually difficult to isolate due to their universality and regularity, and are usually clustered near the root of the tree; while low-quality data points with intermittent signals, severe obstruction, and missing acquisition are relatively easy to isolate due to their particularity and randomness, and are segmented near the treetop. In this way, the model can learn the distribution pattern of normal signal quality characteristic data, thereby possessing the ability to identify abnormal signal quality. The trained global data quality detection model can uniformly detect and evaluate the quality of various sensor signals.

[0083] The data quality features of each sensor at the current moment are input into the already trained global data quality detection model. The model calculates the isolation score of the current signal quality within the global quality distribution based on the learned distribution of normal signal quality features. This score is the data quality anomaly score. A higher score indicates a greater difference between the current sensor signal and the normal signal quality pattern, and thus a higher degree of anomaly; a lower score indicates that the current sensor signal is closer to the normal signal quality pattern, and a lower degree of anomaly. In this way, the quality anomaly of each sensor signal can be quantitatively assessed.

[0084] Based on the data quality anomaly scores of each sensor, adjustment coefficients are assigned according to the varying sensitivity of different health indicators to data quality. Different health indicators have different degrees of dependence on data quality. For example, some health indicators based on real-time sensor monitoring data, such as heart rate and body temperature, are highly sensitive to data quality; poor sensor signal quality will lead to inaccurate assessments of these health indicators. Conversely, some health indicators based on long-term data trend analysis are relatively less sensitive to data quality. Therefore, for health indicators relying on low-quality sensors, due to their lower data reliability, a suppressive adjustment coefficient is assigned to reduce their impact in subsequent weight adjustments. For health indicators relying on high-quality sensors, a maintenance or enhancement adjustment coefficient is assigned to maintain or enhance their impact in subsequent weight adjustments to ensure the accuracy of health assessments.

[0085] By combining the data quality anomaly scores from various sensors with the sensitivity distribution of each health indicator, a data quality adjustment coefficient vector with the same dimension as the base weight vector is generated. During generation, the dependence of each health indicator on the data quality of different sensors and the magnitude of the data quality anomaly scores for each sensor are fully considered. For example, for health indicators that rely on data from multiple sensors with high anomaly rates, their corresponding data quality adjustment coefficients will be relatively small; while for health indicators that rely on a few high-quality sensors, their corresponding data quality adjustment coefficients will be relatively large. In the final generated data quality adjustment coefficient vector, each element corresponds to a weight adjustment coefficient for a health indicator based on the current data collection quality. This accurately reflects the impact of data quality on the weights of each health indicator, providing an important basis for generating scientifically sound pet health reports.

[0086] Furthermore, the process of generating a historical pattern adjustment coefficient vector based on historical pattern factors by comparing the deviation of the current indicator value with historical data from the same period specifically includes: Extract individual historical response pattern parameters and seasonality parameters contained in the historical pattern factors from the weight factor library; Collect health indicator data sequences of pets during the same historical period, group and archive them according to seasonal, monthly and diurnal time labels, and construct a seasonal historical database that reflects the normal historical fluctuation pattern of individuals; The real-time values ​​of each health indicator at the current moment are compared with the distribution range of the corresponding indicators in the same period of the historical seasonal database, and the degree of deviation of each health indicator from the average and fluctuation range of the same period of the historical period is calculated. Based on the degree of deviation of each health indicator, and combined with the historical abnormal recovery trajectory and periodic change pattern recorded in the historical response pattern parameters and seasonal pattern parameters, maintenance adjustment coefficients are assigned to indicators that conform to historical fluctuation patterns, enhancement adjustment coefficients are assigned to indicators that significantly deviate from historical patterns, and higher historical pattern sensitivity is assigned to indicators that follow seasonal patterns. By combining the deviation and sensitivity distribution of various health indicators, a historical pattern adjustment coefficient vector with the same dimension as the basic weight vector is generated. Each element in the historical pattern adjustment coefficient vector corresponds to a weight adjustment coefficient of a health indicator based on the deviation of the current value from the historical period.

[0087] In this embodiment, two key parameters from the historical pattern factors are extracted from the weighting factor library: individual historical response pattern parameters and seasonality parameters. Individual historical response pattern parameters record how pets reacted and recovered from past health problems. For example, after a pet contracted a disease, the time it took for its various physiological indicators to return to normal levels, and the trends in these indicators during recovery. Seasonality parameters reflect the changes in pet health indicators with seasons, months, and time of day. For instance, in summer, a pet's heart rate may be relatively faster due to higher temperatures; in winter, due to reduced activity, its weight may increase. By extracting these parameters, a comprehensive understanding of the historical variation characteristics of pet health indicators can be obtained.

[0088] To accurately compare current health indicators with historical data from the same period, it is necessary to collect a series of health indicator data for pets during the same historical time period. This data should cover various health indicators such as heart rate, body temperature, blood pressure, and weight for pets in different seasons, months, and day / night cycles. The collected data should be meticulously grouped and archived according to seasonal, monthly, and day / night cycle labels. In this way, a seasonal historical database that comprehensively reflects the normal historical fluctuation patterns of an individual pet will be constructed. The real-time values ​​of each health indicator at the current moment will be compared indicator-by-indicator with the corresponding historical distribution range in the seasonal historical database. Specifically, for each health indicator, its historical mean and fluctuation range will be identified. Then, the difference between the current indicator value and the historical mean, as well as the relationship between the current indicator value and the historical fluctuation range, will be calculated to determine the degree of deviation of each health indicator relative to the historical mean and fluctuation range. For example, if the current heart rate of a pet is significantly higher than the historical average heart rate and exceeds the normal fluctuation range of the historical heart rate, then the pet's current heart rate indicator can be considered to have a large deviation.

[0089] Based on the degree of deviation of each health indicator, and combined with the historical response pattern parameters and seasonality parameters recording past abnormal recovery trajectories and periodic changes, adjustment coefficients are assigned. For indicators that conform to historical fluctuation patterns, it indicates that the pet's current physical condition is consistent with similar past performance, and the changes in its health indicators may be normal physiological fluctuations. Therefore, a maintenance adjustment coefficient is assigned to maintain its relative stability in subsequent weight adjustments. For indicators that significantly deviate from historical patterns, this may mean that the pet has developed new health problems or that existing health problems have worsened, requiring more attention. Therefore, an enhancement adjustment coefficient is assigned to increase its influence in subsequent weight adjustments. For indicators with obvious seasonal patterns, such as the faster heart rate in summer mentioned earlier, a higher historical pattern sensitivity is assigned to ensure that the influence of seasonal factors is more fully reflected in the weight adjustments.

[0090] By comprehensively considering the deviation and sensitivity distribution of various health indicators, a historical pattern adjustment coefficient vector with the same dimension as the basic weight vector is generated. During the generation process, the magnitude of the deviation of each health indicator and its corresponding historical pattern sensitivity are fully considered. For example, health indicators with larger deviations and higher historical pattern sensitivity will have relatively larger historical pattern adjustment coefficients; conversely, health indicators with smaller deviations and lower historical pattern sensitivity will have relatively smaller historical pattern adjustment coefficients. In the final generated historical pattern adjustment coefficient vector, each element corresponds to a weight adjustment coefficient for a health indicator based on its current value and the deviation from the historical period. This accurately reflects the impact of historical patterns on the weights of each health indicator, providing an important basis for generating scientifically sound pet health reports.

[0091] In some embodiments, in step S104 above, the step of weighted summing of the standardized scores of each health indicator based on dynamic weights, identifying abnormal changes in the indicators using a trend analysis algorithm, and determining the pet's health risk level based on the weighted score and the number of abnormal indicators specifically includes: Based on dynamic weights, the standardized scores of each health indicator are weighted and summed to calculate the comprehensive health score. Trend analysis was performed on the time series data of various health indicators to calculate the month-on-month and year-on-year change rates of each health indicator compared with the historical average and recent average, and to identify significant trends of continuous increase, continuous decrease or abnormal fluctuation. Based on the overall health score and the number of abnormal indicators marked as showing significant trends, the pet's health status is classified into risk levels using preset risk level determination rules.

[0092] In this embodiment, the standardized scores of each health indicator are weighted and summed based on their dynamic weights. The dynamic weights are calculated using a hierarchical multi-factor weighted model, drawing on multi-dimensional information including the pet's physiological, behavioral, environmental, and historical data. This accurately reflects the influence of different factors on each health indicator. The standardized scores are values ​​obtained after standardizing each health indicator, eliminating differences in units and numerical ranges between different indicators. During the weighted summation, the standardized score of each health indicator is multiplied by its corresponding dynamic weight, and then the weighted scores of all health indicators are summed to calculate the overall health score. This overall health score is a comprehensive value that considers multiple factors and health indicators, providing a more complete picture of the pet's current overall health status. For example, if the pet's heart rate indicator has a high dynamic weight and a high standardized score, then the heart rate indicator will have a significant impact on the overall health score during the weighted summation.

[0093] Trend analysis was performed on the time series data of various health indicators. The time series data recorded the changes in the values ​​of various health indicators of pets at different points in time. By analyzing this data, long-term trends and short-term fluctuations of health indicators can be identified. Specifically, the month-on-month and year-on-year change rates of each health indicator compared to historical averages and recent mean values ​​were calculated. The month-on-month change rate reflects the changes in health indicators within adjacent time periods, enabling timely detection of short-term abnormal fluctuations; the year-on-year change rate reflects the changes in health indicators compared to the same period last year, helping to identify seasonal or long-term trends. By calculating these change rates, it is possible to identify whether there are significant trends of continuous increase, continuous decrease, or abnormal fluctuations in each health indicator. For example, if the month-on-month change rate of a pet's weight is consistently positive over a certain period, and the year-on-year change rate is also significantly higher than the same period in previous years, then it can be considered that the pet's weight has a significant trend of continuous increase, which may indicate a health problem.

[0094] Based on the calculated overall health score and the number of abnormal indicators marked with significant trends, a pre-defined risk level classification rule is used to categorize the pet's health status into risk levels. This rule, derived from extensive data analysis and expert experience, comprehensively considers both the overall health score and the number of abnormal indicators. Generally, a lower overall health score indicates a poorer overall health condition; a higher number of abnormal indicators suggests a greater likelihood of health problems. For example, a low overall health score and a large number of abnormal indicators classify the pet as high-risk, prompting the owner to take immediate action, such as taking the pet to the vet for a checkup. A high overall health score and a small number of abnormal indicators classify the pet as low-risk, allowing the owner to continue monitoring the pet's health. A score in between is classified as medium-risk, reminding the owner to pay attention to changes in the pet's health and adjust feeding methods accordingly. This risk level classification provides pet owners with clear and intuitive health information, helping them understand their pet's health status and make appropriate decisions.

[0095] In some embodiments, in step S105 above, the step of generating a health report based on health risk level, combined with dynamic weighting and trend analysis, using natural language generation technology, including data summary, anomaly interpretation, trend analysis, and personalized suggestions, and embedding the weighting allocation and adjustment reasons from the decision-making process into the health report, specifically includes: Based on health risk levels, combined with dynamic weights and trend analysis, data summaries are generated for each core health indicator, and the indicator values ​​are converted into textual descriptions. Based on the degree of influence of each factor in the dynamic weight on health indicators, attribution explanations are generated for abnormal indicators to produce abnormal interpretation texts. Based on the significant changing trends identified by trend analysis, and combined with historical pattern factors and behavioral factors, trend analysis is interpreted to generate the potential health implications behind the changes in indicators. Based on the health risk level, the guidance of dynamic weights, and the characteristics of abnormal indicators, personalized action suggestions are generated and marked with priorities and basis. A health report is constructed based on textual descriptions, interpretations of anomalies, the potential health implications behind changes in indicators, and personalized action recommendations. Dynamic weights, the contribution of each health indicator to the final conclusion, and the reasons for weight adjustments are embedded in the health report.

[0096] In this embodiment, based on the determined health risk level and the results of dynamic weighting and trend analysis, data summaries are generated for each core health indicator. During this process, the system references pre-defined rules and language templates to transform the indicator values ​​into easily understandable textual descriptions. Furthermore, considering that dynamic weights reflect the degree of influence of different factors on health indicators, the descriptions highlight the impact of factors with higher weights on the indicator values.

[0097] Based on the influence of each factor in the dynamic weights on health indicators, the system generates anomaly interpretation texts for abnormal indicators. The system deeply analyzes the correlation between each factor in the dynamic weights library and the abnormal indicators to identify the main factors causing the anomalies. For example, when a pet's activity level is abnormally low, the system examines the impact of individual factors, behavioral factors, and environmental factors in the dynamic weights on the activity level. If it finds that the pet's age is high among individual factors, the pet's recent activity time has significantly decreased among behavioral factors, and the weather is consistently rainy among environmental factors, the system will synthesize these factors to generate anomaly interpretation text, such as "The pet's abnormally low activity level may be due to the pet's advanced age, declining physical function, reduced recent activity time, and rainy weather that is unfavorable for outdoor activities, among other factors." Through this attribution explanation, pet owners can clearly understand the reasons for the abnormal indicators.

[0098] Based on significant trends identified through trend analysis, the system interprets these trends using historical pattern factors and behavioral factors to generate the potential health implications behind the changes in indicators. The system provides in-depth analysis of significant trends such as continuous increases, continuous decreases, or abnormal fluctuations. Taking a pet's weight as an example, if the weight shows a continuous upward trend, the system will combine historical pattern factors to examine the pet's past weight changes, as well as behavioral factors such as the pet's diet and exercise. If it finds that the pet's recent food intake has increased while its exercise has decreased, the system will generate an interpretation text, such as, "The pet's weight is continuously increasing. Combined with its recent pattern of increased food intake and decreased exercise, as well as its past weight gain patterns, there may be a risk of obesity. This could put a burden on the pet's joints, heart, and other organs, affecting its health." Through this interpretation, pet owners can understand the potential health problems represented by the changes in the indicators.

[0099] Based on health risk levels, dynamic weighting, and the characteristics of abnormal indicators, personalized action recommendations are generated, with priorities and justifications clearly indicated. The system will formulate corresponding action recommendations based on different health risk levels. For high-risk levels, the action recommendations will be more urgent and comprehensive, such as "Take your pet to the vet immediately for a comprehensive check-up, focusing on the function of organs such as the heart and liver." For medium-risk levels, the recommendations will be relatively milder, such as "Adjust your pet's diet, increase exercise, and closely monitor your pet's health." For low-risk levels, the recommendations will mainly focus on daily health care, such as "Maintain your pet's current healthy lifestyle and have regular check-ups." Simultaneously, the action recommendations will be personalized based on the guidance of dynamic weighting and the characteristics of abnormal indicators. For example, if the abnormal indicators are mainly related to diet, the action recommendations will focus on dietary adjustments; if environmental factors have a significant impact on health indicators in the dynamic weighting, the action recommendations will include measures to improve environmental conditions. Furthermore, each action recommendation will be marked with its priority and justification, such as "Recommendation 1: Take your pet to the vet immediately (Priority: High; Justification: Pet's body temperature remains high, indicating a risk of infection)," allowing pet owners to clearly understand which measures to take first and why.

[0100] A health report is constructed based on the generated data summary, anomaly interpretation text, potential health implications behind indicator changes, and personalized action recommendations. The system integrates these elements in a logical order to form a complete and clear report. Simultaneously, dynamic weights, the contribution of each health indicator to the final conclusion, and the reasons for weight adjustments are embedded in the health report. By incorporating this information, pet owners can understand how the system makes decisions, enhancing their trust and comprehension of the report. Finally, the constructed health report is output to the user in a structured format for easy viewing and use by pet owners.

[0101] In some embodiments, step S106 above, which involves optimizing the weight strategy online using a reinforcement learning algorithm based on user interaction feedback and subsequent diagnostic results, specifically includes: The system collects user interaction feedback behavior on health reports and gathers subsequent veterinary diagnostic results, which together constitute a set of feedback signals. The interaction feedback behavior includes adopting suggestions, ignoring prompts, and manual correction operations. The current environmental characteristics, behavioral characteristics, physiological index deviation, and user historical feedback tendency parameters are used as state descriptions to construct a state space. Adjusting the weights of various health indicators is made an optional action, thus constructing an action space; Convert user adoption and correct diagnosis into positive reward signals, and convert user ignore, manual correction, and false positives / missed reports into negative reward signals. Design a reward function based on user feedback behavior and diagnosis results. Based on the state space and action space, a Q-learning algorithm is used to update the weights online, and the state-action value table is updated according to the reward signal output by the reward function.

[0102] In this embodiment, user interaction feedback on health reports is collected, along with subsequent veterinary diagnostic results, forming a combined set of feedback signals. User interaction feedback encompasses various types, such as adopting suggestions (users taking action based on personalized recommendations in the health report, like adjusting the pet's diet or increasing exercise); ignoring prompts (users failing to pay attention to or respond to prompts in the health report); and manual correction (users manually modifying content in the health report, such as health risk levels and personalized recommendations). The subsequent veterinary diagnostic results represent an authoritative assessment of the pet's health by a professional veterinarian, including information on whether the pet is ill, the type of illness, and its severity. By integrating this feedback information, the system can comprehensively understand the user's level of acceptance of the health report and the pet's actual health condition, providing a basis for subsequent weighting strategy optimization.

[0103] The state space is constructed by using various information from the current moment as a state description. Specifically, it collects environmental characteristics, such as current indoor temperature, humidity, and air quality, as environmental factors affect the pet's physiological state; behavioral characteristics, including the pet's activity level, activity time distribution, and sleep duration, which reflect the pet's daily state; physiological indicator deviation, i.e., the degree to which the pet's current physiological indicators (such as body temperature, heart rate, and blood pressure) deviate from the normal range, with larger deviations indicating potential health problems; and user historical feedback tendency parameters, which summarize the patterns of user behavior in past health reports, such as whether users tend to adopt suggestions, ignore prompts, or manually correct them. Integrating this information as a state description can comprehensively depict the current pet health monitoring context and the user's interaction with the system, providing rich state information for weighting strategy optimization.

[0104] The system constructs an action space by allowing users to adjust the weights of various health indicators as optional actions. In practical applications, different health indicators have varying degrees of impact on pet health and change with different factors. For example, in high-temperature environments, the impact of ambient temperature on pet health may increase, requiring an appropriate increase in the weights of related health indicators. The system sets up a series of operations to adjust the weights of each health indicator, such as increasing the weight of a health indicator, decreasing the weight of a health indicator, or keeping the weight of a health indicator unchanged. These operations constitute the action space, from which the system can select appropriate actions to adjust the weighting strategy based on the current state.

[0105] A reward function is designed based on user feedback and diagnostic results, transforming different feedback scenarios into corresponding reward signals. The specific rules are as follows: When a user adopts the suggestions in the health report, and subsequent veterinary diagnoses show that the pet's health has improved or is consistent with the report's assessment, the system converts this into a positive reward signal, awarding a certain reward value to encourage the system to continue generating health reports that meet user needs and the pet's actual condition. When a user ignores the prompts in the health report, and subsequent diagnoses show that the pet's health has not deteriorated or is not significantly different from the report's assessment, the system neither rewards nor penalizes the user. When a user manually corrects the health report, or when the system makes false reports (misclassifying a healthy pet as sick) or misses reports (misclassifying a sick pet as healthy), the system converts these situations into negative reward signals, awarding corresponding penalty values ​​to prompt the system to reflect on and adjust its weighting strategy to prevent similar errors from recurring. Through this reward function design, the system can evaluate the effectiveness of its weighting strategy based on user feedback and diagnostic results.

[0106] Based on the constructed state and action spaces, the system employs the Q-learning algorithm to update the weight strategy online. Q-learning is a commonly used reinforcement learning algorithm that learns the optimal strategy by continuously updating the state-action value table. At each time step, the system selects an action from the action space based on the current state and then updates the state-action value table according to the reward signal output by the reward function. Specifically, the system records the reward value obtained after performing an action in the current state, as well as the new state transitioned to after performing the action. Then, based on this information, it updates the value of the corresponding state-action pair in the state-action value table according to the rules of the Q-learning algorithm. Over time and through continuous learning, the state-action value table gradually converges, allowing the system to select the action with the highest value based on the current state—the optimal weight adjustment operation. This enables online optimization of the weight strategy, making the health reports generated by the system more accurate and meeting user needs.

[0107] In some embodiments, the method further includes, in steps S101 to S106 above: Collect multi-source monitoring data and corresponding health results of pets within a preset historical period to construct a historical training sample set; Using known health outcomes as the optimization objective, the optimal weight vector allocation is solved in reverse under the condition of satisfying medical prior constraints, and the obtained optimal weight vector is used as the training label. Based on the historical training sample set, the contextual features of historical moments are used as input features, and the optimal weight vector is used as the output target to construct a training sample set specific to each individual. Based on the training sample set, the gradient boosting tree algorithm is used for iterative learning. By fitting the prediction residual of the previous tree with multiple decision trees in turn, the nonlinear mapping relationship from context features to weight vector is learned, and an individual-specific weight prediction model is generated. The trained individual-specific weight prediction model is deployed to the real-time system. The individual-specific weight prediction model is used to output a personalized optimal weight allocation scheme based on the current context features when new real-time data is input.

[0108] In this embodiment, multi-source monitoring data and corresponding health results of the pet are collected over a preset historical period. The multi-source monitoring data covers various aspects, including the pet's physiological data, such as body temperature, heart rate, and blood pressure; behavioral data, such as activity level, activity time distribution, and sleep duration; environmental data, such as indoor temperature, humidity, and air quality; and historical record data, including the pet's past medical records and vaccination status. Simultaneously, the actual health results of the pet at the time of data collection are recorded, such as whether it is sick, the type of sickness, and the severity of the sickness. The collected multi-source monitoring data and corresponding health results are organized and matched to construct a historical training sample set. This sample set provides a rich data foundation for subsequent model training, reflecting the pet's health status in different states and the correlation between various factors.

[0109] Using known health outcomes as the optimization objective, the optimal weight vector allocation is solved in reverse under the condition of satisfying medical prior constraints. Medical prior constraints refer to the reasonable range and rules set for the weight allocation of various health indicators based on professional knowledge and experience in the medical field. For example, certain key physiological indicators should have a larger weight in health assessment, and the relative relationships of the weights of various indicators under different health conditions should conform to common medical sense. By inversely deriving from the health outcomes while satisfying these constraints, the weight vector allocation can most accurately reflect the pet's health status when evaluating multi-source monitoring data. The solved optimal weight vector is used as training labels. These labels provide a clear objective for model training, guiding the model to learn the optimal weight allocation method that conforms to medical reality.

[0110] Based on historical training sample sets, a personalized training sample set is constructed, using contextual features from historical moments as input features and the optimal weight vector as the output target. Contextual features are a comprehensive description of the pet's environment and state. In addition to the physiological, behavioral, and environmental data mentioned earlier, they can also include time factors, such as different times of day and seasons, as these factors may affect the pet's health and the importance of various indicators. These contextual features from historical moments are organized and feature-engineered, serving as input features; simultaneously, the optimal weight vector obtained at the corresponding moment is used as the output target. This constructed personalized training sample set establishes a correspondence between contextual features and the optimal weight vector for each pet's unique situation, providing targeted data for subsequent training of the personalized model.

[0111] Based on the constructed training sample set, the gradient boosting tree algorithm is used for iterative learning. The gradient boosting tree algorithm is a powerful machine learning algorithm that continuously improves the model's prediction accuracy by successively fitting the prediction residuals of the previous tree with multiple decision trees. During training, the algorithm inputs the training sample set into the model, using historical context features as input, and attempts to predict the corresponding optimal weight vector. Initially, the model may predict inaccurately, producing some prediction residuals. Then, the algorithm constructs a second decision tree specifically to fit the residuals produced by the first tree, allowing the second tree's predictions to correct the errors of the first tree. This process continues, and through iterative fitting of multiple decision trees, the model gradually learns the non-linear mapping relationship from context features to weight vectors. As training progresses, the model's prediction accuracy for the optimal weight vector continuously improves, ultimately generating an individual-specific weight prediction model. This model can accurately predict the optimal weight allocation scheme suitable for the current state of the pet based on the input context features.

[0112] The trained, individual-specific weight prediction model is deployed to a real-time system. In this system, when new real-time data is input, the system first extracts the current contextual features, including the pet's real-time physiological, behavioral, and environmental data, as well as time information. These contextual features are then input into the individual-specific weight prediction model, which quickly outputs a personalized optimal weight allocation scheme based on learned non-linear mapping relationships. This scheme is used for subsequent weighted evaluation of pet health indicators, ensuring that health assessments fully consider the pet's current individual characteristics and environment, improving accuracy and personalization. By deploying the model to a real-time system, real-time and accurate monitoring and evaluation of pet health can be achieved, providing pet owners with more valuable health reports and recommendations.

[0113] Reference Figure 2An embodiment of the present invention provides a pet health report intelligent generation system 2 based on multi-factor weights, wherein system 2 specifically includes: The data acquisition module 201 is used to collect the pet's physiological data, behavioral data, environmental data and historical record data, and to perform data cleaning, filtering and feature extraction to generate multi-dimensional quantitative features; The weighting factor module 202 is used to construct a weighting factor library based on multi-dimensional quantitative features, including environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors. The weight adjustment module 203 is used to calculate the adjustment coefficients of each factor in the weight factor library in parallel using a hierarchical multi-factor weighted model, synthesize the original weights through element-wise multiplication, and output the dynamic weights of each health indicator after normalization. The health assessment module 204 is used to perform weighted summation of the standardized scores of each health indicator based on dynamic weights, combine trend analysis algorithms to identify abnormal changes in the indicators, and determine the pet's health risk level based on the weighted score and the number of abnormal indicators. The report generation module 205 is used to generate a health report based on the health risk level, combined with dynamic weights and trend analysis, using natural language generation technology. The report includes data summaries, anomaly interpretations, trend analysis, and personalized suggestions, and embeds the weight allocation and adjustment reasons in the decision-making process into the health report. The strategy optimization module 206 is used to output the health report to the user in a structured format and optimize the weight strategy online using a reinforcement learning algorithm based on user interaction feedback and subsequent diagnostic results.

[0114] It is understandable that, such as Figure 1 The content of the illustrated embodiment of the intelligent pet health report generation method based on multi-factor weighting is applicable to the embodiment of this intelligent pet health report generation system based on multi-factor weighting. The specific functions implemented by the embodiment of this intelligent pet health report generation system based on multi-factor weighting are the same as those shown below. Figure 1 The illustrated embodiment of the intelligent pet health report generation method based on multi-factor weighting is the same, and the beneficial effects achieved are the same as those shown. Figure 1 The beneficial effects achieved by the illustrated embodiment of the intelligent pet health report generation method based on multi-factor weights are also the same.

[0115] 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.

Claims

1. A method for intelligently generating pet health reports based on multi-factor weighting, characterized in that, The method specifically includes: Collect pet physiological data, behavioral data, environmental data, and historical data, and perform data cleaning, filtering, and feature extraction to generate multi-dimensional quantitative features; Based on multi-dimensional quantitative characteristics, a weight factor library is constructed that includes environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors. A hierarchical multi-factor weighted model is used to calculate the adjustment coefficients of each factor in the weight factor library in parallel, and the original weights are synthesized by element-wise multiplication. After normalization, the dynamic weights of each health indicator are output. Based on dynamic weights, the standardized scores of each health indicator are weighted and summed, and trend analysis algorithms are used to identify abnormal changes in the indicators. The health risk level of the pet is determined based on the weighted score and the number of abnormal indicators. Based on health risk levels, combined with dynamic weights and trend analysis, a health report is generated using natural language generation technology, which includes data summaries, anomaly interpretations, trend analysis, and personalized suggestions. The weight allocation and adjustment reasons in the decision-making process are also embedded into the health report. The health report is output to the user in a structured format, and the weight strategy is optimized online using a reinforcement learning algorithm based on user interaction feedback and subsequent diagnostic results.

2. The method according to claim 1, characterized in that, The process involves collecting physiological, behavioral, environmental, and historical data of the pets, followed by data cleaning, filtering, and feature extraction to generate multi-dimensional quantitative features, specifically including: Physiological signals such as body temperature, heart rate and respiratory rate of pets were collected. After noise reduction using a sliding window filtering algorithm, the mean, variance and fluctuation coefficient were extracted as physiological features. Collect behavioral data on pets' activity levels, vocalization frequency, food intake, and water intake. After removing outliers caused by equipment malfunctions through anomaly detection, calculate the month-on-month change rate, year-on-year change rate, and trend slope as behavioral trend characteristics. Data on room temperature, humidity, light intensity, and noise levels in the environment are collected, and instantaneous fluctuations are smoothed using a moving average algorithm, while continuous values ​​are converted into discrete state codes as environmental features. A historical archive database containing information on breed, age, gender, medical history, and medication records was constructed, and an individual's health baseline was dynamically calculated based on historical monitoring data as an archive feature.

3. The method according to claim 1, characterized in that, The aforementioned weight factor library, constructed based on multi-dimensional quantitative features, includes environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors. Specifically, it includes: Based on multi-dimensional quantitative features, environmental factors are constructed by extracting parameters such as room temperature, light intensity, noise level, and time period. These environmental factors are used to reflect the adjustment needs of the external environment on the weight of the indicators. Based on multi-dimensional quantitative features, individual factors are constructed by extracting parameters such as variety, age, gender, medical history, and health baseline. These individual factors are used to reflect the benchmark impact of individual differences on the weight of indicators. Based on multi-dimensional quantitative features, behavioral factors are constructed by extracting parameters of current activity status and recent trend changes. These behavioral factors are used to capture the basis for real-time correction of indicator weights by behavioral patterns. Based on multi-dimensional quantitative features, data quality factors are constructed by extracting parameters such as sensor signal strength, occlusion degree, acquisition duration, and data integrity. These data quality factors are used to evaluate the degree to which data credibility inhibits the weight of indicators. Based on multi-dimensional quantitative features, historical response patterns and seasonality parameters of individuals are extracted to construct historical pattern factors. These historical pattern factors are used to identify the need for enhanced indicator weights due to historical deviations. Environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors are integrated to construct a weighted factor library.

4. The method according to claim 1, characterized in that, The method employs a hierarchical multi-factor weighted model to calculate the adjustment coefficients of each factor in the weight factor library in parallel, synthesizes the original weights through element-wise multiplication, and outputs the dynamic weights of each health indicator after normalization. Specifically, this includes: The built-in mapping table is queried based on the pet's breed, age, and medical history to obtain the basic weight vector of each health indicator; Calculate the adjustment coefficient vector of each factor in the weight factor library for each indicator, and multiply the basic weight vector of each health indicator with the adjustment coefficient vector of each factor for each indicator element by element to synthesize the original weight vector. The original weight vector is normalized so that the sum of the weights of all health indicators is one, and the dynamic weights are output.

5. The method according to claim 4, characterized in that, The calculation of the adjustment coefficient vector of each factor in the weight factor library for each indicator specifically includes: Based on environmental factors, an environmental adjustment coefficient vector is generated through predefined rule mapping. Based on individual factors, an individual adjustment coefficient vector is generated according to the degree of deviation between real-time indicators and the individual's health baseline; Based on behavioral factors, the isolated forest algorithm is introduced to identify the degree of abnormality between the current behavior and the individual's historical patterns, and to generate a behavioral adjustment coefficient vector. Based on the data quality factor, the isolated forest algorithm is introduced to detect the degree of anomaly in sensor signals and generate a data quality adjustment coefficient vector; Based on historical pattern factors, a historical pattern adjustment coefficient vector is generated by comparing the deviation of the current indicator value from the historical data of the same period.

6. The method according to claim 5, characterized in that, The process involves weighted summation of standardized scores for each health indicator based on dynamic weights, combined with trend analysis algorithms to identify abnormal changes in the indicators, and determining the pet's health risk level based on the weighted score and the number of abnormal indicators. Specifically, this includes: Based on dynamic weights, the standardized scores of each health indicator are weighted and summed to calculate the comprehensive health score. Trend analysis was performed on the time series data of various health indicators to calculate the month-on-month and year-on-year change rates of each health indicator compared with the historical average and recent average, and to identify significant trends of continuous increase, continuous decrease or abnormal fluctuation. Based on the overall health score and the number of abnormal indicators marked as showing significant trends, the pet's health status is classified into risk levels using preset risk level determination rules.

7. The method according to claim 6, characterized in that, Based on health risk levels, combined with dynamic weighting and trend analysis, a health report is generated using natural language generation technology. This report includes data summaries, anomaly interpretations, trend analysis, and personalized suggestions. The reasons for weight allocation and adjustments during the decision-making process are embedded into the health report. Specifically, this includes: Based on health risk levels, combined with dynamic weights and trend analysis, data summaries are generated for each core health indicator, and the indicator values ​​are converted into textual descriptions. Based on the degree of influence of each factor in the dynamic weight on health indicators, attribution explanations are generated for abnormal indicators to produce abnormal interpretation texts. Based on the significant changing trends identified by trend analysis, and combined with historical pattern factors and behavioral factors, trend analysis is interpreted to generate the potential health implications behind the changes in indicators. Based on the health risk level, the guidance of dynamic weights, and the characteristics of abnormal indicators, personalized action suggestions are generated and marked with priorities and basis. A health report is constructed based on textual descriptions, interpretations of anomalies, the potential health implications behind changes in indicators, and personalized action recommendations. Dynamic weights, the contribution of each health indicator to the final conclusion, and the reasons for weight adjustments are embedded in the health report.

8. The method according to claim 1, characterized in that, The online optimization of the weight strategy based on user interaction feedback and subsequent diagnostic results using a reinforcement learning algorithm specifically includes: The system collects user interaction feedback behavior on health reports and gathers subsequent veterinary diagnostic results, which together constitute a set of feedback signals. The interaction feedback behavior includes adopting suggestions, ignoring prompts, and manual correction operations. The current environmental characteristics, behavioral characteristics, physiological index deviation, and user historical feedback tendency parameters are used as state descriptions to construct a state space. Adjusting the weights of various health indicators is made an optional action, thus constructing an action space; Convert user adoption and correct diagnosis into positive reward signals, and convert user ignore, manual correction, and false positives / missed reports into negative reward signals. Design a reward function based on user feedback behavior and diagnosis results. Based on the state space and action space, a Q-learning algorithm is used to update the weights online, and the state-action value table is updated according to the reward signal output by the reward function.

9. The method according to any one of claims 1 to 8, characterized in that, The method further includes: Collect multi-source monitoring data and corresponding health results of pets within a preset historical period to construct a historical training sample set; Using known health outcomes as the optimization objective, the optimal weight vector allocation is solved in reverse under the condition of satisfying medical prior constraints, and the obtained optimal weight vector is used as the training label. Based on the historical training sample set, the contextual features of historical moments are used as input features, and the optimal weight vector is used as the output target to construct a training sample set specific to each individual. Based on the training sample set, the gradient boosting tree algorithm is used for iterative learning. By fitting the prediction residual of the previous tree with multiple decision trees in turn, the nonlinear mapping relationship from context features to weight vector is learned, and an individual-specific weight prediction model is generated. The trained individual-specific weight prediction model is deployed to the real-time system. The individual-specific weight prediction model is used to output a personalized optimal weight allocation scheme based on the current context features when new real-time data is input.

10. A pet health report intelligent generation system based on multi-factor weighting, characterized in that, The system specifically includes: The data acquisition module is used to collect pet physiological data, behavioral data, environmental data, and historical data, and to perform data cleaning, filtering, and feature extraction to generate multi-dimensional quantitative features. The weighting factor module is used to build a weighting factor library based on multi-dimensional quantitative features, including environmental factors, individual factors, behavioral factors, data quality factors, and historical pattern factors. The weight adjustment module is used to calculate the adjustment coefficients of each factor in the weight factor library in parallel using a hierarchical multi-factor weighted model, synthesize the original weights through element-wise multiplication, and output the dynamic weights of each health indicator after normalization. The health assessment module is used to perform weighted summation of standardized scores for each health indicator based on dynamic weights, combine trend analysis algorithms to identify abnormal changes in the indicators, and determine the pet's health risk level based on the weighted score and the number of abnormal indicators. The report generation module is used to generate health reports based on health risk levels, combined with dynamic weights and trend analysis, using natural language generation technology. The reports include data summaries, anomaly interpretations, trend analysis, and personalized suggestions, and embed the weight allocation and adjustment reasons from the decision-making process into the health reports. The strategy optimization module is used to output health reports to users in a structured format and optimize the weight strategy online using reinforcement learning algorithms based on user interaction feedback and subsequent diagnostic results.

Images