Pet daily management system and method based on AI agent

Abstract

The application discloses a pet daily management system and method based on an AI agent, and relates to the technical field of pet management. The system at least comprises an edge AI computing power box and a pet intelligent sensing execution terminal in communication connection with the edge AI computing power box; the pet intelligent sensing execution terminal is used for collecting pet monitoring data and receiving control instructions issued by the edge AI computing power box to perform linkage operation; the pet intelligent sensing execution terminal at least comprises a high-definition night vision camera, a WiFi intelligent water drinking device, a WiFi intelligent feeding device and a pet intelligent electronic collar; the edge AI computing power box is internally provided with an AI agent, and the AI agent is used for realizing automatic feeding, health monitoring and behavior intervention in combination with the pet monitoring data. The application is used for solving the problem that the existing pet intelligent devices cannot realize integrated and overall offline management and control.

Description

Technical Field

[0001] This invention relates to the field of pet management technology, and in particular to a pet daily management system and method based on an AI intelligent agent. Background Technology

[0002] As the scale of companion animal breeding in China continues to expand, the demand for intelligent pet management from families and pet stores continues to grow. Various independent intelligent products such as feeding, watering, and video monitoring have been launched on the market. Some of these devices are equipped with simple AI (Artificial Intelligence) programs, which can complete basic pet data collection and simple behavior recognition.

[0003] However, most existing smart pet products are independent of each other and lack a unified scheduling platform. Multi-source monitoring data cannot be processed collaboratively, making it difficult to complete pet health assessment and behavior control in an integrated manner. In addition, they are highly dependent on the external network, and their intelligent control capabilities are greatly limited after the network is disconnected, resulting in poor overall pet management functionality. Summary of the Invention

[0004] The purpose of this invention is to provide a pet daily management system and method based on AI intelligent agents, which solves the problem that existing pet intelligent devices cannot achieve integrated and coordinated offline management.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a pet daily management system based on an AI intelligent agent, the system comprising: an edge AI computing power box and a pet intelligent perception and execution terminal that is communicatively connected to the edge AI computing power box; The pet intelligent sensing and execution terminal includes at least a high-definition night vision camera, a WiFi intelligent drinking device, a WiFi intelligent feeding device, and a pet intelligent electronic collar; The edge AI computing power box has a built-in AI agent; the AI ​​agent is used to realize automatic feeding, health monitoring and behavior intervention based on pet monitoring data collected by the pet intelligent perception execution terminal.

[0006] Optionally, the pet monitoring data includes at least video and audio data collected by a high-definition night vision camera and wearable sensor data collected by a smart electronic pet collar; The AI ​​agent should include at least a pet behavior and health prediction model, a health indicator extraction module, a health indicator quantification module, and a triple verification module. The pet behavior and health prediction model is used to extract and fuse features from video data, audio data, and pet wearable sensor data to obtain unified fused features, and output pet predicted behavior and health prediction data based on the unified fused features; the health prediction data includes at least a physiological fused prediction feature vector. The health indicator extraction module includes two data input channels. One data input channel receives the physiological fusion prediction feature vector and performs dimensional splitting and parsing on the physiological fusion prediction feature vector to extract the predicted values ​​of each independent physiological indicator. The other data input channel receives drinking water data collected by the WiFi smart drinking device and food data collected by the WiFi smart feeding device as the raw values ​​of the dietary indicators. The health indicator quantification module is used to calculate the abnormal score of each physiological indicator by combining the predicted or raw values ​​of each physiological indicator with the historical baseline mean and standard deviation of the health indicators of the corresponding individual pet, and to obtain the initial screening abnormal indicators and their corresponding abnormal scores. The triple verification module is used to perform triple verification on the initial screening abnormal indicators and their corresponding abnormal scores in sequence to obtain a valid health abnormality judgment result.

[0007] Optionally, the pet behavior health prediction model includes an encoder and a decoder, with the decoder being a Transformer decoder; the encoder includes a feature extraction module and a feature fusion module; the decoder includes a pet behavior recognition output branch and a pet health assessment output branch; The feature extraction module is used to extract visual features from video data through the YOLOv8 backbone network, extract audio features from audio data through the MFCC algorithm, and normalize the pet's wearable sensor data to obtain vital sign sensor features. The feature fusion module is used to map visual features, audio features, and vital sign sensing features to the same feature dimension through linear projection, and to complete cross-modal fusion through a multi-head self-attention mechanism to output unified fused features. The pet behavior recognition output branch is used to process the unified fusion features to obtain the predicted pet behavior; the predicted pet behavior includes normal pet behavior, canine misbehavior, and feline misbehavior. The pet health assessment output branch is used to process the unified fusion features to obtain health prediction data.

[0008] Optionally, the triple verification module includes a multi-device cross-verification unit, an individual baseline comparison unit, and a veterinary expert rule base verification unit.

[0009] Optionally, the multi-device cross-validation unit is used to retrieve synchronously collected data from different types of associated devices for the initial screening of abnormal indicators, and to calculate the cross-validation score between the multiple devices. When the cross-validation score is greater than or equal to the first preset value, the cross-validation is deemed to have passed and proceeds to the next level of validation; otherwise, it is deemed to be a single-device sensor error and the corresponding initial screening abnormal indicator is directly filtered out.

[0010] Optionally, the individual baseline comparison unit is used to continuously collect the abnormal scores of the abnormal indicators that have passed the cross-validation of multiple devices for a preset number of detection cycles and calculate the average value; when the average value is greater than or equal to the second preset value and the indicator is determined to be abnormal for a preset number of consecutive detection cycles, the time-series baseline comparison verification is passed and enters the third verification; otherwise, it is determined to be a short-term data fluctuation and the corresponding abnormal indicator is filtered out.

[0011] Optionally, the system also includes a user terminal APP; The veterinary expert rule base verification unit is used to match abnormal indicators that have passed the second level of verification with a preset health assessment expert rule base to calculate the rule matching degree. When the rule matching degree is greater than or equal to the third preset value, the health abnormality is confirmed, a valid health abnormality judgment result is obtained, and the pet smart electronic collar is triggered to issue an audible and visual alarm or the user terminal APP is triggered to push medical information. When the matching degree is less than the third preset value, it is judged as a false positive and the corresponding abnormal indicator is filtered out.

[0012] Optionally, the health prediction data also includes a mood prediction feature vector and a sleep state prediction feature vector; The pet health assessment output branch includes three types of output heads, which respectively output physiological fusion prediction feature vector, emotion prediction feature vector, and sleep state prediction feature vector based on unified fusion features. The AI ​​agent also includes an emotion prediction module and a sleep state prediction module; The emotion prediction module is used to perform weighted fusion calculation on four types of features in the emotion prediction feature vector: activity level, body posture, pet barking, and vital signs, to obtain a comprehensive emotional score for the pet. Based on the comprehensive emotional score, the pet is classified into three emotional states: pleasant and calm, anxious and fearful, and aggressive and stressed. The emotional state corresponding to the comprehensive emotional score is written into the health report. The sleep state prediction module is used to calculate the pet's exercise volume based on the triaxial acceleration data and heart rate data in the sleep state prediction feature vector. It combines the resting heart rate interval threshold to classify the three sleep states as deep sleep, light sleep, and wakefulness, and writes the pet's exercise volume and sleep state into a health report.

[0013] Optionally, the edge AI computing power box also includes an audio codec module, and the AI ​​agent also includes a voice intervention scheduling unit; The voice intervention dispatch unit receives predicted pet behaviors. When the predicted pet behavior is a canine or feline undesirable behavior, it retrieves a pre-recorded voice prompt from the owner and plays it in real time through the audio codec module to complete the pet behavior intervention.

[0014] Compared with existing technologies, the pet daily management system based on AI intelligent agents provided by this invention unifies the communication of various pet sensing terminals, such as high-definition night vision cameras, WiFi smart drinking devices, WiFi smart feeding devices, and smart electronic collars, into a unified network. All pet monitoring data is collected locally at the edge, and the AI ​​intelligent agent built into the computing box completes the integrated computing processing, breaking down data silos between different devices. The entire automatic feeding, health monitoring, and behavior intervention logic all run locally at the edge, without relying on real-time computing power in the cloud. Even when the external network is disconnected, the core control functions can still be executed normally. This solves the defects of existing smart individual products, such as being fragmented, unable to coordinate multi-source data, highly dependent on the external network, and the failure of intelligent management functions after network disconnection, realizing integrated offline management of pet devices.

[0015] Secondly, this invention also provides a pet daily management method based on an AI agent, implemented using any of the above-mentioned AI agent-based pet daily management systems, the method comprising: Obtain pet monitoring data; By combining pet monitoring data, automatic feeding, health monitoring, and behavioral intervention can be achieved. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 One of the structural diagrams of a pet daily management system based on an AI agent provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of an AI agent provided in one embodiment of the present invention; Figure 3 A second schematic diagram of the structure of a pet daily management system based on an AI agent, provided as an embodiment of the present invention; Figure 4 This is a flowchart illustrating a pet daily management method based on an AI agent, as provided in one embodiment of the present invention.

[0017] Figure labeling: 1-Edge AI computing power box; 10-AI intelligent agent; 100-Pet behavior and health prediction model; 101-Health indicator extraction module; 102-Health indicator quantification module; 103-Triple verification module; 104-Emotion prediction module; 105-Sleep state prediction module; 106-Automatic feeding control module; 107-Lost location and control unit; 1001-Encoder; 1002-Decoder; 10011-Feature extraction module; 10012-Feature fusion module; 10021-Pet behavior recognition output branch; 10022-Pet health assessment output branch; 11-Audio encoding and decoding module; 2-Pet intelligent sensing execution terminal; 20-High-definition night vision camera; 21-WiFi intelligent drinking device; 22-WiFi intelligent feeding device; 23-Pet intelligent electronic collar; 3-User terminal APP. Detailed Implementation

[0018] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0019] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0020] In this invention, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between the associated objects, indicating that three relationships can exist.

[0021] like Figure 1 As shown, this embodiment of the invention provides a pet daily management system based on an AI intelligent agent. The system includes at least: an edge AI computing power box 1 and a pet intelligent perception and execution terminal 2 that is communicatively connected to the edge AI computing power box 1. The pet intelligent sensing and execution terminal 2 includes at least a high-definition night vision camera 20, a WiFi (Wireless Fidelity) intelligent drinking device 21, a WiFi intelligent feeding device 22, and a pet intelligent electronic collar 23; The edge AI computing power box 1 has a built-in AI agent 10; the AI ​​agent 10 is used to realize automatic feeding, health monitoring and behavior intervention based on pet monitoring data collected by the pet intelligent perception execution terminal.

[0022] In this embodiment, the system establishes an edge AI computing power box that communicates uniformly with various pet smart sensing and execution terminals. It integrates high-definition night vision cameras, WiFi smart drinking devices, WiFi smart feeding devices, and pet smart electronic collars into a complete set of pet smart sensing and execution terminals. This breaks down the fragmented nature of existing individual smart devices, unifying all pet monitoring data to the local edge AI computing power. Simultaneously, the system internally deploys an AI agent for pet analysis and management within the edge AI computing power box, enabling centralized processing of all dimensions of monitoring data. This allows for simultaneous computation related to daily pet management, health monitoring, and behavioral intervention, eliminating the need for external cloud-based AI analysis. This overcomes multiple shortcomings of existing devices, such as lack of a unified scheduling platform, difficulty in coordinating multi-source data processing, and failure of control functions after network outages. Ultimately, it achieves integrated offline management of pet devices, resolving the technical problems of poor management integrity and insufficient offline control capabilities in existing pet smart devices.

[0023] In one optional embodiment, the edge AI computing power box can use Rockchip RK3588, which has a built-in NPU (Neural Processing Unit) with 6 TOPS (TeraOperations Per Second) INT8 (Integer 8-bit) computing power to provide local AI inference capabilities. It is equipped with 8GB (Gigabyte) LPDDR4 (Low Power Double Data Rate 4) memory and 64 GB EMMC (Embedded Multimedia Card) storage to meet the data storage and computing needs of the edge. The specific implementation details are subject to the actual deployment.

[0024] The edge AI computing power box has built-in dual-band WiFi 6, Bluetooth 5.2, and Ethernet communication interfaces, enabling local area network devices to network and communicate with the external network.

[0025] like Figure 2 As shown, the edge AI computing power box can also be equipped with an audio codec module 11, which supports voice assistant and voice intervention functions.

[0026] The edge AI computing box has reserved HDMI (High-Definition Multimedia Interface) and USB (Universal Serial Bus) universal peripheral interfaces to support function expansion.

[0027] In one optional embodiment, the recommended deployment scheme for high-definition night vision cameras is 2-4 units for a home setting, installed in the living room, pet play area, feeding and watering area, and pet toilet area respectively; for a pet shop setting, 10 cameras can be installed per pet store. Deployment density is one unit per enclosed space. Users can adjust the number of devices and installation locations according to the actual scenario. The high-definition night vision camera can be a 2-megapixel starlight-level night vision camera, supporting H.265 encoding and a frame rate of 30fps. It is networked with the edge AI computing power box via WiFi, and the collected video and audio data are processed and stored locally only on the edge AI computing power box.

[0028] Correspondingly, pet monitoring data includes at least video and audio data collected by high-definition night vision cameras.

[0029] In one alternative embodiment, the WiFi smart drinking device includes a main unit and a detachable drinking basin; for example, the drinking basin can be made of stainless steel. The main unit is equipped with a composite physical filter cartridge of PP cotton and activated carbon, a TDS water quality sensor, an ultrasonic level sensor, and a miniature water pump. The filter cartridge, TDS water quality sensor, ultrasonic level sensor, and miniature water pump are connected to the drinking basin through the internal water circuit, and the filtered water is input into the drinking basin. The level sensor is used to collect the amount of water in the drinking basin; Miniature water pumps are used to automatically replenish water into drinking basins; The TDS water quality sensor has a detection range of 0–9990 ppm and a detection accuracy of ±2%. Ultrasonic liquid level sensor with a liquid level detection accuracy of ±1mm; WiFi smart drinking water devices connect to an edge AI computing power box via a WiFi local area network, uploading water quality, water quantity, and drinking water volume data to the edge AI computing power box in real time, and receiving automatic water replenishment control commands issued by the edge AI computing power box.

[0030] Correspondingly, pet monitoring data includes, but is not limited to, the aforementioned water quality, water quantity, and drinking water volume data, as well as other drinking water data.

[0031] In one optional embodiment, the WiFi smart feeding device has a built-in weighing sensor and a spiral feeding structure; the weighing sensor is used to monitor the remaining grain and the amount of food fed at one time in real time; for example, the weighing sensor (e.g., weighing accuracy ±1g).

[0032] The spiral feeding system is used for timed and quantitative feeding. For example, the spiral feeding system can be configured with differentiated feeding plans for pet cats and dogs.

[0033] The WiFi smart feeding device connects to an edge AI computing box via a WiFi local area network, uploading data on food intake, feeding frequency, and remaining food supply to the edge AI computing box in real time, and receiving feeding control commands from the edge AI computing box.

[0034] Correspondingly, pet monitoring data includes, but is not limited to, the aforementioned food data such as food intake, feeding frequency, and remaining food reserves.

[0035] In one alternative embodiment, the pet smart electronic collar is equipped with an adjustable silicone strap, for example, an adjustable silicone strap that fits full-size cats and dogs with a neck circumference of 15cm to 60cm.

[0036] For example, the pet smart electronic collar has an overall IP67 dustproof and waterproof rating.

[0037] The pet smart electronic collar can be equipped with an ESP32-C3 main control chip, a GPS / BeiDou dual-mode positioning module, a photoelectric heart rate sensor, a contact body temperature sensor, a three-axis accelerometer, a high-brightness red light, a speaker, and a 1000mAh lithium battery. The data acquisition functions and execution purposes of each hardware component are clearly defined. ESP32-C3 main control chip: Supports dual communication modes of WiFi and NB-IoT (Narrow Band Internet of Things), and is responsible for all data transmission and reception and command scheduling of the collar, but has no physiological data acquisition function; GPS / BeiDou dual-mode positioning module with positioning accuracy of ±5m, specifically designed to collect real-time location coordinates of pets; Photoelectric heart rate sensor to collect real-time heart rate data of pets individually; Contact-type body temperature sensor, which collects pet's surface body temperature data individually; A three-axis accelerometer is used to collect temporal motion sensing data on pet activity levels, limb posture, respiratory rate, sleep duration, and sleep depth. High-brightness red light and speaker: These are hardware actuators used only to realize sound and light alarms and remote pet finding sound and light prompts, and do not participate in any monitoring data collection. 1000mAh lithium battery: Powers the entire collar hardware, supports magnetic charging, and has a single-charge battery life of no less than 7 days. It has no data collection capability. It should be noted that the pet's respiratory status is not directly collected by sensors in the pet's smart electronic collar, but is estimated jointly through heart rate variability and visual body posture characteristics. The specific implementation logic is as follows: based on the physiological pattern of respiratory sinus arrhythmia, respiratory rhythm characteristics are extracted from the heart rate variability data collected by the heart rate sensor. At the same time, the visual respiratory characteristics are obtained by analyzing the image through a high-definition night vision camera and tracking the frequency of chest and abdominal fluctuations. The real-time respiratory rate and respiratory stability of the pet are estimated by weighted fusion of the two types of characteristics. The respiratory-related results are only used as an auxiliary reference dimension for health assessment and do not constitute a basis for clinical diagnosis.

[0038] Correspondingly, all time-series monitoring data collected by the positioning module, heart rate sensor, body temperature sensor, and triaxial accelerometer are aggregated to form pet wearable sensor data, which is a component of the system's complete pet monitoring data.

[0039] In an exemplary embodiment, see Figure 2 The AI ​​agent 10 includes at least a pet behavior health prediction model 100, a health indicator extraction module 101, a health indicator quantification module 102, and a triple verification module 103; among them, this system covers a complete range of pet health assessment dimensions, and the corresponding health indicators are divided into four categories: physiological indicators, dietary indicators, emotional indicators, and sleep indicators. The pet behavior and health prediction model 100 is used to extract and fuse features from video data, audio data, and pet wearable sensor data to obtain unified fused features, and output pet predicted behavior and health prediction data based on the unified fused features; the health prediction data includes at least a physiological fused prediction feature vector. The health indicator extraction module 101 includes two data input channels. One data input channel receives the physiological fusion prediction feature vector and performs dimensional splitting and parsing on the physiological fusion prediction feature vector to extract the predicted values ​​of each independent physiological indicator. The other data input channel receives drinking water data collected by the WiFi smart drinking device and food data collected by the WiFi smart feeding device as the raw values ​​of the dietary indicators. The health indicator quantification module 102 is used to calculate the abnormal score of each physiological indicator by combining the predicted or original value of each physiological indicator with the historical baseline mean and standard deviation of the health indicators of the corresponding individual pet, and to obtain the initial screening abnormal indicators and their corresponding abnormal scores. The triple verification module 103 is used to perform triple verification on the initial screening abnormal indicators and their corresponding abnormal scores in sequence to obtain a valid health abnormality judgment result.

[0040] Regarding the implementation process of the health indicator extraction module, for example, the output branch of the pet health assessment outputs a physiological fusion prediction feature vector as follows: The physiological fusion prediction feature vector is packaged and stored at once, containing the predicted values ​​of three types of physiological indicators without separate boundaries. After receiving the physiological fusion prediction feature vector, the first input channel of the health indicator extraction module performs splitting and extraction: (1) The physiological fusion prediction feature vector is decomposed into dimensions, and three independent numerical vectors are separated according to a fixed dimension index. The first dimension is analyzed and extracted: the predicted value of heart rate; the second dimension is analyzed and extracted: the predicted value of body surface temperature; the third dimension is analyzed and extracted: the predicted value of daily cumulative activity. (2) After splitting, the predicted values ​​of the three independent physiological indicators of heart rate, body temperature and daily activity are output to the downstream health indicator quantification module for calculating the abnormal score of each indicator separately, so as to realize the independent abnormal judgment of a single indicator.

[0041] See Figure 2 The pet behavior health prediction model 100 includes an encoder 1001 and a decoder 1002, with the decoder being a Transformer decoder; the encoder 1001 includes a feature extraction module 10011 and a feature fusion module 10012; the decoder 1002 includes a pet behavior recognition output branch 10021 and a pet health assessment output branch 10022. The feature extraction module 10011 is used to extract visual features from video data through the YOLOv8 backbone network, extract audio features from audio data through the MFCC (Mel-Frequency Cepstral Coefficients) algorithm, and normalize the pet's wearable sensor data to obtain vital sign sensor features. The feature fusion module 10012 is used to map visual features, audio features, and vital sign sensing features to the same feature dimension through linear projection, and to complete cross-modal fusion through a multi-head self-attention mechanism to output unified fused features. The pet behavior recognition output branch 10021 is used to process the unified fusion features to obtain the predicted pet behavior; the predicted pet behavior includes normal pet behavior, canine misbehavior, and feline misbehavior. The pet health assessment output branch 10022 is used to process the unified fusion features to obtain health prediction data.

[0042] In a specific embodiment of a cat and dog home-raising scenario, the above-mentioned AI agent architecture is implemented in a home scenario where cats and dogs are raised simultaneously. All algorithms of the entire pet behavior and health prediction model are adapted to the local computing power environment of the edge AI computing box. Lightweight deployment is achieved through INT8 quantization and compression. The Transformer decoder, feature extraction module, feature fusion module, and dual-output branch architecture from the previous embodiment are reused. The data processing flow of the cat and dog pet behavior and health prediction model is as follows: (1) The cat and dog pet behavior health prediction model is developed based on the YOLOv8 open-source basic visual large model and the Transformer multimodal fusion framework. Dedicated sub-models are trained for pet cats and pet dogs respectively. The underlying lightweight feature extraction network is shared. The upper layer of the Transformer decoder is configured with cat and dog differentiated classification heads and health assessment rules. Based on the feature extraction module and feature fusion module, multi-class heterogeneous data fusion is completed, which serves as the data foundation for pet behavior classification and health assessment.

[0043] Among them, visual features The feature extraction module extracts cat and dog posture, behavior, and surface visual features based on the YOLOv8 backbone network, with dimensions... .

[0044] Vital signs and sensory characteristics The feature extraction module normalizes wearable sensor data such as heart rate, body temperature, activity level, and sleep uploaded from the pet smart electronic collar to generate a feature-rich data set. .

[0045] Audio features The feature extraction module extracts cat and dog barking sounds and environmental audio features using the MFCC algorithm, with dimensions... .

[0046] (2) Use the formula: (1) Linear projection is applied to three types of features—video features, audio features, and vital sign sensing features—and uniformly mapped them to a 512-dimensional common feature space to address the feature dimension mismatch problem. Among these, , , The weight matrix is ​​a learnable matrix. , , This is a bias term.

[0047] (3) Adopt the standard Transformer self-attention computation paradigm: (2) (3) (4) This study explores the correlations among three different modalities of data: video features, audio features, and vital sign sensing features, to achieve cross-modal feature fusion and scaling factor. To avoid the vanishing softmax gradient, the final output is a unified fused feature. Then, the data is split and sent to the pet behavior recognition output branch and the pet health assessment output branch of the decoder, respectively. It is a query vector; It is a key vector; These are value vectors; the three are the three core feature matrices of the self-attention mechanism, all composed of concatenated unified features. After three sets of independent learnable weight matrices , , The linear projection yields the following calculation relationship: , , The three will map the same multimodal feature to three different semantic subspaces, which will respectively undertake three different functions: query matching, identity identification, and information carrying, thereby realizing the mining of cross-modal correlation within the feature; It is a scaled dot product self-attention function; It is a feature concatenation function; It is an intermediate feature vector formed by concatenating three types of features: video features, audio features, and vital sign sensing features.

[0048] It should be noted that during the pre-training process of the cat and dog pet behavior health prediction model, a weighted cross-entropy loss function is used to differentiate between canine undesirable behaviors, feline undesirable behaviors, and normal pet behaviors based on the differentiated labels. : (5) To address the imbalance in the distribution of cat and dog samples, higher category weights are assigned to low-frequency behaviors such as cat stress and dog jumping on people. Where N is the total number of samples and C is the total number of behavioral categories, for example, the total number of behavioral categories includes 4 types of canine bad behavior, 4 types of cat bad behavior and normal pet behavior; Let i be the true label of the i-th sample. Weights are used to predict probabilities for the model. It is inversely proportional to the total number of samples in the corresponding category, thus improving the accuracy of identifying undesirable behaviors of cats and dogs in niche areas.

[0049] The beneficial effects of this embodiment are analyzed as follows: 1) Existing market feeding, watering, camera, and wearable devices are independent of each other, and various monitoring data cannot be shared. Moreover, they rely on only single-dimensional information for identification and judgment, resulting in poor accuracy in differentiating pets such as cats and dogs. This solution integrates a multimodal feature fusion model, health data processing, and anomaly verification functions into the same AI intelligent body. The pet behavior and health prediction model uniformly receives three types of heterogeneous data: video, audio, and wearable sensor data. It completes cross-modal feature fusion through linear projection and multi-head self-attention, and outputs two prediction results for behavior and health based on comprehensive multi-dimensional information. At the same time, the health indicator extraction module synchronously accesses the dietary monitoring data collected by the watering and feeding devices, realizing unified collection and processing of monitoring data for all types of pets, breaking the data silo problem of independent devices, and improving the comprehensive accuracy of pet behavior and health status identification by relying on multimodal fusion algorithms; 2) Existing health monitoring solutions only use data from a single device to determine health abnormalities, which is easily affected by environmental and equipment errors, resulting in a large number of misjudgments. In this embodiment, the dual-channel health indicator extraction module simultaneously integrates two types of complete health indicators: physiological indicators and dietary indicators, covering all dimensions of pet physiological and feeding health information. The health indicator quantification module calculates standardized abnormal scores based on the pet's specific historical baseline to complete the initial screening. Then, the triple verification module reviews the abnormal indicators of the initial screening layer by layer, forming a complete and systematic health abnormality filtering process. This avoids misdiagnosis and false alarm problems caused by single-dimensional judgment and improves the reliability of pet health monitoring results.

[0050] In an exemplary embodiment, the triple verification module includes a multi-device cross-verification unit, an individual baseline comparison unit, and a veterinary expert rule base verification unit.

[0051] Specifically, the multi-device cross-validation unit is used to retrieve synchronously collected data from different types of associated devices for the initial screening of abnormal indicators, and to calculate the cross-validation score between multiple devices. When the cross-validation score is greater than or equal to the first preset value (e.g., 0.8), the cross-validation is considered successful and proceeds to the next level of validation; otherwise, it is considered a single-device sensor error and the corresponding initial screening abnormal indicator is directly filtered out.

[0052] In Implementation Method 1, all initial screening abnormal indicators are first quantified using a health abnormality scoring system to obtain a single-indicator abnormality score. Then, it is sent to a multi-device cross-validation unit for multi-device cross-validation, and the calculation is based on the statistical 95% confidence interval (2σ principle), as shown in the following formula: (6) in, It is the real-time detection value of the kth physiological indicator; This is the historical baseline mean of this indicator for that individual pet; It is calculated based on the basic information of the pet and normal data over a historical period (such as the past 7 days); The range of values ​​is A higher score indicates a higher degree of abnormality; the initial screening threshold for a single physiological indicator is [not specified]. For example, it can be set to 1. If the deviation exceeds twice the standard deviation, it can be judged as an abnormality in the initial screening.

[0053] For the abnormal indicators identified in the initial screening, the multi-device cross-validation unit calculates the multi-device cross-validation score. The formula is as follows: (7) Where M represents the number of associated devices. The abnormal score for the k-th associated indicator corresponding to the m-th device; the verification rule matches the first preset value. A value ≥0.8 indicates that the cross-verification of multiple devices has passed.

[0054] Next, we will explain the multi-device cross-validation unit using a specific scenario of cats and dogs living at home. This system includes all pet intelligent sensing execution terminals: high-definition night vision cameras, WiFi smart drinking devices, WiFi smart feeding devices, and pet smart electronic collars. The associated device group is defined as a complete set of pet intelligent sensing execution terminal devices capable of synchronously collecting data related to the same health dimension and mutually verifying the authenticity of initial screening abnormal indicators. M represents the total number of terminal devices participating in the calculation in the current round of cross-validation, including the core terminal that outputs the original data of the indicator and the associated terminals used for verification. A typical verification scenario is illustrated below: Scenario 1: The initial screening abnormal indicators are activity level / heart rate Core pet intelligent sensing and execution terminal: pet intelligent electronic collar (outputs heart rate and three-axis activity data from the pet's wearable sensor data, serving as the first device participating in the calculation, m=1); Associated pet intelligent perception execution terminal: 20 high-definition night vision cameras (relying on the YOLOv8 backbone network to analyze video data, extract visual features of pet activity, calculate the abnormal activity score in the same dimension, and serve as the second device participating in the calculation, m=2). The total number of terminals participating in the verification in this group is M=2, and the output of the pet smart electronic collar will be... High-definition night vision camera output Substitute them into formula (7) to calculate the mean, and obtain the cross-validation score of the indicator. .

[0055] Scenario 2: The initial screening abnormal indicator is the average daily water intake. Core pet intelligent sensing and execution terminal: WiFi smart water drinking device (outputs single and daily water consumption values ​​in the water drinking data, m=1); Associated pet intelligent sensing and execution terminal: High-definition night vision camera (analyzes video and audio data to identify the frequency of pet drinking behavior and generates anomaly scores for drinking dimension, m=2); The total number of terminals participating in the verification in this group is M=2. The scores output by the two terminals are used together to calculate the cross-validation score.

[0056] Scenario 3: The initial screening abnormal indicator is daily food intake. Core pet intelligent sensing and execution terminal: WiFi intelligent feeding device (outputs food data including single feeding weight and feeding frequency, m=1); Associated pet intelligent sensing execution terminal: High-definition night vision camera (analyzes video data to identify pet eating scenes, matches eating frequency characteristics to generate corresponding abnormal scores, m=2); The total number of terminals participating in the verification in this group is M=2.

[0057] Scenario 4: The initial screening abnormal indicators are systemic metabolic syndrome abnormal indicators. Simultaneously, four types of pet smart sensing and execution terminals were activated for cross-verification: pet smart electronic collar (m=1), WiFi smart drinking device (m=2), WiFi smart feeding device (m=3), and high-definition night vision camera (m=4), with a total of M=4 terminals participating in the verification. Each of the four terminals outputs corresponding metabolic dimensions. Substituting all values ​​into the formula and calculating the average yields... .

[0058] It should be noted that relying solely on the output values ​​of a single pet intelligent sensing terminal is susceptible to hardware interference such as dirty sensors, loose device fittings, and liquid level detection errors, which can lead to false initial screening anomalies. Only by using a multi-device cross-validation unit to simultaneously retrieve data from multiple terminals within the same dimension and jointly calculate the score can the results be accurately determined. The verification is considered successful only if the value is greater than or equal to the first preset value of 0.8. By filtering out single-device errors through cross-validation of data from multiple terminals, the misdiagnosis rate of health detection is reduced.

[0059] Specifically, the individual baseline comparison unit is used to continuously collect the abnormal scores of anomalies for a preset number (e.g., 3) detection cycles and calculate the average for anomalies that have passed multi-device cross-validation. When the average is greater than or equal to the second preset value (e.g., 0.9) and the anomalies are determined to be abnormal for a preset number of consecutive detection cycles, the time-series baseline comparison verification passes and enters the third verification. Otherwise, it is determined to be short-term data fluctuation and the corresponding abnormal indicators are filtered out.

[0060] In the first specific implementation method, the time-series average of the single-index anomaly scores for three consecutive detection cycles is calculated. This is used to exclude short-term, instantaneous data fluctuations, and the calculation formula is as follows: (8) in, The abnormal score of the k-th health indicator in the t-th detection period; the verification rule matches the second preset value. When the value is ≥0.9 and the indicator is marked as abnormal for three consecutive periods, the time series baseline comparison verification is passed.

[0061] Specifically, see Figure 3 The system also includes a user terminal APP3; a veterinary expert rule base verification unit is used to match the abnormal indicators that have passed the second verification with the preset health assessment expert rule base and calculate the rule matching degree; when the rule matching degree is greater than or equal to the third preset value (e.g., 0.7), the health abnormality is confirmed, a valid health abnormality judgment result is obtained, and the pet smart electronic collar is triggered to issue an audible and visual alarm or the user terminal APP3 is triggered to push medical information; when the matching degree is less than the third preset value (e.g., 0.7), it is judged as a false positive and the corresponding abnormal indicator is filtered out.

[0062] For example, the pre-defined health assessment expert rule base can be a non-invasive expert rule base built on "Canine and Feline Clinical Pathology".

[0063] In the first specific implementation method, after the veterinary expert rule base verification unit matches the canine and feline clinical expert rule base, it uses the following formula: (9) Calculate the matching degree R of the health indicator rules, verify that the rules match the third preset value, and determine that the verification is successful when R≥0.7.

[0064] Only abnormal data that passes the above three verifications will trigger subsequent audio-visual alarms and APP push notifications. If any verification fails, the abnormal indicators and related abnormal data will be directly filtered out. In this way, the misdiagnosis rate of health detection is reduced from the root of the algorithm.

[0065] The beneficial effects of this exemplary embodiment: The triple verification module in this embodiment consists of three levels of units connected in series: a multi-device cross-verification unit, an individual baseline comparison unit, and a veterinary expert rule base verification unit. It has the following advantages: 1) The multi-device cross-verification unit (first level of verification) achieves the first layer of filtering for false screening anomalies caused by hardware / sensor errors of a single pet intelligent sensing execution terminal. Existing solutions rely solely on the values ​​of a single device to judge health indicators, which is prone to hardware interference and false alarms such as loose collars, dirty water level probes, and camera recognition deviations. This unit can retrieve synchronous data from multiple pet intelligent sensing execution terminals that can corroborate the same health dimension, calculate cross-verification scores, and only proceed to the next level of verification when the data synchronization of multiple devices is abnormal. When the value of a single device exceeds the standard but the data of the corroborating device is normal, it is directly filtered out, eliminating invalid abnormal indicators caused by sensor failures from the hardware source and significantly reducing the base number of false judgments at the hardware level. 2) The individual baseline comparison unit (second level of verification) achieves the second layer of filtering for temporary indicator exceedances caused by instantaneous short-term data fluctuations. Pets may experience transient states such as short-term strenuous exercise, small amounts of food or water consumed at one time, or brief stress, which may cause their health scores to temporarily exceed the threshold in a single testing cycle. However, these are not considered persistent health issues. This unit continuously collects abnormal scores from three testing cycles and calculates the time-series average. Only when multiple consecutive cycles show synchronous abnormalities and the average is ≥0.9 is the score released. This accurately distinguishes between transient data fluctuations and persistent health risks, preventing temporary fluctuations from triggering invalid health alarms. 3) The veterinary expert rule base verification unit (third-level verification) binds quantitative values ​​with the clinical pathology logic of cats and dogs, filtering out false positives that are simply numerically excessive but have no clinical significance, while simultaneously completing the closed-loop handling of abnormalities. The first two layers only perform mathematical quantitative verification, which cannot distinguish whether numerical deviations conform to the pathological characteristics of pet diseases. This unit relies on canine and feline clinical pathology to build a professional non-invasive rule base, calculates the pathological rule matching degree, and only outputs a valid health abnormality judgment result when the abnormal indicator matches the corresponding cat and dog disease pathological pattern. It also links the pet's smart electronic collar with sound and light alarms and the user terminal APP to push medical information. Data that exceeds the standard in pure numbers but does not conform to the clinical pathological logic is directly judged as false positives and filtered, so that health judgment is upgraded from simple mathematical calculation to a credible conclusion with veterinary professional reference value. 4) The triple verification serial architecture works together as a whole, with layered and progressive, classified filtering to achieve multi-dimensional three-dimensional misdiagnosis interception: The three-layer unit reviews the initial screening of abnormal indicators from three completely independent dimensions: "device hardware reliability → time series continuity → veterinary clinical rationality", and removes misjudgments of different causes in a layered and targeted manner. Compared with the traditional health monitoring solution with a single 2σ threshold scoring and single-layer verification, multi-layer superimposed filtering can significantly reduce the overall misdiagnosis rate of health detection from the root, and solve the core defect of frequent misjudgments in existing pet health monitoring.5) Tiered verification operates independently, balancing filtering accuracy with no missed anomalies: The three-tiered verification employs a unidirectional serial logic of "direct filtering if not met, and confirmation of anomaly only when all criteria are met." Each layer only removes invalid data of the corresponding type, preventing the false filtering of genuine, continuous, multi-device synchronized, and pathologically consistent health anomalies. This significantly reduces false alarm pushes while ensuring that genuine pet illness anomalies are fully preserved and trigger veterinary reminders, balancing noise reduction and health monitoring reliability. 6) The triple verification adopts a complete and standardized health anomaly judgment chain, outputting standardized results that can be directly linked to control devices: A standardized algorithm chain is formed from quantitative scoring of health indicators → cross-verification of multiple devices → temporal stability verification → clinical pathology matching, uniformly outputting valid health anomaly judgment results. The judgment results can directly drive audible and visual alarms and APP veterinary push notifications, creating a closed loop of "indicator collection - anomaly review - treatment push," solving the problems of existing solutions that only provide numerical values, lack reliable judgment conclusions, and cannot automatically tiered alarms.

[0066] In an exemplary embodiment, the health prediction data also includes a mood prediction feature vector and a sleep state prediction feature vector; see [link to example]. Figure 2 The pet health assessment output branch includes three types of output heads, which output physiological fusion prediction feature vectors, emotion prediction feature vectors, and sleep state prediction feature vectors based on unified fusion features, respectively. Combination Figure 2 It can be seen that the AI ​​agent 10 also includes an emotion prediction module 104 and a sleep state prediction module 105; The emotion prediction module 104 is used to perform weighted fusion calculation on four types of features in the emotion prediction feature vector: activity level, body posture, pet barking, and vital signs, to obtain a comprehensive emotional score for the pet. Based on the comprehensive emotional score, the pet is classified into three emotional states: pleasant and calm, anxious and fearful, and aggressive and stressed. The emotional state corresponding to the comprehensive emotional score is written into the health report. The sleep state prediction module 105 is used to calculate the pet's exercise volume based on the triaxial acceleration data and heart rate data in the sleep state prediction feature vector, and to divide the sleep state into three types: deep sleep, light sleep, and wakefulness, combined with the resting heart rate interval threshold, and to write the pet's exercise volume and sleep state into a health report.

[0067] In the second specific implementation method, the specific algorithm in the emotion prediction module is illustrated by example: The sentiment prediction module uses a multi-dimensional weighted fusion algorithm with a total weight of 1 to perform sentiment quantification. It relies entirely on the four types of feature data within the sentiment prediction feature vector for calculation. The calculation formula is as follows: (10) in, The pet's overall emotional score ranges from 0 to 10. A higher score indicates stronger negative emotions (anxiety, fear, aggression) in the pet. To normalize the score for activity level, the weights are... ; Score abnormal posture visual features with fixed weights ; Scoring is applied to abnormal pet vocalizations using fixed weights. ; For abnormal vital signs, a fixed weight is assigned. .

[0068] when ∈[0,2]: The pet is in a happy or calm state; when ∈(2,5]: The pet is in a state of anxiety or fear; when ∈(5,10]: The pet is in an aggressive or stressed state; after the emotion prediction module calculates the comprehensive score and matches the corresponding emotion category, the score value and emotion state are simultaneously entered into the pet's weekly health report.

[0069] In the second specific implementation method, the specific algorithm in the sleep state prediction module is illustrated by example: The sleep state prediction module, based on the raw triaxial acceleration and heart rate data carried by the sleep state prediction feature vector, first calculates the pet's activity level A per minute using the following formula: (11) in, , , The value of the triaxial acceleration at time t is in g (gravitational acceleration, 1g = 9.8m / s2); T is the total number of built-in sampling points per minute, with a default configuration of T = 60.

[0070] Based on the activity level (A) per minute and the individual pet's resting heart rate range, three sleep state determination rules are established: Deep sleep: A < 0.1g, and heart rate falls within 85%~95% of resting heart rate, with this state lasting ≥ 10min; Light sleep: 0.1g≤A<0.5g, and heart rate falls within 92%~100% of resting heart rate, with this state lasting ≥5min; Awake: A ≥ 0.5g, or heart rate exceeding 100% of resting heart rate.

[0071] In practice, the sleep state prediction module summarizes the amount of exercise and the total duration of deep / light sleep throughout the day. The average daily amount of exercise and the statistical data of the duration of various types of sleep are written into the pet's weekly health report as a core reference indicator for assessing the pet's long-term physical health.

[0072] The beneficial effects of this exemplary embodiment are as follows: 1) The emotion prediction module adopts a four-dimensional multimodal weighted fusion algorithm, which simultaneously integrates multi-dimensional information such as activity level, body posture, vocalizations, and vital signs to make a comprehensive judgment. This overcomes the limitations of a single activity level for rough assessment and adapts to the differentiated behavioral characteristics of different pets (e.g., cats and dogs), significantly improving the accuracy of identifying negative emotions such as cat stress and canine aggression. Unlike traditional devices that simply use a binary judgment of normal / abnormal, this solution outputs a continuous emotion score in the range of 0 to 10 and divides it into three emotion levels. This allows for a direct differentiation of the severity of a pet's anxiety, facilitating accurate assessment of the pet's psychological state by owners and veterinarians. The system automatically saves the emotion score and corresponding status daily and writes them into a weekly health report. Over a long period, this data can be accumulated to track the cyclical changes in the pet's emotions, identify hidden psychological problems such as anxiety and depression in advance, and address the shortcomings of existing devices that can only capture momentary abnormalities and cannot track long-term emotional changes. 2) The sleep state prediction module combines exercise volume and the pet's personalized resting heart rate to determine sleep state. Combined with duration filtering rules, it accurately distinguishes between deep sleep, light sleep, and wakefulness, filtering out misjudgments caused by brief periods of lying down. This addresses the shortcomings of traditional products that can only count total duration and cannot identify insufficient deep sleep, which is an early warning sign of endocrine and urinary tract diseases in cats and dogs. The system automatically summarizes all sleep-related indicators daily, including average daily exercise volume and the duration of both deep and light sleep, and enters them into a health report, transforming raw sensor data into standardized feeding assessment criteria for veterinarians. The entire sleep calculation process is completed locally at the edge, continuously storing time-series sleep data offline. Once the network is restored, the data is fully synchronized to the user's app, preventing data loss during network outages. 3) This embodiment adds two new assessment dimensions—psychological emotion and sleep structure—to the existing triple physiological verification, constructing a comprehensive pet health evaluation system integrating physical physiology, psychological emotion, and sleep state, thus addressing the industry's shortcomings in existing health monitoring dimensions. Physiological abnormality records, daily mood data, and full-cycle sleep statistics are uniformly summarized to generate a weekly health report, along with personalized feeding suggestions. Compared to scattered, instantaneous alerts, this provides families, pet stores, and veterinarians with a long-term, quantitative, and complete pet health record, significantly improving the professionalism of intelligent pet management. Emotional weighting calculations and sleep time-series statistics all rely on local edge computing power, requiring no cloud-based computation. It shares local inference resources with the multimodal fusion and triple-verification modules, fully realizing the goal of integrated offline system management.

[0073] In an exemplary embodiment, see Figure 3The edge AI computing power box 1 also includes an audio codec module 11, and the AI ​​intelligent agent also includes a voice intervention scheduling unit; The voice intervention dispatch unit receives predicted pet behaviors. When the predicted pet behavior is a canine or feline undesirable behavior, it retrieves a pre-recorded voice prompt from the owner and plays it in real time through the audio codec module to complete the pet behavior intervention.

[0074] For example, the total end-to-end delay from identifying undesirable pet behavior to playing the voice message can be less than or equal to 5 seconds; if the identification result is normal pet behavior, no voice intervention will be performed.

[0075] The beneficial effects of this exemplary embodiment are as follows: 1) Relying on the edge AI computing power box to complete behavior reasoning locally, it achieves end-to-end ultra-low latency real-time intervention of ≤5 seconds. The entire recognition and voice transmission process does not require uploading to the cloud for computing and transmission, eliminating the lag problems caused by network transmission and cloud reasoning. Even when the external network is disconnected and the device is in offline operation mode, it can still immediately play the stopping voice when the pet's bad behavior is recognized, solving the defects of existing cloud intervention solutions, such as high latency and direct failure of intervention function after network disconnection. 2) It uses the owner's exclusive voice recorded in advance by the user as the intervention prompt sound, which is different from the uniform and standardized prompt sounds of devices on the market. It is more adaptable to the deterrence and soothing of pets such as cats and dogs, and can significantly improve the actual effect of bad behavior correction, overcoming the industry pain point that the effect of general voice intervention is weak and difficult to correct pet bad habits. 3) It has intelligent triggering logic, which only starts voice playback when the pet behavior recognition outputs the result of bad dog or cat behavior; if it is determined to be normal pet behavior, no voice prompt is output, which will not generate unnecessary home noise to disturb pets and residents, and the management logic is more intelligent and humanized. 4) The system features an integrated hardware and software design. The voice intervention scheduling unit is natively paired with the built-in audio codec module of the edge AI computing power box. No external audio peripherals are required. The behavior recognition results can be directly transmitted to the scheduling unit within the AI ​​agent and drive the voice module. The link is simple, highly integrated, and compatible with the overall architecture of the system's local offline integrated management.

[0076] In an exemplary embodiment, the AI ​​agent also includes an automatic feeding control module 106. The automatic feeding control module 106 establishes command communication with the WiFi smart water device and the WiFi smart feeding device, respectively, collects data on the remaining water and food levels in real time, and supports differentiated timed feeding, automatic water replenishment, and graded remaining food level reminders for cats and dogs.

[0077] In the specific control logic, the system sets fixed and unmodifiable two-level residual warning thresholds: when the residual food level of the WiFi smart feeding device is ≤20%, a level three replenishment reminder is triggered; when the residual food level of the WiFi smart feeding device is ≤5%, a level two emergency replenishment reminder is triggered. The WiFi smart drinking device simultaneously pushes a water shortage reminder based on the residual data of the liquid level sensor. Both types of residual reminder messages are uniformly pushed to the user terminal APP.

[0078] Meanwhile, the automatic feeding control module supports configuring differentiated timed and quantitative feeding plans for pet cats and dogs, and sends start and stop commands to the spiral feeding mechanism of the WiFi smart feeding device; when the liquid level of the WiFi smart drinking device is lower than the safety threshold, it automatically sends a water replenishment command to start the built-in micro water pump to complete the automatic water replenishment.

[0079] The beneficial effects of this embodiment are as follows: Existing pet feeding and watering devices only have local indicator lights to show the remaining amount, without tiered remote APP reminders. Users may easily forget to replenish food and water, leading to pets going hungry and dehydrated. This module uniformly completes the remaining amount monitoring and feeding instruction issuance locally by the edge AI computing box. The threshold is fixed and stable, and it distinguishes between cat and dog feeding programs. Tiered push reminders eliminate the need for frequent manual checks of the device, achieving fully automatic diet management. It is compatible with the system's integrated offline operation architecture and can still perform timed feeding and automatic water replenishment operations normally even when the network is disconnected.

[0080] In an exemplary embodiment, the AI ​​agent also includes a lost pet location management unit 107. This unit communicates and works in conjunction with the pet's smart electronic collar to enable remote pet retrieval and management. When a user manually triggers the retrieval mode command via the user terminal APP, the management unit sends a communication switching command to the pet's smart electronic collar. The collar automatically disconnects from the local area network (WiFi) and switches to the NB-IoT cellular network, continuously uploading GPS / BeiDou dual-mode real-time location data to the cloud and the APP. Simultaneously, it supports remotely issuing audible and visual alarm commands from the user, driving the collar's built-in bright red light and speaker to provide continuous audible and visual alerts, assisting in quickly locating the lost pet.

[0081] The unique benefits of this embodiment are as follows: Existing pet collars only provide local area network positioning at home and lack cellular communication capabilities when the pet leaves home, making it impossible to track the pet if it gets lost; This unit relies on the edge terminal to issue remote control commands and automatically switches the cellular communication link. Even in offline or network-disconnected scenarios, it can still independently upload its location after receiving a retrieval command from the APP, forming a closed loop of full-scenario control for home monitoring and preventing the pet from getting lost when going out, thus improving the system's full-lifecycle pet management capabilities.

[0082] In an exemplary embodiment, the AI ​​agent also includes a lightweight YOLOv8n detection model. A detailed description is provided of a local real-time privacy filtering mechanism for non-target data based on the YOLOv8n lightweight detection model.

[0083] The YOLOv8n lightweight detection model performs local pre-screening during the camera and audio data acquisition stage, avoiding the privacy leakage risk caused by uploading all audio and video data to the cloud from the source. The specific implementation logic is as follows: (1) Real-time detection of lightweight targets The YOLOv8n lightweight detection model is used to perform target recognition on each frame of the camera, and output the target category and confidence level. The YOLOv8n lightweight detection model can achieve a single-frame inference latency of ≤10ms after INT8 quantization, with extremely low computing power overhead, and is fully compatible with the local hardware performance of edge AI computing boxes.

[0084] (2) Standardized audio and video filtering rules Image frame filtering: If the detection result of the video data contains a pet target and the confidence level is equal to or equal to the first preset confidence level value (e.g., 0.75), the complete image of the frame is retained and sent to the subsequent AI inference link; if only human body, other irrelevant animals, or pure environmental images are identified, the frame is directly deleted at the edge local regardless of the confidence level value, without caching or uploading to the cloud. Audio filtering: Only retain cat and dog barks and user voice command audio segments with a confidence level greater than or equal to the second preset confidence level value (e.g., 0.8), while other environmental noises and irrelevant human voice audio are removed locally in real time.

[0085] The beneficial effects of this exemplary embodiment: Existing pet camera devices directly upload all captured images and audio to the cloud for complete storage, posing a risk of privacy leakage of family members' faces and daily life scenes; This embodiment completes the pre-screening of data locally at the edge, and irrelevant human figures, ambient sounds, and non-target images are not transmitted to the outside throughout the process, only retaining the effective data required for pet monitoring, and meeting privacy compliance requirements with local storage logic, thus solving the industry defects of existing products in lack of privacy control and high data security risks.

[0086] In an exemplary embodiment, the system further includes an offline operation and low-latency inference optimization module. This module leverages three mechanisms—INT8 quantization compression, local encrypted storage, and breakpoint resumption—to address the issues of core function paralysis and high cloud transmission latency in existing solutions after external network interruptions. The optimization logic is as follows: (1) INT8 linear quantization compression algorithm Lossless quantization compression is performed on the trained FP32 floating-point model. The calculation formula is as follows: (12) in, These are the model's original floating-point weights and activation values; is the quantized integer parameter; S is the quantization scaling factor, and Z is the zero-point offset; This is a numerical truncation function that limits the output range to the standard INT8 value [-128, 127]. `round` is a rounding function.

[0087] After quantization, the overall size of the model can be compressed by 75%, the inference speed is increased by 3 times, and it can run smoothly on low-power edge AI computing boxes.

[0088] (2) Offline data storage and breakpoint synchronization mechanism In offline mode with the external network disconnected, all collected pet monitoring data, AI inference results, and health assessment logs are stored on edge local EMMC storage media using AES-256 (Advanced Encryption Standard 256-bit) encryption algorithm. Local storage adopts a cyclic overwrite strategy with a default storage period of 30 days. When local storage space is occupied to 80%, the oldest non-critical data is automatically overwritten to ensure uninterrupted system operation during offline periods. When the device detects external network connectivity again, it automatically synchronizes all encrypted data from the offline phase to the cloud and user APP through a breakpoint resume mechanism. After synchronization is completed, a local backup is retained, and redundant data is not uploaded repeatedly.

[0089] The beneficial effects of this exemplary embodiment: Existing pet AI solutions rely entirely on cloud servers for inference operations, which not only results in high latency due to network transmission, but also causes behavior recognition and health monitoring to stop completely when the network is disconnected; This embodiment deploys all AI inference locally at the edge, quantifies the model to adapt to the limited computing power of the device, fully retains all core functions such as automatic feeding, behavior recognition, and health monitoring in the offline state, and encrypts and stores data locally to balance offline availability and data security, thus achieving the core invention goal of integrated offline management and control.

[0090] In an exemplary embodiment, the system also includes a model iteration module. The model iteration module has a built-in model iteration mechanism with a fixed period and user-authorized authorization, which solves the technical defects of existing platforms such as forced upgrades, unauthorized use of training data, and model instability caused by frequent iterations. The complete iteration process is as follows: 1) The model iteration cycle is fixed at 3 months, and the underlying system parameters are locked, so users cannot modify the iteration cycle themselves, thus avoiding the problems of fluctuation in recognition accuracy and decrease in model stability caused by frequent updates.

[0091] 2) The model training data sources are divided into two categories: the first category is publicly available cat and dog behavior and health open source datasets, which do not involve any user's private pet data; the second category is anonymized pet historical monitoring data after the user manually and actively enables authorization, and the data authorization switch is off by default.

[0092] 3) After the incremental training is completed in the cloud and an encrypted upgrade installation package is generated, an upgrade prompt pop-up will only be pushed to the user's terminal APP. Users can choose to upgrade immediately, postpone the upgrade, or permanently turn off the upgrade push. The edge AI computing box will not perform any model update operations before receiving confirmation from the user.

[0093] The beneficial effects of this exemplary embodiment: Most existing smart pet devices force model upgrades, leaving users with no choice, and the platform lacks clear data authorization boundaries, posing compliance risks by using users' pet data for training without authorization; this embodiment ensures model stability by fixing a low-frequency iteration cycle, and the upgrade operation is entirely up to the user's discretion. The collection and use of historical pet data for training must be authorized by the user, balancing continuous model optimization, user autonomy, and data privacy compliance requirements.

[0094] Based on the aforementioned system hardware architecture, AI agent modules, and algorithm implementation, this system is adaptable to two typical scenarios: family cat and dog ownership and pet shop boarding. Three application examples are as follows: Application scenarios for family dogs: The system is adapted to the specific physiological baseline of dogs and four typical undesirable behaviors (destroying furniture, barking continuously, jumping on people, and urinating / defecating indiscriminately), and a differentiated timed feeding strategy is configured three times a day. When the pet behavior health prediction model identifies undesirable behaviors such as destroying furniture or barking at night, the edge device triggers a pre-stored owner's voice intervention within 5 seconds and pushes a reminder to the user's terminal APP. The system summarizes the dog's activity level, sleep duration, and sleep depth data weekly to generate a personalized health report and provides suggestions for optimizing exercise and rest.

[0095] Application scenarios for domestic cat owners: The system is adapted to the baseline of feline physiological indicators and optimizes the recognition model for four types of undesirable feline behaviors: scratching furniture, entering restricted areas, inappropriate urination and defecation, and stress. It adopts a feeding scheme of four small, frequent meals per day. It provides real-time voice intervention after recognizing scratching or boundary-crossing behaviors; it identifies the cat's stress state based on activity level, body posture, audio, and sleep data and pushes soothing suggestions; the weekly health report focuses on analyzing water intake and excretion-related indicators, and provides targeted urinary system protection and water optimization solutions.

[0096] In a centralized pet boarding scenario: a sensing terminal (camera and water / feeding equipment) is deployed for every 10 square meters of boarding space. Each boarded pet wears a unique smart electronic collar, and an edge AI computing box centrally manages all the equipment in the space. The system establishes an independent health baseline and personalized feeding plan for each pet, continuously monitoring pet health indicators, sleep, and fighting / stress behavior. Once disease, conflict, or stress abnormalities are identified, a tiered alarm is immediately pushed to the management personnel's APP, achieving centralized intelligent management of multiple pets.

[0097] This invention also provides a pet daily management method based on an AI agent, implemented using the pet daily management system based on an AI agent as described in any of the above embodiments, the method comprising: Step S1: Obtain pet monitoring data; Specifically, the pet monitoring system uses a high-definition night vision camera 20 to collect video and audio data, a WiFi smart drinking water device 21 to collect drinking water data, a WiFi smart feeding device 22 to collect food data, and a pet smart electronic collar 23 to collect wearable sensor data. All monitoring data is transmitted to the edge AI computing box 1 via a local area network.

[0098] Step S2: Combine pet monitoring data to achieve automatic feeding, health monitoring and behavior intervention.

[0099] Step S2 specifically includes the complete serial execution of sub-steps S21 to S28: S21 Edge Local Front-End Privacy Data Filtering: The AI ​​agent uses the YOLOv8n lightweight detection model to filter each frame of the camera footage and the captured audio in real time. Image filtering rules: Only video frames that detect cats / dogs and have a confidence level of conf≥0.75 are retained and sent to downstream inference. Images containing only people, irrelevant animals, or purely environmental elements are directly deleted locally, not cached, and not uploaded. Audio filtering rules: Only retain cat and dog barks and user voice clips with a confidence level ≥ 0.8; other environmental noises and irrelevant human voices are removed in real time. Filtered valid audio and video data, along with complete wearable sensor data, are fed into the pet behavior and health prediction model.

[0100] S22 Cat and Dog Exclusive Multimodal Fusion Inference Computing: The feature extraction module extracts visual features through the YOLOv8 backbone network and audio features through the MFCC algorithm, and normalizes the wearable sensor data to obtain vital sign sensing features. The three types of heterogeneous features are unified into a 512-dimensional common feature space by linear projection formula, and cross-modal fusion is performed by multi-head self-attention mechanism to output unified fused features. The fused features are split into two output branches of the decoder: ① Pet behavior recognition output branch operation, outputting predicted pet behavior (divided into normal behavior, canine bad behavior, and feline bad behavior); ② The pet health assessment output branch outputs physiological fusion prediction feature vector, emotion prediction feature vector, and sleep state prediction feature vector through three types of output heads; S23 three parallel data processing links operate synchronously: Link 1: Real-time intervention and control of undesirable pet behaviors: 1) The AI ​​agent's voice intervention scheduling unit receives the predicted pet behavior output by S2.2; 2) If the dog / cat's behavior is identified as bad behavior, the pre-stored owner's voice will be retrieved and the stop voice will be played in real time through the built-in audio codec module of the edge AI computing box, with a total end-to-end latency of ≤5s; if the behavior is normal, the voice playback will not be triggered. 3) Synchronously monitor the user terminal APP3 to send the retrieval command: After the user triggers the retrieval mode, the lost location control unit sends a switching command to the pet smart electronic collar 23. The device disconnects WiFi, switches to NB-IoT cellular network, and continuously uploads GPS / Beidou dual-mode positioning. The user can remotely send commands to drive the collar's bright red light and the speaker's sound and light alarm to assist in finding the pet.

[0101] Link Two: Full Process Judgment of Pet Health Triple Verification 1) Health indicator extraction module with dual-channel value acquisition: One data input channel receives the physiological fusion prediction feature vector and performs dimensional splitting and parsing on the physiological fusion prediction feature vector to extract the predicted values ​​of each independent physiological indicator; the other data input channel receives the drinking water data collected by the WiFi smart drinking device and the food data collected by the WiFi smart feeding device as the raw values ​​of the dietary indicators. 2) The health indicator quantification module calculates the abnormal score of each physiological indicator according to formula (6). When the indicator exceeds twice the standard deviation, it is judged as an abnormal indicator in the initial screening. 3) Triple verification module performs sequential verification layer by layer: ① First level: Multi-device cross-validation unit, retrieves data from related terminals in the same dimension, calculates the cross-validation score according to formula (7), if the cross-validation score is ≥0.8, proceed to the next level, otherwise filter out abnormalities; ② Second layer: Individual baseline comparison unit, take the average abnormal score for 3 consecutive periods (Formula 8), only those with an average abnormal score ≥ 0.9 and continuous abnormality for three periods are allowed, and those with short-term fluctuations are filtered out. ③ The third layer: Veterinary expert rule base verification unit, matching the "Canine and Feline Clinical Pathology" rule base, calculates the matching degree according to formula (9), and the matching degree ≥ 0.7 confirms the real health abnormality; 4) If all three verifications pass, the smart electronic collar of the pet will trigger an audio and visual alarm and push a medical reminder to the user's terminal APP; if any one verification fails, the abnormal data in this group will be discarded.

[0102] Link 3: Archive of Emotional Levels and Sleep Status Statistics 1) The emotion prediction module reads the emotion prediction feature vector, uses formula (10) to weight and fuse to calculate the comprehensive emotion score, divides the emotions into three categories: pleasant and calm, anxious and fearful, and aggressive and stressful according to the score interval, and records the score and state; 2) The sleep state prediction module reads the triaxial acceleration and heart rate sensor data, calculates the pet's exercise volume A per minute according to formula (11), and distinguishes between deep sleep, light sleep and wakefulness by combining the resting heart rate threshold and duration. 3) Summarize the total amount of exercise, the duration of deep and light sleep, and the mood level every day, and write them into the weekly health report.

[0103] The S24 automatic feeding control module is linked to execute dietary management: 1) Real-time reading of remaining water and feeding equipment data from WiFi smart drinking and feeding devices; 2) When the remaining grain storage is ≤20%, a Level 3 restocking reminder will be sent via the app; when the remaining grain storage is ≤5%, a Level 2 emergency replenishment reminder will be sent; when the drinking water level is below the threshold, an instruction will be automatically issued to start the micro water pump to replenish water. 3) Based on the different needs of cats and dogs, timed and quantitative feeding instructions are issued to drive the spiral feeding structure to complete automatic feeding; all timed feeding and automatic water replenishment functions will still be executed normally even when the network is disconnected or offline.

[0104] S25 offline encrypted storage and external network breakpoint synchronization: In offline mode with the external network disconnected, all monitoring data, AI inference results, health logs, and emotional sleep records are encrypted with AES-256 and stored on the EMMC medium of the edge AI computing box; a 30-day cyclic overwrite strategy is adopted locally, and the oldest non-critical data is automatically overwritten when the storage space reaches 80%; After the external network is restored, the encrypted data is synchronized to the cloud and the user's APP through the breakpoint resume mechanism, and the local backup is retained simultaneously, without uploading redundant data repeatedly.

[0105] S26 Weekly Health Report Automatically Generated: At fixed times each week, physiological triple verification records, daily mood data, sleep duration, food and water intake statistics, and behavior recognition records are summarized to generate a standardized weekly health report, along with optimization suggestions for feeding, exercise, drinking water, and grooming, which is then pushed to the user's terminal APP; the report serves as a reference for non-invasive veterinary diagnosis.

[0106] S27 Fixed-Period Controllable Model Iterative Management: The system has a fixed 3-month model iteration cycle, and the underlying parameters are locked and cannot be modified. The training data sources are divided into public cat and dog datasets and historical data that users actively authorize to be anonymous (authorization is turned off by default). After the encrypted upgrade package is generated in the cloud, a pop-up window is only pushed to the APP, allowing users to choose to turn off the upgrade immediately / delay / permanently. The model update will not be executed if the user does not confirm.

[0107] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0108] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A pet daily management system based on an AI intelligent agent, characterized in that, The system includes at least: an edge AI computing power box and a pet intelligent sensing and execution terminal that is communicatively connected to the edge AI computing power box; The pet intelligent sensing and execution terminal includes at least a high-definition night vision camera, a WiFi intelligent drinking device, a WiFi intelligent feeding device, and a pet intelligent electronic collar; The edge AI computing power box has a built-in AI agent; the AI ​​agent is used to realize automatic feeding, health monitoring and behavior intervention based on the pet monitoring data collected by the pet intelligent perception execution terminal.

2. The pet daily management system based on AI intelligent agent according to claim 1, characterized in that, The pet monitoring data includes at least video and audio data collected by the high-definition night vision camera and pet wearable sensor data collected by the pet smart electronic collar. The AI ​​agent includes at least a pet behavior and health prediction model, a health indicator extraction module, a health indicator quantification module, and a triple verification module. The pet behavior and health prediction model is used to extract and fuse features from the video data, audio data, and pet wearable sensor data to obtain unified fused features, and output pet predicted behavior and health prediction data based on the unified fused features. The health prediction data includes at least a physiological fusion prediction feature vector; The health indicator extraction module includes two data input channels. One data input channel receives the physiological fusion prediction feature vector and performs dimensional splitting and parsing on the physiological fusion prediction feature vector to extract the predicted values ​​of each independent physiological indicator. The other data input channel receives drinking water data collected by the WiFi smart drinking device and food data collected by the WiFi smart feeding device as the raw values ​​of the dietary indicators. The health indicator quantification module is used to calculate the abnormal score of each physiological indicator by combining the predicted or original value of each physiological indicator with the historical baseline mean and standard deviation of the health indicators of the corresponding individual pet, and to obtain the initial screening abnormal indicators and their corresponding abnormal scores. The triple verification module is used to perform triple verification on the initial screening abnormal indicators and their corresponding abnormal scores in sequence to obtain a valid health abnormality judgment result.

3. The pet daily management system based on AI intelligent agents according to claim 2, characterized in that, The pet behavior health prediction model includes an encoder and a decoder, wherein the decoder is a Transformer decoder; the encoder includes a feature extraction module and a feature fusion module. The decoder includes a pet behavior recognition output branch and a pet health assessment output branch; The feature extraction module is used to extract visual features from the video data through the YOLOv8 backbone network, extract audio features from the audio data through the MFCC algorithm, and normalize the pet's wearable sensor data to obtain vital sign sensor features. The feature fusion module is used to uniformly map the visual features, audio features, and vital sign sensing features to the same feature dimension through linear projection, and complete cross-modal fusion through a multi-head self-attention mechanism to output the unified fused features. The pet behavior recognition output branch is used to process the unified fusion features to obtain the predicted pet behavior; The predicted pet behaviors include normal pet behaviors, undesirable canine behaviors, and undesirable feline behaviors. The pet health assessment output branch is used to process the unified fusion features to obtain the health prediction data.

4. The pet daily management system based on AI intelligent agent according to claim 2, characterized in that, The triple verification module includes a multi-device cross-verification unit, an individual baseline comparison unit, and a veterinary expert rule base verification unit.

5. The pet daily management system based on AI intelligent agents according to claim 4, characterized in that, The multi-device cross-validation unit is used to retrieve synchronously collected data from different types of associated devices for the initial screening anomaly indicators and calculate the cross-validation score between multiple devices. When the cross-validation score is greater than or equal to the first preset value, the cross-validation is deemed to have passed and proceeds to the next level of validation; otherwise, it is determined to be a single-device sensor error and the corresponding initial screening abnormal indicator is directly filtered out.

6. The pet daily management system based on AI intelligent agent according to claim 4, characterized in that, The individual baseline comparison unit is used to continuously collect the abnormal scores of abnormal indicators that have been cross-validated by multiple devices for a preset number of detection cycles and calculate the average value. When the average value is greater than or equal to the second preset value and the indicator is determined to be abnormal for the preset number of consecutive detection cycles, the time-series baseline comparison verification is passed and the third verification is entered; otherwise, it is determined to be a short-term data fluctuation and the corresponding abnormal indicator is filtered out.

7. The pet daily management system based on AI intelligent agent according to claim 4, characterized in that, The system also includes a user terminal APP; The veterinary expert rule base verification unit is used to match abnormal indicators that have passed the second verification with a preset health assessment expert rule base to calculate the rule matching degree. When the matching degree of the rule is greater than or equal to the third preset value, a health abnormality is confirmed, the valid health abnormality judgment result is obtained, and the pet smart electronic collar is triggered to issue an audible and visual alarm or the user terminal APP is triggered to push medical information; when the matching degree is less than the third preset value, it is determined to be a false positive and the corresponding abnormal indicators are filtered out.

8. The pet daily management system based on AI intelligent agent according to claim 3, characterized in that, The health prediction data also includes emotion prediction feature vectors and sleep state prediction feature vectors; The pet health assessment output branch includes three types of output heads, which respectively output the physiological fusion prediction feature vector, the emotion prediction feature vector, and the sleep state prediction feature vector based on the unified fusion feature. The AI ​​agent also includes an emotion prediction module and a sleep state prediction module; The emotion prediction module is used to perform weighted fusion calculation on four types of features in the emotion prediction feature vector: activity level, body posture, pet barking, and vital signs, to obtain a comprehensive emotional score for the pet. Based on the comprehensive emotional score, the pet is classified into three emotional states: pleasant and calm, anxious and fearful, and aggressive and stressed. The emotional state corresponding to the comprehensive emotional score is written into the health report. The sleep state prediction module is used to calculate the pet's exercise volume based on the triaxial acceleration data and heart rate data in the sleep state prediction feature vector, and to classify the sleep state into three types: deep sleep, light sleep, and wakefulness, combined with the resting heart rate interval threshold, and to write the pet's exercise volume and sleep state into the health report.

9. The pet daily management system based on AI intelligent agent according to claim 3, characterized in that, The edge AI computing power box also includes an audio encoding and decoding module, and the AI ​​intelligent agent also includes a voice intervention scheduling unit; The voice intervention scheduling unit receives the predicted pet behavior. When the predicted pet behavior is a bad dog behavior or a bad cat behavior, it retrieves the pre-recorded owner's voice prompt and plays it in real time through the audio codec module to complete the pet behavior intervention.

10. A pet daily management method based on AI intelligent agents, characterized in that, The pet daily management system based on an AI agent, as described in any one of claims 1 to 9, is implemented using the method comprising: Obtain pet monitoring data; By combining the pet monitoring data, automatic feeding, health monitoring, and behavioral intervention can be achieved.

Images