Intelligent management and human-computer interaction system for pet chronic diseases

Abstract

The present application relates to the field of pet health monitoring, and discloses a pet chronic disease intelligent management and man-machine interaction system, comprising an Internet of Things hardware sensing layer, an edge cloud computing layer and a user interaction terminal; the hardware sensing layer switches the collection mode according to the parameter change to generate event level data packets; the edge cloud computing layer extracts an asynchronous feature vector to input an individual baseline model, outputs a risk deviation index and locates a dominant abnormal feature, and generates a hardware scheduling control frame downward; the user interaction terminal reconstructs a graphical interface according to the index, renders a modal view when a blocking threshold is reached, and collects a structured label through the modal view. Through event triggering and dynamic sampling of hardware nodes, environmental interference is filtered out, the present application accurately quantifies the small metabolic deviation of pets in the incubation period of chronic diseases by using a special individual baseline model, and a bidirectional intelligent intervention closed loop of downward scheduling sensors and upward forced user interaction interception is constructed.

Description

Technical Field

[0001] This invention relates to the field of pet health monitoring technology, specifically to a pet chronic disease intelligent management and human-computer interaction system. Background Technology

[0002] With the increasing number of pets, early intervention for chronic pet diseases such as urinary tract diseases and metabolic disorders has gradually become a focus of pet health management. Current intelligent pet health management technologies have undergone several stages of development in data acquisition and analysis, but still have many limitations.

[0003] Early systems focused on the continuous collection of basic physiological data. Patent document US20150130617 discloses an animal behavior and health monitoring system that continuously collects movement behavior and physiological parameters through wearable sensors, but it lacks intelligent analysis models and human-computer interaction intervention mechanisms for chronic diseases. Later, the industry introduced more in-depth physiological indicator monitoring and cloud computing architectures. For example, patent document US20250241276 discloses a system integrating wearable biosensors and machine learning algorithms, which monitors pet blood sugar, blood pressure, and other data to identify early signs of chronic diseases and provides a health dashboard and natural language interaction functions. Patent document CN119252477A discloses an intelligent system using a hybrid edge computing and cloud computing architecture, which uses wearable devices to process physiological data such as heart rate and body temperature, performs trend analysis in the cloud, and provides multi-terminal collaborative interaction. However, these solutions are highly dependent on hardware devices worn on the pet. In practical applications, wearable devices not only easily cause stress reactions in pets, but also pose problems such as high maintenance costs and data susceptibility to environmental interference in home environments where multiple pets live together.

[0004] To reduce reliance on wearable devices, some existing technologies are shifting towards using smart home devices and fixed behavioral pattern libraries for health assessment. Patent document CN115104548B discloses a pet behavior adjustment method based on multimedia information technology. This system has a built-in chronic disease behavior atlas database and uses environmental stimulus intervention and human-computer two-way interaction by matching abnormal behaviors. Patent document US11894144 discloses an animal health decision support system that integrates clinical, behavioral, and environmental data for trend prediction and allows users to upload multimodal data for conversational queries. While these technologies introduce disease knowledge bases and decision support functions, their judgment logic generally relies on fixed group standard behavioral atlases. Because different pets vary significantly in breed, physical condition, and daily routines, using fixed prior atlases as standards often fails to effectively capture subtle behavioral and metabolic deviations in specific individuals during the incubation period of chronic diseases, easily leading to missed early minor abnormalities.

[0005] At the system early warning mechanism and interaction level, patent document CN113593702A discloses a pet health early warning and diagnosis system. This system constructs a risk assessment model using devices such as smart feeders and triggers tiered alerts via application push notifications or SMS messages based on risk levels. However, current system architectures generally suffer from one-way data collection and shallow interaction. At the underlying perception level, each hardware node operates independently, lacking the ability to dynamically adjust the sampling parameters of associated sensors by issuing reverse instructions when suspected abnormal data is detected, resulting in a waste of system computing and power resources. At the interaction level, routine message push notifications are easily ignored by users. Existing systems lack mandatory intervention mechanisms that escalate with risk levels, failing to forcibly intercept routine user operations when the pet is in a high-risk deviation state, thus failing to guide users to supplement necessary structured human feedback. This makes it difficult for the algorithm model to obtain effective correction labels and prevents the formation of a complete hardware scheduling and human-computer interaction closed loop.

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent management and human-computer interaction system for pet chronic diseases. It solves the problems of low behavior recognition accuracy caused by fragmented data from a single sensor, inability to detect early minor abnormal deviations in individual pet chronic diseases due to reliance on fixed graphs and warning thresholds, and the lack of dynamic sampling scheduling and effective human-computer interaction intervention loop at the hardware layer in existing technologies. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent management and human-computer interaction system for pet chronic diseases. It solves the problems of low accuracy in behavior recognition due to fragmented data from a single sensor, inability to detect early minor abnormalities in individual pet chronic diseases due to reliance on fixed warning thresholds, and the lack of dynamic sampling scheduling and effective human-computer interaction intervention loop at the hardware layer in existing technologies.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: a pet chronic disease intelligent management and human-computer interaction system, including an Internet of Things hardware sensing layer, which includes heterogeneous sensing nodes. The heterogeneous sensing nodes switch to event acquisition mode according to environmental trigger parameters or weight changes, acquire continuous weight signals and identification information, and generate event-level data packets with timestamps. The edge cloud computing layer is used to receive the event-level data packets, generate an asynchronous behavior feature matrix, extract feature vectors from the asynchronous behavior feature matrix, input them into the individual baseline model, and output a continuous chronic risk deviation index. The contribution of each feature dimension in the feature vector to the continuous chronic risk deviation index is calculated, and the feature dimension with the highest contribution and reaching or exceeding the preset contribution threshold is determined as the dominant abnormal feature. A hardware scheduling control frame is generated and sent to the IoT hardware perception layer. The user interaction terminal is used to reconstruct the graphical user interface based on the continuous chronic risk deviation index, and render a modal view when the continuous chronic risk deviation index reaches a preset blocking threshold, and collect structured tags through the modal view.

[0009] Furthermore, the heterogeneous sensing nodes include a smart water dispenser node, a smart feeder node, and a smart litter box node; each heterogeneous sensing node embeds a weighing sensor, a microwave radar, and an RFID reader; in low-power polling mode, when the intensity of the reflected signal received by the microwave radar reaches a preset environmental trigger parameter threshold or the weight change detected by the weighing sensor exceeds a preset weight jump threshold, the heterogeneous sensing node switches to the event acquisition mode and modifies the sampling frequency value of the weighing sensor; the heterogeneous sensing node calculates the absolute mass change value based on the continuous weight signal; the smart litter box node, based on the absolute mass change value and the activity dwell time extracted by the microwave radar, segments a single monitoring action into urination and defecation events; the smart water dispenser node generates a drinking event based on the absolute mass change value; the smart feeder node generates a feeding event based on the absolute mass change value; the local control unit of the heterogeneous sensing node encapsulates the corresponding event type, absolute mass change value, and timestamp into an event-level data packet.

[0010] Furthermore, the edge cloud computing layer combines the identity recognition success factor determined by the identity recognition information, the weighing signal variance mapping output value of the continuous weight signal, and the event classification probability value extracted from the event-level data packet to perform a weighted summation to obtain the confidence score of the event-level data packet; the edge cloud computing layer compares the confidence score with a confidence threshold, and removes the corresponding event-level data packet when the confidence score is lower than the confidence threshold.

[0011] Further, the asynchronous behavior feature matrix includes a hydration-to-excretion ratio component and an intake-to-consumption ratio component; the edge cloud computing layer extracts the absolute mass change value corresponding to the drinking event, and calculates the comprehensive water intake by combining the absolute mass change value corresponding to the eating event with a preset food moisture content parameter; it extracts the absolute mass change value corresponding to the urination event to calculate the total urine volume, and uses the ratio of the comprehensive water intake to the total urine volume as the hydration-to-excretion ratio component; the edge cloud computing layer calculates the total food intake calorie equivalent based on the absolute mass change value corresponding to the eating event and a preset unit mass calorie conversion parameter; it determines the activity intensity parameter of each activity event based on the change in reflected signal intensity collected by the microwave radar, determines the activity duration based on the start and end timestamps of each activity event, and accumulates the product of the activity intensity parameter and the activity duration of each activity event to obtain the activity integral; the ratio of the total food intake calorie equivalent to the activity integral is used as the intake-to-consumption ratio component.

[0012] Furthermore, the edge cloud computing layer extracts the asynchronous behavior feature matrix within the historical monitoring period to construct a historical normal sample set; the edge cloud computing layer applies a single-class support vector machine algorithm with radial basis kernel function, inputs the historical normal sample set to establish the decision boundary of the normal behavior pattern of the target pet individual, and generates the individual baseline model.

[0013] Furthermore, the edge cloud computing layer inputs the acquired feature vector into the individual baseline model, and truncates and calculates the spatial deviation distance of the feature vector relative to the hyperplane of the decision boundary in the feature space; when the spatial deviation distance is less than or equal to the boundary margin threshold, the continuous chronic risk deviation index is determined to be zero; when the spatial deviation distance is greater than the boundary margin threshold, the edge cloud computing layer applies a nonlinear smoothing mapping function to convert the distance value of the excess part into the continuous chronic risk deviation index in a percentage range.

[0014] Furthermore, the edge cloud computing layer replaces each feature dimension in the feature vector with a benchmark value corresponding to the mean center of the historical normal sample set to construct alternative feature vectors corresponding to each feature dimension, and calculates the absolute value of the difference between the output of the model decision function before and after the replacement, as the basic value of the contribution of the corresponding feature dimension; when the sum of the basic values ​​of the contribution of all feature dimensions is greater than zero, the normalized contribution of the corresponding feature dimension is determined according to the proportion of the basic value of the contribution of each feature dimension to the sum of the basic values ​​of the contribution of all feature dimensions, and the feature dimension with the highest normalized contribution that reaches or exceeds the preset contribution threshold is determined as the dominant abnormal feature; when the sum of the basic values ​​of the contribution of all feature dimensions is equal to zero, the current attribution result is marked as no clear dominant abnormal feature; the edge cloud computing layer uses a preset rule engine to establish a rule mapping relationship between the dominant abnormal feature and the physiological metabolic imbalance type.

[0015] Furthermore, the hardware scheduling control frame includes target node identifier, target sensor type, and sampling mode fields; when the dominant anomaly characteristic points to the hydration-to-excretion ratio component, the edge cloud computing layer sends the hardware scheduling control frame to the smart water dispenser node and the smart litter box node to switch to the event-triggered high-frequency sampling mode; when the dominant anomaly characteristic points to the intake-to-consumption ratio component, the edge cloud computing layer sends the hardware scheduling control frame to the smart feeder node to adjust the sampling frequency value of the corresponding weighing sensor.

[0016] Furthermore, the user interaction terminal is configured with a warning threshold and a blocking threshold, and the blocking threshold is greater than the warning threshold; when the continuous chronic risk deviation index is lower than the warning threshold, the graphical user interface of the user interaction terminal maintains a preset conventional layout structure; when the continuous chronic risk deviation index is greater than or equal to the warning threshold and less than the blocking threshold, the user interaction terminal renders the data display component associated with the dominant abnormal feature at the top; when the continuous chronic risk deviation index is greater than or equal to the blocking threshold, the user interaction terminal renders the modal view covering the entire screen to intercept application layer touch events, and closes the modal view after collecting the structured tag to enter a silent cooling period.

[0017] Further, the edge cloud computing layer locates the original time node that generates the continuous chronic risk deviation index, extracts the asynchronous behavior feature matrix stored within the time interval corresponding to the original time node, and generates labeled incremental samples that are stored in the calibration sample pool. When the number of incremental samples in the calibration sample pool reaches the update trigger threshold, the edge cloud computing layer expands the incremental samples labeled as normal physiological fluctuations into the historical normal sample set. The edge cloud computing layer calculates the time weight coefficient of the historical normal sample set, removes samples whose time weight coefficient is lower than the retention threshold, and then recalculates the support vector set and bias parameters in the individual baseline model.

[0018] Furthermore, the IoT hardware sensing layer is configured with a multi-target concurrent conflict handling mechanism; when the RFID reader reads multiple RFID tags within a preset time window, the heterogeneous sensing node extracts the RFID signal reception strength indication value corresponding to each RFID tag and generates a candidate identity sequence in descending order of the RFID signal reception strength indication value.

[0019] The preset baseline weight corresponding to each RFID tag is determined based on the most recent preset number of valid weighing results for the pet corresponding to the RFID tag, which are unique in identity, have stable weighing signals, and have an event confidence level that reaches a confidence threshold. Preferably, the median of the valid weighing results is used as the preset baseline weight.

[0020] For heterogeneous sensing nodes capable of acquiring the weight carried by a pet, the heterogeneous sensing nodes sequentially match and verify the pet's resting weight value determined by the continuous weight signal with the preset baseline weight corresponding to each candidate RFID tag, according to the candidate identity sequence. The pet's resting weight value is determined based on the difference between the stable weight during the target's stay and the initial stable weight before the target enters.

[0021] The matching error is the absolute difference between the pet's resting weight and the preset benchmark weight; the weight tolerance threshold is determined based on the maximum measurement error of the weighing sensor and the maximum allowable fluctuation of the pet's historical normal weight, preferably the larger of the two.

[0022] For heterogeneous sensing nodes that cannot obtain the weight carried by the pet, the body contour features extracted by the microwave radar can be used for auxiliary verification.

[0023] When the matching error is not greater than the weight tolerance threshold, the corresponding candidate RFID tag is determined as the primary identity tag and the target pet individual is locked; when the current candidate RFID tag does not meet the matching verification condition, the next candidate RFID tag in the candidate identity sequence is verified; when none of the candidate RFID tags meet the matching verification condition, the current event is marked as a multi-target identity recognition uncertain event, and the corresponding event-level data packet is removed or downweighted.

[0024] Furthermore, the smart litter box node and the smart feeder node have a built-in zero-point drift compensation mechanism for the weighing sensor; when the heterogeneous sensing node is in the low-power polling mode for more than the reference sleep time and the microwave radar does not detect an active target, it automatically triggers a zero-point calibration command, collects the basic weight data under the current no-load state, and applies a moving average filtering algorithm to update the zero-point reference value of the weighing sensor.

[0025] Furthermore, the heterogeneous sensing node supports adaptive dynamic adjustment of the environmental trigger parameter threshold; the heterogeneous sensing node records the frequency of invalid wake-up events within a historical monitoring period, and when the frequency of invalid wake-up events within a specific time period exceeds a preset false detection tolerance threshold, it automatically increases the environmental trigger parameter threshold corresponding to that time period, or reduces the detection distance parameter of the microwave radar.

[0026] Furthermore, the edge cloud computing layer is configured with a trend prediction module; the trend prediction module extracts the continuous chronic risk deviation index from the historical continuous time series, applies a long short-term memory network to construct an index evolution prediction model, and outputs the predicted deviation index within a future preset time window; when the growth slope of the predicted deviation index is greater than a preset deterioration rate threshold but has not yet reached the warning threshold, the edge cloud computing layer sends a pre-notification instruction to the user interaction terminal.

[0027] Furthermore, a network outage disaster recovery synchronization mechanism is established between the IoT hardware sensing layer and the edge cloud computing layer; when the heterogeneous sensing node detects a WAN connection interruption, it writes the generated event-level data packets into a local non-volatile memory to maintain the timestamp sequence; when the network connection is restored, the heterogeneous sensing node performs breakpoint resume transmission to the edge cloud computing layer in the form of batch compressed data frames according to the timestamp order.

[0028] Furthermore, when rendering the modal view covering the entire screen, the user interaction terminal simultaneously invokes a structured guidance script library matching the dominant abnormal features. The user interaction terminal generates a cascading questionnaire containing preset options based on the physiological metabolic imbalance type. The next level of questions in the cascading questionnaire is loaded based on the selection results of the previous level question. The user interaction terminal encapsulates the collected questionnaire options and writes them into the structured tags. The structured tags include at least one of the following: abnormal observation tags, dietary change tags, excretion change tags, medication status tags, medical confirmation tags, environmental stress tags, hardware failure tags, and normal physiological fluctuation tags.

[0029] The intelligent management and human-computer interaction system for pet chronic diseases provided by this invention has the following beneficial effects: This invention integrates microwave radar, weighing sensors, and RFID readers at the IoT hardware sensing layer, fusing continuous weight data with physical activity parameters. The system can accurately segment a pet's continuous behavior into independent events such as defecation, urination, drinking, and eating using calculated absolute mass changes and activity dwell time, and dynamically adjust the sensor sampling frequency upon identifying a specific event. This mechanism effectively removes invalid data from single sensors due to environmental interference, improves the accuracy of low-level behavioral feature extraction, and reduces the conventional power consumption of edge nodes.

[0030] This invention applies a single-class support vector machine algorithm with radial basis function kernel function to the edge cloud computing layer. It utilizes the pet's historical normal behavior data to establish decision boundaries and generate a unique individual baseline model. The system inputs the feature vector of the current behavior into this individual baseline model, truncates the deviation distance in the computational space, and transforms it into a continuous chronic risk deviation index. This calculation logic avoids the individual difference errors introduced by using fixed physiological parameter thresholds, and can more objectively quantify the degree of subtle metabolic imbalances and behavioral deviations that occur in specific pets during the incubation period of chronic diseases.

[0031] This invention constructs a two-way closed-loop control mechanism based on dominant anomaly features. When a specific anomaly feature is located, the system generates a hardware scheduling control frame to selectively increase the sampling frequency of sensors associated with the anomaly data source to obtain high-precision data; it also triggers a tiered response mechanism at the user interaction terminal, forcibly intercepting user operations through a full-screen modal view when the risk index reaches the blocking threshold. This design achieves on-demand allocation of computing power and acquisition resources at the perception layer, and ensures that structured external human feedback tags can be obtained in a timely manner for subsequent model correction when the pet is in a high-risk state. Attached Figure Description

[0032] Figure 1 This is a diagram illustrating the overall architecture of the intelligent management and human-computer interaction system for pet chronic diseases, as described in this invention. Figure 2 This is a flowchart of the intelligent management and human-computer interaction method for pet chronic diseases according to an embodiment of the present invention; Figure 3 This is a flowchart of the heterogeneous sensing node event acquisition and structured segmentation method according to an embodiment of the present invention; Figure 4 This is a flowchart of the individual baseline modeling and continuous chronic risk deviation index inference method according to an embodiment of the present invention; Figure 5 This is a flowchart of the dominant anomaly characteristic attribution and hardware reverse scheduling intervention method according to an embodiment of the present invention; Figure 6 This is a flowchart of the model delay self-correction update method based on structured feedback labels according to an embodiment of the present invention; Figure 7 This is a graph showing the evolution of spatial deviation distance of the individual baseline model in an embodiment of the present invention. Figure 8 This is a tiered intervention chart based on the continuous chronic risk deviation index, as described in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] like Figure 1 As shown, the present invention provides a pet chronic disease intelligent management and human-computer interaction system, which may include: an Internet of Things hardware sensing layer, an edge cloud computing layer, and a user interaction terminal.

[0035] The IoT hardware sensing layer comprises multiple heterogeneous sensing nodes. These nodes include smart water dispenser nodes, smart feeder nodes, smart litter box nodes, activity detection nodes, and identification nodes. Each heterogeneous sensing node is equipped with a weighing sensor, proximity sensor, microwave radar, RFID reader, communication module, and local microcontroller unit.

[0036] Among them, the weighing sensor is used to collect continuous weight signals during drinking, eating, and excretion; the proximity sensor is used to detect the status of the pet's approach point; the microwave radar is used to collect the pet's reflected signal intensity, activity dwell time, and body shape features; the RFID reader is used to read the RFID tag carried by the pet; the communication module is used to upload event-level data packets to the edge cloud computing layer; and the local microcontroller unit is used to perform sampling mode switching, weight signal preprocessing, event type identification, and data encapsulation.

[0037] The edge cloud computing layer establishes a data communication connection with the IoT hardware sensing layer to receive and process asynchronous discrete data from various heterogeneous sensing nodes. The edge cloud computing layer performs event-level data cleaning, asynchronous behavioral feature matrix construction, individual behavior baseline modeling, continuous chronic risk deviation index calculation, abnormal feature contribution extraction, hardware scheduling control frame generation, and model self-correction update.

[0038] The user interface terminal establishes a data communication connection with the edge cloud computing layer to receive continuous chronic risk deviation index, dominant abnormal characteristics, and hardware status information output by the edge cloud computing layer. Based on the received data parameters, the user interface terminal performs graphical user interface reconstruction, health risk alerts, electronic health record export, and structured feedback collection.

[0039] like Figure 2 As shown, this embodiment of the invention provides a method for intelligent management and human-computer interaction of pet chronic diseases, including the following steps: S10, each heterogeneous sensing node in the IoT hardware sensing layer switches to event acquisition mode according to environmental trigger parameters, extracts and cleans the weight signal of the event interval to generate event-level data packets and uploads them. The edge cloud computing layer receives the event-level data packets and calculates the confidence score of a single event by combining the identity recognition information collected by the heterogeneous sensing nodes. The edge cloud computing layer removes or downweights event data packets with confidence scores lower than the confidence threshold. S20, the edge cloud computing layer uses a sliding time window and update step size to perform spatiotemporal alignment and aggregation of asynchronous discrete events, and constructs an asynchronous behavior feature matrix that includes at least the hydration-to-excretion ratio component, the intake-to-consumption ratio component, the diurnal rhythm offset component, and the event comprehensive confidence component; in an extended implementation, the asynchronous behavior feature matrix may also include the digestion-to-excretion ratio component. S30, the edge cloud computing layer inputs the feature vector extracted by the current sliding time window into the individual baseline model trained based on the single-class support vector machine algorithm, and calculates the continuous chronic risk deviation index in combination with the boundary margin threshold. It also measures the contribution of each physiological semantic feature component involved in abnormal attribution to the deviation value of the continuous chronic risk deviation index in order to extract the dominant abnormal features. S40, the edge cloud computing layer generates a hardware scheduling control frame carrying sampling scheduling parameters and optional hardware intervention parameters based on the dominant anomaly characteristics, and sends it to the IoT hardware perception layer to dynamically modify the sampling mode, sampling frequency, wake-up condition, duration and sleep recovery condition of the heterogeneous perception nodes associated with the dominant anomaly characteristics, and further triggers the physical inducement or environmental adjustment action of the corresponding node when the preset intervention conditions are met. S50, the user interaction terminal performs hierarchical reconstruction of the graphical user interface based on the continuous chronic risk deviation index, and renders a full-screen intervention view when the blocking threshold is reached, so as to collect at least one of the following interactive data: user informed confirmation, electronic health record export operation or structured feedback form. S60, the edge cloud computing layer extracts feedback tags based on structured feedback forms or follow-up forms of subsequent early warning events, and performs a delayed self-correction operation with tag triage logic on the individual baseline model in combination with the unique identifier of the early warning event; if no valid feedback tags are obtained, only the informed confirmation log or electronic health record export record is archived, and the individual baseline model self-correction update is not triggered.

[0040] See Figure 3 The IoT hardware sensing layer of this invention is responsible for performing asynchronous discrete behavior data collection and cleaning operations in a multi-pet concurrent environment. The specific processing flow is as follows: S110, each heterogeneous sensing node in the IoT hardware sensing layer performs dynamic state switching between low-power polling mode and event acquisition mode, and acquires continuous weight signals before and after the event interval in event acquisition mode.

[0041] In this embodiment, the smart water dispenser node, smart feeder node, and smart litter box node maintain low-power polling mode during periods when no target enters the sensing range. In low-power polling mode, the weighing sensors of each heterogeneous sensing node are in low-frequency sampling or low-power comparator listening mode, the RFID reader is in low-frequency wake-up mode, and the microwave radar and proximity sensor send detection signals at preset time intervals. When the intensity of the reflected signal received by the microwave radar or proximity sensor reaches a preset environmental trigger parameter threshold, or when the weight change detected by the weighing sensor in low-frequency sampling or low-power comparator listening mode exceeds a preset weight jump threshold, the heterogeneous sensing node switches from low-power polling mode to event acquisition mode. Preferably, the aforementioned environmental trigger parameter threshold and weight jump threshold are determined by environmental background noise sampling calibration under no-load conditions before the device leaves the factory to filter out false triggers caused by normal environmental vibrations.

[0042] Upon entering event acquisition mode, the heterogeneous sensing node increases the sampling frequency of its weighing sensors to a preset high-frequency operating level to acquire continuous weight signals, including the initial stable weight before the target enters, the dynamic fluctuation weight during the target's stay, and the final stable weight after the target leaves. To ensure a closed-loop physical implementation of low-power characteristics, when the weight fluctuation amplitude detected by the weighing sensors of the heterogeneous sensing node within a preset time period is less than the stationary judgment threshold, and the microwave radar does not detect any target activity characteristics, the single event is determined to be over, and the heterogeneous sensing node exits event acquisition mode and returns to low-power polling mode.

[0043] S120, the heterogeneous sensing node performs filtering and smoothing processing on the continuous weight signal, calculates the absolute mass change value, and performs structured segmentation for complex events.

[0044] To eliminate high-frequency noise from mechanical vibrations caused by pets stepping on mechanical parts or water surface sloshing, heterogeneous sensing nodes extract the data sequence of dynamically fluctuating weights and smooth it using median filtering or Kalman filtering algorithms. For the specific calculation process of the digital filtering algorithm, those skilled in the art can select a conventional discrete filtering function based on the hardware computing power of the microcontroller unit; the implementation details of the filtering algorithm are well-known in the field and will not be elaborated here. The heterogeneous sensing nodes calculate the absolute mass change value corresponding to a single event based on the difference between the final stable weight and the initial stable weight. For smart water dispenser nodes and smart feeder nodes, the absolute mass change value directly corresponds to the mass of a single drinking session and the mass of a single feeding session.

[0045] For the intelligent litter box node, the local microcontroller unit needs to perform structured segmentation of excretion events by combining the initial stable weight, the final stable weight, the event duration, the weight change amplitude, and radar action characteristics. Specifically, the intelligent litter box node calculates the positive difference between the final stable weight and the initial stable weight as the excrement mass. Combining the target's activity dwell time parameter extracted by microwave radar within the chamber, the local microcontroller unit classifies excretion events into urination events, defecation events, and mixed events. In terms of specific classification logic, if the absolute mass change of excrement mass is within a preset weight range and the activity dwell time is less than a preset duration threshold, it is determined as a urination event; if the absolute mass change is greater than the upper limit of the preset weight range, or the activity dwell time is greater than or equal to the duration threshold, it is determined as a defecation event; if two or more independent positive weight jumps are extracted within a single event interval and the radar signal is not interrupted, it is determined as a mixed event. If the excrement mass shows a negative jump, the local microcontroller unit marks the corresponding event as a litter cleaning event. If a positive weight jump is detected and no radio frequency identification information of the target pet is entered, the local microcontroller unit will mark it as a sand-adding event or an abnormal interference event.

[0046] S130, heterogeneous sensing nodes collect the identification information of individual target pets to determine the pet's identity and monitor the overlapping state of multiple targets.

[0047] While collecting data, the heterogeneous sensing nodes acquire the identification information of individual target pets to determine their identities. This identification information includes at least one of the following: RFID tag identifiers read by the RFID reader, the pet's resting weight, and body contour features extracted by microwave radar. Considering data interference in multi-pet environments, when the RFID reader parses more than two independent RFID tags within a single event cycle, the heterogeneous sensing nodes initiate a multi-target concurrent conflict handling mechanism. If a single target pet is successfully identified through candidate identity sorting and matching verification, the identification result is marked as successful. If none of the candidate identities pass matching verification, or if the variance of the continuous weight signal exceeds a preset multi-target aliasing threshold and target separation cannot be achieved, the multi-target aliasing state is recorded.

[0048] S140, the edge cloud computing layer receives event-level data packets with timestamps, calculates confidence scores based on multi-source information, and performs data removal and weight reduction cleaning.

[0049] After receiving event-level data packets, the edge cloud computing layer calculates a confidence score to quantify the data reliability of a single event, combining the identity recognition results, weighing signal variance, and event classification labels. The formula for calculating the confidence score is: ; in, Indicates the confidence score; This indicates a successful identification factor, which occurs when a single target is directly identified, or when a single individual pet is successfully identified through a multi-target concurrent conflict handling mechanism. The value is 1; when no target is identified or multi-target concurrent conflict handling fails, The value is 0; This represents the variance of the weighing signal; This represents a mapping function that is negatively correlated with the signal variance. It is used to map the weighing signal variance to a preset numerical range, so that the larger the variance value, the smaller the corresponding mapping output value, thereby quantifying the confidence decay caused by data jitter. This represents the event classification probability value, which is obtained by the local microcontroller unit by inputting the duration of the current event and the absolute quality change value into a preset historical statistical distribution model (such as a Gaussian mixture model or a two-dimensional joint histogram statistical model) to calculate the matching probability. , and These represent the corresponding preset weighting coefficients, and their sum is 1. The preset weighting coefficients can be pre-set by technicians according to the data confidence priority of different sensors.

[0050] The edge cloud computing layer compares the calculated confidence score with a preset confidence threshold. Data packets marked as sand-cleaning events, sand-adding events, or abnormal interference events are directly discarded by the edge cloud computing layer. For event data packets with confidence scores below the confidence threshold, the edge cloud computing layer performs numerical weighting processing during subsequent feature construction to prevent invalid data caused by multi-target aliasing from contaminating the historical behavioral baseline of individual pets.

[0051] The edge cloud computing layer receives low-level data transmitted from the IoT hardware sensing layer. Because the event-level data packets reported by various heterogeneous sensing nodes are asynchronous and discrete in terms of time triggering, and the dimensions of different physical quantities (such as weight, frequency, and duration) differ significantly, the edge cloud computing layer needs to perform spatiotemporal alignment and aggregation operations on event-level data packets with varying timestamps and physical dimensions to construct an asynchronous behavioral feature matrix with a unified dimension. The specific processing is as follows: S210, the edge cloud computing layer sets a fixed-length sliding time window and update step size, and performs time domain segmentation and alignment operations on asynchronous discrete events within the current time window.

[0052] In this embodiment, considering that the physiological metabolic patterns of individual pets mostly follow a natural day cycle, as a preferred approach, the edge cloud computing layer sets the sliding time window length to 24 hours and the update step size to a fixed value within the range of 1 to 4 hours. Based on the timestamps carried in the event-level data packets, the edge cloud computing layer extracts all drinking, eating, urinating, defecating, and activity events belonging to the same target pet and falling within the current sliding time window into a temporary queue. For persistent events that cross the boundaries of the sliding time window, the edge cloud computing layer extracts the duration of the event falling within the current sliding time window and calculates its proportion of the total duration of a single event. Then, based on this time proportion, it linearly allocates the absolute quality change value of the event to ensure the accuracy of data aggregation at the time boundaries.

[0053] S220, the edge cloud computing layer, calculates and extracts the hydration-to-excretion ratio and the intake-to-consumption ratio for event data in the temporary queue, thereby transforming discrete behavioral events into continuous values ​​that characterize the pet's physiological metabolic balance.

[0054] The hydration-to-excretion ratio reflects the balance between fluid intake and loss in an individual pet. The edge computing layer extracts and sums the absolute mass changes corresponding to all drinking events within the current sliding time window, and combines this with the absolute mass changes from eating events and a preset food moisture content parameter to calculate the total water intake. Simultaneously, the edge computing layer extracts and sums the absolute mass changes corresponding to all urination events to obtain the total urine output. The formula for calculating the hydration-to-excretion ratio is as follows: ; in, Indicates the ratio of hydrated to excreted components; This indicates the total number of drinking water incidents; Indicates the first The absolute quality change value of the second drinking water incident; This represents the total number of eating events; Indicates the first The absolute change in mass after each feeding event; This represents the preset food moisture content parameter, the value of which can be assigned by the background table after the user enters the food type through the user interaction terminal. This indicates the total number of urination events; Indicates the first The absolute change in quality during each urination event; This represents a preset minimal smoothing constant. The purpose of introducing a minimal smoothing constant is to prevent calculation anomalies where the denominator is zero if no urination event occurs within the current sliding time window.

[0055] Similarly, the intake-to-expenditure ratio is used to quantify the relative relationship between a pet's energy intake and activity expenditure within the current sliding time window. The edge cloud computing layer calculates the total calorie equivalent of food intake within the current sliding time window based on the food quality corresponding to the feeding event and preset calorie conversion parameters per unit mass; simultaneously, it calculates the integral of activity within the current sliding time window based on the amplitude of movement, duration of activity, and frequency of activity collected by microwave radar, activity detection nodes, or wearable activity sensors. The formula for calculating the intake-to-expenditure ratio is as follows: ; in, This indicates the ratio of intake to expenditure; This represents the total number of feeding events; Indicates the first The absolute change in mass after each feeding event; This indicates the preset conversion factor for heat per unit mass; This indicates the total number of events.

[0056] Indicates the first The activity intensity parameter of the event is obtained by heterogeneous sensing nodes when microwave radar is used to collect activity data. The absolute value of each reflection signal intensity change is taken and the average value is calculated to obtain the original activity intensity of the event.

[0057] Heterogeneous sensing nodes normalize the original activity intensity based on low and high activity reference values ​​from the target pet's historical normal activity samples. When the original activity intensity is not higher than the low activity reference value, the activity intensity parameter is set to 0; when the original activity intensity is not lower than the high activity reference value, the activity intensity parameter is set to 1; and when the original activity intensity is between the two, a linear conversion is performed between 0 and 1. As a preferred method, the low and high activity reference values ​​are determined based on the lower and higher quantiles of the original activity intensity in the historical normal activity samples, respectively.

[0058] When the activity detection node can output the motion amplitude or activity frequency, the motion amplitude or activity frequency is converted to the 0 to 1 range in the same way; when more than two activity parameters are obtained for the same activity event, the arithmetic mean of the normalized values ​​of each activity parameter is taken as the activity intensity parameter of the activity event.

[0059] Indicates the first The duration of a sub-event is determined by the time difference between the start timestamp and the end timestamp of the sub-event. This represents the minimum effective activity quantity with the same dimensions as the activity quantity integral, used to avoid a denominator of zero when no effective activity event is detected within the current sliding time window.

[0060] In another alternative implementation, the edge cloud computing layer can further construct a digestion-to-excretion ratio component to quantify the relationship between dry matter intake and total excretion corresponding to defecation events. This digestion-to-excretion ratio component does not replace the aforementioned intake-to-consumption ratio component, but rather serves as an extended feature dimension in the asynchronous behavioral feature matrix to participate in subsequent model training or assist in anomaly interpretation.

[0061] The formula for calculating the ratio of digestion to excretion is as follows: ; in, Indicates the ratio of digestion to excretion; This indicates the preset food moisture content parameter; This represents the total number of feeding events; Indicates the first The absolute change in mass after each feeding event; This indicates the total number of defecation events; Indicates the first The absolute change in quality of each defecation event; This represents a preset minimal smoothing constant.

[0062] S230, the edge cloud computing layer combines timestamp features to extract the diurnal rhythm offset component, and aggregates the confidence score of event-level data packets to calculate the comprehensive confidence component of the event.

[0063] To quantify daily routines, the edge cloud computing layer calculates the frequency of event triggers and the duration of activity during the day and night periods based on the time stamps in the event-level data packets. The boundary between the day and night periods is dynamically set based on the local sunrise and sunset timestamps. The edge cloud computing layer defines the ratio of the cumulative nighttime activity duration to the total cumulative daily activity duration as the circadian rhythm offset component. This component reflects whether the pet exhibits characteristics that deviate from its normal circadian rhythm, such as frequent nighttime awakenings or restless pacing.

[0064] Meanwhile, the edge cloud computing layer extracts the confidence scores of all selected events within the current sliding time window and calculates the overall event confidence component using an arithmetic mean or weighted average algorithm. This overall event confidence component serves as an evaluation indicator of the overall reliability of the data within the current time window and will be used to measure the validity of the sample data in the subsequent individual baseline modeling stage.

[0065] S240, the edge-cloud computing layer performs normalization processing on the calculated feature components of each dimension, and then sequentially concatenates and assembles them into an asynchronous behavioral feature matrix of a unified dimension. Specifically, for the first... Each sliding time window generates a corresponding feature vector from the edge cloud computing layer: ; in, Indicates the first The hydration-to-excretion ratio within a sliding time window. This indicates the ratio of intake to expenditure. Indicates the diurnal rhythm shift component. This represents the overall confidence component of the event. (Continuous) The feature vectors corresponding to the sliding time windows are arranged in chronological order to form an asynchronous behavior feature matrix: ; in, This represents the asynchronous behavior feature matrix corresponding to the current moment. Each row corresponds to a sliding time window, and each column corresponds to a feature component.

[0066] Because the hydration-to-excretion ratio, intake-to-consumption ratio, diurnal rhythm shift, and event comprehensive confidence component have different physical dimensions and numerical scales, directly calculating mathematical distances can easily lead to measurement biases in subsequent model training. Therefore, the edge cloud computing layer employs a max-min normalization algorithm or a standard deviation normalization algorithm to linearly map the aforementioned feature components to the same dimensionless numerical interval. For the specific calculation implementation of the normalization algorithm, those skilled in the art can consult conventional mathematical statistical analysis methods; its data mapping rules are well-known techniques in the field and will not be elaborated upon here.

[0067] See Figure 4 After constructing the multi-dimensional asynchronous behavioral feature matrix, considering the often insidious early onset of chronic diseases in pets, the system struggles to obtain sufficient abnormal disease data with clear pathological labels in real-world home deployment scenarios. Due to this limitation, conventional supervised learning models often fail to converge. Therefore, the edge cloud computing layer employs a single-classification algorithm based on anomaly detection logic to independently construct an adaptive individual baseline model for each pet and infer continuous risk indicators. The specific processing steps are as follows: S310, the edge cloud computing layer extracts the set of asynchronous behavioral feature matrices that are determined to be in a normal physiological state within the historical monitoring period, performs data quality filtering and extracts feature vectors of the physiological dimension, and constructs a historical normal sample set for individual baseline modeling.

[0068] In this embodiment, the edge cloud computing layer uses data generated during a preset observation period after the system's initial deployment, or data generated during the period when users actively confirm the pet's health status through user interaction terminals, as baseline data. The edge cloud computing layer extracts the asynchronous behavioral feature matrix within this time period and removes matrix samples whose event comprehensive confidence component is lower than a set validity threshold, retaining high-quality data to form a historical normal sample set. After quality filtering, the edge cloud computing layer uses the event comprehensive confidence component as a sample selection parameter, sample weighting parameter, or training weight parameter, rather than as a target dimension for physiological abnormality semantic attribution. That is, the event comprehensive confidence component is used to evaluate the reliability of the current window data and to control whether a sample enters the training set, whether it participates in training with reduced weight, and the credibility of the current inference result. For the physiological dimension feature vector used for individual baseline model training and dominant abnormal feature attribution, the edge cloud computing layer preferably extracts the hydration-to-excretion ratio component, the intake-to-consumption ratio component, and the diurnal rhythm shift component; in extended embodiments, other physiological semantic features such as the digestion-to-excretion ratio component can also be added. As a preferred approach, the aforementioned validity threshold can be set to a real number between 0.6 and 0.8 to filter out low-quality feature vectors caused by the mixing of multiple pets, and to prevent non-physiological data fluctuations from contaminating the individual risk deviation inference model.

[0069] S320, the edge cloud computing layer, uses historical normal sample sets to train a single-class support vector machine model, establishing the decision boundary for the normal behavior pattern of the target pet individual.

[0070] The edge cloud computing layer employs a single-class support vector machine algorithm with radial basis function kernel to construct individual baseline models. The basic principle of this algorithm is to utilize only sample data from the normal class, mapping the data to a high-dimensional feature space and finding a hyperplane with the largest distance to the origin, thereby effectively isolating normal and abnormal data. During training, the optimization objective of the individual baseline model is to minimize the norm of the model weight vector and the sum of the slack variables.

[0071] To handle the nonlinear characteristics of the data, the edge cloud computing layer introduces a radial basis function (RBF) kernel to map low-dimensional feature vectors from the historical normal sample set to a high-dimensional feature space. The specific formula for calculating the RBF kernel is as follows: ; in, Represents the radial basis function kernel value; and Represents any two feature vectors in the historical normal sample set; This represents the Euclidean distance between two eigenvectors. Represents the square of the Euclidean distance; This represents an exponential function with the natural constant e as its base. This represents the Gaussian kernel bandwidth parameter, used to control the effective range of a single training sample. Its value range is set to a real number greater than 0.

[0072] For the parameter solving and optimization iteration calculation of this single-class support vector machine model, those skilled in the art can use conventional quadratic programming algorithms such as sequential minimum optimization. The underlying iterative process is well-known in the field and will not be elaborated upon here. Regarding the setting of model training parameters, the edge cloud computing layer will control the penalty parameters for the upper limit ratio of support vectors and the lower limit ratio of training error, dynamically optimizing these parameters based on the sample size of the historical normal sample set and the system's expected false alarm rate requirements. After training, the individual baseline model retains a small number of support vectors constituting the decision boundary and their corresponding weight coefficients.

[0073] S330, the edge cloud computing layer, inputs the feature vector extracted by the current sliding time window into the trained individual baseline model and calculates its spatial deviation distance relative to the decision boundary.

[0074] During the real-time monitoring and inference phase, the edge cloud computing layer extracts the latest feature vector generated within the current sliding time window and inputs it into the individual baseline model. The individual baseline model calculates the inner product of the kernel function between the input vector and each support vector, outputting the functional distance of the feature vector relative to the hyperplane in the feature space. To ensure the accuracy of the distance metric logic, the edge cloud computing layer extracts the positive offset portion as the spatial deviation distance through truncation calculation. The specific formula for calculating the spatial deviation distance is as follows: ; in, Indicates spatial deviation distance; The bias parameter represents the individual baseline model and is used to characterize the intrinsic distance from the origin to the decision boundary; This represents the total number of support vectors; Indicates the first Lagrange multipliers corresponding to each support vector; Indicates the first Support vectors; This represents the feature vector extracted during the current sliding time window; Represents the radial basis function kernel. When the eigenvectors When the value falls inside the decision boundary, the sum of its functional distance inner product is greater than or equal to the bias parameter. At this time, the spatial deviation distance A value truncated to 0 indicates that the current behavior pattern of the pet conforms to its historical routines. When the feature vector... When the value falls outside the decision boundary, the inner product sum is less than the bias parameter. The algorithm outputs a positive spatial deviation distance, indicating the detection of abnormal physiological characteristic deviations.

[0075] S340 introduces a boundary margin threshold in the edge cloud computing layer to filter out daily behavioral jitters, and uses a smooth mapping function to transform spatial deviation distance into a numerical continuous chronic risk deviation index.

[0076] Considering that pets may experience normal behavioral fluctuations in their daily lives due to changes in ambient temperature, dietary adjustments, or short-term activity changes, the edge cloud computing layer sets a boundary margin threshold to distinguish between normal physiological fluctuations and abnormal deviations that require the output of risk values.

[0077] The edge cloud computing layer divides the historical normal sample set into model training samples and normal calibration samples. It uses the model training samples to train the individual baseline model and inputs the normal calibration samples one by one into the trained individual baseline model to obtain the spatial deviation distance corresponding to each normal calibration sample.

[0078] The edge cloud computing layer sorts the spatial deviation distances according to their numerical values ​​and selects the quantile corresponding to the preset normal sample acceptance ratio as the boundary margin threshold. As a preferred approach, the preset normal sample acceptance ratio is set to 95%, that is, the 95th percentile of the spatial deviation distance corresponding to the normal calibration sample is selected as the boundary margin threshold, so that most normal physiological fluctuations fall within the boundary margin threshold.

[0079] When the number of normal calibration samples does not reach the preset calibration quantity, the edge cloud computing layer uses the device's preset initial boundary margin threshold; when the number of normal calibration samples reaches the preset calibration quantity, the boundary margin threshold is updated based on the target pet's own normal calibration samples.

[0080] The edge cloud computing layer compares the calculated spatial deviation distance with the boundary margin threshold. When the spatial deviation distance is less than or equal to the boundary margin threshold, the system determines that the current behavior is a normal physiological behavioral fluctuation, and the risk assessment result remains zero. When the spatial deviation distance is greater than the boundary margin threshold, the edge cloud computing layer applies a nonlinear smoothing mapping function to convert the excess distance value into a continuous chronic risk deviation index in percentage or normalized interval. The formula for calculating the continuous chronic risk deviation index is as follows: ; in, This represents the continuous chronic risk deviation index, with values ​​ranging from 0 to 100. Indicates the boundary margin threshold; This represents the exponential scaling factor, used to adjust the steepness of the increase in risk value as the deviation distance increases. Its value is set to a real number greater than 0, and the specific value can be fitted and calibrated using historical monitoring samples, long-term user feedback samples, equipment experimental samples, or samples confirmed by a veterinarian. Through this inference mechanism, the system can continuously output continuous numerical indicators reflecting the gradual trend of pet health status, providing a quantitative judgment benchmark for subsequent user interface display of graded early warnings and hardware reverse scheduling.

[0081] Spatial deviation distance represents the degree of deviation of the current feature vector from the individual baseline model decision boundary; the boundary margin threshold is used to determine whether to start outputting a non-zero continuous chronic risk deviation index. When the spatial deviation distance is not greater than the boundary margin threshold, the continuous chronic risk deviation index is zero; the continuous chronic risk deviation index is calculated only when the spatial deviation distance is greater than the boundary margin threshold, based on the portion exceeding the boundary margin threshold.

[0082] The exponential scaling factor controls the rate at which the continuous chronic risk deviation index increases with spatial deviation distance after exceeding the boundary margin threshold; it is not involved in determining the boundary margin threshold. A larger exponential scaling factor results in a faster increase in the continuous chronic risk deviation index; a smaller exponential scaling factor results in a more gradual increase in the continuous chronic risk deviation index.

[0083] See Figure 5 After obtaining the continuous chronic risk deviation index, in order to upgrade the system from one-way data monitoring to closed-loop physical intervention, the edge cloud computing layer needs to trace and attribute the specific characteristic components that cause the high-risk index, and perform targeted reverse control scheduling on the underlying sensing nodes based on the attribution results. The specific processing procedure is as follows: S410, the edge cloud computing layer continuously monitors the continuous chronic risk deviation index and initiates the dominant abnormal feature extraction mechanism when it reaches or exceeds the preset intervention trigger threshold.

[0084] In this embodiment, the continuous chronic risk deviation index is configured with an early warning threshold, an intervention trigger threshold, and a blocking threshold, wherein the early warning threshold is lower than the intervention trigger threshold, and the intervention trigger threshold is lower than the blocking threshold. As a specific implementation, the early warning threshold is set to 60, the intervention trigger threshold is set to 75, and the blocking threshold is set to 80.

[0085] When the continuous chronic risk deviation index reaches or exceeds the intervention trigger threshold, the edge cloud computing layer extracts the feature vector of the current input individual baseline model and calculates the contribution of each physiological semantic feature component involved in anomaly attribution to the output offset of the individual baseline model's decision function. The edge cloud computing layer preferably adopts a contribution calculation method based on model decision function perturbation to measure the actual impact of each feature component on the anomaly determination result of the single-class support vector machine model.

[0086] In other implementations, the above thresholds may be adjusted according to the individual health condition of the pet or the expected sensitivity of the system, but the warning threshold should be kept lower than the intervention trigger threshold and the intervention trigger threshold should be kept lower than the blocking threshold.

[0087] Specifically, let the decision function of the individual baseline model be... The feature vector extracted by the current sliding time window is The baseline vector corresponding to the mean center of the historical normal sample set is ,in This represents the total number of feature dimensions involved in anomaly attribution.

[0088] For the The feature components corresponding to each feature dimension are used to construct alternative feature vectors at the edge cloud computing layer. , where only the current feature vector The first in each feature component Replace with the baseline value corresponding to the mean center of the historical normal sample set. The remaining feature components remain unchanged. The edge cloud computing layer calculates the absolute value of the difference between the model decision function outputs before and after the replacement, and uses the absolute value as the first... The base value of the contribution of each feature dimension.

[0089] The feature contribution can be expressed as: ; in, Indicates the first The normalized contribution of each feature dimension to the current model's decision output bias; This indicates the feature dimension number whose contribution is to be calculated. This represents the feature dimension traversal index used for denominator normalization summation; This means only the current feature vector will be used. The first in each feature component Replace with the baseline value corresponding to the mean center of the historical normal sample set. The resulting alternative feature vector; This represents the total number of feature dimensions involved in anomaly attribution.

[0090] When the sum of the base values ​​of the contribution corresponding to all feature dimensions is equal to zero, it indicates that replacing each feature component with the corresponding baseline value has not caused any change in the output of the model decision function. The edge cloud computing layer does not calculate the normalized contribution and marks the current attribution result as having no clear dominant abnormal feature.

[0091] When the sum of the base values ​​of the contribution values ​​corresponding to all feature dimensions is greater than zero, the edge cloud computing layer calculates the normalized contribution of each feature dimension according to the above formula, and the sum of the normalized contribution values ​​of each feature dimension is 1. The contribution threshold is determined based on the total number of feature dimensions involved in anomaly attribution and is set to a value higher than the value corresponding to the average contribution of each feature dimension, in order to distinguish between dominant anomaly features and general correlation features.

[0092] As a preferred approach, the contribution threshold is set to 0.40 when there are three feature dimensions involved in anomaly attribution, and to 0.30 when there are four feature dimensions involved in anomaly attribution. The contribution threshold can also be calibrated using equipment test samples, user-confirmed anomaly samples, or veterinarian-confirmed anomaly samples to ensure that the feature dimensions corresponding to the confirmed anomaly types reach or exceed the contribution threshold.

[0093] The edge cloud computing layer sorts the normalized contributions of each feature dimension in descending order, and identifies the feature dimension with the highest normalized contribution that reaches or exceeds the contribution threshold as the dominant anomaly feature. When the feature dimension with the highest normalized contribution is still below the contribution threshold, the current anomaly state is marked as a multi-feature joint anomaly, without forcibly identifying a single dominant anomaly feature.

[0094] In one alternative implementation, if computing resources are limited, the edge cloud computing layer may also use the normalized deviation ratio of the current feature component relative to the historical normal mean center as an approximate contribution, but this approximate contribution is preferably only used for auxiliary sorting or pre-screening of low computing power nodes.

[0095] S420 establishes a rule-based mapping relationship between dominant abnormal features and physiological metabolic imbalance types in the edge cloud computing layer.

[0096] After extracting the dominant abnormal features and their deviation directions, the edge cloud computing layer uses a pre-built rule engine to perform feature semantic transformation. If the dominant abnormal feature points to the hydration-to-excretion ratio component, the edge cloud computing layer further extracts the water intake sub-component, food water content sub-component, and urine excretion sub-component within the current sliding time window, and calculates their deviation directions relative to the mean of the corresponding sub-component in the historical normal sample set. If the water intake sub-component decreases but the urine excretion sub-component does not decrease synchronously, the system maps it to a water deficiency anomaly; if the water intake sub-component increases and the urine excretion sub-component increases synchronously, the system maps it to a polydipsia and polyuria anomaly; if the urine excretion sub-component decreases but the water intake sub-component does not decrease, the system maps it to a reduced urine excretion anomaly; if both the water intake and urine excretion sub-components show abnormalities but in different directions, the system marks it as an abnormality in water metabolism and maintains the event-triggered high-frequency sampling state of the water dispenser node and litter box node to obtain supplementary data for subsequent continuous windows. If the dominant abnormal feature points to a deviation in the intake-to-expenditure ratio, the edge cloud computing layer maps it to an imbalance between energy intake and activity expenditure. In the extended implementation where the asynchronous behavior feature matrix includes a digestion-to-excretion ratio, if the dominant abnormal feature points to a deviation in the digestion-to-excretion ratio, the edge cloud computing layer maps it to an abnormal digestion-to-excretion relationship. If the dominant abnormal feature points to a positive increase in the diurnal rhythm shift component, the edge cloud computing layer maps it to a circadian rhythm disorder, suggesting that the pet may have increased nighttime activity due to hidden pain.

[0097] The S430 edge cloud computing layer generates corresponding hardware control strategy packages based on the physiological metabolic imbalance type obtained from the mapping, and sends reverse scheduling instructions to specific heterogeneous sensing nodes through the Internet of Things communication protocol.

[0098] As a preferred approach, different types of dominant anomaly features are preferentially associated with the sampling scheduling actions of different nodes. The edge-cloud computing layer generates hardware scheduling control frames based on the dominant anomaly features. The hardware scheduling control frames include at least the target node identifier, target sensor type, sampling mode field, sampling frequency level field, wake-up condition field, duration field, recovery condition field, and priority field.

[0099] When the dominant anomaly points to the hydration-to-excretion ratio component, the edge cloud computing layer sends a sampling scheduling control frame to the smart water dispenser and smart litter box nodes, switching them from a low-frequency polling mode to an event-triggered high-frequency sampling mode. In this mode, the smart water dispenser and smart litter box nodes use microwave radar, proximity sensors, RFID readers, or weight jump signals as wake-up conditions, and increase the sampling frequency of the weighing sensor during the event duration to obtain high temporal resolution data on single-time water intake, urine output, and excretion frequency. Meanwhile, smart feeder nodes with lower correlation to the dominant anomaly maintain a low-frequency polling mode or a dormant mode.

[0100] When the dominant anomaly indicates an intake-to-consumption ratio, the edge cloud computing layer sends sampling scheduling control frames to the smart feeder nodes and activity detection nodes. This causes the smart feeder nodes to increase the sampling frequency of weighing data before and after feeding, and the activity detection nodes to increase the sampling frequency of movement intensity, activity duration, and activity frequency. Meanwhile, smart water dispenser nodes or smart litter box nodes with low correlation to this dominant anomaly maintain a low-frequency polling mode or a dormant mode.

[0101] When the dominant abnormal feature points to the diurnal rhythm offset component, the edge cloud computing layer sends nighttime enhanced sampling control frames to activity detection nodes, proximity detection nodes, or smart litter box nodes, so that the relevant nodes increase the frequency of event trigger detection during the nighttime period to capture abnormal nighttime activities, nighttime drinking, or nighttime excretion behaviors.

[0102] Building upon the aforementioned sampling and scheduling actions, the edge cloud computing layer can also issue optional physical induction or environmental adjustment commands based on specific anomaly types. For example, when the second-level sub-component judgment result corresponding to the hydration-excretion ratio component indicates a water deficiency anomaly, the system can further issue a water flow mode adjustment command to the smart water dispenser node; when the anomaly result corresponding to the intake-consumption ratio component indicates insufficient food intake, the system can further adjust the food dispensing frequency of the smart feeder node; when the excretion frequency increases significantly, the system can further adjust the cleaning waiting time of the smart litter box node. Physical induction or environmental adjustment commands, as additional control measures to the sampling and scheduling actions, do not replace dynamic scheduling at the sampling granularity based on the dominant anomaly characteristics.

[0103] In order to ensure the integrity of the system control logic and the balance of hardware power consumption, the edge cloud computing layer maintains polling calculations of the continuous chronic risk deviation index and the contribution of abnormal features while the hardware is in intervention mode.

[0104] When the continuous chronic risk deviation index falls below the intervention trigger threshold minus the anti-jitter hysteresis value, and the normalized contribution of the dominant abnormal feature that originally triggered the current hardware scheduling action is lower than the contribution threshold determined in S410 within three consecutive sliding time windows, the edge cloud computing layer determines that the current abnormal trend has stabilized and issues a normalization command to the relevant heterogeneous sensing nodes, so that the relevant heterogeneous sensing nodes are restored from the event-triggered high-frequency sampling mode or enhanced sampling mode to the low-frequency polling mode or dormant mode.

[0105] When the continuous chronic risk deviation index falls back only within a single sliding time window, or when the normalized contribution of the dominant abnormal feature that originally triggered the current hardware scheduling action still reaches or exceeds the contribution threshold, the edge cloud computing layer maintains the current sampling scheduling state to avoid frequent switching of sampling modes due to short-term data fluctuations.

[0106] The S450 edge cloud computing layer generates a comprehensive health report that includes source attribution results and intervention records, and pushes it to the user's interactive terminal.

[0107] While performing hardware reverse scheduling, the edge cloud computing layer packages the continuous chronic risk deviation index, the specific names of the dominant abnormal features, the data deviation curve, and the executed hardware control strategy package within the current time window. The edge cloud computing layer encapsulates the above data into a visualized comprehensive health report and pushes it to the user's interactive terminal. Users can intuitively view the deviation dimensions of physiological indicators and the system's intervention actions. This report can also serve as objective data support for subsequent veterinary assessments or health management processes.

[0108] After the edge cloud computing layer outputs the continuous chronic risk deviation index and hardware reverse scheduling strategy, the user interaction terminal synchronously executes the dynamic view reconstruction logic of the graphical user interface to facilitate the graded display and interaction of the pet's abnormal health status. The specific processing procedure is as follows: S510, the user interaction terminal receives real-time data packets sent from the edge cloud computing layer, and performs risk status classification judgment based on the numerical range of the continuous chronic risk deviation index.

[0109] In this embodiment, the continuous chronic risk deviation index is configured with an early warning threshold, an intervention trigger threshold, and a blocking threshold. The early warning threshold is set to 60, the intervention trigger threshold is set to 75, and the blocking threshold is set to 80. Furthermore, the early warning threshold is less than the intervention trigger threshold, and the intervention trigger threshold is less than the blocking threshold.

[0110] When the continuous chronic risk deviation index is below 60, the user terminal determines that the pet is in the normal monitoring level, and the graphical user interface maintains the normal layout structure. When the continuous chronic risk deviation index is greater than or equal to 60 and less than 75, the user terminal determines that the pet is in the sub-health warning level, and executes the top rendering of the data display component associated with the dominant abnormal feature. When the continuous chronic risk deviation index is greater than or equal to 75 and less than 80, the user terminal continues to maintain the sub-health warning level interface, while the edge cloud computing layer starts the extraction of dominant abnormal features and the sampling scheduling of associated hardware nodes. When the continuous chronic risk deviation index is greater than or equal to 80, the user terminal determines that the pet is in the high-risk intervention level and initiates the blocking intervention process.

[0111] In the S520, between the normal monitoring level and the sub-health early warning level, the user interaction terminal executes the dynamic view reconstruction logic of the graphical user interface.

[0112] As a preferred approach, when the system is in a normal monitoring phase, the graphical user interface of the user terminal maintains a preset, conventional layout. When the continuous chronic risk deviation index jumps to the sub-health warning level, the user terminal triggers a partial interface redraw mechanism. The front-end application parses the dominant abnormal feature identifier carried in the data packet and adjusts the rendering level of the data display components associated with that dominant abnormal feature.

[0113] At the underlying code level, the terminal modifies the stacking order parameters or sorting attributes of the front-end component tree to place the data card containing the dominant anomaly characteristics at the top of the visible area and automatically expand the corresponding continuous deviation curve chart. Simultaneously, the color temperature parameter of the interface background smoothly transitions to the preset warning color scheme. For the dynamic modification of the front-end rendering and component styles of mobile application interfaces, those skilled in the art can use conventional responsive frameworks for data binding and view layer updates; the front-end rendering mechanism is a well-known technology in the field and will not be elaborated upon here.

[0114] S530: When the health status is determined to be at the high-risk intervention level, the user interaction terminal suspends the normal component reordering logic and directly initiates the blocking intervention process.

[0115] Since some chronic health risk states require a two-way collaboration between physical environmental intervention and human intervention, when the continuous chronic risk deviation index exceeds the blocking threshold, the terminal renders a full-screen modal view on top of the current graphical user interface. The terminal forcibly loads the dominant abnormal characteristics, comprehensive health report, and hardware reverse scheduling records already executed by the system into the visible area of ​​this modal view.

[0116] S540, the user interaction terminal intercepts the touch event at the modal view layer, forcing the user to perform an informed feedback operation to release the blocking state. In one implementation, the informed feedback operation includes at least one of informed confirmation, exporting electronic health records, and filling out a structured feedback form.

[0117] During the duration of the full-screen modal view, the user interaction terminal intercepts touch events at the application interface layer through the full-screen modal view, blocking user touch operations on the underlying regular business modules within the application, thus forming an application-layer blocking interaction. When the terminal has device management permissions, is in dedicated terminal mode, or is a supervised device, the system can further invoke the system-level single-application locking mechanism to enhance the blocking intervention effect. To restore underlying business interaction permissions, the user needs to perform a preset informed feedback operation within the modal view, such as scrolling the page containing the comprehensive health report to the bottom to simulate a complete reading and triggering an interactive control with informed confirmation attributes. Structured tags include at least one of the following: abnormal observation tags, dietary change tags, excretion change tags, medication status tags, medical visit confirmation tags, environmental stress tags, hardware failure tags, and normal physiological fluctuation tags.

[0118] In S550, after a user completes a specific feedback operation, the user interaction terminal closes the full-screen modal view, restores the normal interaction permissions of the underlying business, and initiates a silent cooling-off period mechanism.

[0119] To avoid a loop of pop-up windows in the modal view caused by a continuous short-term exceedance of the chronic risk deviation index, the user interaction terminal enters a preset silent cooling-off period after the blocking state is lifted. During the silent cooling-off period, the terminal will not repeatedly trigger the blocking intervention process for the same dominant abnormal feature, and will only provide visual prompts on the sub-health warning interface. At the same time, the terminal will encrypt and package the blocking interaction log, which includes the user's operation type, operation content identifier, and timestamp, and send it back to the edge cloud computing layer for archiving. Through the above-mentioned hierarchical reconstruction and blocking interaction mechanism, the system achieves a hierarchical presentation of data and a logical closed loop of mandatory informed intervention on the software application side.

[0120] See Figure 6 As pets age, their environment changes with the seasons, and their diets adjust daily, their regular physiological and metabolic characteristics naturally evolve. This natural evolution can cause conceptual drift at the data level in the decision boundaries established in the early stages of the system, leading to an increase in the false alarm rate of the individual baseline model over long-term operation. To maintain the long-term effectiveness of the anomaly inference logic, a closed-loop data feedback and parameter correction mechanism has been established at the edge cloud computing layer. The specific processing procedure is as follows: S610: The user interaction terminal receives the actual health status assessment results submitted by the user for historical warning events, converts them into structured tags, and uploads the structured tags, the unique identifier of the warning event, and the submission timestamp to the edge cloud computing layer.

[0121] In this embodiment, after a user takes their pet for veterinary treatment or observes it for several days, they can submit an evaluation result through a feedback form built into the user interaction terminal. The evaluation result includes categories such as abnormal conditions confirmed by the veterinarian, normal physiological fluctuations observed by the user, environmental stress interference, or hardware failure. The user interaction terminal converts the above form options into preset standard dictionary codes to form structured tags, and then packages and uploads the structured tags, the unique identifier of the warning event, and the user's submission timestamp together to the edge cloud computing layer.

[0122] S620, the edge cloud computing layer performs time backtracking and feature alignment based on the unique identifier of the warning event, locates the original time node that triggered the warning, extracts the asynchronous behavior feature matrix within the corresponding time interval, and writes it into the calibration sample pool after binding it with structured labels.

[0123] Considering the several-day time lag between the system triggering a high-risk alert and the user actually obtaining veterinary confirmation or ongoing observation results, directly updating the model using data within the submission timestamp window would lead to a temporal misalignment between features and the actual pathological state. Therefore, the edge cloud computing layer uses the unique identifier of the received alert event as a precise query anchor, directly mapping and locating the original time node that triggered the alert in the cloud database. The edge cloud computing layer extracts the asynchronous behavioral feature matrix stored within this time interval, strongly binds it to the uploaded structured labels in the time domain, generates incremental samples with clear positive and negative labels, and stores them in a local calibration sample pool.

[0124] The S630 edge cloud computing layer continuously monitors the data size and time span of the calibration sample pool, and triggers the sample set reconstruction process when the number of incremental samples reaches the update trigger threshold, or when the time interval since the last model training exceeds the maximum reconstruction cycle.

[0125] As a preferred approach, to avoid excessive shifts in the model's decision boundary caused by a single sample, the edge cloud computing layer sets an update trigger threshold and a maximum reconstruction period. The update trigger threshold is set to 5 to 15 samples, and the maximum reconstruction period is set to 30 to 90 days. When the number of incremental samples accumulated in the calibration sample pool reaches the preset update trigger threshold, or when the time interval since the last individual baseline model training exceeds the maximum reconstruction period, the edge cloud computing layer initiates a self-calibration process.

[0126] To address the underlying characteristics of the single-classification model, the edge cloud computing layer performs differentiated processing on incremental samples with different labels: incremental samples labeled as normal physiological fluctuations and environmental stress interference are selected, written into the candidate baseline sample pool, and assigned a state to be verified; in subsequent consecutive sliding time windows, if the continuous chronic risk deviation index falls back to the preset safe range, and the contribution of the corresponding dominant abnormal feature is continuously lower than the preset contribution threshold, then the incremental sample is expanded into the historical normal sample set as a low-weight positive class instance.

[0127] When selecting samples marked as abnormal states confirmed by veterinarians, the edge cloud computing layer does not incorporate them into the historical normal sample set. Instead, it isolates and retains them as negative validation samples. The edge cloud computing layer can then reverse-search and eliminate potential early misclassification data from the historical normal sample set by calculating the similarity between the feature vectors of this negative validation sample and those in the historical normal sample set. For samples marked as hardware failures, the edge cloud computing layer marks them as invalid samples and removes them from the model training sample set.

[0128] The S640 edge cloud computing layer performs splitting processing on incremental samples based on different structured labels and retrains individual baseline models based on the updated sample set.

[0129] To prevent the historical normal sample set from expanding indefinitely, and to ensure that the model can promptly eliminate (or downweight) early physiological features of the pets, the edge cloud computing layer calculates a time weight coefficient for each feature vector in the sample set. The formula for calculating the time weight coefficient is as follows: ; in, Indicates the first Time weighting coefficients for each sample; Indicates the system timestamp that triggered the self-correction operation; Indicates the first Each sample generates a timestamp when it is added to the database; This represents the preset time decay rate parameter, the specific value of which is calculated from the expected characteristic half-life of the system, and is a real number greater than 0.

[0130] The edge cloud computing layer iterates through the historical normal sample set, discarding outdated samples whose time weight coefficients are below a set retention threshold, typically ranging from 0.1 to 0.2. Using the filtered historical normal sample set as input, the edge cloud computing layer employs the same data mapping rules and optimization methods as described above to recalculate the support vector set and bias parameters in the single-class support vector machine model. Through these operations, the model's hyperplane in the feature space is dynamically fine-tuned, and the decision boundary enveloping the normal behavior pattern adaptively shifts accordingly.

[0131] The S650 edge cloud computing layer persistently saves the corrected model parameters and overwrites the original individual baseline model.

[0132] After the model parameters are updated, the edge cloud computing layer clears the corresponding data in the currently processed calibration sample pool, releasing memory usage and allowing the system to smoothly transition to the next monitoring and iteration cycle. By introducing a structured feedback mechanism with delayed alignment logic, the system can continuously absorb deviation characteristics from the actual environment during operation, enabling the inference logic of the continuous chronic risk deviation index to adaptively adjust to changes in the long-term behavioral baseline of individual pets.

[0133] Specific application examples: This example uses a 7-year-old spayed male cat as the target pet, a group that is at high risk for feline lower urinary tract syndrome (FUTS). Early symptoms of this disease include frequent trips to and from the litter box with very little urine output, and increased water intake with little change. The following is in conjunction with the appendix... Figure 7 and attached Figure 8 The technical execution process of the present invention is illustrated by a 15-day continuous monitoring time series. (See appendix.) Figure 7 and attached Figure 8 In the diagram, the horizontal axis (X-axis) represents the sliding time window number (days), which is the continuous time series of the system with a fixed length of 24 hours and a sliding time window that moves according to the natural day.

[0134] During the routine monitoring period from day 1 to day 7, the target pet was in a normal physiological state. The smart litter box node in the IoT hardware sensing layer detected the pet entering via microwave radar, and the RFID reader simultaneously read the RFID tag to identify the pet. After extracting continuous weight signals, the edge-cloud computing layer calculated the confidence score for a single event using a formula: ; in, Indicates the confidence score; , and These represent the corresponding preset weight coefficients; Indicates the success factor of identity verification; This represents the variance of the weighing signal; A mapping function that is negatively correlated with the signal variance; This represents the event classification probability value. Because the confidence score is higher than a preset threshold, the edge cloud computing layer retains this event-level data packet. The edge cloud computing layer extracts data within the sliding time window and calculates the hydration-excretion ratio component: ; in, Indicates the ratio of hydrated to excreted components; This indicates the total number of drinking water incidents; Indicates the first The absolute quality change value of the second drinking water incident; This indicates the preset food moisture content parameter; This represents the total number of feeding events; Indicates the first The absolute change in mass after each feeding event; This indicates the total number of urination events; Indicates the first The absolute change in quality during each urination event; This represents a preset minimal smoothing constant. After constructing an asynchronous behavior feature matrix by combining other feature components, the edge-cloud computing layer inputs the feature vector into the individual baseline model and calculates its spatial deviation distance relative to the decision boundary.

[0135] See Figure 7 The vertical axis (Y-axis) represents the spatial deviation distance. The actual spatial deviation distance (solid line with hollow square markers) in the graph fluctuated between 0.3 and 1.1 from day 1 to day 7, all less than the boundary margin threshold (set to 1.2, represented by a horizontal dotted line). According to the formula: when hour, =0. Where, Indicates spatial deviation distance; Indicates the boundary margin threshold; This represents the continuous chronic risk deviation index.

[0136] Reference Figure 8 The vertical axis (Y-axis) represents the risk deviation index on a percentage basis (0-100). Given the above conditions, the continuous chronic risk deviation index (dashed line with a solid circle marker) for the previous 7 days outputs a value of 0. During this period, the user's interactive terminal determines the health status to be at the normal monitoring level, and the graphical user interface maintains the preset conventional layout structure.

[0137] From day 8 onwards, the target pet's absolute mass change (volume per urination) begins to decrease. Due to the decrease in the cumulative mass value corresponding to each urination event, the hydration-to-excretion ratio increases. (Refer to...) Figure 7 From day 8 onwards, the actual spatial deviation curve began to rise and crossed the boundary margin threshold indicated by the horizontal dotted line. At this point, the edge cloud computing layer enables a non-linear smoothing mapping formula: ; in, This represents the exponential scaling factor. In this embodiment, the exponential scaling factor... Set to 0.505, boundary margin threshold. Set to 1.2. (Refer to...) Figure 8 After calculation using the nonlinear smoothing mapping formula, the continuous chronic risk deviation index increases with the increase of spatial deviation distance. On days 10 and 11, the dashed curve with a solid circle marker exceeded the warning threshold of 60 (represented by a horizontal dotted line). The user interaction terminal determined that the system had entered the sub-health warning level, triggering a local interface redrawing mechanism, which placed the data card containing the abnormal hydration excretion ratio component at the top of the visible area's first screen.

[0138] On day 12, the target pet exhibited frequent entry and exit from the smart litter box but with extremely low-quality excrement. (Refer to...) Figure 7 At this point, the actual spatial deviation reaches a local peak of 4.2. Substituting the spatial deviation of 4.2, the boundary margin threshold of 1.2, and the exponential scaling factor of 0.505 into the nonlinear smoothing mapping formula, we obtain: ; Therefore, refer to Figure 8 The calculated continuous chronic risk deviation index for that day was approximately 78. This value exceeded the system's intervention trigger threshold of 75, but did not reach the blocking threshold of 80 (represented by a horizontal dash). At this point, the edge cloud computing layer initiated the dominant abnormal feature extraction mechanism, calculating that the hydration excretion component had the highest contribution among the various physiological semantic feature components, establishing a rule mapping, and identifying it as a reduced urination type abnormality. Based on the judgment result, the edge cloud computing layer sent a hardware scheduling control frame to the smart litter box node, dynamically changing the node from a low-power polling mode to an event-triggered high-frequency sampling mode to improve the temporal resolution of weight monitoring. Because the index did not reach the blocking threshold, the user interaction terminal did not render a full-screen intervention view, but continued to provide visual prompts on the sub-health warning interface, suggesting that the user arrange veterinary treatment.

[0139] Days 13 to 15 are the recovery period after treatment. After obtaining the veterinary diagnosis (confirming early lower urinary tract obstruction), the target pet's physiological indicators gradually recover. (Refer to...) Figure 7 and Figure 8 From day 13 to day 15, both the actual spatial deviation distance and the continuous chronic risk deviation index showed a significant decrease and fell back below the warning threshold. Users submitted assessment results through a structured feedback form on the user interaction terminal, which was then converted into an abnormal status label confirmed by a veterinarian. The edge cloud computing layer received this label and the unique identifier of the warning event, executed time backtracking and feature alignment logic, and accurately located the asynchronous behavioral feature matrix set from day 8 to day 12.

[0140] The edge cloud computing layer isolates these incremental samples and retains them as negative validation samples, refraining from merging them into the historical normal sample set. This self-correction process avoids the contamination of the individual baseline model's decision boundary by outdated data containing pathological features, maintaining the accuracy of subsequent risk inference logic.

Claims

1. A pet chronic disease intelligent management and human-computer interaction system, characterized in that, include: The IoT hardware sensing layer includes heterogeneous sensing nodes. These nodes switch to event acquisition mode based on environmental trigger parameters or weight changes to acquire continuous weight signals and identification information, and generate event-level data packets with timestamps. The edge cloud computing layer is used to receive the event-level data packets to generate an asynchronous behavior feature matrix, extract feature vectors from the asynchronous behavior feature matrix and input them into the individual baseline model to output a continuous chronic risk deviation index; calculate the contribution of each feature dimension in the feature vector to the continuous chronic risk deviation index, and determine the feature dimension with the highest contribution that reaches or exceeds a preset contribution threshold as the dominant abnormal feature, and generate a hardware scheduling control frame to be sent to the IoT hardware perception layer. The user interaction terminal is used to reconstruct the graphical user interface based on the continuous chronic risk deviation index, and render a modal view when the continuous chronic risk deviation index reaches a preset blocking threshold, and collect structured tags through the modal view.

2. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 1, characterized in that, The heterogeneous sensing node is equipped with a local control unit and a weighing sensor, microwave radar and radio frequency identification reader connected to the local control unit. The weighing sensor is used to collect continuous weight signals, the microwave radar is used to collect reflected signal intensity and activity dwell time, and the radio frequency identification reader is used to read radio frequency tag identification. The local control unit is used to determine the identification information based on the RFID tag identifier, the pet's resting weight value determined by the continuous weight signal, and / or the body shape contour features extracted by the microwave radar.

3. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 2, characterized in that, The heterogeneous sensing nodes include smart water dispenser nodes, smart feeder nodes, and smart litter box nodes. The heterogeneous sensing node calculates the absolute mass change value based on the continuous weight signal; The smart litter box node generates a urination event or a defecation event based on the absolute mass change value and the activity dwell time; the smart water dispenser node generates a drinking event based on the absolute mass change value; and the smart feeder node generates a feeding event based on the absolute mass change value. The local control unit of the heterogeneous sensing node encapsulates the corresponding event type, absolute quality change value, and timestamp into an event-level data packet.

4. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 1, characterized in that, The edge cloud computing layer performs a weighted summation based on the identity recognition success factor determined by the identity recognition information, the weighing signal variance mapping output value of the continuous weight signal, and the event classification probability value extracted from the event-level data packet to obtain the confidence score of the event-level data packet; when the confidence score is lower than the confidence threshold, the corresponding event-level data packet is removed.

5. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 3, characterized in that, The asynchronous behavior feature matrix includes a hydration-to-excretion ratio component and an intake-to-consumption ratio component. The edge cloud computing layer extracts the absolute mass change value corresponding to the drinking water event, and calculates the comprehensive water intake based on the absolute mass change value corresponding to the eating event and the preset food moisture content parameter; Extract the absolute mass change value corresponding to the urination event to calculate the total urine volume, and use the ratio of the total water intake to the total urine volume as the hydration-excretion ratio component; The total calorie equivalent of the food intake is calculated based on the absolute mass change value corresponding to the food intake event and the preset unit mass calorie conversion parameter. The activity intensity parameters of each activity event are determined based on the change in the intensity of the reflected signal collected by the microwave radar. The duration of the activity is determined based on the start and end timestamps of each activity event. The product of the activity intensity parameters and the duration of the activity event is accumulated to obtain the activity quantity integral. The ratio of the total calorie intake to the integral of the activity level is used as the intake-to-expenditure ratio.

6. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 1, characterized in that, The edge cloud computing layer extracts the asynchronous behavior feature matrix within the historical monitoring period to construct a historical normal sample set; A single-class support vector machine algorithm with radial basis function is applied, and the historical normal sample set is input to establish the decision boundary of the normal behavior pattern of the target pet individual, thereby generating the individual baseline model.

7. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 6, characterized in that, The edge cloud computing layer inputs the acquired feature vector into the individual baseline model and truncates and calculates the spatial deviation distance of the feature vector relative to the hyperplane of the decision boundary in the feature space; When the spatial deviation distance is less than or equal to the boundary margin threshold, the continuous chronic risk deviation index is determined to be zero. When the spatial deviation distance is greater than the boundary margin threshold, a nonlinear smoothing mapping function is applied to convert the excess distance value into the continuous chronic risk deviation index in a percentage range.

8. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 6, characterized in that, The edge cloud computing layer replaces each feature dimension in the feature vector with the benchmark value corresponding to the mean center of the historical normal sample set, so as to construct alternative feature vectors corresponding to each feature dimension, and calculates the absolute value of the difference between the output of the model decision function before and after the replacement, as the basic value of the contribution of the corresponding feature dimension. When the sum of the base values ​​of the contribution of all feature dimensions is greater than zero, the normalized contribution of the corresponding feature dimension is determined according to the proportion of the base value of the contribution of each feature dimension to the sum of the base values ​​of the contribution of all feature dimensions, and the feature dimension with the highest normalized contribution and reaching or exceeding the preset contribution threshold is determined as the dominant abnormal feature. When the sum of the base values ​​of the contribution of all feature dimensions is equal to zero, the current attribution result is marked as having no clear dominant abnormal feature; The edge cloud computing layer uses a pre-built rule engine to establish a rule mapping relationship between the dominant abnormal features and the types of physiological metabolic imbalances.

9. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 5, characterized in that, When the dominant abnormal feature points to the hydration-excretion ratio component, the edge cloud computing layer sends a hardware scheduling control frame to the smart water dispenser node and the smart litter box node to increase the sampling frequency of the corresponding node. When the dominant abnormal feature points to the intake-to-consumption ratio, the edge cloud computing layer sends a hardware scheduling control frame to the smart feeder node to increase the sampling frequency of the weighing sensor.

10. The intelligent management and human-computer interaction system for pet chronic diseases according to claim 1, characterized in that, The user interaction terminal is specifically used for: When the continuous chronic risk deviation index is lower than the warning threshold, the graphical user interface maintains the preset conventional layout structure; When the continuous chronic risk deviation index is greater than or equal to the warning threshold and less than the blocking threshold, the data display component associated with the dominant abnormal feature will be rendered at the top. When the continuous chronic risk deviation index is greater than or equal to the blocking threshold, a modal view covering the entire screen is rendered to intercept application-layer touch events, and a structured feedback form is loaded through the modal view to collect structured tags. After the structured tags are collected, the modal view is closed and enters a silent cooling-off period.

Images