A pet health anomaly prompting method, system and device based on co-occurrence corroboration of multi-source evidence and a storage medium

Abstract

The application discloses a pet health anomaly prompt method, system and device based on multi-source evidence co-occurrence verification and a storage medium, and relates to the technical field of intelligent pet monitoring. The method independently detects anomalies of each data source to obtain source-specific anomaly evidence of each data source. The data sources include a sound data source and respective data sources of behavior data, physiological data and environmental data. Co-occurrence determination is performed to determine whether the source-specific anomaly evidence from at least two different data sources corresponds to the same time window pre-configured according to the data sources, and a health anomaly prompt for a target pet is output when the co-occurrence is established. The time window width is configured according to the response delay characteristic difference of the monitoring indicators relative to the health anomaly, and trend anomaly evidence is generated on at least two time scales to participate in the co-occurrence determination. Single-source transient interference is filtered due to the lack of verification of other data sources within the same time window, false positives are reduced, cold start is available, and end-side deployment is supported.

Description

Technical Field

[0001] This application relates to the field of intelligent pet monitoring technology, and in particular to a method, system, device and storage medium for alerting pet health abnormalities based on the co-occurrence of multiple sources of evidence. Background Technology

[0002] As pet ownership expands in households, pets are increasingly becoming like family members, and their health is receiving more and more attention. However, pets cannot express discomfort verbally; their vocalizations, littering behavior, eating and drinking habits, weight changes, and even the concentration of gases in their environment often convey information about changes in their health before obvious symptoms appear. In recent years, pet monitoring products, such as smart litter boxes, smart feeders, and wearable monitoring devices, have developed rapidly. Related patent documents have proposed various health monitoring solutions, which can be broadly categorized as follows.

[0003] CN112369343B discloses a pet health monitoring scheme based on smart litter box data: It collects indicators such as the cat's weight, frequency of littering, duration of littering, and amount of feces, compares the daily value with its seven-day average to obtain the rate of change, and then classifies the health status into three levels according to fixed percentage threshold ranges (e.g., 0-5%, 5%-10%, and greater than 10%). This scheme has the following drawbacks: First, the judgment is based on a single scale and single-point ratio; the "daily to seven-day average" comparison method cannot distinguish between transient fluctuations and continuous trend changes, and even daily dietary fluctuations may lead to misjudgments. Second, the fixed percentage threshold does not consider the differences in normal fluctuation ranges between individuals, breeds, and ages, resulting in insufficient cross-individual reliability. Third, each indicator is compared independently in parallel, without any cross-modal verification between indicators; any indicator exceeding its limit will adjust the health level, and interference from a single measurement is directly transmitted to the output.

[0004] CN121596794A discloses another on-site health monitoring solution: collecting multidimensional health characteristics to establish a personalized health baseline (mean ± standard deviation) for the target pet, dynamically adjusting the judgment threshold based on this baseline, and then weighting and summing the abnormalities in each dimension according to fixed weights to obtain a comprehensive score, which is then mapped to a health risk level. The shortcomings of this solution are: firstly, the judgment reference is a health baseline specific to the target individual, relying on the long-term accumulation of individual data, and lacking a judgment reference in the initial stage of device use (cold start period); secondly, the weighted summation of the comprehensive score linearly superimposes abnormalities from various sources, and a single source of instantaneous interference (such as accidental weighing or insufficient feeding at one time) can lower the comprehensive score and cause false alarms; thirdly, the solution does not address the sound dimension, completely missing the important chief complaint signal of pet vocalizations.

[0005] CN122207608A discloses a method and system for monitoring the health of excrement. Its detection layer scheme (see dependent claims) collects multimodal data (images, weight, humidity, odor) from the excretion scene, and obtains a health index by nonlinearly combining the extracted features after they deviate from the baseline. The shortcomings of this scheme are: first, its "multi-source" is limited to excrement-related collection items, and vocal signals are still missing; second, its fusion is still a weighted nonlinear combination, and false detections of any modality can still be transmitted to the health index, lacking the corroborating constraint of "consistency among multiple sources before prompting"; third, it does not limit the temporal relationship between the evidence from each modality, allowing temporally unrelated false detections to be freely superimposed.

[0006] CN121148665A discloses a pet monitoring scheme: multimodal monitoring data is processed in parallel by an emotion recognition model and a disease recognition model, then aggregated through a structured context interface, and the monitoring results are output by a large language model. Its drawbacks are: first, the judgment relies on large model inference, resulting in high costs due to edge computing power requirements, response latency, and cloud dependency; second, after the two models output their results, the large language model is used for aggregation, lacking temporal co-occurrence constraints at the evidentiary level, and false alarm control depends entirely on the model's generalization ability; third, its output of disease diagnosis information poses risks in terms of patent subject matter eligibility and pet medical compliance.

[0007] CN121709236A discloses another approach: multimodal monitoring data is analyzed using a large model to obtain the anomaly probability of each modality, and the health risk value is obtained by weighted summation of the anomaly probabilities of each modality, which is then used for disease early warning. Its drawbacks are: the weights of each modality are fixed, and the judgment performance degrades significantly when a modality is missing; the numerical coupling of the weighted summation cannot express the verification logic of "consistency across multiple sources is reliable," and a high anomaly probability in a single modality can inflate the total risk value.

[0008] CN121938610A discloses a collar-based solution: the collar collects data from multiple sources and uploads it to the cloud, combining it with a large model and a veterinary knowledge base to generate a health report. Its drawbacks are: the judgment is completed in the cloud, relying on network and large model reasoning, lacking a deterministic anomaly judgment structure, and its real-time performance, offline availability, and data privacy are all constrained.

[0009] CN120260586A discloses a method for analyzing vocalizations: normal / abnormal vocalizations are manually labeled and classified into binary categories, and a deep network is used to locate abnormal segments in the vocalization flow using a sliding window. Its drawbacks are: high cost of manual labeling, subjective abnormality boundaries; and significant individual, breed, and age differences in vocalization acoustic characteristics, making the general judgment boundary obtained based on labeling training unreliable across individuals.

[0010] CN120731890A discloses another scheme for utilizing vocalizations: adjusting the weights of vocalization parameters according to vocalization type, calculating emotional intensity and health values, and then obtaining a multidimensional abnormality risk level from the numerical values. Its drawbacks are: emotional intensity and health values ​​are numerically coupled, meaning a single modal abnormality can affect the total score; the judgment result is influenced by environmental context factors; fusion occurs at the numerical level rather than the evidentiary level, and single-source interference cannot be structurally isolated.

[0011] CN122397627A discloses a smart litter box excrement health analysis scheme: excrement images are collected in three stages: during littering, before leaving after littering, and after automatic cleaning. Sub-results are obtained at each stage, and the analysis results are output by weighting the sub-results according to confidence level. The shortcomings of this scheme are: first, the criterion is based on a single image modality, and the vocal complaint signal is completely missing; second, the confidence level weighted fusion route is not protected against single-source interference—the weighted calculation of the sub-results at each stage cannot express the verification logic of "consistency among multiple sources is reliable," and false detections in a single stage are propagated to the final output according to the weight; third, there is no cross-modal time window co-occurrence constraint.

[0012] Another approach utilizes multi-sensor information for pet identification and measurement. CN116267670A discloses a feeding system that identifies pets through image recognition and associates the identified individuals with their feeding data. CN119014337A (authorization announcement number CN119014337B) discloses a feeder that identifies pets through image sensors, combining weighing and proximity sensing to obtain the weight, frequency, and duration of a single pet's meals using a "head-down posture × weight data collaborative judgment" method. It should be noted that in this type of approach, the "collaboration" of multiple sensors serves the accurate attribution and reliability of the same measurement value (the pet's food intake), essentially a consistency check of the same measurement value. Its purpose and structure differ from cross-modal time window co-occurrence verification at the anomaly evidence layer; furthermore, its entire description does not address anomaly detection, health prompts, or audio dimensions.

[0013] In summary, existing technologies share the following common problems: First, some solutions rely on the long-term accumulation of individual statistics for judgment, lacking usable references during the cold start period; second, they commonly use weighted summation, weighted fusion, comprehensive scoring, or large model aggregation to couple the numerical results from each source into a single output, meaning that any instantaneous interference from any data source will be transmitted to the final result, i.e., existing solutions are not protected against single-source interference; third, the sound dimension is generally missing or relies on costly manual annotation; fourth, solutions relying on large models bring about issues of end-side deployment costs, latency, and availability during network outages; fifth, there is a lack of temporal co-occurrence constraints at the evidentiary level between the results from different sources, allowing temporally unrelated false alarms to be freely superimposed into alarm outputs.

[0014] In view of this, there is a need for a multi-source verification scheme that does not rely on long-term accumulation of individual baselines, does not rely on large model inference and manual annotation, and is not protected against single-source interference, so as to provide reliable, interpretable and edge-deployable prompts for pet health abnormalities under cold start conditions. Summary of the Invention

[0015] The technical problem to be solved by this application is: how to reliably map sound data and multi-source monitoring data such as behavior, physiology, and environment into pet health abnormality prompts without relying on long-term accumulation of individual statistical references, large model inference, and manual annotation, suppress false alarms caused by single-source transient interference, supplement cold start capability, and support low-cost deployment on the edge.

[0016] In a first aspect, this application provides a method for alerting pet health abnormalities, comprising: acquiring sound data of a target pet, and at least one type of monitoring data related to the target pet; wherein the at least one type of monitoring data belongs to at least one type of behavioral data, physiological data, and environmental data of the target pet; independently detecting abnormalities in each data source to obtain source-specific abnormal evidence for each data source; wherein the sound data corresponds to a sound data source, and each type of data in the at least one monitoring data corresponds to an independent data source; performing co-occurrence determination, the co-occurrence determination comprising: determining whether source-specific abnormal evidence from at least two different data sources corresponds to the same time window, and if so, determining that co-occurrence is established; wherein the time window is pre-configured according to the data source, and the time window spans the data collection cycles of different data sources; and outputting a health abnormality alert for the target pet when the co-occurrence is established.

[0017] In one possible implementation, anomaly detection is performed independently on each data source to obtain source-specific anomaly evidence for each data source, including: detecting vocalization events in the sound data to obtain at least one vocalization event; extracting acoustic features of the vocalization event, wherein the acoustic features include at least one of fundamental frequency features, formant features, energy distribution features, and rhythm features; comparing the acoustic features with a preset reference statistical range, and determining the vocalization event as an anomalous vocalization event when the deviation of the acoustic features from the preset reference statistical range meets a preset deviation condition; wherein the preset reference statistical range is a population-level statistical range obtained by pre-statistical analysis of acoustic features from samples of multiple pets; and obtaining source-specific anomaly evidence for the sound data source based on the anomalous vocalization event.

[0018] In one possible implementation, at least one type of the behavioral data, physiological data, and environmental data includes at least one of the following monitoring data sources: weight data, food intake data, water intake data, toilet frequency data, toilet duration data collected by a stationary monitoring device where the target pet is located, and environmental gas concentration data of the environment where the target pet is located; at least one of the target pet's heart rate data, body temperature data, activity level data, and respiratory data collected by a wearable monitoring device; and image data obtained by photographing the target pet, wherein the image data is used to extract the target pet's appearance features and / or body shape features; wherein each of the monitoring data sources belongs to the behavioral data, the physiological data, or the environmental data, and each participates in the source anomaly detection as an independent data source.

[0019] In one possible implementation, the method further includes: determining trend features at at least two time scales for the monitoring indicators represented by the at least one monitoring data; wherein the trend feature at each time scale represents the direction and rate of change of the corresponding monitoring indicator within that time scale; when the direction of change represented by the trend features at at least two time scales is consistent, and the rate of change satisfies a preset rate condition at the corresponding time scale, generating trend anomaly evidence; the trend anomaly evidence, as one of the source anomaly evidences, participates in the co-occurrence determination.

[0020] In one possible implementation, the co-occurrence determination includes: configuring a determination time window for each data source; wherein the window width of the determination time window corresponding to each data source is a preset value, and the window widths corresponding to at least two data sources are configured differently according to the response delay characteristics of the monitored indicators of the corresponding data sources relative to health abnormalities; when the source abnormal evidence from at least two different data sources falls into mutually overlapping determination time windows, co-occurrence is determined; wherein the at least two different data sources participating in co-occurrence include the sound data source, and at least one of the behavioral data source and the physiological data source.

[0021] In one possible implementation, the output of the health abnormality alert for the target pet includes: determining alert level parameters, the alert level parameters including at least two of the following: the number of sources of co-occurring sub-source abnormal evidence; the number of windows in which the co-occurrence is sustained; and a trend rate level determined based on trend abnormal evidence, the trend abnormal evidence being predetermined based on the trend characteristics of monitoring indicators at at least two time scales; determining the level of the health abnormality alert according to preset logical conditions based on each of the determined alert level parameters; and outputting the health abnormality alert with the determined level.

[0022] In one possible implementation, the method further includes: when an event of the target pet using the toilet or eating is detected based on the at least one monitoring data, triggering audio acquisition to obtain audio data; and performing anomaly detection on the audio data to obtain source anomaly evidence of the sound data source.

[0023] In one possible implementation, the method is applied to a monitoring environment with multiple pets. The method further includes: when an event related to any of the multiple pets is detected, determining the attributing individual of the event from a pre-configured set of attribution methods; wherein the event is an event detected through at least one monitoring data source, and the set of attribution methods includes at least one of the following: identifying the identity information represented by an electronic tag carried by the pet; distinguishing individuals based on weighing characteristics determined from weighing data; and distinguishing individuals based on appearance characteristics obtained from pet image recognition; binding the source-specific anomaly evidence associated with the event to the attributing individual to obtain individualized anomaly evidence belonging to the attributing individual; maintaining trend characteristics for each attributing individual, the trend characteristics being determined at at least two time scales based on historical monitoring data bound to the attributing individual; the co-occurrence determination is performed individually, including: for the same attributing individual, determining whether anomaly evidence from at least two different data sources in the source-specific anomaly evidence belonging to the attributing individual corresponds to the same time window; and outputting the health anomaly alert carrying the individual identifier of the attributing individual where co-occurrence is established.

[0024] In one possible implementation, the method further includes: acquiring environmental gas concentration data of the environment in which the target pet is located; wherein the environmental gas concentration data includes ammonia concentration data and / or total volatile organic compound (TVOC) concentration data; generating gas trend anomaly evidence based on the direction and rate of change of the environmental gas concentration data over at least two time scales; wherein the gas trend anomaly evidence is source-specific anomaly evidence from the data source corresponding to the environmental data; determining toilet behavior anomaly evidence based on toilet behavior monitoring data; wherein the toilet behavior monitoring data includes toilet frequency data and / or toilet duration data, and the toilet behavior anomaly evidence is source-specific anomaly evidence from the data source corresponding to the behavior data; determining co-occurrence when the gas trend anomaly evidence and the toilet behavior anomaly evidence correspond to the same time window, and outputting urinary system-related health tips for the target pet.

[0025] In one possible implementation, the method further includes: after outputting the health anomaly alert, receiving confirmation or denial information input by the user in response to the health anomaly alert; in response to the confirmation or denial information, correcting a preset reference statistical range and / or trend determination condition parameters; wherein the preset reference statistical range is used to determine the deviation of monitoring data from the statistical range, and the trend determination condition parameters are parameters used to determine anomalies based on the trend characteristics of monitoring indicators at at least two time scales; wherein the confirmation or denial information is explicit binary feedback input by the user.

[0026] In one possible implementation, the anomaly detection and co-occurrence determination for each data source are performed independently by an edge device; the output of the health anomaly alert for the target pet includes: the edge device outputting the health anomaly alert locally.

[0027] In one possible implementation, the method further includes: the edge device uploading the source anomaly evidence and / or the at least one monitoring data to a cloud device; the cloud device performing a multi-timescale trend review based on the uploaded source anomaly evidence and / or the at least one monitoring data, and updating the trend reference range and / or grade mapping parameters based on the review results; and the cloud device distributing the updated trend reference range and / or grade mapping parameters to the edge device.

[0028] Secondly, this application provides a method for alerting pet health abnormalities using a stationary monitoring device, comprising: acquiring the weight data of a target pet, as well as feeding-related data and / or toilet-related data of the target pet, through the device body of the stationary monitoring device using a conventional weighing method; acquiring environmental gas concentration data of the environment in which the target pet is located; triggering audio acquisition when an feeding event or toilet-related event of the target pet is detected based on the feeding-related data or the toilet-related data, obtaining audio data; and processing the weight data, the feeding-related data and / or the toilet-related data, the environmental gas concentration data, and the... Each audio data source is independently subjected to anomaly detection to obtain source-specific anomaly evidence for each data source; based on the direction and rate of change of the feeding-related data and / or the toilet-related data over at least two time scales, trend anomaly evidence is generated, which serves as one of the source-specific anomaly evidences; it is determined whether the source-specific anomaly evidences from at least two different data sources correspond to the same time window. If so, co-occurrence is determined, wherein the time window is pre-configured according to the data source and spans the data collection cycles of different data sources; when co-occurrence is determined, a health anomaly alert for the target pet is output.

[0029] Thirdly, this application provides a stationary pet health abnormality alert system, comprising: a device body, including a weighing module, an environmental gas sensing module, and an event detection module; the weighing module is used to collect the weight data of the target pet, as well as the target pet's eating-related data and / or toilet-related data, using a conventional weighing method; the environmental gas sensing module is used to collect environmental gas concentration data of the environment in which the target pet is located; the event detection module is used to detect eating events or toilet-related events occurring in the target pet based on the eating-related data or the toilet-related data; and an audio acquisition module is used to detect the eating event or toilet-related event when the event detection module detects the eating event or toilet-related event. Triggered, audio data is collected; a processor is used to independently perform anomaly detection on the weight data, the feeding-related data and / or the toilet-related data, the environmental gas concentration data, and the audio data corresponding to their respective data sources, to obtain source-specific anomaly evidence for each data source; generate trend anomaly evidence based on trend characteristics at at least two time scales; determine whether source-specific anomaly evidence from at least two different data sources corresponds to the same time window pre-configured according to the data source, and output a health anomaly prompt for the target pet when co-occurrence is established; a communication module is used to send the source-specific anomaly evidence and / or the collected data to external devices, and / or receive the sent parameters.

[0030] Fourthly, this application provides a pet health abnormality alert device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of any of the embodiments of the methods in the first or second aspects described above.

[0031] Fifthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the embodiments of the methods described in the first or second aspects.

[0032] Compared with the prior art, the beneficial effects of this application include: (1) High reliability and few false alarms. This application independently detects anomalies from each data source to form source-specific anomaly evidence. An alert is only output when source-specific anomaly evidence from at least two different data sources co-occurs within the same time window. Single-source transient interferences such as accidental weighing, insufficient food intake at one time, or an occasional scream are naturally filtered out because they lack corroboration from other data sources within the same time window. At the same time, the judgment process is deterministic, and each alert can be traced back to the specific evidence combination and time window relationship, making it interpretable and auditable.

[0033] (2) It can be used immediately after cold start. The statistical range referenced is the population-level reference statistical range obtained in advance based on statistics of multiple pet samples. Combined with trend characteristic judgment, it does not rely on the long-term accumulation of target individual-specific statistics. After the equipment is deployed, it can provide monitoring and protection, and there is no gap period without reference.

[0034] (3) Differentiated time windows fit modal characteristics. The judgment time windows of each data source are configured differently according to the response delay characteristics of their monitoring indicators relative to health abnormalities, which conforms to the natural differences of different modal signals on the time scale (for example, changes in behavioral indicators gradually appear over several days, while vocalizations are instantaneous signals), reducing the omissions and mismatches caused by a one-size-fits-all window width.

[0035] (4) Improved trend detection reliability through consistency of trends across multiple time scales. Evidence of trend anomalies is generated only when the trend characteristics of short-term and long-term time scales are in the same direction and the rate of change of each scale meets the preset conditions. This can distinguish between transient fluctuations and continuous changes and suppress misjudgments caused by daily fluctuations.

[0036] (5) Event-triggered audio reduces power consumption and privacy exposure. Audio collection is triggered only when toileting or eating events are detected, which significantly reduces the duty cycle and storage overhead of the audio module compared to continuous monitoring, avoids 24 / 7 recording of the home environment, and supplements the subjective voice signal, so that the audio evidence and behavioral evidence are naturally aligned in time.

[0037] (6) Edge-side deployment friendly. Anomaly detection and co-occurrence determination are deterministic rule logics with no large model inference dependency. They can be executed on low-computing-power edge devices and output prompts locally. They have low latency and are available even when the network is down. The cloud only undertakes non-real-time tasks such as trend verification and parameter updates.

[0038] (7) It can be superimposed with individual adaptation schemes. The scheme in this application does not rely on individual-specific references and can be used independently as a cold start base. It can also be superimposed and deployed with adaptation schemes based on individual data (such as individual voiceprint adaptation schemes) to form a complete evolution path of "cold start base + personalized enhancement". The two schemes do not conflict with each other and complement each other. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments are briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the overall process of a pet health abnormality alert method provided in an embodiment of this application; Figure 2This is a timing diagram illustrating the time window alignment for co-occurrence determination provided in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of a stationary pet health abnormality alert system provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating event attribution and evidence binding in a multi-pet environment, as provided in an embodiment of this application.

[0041] Explanation of reference numerals in the attached drawings: 301—Equipment body; 302—Weighing module; 303—Ambient gas sensing module; 304—Event detection module; 305—Audio acquisition module; 306—Processor; 307—Communication module; 308—Cloud device; 309—User terminal. Detailed Implementation

[0042] The technical solution of this application will be further clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application. The specific values ​​(window width, threshold, sampling frequency, parameter level, etc.) appearing in the embodiments are all exemplary configurations, and those skilled in the art can adjust them according to the specific sensor model, pet breed and application scenario. In this document, "and / or" describes the association relationship of the associated objects, indicating that there can be three relationships. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, and B exists alone. "At least one (at least one)" means any one or any combination of any of the listed objects. "Preset" refers to parameters, ranges or conditions that are pre-configured before use, which can be fixed at the factory or configured and issued by the user or server.

[0043] This embodiment provides an overall flow of a method for alerting pets to abnormal health conditions, such as... Figure 1 As shown, it includes steps S101 to S105.

[0044] S101, Multi-source data acquisition. Acquire sound data of the target pet and at least one monitoring data related to the target pet. Sound data is collected by a microphone, with an example configuration of an omnidirectional electret microphone or MEMS microphone, sampling frequency 16 kHz, quantization bit depth 16 bit, mono. Monitoring data belongs to at least one type of target pet's behavioral data, physiological data, and environmental data. Example configurations include: weight data, food intake data, water intake data, toileting frequency data, and toileting duration data collected by a weighing sensor (resistance strain gauge, conventional sampling frequency 1 Hz, encrypted to 10 Hz when an event is detected) of a stationary monitoring device; environmental gas concentration data collected by an environmental gas sensor (ammonia electrochemical sensor and / or semiconductor metal oxide TVOC sensor, sampling period 30 s); heart rate, body temperature, activity level, and respiration data collected by a wearable device's PPG heart rate sensor (sampling frequency 25–50 Hz), infrared body temperature sensor (sampling period 1 min), and triaxial accelerometer (sampling frequency 25 Hz); and image data captured by a camera during events.

[0045] S102, Source-Specific Anomaly Detection. Anomaly detection is performed independently for each data source, yielding source-specific anomaly evidence for each source. Each piece of source-specific anomaly evidence carries at least: an evidence timestamp, a data source identifier, and an evidence type (e.g., abnormal vocalization evidence, abnormal toilet behavior evidence, abnormal gas trend evidence). "Independently" means that the anomaly detection process for any data source does not read data from other data sources; the sources only converge at the evidence layer, structurally ensuring that a false detection from a single data source will not change the judgment results of other data sources. Sound data corresponds to a sound data source, and each type of data in the monitoring data corresponds to an independent data source (e.g., weight data corresponds to a weight data source, ammonia concentration data corresponds to an environmental gas data source).

[0046] S103 (optional), Trend analysis. For the monitoring indicators represented by the monitoring data, determine the trend characteristics at at least two time scales, and generate trend anomaly evidence when the trends at multiple scales are in the same direction and the rates all meet preset conditions, as one of the sources of anomaly evidence (see Example 4 for details).

[0047] S104, Co-occurrence determination. Determine whether anomaly evidence from at least two different data sources corresponds to the same time window; the time window is pre-configured according to the data source, and the determination method is detailed in Example 5.

[0048] S105, Hierarchical Output. When co-occurrence is established, the prompt level parameter is determined and the level of health abnormality prompt is determined according to preset logic conditions, and a health abnormality prompt with the determined level is output (see Example 6 for details).

[0049] The execution subject of the method in this embodiment can be the main control board of the stationary monitoring device, the home gateway, the user terminal or the cloud server, or a combination of the above devices (see Embodiment Eleven for end-to-cloud collaborative deployment).

[0050] This embodiment provides an implementable process for anomaly detection of sound data sources, including steps S201 to S205.

[0051] S201, Calling Event Detection. The audio stream is framed (frame length 25 ms, frame shift 10 ms), and the short-term energy of each frame is calculated. When the energy exceeds an adaptive threshold and its duration is not less than 100 ms, it is segmented into candidate calling segments. Adjacent candidate segments are merged when the interval is less than 300 ms, resulting in at least one calling event. The adaptive threshold can be the ambient noise floor plus a fixed margin (example: mean noise floor + 12 dB).

[0052] S202, Acoustic Feature Extraction. For each call event, extract at least one of the following acoustic features: fundamental frequency features (estimate the fundamental frequency trajectory using autocorrelation method, and take the fundamental frequency mean and range); formant features (estimate the frequencies of the first two formants using linear predictive coding method); energy distribution features (short-time energy mean, peak value, and spectral centroid); rhythm features (duration of the call event, number of calls per unit time, and sequence of adjacent calls).

[0053] S203, Population-level Reference Statistical Range. The acoustic characteristics of samples from various pets are statistically analyzed beforehand to obtain a population-level statistical range, which serves as a preset reference statistical range. Example: For 300 domestic cats and dogs of different breeds and ages, the acoustic characteristics of their vocalizations are extracted using the procedures described in S201-S202 above. For each feature dimension, the quantile interval from the 5th percentile to the 95th percentile is taken as the reference statistical range for that feature dimension. This reference is a population-level statistical result and is not specifically constructed for the target individual.

[0054] S204, Deviation Detection. The acoustic characteristics of the call event are compared with a preset reference statistical range. If any feature dimension exceeds the corresponding reference statistical range, the preset deviation condition is met, and the call event is identified as an abnormal call event. To suppress occasional noise, it can be further required that N consecutive call events (N=3 in the example) meet the deviation condition before anomaly is confirmed.

[0055] S205, Evidence Output. The confirmed abnormal call events (carrying timestamps and deviation feature terms) are output as source-specific abnormal evidence of the sound data source and participate in the co-occurrence determination in Example 5.

[0056] This embodiment provides feasible sensor configurations and acquisition methods for various monitoring data sources.

[0057] (a) On-site monitoring equipment. An example configuration is a smart litter box with a smart feeder and a smart water dispenser. A bottom-mounted or pedal-mounted weight sensor (resistance strain gauge type, with 24-bit analog-to-digital converter, standard sampling 1 Hz) in the litter box collects weight data and detects entry, exit, and dwell time during littering based on the weight signal, obtaining littering frequency and duration data. The feeder's bowl is weighed at the bottom (sampling 1 Hz) to obtain food intake data based on the weight difference before and after feeding. The water dispenser uses a flow meter or water level sensor (sampling period 10 s) to integrate water intake data. An ammonia sensor (electrochemical type, sampling period 30 s) and / or a TVOC sensor (semiconductor metal oxide type, sampling period 30 s) are placed inside the litter box cavity to collect ambient gas concentration data, which is attributed to environmental data.

[0058] (b) Wearable monitoring device. In the form of a collar or shoulder strap, integrating a PPG sensor (green light, sampling frequency 25–50 Hz) to collect heart rate data; an infrared thermometer (sampling period 1 min) to collect body temperature data; a triaxial accelerometer (sampling frequency 25 Hz) to obtain activity level data through activity intensity integration; and respiratory data to obtain the periodicity of the acceleration waveform envelope. All of the above belong to physiological data.

[0059] (c) Image data. Captured by a home camera or the built-in camera of a device at the time of the event, extracting the target pet's appearance features (fur color histogram, pattern texture features) and / or body features (profile aspect ratio, back curve shape) from the images, for the source anomaly detection of the image data source itself (e.g., persistent abnormal body features can form evidence of body abnormality), and attributed to behavioral data or physiological data representation.

[0060] Each monitoring data source is categorized as behavioral data, physiological data, or environmental data, and each is treated as an independent data source for anomaly detection as described in Example 2. Example categorization: weight, food intake, water intake, toilet frequency, toilet duration, activity level, and appearance and body shape are categorized as behavioral data; heart rate, body temperature, and respiration are categorized as physiological data; and environmental gas concentrations are categorized as environmental data.

[0061] For each monitoring indicator (e.g., 24-hour water intake, daily toilet frequency, daily weight), trend characteristics are determined on both short-term and long-term time scales. Example configuration: short window width 24 h, long window width 7 d; within the window, sample sequences are aggregated daily, and the slope of the least squares linear fit is used as the direction of change. The rate of change (unit: % / day) is normalized to the percentage of the absolute value of the slope relative to the median level of the indicator over the past 7 days.

[0062] The conditions for generating evidence of trend anomalies (preset rate conditions) are: the direction of change is consistent at at least two time scales, and the rate of change at each time scale meets the preset rate threshold for the corresponding scale. Example: For the drinking water index, the short-window rate threshold is +15% / day, and the long-window rate threshold is +5% / day. When the change rate in the short window (24 h) is +22% / day, the change rate in the long window (7 d) is +8% / day, and both windows show an upward trend, evidence of an upward trend in drinking water is generated. Conversely, situations where there is only a sudden increase at a single point on a given day while the long window remains stable (the long window rate does not exceed the limit), or a sudden increase on one day followed by a decline the next day (inconsistent directions), do not generate evidence of trend anomalies, thus distinguishing between transient fluctuations and persistent changes.

[0063] The trend anomaly evidence further carries a trend rate level: it is calculated as a multiple exceeding the rate threshold; exceeding the threshold by 1 to 2 times is considered medium-level, and exceeding it by more than 2 times is considered high-level (example configuration). Trend anomaly evidence, as one of the sources of anomaly evidence, participates in the co-occurrence determination in Example 5. The above window width and threshold are examples and can be configured into parameter tables according to indicator type.

[0064] Configure a judgment time window for each data source, with a preset window width. The window widths for at least two data sources are differentiated based on the response delay characteristics of the monitored indicators relative to health abnormalities. The differentiation is based on the following: different modal signals have different time scales from the occurrence of an anomaly to its observability—changes in behavioral indicators (toilet frequency, food intake) gradually appear over several days, so a longer window width is used (example: 24 h); the response of environmental gas concentration to changes in toilet behavior is a cumulative increase, requiring several hours to distinguish from normal fluctuations, so a medium window width is used (example: 6 h); vocalizations are immediate signals, so a shorter window width is used (example: 5 min).

[0065] One implementation of co-occurrence determination is as follows: For each piece of anomaly evidence e (timestamp t_e, data source d), its determination time window is an interval [t_e−W_d / 2, t_e+W_d / 2] centered on t_e and with a window width W_d; when the determination time windows of anomaly evidence from at least two different data sources intersect, co-occurrence is determined. An equivalent implementation can also be the anchor method: using a pre-determined anomaly evidence time point t0 as the anchor, when another data source has anomaly evidence with a timestamp falling within [t0−W_d, t0+W_d], co-occurrence with the anchor is determined.

[0066] Example (urinary scenario) Figure 2The evidence of abnormal toilet behavior is timed at 20:00 on the 8th day (window width 24 hours, judgment time window [08:00, next day 08:00]); the evidence of abnormal ammonia trend is timed at 15:00 on the 8th day, falling within the judgment time window of 6 hours before and after the behavioral anchor point 20:00 [14:00, next day 02:00]; the evidence of abnormal vocalizations is timed at 20:00 on the 8th day, falling within the judgment time window of 5 minutes before and after the behavioral anchor point 20:00 [19:55, 20:05]. The three judgment time windows have a common intersection [19:55, 20:05], and are both valid (3 data sources).

[0067] Furthermore, the combination of data sources involved in co-occurrence can be constrained: the at least two different data sources involved in co-occurrence include an audio data source, as well as at least one of a behavioral data source and a physiological data source; that is, it requires that the immediate audio complaint signal and the behavioral or physiological signal jointly corroborate each other, further narrowing the triggering surface of co-occurrence.

[0068] The prompt level parameters include at least two of the following: the number of sources of co-occurring anomaly evidence (denoted as K); the number of windows in which the co-occurrence is sustained (denoted as M); and the trend rate level determined based on the trend anomaly evidence (denoted as G, which is either medium or high; this parameter is defaulted if there is no trend anomaly evidence). The level is determined according to preset logical conditions, and the truth table for example is as follows: L1 (Attention Level): K=2 and M=1, with no evidence of abnormal trends. Output example: Displayed on the application's summary page, without interfering with push notifications.

[0069] L2 (Recommended Attention Level): Meets either of the following conditions—K=2 and one of the co-occurring pieces of evidence is evidence of trend anomaly (the trend direction is the same as the co-occurring source indicator); or K≥3. Example output: Push notification, text generated according to the "Recommended Attention" template.

[0070] L3 (Consult as early as possible): K≥3, G=High-end, and M≥2. Output method example: Push notification enhanced reminder (repeated reminder or device audio and visual indication). If multiple conditions are met simultaneously, the highest level will be output.

[0071] The co-occurrence duration window number M is calculated as follows: The reassessment period is 1 / 4 of the maximum window width among the participating data sources (e.g., reassessing every 6 hours for a 24-hour window). The count is accumulated when co-occurrence is true in all consecutive reassessment periods; M=2 when co-occurrence is true in both the current window and the previous reassessment period. The level mapping relationship (the threshold values ​​of each condition in the truth table above) can be maintained as level mapping parameters (see Example 11 for updates under end-cloud collaboration). The source-specific anomaly evidence from each data source participates in the judgment only in the form of "existence / non-existence and count" under the above logical conditions. The level is entirely determined by the logical conditions and does not involve numerical fusion calculations between the source evidence.

[0072] This embodiment uses a stationary monitoring device (example form: smart litter box, also applicable to smart feeders) to illustrate the complete mechanism of event-triggered audio acquisition.

[0073] Equipment structure such as Figure 3 As shown, the device includes a main body 301 (containing a weighing module 302, an ambient gas sensing module 303, and an event detection module 304), an audio acquisition module 305, a processor 306, and a communication module 307. The weighing module 302 continuously acquires weight data and toilet-related data (entry and exit edges, dwell time) in a conventional weighing method (pedal / bottom-mounted weighing sensor, conventional sampling 1 Hz, event segment encryption 10 Hz); in feeder mode, it acquires feeding-related data (weight decrease curve at the bottom of the bowl).

[0074] Event detection conditions (examples): Toileting event = weight signal appears at the entrance edge and lingers for no less than 10 s and then exits at the exit edge; Eating event = the rate of weight drop at the bottom of the bowl exceeds the threshold (example 2 g / s) and lasts for no less than 20 s.

[0075] Audio triggering mechanism: The audio acquisition module 305 is normally in a power-off or sleep state; when the event detection module 304 detects a toileting event or eating event, it wakes up the microphone at the event start time t_e, with a recording window of [t_e−10 s, t_e+50 s], for a total of 60 s, sampling at 16 kHz and 16 bits; it immediately returns to sleep after recording is completed. The recording is processed using the vocalization event detection and deviation judgment method of Example 2 to obtain evidence of abnormal vocalizations, with the timestamp representing the relative time point within the event, and the data source identified as the sound data source.

[0076] The technical benefits of this mechanism (qualitative explanation): First, based on an estimated 20 events per day, each lasting 60 seconds, the duty cycle of the audio module decreases from 100% during continuous monitoring to approximately 1.4%, significantly reducing power consumption and storage overhead. Second, it avoids 24 / 7 recording of the home environment, with audio acquisition occurring only within the event window, reducing privacy exposure. Third, the recording time naturally aligns with toilet / eating events, resulting in a strong temporal correlation between auditory and behavioral evidence, thus improving the effectiveness of co-occurrence determination.

[0077] This embodiment is applied to monitoring environments with multiple pets (e.g., 2-4 pets). Without individual differentiation, co-occurrence determination may result in cross-individual mismatches (e.g., abnormal vocalizations of individual A and abnormal toileting behavior of individual B may incorrectly co-occur). Figure 4 As shown, this embodiment performs event individual attribution before or in parallel with source anomaly detection, and binds source anomaly evidence to the attributed individual.

[0078] The pre-configured attribution methods include the following three, which can be used individually or in combination: (1) Electronic tagging method. Pets wear low-frequency electronic tags (e.g., 134.2 kHz RFID tags conforming to ISO 11784 / 11785) or 2.4 GHz active tags; card readers (low-frequency reading distance not greater than 10cm) are placed at the equipment entrance or passage. If the tag identification information is read within 5 seconds before or after the event, the event is attributed to the individual corresponding to that tag.

[0079] (2) Weighing characteristics. In the registration mode, the weight of each pet is collected and registered separately (example: each pet is placed in the device for 10 minutes, and the average weight of the platform is taken); when an event occurs, the weighing characteristics of the pedal weight curve (entry platform weight, weight fluctuation pattern) are extracted and compared with the registered weighing characteristics of each pet: if the platform weight difference is less than 200 g and only one pet matches, it is assigned to that pet; when multiple pets fall into the range at the same time, the entry and exit weight pairs are logically disambiguated (if only one entry and exit pair is not closed in the entry and exit sequence at the same time, it is assigned to that pet).

[0080] (3) Image appearance method. The event triggers the capture of the image, extracts the appearance features (fur color histogram, pattern texture), and compares them with the pre-registered pet appearance templates: if the similarity is not less than 0.8 and is the only one that exceeds the threshold, it belongs to that pet.

[0081] Combination strategy: Any method that yields a unique attribution is adopted; when there is a conflict between the results of the methods, the method is selected according to the reliability priority (electronic tag > weighing feature > image appearance), or it is marked as "attribution uncertain" and does not participate in the individualized co-occurrence determination.

[0082] Evidence binding and individualized processing: The individual identifier of the pet is written into the event, along with the vocal evidence obtained from the audio data collected triggered by the event (trigger chain inheritance); trend characteristics are maintained separately for each individual, that is, each pet uses only the historical monitoring data bound to that individual to independently calculate the 24 h / 7 d dual-scale trend characteristics; co-occurrence determination is performed separately for each individual, that is, the vocal evidence of individual A is only co-occurred with the behavioral / physiological / environmental evidence of individual A within a time window; the output health abnormality prompt carries the individual identifier of the pet to which the co-occurrence is established (example text: "Dodo's health abnormality prompt").

[0083] It should be noted that the difference in processing in this embodiment is that: the object of attribution is the monitored abnormal related event, the content bound is the source abnormal evidence, and the subsequent execution is individual trend feature maintenance and co-occurrence gating by individual; electronic tags, weighing features, and image appearance are only parallel optional attribution methods.

[0084] This embodiment provides an end-to-end process for outputting health alerts related to the urinary system based on the co-occurrence of evidence of abnormal environmental gas trends and abnormal toilet behavior. Scenario: A smart litter box (weighing module 302 + environmental gas sensor module 303 + event-triggered audio acquisition module 305 + processor 306 + communication module 307) and an application terminal; the monitored object is a 3-year-old domestic cat named "Duoduo".

[0085] Historical stable period (days 1-7): about 4 times a day, each time lasting 2-4 minutes; ammonia concentration in the basin is stable at a low level of 0-5 ppm.

[0086] Data from day 8: 7 toilet visits per day, with each visit lasting 1-2 minutes; ammonia concentration increased from 4 ppm to 9 ppm within 6 hours; an abnormal cry was detected during a toilet visit in the evening, triggered by a recording.

[0087] Execution chain: Step 1: Evidence of Abnormal Toileting Behavior. A daily toilet frequency of 7 times exceeds the preset reference statistical range (population-level statistical example: the upper bound of this range is taken as the 95th percentile, which is 6 times / day), and the duration of each toilet visit is shifted downwards. This constitutes evidence of abnormal toilet behavior, with the timestamp taken as the time of the last toilet visit at 20:00. This evidence is considered source-specific anomaly evidence corresponding to the behavioral data source.

[0088] Step 2: Evidence of Anomaly in Gas Trends. Following Example 4, a dual-scale trend assessment was performed on ammonia concentration: the short-window (24 h) change rate was +120% / day, exceeding the short-window rate threshold (example +50% / day); the long-window (7 d) change rate was +18% / 7 days, exceeding the long-window rate threshold (example +10% / 7 days). Since both windows showed a consistent upward trend, evidence of anomaly in gas trends was generated, with a high trend rate level and the evidence time stamp set to 15:00. This evidence represents source-specific anomaly evidence corresponding to the environmental data source.

[0089] Step 3, Evidence of Abnormal Vocalizations. Abnormal vocalizations (fundamental frequency mean and rhythmic characteristics exceeding the population-level reference statistical range) were detected in the recording triggered by the toilet incident at 20:00 (19:59:50~20:00:50), forming source-specific evidence of anomalous vocalizations from the sound data source, with a timestamp of 20:00.

[0090] Step 4, Co-occurrence Determination. Based on the differentiated window widths of Example 5 (toileting behavior window 24 h, ammonia window 6 h, vocalization window 5 min): Ammonia evidence (15:00) falls within the behavior anchor point ±6 h window; vocalization evidence (20:00) falls within the behavior anchor point ±5 min window; the three-source co-occurrence is established, including both sound and behavioral data sources, and also satisfies the combination constraints of Example 5. According to Example 6, the number of co-occurrence duration windows is counted: In the previous reassessment period, co-occurrence of gas and behavior was established; in this period, M=2.

[0091] Step 5, Tiered Output. K=3 (≥3), G=High-end, M=2 (≥2), meeting the L3 condition, output a three-level prompt (consultation recommended as early as possible). Example push notification text: "Dodo's littering frequency has increased significantly today, with shorter durations per littering session. Ammonia concentration in the litter box continues to rise, and abnormal vocalizations are accompanied by Dodo's littering. We recommend monitoring Dodo's urinary system health. If the situation persists, please consult a veterinarian promptly." The prompt does not include a disease name or any treatment instructions.

[0092] Two additional points need clarification: First, environmental gas data sources do not trigger prompts independently—even if ammonia trend abnormality evidence exists alone, no prompts are output if no toilet behavior abnormality evidence co-occurs within the same time window. That is, gas evidence is always used in conjunction with behavioral event evidence. Second, even without vocalization evidence, the co-occurrence of gas trend abnormality evidence and toilet behavior abnormality evidence is still established and output according to this process (at this time, the level is reduced to L1 or L2 according to Example 6). The role of vocalization evidence is to increase the number of co-occurring sources, thereby raising the level.

[0093] After outputting a health anomaly alert, the application provides two explicit feedback buttons: "Compliant" and "Disagree" (binary input, no implicit behavioral inference). When a user clicks "Disagree" (determined as a false alarm), the application records the data source and evidence type involved in the co-occurrence and corrects the corresponding parameters. Examples include: (a) if the false alarm originates from frequent deviations in sound evidence, the preset reference statistical range of the sound source is narrowed (e.g., the quantile interval is adjusted from the 5th to the 95th percentile to the 10th to the 90th percentile) or the number of consecutive events N in the preset deviation condition is increased; (b) if the false alarm originates from frequent exceedances in the trend rate, the rate threshold of the corresponding indicator is increased (e.g., the short window threshold for water intake is increased from +15% / day to +20% / day); (c) the judgment time window width of the corresponding data source is widened. When a user clicks "Compliant," the current parameters are maintained and the feedback is recorded for manual review.

[0094] The above correction is triggered by explicit binary feedback from user input, and the correction is limited to preset reference statistical ranges and trend determination parameters; the correction does not involve any adjustments to the interface display layout. The corrected parameters will take effect in the next determination and can be synchronized to other devices under the same account after server-side consistency verification.

[0095] The edge device (example: cat litter box controller, ARM Cortex-A53 SoC, 512 MB memory) performs source anomaly detection, co-occurrence determination and hierarchical output according to Examples 1 to 6. Health anomaly prompts are generated locally and pushed to user terminal 309 via communication module 307 (Wi-Fi) with a latency of seconds. When the network is disconnected, the edge device works independently using the parameters issued most recently. Evidence and data are cached locally (with a capacity of not less than 7 days). They are retransmitted after the network is restored, and the determination is not interrupted.

[0096] The cloud device 308 receives source-specific anomaly evidence and / or monitoring data uploaded from the edge device and performs multi-timescale trend verification: It recalculates long-window trend characteristics using the full historical data stored in the cloud (e.g., 30-90 days) to correct estimation biases during periods of insufficient long-window historical data at the edge device; and updates the trend reference range and / or level mapping parameters (i.e., the rate threshold table in Example 4 and the truth table condition threshold in Example 6) based on the verification results, and then distributes the updated parameters to the edge device. Example distribution cycle: once daily, or event-driven distribution when the verification result deviates from the current parameters by more than a threshold; parameter updates are implemented using a gray-scale approach and retain rollback versions.

[0097] The technical benefits of this division of labor are: ensuring real-time performance and availability even when the network is down, while ensuring long-term trend accuracy and parameter iteration capabilities in the cloud. The two are decoupled through an interface that allows "evidence / data to go up and parameters to go down".

[0098] The sound and monitoring data used in all embodiments of this application are directionally collected from sensors deployed in user home environments, including: weighing sensors, ambient gas sensors, event-triggered microphones, and event-capturing cameras on stationary devices, and PPG, temperature measurement, and inertial sensors on wearable devices; publicly available online materials are not used as data sources for judgment. Data collection and use are conducted under user authorization and privacy agreements, and audio collection only occurs within the event-triggered window of Embodiment Seven. Data used for verification can be collected in a controlled test environment according to the above sensor configuration and summarized into a preset reference statistical range using a population-level statistical method, without constructing a specific reference for any particular individual.

[0099] The term "multi-sensor collaboration / fusion" also exists in existing technical literature, but its structure differs from that of this application and needs to be distinguished.

[0100] Firstly, one type of "cooperation" in existing technologies is consistency verification that serves the same measurement value. Taking CN119014337B as an example, it identifies the pet's identity through an image sensor and obtains the pet's food weight by "cooperative judgment of head-down posture × weight data"—the information from both image and weight sensors jointly serves to accurately assign the same measurement value of "how much this pet ate." The essence of cooperation is to improve the consistency verification of the reliability of the same measurement value.

[0101] Secondly, another type of "fusion" in existing technologies couples the results from various sources into a single score at the numerical level. For example, CN122397627A fuses the three-stage image sub-results by weighting them according to confidence level, CN121596794A sums the anomalies of each dimension by weighting them into a comprehensive score, and CN121709236A sums the probabilities of anomalies of each modality by weighting them into a risk value. In this type of structure, the results from various sources are combined into a single value through weighted operations, and the instantaneous error of any data source will be transmitted to the final output according to the weight, leaving no defense against single-source interference.

[0102] The structure of this application differs from both of the above two types: each data source independently performs anomaly detection, first forming discrete source-specific anomaly evidence (timestamp + data source identifier + evidence type), and then performing cross-modal co-occurrence determination at the evidence layer according to the time windows configured differently by the data sources—an alert is only output when the source-specific anomaly evidence from at least two different data sources overlaps in the time window. In other words, the collaboration / fusion of existing technologies is a numerical layer operation serving the same measurement value or a single score; the co-occurrence gating of this application is an evidence layer time constraint serving the anomaly determination of different data sources, without the computational step of merging the results of each source into a single value. No confidence weighting or weighted summation is performed between the source-specific anomaly evidence from each data source; the alert level is determined only by triggering preset logical conditions. This structural difference directly leads to a difference in effect: single-source transient interference is structurally filtered out due to the lack of corroboration from other data sources within the same time window, without relying on any numerical compensation methods.

[0103] The health abnormality prompts and their levels output in the various embodiments of this application are intended to assist users in monitoring their pets' health status. They do not constitute any medical advice, nor do they replace the medical opinions of a licensed veterinarian. Whether a user consults a veterinarian and takes further action after receiving the prompts is up to the user to judge and decide.

Claims

1. A method for indicating abnormal pet health, characterized in that, The method includes: Acquire sound data of the target pet, as well as at least one type of monitoring data related to the target pet; wherein the at least one type of monitoring data belongs to at least one of the target pet's behavioral data, physiological data, and environmental data. Anomaly detection is performed independently on each data source to obtain source-specific anomaly evidence for each data source; wherein, the sound data corresponds to a sound data source, and each type of data in the at least one monitoring data corresponds to an independent data source; The co-occurrence determination includes: determining whether the anomaly evidence from at least two different data sources corresponds to the same time window; if so, the co-occurrence is determined to be established; wherein, the time window is pre-configured according to the data source, and the time window spans the data collection cycle of different data sources; When the co-occurrence is established, output a health abnormality alert for the target pet.

2. The pet health abnormality alert method according to claim 1, characterized in that, The process of independently detecting anomalies in each data source to obtain source-specific anomaly evidence for each data source includes: The sound data is subjected to call event detection to obtain at least one call event; The acoustic features of the call event are extracted; wherein the acoustic features include at least one of the following: fundamental frequency features, formant features, energy distribution features, and rhythmic features; The acoustic features are compared with a preset reference statistical range. When the deviation of the acoustic features from the preset reference statistical range meets a preset deviation condition, the vocalization event is determined as an abnormal vocalization event. The preset reference statistical range is a population-level statistical range obtained by statistically analyzing the acoustic features of samples from multiple pets. Based on the abnormal noise event, evidence of source anomalies in the sound data source is obtained.

3. The method for indicating abnormal pet health according to claim 1, characterized in that, At least one type of behavioral data, physiological data, and environmental data includes at least one of the following monitoring data sources: The on-site monitoring device for the target pet collects data on weight, food intake, water intake, toilet frequency, toilet duration, and environmental gas concentration in the environment where the target pet is located. At least one of the following: heart rate data, body temperature data, activity level data, and respiratory data of the target pet collected by the wearable monitoring device; Image data obtained by photographing the target pet, the image data being used to extract the appearance features and / or body features of the target pet; Each of the monitoring data sources belongs to the behavioral data, the physiological data, or the environmental data, and each participates in the source anomaly detection as an independent data source.

4. The pet health abnormality alert method according to claim 1, characterized in that, The method further includes: For the monitoring indicators represented by the at least one monitoring data, trend characteristics are determined at at least two time scales; wherein, the trend characteristics at each time scale represent the direction and rate of change of the corresponding monitoring indicator within that time scale; When the trend features at at least two time scales represent the same direction of change, and the rate of change satisfies the preset rate condition at the corresponding time scale, evidence of trend anomaly is generated. The trend anomaly evidence, as one of the source anomaly evidences, participates in the co-occurrence determination.

5. The method for indicating abnormal pet health according to claim 1, characterized in that, The co-occurrence determination includes: Configure a judgment time window for each data source; wherein, the window width of the judgment time window for each data source is a preset value, and the window widths for at least two data sources are configured differently according to the response delay characteristics of the monitored indicators of the corresponding data sources relative to health abnormalities. When anomaly evidence from at least two different data sources falls into overlapping decision time windows, co-occurrence is established. The at least two different data sources involved in co-occurrence include the sound data source, as well as at least one of the behavioral data source and the physiological data source.

6. The method for indicating abnormal pet health according to claim 1 or 5, characterized in that, The output includes health abnormality alerts for the target pet, including: The alert level parameters are determined, including at least two of the following: the number of sources of co-occurring sub-source anomaly evidence; the number of windows in which co-occurrence is sustained; and the trend rate level determined based on trend anomaly evidence, which is predetermined based on the trend characteristics of the monitoring indicators at at least two time scales. Based on each of the determined prompt level parameters, the level of the health abnormality prompt is determined according to preset logical conditions; Output the health abnormality alert with the determined level.

7. The method for indicating abnormal pet health according to claim 1 or 3, characterized in that, The method further includes: When a toilet-use or eating event of the target pet is detected based on at least one of the monitoring data, audio acquisition is triggered to obtain audio data. Anomaly detection is performed on the audio data to obtain source-specific anomaly evidence of the sound data source.

8. The method for indicating abnormal pet health according to claim 1, characterized in that, The method is applied to monitoring environments with multiple pets, and the method further includes: When an event related to any of the multiple pets is detected, the individual to which the event belongs is determined from a variety of pre-configured attribution methods; wherein, the event is an event detected through the at least one monitoring data, and the multiple attribution methods include at least one of the following: identifying the identity information represented by the electronic tag carried by the pet; distinguishing individuals based on weighing characteristics determined by weighing data; and distinguishing individuals based on appearance characteristics obtained by pet image recognition; By binding the source-specific anomaly evidence associated with the event to the attributing individual, individualized anomaly evidence belonging to the attributing individual is obtained. The trend characteristics of the associated individuals are maintained separately for each individual, and the trend characteristics are determined at at least two time scales based on historical monitoring data associated with the associated individuals; The co-occurrence determination is performed on an individual basis, including: for the same attributed individual, determining whether the anomalous evidence from at least two different data sources among the source anomalous evidence belonging to the attributed individual corresponds to the same time window; The output of the health abnormality notification carries the individual identifier of the individual to which the co-occurrence was established.

9. The method for indicating abnormal pet health according to claim 1, 3, or 4, characterized in that, The method further includes: Obtain environmental gas concentration data of the environment in which the target pet is located; wherein, the environmental gas concentration data includes ammonia concentration data and / or total volatile organic compound concentration data; Based on the direction and rate of change of the environmental gas concentration data over at least two time scales, evidence of gas trend anomalies is generated; wherein, the evidence of gas trend anomalies is source-specific anomaly evidence corresponding to the environmental data source. Evidence of abnormal toilet behavior is determined based on toilet behavior monitoring data; wherein, the toilet behavior monitoring data includes toilet frequency data and / or toilet duration data, and the evidence of abnormal toilet behavior is source-specific abnormal evidence from the data source corresponding to the behavior data; When the evidence of abnormal gas trends and the evidence of abnormal toileting behavior correspond to the same time window, a co-occurrence is determined, and a health alert related to the urinary system for the target pet is output.

10. The method for indicating abnormal pet health according to claim 1, characterized in that, The method further includes: After outputting the health abnormality prompt, the system receives confirmation or denial information from the user regarding the health abnormality prompt. In response to the confirmation information or the denial information, the preset reference statistical range and / or trend judgment condition parameters are corrected; wherein, the preset reference statistical range is used to determine the deviation of the monitoring data from the statistical range, and the trend judgment condition parameters are parameters on which the anomaly judgment is based based on the trend characteristics of the monitoring indicators at at least two time scales. The confirmation information or the denial information is an explicit binary feedback input by the user.

11. The method for indicating abnormal pet health according to claim 1, characterized in that, The independent anomaly detection for each data source and the co-occurrence determination are performed by the edge device. The output of the health abnormality prompt for the target pet includes: the health abnormality prompt being output locally by the edge device.

12. The pet health abnormality alert method according to claim 11, characterized in that, The method further includes: The edge device uploads the source anomaly evidence and / or at least one monitoring data to the cloud device; The cloud device performs multi-timescale trend verification based on the uploaded source anomaly evidence and / or at least one monitoring data, and updates the trend reference range and / or level mapping parameters based on the verification results. The cloud device will send the updated trend reference range and / or the level mapping parameters to the edge device.

13. A method for indicating abnormal pet health in a stationary monitoring device, characterized in that, The method includes: The stationary monitoring device acquires the target pet's weight data, as well as its feeding and / or toileting data, through a weighing sensor. Obtain the ambient gas concentration data of the environment in which the target pet is located; When an eating or toileting event is detected in the target pet based on the feeding-related data or the toileting-related data, audio acquisition is triggered to obtain audio data. Anomaly detection is performed independently on the data sources corresponding to the weight data, the eating-related data and / or the toilet-related data, the environmental gas concentration data and the audio data, to obtain source-specific anomaly evidence for each data source. Based on the direction and rate of change of the eating-related data and / or the toilet-related data over at least two time scales, evidence of trend anomalies is generated; the evidence of trend anomalies is one of the sources of anomalies. Determine whether the anomaly evidence from at least two different data sources corresponds to the same time window. If so, determine that co-occurrence is established. The time window is pre-configured according to the data source and spans the data collection cycles of different data sources. When the co-occurrence is established, output a health abnormality alert for the target pet.

14. A stationary pet health abnormality alert system, characterized in that, include: The device body includes a weighing module, an ambient gas sensing module, and an event detection module. The weighing module is used to collect the weight data of the target pet, as well as the pet's eating-related data and / or toilet-related data, through the weighing sensor. The ambient gas sensing module is used to collect the ambient gas concentration data of the environment in which the target pet is located. The event detection module is used to detect eating events or toilet-related events that occur in the target pet based on the eating-related data or the toilet-related data. The audio acquisition module is used to acquire audio data when the event detection module detects the eating event or the toileting event. The processor is configured to independently perform anomaly detection on the data sources corresponding to the weight data, the eating-related data and / or the toilet-related data, the environmental gas concentration data, and the audio data, respectively, to obtain source-specific anomaly evidence for each data source; and generate trend anomaly evidence based on trend characteristics at at least two time scales. Determine whether the anomaly evidence from at least two different data sources corresponds to the same time window pre-configured by data source, and output a health anomaly prompt for the target pet when co-occurrence is established; The communication module is used to send the source anomaly evidence and / or the collected data to external parties, and / or receive the parameters sent out.

15. A pet health abnormality alert device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the pet health abnormality alert method as described in any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the pet health abnormality alert method as described in any one of claims 1 to 13.

Images