Intelligent captive panda health monitoring system

Abstract

The invention belongs to the technical field of giant panda health monitoring, and discloses an intelligent captive giant panda health monitoring system, which adopts an air, ground and environment three-source acquisition architecture, realizes uninterrupted data acquisition through equipment such as a micro unmanned aerial vehicle, a thermal infrared imager and an intelligent food bowl, adds time-space stamps to all acquired data and completes multi-source mutual verification, and realizes intelligent captive giant panda health monitoring. The data accuracy is ensured; meanwhile, local alarm can be quickly triggered according to an abnormal condition by means of edge calculation and a hierarchical priority transmission mechanism, and an abnormal level and a possible reason can be timely determined in cooperation with a four-level early warning mechanism, so that the response speed to sudden abnormity is improved; three types of data of physiology, behavior and environment are fused through a double-model architecture to perform causal reasoning, the adaptability of behavior judgment is improved in combination with a scenarized behavior recognition algorithm, meanwhile, a main baseline and a sub-baseline are constructed according to the age, the physiological cycle and other dimensions of pandas, and the exclusive behavior mode of each panda is recorded, so that health judgment better fits the actual conditions of individuals.

Description

An intelligent health monitoring system for captive giant pandas Technical Field

[0001] This invention belongs to the field of giant panda health monitoring technology, specifically an intelligent health monitoring system for captive giant pandas. Background Technology

[0002] Captive breeding of giant pandas is a crucial way to maintain their population. Health monitoring is essential for disease prevention and control, improving reproductive success rates, and ensuring population quality. It requires precise data to support full life-cycle management. Health monitoring of captive giant pandas is a key link in species conservation. However, current technologies still have the following technical problems: Traditional monitoring mainly relies on keepers' regular patrols and manual recording, which cannot cover critical periods such as nighttime and breeding season, and the proportion of effective 24-hour monitoring data is not high; for sudden abnormalities such as collapsing or refusing to eat, the average detection time is too long, and it is easy to miss the best intervention opportunity; in addition, manual recording is greatly affected by subjective factors, and the data error rate is high, making it difficult to provide reliable support for dynamic health assessment.

[0003] Most existing systems only collect data from a single dimension in isolation, such as collecting behavioral videos or physiological indicators, without establishing a correlation mechanism between physiological, behavioral, and environmental data. For example, traditional hormone testing has a significant lag and cannot be linked with real-time behaviors such as anxiety and frequent wall-rubbing, as well as environmental parameters such as sudden changes in temperature and humidity, resulting in low accuracy in predicting estrus and easy to miss early signs of disease.

[0004] Behavior recognition mainly relies on a single visual algorithm, which does not take into account individual differences. For example, older pandas move slowly and are easily misjudged as abnormal, resulting in low accuracy. Moreover, the use of uniform health standards, such as setting the same body temperature threshold regardless of age, does not take into account the baseline differences of giant pandas at different growth stages and different physiological cycles. This leads to a high misjudgment rate of health abnormalities and increases the cost of ineffective interventions. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent health monitoring system for captive giant pandas to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent health monitoring system for captive giant pandas, comprising a data acquisition module, a data preprocessing module, a health identification module, a health baseline construction module, an early warning response module, a decision-making module, and an operation and maintenance module; preferably, the data acquisition module adopts a three-source acquisition architecture of air, ground, and environment, with micro-drones deployed in the air layer to achieve real-time positioning and activity trajectory mapping of giant pandas; a monitoring unit is set up in the ground layer, integrating an infrared thermal imager with a sampling rate of 10 frames / second and a millimeter-wave radar with a detection range of 5-20m, simultaneously capturing typical behaviors including feeding and estrus through a 4K panoramic camera; the environment and diet linkage layer is located in the enclosure. Temperature, humidity, and ammonia sensors with a sampling frequency of once every 30 seconds, along with a smart feeding bowl with a weight accuracy of ±2g, are deployed. All data collected from these devices are appended with millisecond-level timestamps and centimeter-level spatial coordinates, providing a spatiotemporal correlation basis for dynamic compression at the edge. Cross-validation of data from infrared thermal imagers and millimeter-wave radar, as well as matching of data from the smart feeding bowl and 4K cameras, enables mutual verification of multi-source data to ensure that the data transmitted to the edge is free from abnormal deviations. An analysis unit based on Granger causality test is integrated at the edge, using a 1-hour sliding window to mine causal relationships between data and add labels, providing causal inference material for the Bayesian network of the health recognition module and structured raw data for the data preprocessing module.

[0007] Preferably, the data preprocessing module receives multi-source data from the acquisition module, adopts an edge computing processing architecture, and has an edge computing chip with a computing power of ≥2 TOPS built into the acquisition terminal. It removes video noise by improving the VIBE background modeling algorithm and repairs instantaneously missing sensor data using Kalman filtering. A two-stage processing mechanism of feature extraction and redundancy removal is constructed to extract behavioral keyframes and physiological feature values, reducing the amount of invalid data processed by the health recognition module and retaining only structured feature data in JSON format. It establishes bidirectional communication with the health recognition module, receiving high-value feature types from the module and implementing feedback-based dynamic compression. When high-value data is identified, the compression ratio is reduced from 10:1 to 5:1, while the compression ratio is increased for data without panda activity at night. A scene keyframe extraction library is established, with classification based on typical behaviors captured by the acquisition module to ensure that the extracted keyframes match the behavior sample library of the health recognition module. A hierarchical priority transmission protocol is adopted, with abnormal behavior data transmission latency ≤100ms. Regular data is transmitted in batches. A simplified YOLOv9 model is integrated at the edge, and its pre-trained samples share part of the dataset with the health recognition module. When high-risk anomalies are identified, a local audible and visual alarm is triggered directly, and the original data is transmitted to the cloud with priority.

[0008] Preferably, the health recognition module receives structured data processed by the data preprocessing module and adopts a dual-model architecture causal inference fusion model combining LSTM and Bayesian networks. The LSTM network analyzes the temporal changes of physiological indicators such as body temperature and respiratory rate with a 2-hour time window. The analysis process is based on the physiological feature values ​​output from the edge end, combined with the causal labels from the acquisition module. The Bayesian network infers the root cause of health abnormalities and outputs the correlation degree of abnormality causes based on the causal label data from the acquisition module. Based on the improved YOLOv9 algorithm with an attention mechanism, a scenario-based behavior recognition algorithm is constructed, whose training set covers behavior samples of 500+ individuals. The sample annotation refers to the individual behavior library initially established by the health baseline construction module. A dynamic adjustment mechanism for scenario and feature weights is constructed, and the adjustment process refers to the environmental parameters of the acquisition module and the scenario-specific sub-baselines of the health baseline construction module. Three types of features are fused: spatiotemporal correlation data, causal label data, and environmental parameters from the acquisition module, and weights are assigned according to the scenario. An evaluation index system is constructed, with index dimensions referring to the physiological and behavioral classifications of the health baseline construction module. A health status vector of 0-100 points is output through a weighted voting algorithm.

[0009] Preferably, the health baseline construction module receives health status data and abnormal cases from the health identification module. Based on continuous monitoring data, it uses an improved K-means clustering algorithm to divide dimensions by age, physiological cycle, and activity scenario to generate a main baseline and sub-baselines. This baseline serves as the benchmark for the health identification module to determine abnormalities. A 14-day sliding window algorithm is used to incorporate the latest structured data from the data preprocessing module in real time. The baseline threshold is adjusted based on the intervention effect feedback from the decision-making module. The baseline deviation is periodically corrected based on physical examination data, and the correction results are synchronously fed back to the health identification module to update the abnormality judgment criteria. The unique behavioral patterns of each panda are recorded and incorporated into the baseline judgment criteria.

[0010] The specific data processing steps of the 14-day sliding window algorithm are as follows: (1) Window data filtering: From the structured data obtained from the data preprocessing module every day, outliers are removed (judged by the 3σ criterion, i.e., data that deviates from the daily mean ± 3 times the standard deviation of the indicator), and date data with ≥ 80% of the effective data are retained and included in the window; (2) Data update logic: For each new day of effective data, the earliest day of data in the window is automatically removed to ensure that the window always maintains 14 days of continuous data; (3) Baseline calculation method: The data in the window are recalculated according to the "mean ± standard deviation" to determine the main baseline and sub-baseline, where the standard deviation coefficient is determined based on the data stability. Qualitative dynamic adjustment (when the data variation coefficient is <5%, the standard deviation coefficient is 1.2; when the variation coefficient is 5%-10%, it is 1.5; when the variation coefficient is >10%, it is 1.8); (4) Quantitative rules for adjusting the intervention effect: if the decision module reports "intervention is effective" (i.e., the health indicators return to the baseline range within 72 hours after the intervention), the baseline threshold is slightly adjusted by "original threshold × 0.95"; if the report reports "intervention is partially effective" (the indicators return to the baseline ±10% range), the threshold is slightly adjusted by "original threshold × 0.98"; if the report reports "intervention is ineffective", the original threshold is retained and the case is marked for subsequent baseline deviation correction.

[0011] Preferably, the early warning response module receives the health status vector and abnormal cause from the health identification module and establishes a four-level graded early warning mechanism: 85-100 points are normal, 70-84 points are general early warning, 50-69 points are emergency early warning, and <50 points are critical early warning; the early warning content includes abnormal indicators, suspected causes and correlation, directly citing the causal reasoning results of the health identification module, and providing attribution basis for the decision-making module; according to the early warning cause, a scenario-based intelligent linkage strategy is matched. When the abnormal cause is a sudden change in temperature and humidity, the air conditioner is automatically adjusted and the temperature deviation is controlled within ±0.5℃, and the adjustment command is sent to the enclosure equipment managed by the operation and maintenance module; when the abnormal cause is insufficient food intake, the intelligent feeding trough adjusts the bamboo shoot ratio, and the feeding trough control signal comes from the dietary linkage layer of the acquisition module; when the abnormal cause is a suspected disease, the isolation door is automatically closed, and the door control status is fed back to the operation and maintenance module for recording; the specific parameter thresholds of the scenario-based intelligent linkage strategy are: (1) sudden change in temperature and humidity judgment criteria: temperature change ≥3℃ or humidity change within 10 minutes. ≥15%, triggering air conditioning adjustment; the adjustment target is "current environmental parameters ±0.5℃ / ±5% humidity", and the air conditioning speed is dynamically adjusted according to the temperature difference (the air speed is set to level 2 when the temperature difference is 1-2℃, and the air speed is set to level 3 when the temperature difference is >2℃); (2) Insufficient food intake judgment and adjustment: when the smart food bowl detects "the food intake in 1 hour is <50% of the average food intake of the panda in the same period in history", the bamboo shoot ratio adjustment is triggered; the adjustment rule is "original bamboo shoot ratio (usually 30%) + 15%, while reducing the proportion of concentrated feed by 10% and the proportion of fruits and vegetables by 5%", and after adjustment, the food intake is monitored once every 30 minutes. If the standard is met twice in a row (≥70% of the same period in history), the adjustment will stop; (3) Response to suspected disease isolation door: when the health identification module outputs "suspected infectious disease correlation ≥80%" (such as suspected digestive tract infection, respiratory tract infection), the isolation door is closed within 30 seconds. After closing, the door control status is fed back through the operation and maintenance module ("closed in place" / "fault"). When there is a fault, a maintenance reminder is triggered and the manual control mode is switched.

[0012] Automatically record the warning trigger time, response operation, and handling result to form a traceability log, providing cases for the operation and maintenance module to optimize equipment linkage strategy and the health identification module to adjust feature weights; regularly count false warnings and missed warnings, and statistically analyze the handling effect based on the feedback from the reference decision module, adjust the health score threshold accordingly, and synchronize the adjustment results to the health identification module.

[0013] Preferably, the decision-making module takes over the abnormal causes and health status of the early warning response module, constructs a multi-dimensional expert knowledge base, and classifies the case classification by referring to the abnormal types of the health identification module and the individual dimensions of the health baseline construction module, and stores the treatment plan according to the disease type, individual characteristics, and abnormal causes; constructs a reinforcement learning decision-making model with the goal of optimal intervention effect, and the training data includes the individual baseline of the health baseline construction module and the abnormal causes of the early warning response module. Combining the individual health baseline, real-time health status, and abnormal causes, it outputs a combination of dietary adjustment, drug recommendation, and environmental optimization; the dietary adjustment instruction is sent to the smart food bowl of the collection module, and the environmental optimization instruction is sent to the enclosure equipment of the operation and maintenance module; the drug recommendation includes the dosage range; the core elements and training details of the reinforcement learning decision-making model are: (1) State space definition: includes 3 types of dimensions, a total of 28 features, including individual baseline features (body temperature, food intake, etc. 8 dimensions), real-time health status features (health score, abnormal indicators, etc. 12 dimensions), and environmental features (temperature, humidity, ammonia concentration, etc. 8 dimensions). All features are normalized to the [0, 1] interval; (2) Action space definition: divided into 3 types of actions, a total of 18 types of actions. The operation includes dietary adjustments (6 types such as bamboo shoot ratio and feeding frequency), drug recommendations (4 types such as antibiotic dosage and probiotic types), and environmental optimization (8 types such as temperature, humidity, and light). Each action corresponds to a specific parameter range (e.g., antibiotic dosage: 20-30 mg / kg for adult pandas and 10-15 mg / kg for juvenile pandas); (3) Reward function design: a multi-objective weighted reward is adopted, and the total reward = health recovery reward (accounting for 60%) + intervention cost reward (accounting for 20%) + individual adaptation reward (accounting for 20%); where the health recovery reward = (dry Prognostic health score - pre-intervention health score) / 100, intervention cost reward = 1 - (drug cost + equipment energy consumption cost) / baseline cost, individual adaptation reward = 1 - (action parameter deviation rate from individual baseline); (4) training parameters: DQN algorithm is used for training, the experience replay pool capacity is set to 10,000, the target network update frequency is every 100 steps, the initial learning rate is 0.001, the ε-greedy exploration rate is linearly decayed from 0.9 to 0.1 (decay step number 5000 steps), and training is continued until the reward function converges (reward fluctuation < 5% for 1000 consecutive steps).

[0014] Changes in health indicators are recorded within 72 hours after intervention, with data obtained from real-time monitoring of the data acquisition and preprocessing modules. If several expected results are not achieved, the plan is automatically adjusted, and the adjustment logic is fed back to the health identification module to optimize feature weights. Effective plans are added to the expert knowledge base, and the post-intervention baseline adjustment rules of the health baseline construction module are updated synchronously.

[0015] Preferably, the operation and maintenance module receives the equipment linkage data from the early warning response module and the execution feedback data from the decision-making module. Under the premise of multi-base data privacy protection, federated learning is used to achieve model evolution. The data from the health identification module and health baseline construction module of each base are combined to update the AI ​​identification model and baseline model parameters every quarter. After the update, the parameters are pushed to the corresponding module. Each base only uploads the model parameters and optimizes the performance through parameter aggregation. The operating status of the sensors, drones and edge computing terminals of the acquisition module is monitored by the equipment heartbeat detection. When a fault occurs, the location is automatically marked and a maintenance reminder is pushed to ensure that the acquisition module and data preprocessing module have no data interruption. The equipment wear prediction model is established based on the usage time and environmental factors to provide early warning of equipment replacement needs 15 days in advance. The specific construction details of the equipment wear prediction model are as follows: (1) Model type: The gradient boosting tree (XGBoost) model is adopted, which is suitable for different equipment types such as sensors, drones and computing terminals. (2) Input feature dimensions: There are 12 features in total, including the equipment usage time. Long (month), average daily working hours (hours), ambient humidity (%RH), ambient temperature (°C), number of historical faults, data acquisition error rate (%), vibration frequency (Hz), power supply voltage stability (V), ammonia concentration (ppm), cleaning and maintenance cycle (days), equipment model weight (assigned according to equipment accuracy, such as infrared thermal imager weight 1.2, temperature and humidity sensor weight 1.0), cumulative data transmission volume (GB); (3) Output and threshold: output "remaining service life (days)", when the predicted remaining service life is ≤15 days, trigger equipment replacement warning; (4) Model training: training data comes from 3 years of equipment operation data from 5 giant panda bases (a total of 2000+ equipment operation records), the ratio of training set to validation set is 8:2, the tree depth is set to 6, the learning rate is 0.05, the number of iterations is 200, and the model prediction accuracy target is ≥90%; (5) Update mechanism: the model is fine-tuned every quarter in combination with the newly added equipment operation data (monthly records of ≥100 equipment) to ensure that the prediction accuracy is not lower than the initial value.

[0016] The system employs AES-256 encryption to transmit collected data and model parameters, protecting individual monitoring data from the data acquisition module and model data from the health identification module. It uses blockchain to store giant panda health records. Permissions are assigned based on roles: keepers can view basic data from the data acquisition module, while veterinarians can view complete health data from the health baseline construction module. An operation log auditing function is implemented. Maintenance data is fed back to the data acquisition module, adjusting the sampling rate of backup sensors during periods of failure to ensure data continuity. Early warning and traceability data are fed back to the health identification module.

[0017] The beneficial effects of this invention are as follows: 1. This invention adopts a three-source acquisition architecture of air, ground, and environment, achieving 24-hour uninterrupted data acquisition through devices such as micro drones, infrared thermal imagers, and smart feeding bowls. All acquired data is stamped with a time and space stamp and undergoes multi-source cross-verification to ensure data accuracy. Simultaneously, relying on edge computing and a hierarchical priority transmission mechanism, abnormal situations can quickly trigger local alarms. Combined with a four-level early warning mechanism, the anomaly level and possible causes can be clearly identified in a timely manner, effectively solving the problem of time gaps in traditional monitoring and significantly improving the response speed to sudden anomalies.

[0018] 2. This invention integrates physiological, behavioral, and environmental data through a dual-model architecture for causal reasoning, and combines it with a scenario-based behavior recognition algorithm to improve the adaptability of behavior judgment. At the same time, it constructs a main baseline and sub-baselines based on the panda's age, physiological cycle, and other dimensions, and records the unique behavior patterns of each panda, making health judgment more in line with the actual situation of the individual, significantly reducing misjudgments and omissions of health abnormalities, and reducing the cost of ineffective intervention.

[0019] 3. This invention integrates veterinary experience to construct a multi-dimensional expert knowledge base. Combining the individual panda's health baseline and real-time health status, it outputs a combination of solutions including dietary adjustments, medication recommendations, and environmental optimization through reinforcement learning. Instructions can be precisely issued to the corresponding modules for execution. In the operation and maintenance phase, while protecting data privacy, federated learning optimizes model performance. Simultaneously, it conducts equipment status monitoring, wear prediction, and encrypted storage of health records, forming a complete closed loop of collection, identification, early warning, decision-making, and optimization. This promotes the transformation of captive panda health management from reliance on manual labor to a more refined and intelligent approach. Attached Figure Description

[0020] Figure 1 is a flowchart of the intelligent captive giant panda health monitoring system of the present invention. Detailed Implementation

[0021] The technical solutions of 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.

[0022] As shown in Figure 1, this embodiment of the invention provides an intelligent health monitoring system for captive giant pandas, including a data acquisition module, a data preprocessing module, a health identification module, a health baseline construction module, an early warning response module, a decision-making module, and an operation and maintenance module. The specific implementation of each module is as follows: The data acquisition module adopts a three-source acquisition architecture of air, ground, and environment. In the air layer, a micro-UAV with a fuselage of <30cm and equipped with a single-photon lidar is deployed to achieve real-time positioning and activity trajectory mapping of giant pandas within a 50m range under the forest canopy, adapting to captive breeding area terrain with a slope of ≤30°. In the ground layer, camouflaged monitoring units simulating dead wood or rocks are set up, integrating an infrared thermal imager with a sampling rate of 10 frames / second and a millimeter-wave radar with a detection range of 5-20m, simultaneously capturing data through a 4K panoramic camera. The system captures typical behaviors such as feeding and estrus. The environment and diet linkage layer deploys temperature, humidity, and ammonia sensors (sampling once every 30 seconds) and smart feeding bowls with a weight accuracy of ±2g in the enclosures. All data collected is appended with millisecond-level timestamps and centimeter-level spatial coordinates, providing a spatiotemporal correlation basis for dynamic compression at the edge. Cross-validation of data from infrared thermal imagers and millimeter-wave radar, and matching of data from smart feeding bowls and 4K cameras, ensures that the data transmitted to the edge is free of abnormal deviations. The edge integrates an analysis unit based on Granger causality testing, using a 1-hour sliding window to mine causal relationships between data and add labels, providing causal inference material for the Bayesian network of the health identification module and structured raw data for the data preprocessing module.

[0023] The specific implementation method and judgment threshold of the multi-source data mutual verification are as follows: (1) Cross-verification between infrared thermal imager and millimeter-wave radar: The “panda body temperature value” output by infrared thermal imager is correlated with the “activity status (stationary / moving)” output by millimeter-wave radar. The rule is that “the body temperature fluctuation in the stationary state should be ≤0.5℃ / 10 minutes, and the body temperature fluctuation in the moving state should be ≤1℃ / 10 minutes”. If the threshold is exceeded, the data in that time period is marked as “suspicious” and the average of the two is used to complete the data; (2) Data matching between smart food bowl and 4K panoramic camera: The "weight change timestamp" of the feeding bowl is compared with the "feeding behavior frame timestamp" of the 4K camera. The deviation is allowed to be ≤3 seconds. If the deviation is >3 seconds, the 4K camera frame is used first for judgment (if the camera captures "chewing action", the feeding bowl data is retained; if it does not capture, the feeding bowl data for that period is removed); (3) Mutual verification pass rate judgment: the mutual verification pass rate of a single device's single data collection must be ≥90%. Otherwise, the device self-test is triggered (such as recalibrating the infrared thermal imager or resetting the radar parameters). If the pass rate is still <85% after the self-test, a maintenance reminder is pushed.

[0024] The data preprocessing module receives multi-source data from the acquisition module and adopts an edge computing architecture. The acquisition terminal has a built-in edge computing chip with a computing power of ≥2 TOPS. It removes video noise by improving the VIBE background modeling algorithm and uses Kalman filtering to repair momentarily missing sensor data, ensuring that the repaired data maintains spatiotemporal consistency. A two-stage processing mechanism of feature extraction and redundancy removal is constructed to extract behavioral keyframes and physiological feature values, reducing the amount of invalid data processed by the health recognition module and retaining only structured feature data in JSON format. A two-way communication is established with the health recognition module to receive high-value feature types, such as respiratory data during pregnancy, and to achieve feedback-based dynamic compression. When high-value data is identified, the compression ratio is reduced from 10:1 to 5:1, while the compression ratio is increased for data from nighttime when there is no panda activity. A scene keyframe extraction library is established, with classification based on typical behaviors captured by the acquisition module to ensure the extraction of keyframes. Matching with the behavior sample library of the health recognition module; the specific parameters of the improved VIBE background modeling algorithm and Kalman filter are as follows: (1) Improved VIBE algorithm parameters: the initial background model sampling frame number is set to 20 frames, the matching threshold is set to 20, the update factor is set to 0.05, the dynamic background is given a pixel-level dynamic weight, the tree branch area matching threshold is increased to 30, the panda activity area is kept at 20, and the background false detection is reduced; (2) Kalman filter parameters: the state vector dimension is set to 2, which is the sensor measurement value and measurement error, the observation matrix H is set to [1, 0], the state transition matrix F is set to [1, 0.1; 0, 1], the process noise covariance Q is set to [0.01, 0; 0, 0.01], and the observation noise covariance R is dynamically adjusted according to the sensor type, the temperature and humidity sensor R=0.5, the ammonia sensor R=0.8, and the smart food bowl weight sensor R=0.2; (3) Edge computing chip adaptation: NVIDIA is used. The Jetson Nano 2GB, through TensorRT quantization during algorithm deployment, increases the frame rate for video noise removal to 25fps, meeting the real-time processing requirements of 4K cameras.

[0025] A hierarchical priority transmission protocol is adopted, with a data transmission delay of ≤100ms for abnormal behavior, supporting the rapid response of the early warning response module; regular data is transmitted in batches; a simplified YOLOv9 model is integrated at the edge, and its pre-trained samples share part of the dataset with the health recognition module. When high-risk anomalies are identified, local audible and visual alarms are triggered directly, and the original data is transmitted to the cloud with priority.

[0026] The specific structure of the simplified YOLOv9 model is as follows: remove the two C2f modules of the original YOLOv9, retain the core feature extraction network, reduce the input image size to 480×480×3, and retain only the high-risk abnormal behavior categories (falling to the ground, refusing to eat, and severe convulsions) in the detection head. The output dimension is (batch_size, 4000, 4+1+3). Training parameters: the pre-training samples share 60% of the dataset with the health recognition module, an additional 20,000 high-risk abnormal behavior samples are added, the learning rate is 0.005, the number of iterations is 150 rounds, the confidence threshold is set to 0.7, and the local alarm trigger condition is that high-risk abnormal behavior is detected for 3 consecutive frames and the confidence level is ≥0.8.

[0027] The health recognition module receives structured data processed by the data preprocessing module and employs a dual-model architecture causal inference fusion model combining LSTM and Bayesian networks. The LSTM network analyzes the temporal changes of physiological indicators such as body temperature and respiratory rate within a 2-hour time window. The analysis process is based on the physiological feature values ​​output from the edge, combined with the causal labels from the acquisition module. The Bayesian network infers the root causes of health abnormalities and outputs the correlation degree of abnormality causes based on the causal label data from the acquisition module. Based on the improved YOLOv9 algorithm with an attention mechanism, a scene-based behavior recognition algorithm is constructed. Its training set covers behavioral samples from over 500 individuals, including differentiated behaviors of elderly and young pandas. The sample annotation references the individual behavior database initially established by the health baseline construction module. A dynamic adjustment mechanism for scene and feature weights is constructed, for example, increasing the weight in cold scenes. The normal weights for high curling behavior are adjusted by referencing the environmental parameters of the data acquisition module and the scenario-specific sub-baselines of the health baseline construction module to ensure that the weight adjustments conform to individual health benchmarks. The specific connection relationship between the LSTM and Bayesian network is a serial fusion architecture: the output of the LSTM network (the feature vector of time-series changes in physiological indicators, with a dimension of 64) serves as one of the input layer nodes of the Bayesian network, which together with the causal labels (with a dimension of 16) and environmental parameter features (with a dimension of 8) from the data acquisition module constitute the input layer of the Bayesian network (a total of 88 nodes). The Bayesian network contains one hidden layer (32 nodes), and the hidden layer and the output layer (the correlation vector of abnormal causes, with a dimension of 12) are fully connected. The conditional probabilities between nodes are calculated using the maximum likelihood estimation method, and the initial probability matrix is ​​obtained by training on more than 1,000 clinical case data.

[0028] The specific steps of dual-model training are as follows: (1) Data partitioning: The ratio of training set, validation set and test set is 7:2:1. The data comes from 180 days of continuous monitoring data of 500+ individuals, including samples of different ages and physiological cycles; (2) LSTM network training: The number of hidden layers is 3, and the number of neurons in each layer is 128, 64 and 32 respectively. The activation function is ReLU, the dropout rate is set to 0.3, the optimizer is Adam, the initial learning rate is 0.001, the learning rate is adjusted by cosine annealing strategy, the number of iterations is 200 rounds, and the convergence criterion is that the validation set loss function (MSE) is lower than 0.005 for 10 consecutive rounds; (3) Bayesian network training: The conditional probability matrix is ​​optimized by EM algorithm based on the training set data, the number of iterations is 50 rounds, and the convergence threshold is 0.001; (4) Fusion training: The LSTM training results and the Bayesian network training results are fused by weighted summation (the weights are 0.6 and 0.4 respectively). The weight coefficients are adjusted based on the test set to ensure that the accuracy of abnormal cause identification is ≥92%.

[0029] The specific improvement scheme of the improved YOLOv9 algorithm is as follows: a spatial attention mechanism is embedded after the C2f module of YOLOv9, and the extraction of key behavioral features is enhanced through channel attention branch and spatial attention branch; the network input is a 640×640×3 RGB image frame, the feature extraction network retains the C2f and SPPF modules of YOLOv9, and the output dimension of the detection head is (batch_size, 8400, 4+1+12), where 4 is the bounding box coordinates, 1 is the confidence score, and 12 is the behavior category, including typical behaviors such as eating, estrus, and falling down.

[0030] Specific logic for scenario-based adaptation: (1) Input data setting: After the environmental parameters are normalized, they are used as the weight adjustment factor of the attention mechanism. For example, when the temperature is <10℃, the weight coefficient of the spatial attention branch on the curling behavior area is increased to 1.5 times; (2) Feature weight dynamic adjustment rules: Based on the sub-baseline of the health baseline construction module, a scenario-weight mapping relationship is established. Specifically, in the low temperature environment scenario, the weight ratio of behavioral features is 30% for eating, 20% for activity, and 50% for resting; in the feeding period scenario, the weight ratio of behavioral features is 60% for eating, 10% for activity, and 30% for resting; in the breeding period scenario, the weight ratio of behavioral features is 20% for eating, 60% for activity, and 20% for resting. The environmental parameters of the acquisition module are read in real time to match the corresponding scenario and the feature weights are adjusted; (3) Training parameters: Mosaic data augmentation is used, the initial learning rate is 0.01, the optimizer is SGD, the number of iterations is 300 rounds, the IOU threshold is set to 0.5, and the target of behavior recognition accuracy is ≥95%.

[0031] The system integrates three types of features: spatiotemporal correlation data from the acquisition module, causal label data, and environmental parameters. Weights are assigned according to the scenario, with the smart feeding bowl data accounting for 40% in the dietary scenario. An evaluation index system is constructed, with index dimensions referencing the physiological and behavioral classifications of the health baseline construction module. A weighted voting algorithm is used to output a health status vector of 0-100 points; 85-100 points are normal, 70-84 points are generally abnormal, 50-69 points are urgent abnormal, and <50 points are critical abnormal.

[0032] The specific weight allocation rules of the weighted voting algorithm are as follows: (1) Basic weight setting: The default weights of the three types of fusion features are “spatiotemporal correlation data (40%), causal label data (35%), and environmental parameters (25%)”, with a total weight of 1; (2) Scenario-based weight adjustment: Dietary scenario: The weight of spatiotemporal correlation data (including smart food bowl data) is increased to 50%, the weight of environmental parameters is reduced to 20%, and the weight of causal label data remains at 30%; Reproduction scenario: The weight of causal label data (including estrus behavior correlation labels) is increased to 45%, the weight of spatiotemporal correlation data is reduced to 35%, and the weight of environmental parameters remains at 20%; Disease suspected scenario: The weight of environmental parameters (including ammonia concentration, temperature and humidity) is increased to 35%, the weight of spatiotemporal correlation data is reduced to 35%, and the weight of causal label data remains at 30%; (3) Voting calculation method: Each type of feature outputs a score (0-100 points) according to the “deviation degree between feature value and baseline”, and the final health status vector = (spatiotemporal correlation data score × corresponding weight) + (causal label data score × corresponding weight) + (environmental parameter score × corresponding weight), and the result is retained to one decimal place.

[0033] Among them, the health baseline construction module receives the health status data and abnormal cases of the health identification module. Based on 6 months of continuous monitoring data, the daily effective data is ≥500 pieces, which are from the historical data of the collection module and the edge module. The improved K-means clustering algorithm is adopted, and the dimensions are divided according to age, physiological cycle and activity scene. The scene classification refers to the behavior and environmental data types of the collection module. The main baseline and sub-baseline are generated. For example, the main baseline of adult males in non-estrus period is body temperature 36.2±0.3℃, and the sleep sub-baseline is body temperature 35.9±0.2℃ and respiratory rate 22±2 times / minute. This baseline is used as the benchmark for the health identification module to judge abnormalities. For example, when the health identification module judges whether a body temperature of 37.5℃ is abnormal, it needs to compare with the corresponding sub-baseline. The specific parameters and steps of the improved K-means clustering algorithm are: (1) Clustering dimension setting: The input feature dimension is 24 dimensions, including physiological indicators (body temperature, respiratory rate, etc. 8 dimensions), behavioral indicators (eating time, activity level, etc. 10 dimensions), and environmental adaptation indicators (6 dimensions); (2) K value determination: The elbow method is used. Then, combining the silhouette coefficient method, the K value is determined for different dimension combinations: age dimension K=4 (childhood, sub-adulthood, adulthood, old age), physiological cycle dimension K=3 (non-estrus period, estrus period, pregnancy period), activity scenario dimension K=5 (eating, resting, activity, reproduction, medical treatment); (3) Improvement points: In view of the defect of random initial cluster centers in traditional K-means, density peak clustering (DPC) algorithm is used to pre-select the initial centers to ensure that the cluster centers are evenly distributed; (4) Distance measurement: weighted Euclidean distance is used, and physiological indicators The weight is 0.5, the weight of the behavior index is 0.3, and the weight of the environment adaptation index is 0.2; (5) Training steps: ① Data standardization (Z-score standardization, μ=0, σ=1); ② Pre-select initial cluster centers based on DPC; ③ Iteratively update cluster centers (iteration times ≤ 100 rounds); ④ Convergence judgment (cluster center offset < 0.001); (6) Baseline threshold setting: The main baseline is the mean of each cluster ± 1.5 times the standard deviation, and the sub-baseline is the mean of each cluster in the subdivided scenario ± 1.2 times the standard deviation.

[0034] A 14-day sliding window algorithm is used to incorporate the latest structured data from the data preprocessing module in real time. The baseline threshold is adjusted based on the intervention effects reported by the decision-making module, such as the recovery of body temperature after drug intervention. For example, after a panda's digestive tract disease was cured, the baseline for "fecal characteristics" was adjusted from "formed → soft stool" to "formed," and the "food intake fluctuation threshold" was reduced from ±10% to ±5%. Baseline deviations are corrected quarterly based on veterinary examination data, and the correction results are simultaneously fed back to the health identification module to update the abnormality judgment criteria. Each panda's unique behavioral patterns are recorded, such as rubbing against walls at 3:00 PM and prioritizing bamboo shoot tips. These long-term behavioral records from the data collection module are incorporated into the baseline judgment criteria to eliminate "misjudgments based on specific behaviors" for the health identification module and to provide a basis for the decision-making module to formulate plans that do not affect the panda's habits.

[0035] The early warning response module receives the health status vector and abnormal cause from the health identification module, establishing a four-level graded early warning mechanism: 85-100 points indicate normal status with no warning; 70-84 points indicate a general warning, pushing an APP reminder containing data trend charts and suspected causes; 50-69 points indicate an emergency warning, triggering SMS + animal enclosure audible and visual alarms, with additional intervention suggestions; <50 points indicate a critical warning, triggering the veterinary emergency dispatch system, simultaneously pushing abnormal data and preliminary treatment plans. The warning content includes abnormal indicators, suspected causes, and correlation, directly referencing the causal reasoning results from the health identification module. For example, a body temperature of 37.5℃, suspected cause: environmental ammonia concentration of 25ppm, causal correlation of 89%, provides attribution basis for the decision-making module; a scenario-based intelligent linkage strategy is matched according to the warning cause, with the abnormal cause being temperature... When humidity changes abruptly, the air conditioning is automatically adjusted, and the temperature deviation is controlled within ±0.5℃. The adjustment command is sent to the enclosure equipment managed by the operation and maintenance module. When the cause of the abnormality is insufficient food intake, the smart feeder adjusts the bamboo shoot ratio. The feeder control signal comes from the diet linkage layer of the acquisition module. When the cause of the abnormality is suspected disease, the isolation door is automatically closed, and the door control status is fed back to the operation and maintenance module for recording. The system automatically records the warning trigger time, response operation, and handling result to form a traceability log, providing cases for the operation and maintenance module to optimize equipment linkage strategies and for the health recognition module to adjust feature weights. False and missed warning cases are statistically analyzed monthly. The statistical basis is used to refer to the handling effect fed back by the decision-making module, and the health score threshold is adjusted accordingly. For example, when the false warning rate of elderly pandas is high, the general warning threshold is lowered from 70 points to 68 points, and the adjustment result is synchronized to the health recognition module.

[0036] The decision-making module takes over the abnormal causes and health status from the early warning and response module, integrates the veterinary experience from 10 giant panda bases, and constructs a multi-dimensional expert knowledge base. Case classification references the abnormality types from the health identification module and the individual dimensions from the health baseline construction module, storing treatment plans according to disease type, individual characteristics, and abnormal causes. For example, for digestive tract abnormalities in pregnant adult females, the cause is insufficient dietary fiber; the corresponding solution is to increase the proportion of bamboo fiber combined with daily probiotic supplementation. A reinforcement learning decision-making model is constructed with the goal of achieving optimal intervention effect. The training data includes the individual baseline from the health baseline construction module and the abnormal causes from the early warning and response module. Combining the individual health baseline, real-time health status, and abnormal causes, the model outputs dietary adjustments. The system combines drug recommendations with environmental optimization. Dietary adjustment commands are sent to the smart food bowl in the data collection module, while environmental optimization commands are sent to the enclosure equipment in the operation and maintenance module. Drug recommendations include dosage ranges; for example, the antibiotic dosage for adult pandas is calculated at 20 mg / kg of body weight, referencing individual weight data from the baseline module. Changes in health indicators are recorded within 72 hours after intervention, with data sourced from real-time monitoring by the data collection and preprocessing modules. If certain pre-intervention effects are not achieved, such as body temperature not returning to the baseline range within 48 hours, the system automatically adjusts the plan, feeding back the adjustment logic to the health recognition module and optimizing feature weights. Effective plans are added to the expert knowledge base, and the post-intervention baseline adjustment rules in the health baseline construction module are updated synchronously.

[0037] Among them, the operation and maintenance module receives the equipment linkage data of the early warning response module and the execution feedback data of the decision module. Under the premise of multi-base data privacy protection, federated learning is used to realize model evolution; the data of the health identification module and health baseline construction module of each base are combined, and the AI ​​identification model and baseline model parameters are updated every quarter. After the update, the parameters are pushed to the corresponding modules to improve the accuracy of identification and baseline; each base only uploads model parameters and optimizes performance through parameter aggregation; the operation status of the sensors, drones and edge computing terminals of the acquisition module is monitored by the equipment heartbeat detection every 30 seconds. When a fault occurs, the location is automatically marked and a maintenance reminder is pushed to ensure that the acquisition module and data preprocessing module have no data interruption; the equipment wear prediction model is established based on the usage time and environmental factors (such as the influence of humidity on the sensor, and the data comes from the environmental parameters of the acquisition module) to warn of equipment replacement needs 15 days in advance and ensure the data acquisition accuracy of the acquisition module; the federated learning adopts a horizontal federated learning architecture. The specific implementation steps are: (1) Initialization: each base deploys a local model with the same structure (health identification module) (1) LSTM-Bayes fusion model, the healthy baseline building module is an improved K-means model), the global model parameters are initialized to the average parameters of the local models of each base; (2) Local training: each base trains the local model based on the local dataset (≥100,000 structured data), the training rounds are 50 rounds, the optimizer is AdamW (weight decay 0.0001), the learning rate is 0.0005; (3) Parameter upload: each base only uploads the parameter update amount (gradient information) of the local model, and transmits it to the federated learning server after encryption by AES-256, the upload frequency is once per quarter; (4) Global aggregation: the server uses the FedAvg algorithm to aggregate parameters, the aggregation weight is the proportion of data volume of each base (data volume weight = base data volume / total data volume of all bases); (5) Model update: the aggregated global model parameters are distributed to each base, covering the local model parameters, and the model iteration is completed; (6) Privacy protection: differential privacy technology is used, and Laplacian noise (ε=1.5) is added when the parameters are uploaded to ensure that the risk of data privacy leakage is ≤0.01.

[0038] AES-256 encryption is used to transmit collected data and model parameters to protect individual monitoring data of the collection module and model data of the health identification module; the health records of giant pandas are stored through blockchain to ensure that they cannot be tampered with; permissions are assigned according to roles, with keepers viewing basic data of the collection module and veterinarians viewing complete health data of the health baseline construction module; an operation log audit function is set up to prevent data leakage or tampering; the specific implementation details of the blockchain storage are as follows: (1) Blockchain type: a consortium blockchain architecture is adopted, and the nodes include major panda bases and regulatory agencies. Node access requires identity authentication; (2) Storage data format: each The health record data includes panda ID, timestamp, health status vector, abnormal cause and intervention record, and is stored in JSON format. The data hash value is uploaded to the blockchain, and the original data is encrypted and stored on the local server of each base; (3) Consensus mechanism: The practical Byzantine fault tolerance algorithm is adopted, the number of consensus nodes is set to 3, and the consensus delay is controlled within 500ms to ensure the real-time data upload to the blockchain; (4) Record query and modification: When the breeder / veterinarian queries, he / she needs to pass the role permission and secondary verification. The health record only allows the addition of records and does not allow the deletion / modification of historical data. The modification operation needs to be approved by more than 3 consensus nodes.

[0039] The operation and maintenance data is fed back to the acquisition module, and the sampling rate of the backup sensor during the fault period is adjusted to ensure data continuity; the early warning and traceability data is fed back to the health identification module, such as the feature weight corresponding to false early warning cases is reduced.

[0040] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent health monitoring system for captive giant pandas, characterized in that, It includes a data acquisition module, a data preprocessing module, a health identification module, a health baseline construction module, an early warning response module, a decision-making module, and an operation and maintenance module. The data acquisition module uses a three-source architecture to collect data, adds spatiotemporal stamps to the data and performs multi-source cross-verification, and generates labels through causal analysis. The data preprocessing module removes video noise and repairs missing data, extracts key frames of behavior, and performs hierarchical transmission and dynamic compression of data. The health identification module uses a dual-model architecture to perform causal inference, combines scenario-based behavior recognition and multi-feature fusion to generate a health status vector of 0-100 points. Health baseline construction module: Establishes a baseline that is adaptable to multiple scenarios and can evolve dynamically. It generates a main baseline and sub-baselines through clustering, updates data based on a sliding window, optimizes the baseline by combining intervention effect feedback and physical examination data, and records panda-specific behaviors. Early warning response module: Establishes a four-level early warning mechanism. The early warning content includes abnormal indicators, causes, and correlations. It matches scenario-based device linkage strategies and records traceability logs to optimize thresholds. Decision module: Generates solutions based on knowledge base and reinforcement learning. It sends instructions to the collection module and operation and maintenance module for execution. It records the intervention effect and adjusts the solution. It provides effect feedback to the health baseline construction module and feature change materials to the health identification module. Operation and maintenance module: Provides equipment support and parameter optimization for the aforementioned modules through federated learning to optimize models, monitor equipment and protect security.

2. The intelligent captive giant panda health monitoring system according to claim 1, characterized in that, The data acquisition module adopts a three-source acquisition architecture: air, ground, and environment. The air layer deploys micro-drones to locate giant pandas and map their activity trajectories. The ground layer is equipped with a monitoring unit that integrates an infrared thermal imager, millimeter-wave radar, and a 4K panoramic camera to capture typical behaviors, including feeding and estrus. The environment and diet linkage layer deploys temperature and humidity sensors, ammonia sensors, and a smart feeding bowl. All data from the acquisition devices are appended with timestamps and spatial coordinates. Cross-validation of data from the infrared thermal imager and millimeter-wave radar, as well as matching of data from the smart feeding bowl and the 4K panoramic camera, enables mutual verification of multi-source data. The edge-end integrates a causal verification and analysis unit to mine causal relationships between data and add labels, providing causal reasoning material for the health identification module and structured raw data for the data preprocessing module.

3. The intelligent captive giant panda health monitoring system according to claim 2, characterized in that, The data preprocessing module adopts an edge computing processing architecture and has a built-in edge computing chip. It removes video noise by improving the VIBE background modeling algorithm and uses Kalman filtering to repair missing sensor data. A two-stage processing mechanism of feature extraction and redundancy removal is constructed to extract behavioral keyframes and physiological feature values, while retaining structured feature data in JSON format; Two-way communication is established with the health recognition module, and the compression ratio is dynamically adjusted according to the high-value feature type; a scene key frame extraction library is established to ensure matching with the behavior sample library of the health recognition module; a hierarchical priority transmission protocol is adopted, and abnormal behavior data is transmitted first; a simplified YOLOv9 model is integrated at the edge, and its pre-trained samples share part of the dataset with the health recognition module. When high-risk anomalies are identified, local sound and light alarms are directly triggered and the original data is transmitted to the cloud first.

4. The intelligent captive giant panda health monitoring system according to claim 3, characterized in that, The health identification module adopts a dual-model architecture causal reasoning fusion model combining LSTM and Bayesian network. The LSTM network analyzes the temporal changes of physiological indicators and combines them with the causal labels from the acquisition module; the Bayesian network infers the root cause of health abnormalities based on the causal label data and outputs the correlation degree of the abnormality cause. A scenario-based behavior recognition algorithm is constructed based on the improved YOLOv9 algorithm with an attention mechanism. The training set covers multiple individual behavior samples. A dynamic adjustment mechanism for scenario and feature weights is constructed, and the sub-baselines of the module are constructed with reference to the environmental parameters and health baseline of the acquisition module. It integrates three types of features: spatiotemporal correlation data, causal label data, and environmental parameters, and assigns weights according to scenarios; An evaluation index system is constructed, referencing the physiological and behavioral classifications of the health baseline construction module, and a health status vector of 0-100 points is output through a weighted voting algorithm.

5. The intelligent captive giant panda health monitoring system according to claim 4, characterized in that, The health baseline construction module uses an improved K-means clustering algorithm to divide dimensions according to age, menstrual cycle, and activity scenario, generating a main baseline and sub-baselines as the benchmark for the health identification module to determine abnormalities; it also uses a sliding window algorithm to incorporate the latest structured data from the data preprocessing module in real time. Adjust the baseline threshold based on the intervention effect feedback from the decision-making module; Baseline deviations are corrected regularly based on physical examination data, and the correction results are simultaneously fed back to the health recognition module; the unique behavioral patterns of each panda are recorded and incorporated into the baseline judgment criteria.

6. The intelligent captive giant panda health monitoring system according to claim 5, characterized in that, The four-level early warning mechanism of the early warning response module is as follows: 85-100 points is normal, 70-84 points is general warning, 50-69 points is emergency warning, and <50 points is critical warning. The warning content includes abnormal indicators, suspected causes, and correlation, referencing the causal reasoning results of the health recognition module. Based on the warning cause, a scenario-based intelligent linkage strategy is matched: the air conditioner is automatically adjusted when the temperature and humidity change suddenly; the intelligent feeding bowl adjusts the food ratio when the food intake is insufficient; and the isolation door is automatically closed when a disease is suspected. Relevant instructions are sent to the corresponding module's control equipment. The warning trigger time, response operation, and handling result are automatically recorded to form a traceability log, providing a case for the operation and maintenance module to optimize equipment linkage strategies and for the health recognition module to adjust feature weights. Regularly analyze cases of false and missed warnings, adjust the health score threshold accordingly, and synchronize it to the health identification module.

7. The intelligent captive giant panda health monitoring system according to claim 6, characterized in that, The decision-making module constructs a multi-dimensional expert knowledge base, storing treatment plans categorized by disease type, individual characteristics, and causes of abnormalities. It builds a reinforcement learning decision model with the goal of achieving optimal intervention results, combining individual health baselines, real-time health status, and causes of abnormalities to output a combined plan of dietary adjustments, medication recommendations, and environmental optimization. Dietary adjustment instructions are sent to the data collection module, and environmental optimization instructions are sent to the operation and maintenance module. Changes in health indicators after intervention are recorded; for cases where expected results are not met, the plan is automatically adjusted, and the adjustment logic is fed back to the health identification module. Effective plans are added to the expert knowledge base, and the post-intervention baseline adjustment rules of the health baseline construction module are updated synchronously.

8. The intelligent captive giant panda health monitoring system according to claim 7, characterized in that, Under the premise of protecting data privacy across multiple bases, the operation and maintenance module adopts federated learning to jointly update the AI ​​recognition model and baseline model parameters with data from each base; each base only uploads model parameters; the module monitors the operating status of sensors, drones and edge computing terminals through device heartbeat detection, automatically marks the location and pushes maintenance reminders when a fault occurs; and establishes a device wear prediction model based on usage time and environmental factors. The system uses AES-256 encryption to transmit collected data and model parameters, and stores giant panda health records on blockchain. It assigns permissions based on roles and sets up an operation log audit function. It feeds back operation and maintenance data to the acquisition module to adjust the sensor sampling rate, and feeds back early warning and traceability data to the health identification module.

9. The intelligent captive giant panda health monitoring system according to claim 8, characterized in that, The multi-source data cross-verification of the acquisition module uses the 3σ criterion to determine outliers. When the infrared thermal imager and millimeter-wave radar are cross-verified, the body temperature fluctuation in a static state must be ≤0.5℃ / 10 minutes. The data timestamp deviation between the smart feeding bowl and the 4K panoramic camera is allowed to be ≤3 seconds.

10. The intelligent captive giant panda health monitoring system according to claim 9, characterized in that, In the sliding window algorithm of the health baseline construction module, the window data filtering needs to remove outliers that deviate from the daily average by ±3 times the standard deviation, and only retain date data with ≥80% of the effective data volume. The standard deviation coefficient is dynamically adjusted with the data variation coefficient during baseline calculation.

Images