Full-automatic wild mammal health monitoring and disease early warning system

Abstract

The invention discloses a full-automatic wild mammal health monitoring and disease early warning system, and belongs to the technical field of ecological monitoring. The system comprises a data acquisition module which is configured to be capable of acquiring behavior, physiology and living environment data of mammals; the data processing module is configured to perform corresponding preprocessing on the acquired different types of data, so that the data can meet the requirements of subsequent data analysis and / or early warning; the data analysis module is configured to identify and analyze the preprocessed data to obtain the type, position, movement speed, attitude information and health information of the mammal; and the disease early warning module is configured to predict the disease risk probability of the mammals and trigger corresponding early warning according to the prediction result and the grading early warning rule. According to the invention, through multi-source data acquisition, intelligent analysis and real-time early warning, the problems of data islands, strong manual dependence, response delay and the like in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of ecological monitoring, and particularly relates to a full-automatic wild mammal health monitoring and disease early warning system. Background Art

[0002] Wild mammals are important indicators of the balance of the ecosystem, and their health status directly affects biodiversity and ecological stability. At present, the mainstream monitoring methods have the following limitations:

[0003] 1. Strong dependence on manual labor: Traditional methods mainly rely on field observations or physical examinations after capture, which require a large amount of manpower and are inefficient, and it is difficult to cover vast or complex terrained habitats.

[0004] 2. Fragmented data: Existing sensor technologies (such as GPS collars) can only collect location or basic physiological data (such as heart rate), lacking synchronous analysis of behavior patterns, sound characteristics, and environmental parameters.

[0005] 3. Lack of real-time performance: Existing systems mostly use periodic data transmission, and cannot achieve instant identification and early warning of abnormal behaviors, resulting in a lag in response when diseases break out.

[0006] 4. Low level of intelligence: Data analysis relies on manual experience, lacking automated models to predict potential disease risks, and it is easy to miss detections or make misjudgments.

[0007] For example, although infrared cameras can capture images of animal activities, the memory card data needs to be manually checked regularly; sound sensors can record animal vocalizations, but they are not associated with behavior or environmental data for analysis. In addition, existing early warning systems are mostly based on static rules and cannot adapt to the dynamic changes of wild animal behaviors, making the early warning results inaccurate. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems in the above related technologies to a certain extent.

[0009] To this end, the object of the present invention is to provide a full-automatic wild mammal health monitoring and disease early warning system, which solves the problems of data islands, strong dependence on manual labor, and response delay in the prior art through multi-source data collection, intelligent analysis, and real-time early warning, and provides a dynamic and scientific management tool for wild animal protection.

[0010] To solve the above technical problems, the present invention is implemented as follows:

[0011] An embodiment of the present invention provides a full-automatic wild mammal health monitoring and disease early warning system, and the system includes:

[0012] A data collection module configured to be able to collect the behavior, physiology, and living environment data of mammals;

[0013] A data processing module, configured to be able to perform corresponding preprocessing on different types of collected data so that it can meet the requirements of subsequent data analysis and / or early warning; the preprocessing is any one or more of cleaning, normalization, fusion, and feature extraction;

[0014] A data analysis module, configured to be able to identify and analyze the preprocessed data to obtain the species, location, movement speed, posture information, and health information of mammals; and,

[0015] A disease early warning module, configured to be able to predict the probability of disease risk of mammals and trigger corresponding early warnings according to the prediction results and the classification early warning rules.

[0016] In addition, the full-automatic wild mammal health monitoring and disease early warning system according to the present invention may further have the following additional technical features:

[0017] In some embodiments, the data acquisition module includes a multi-type sensor network; the multi-type sensor network includes a plurality of infrared cameras, a plurality of sound sensors, and a plurality of environmental sensors.

[0018] In some embodiments, the preprocessing of the images acquired by the infrared cameras by the data processing module includes: performing normalization processing on the acquired original images, adjusting the image resolution to a preset standard resolution, and enhancing the contrast under low light conditions by using histogram equalization to obtain the preprocessed infrared image data.

[0019] In some embodiments, the preprocessing of the data processing module for the sound data includes voiceprint feature extraction, and the extraction content includes: performing frame division and windowing, FFT transformation, and cepstrum analysis on the sound data to obtain MFCC feature vectors.

[0020] In some embodiments, in the frame division and windowing, the frame division standard is: the frame length is 25 ms, the frame shift is 10 ms, and the windowing is Hamming windowing;

[0021] The content of the FFT transformation includes: calculating the spectrum of each frame and then mapping it to the Mel scale through a Mel filter bank;

[0022] The content of the cepstrum analysis includes: performing DCT transformation on the logarithmic energy and taking the first 13 coefficients as the MFCC feature vectors.

[0023] In some embodiments, the data analysis module uses a pre-trained YOLOv5s model for target detection for the preprocessed infrared image data;

[0024] Extract multi-scale features through CSPDarknet53 in the YOLOv5s model, enhance the feature fusion ability with PANet, and output the fused feature data; then process the fused feature data through non-maximum suppression in the YOLOv5s model to filter out redundant detection boxes and output the animal species and location information.

[0025] In some of these embodiments, when the data analysis module calculates the movement speed of mammals, based on the centroid coordinates of the mammals in consecutive frames, the optical flow method is used to calculate the displacement, and then the movement speed is obtained by combining with the time.

[0026] In some of these embodiments, when the data analysis module performs posture anomaly detection on mammals, the pre-trained HRNet is loaded using the DNN module of OpenCV to output the key point coordinates of the mammals, and then the percentage of the joint angle deviating from the normal range is calculated according to the coordinates to determine whether the posture of the mammals is abnormal.

[0027] In some of these embodiments, for the voiceprint features extracted by the data processing module, DTW is used for voiceprint matching to calculate the similarity between the voiceprint features and the similar voice features in the disease voice library, and then whether the corresponding disease is suffered is judged according to the similarity.

[0028] In some of these embodiments, the disease warning module performs warning recognition through the LightGBM machine learning model; the LightGBM machine learning model receives the health information output by the data analysis module, outputs the disease risk probability; and performs corresponding warnings according to the disease risk probability.

[0029] Compared with the prior art, the present invention has at least the following beneficial effects:

[0030] In the embodiments of the present invention, the provided full-automatic wild mammal health monitoring and disease warning system avoids the limitations of a single sensor through the collaborative analysis of image, sound, and environmental data;

[0031] In the embodiments of the present invention, the provided full-automatic wild mammal health monitoring and disease warning system combines a static rule base (such as the judgment standard for canine distemper) with dynamic model prediction, taking into account the advantages of empirical knowledge and data-driven.

[0032] The additional aspects and advantages of the present invention will be partially given in the following description, partially become apparent from the following description, or be understood through the practice of the present invention. Brief Description of the Drawings

[0033] Figure 1Block diagram of the full - automatic wild mammal health monitoring and disease early - warning system disclosed in an embodiment of the present invention. Detailed implementation manners

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0035] Next, the embodiments of the present invention will be described in detail through specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0036] In some embodiments of the present invention, a full - automatic wild mammal health monitoring and disease early - warning system is provided. Through multi - source data collection, intelligent analysis, and real - time early - warning, it solves problems such as data islands, strong artificial dependence, and response delays in the prior art, providing a dynamic and scientific management tool for wildlife protection.

[0037] Please refer to Figure 1 As shown, in some embodiments of the present invention, the full - automatic wild mammal health monitoring and disease early - warning system includes: a data collection module, a data processing module, a data analysis module, and a disease early - warning module.

[0038] In the above - mentioned embodiment, the data collection module is used to collect the behavior, physiology, and living environment data of animals in real - time through a multi - type sensor network. The multi - type sensor network includes a number of infrared cameras, a number of sound sensors, and a number of environmental sensors.

[0039] In this embodiment, the infrared cameras are installed in hot - spot areas of animal activities, such as near water sources and nests; with wide - angle lenses and night - vision functions, they can capture animal activity images at a frequency of 10 seconds / frame. The sound sensors are arranged in the forest canopy or on the ground to collect animal sounds, such as calls for help and coughs, and environmental noises, and the sampling rate can be 44.1 kHz. The environmental sensors at least include a number of temperature and humidity sensors, a number of air pressure sensors, and a number of light sensors; they can record environmental parameters every 5 minutes.

[0040] In the above - mentioned embodiment, the sensor nodes are built - in with edge - computing chips to compress and encrypt the original data, reducing the occupancy of transmission bandwidth.

[0041] In some embodiments of the present invention, a data processing module is configured to preprocess multi-source data, including cleaning, normalization, fusion, and feature extraction. The preprocessing of the images collected by the infrared camera by the data processing module includes: performing normalization processing on the collected original images, which can be adjusted to a resolution of 640×640, and enhancing the contrast under low-light conditions by using histogram equalization to obtain the preprocessed infrared image data. The data processing module extracts voiceprint features. Specifically: performing frame division and windowing, FFT transformation, and cepstrum analysis on the audio signal to obtain MFCC feature vectors. In frame division and windowing, the frame division standard is: frame length 25 ms, frame shift 10 ms, and the windowing is applying a Hamming window. In the FFT transformation, the spectrum of each frame is calculated and mapped to the Mel scale through a Mel filter bank (40 triangular filters). In cepstrum analysis, a DCT transformation is performed on the logarithmic energy, and the first 13 coefficients are taken as the MFCC feature vectors.

[0042] The data processing module is also configured to perform data standardization on environmental data and physiological data. Specifically: performing Z-score normalization on environmental parameters (temperature, humidity, light) and physiological data (thermal imaging body temperature).

[0043] In some embodiments of the present invention, a data analysis module is configured to identify and analyze data to obtain information such as the type, location, movement speed, posture information, and health index of the detection target.

[0044] The data analysis module of the present invention uses a pre-trained YOLOv5s model for target detection on the preprocessed infrared image data. The YOLOv5s model structure includes Backbone (CSPDarknet53), Neck (PANet), and Head (Anchor-Based Detection), and the loss function is CIoU Loss (Complete Intersection over Union Loss). The loss function formula is as follows:

[0045]

[0046] Among them, ρ is the distance between the center points of the predicted box and the ground truth box, c is the diagonal length of the smallest bounding box, and v is the consistency measure of the length ratio.

[0047] In the above embodiments, YOLOv5 achieves a balance between high speed and accuracy through single-stage detection. Among them, CSPDarknet53 extracts multi-scale features, and PANet enhances the feature fusion ability to output the fused feature data. Then, the fused feature data is processed through non-maximum suppression (NMS, threshold = 0.5) in YOLOv5 to filter out redundant detection boxes, and the animal species labels and bounding box coordinates are output. That is, YOLOv5 finally outputs the species (such as Amur tigers, sika deer) and position information of the animals. Specifically, the species and position information of the animals in each frame of the image are output, and the confidence level > 90%. Compared with traditional manual annotation, the recognition method of the present invention using YOLOv5 can increase the speed by 50 times (real-time processing at 30 FPS) and supports multi-object detection in complex backgrounds.

[0048] When the data analysis module of the present invention analyzes the behavioral characteristics of animals, it is necessary to calculate the movement speed of the animals and then perform posture anomaly detection.

[0049] In the above embodiments, when calculating the movement speed of the animals, based on the centroid coordinates of the animals in consecutive frames, the optical flow method (Lucas-Kanade algorithm) is used to calculate the displacement, and the formula is:

[0050] I x V x +I y V y =-I t

[0051] Among them, I x 、I y are spatial gradients, I t is the temporal gradient, and V x 、V y are velocity components.

[0052] In the above embodiments, when performing posture anomaly detection on the animals, the pre-trained HRNet (High-Resolution Net) is loaded using the DNN module of OpenCV to output the coordinates of the key points of the animals (such as limbs, head), and the percentage of the joint angle deviating from the normal range is calculated (for example, if the bending angle of the hind limb < 100°, it is determined as lameness). The present invention captures spatio-temporal motion features through the optical flow method, and HRNet improves the posture estimation accuracy through high-resolution feature retention.

[0053] In some embodiments of the present invention, according to the previously calculated movement speed and posture anomaly detection results, behavioral indicators are generated. For example, if the movement speed = 1.2 m / s and the posture anomaly degree = 15%, it is marked as an abnormal behavior frame. Compared with the traditional threshold method, the false detection rate of the present invention can be reduced by 40%, and micro-posture changes (such as head drooping, etc.) can be recognized.

[0054] The data analysis module of the present invention can also perform dynamic tracking and anomaly marking. Initialize the tracker, assign a unique ID to the detected animal target, and use the DeepSORT algorithm (combining Kalman filtering and CNN appearance features) for cross-frame association. Anomaly behavior aggregation: Statistically analyze the frequency of abnormal behaviors of a single target within a time window (such as 10 minutes) (such as abnormal postures in 5 consecutive frames), and trigger the marking of abnormal events. The DeepSORT algorithm can solve the occlusion problem through dual clues of motion and appearance, and then predict the target motion trajectory through the Kalman filter in the DeepSORT algorithm. Finally, output the animal trajectory with ID and the timestamp of abnormal events (such as lameness of ID-003 from 12:05 to 12:15).

[0055] Through the collaborative work of Kalman filtering and CNN appearance features, the DeepSORT algorithm in the present invention combines the motion information and appearance information of the target. Kalman filtering uses motion information for fast state prediction, and CNN appearance features use the unique appearance of the target for precise matching. This combination enables the DeepSORT algorithm to maintain a high tracking accuracy and stability in complex scenarios, such as target occlusion and interference from similar targets. The tracking accuracy, MOTA = 85%, is better than the traditional KCF algorithm, and it can achieve long-term accurate tracking in complex scenarios.

[0056] In some embodiments of the present invention, the data analysis module of the present invention can also perform voiceprint matching on the voiceprint features extracted by the data processing module, which is implemented using DTW (Dynamic Time Warping). Calculate the similarity between the test audio and the disease voice database, and the formula is:

[0057]

[0058] where π is the optimal alignment path. If the similarity > 75%, it is determined as a successful match. MFCC simulates the auditory characteristics of the human ear, and DTW solves the problem of inconsistent time series lengths. Output the voiceprint matching degree (such as a cough sound matching the canine distemper database, similarity = 82%). Compared with the traditional FFT spectrum analysis, the disease voice recognition accuracy is increased by 35%.

[0059] The data analysis module of the present invention can also calculate the health index of animals based on the standardized environmental data and the physiological data of animals. Use a multiple linear regression model for calculation, and the formula is:

[0060] H = 0.3·T norm + 0.2·H norm + 0.5·A norm

[0061] where T norm is the standardized body temperature, and H normis the humidity deviation value, A norm is the activity score.

[0062] If the health index H < 0.6 and persists for 30 minutes, it is determined to be in a sub-healthy state.

[0063] Linear regression quantifies the influence weights of multiple factors, and the dynamic threshold adapts to environmental changes; a comprehensive health index is generated, and corresponding warnings are triggered according to the comprehensive health index. For example, when H = 0.45, a secondary warning is triggered. The present invention integrates multi-dimensional data, and finally the F1-score of disease prediction is increased to 92% (the single-dimensional index is only 78%).

[0064] In the foregoing embodiment, the disease warning module is used to formulate a hierarchical warning rule, predict the probability of an animal's disease risk, and trigger a corresponding warning according to the prediction result.

[0065] The disease warning module performs warning identification through the LightGBM machine learning model. The LightGBM model is obtained by training based on historical data (including normal and diseased samples), and outputs the disease risk probability (0%-100%) of the current monitoring target. The model supports dynamic update and can be optimized online when new disease patterns or data are added.

[0066] In this embodiment, for the hierarchical warning rule, the first-level warning is for high risk, which is triggered when the risk probability > 80% and at least two disease judgment criteria are met. The system automatically matches the disease types in the rule library (such as canine distemper, respiratory infection, etc.), and pushes an emergency notice (such as "suspected infectious disease") through the user interaction module. When the first-level warning is triggered, the administrator immediately activates the emergency plan to prevent the spread of the epidemic. The emergency plan can be means such as isolating the diseased animal and disinfecting the habitat. The second-level warning is for medium risk, which is triggered when the risk probability is 50%-80% or only one disease judgment criterion is met, marked as "sub-healthy state". After being marked as the second-level warning, it is necessary to manually review the data and increase the monitoring frequency (such as checking sensor data or on-site investigation). The warning information is transmitted to the administrator terminal in real time (such as a Web dashboard, a mobile APP, etc.). The system records the response result and feedbacks it to the model to achieve closed-loop optimization.

[0067] As an example scenario, in the Amur tiger reserve, the system detects a certain Amur tiger individual. A body temperature > 39°C is judged as a physiological abnormality, a 75% decrease in activity is judged as a behavioral abnormality, a cough sound with a matching degree of no less than 82% with the canine distemper library is judged as a voiceprint match, and the environmental humidity continuously higher than 80% is judged as an environment-related disease. Final warning result: risk probability 85% → The first-level warning (canine distemper) is triggered, and the administrator immediately implements isolation measures to effectively block the spread of the epidemic. If the risk probability is 50%-80%, the second-level warning (such as sub-healthy state) is triggered, and manual recheck is recommended.

[0068] The web dashboard displays a real-time health map, marks the risk levels with colors (red / yellow / green), and supports historical data query and trend analysis (such as the disease outbreak probability curve in a certain area). Administrators can adjust the sensor sampling frequency or update the warning rule library through the API. The mobile APP can push warning notifications and support one-key activation of emergency responses (such as dispatching patrol teams, etc.).

[0069] The present invention has an edge-cloud collaboration function: pre-computation by sensor nodes reduces the transmission load, and complex analysis is achieved by the cloud model. It has an adaptive learning function: the model supports online updates (such as adding a new disease voiceprint library) to adapt to the dynamic changes in wild animal behavior.

[0070] For the parts not detailed in the present invention, reference can be made to the prior art or they are well-known technologies to those skilled in the art. This embodiment does not make any limitations in this regard and will not be described in detail herein.

[0071] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims, and all of them fall within the protection scope of the present invention.

Claims

1. An automatic wild mammal health monitoring and disease warning system, characterized in that, The system includes: A data acquisition module configured to be able to acquire the behavior, physiology, and living environment data of mammals; A data processing module configured to be able to perform corresponding preprocessing on the acquired different types of data so that it can meet the requirements of subsequent data analysis and / or early warning; the preprocessing is any one or more of cleaning, normalization, fusion, and feature extraction; A data analysis module configured to be able to identify and analyze the preprocessed data to obtain the species, location, movement speed, posture information, and health information of mammals; and, A disease early warning module configured to be able to predict the probability of the disease risk of mammals and trigger corresponding early warnings according to the prediction results and the classification early warning rules.

2. The fully automatic wild mammal health monitoring and disease warning system according to claim 1, wherein, The data acquisition module includes a multi-type sensor network; the multi-type sensor network includes a number of infrared cameras, a number of sound sensors, and a number of environmental sensors.

3. The fully automatic wild mammal health monitoring and disease warning system according to claim 2, characterized in that, The preprocessing of the images acquired by the infrared cameras by the data processing module includes: performing normalization processing on the acquired original images, adjusting the image resolution to a preset standard resolution, and enhancing the contrast under low light conditions by using histogram equalization to obtain the preprocessed infrared image data.

4. The fully automatic wild mammal health monitoring and disease early warning system according to claim 2, characterized in that, The preprocessing of the sound data by the data processing module includes voiceprint feature extraction, and the extraction content includes: performing frame division and windowing, FFT transformation, and cepstrum analysis on the sound data to obtain MFCC feature vectors.

5. The fully automatic wild mammal health monitoring and disease early warning system according to claim 4, wherein In the frame division and windowing, the frame division standard is: frame length 25ms, frame shift 10ms, and the windowing is adding a Hamming window; The content of the FFT transformation includes: calculating the spectrum of each frame and then mapping it to the Mel scale through a Mel filter bank; The content of the cepstrum analysis includes: performing DCT transformation on the logarithmic energy and taking the first 13 coefficients as MFCC feature vectors.

6. The fully automatic wild mammal health monitoring and disease warning system according to claim 3, characterized in that The data analysis module uses the pre-trained YOLOv5s model to perform object detection on the preprocessed infrared image data; Extract multi-scale features through CSPDarknet53 in the YOLOv5s model, enhance the feature fusion ability by PANet, and output the fused feature data; then process the fused feature data through non-maximum suppression in the YOLOv5s model, filter out redundant detection frames, and output the animal species and location information.

7. The fully automatic wild mammal health monitoring and disease early warning system according to claim 1, characterized in that, When the data analysis module calculates the movement speed of mammals, based on the centroid coordinates of the moving object in consecutive frames, it uses the optical flow method to calculate the displacement amount and then combines it with time to obtain the movement speed.

8. The fully automatic wild mammal health monitoring and disease warning system according to claim 1, characterized in that, When the data analysis module performs abnormal posture detection on mammals, it uses the DNN module of OpenCV to load the pre-trained HRNet, thereby outputting the key point coordinates of mammals, and then calculating the percentage of the joint angle deviating from the normal range according to the coordinates to determine whether the posture of mammals is abnormal.

9. The fully automatic wild mammal health monitoring and disease warning system according to claim 2, characterized in that The data analysis module uses DTW to perform voiceprint matching on the voiceprint features extracted by the data processing module, thereby calculating the similarity between the voiceprint features and the similar voice features in the disease voice library, and then judging whether the corresponding disease is suffered according to the similarity.

10. The fully automatic wild mammal health monitoring and disease warning system according to claim 1, characterized in that The disease warning module performs warning identification through the LightGBM machine learning model; the LightGBM machine learning model receives the health information output by the data analysis module, outputs the disease risk probability, and issues corresponding warnings according to the disease risk probability.

Images