PHSM, or Pig Heat Stress Machine, is a capstone project develop to help piggery owners to detect and automatically resolve any possible heat stress affecting the pigs. PHS machine gathers all information & details that is being monitored by the web-based system of PHSM.
PHS machine doesn't need any internet connection to function. The PHS runs on local network on the piggery and can be used to scan and connect other PHS machine on the same network.
The machine uses machine learning & image processing technology to detect and resolve the heatstress problem on pigs
The machine uses Mongodb for database that uses json like documents/record for storing data. NextJs that include Front-end (React.js) for dynamic and reusable web components. And Flask for realtime detection monitoring and PHS machine controls.
PHS Machine utilizes Raspberry Pi 4B Architecture:aarch64 only. Due to fund limitation, we only tested the system to run on Pi 4B 8Gb running version Debian GNU/Linux 11 (bullseye) 64bit.
We also attempted to install & run PHS but fails to run or not getting isntalled at all due to some dependency has been compiled and designed to function only on 64bit os & system.
We also attempted to run it using Ubuntu 64bit (PHS does not work) but it lacks the advance IO and settings capability compared to bullseye.
The image above shows the Conceptual Framework of PHSM. The PHS rely on it's 2 sensor. The Camera and Thermal Camera.
Any 0biects that has heat is emitting infrared radiation. Infrared radiation is one of the electromagnetic waves in the electromagnetic spectrum.
Human cannot see infrared on the naked eye. But we can feel it in the form of heat. Thermal scanners or Thermal cameras are being used to see the world of infrared. There's a wide variety of usage for thermal imaging including health, security, research and science. PHS uses thermal imaging to identify the intensity of thermal radiation emitting from the pigs body & uses Convolutional Neural Network for predicting or classifying it.
PHS uses Fan & Mist to help reduce the pigs body temperature. PHS decide when to activate these components depending on the analyzed data from the camera & thermal camera.
PHS used a custom trained Yolov5 that is only trained to detect pig. The model is trained using the dataset we created in roboflow. It contains 11,374 images of pigs.
- Training Set : 9.9k images ( 87% )
- Validation Set : 944 images ( 8% )
- Testing Set Set : 482 images ( 4% )
The result of yolov5 custom trained model is shown bellow
PHS implements custom made CNN classify between HeatStress & Normal. It is trained with 5,072 Normal temperature & 5,072 HeatStress tempareture. The following graph shows the training accuracy & loss graph.
accuracy
PHS uses Adafruit MLX90640 Thermal Camera which contains a 24x32 array of IR thermal sensors & 110° field of view. It can measure -40°C to 300°C with an accuracy of +- 2°C (in the 0-100°C range).
MLX90640 can be used in Arduino boards and in Raspberry Pi. The sensor communicates through I2C interface and can run from 3.3V up to 5v.
PHS also uses a camera to properly identify the pigs. The camera matches the 110° FOV of thermal camera. The images from this camera are then feed to Yolov5 object detection model. The model then returns coordinates where the system extract the data from the same coordinates in the thermal camera's thermal data.
We both modified this cameras & combine it in a single casing to achieve thesame FOV alignment that will produce identical aligned image & thermal image.
We are open for contribution if you have a better idea to improve the system. We made the training scripts & datas available here.