PHS Machine

Pig Heat Stress Machine (PHSM)

PHSM, or Pig Heat Stress Machine, is a capstone project developed to help piggery owners detect and automatically resolve any possible heat stress affecting their pigs. The PHSM gathers all relevant information and details that are monitored by its web-based system.

Since the beginning of farming, livestock management has been crucial for maintaining healthy and productive animals. Over time, advancements in technology have provided farmers with tools to improve animal welfare and productivity. The PHSM is a modern solution designed to address one of the significant challenges in pig farming: heat stress.

Pigs do not sweat, which means they have a limited ability to cool themselves down. Their natural response to heat is to lay in mud or water to cool off. However, in modern pig farming, pigs are often contained within closed walls, giving them limited options to regulate their body temperature. This makes them particularly vulnerable to heat stress, especially in the Philippines, where the tropical climate is predominantly hot.

The PHSM mitigates this risk by monitoring the pigs' real-time body temperature and automatically resolving heat stress. It uses various devices, such as water misters and fans, to prevent or address heat stress.

Object Detection & Classification

PHS Machine identifies the pig & classify it's temperature using an Object detection model & a custom Convolutional Neural Network.

PHSM used YOLO v5 because of it's great performance and accuracy when it comes to realtime object detection. Since YOLO v5 have a pretrained weights to detect various objects/animals by default, we custom trained the model to detect only pig. We achieved this by creating our own training dataset ( PAB-2022-OLFU-Capstone ) in YOLO format using Roboflow & trained the model in Google Colab producing the best trained weight ( best.pt ).

The "Box" indicates how well the algorithm can detect the object's center and how well the anticipated bounding box encompasses the object with the final result of the Box is 0.03. The probability that an object exists in a proposed region of interest is measured by objectness and its final result is 0.020. 77.9 is the value of the mAP that the custom trained YOLOv5 has. The other results are shown below.

6 more

PHSM uses a custom CNN model to classify thermal images between Heat Stress and Normal. We opted for a custom CNN due to its strong capability to automatically extract features from images via convolutional layers. CNNs are adept at learning spatial hierarchies of features, which aids in identifying patterns, edges, textures, and objects. Their ability to detect subtle patterns makes them ideal for analyzing thermal images, where minor temperature variations are important.

The CNN was trained using 10,144 raw thermal images (5,072 normal, 5,072 heatstress). The training accuracy & result graph shows a consistently increasing accuracy and a steadily decreasing loss, indicating effective learning and improved performance. mai_Net.h5

The training for detecting heat stress consists of 10 epochs with a final validation accuracy of 94.78% and a validation loss of 0.1605 without overfitting and underfitting.

Making it work

PHS Machine uses 2 cameras, A 5MP Normal camera with 110° FOV and a thermal camera Adafruit MLX90640 24x32 IR Thermal Camera ( 110° FOV ). Since the thermal camera produces an array of thermal readings, we first convert it into a black and white image retaining its original values. This image is then resized to match the width and height of the normal camera image.

The PHS Machine utilizes two cameras: a 5MP normal camera with a 110° field of view and a thermal camera, the Adafruit MLX90640 24x32 IR Thermal Camera (also with a 110° field of view).

The thermal camera generates an array of thermal readings, which we first convert into a black and white image while preserving the original temperature values. This image is then resized to match the dimensions of the normal camera image, ensuring alignment for accurate combined analysis.

By resizing the raw thermal image to match the normal camera dimensions, the pixels containing temperature data will align precisely with those in the normal image. The raw thermal image itself lacks distinctive features, so for visualization purposes, we apply a color map and overlay it with the normal image to ensure exact alignment.

With these two images aligned, PHSM feeds the normal image into the YOLOv5 model to identify the pigs in the frame. YOLOv5 returns results including the coordinates (xmin, ymin, xmax, ymax) and confidence scores.

For all results with confidence scores higher than 0.5, we crop the raw thermal image based on the coordinates of each detection. This allows us to extract the precise thermal data for each specific pig. We use the raw thermal image because it contains the original temperature values in its pixels.

Using the processed thermal image is not ideal due to the application of color mapping, which converts the raw thermal data into a color scale. While this colorization makes it easier for humans to visually interpret temperature variations, it does not retain the original temperature values. This loss of original temperature precision can negatively impact the accuracy of the CNN model.

Actions are triggered based on the location of the pig within the scene. PHS virtually divides the scene into rows and columns, and each user-defined action can be bound to these specific cells. Depending on the pig's coordinates, the specific action defined for the cell where the pig is located will be activated.

While not all actions are bound to specific cells, they can also be triggered by events, such as a heatstress event. For example, the fan action is not bound to any specific cell; instead, it activates based on the "heatstress" event type. When a heatstress event occurs, the fan will activate to regulate the air inside the piggery.

PHSM have pre-defined actions available but also allows user to create their own actions and use available GPIOs.

Demo

Below show's the demo of PHSM fully working & detecting heatstress in a pig farm in San Miguel Bulacan.

Screenshots

10 more

PHS Machine is one of the hardest project I ever did due to the technologies/algorithm used. I thank my groupmates because without them, this capstone will not be possible.

Contributors

Thank's to my groupmates who contributed and made this Capstone Project possible.

• Jerbee
• Rojan
• Andrei
• Jason

Special Thanks

• Zalven

PHSM manual here 👉 PHS Docs