DESIGN AND IMPLEMENTATION OF SOFTWARE AND ELECTRICAL HARDWARE ON DEEP-LEARNING BASED FISH APPETITE DETECTION SYSTEM

• Main Author: D'sky, Agape
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73866
• Institution: Institut Teknologi Bandung

This final project book contains design, implementation, and testing of electrical hardware, inference models, and integration algorithms from SmartFishSense, a tool for detecting fish appetite when a stimulus is given based on artificial intelligence in the form of deep learning. The three parts of this system are primarily designed to be integrated with the Raspberry Pi 4 model B. These three parts are needed in the product as derived from existing problems, extracted into several specification points, and finally made into realization in several subsections of the product. The design of the electrical hardware subsystems, inference models, and integration algorithms is carried out separately and modularly before being integrated into a single unit. Each subsystem has its own functionality and an interface is defined for each subsystem, so testing can be done on a smaller scale first. Each subsystem is made with different methods and media. Electrical hardware is compiled using Altium, starting from making schematics, preparing PCB layouts, and creating documents for fabrication. Deep learning models (using a mixed multi-modal inference model between CNN and ANN) are built using Tensorflow, starting from designing the model architecture, training, to model validation. The integration algorithm is written in the Python programming language which is implemented on the Raspberry Pi. Broadly speaking, the products made have been able to meet the needs in design, namely to detect fish appetite and reduce the amount of wasted fish feed. However, this is still not enough because it must be limited by several factors, such as lighting, variations of fish and their behavior, and so on. In the future, disturbances caused by these factors can still be corrected by enriching the model training data.