FPGA-Based Bayesian Neural Network Accelerator: A Simulation-Driven Framework for Real-Time Risk Prediction Using IR Sensors

This research proposes a simulation-oriented framework that leverages a VHDL-based Bayesian Neural Network (BNN) for real-time risk analysis using analog infrared (IR) sensor data. The design integrates Python scripting for live data acquisition and visualization, enabling an interactive loop between physical sensor input and FPGA-simulated inference. Utilizing Monte Carlo Dropout for modeling uncertainty, the VHDL-based accelerator outputs class-wise probabilistic predictions, classifying input scenarios into risk levels—low, moderate, and high. Real-time IR readings captured via Arduino are mapped into calibrated percentage risk scores and written to a file-based interface that communicates with the VHDL simulation. Prediction results are then read and graphically visualized using Python, with dynamic updates on a live graph displaying confidence levels for each output class. The proposed system eliminates the need for real hardware acceleration while maintaining simulation fidelity, making it ideal for prototyping embedded inference systems with limited resources. This work demonstrates the potential of hybrid simulation frameworks in enabling low-cost, scalable, and interpretable neural decision-making pipelines.