PI Control for 3ph Motor Control Load (On-Going)
This block diagram provides an overview of an Active Front End (AFE), a high-performance AC-to-DC power converter. A digital controller acts as the brain, using PI control loops to precisely command a power inverter. This process allows the system to draw clean, sinusoidal current from the grid, ensuring high power quality. The result is a stable and regulated DC voltage output suitable for various industrial applications.
FFT for Active Harmonics Cancellation on Hysteris Control System (On-Going)
Our system uses a high-speed, interrupt-driven process to sample the distorted waveform in real-time. A 256-point Fast Fourier Transform (FFT) then precisely extracts the magnitude and phase of each unwanted harmonic, providing the exact data needed for active cancellation.
Power Optimized High Accuracy FPGA for NN Function (On-Going)
This project accelerates a Vision Transformer (ViT) on an AMD Xilinx FPGA by designing a custom, high-speed GELU activation function. Using a Chebyshev polynomial approximation for the tanh function, the hardware core achieves exceptional accuracy and power efficiency. This work enables the deployment of complex neural networks in high-throughput, power-constrained edge computing environments.
HIgh-speed SPI Interface for Real-Time data transfer for Inverter (On-Going)
This project showcases an advanced mixed-signal processing system built on an Intel Cyclone V SoC FPGA. It utilizes a hardware-software co-design approach, where a Linux-powered ARM Hard Processor System (HPS) works in tandem with custom logic implemented on the FPGA fabric. The primary function of this powerful combination is to generate high-precision analog signals by interfacing with an external Texas Instruments DAC Evaluation Module.
Optimized Neural Network Activation for Edge Devices
This project addresses the challenges of implementing neural networks on edge devices with limited computing power and memory. Using continued fractions for the hyperbolic tangent activation function, the design reduces memory usage and speeds up calculations on an Intel Cyclone V FPGA. The 32-bit floating-point blocks ensure efficient resource use while maintaining 7-digit accuracy. The design was validated by comparing functional simulations with MATLAB calculations, confirming its precision and performance in real-time applications.
Neural Network Controller for Real-Time Solar Energy Integration
This project focuses on implementing a neural network controller for real-time solar energy integration into the power grid using an Intel FPGA. To meet the high accuracy demands of real-time control, 32-bit floating-point blocks were employed, achieving a precision of 7 decimal digits. The design was validated through functional simulations and compared with MATLAB calculations, ensuring its effectiveness for embedded environments like microcontrollers and digital signal processors.
Novel Weight-Dropout Approach for Neural Network Controller
This project introduces a weight-dropout method to enhance the efficiency of neural network controllers in real-time solar inverter systems. By dropping small-magnitude non-diagonal weights during training, the approach minimizes connections while ensuring convergence. Integrated with Levenberg-Marquardt and Forward Accumulation Through Time algorithms, the technique improves trajectory tracking. Implemented on an Intel Cyclone V FPGA, this method significantly reduces computation and resource needs, making it ideal for embedded environments.
32-bit Fixed-Point Neural Network Controller for Real-Time Solar Inverter
This project implements a 32-bit fixed-point neural network controller for real-time control of a solar inverter using FPGA for faster computation. Two activation function designs, based on LookUp Table (LUT) and CORDIC methods, were implemented and verified on Intel Altera Cyclone V boards. The design achieved 4-digit accuracy compared to MATLAB double format, demonstrating its feasibility and efficiency for real-time control in embedded systems
Hardware Authentication for Smart Inverters Using PWM and Transient Response
This project introduces a cost-effective method to detect counterfeit solar inverters using Pulse Width Modulation (PWM) and transient response analysis. By comparing transient responses, significant differences serve as analog signatures for authentication. Extensive statistical analysis demonstrates the method's effectiveness in distinguishing genuine from counterfeit PCBs, offering a valuable tool for enhancing the security of inverter-based power systems.
RNN-Based Inverter Control Using FPGA for Enhanced Performance
This research introduces an innovative approach to implementing a Recurrent Neural Network (RNN) on an FPGA to overcome the limitations of traditional microcontroller-based systems. The design optimizes real-time energy performance by leveraging the FPGA’s parallel processing capabilities. The 32-bit fixed and floating-point RNN model was validated through timing, resource analysis, and accuracy checks against MATLAB and ModelSim simulations. This demonstrated a cutting-edge solution for neural network applications in power systems.