Computer Vision based Framework for Power Converter Identification and Analysis

Abstract

This paper proposes a computer vision-based framework to identify the topology of a hand-drawn image or schematic of a power converter circuit and perform automated simulations. For component detection, a deep learning-based model, i.e., YOLOR, the state-of-the-art object detection model, is used with a model accuracy mAP0.5 of 91.6%. In order to trace the wire connections in the circuit diagram, a classical Hough transform algorithm is used. The nodes of the circuit diagram are identified with K-Means clustering of the point-of-intersections between the horizontal and vertical lines. With the help of the position of the components detected and the nodes, a netlist of the circuit diagram is generated that can be fed into any spice-based circuit simulator. An automated simulation of the schematic of the power converter is done with the help of PySpice - an open-source python module, to simulate the electronic circuit that runs a spice-based simulation engine, i.e., ngspice and xyce on the backend. The proposed methods have been verified using the main non-isolated DC-DC converters (buck, boost, and buck-boost). It is envisioned that this framework can also act as an educational tool. Moreover, the proposed concepts can be extended to create fully automated and optimal power converter designs for practical applications.