Sepehr Karimi, Ph.D.

Sepehr Karimi, Ph.D.

Director of AI and IT Solutions

Deep Learning for Intelligent Fault Detection in Analog Circuits

Overview

This research presents a novel approach to fault detection in analog circuits using deep learning techniques. The paper, titled “A deep network solution for intelligent fault detection in analog circuit”, was published in the journal Analog Integrated Circuits and Signal Processing in 2021.

Link to full paper

Key Contributions

  • Developed a new fault detection method based on Convolutional Neural Networks (CNN)
  • Utilized Power Spectrum Density (PSD) of faulty signals as input to the CNN
  • Achieved superior performance compared to state-of-the-art methods

Deep Learning Approach

CNN Architecture

The proposed method employs a deep CNN with the following structure:

  • Input: Gray-scale images (129x8x1) derived from signal PSDs
  • 3 Convolutional layers
  • 2 Max-pooling layers
  • 1 Fully-connected layer
  • Output: SoftMax layer for classification

Key Features

  1. Feature Extraction: The CNN automatically extracts discriminative features from the PSD images, eliminating the need for manual feature engineering.

  2. Hierarchical Learning: The network learns both general and detailed features through its multiple layers.

  3. Adaptability: The model can be trained on various analog circuits, making it versatile for different fault detection scenarios.

Results

The proposed deep learning method outperformed existing techniques:

  • First Circuit Under Test (CUT):
    • Scenario 1: 99.8% accuracy (vs. 98.73% for next best method)
    • Scenario 2: 97.8% accuracy (vs. 94.3% for next best method)
  • Second CUT: 93% accuracy (vs. 87.9% for next best method)

Implications

This research demonstrates the potential of deep learning in analog circuit fault detection, offering:

  1. Improved accuracy in fault classification
  2. Reduced need for domain expertise in feature selection
  3. Potential for application in various types of analog circuits

Future Directions

Potential areas for further research include:

  1. Expanding the approach to more complex analog circuits
  2. Investigating transfer learning for quick adaptation to new circuit types
  3. Exploring real-time fault detection using this deep learning approach