Analog Circuit Fault Diagnosis Using a Novel Variant of a Convolutional Neural Network

Liang Han, Feng Liu, Kaifeng ChenAlgorithms 2022 · DOI

TL;DR

This paper introduces MSCNN-SK, a novel convolutional neural network variant for analog circuit fault diagnosis. By incorporating a multi-scale average difference layer and a dynamic convolution kernel selection mechanism, MSCNN-SK significantly improves the accuracy and discriminative power of fault detection in complex analog circuits compared to existing methods.

What Problem Does This Paper Address?

Analog circuits are crucial components in modern electronic systems, and their reliable operation hinges on accurate fault diagnosis. However, diagnosing faults in analog circuits is challenging due to the complexity of fault mechanisms and the diversity of failure modes. Traditional machine learning methods often struggle with nonlinear fitting and extracting sensitive fault features. While deep learning (DL) approaches like Convolutional Neural Networks (CNNs) have shown promise, existing CNN-based methods for analog circuit fault diagnosis have limitations: they often use raw output signals as input, which only captures numerical characteristics, and they employ fixed receptive fields, limiting their ability to adapt to varying fault feature scales. This paper aims to address these limitations by proposing a more robust and adaptive CNN architecture.

Methods

The proposed method, MSCNN-SK (Multi-Scale Convolutional Neural Network with Selective kernel), builds upon the 1DCNN architecture and introduces two key innovations:

  • Multi-Scale Order Mean Difference Layer: This layer computes 0th, 1st, and 2nd order mean difference sequences from the input time-domain response signals. This approach extracts the average rate of change of the signal at different scales, providing richer information for fault characterization and helping to filter out noise.
  • Convolution Kernel Dynamic Selection Mechanism: Inspired by the dynamic receptive fields in the human visual cortex, this mechanism allows the convolutional layer to adaptively select the size of its receptive field. It achieves this through a branching structure that processes the input with kernels of different sizes (3x3, 5x5, 7x7), fuses the results, and then uses a learned weighting mechanism (Softmax) to dynamically emphasize features from different receptive field sizes. This allows the network to better capture fault features of varying scales without manual trial-and-error tuning of kernel sizes.

The overall MSCNN-SK network integrates these novel layers with standard CNN components like convolutional layers, pooling layers, and fully connected layers for feature extraction, fusion, and final fault diagnosis.

Datasets

The paper evaluates the proposed MSCNN-SK network using data from two benchmark analog circuits:

  • Sallen-Key Band-Pass Filter Circuit: This circuit has 5 resistors, 2 capacitors, and 1 operational amplifier. The dataset consists of 18,000 samples, each with 150 time-domain points, representing 9 fault modes (including no fault).
  • Four-Opamp Biquad High-Pass Filter Circuit: This more complex circuit includes 10 resistors, 2 capacitors, and 4 operational amplifiers. The dataset comprises 26,000 samples, each with 300 time-domain points, covering 13 fault modes (including no fault).

The data for both circuits was generated through Monte Carlo simulations, accounting for component tolerances.

Key Insights and Ideas

  • Multi-scale analysis is crucial for analog circuit fault diagnosis: Simply using raw signals limits feature extraction. By analyzing the signal's rate of change at different scales (0th, 1st, and 2nd order mean differences), richer fault characteristics can be captured.
  • Adaptive receptive fields are key to robustness: Analog circuit fault features can vary in scale. A dynamic selection mechanism for convolution kernels allows the network to adapt its "view" to better capture these diverse features, outperforming fixed-kernel approaches.
  • Improved feature discriminability leads to better diagnosis: The t-SNE visualizations clearly show that MSCNN-SK extracts features that are more separated and concentrated for each fault mode, indicating higher discriminative power and making the subsequent classification task easier.

Key Results

The MSCNN-SK network demonstrated superior performance in fault diagnosis compared to several benchmark DL models (BPNN, RNN, LSTM, and 1DCNN):

  • Sallen-Key Band-Pass Filter Circuit: MSCNN-SK achieved an average fault diagnosis accuracy of 100%, matching the 1DCNN but outperforming BPNN, RNN, and LSTM (which were also close to 100% but had slightly lower averages and higher standard deviations).
  • Four-Opamp Biquad High-Pass Filter Circuit: For this more complex circuit, MSCNN-SK achieved the highest average fault diagnosis accuracy of 99.87%. This significantly outperformed RNN (92.13%) and was slightly better than BPNN, LSTM, and 1DCNN (around 99%). MSCNN-SK also exhibited the smallest standard deviation, indicating greater consistency.
  • Feature Visualization: t-SNE analysis showed that MSCNN-SK extracted features with the highest separation and concentration for both benchmark circuits, confirming its ability to generate highly discriminative representations.

Limitations and Future Work

  • Limitations: The paper notes that hyperparameter tuning for all tested models was determined through trial and error, which can be time-consuming. While MSCNN-SK shows excellent results, the complexity of the dynamic selection mechanism might introduce computational overhead compared to simpler CNNs.
  • Future Work: The authors propose applying the MSCNN-SK network to more analog circuit fault diagnosis scenarios to further evaluate its generalization ability. They also suggest exploring its use in other fault classification tasks beyond analog circuits.

Summary updated

BibTeX
@article{han2022analog,
  author = {Liang Han and Feng Liu and Kaifeng Chen},
  title = {{Analog Circuit Fault Diagnosis Using a Novel Variant of a Convolutional Neural Network}},
  journal = {Algorithms},
  year = {2022},
  doi = {10.3390/a15010017},
}
Figure