Abstract
Owing to the tolerance of circuit components in analogue circuits, feature overlap occurs between different component faults, especially in the case of early-stage faults. To address this issue, we present an improved model that utilizes the channel attention mechanism in conjunction with a fully convolutional network. The proposed model conducts end-to-end diagnosis by utilizing the response signals generated from the tested circuit. To diagnose different fault categories, we add a global average pooling layer to the last layer of the fully convolutional network to obtain the probabilities of various fault categories. Furthermore, we incorporate a channel attention mechanism, which allocates distinct weight coefficients to each channel to amplify crucial features while dampening less significant ones. The proposed method is validated with a four-opamp biquad high-pass filter circuit and a leapfrog filter circuit. The results demonstrate that the proposed model achieves a fault diagnosis accuracy of 95.24% in Case 1 and 87.09% in Case 2. This method exhibits higher diagnostic accuracy and reliability than existing methods.
Introduction
In the domain of electronics, analog circuits present a significant diagnostic challenge compared to their digital counterparts. They are highly sensitive to environmental factors and manufacturing tolerances. While "hard faults," such as short circuits, are relatively straightforward to identify, "soft faults"—where a component's value drifts slightly outside its intended range—are notoriously difficult to detect.
This paper specifically addresses incipient fault detection. When a component like a capacitor or resistor drifts by a mere 20%, the resulting change in the circuit's output signal often overlaps significantly with the "noise" generated by standard 5-10% component tolerances. Traditional diagnostic methods often rely on manual feature extraction (e.g., Fourier or Wavelet transforms), which may lack the precision required to distinguish these subtle, non-linear deviations from normal operating variance.
Proposed Methodology: SE-FCN
To overcome these challenges, we propose the SE-FCN (Squeeze-and-Excitation Fully Convolutional Network) architecture. This model is designed for end-to-end diagnosis, processing raw time-domain signals directly without the need for manual feature engineering.
1. 1D Convolutional Layers
The model utilizes 1D convolutional layers to work directly on time-domain pulse responses. This preserves the temporal relationships within the signal data and avoids the information loss associated with transforming signals into images or frequency charts.
2. Channel Attention (SE-Blocks)
The core innovation is the integration of Squeeze-and-Excitation (SE) blocks. These blocks "squeeze" global information from each channel and "excite" it by assigning learned weight coefficients. This mechanism allows the network to automatically amplify informative features related to specific faults while dampening the noise caused by component tolerances. Research findings indicated that the placement of these blocks is critical; inserting SE-blocks after the first three MaxPool layers yielded the highest diagnostic accuracy, whereas adding them to every layer led to a performance decline due to over-parameterization.
3. Fully Convolutional Design and Global Average Pooling (GAP)
By replacing traditional dense layers with convolutional ones, the model reduces the total parameter count and maintains structural data relationships. A Global Average Pooling (GAP) layer is added at the final stage to output probabilities for various fault categories efficiently.
Datasets and Data Generation
The model was validated using two benchmark circuits simulated in PSpice:
- Case 1: Four-opamp biquad high-pass filter. This moderately complex circuit involved testing 12 different fault conditions (drifts in C1, C2, R1, R2, R3, and R4).
- Case 2: Leapfrog filter circuit. A more complex circuit used to test scalability, involving 20 different fault conditions across 10 components.
Data was generated using Monte Carlo analysis to produce 26,000 samples. This process simulates real-world conditions where every component has a slight random variance within its tolerance range, creating the "overlap" problem that makes incipient faults so difficult to classify.
Results and Discussion
The SE-FCN model demonstrated superior performance compared to existing state-of-the-art methods. In Case 1, the model achieved an accuracy of 95.24%. In the more complex Case 2 (Leapfrog circuit), it achieved 87.09% accuracy, significantly outperforming standard 1DCNNs (by nearly 13%) and ResNet50 (by approximately 3%).
The attention mechanism proved vital in resolving the "tolerance overlap" issue. By mapping signals into a high-dimensional space, the model successfully separated fault categories that appeared indistinguishable in raw form. However, in the complex Leapfrog circuit, some fault categories (such as F0 and R3↑) still exhibited partial overlap, suggesting room for further refinement. Despite the added complexity of the attention mechanism, the model maintained high computational efficiency, classifying 1,000 samples in just 3.67 seconds.
Conclusion
The proposed SE-FCN model offers a robust, end-to-end solution for incipient fault detection in analog circuits. By leveraging channel attention, the model effectively navigates the complexities of component tolerances and feature overlap. Future work will focus on validating the model with physical hardware to account for real-world environmental noise and extending the approach to detect concurrent multiple-component faults.
Summary updated