Abstract
Aiming at the analog circuit performance online evaluation demand of the largest interval principle and underlying geometric structure, two online methods of dimension reduction are proposed for analog circuit performance evaluation from the angle of feature extraction. First, a supervised method of dimension reduction based on Fisher's Linear Discriminant Analysis (LDA) is presented to increase the classification distance largely. However, since incomplete classification can significantly impact evaluation accuracy, another feature extraction strategy using Locality Preserving Projections (LPP) is proposed to detect the essential manifold structure. In this paper, fault diagnosis is recognized via the Global and Local Preserving based Semi-supervised Support Vector Machine (semi-supervised Global LSSVM). The experiment utilizes a typical Sallen-key low-pass circuit as the diagnosis object. The diagnosis speed and accuracy are validated via numerical simulation, comparing the proposed method with traditional Support Vector Machines (SVM).
Introduction
Diagnosing faults in analog circuits is notoriously difficult compared to digital circuits due to component tolerances, circuit nonlinearities, and the increasing complexity of modern integrated systems. Traditional supervised learning methods, like standard SVMs, require large amounts of labeled data, which is expensive and time-consuming to generate in real-world testing environments. Conversely, purely unsupervised methods often struggle with noise and uncertainty.
This paper addresses these challenges by proposing a semi-supervised learning approach. By leveraging a small amount of labeled data alongside a larger pool of unlabeled data, the "Global LSSVM" model aims to improve diagnostic speed and generalization, making it suitable for real-time online evaluation.
Methodology
The core of the proposed method is a hybrid feature extraction strategy that feeds into a Least Squares Support Vector Machine (LSSVM). The workflow balances global classification boundaries with fine-grained local data clusters.
Feature Extraction and Classification
- Global Feature Extraction (LDA): Fisher’s Linear Discriminant Analysis is used to maximize the distance between different fault classes, capturing the global structure of the data.
- Local Feature Extraction (LPP): Locality Preserving Projections detect the essential manifold structure—the local geometric relationships between data points that global methods might ignore.
- Semi-supervised Global LSSVM: By combining LDA and LPP, the model respects the underlying shape of the data. Manifold regularization is applied to ensure the classification function is smooth across the data distribution, preventing overreaction to noise.
Experimental Setup
The researchers evaluated their method using a simulated 1 kHz Sallen-key low-pass filter circuit. Nodal voltages from circuit simulations were normalized to a range of [-1, 1] during pretreatment to ensure optimal SVM performance.
Dataset and Simulation
- Fault Types: Eight specific "soft fault" categories (F1-F8) were defined, representing ±50% variations in resistors (R2, R3) and capacitors (C1, C2).
- Data Volume: A total of 60 datasets were used—20 for training and 40 for testing.
- Simulation Tools: Circuit behavior was simulated using Multisim 12.0, and the algorithm was implemented in MATLAB using the LSSVM Toolbox.
Results and Discussion
The paper compares the proposed Semi-supervised Global LSSVM against a traditional SVM. While the traditional SVM showed slightly higher accuracy in this specific small-sample test, the Global LSSVM demonstrated a significant improvement in computational speed.
| Evaluation Algorithm | Training Samples | Testing Samples | Accuracy | Computational Speed |
|---|---|---|---|---|
| SVM | 20 | 40 | 78.45% | 98.36% |
| Semi-supervised Global LSSVM | 20 | 40 | 75.25% | 99.86% |
The authors highlight that for larger datasets (over 1,000 samples), the efficiency gains of the semi-supervised approach become even more pronounced, making it better suited for real-time online evaluation where speed is critical.
Conclusion
The Semi-supervised Global LSSVM offers a pragmatic middle ground for analog circuit fault diagnosis, avoiding the laborious labeling process of fully supervised methods while maintaining better stability than purely unsupervised ones. Future research will focus on the diagnosis of multiple simultaneous faults, which represent a more complex and common real-world challenge.
Summary updated