Soft fault diagnosis in analog electronic circuits using supervised machine learning

Analog circuits are commonly used in a wide range of industrial applications, and their assessment is of great importance to ensure proper functionality and prevent faults. However, this task is not as fully developed and is significantly less advanced compared to the assessment of digital circuits, as soft faults are particularly difficult to detect in analog circuits. This study addresses the application of supervised classification techniques for the detection and classification of soft faults in analog circuits. A feature extraction methodology is proposed based on voltage measurements at key circuit points and across different frequencies, enabling precise characterization of system behavior. From this feature, a benchmark employing different machine learning methods was used. The evaluated classifiers include k-Nearest Neighbors (KNN), Naïve Bayes (NB), Discriminant Analysis Classifier (DAC), Classification Decision Tree (CDT), Random Forest (RF), Support Vector Machines (SVM) and Artificial Neural Networks (ANN). Each model was optimized through hyperparameter tuning and validated using cross-validation techniques. The results indicate that ANN and SVM achieved the best performance, attaining an accuracy of 97.92 % and 97.22 % on test data, with a global Matthews Correlation Coefficient (MCC) of 97.76 % and 97.01 %, respectively. Although RF obtained the highest training accuracy (99.39 %), its performance significantly dropped during testing (93.06 %, MCC of 92.52 %), indicating overfitting. Additionally, models such as KNN and DAC demonstrated solid performance, whereas NB and CDT were the least effective. These findings highlight the importance of carefully selecting both the feature set and the classification model for fault detection in electronic circuits. A Sallen-Key band-pass filter was used as the circuit under test (CUT), as soft fault classification in this type of circuit is particularly challenging. This study demonstrates that it is possible to accurately predict faults in circuits similar to the one analyzed.