Fault Diagnosis in Analog Circuits Using Stacking Ensemble Machine Learning Approach

Diagnosing faults on complex analog circuits for modern industrial systems is a difficult task to accomplish. This is especially true in the context of ensuring reliable fault diagnosis accuracy. The task gets even harder with the variation of circuit component parameters due to things like aging, temperature variations, and flaws in manufacturing. As requirements for reliability increase, the methodology for testability advancements seem to be stalled. This research presents a powerful ensemble stacking method for analog soft fault diagnosis. The proposed system outperforms standard machine learning methods in classification accuracy. The new method combines many models instead of single classifiers, including Decision Trees, Logistic Regression, SVM, and CatBoost, and takes advantage of their robustness. To test the effectiveness of our strategy, we designed three different benchmark datasets using a leapfrog low-pass filter, a bandpass filter, and a four-opamp high-pass filter with an aim to replicate real world distortion of circuit parameters. These datasets consist of all possible fault and non-fault scenarios, which exceed 10,000. Changes in component tolerances, temperature drifting, and process variability are all present in these datasets. While CatBoost reaches up to 95.67% accuracy, it is still not high enough for tasks that are based around precision. Our new method performs far better, achieving 99.97% accuracy on the third circuit and 100% accuracy on the other two. Despite the high results of the fault detection reliability, there are still complications to deployment such as differences in temperature, aging of components, and real-world noise. Moving forward, we will validate our method under these conditions to ensure that the model is functional outside of a simulation. This attempt to balance accuracy in testing and environmental robustness is an effort to improve AI driven testability of analog systems.