Fault detection for plasma-enhanced chemical vapor deposition process using feature extraction

This paper presents a simple fault detection approach for plasma-enhanced chemical vapor deposition (PECVD) process by feature neurons. PECVD using the RF-induced plasma to create and sustain CVD reaction is an important process for the deposition of amorphous silicon and silicon nitride in the construction of TFT layers. Stable RF power is the basic requirement and any malfunction on RF power is possible to cause the incomplete deposition of thin film and, in addition, to affect the next production in sequence. In this study, delivery power is the main parameter investigated. Kohonen network is used to construct the feature neurons to draw out the signal characteristics of delivery power. Incorporating with the fuzzy c-mean algorithm and ellipsoidal calculus, this approach establishes the ellipsoidal threshold limits and can extract the process drifts and abnormal deviations in the process characteristics by limit checking. This fault detection system was implemented to check both the normal and fault-induced delivery powers and precisely discovered the equipment malfunctions.