Fault Detection on Variable Length Multivariate Time Series from Semiconductor Manufacturing

Fault detection in semiconductor manufacturing plays a critical role in ensuring product quality and reducing production costs. This research paper addresses the challenge of detecting faults in variable length multivariate time series data obtained from semiconductor manufacturing processes. The study compares the performance of two deep learning architectures, namely Convolutional Transformers and 1D Residual Networks (ResNet), for supervised fault detection using raw sensor data. Convolutional Transformers have gained attention for their ability to capture spatial and temporal dependencies in time series data, while 1D ResNet models have demonstrated strong performances in various classification tasks. To evaluate the effectiveness of the models, a comprehensive experimental framework is constructed using a real-world semiconductor manufacturing dataset, comprising variable length multivariate time series data from a plasma etching process. The results of the experiments indicate that both Convolutional Transformers and 1D ResNet models achieve promising fault detection performances on variable length multivariate time series. However, the performance comparison between these two architectures highlights their relative strengths and weaknesses. Furthermore, the impact of various factors such as network architecture, input representation, and training parameters on the performance of fault detection is investigated. This analysis provides valuable insights into the design considerations and trade-offs involved in developing effective deep learning models for semiconductor manufacturing fault detection.