Book Description

As the technology trends moving forward rapidly in semiconductor manufacturing industry, the importance of prognostics and health management cannot be neglected. Any kind of failure happens during the manufacturing process will cause huge lost of the profit. The traditional human inspection and experience of detecting operating faults is obsolescent because more and more signals are used to control the manufacturing process in semiconductor industry to fit the requirement of product which will make the failure definition becoming more complicated. Condition Based Monitoring enabled prognostics have been widely accepted by many industries. However, in real deployment, equipment or process fault detection accuracy is still a big challenge. From data-driven modeling point of view, the loss of accuracy comes from several aspects including data quality, individual equipment behavior variation, external input material variation, environment difference, operation condition and even modeling inaccuracy. Many researches focus on applying new algorithm or improving existing methods to extract information from the data and detect failure of equipment. They made great breakthrough and contribution on improving the fault detection algorithm calculation efficient and accuracy. However, sometimes the low accuracy of fault detection result is because of the data characteristic instead of the algorithm itself. For example, recipe change will affect machine operating status to cause shift and drift in collected signals, which is called multiple regimes. Every regime is one kind of class which contains its specific characteristic. With multiple regimes identification, uncorrelated cycle data can be separated to different groups to avoid the confusion. Considering the learning and classification ability of SOM, it will be applied to identify multiple regimes. By learning each regime's pattern, SOM can classify different regimes to reduce the impact of data shift and drift. The key development in this research is to improve fault detection method based on multiple regimes identification. Three one-class fault detection methods PCA-MSPC, FD-kNN and 1-SVM will be applied in each regime respectively. Due to the reason that the operation will always be aborted immediately when an error is detected, so the quantity of faulty data is usually limited. In order to deal with this issue, one class fault detection method which only needs normal condition data will be applied in this research. Besides, in order to handle different data characteristic, three fault detection methods are applied in the system as comparison. In the semiconductor manufacturing process case study given in this work, the one-class fault detection system based on multiple regimes identification, named local model, showed superior performance than global model. The faults detected rate is enhanced up to 40% by using local-based fault detection. In this case study, a number of practical concerns were considered including data quantity limitation and multiple regimes issue, and the fault detection comparison result between local and global model for all three methods will also be given.