In this thesis, a fault detection and classification method for unbalanced profile data is proposed. The maintenance of manufacturing equipment via fault detection and classification is a key to improving qualities of products in modern manufacturing system. Profile signal characterized by manufacturing process is used as the dataset for fault detection and classification. The signal, composed of steps that portray similar aspects, well reflects the status of process equipment. However, there are some problems in the utility of profile data.
The signals obtained from manufacturing process are difficult to be analyzed without preprocessing for the following reasons. The dimension of signals has been expanding recently. This phenomenon demands tremendous cost for storage and lowers the performance of detecting and classifying faults. Moreover, the physical length of profile signal differs from each other; statistical analysis regarding a sampled time as a variable cannot be adapted. Therefore, feature extraction method should be applied before the use of fault detection and classification. Wavelet based feature scoring algorithm and common feature selection schemes are proposed for preprocessing profile signal data.
If the distribution of fault is known, more sensitive fault detection is possible through 2-class classification methods. When a control chart is initially constructed, there are only few or no fault data. Hence, 2-class classification method cannot be applied or does not perform well due to the unbalanced data at the early stage of manufacturing cycle.
Hybrid process control method, which uses one-class classification based control chart when there are no fault data, makes possible the use of 2-class classification even if there exists only one fault data. SOMBoost is designed for overcoming such disadvantages caused by unbalanced data.
Profile data, which resembles semiconductor process signal, is used in analyzing the performance of hybrid process control. Process control framework based on hybrid process control method performs better in experiments than conventional methods.