The accuracy of data-mining based predictive maintenance often relies on extensive process and machine knowledge to enable appropriate feature selection and data preprocessing. Measurement data obtained may be asynchronous and result in inaccurate features, affecting the accuracy of maintenance prediction. To overcome this drawback, this paper introduces an approach to automatically select a feature subset through a genetic algorithm. The full feature set is created based on different sliding windows characterizing different time shifts on adopted statistical metrics of the measurement data. The fitness function of the genetic algorithm is then developed based on the preliminary fitting of a hidden Markov model (HMM) on the selected subset of features and assumed machines’ condition in the training data. Ultimately the fittest subset of features is used to enable HMM-based predictive maintenance. The proposed approach is evaluated using data from semiconductor wafer production equipment, recorded over a period of one year.     «

The accuracy of data-mining based predictive maintenance often relies on extensive process and machine knowledge to enable appropriate feature selection and data preprocessing. Measurement data obtained may be asynchronous and result in inaccurate features, affecting the accuracy of maintenance prediction. To overcome this drawback, this paper introduces an approach to automatically select a feature subset through a genetic algorithm. The full feature set is created based on different sliding wi...     »