A hybrid memetic algorithm for simultaneously selecting features and instances in big industrial iot data for predictive maintenance

In Industry 4.0, various types of IoT sensors which are installed on machines to collect data for predictive maintenance. As the collected data increases, there are more missing values and noisy data. Related studies have already proposed various methods to solve the problems in big data. Among them, most studies focused on either feature selection or instance selection for data preprocessing before training forecast models. Metaheuristic algorithm is one of the mainstream methods in data preprocessing. However, most of these studies rarely considered feature and instance selection simultaneously. In addition, they seldom focused on noisy data. Therefore, this work combines the UCI datasets with noisy data to simulate the real situation. Memetic algorithm (MA) has excellent performance in machine learning of data selection, and variable neighborhood search (VNS) was also proved to be widely applied to the systematic change of local search algorithms. This work proposes a hybrid MA and VNS to find a new subset that maximizes the accuracy of the classifier while preserving the minimum amount of data by feature and instance selection simultaneously. Experimental results show that the proposed method can efficiently reduce the amount of data and the ratio of noisy data. By comparison with other metaheuristic algorithms, the proposed method has good performance by an excellent balance between exploration and exploitation.