Feature Selection in Machine Learning for Knocking Noise detection
Abstract
The search for effective methods to obtain an accurate detection of faults in cyber-physical systems grows constantly. Usually, a considerable amount of data generated by sensors is the source of any data-based analysis. In this context, the application of Machine Learning algorithms to identify faults has gained popularity and acceptance due to the high performance and low cost compared to other techniques. To improve the performance of such anomaly detection algorithms and have greater accuracy for failure identification, some strategies can be addressed, such as selecting the features that best describe the failure. For this, Features Selection is performed to identify significant features in a dataset. In this paper we present a comparison of 6 feature selection algorithms that are used to select the best features to detect the knocking noise fault in automotive combustion engines. By collecting and using data from an engine electronic control unit (ECU), we show that features selection can reduce the number of selected features in a failure classifier by 55% (from 9 to 5) with an improvement of the detection precision by 2%.
