Classification of fuel type for predictive maintenance in marine and industrial engines using time series feature extraction based on hypothesis tests and automated machine learning

• A machine learning method based on tsfresh and AutoML was carried out. • Best test F1 score ranging from 0.688 to 0.995, depending on the number of features and downsampling frequency of rotational speed, was achieved. • The Fourier and wavelet transform applied to the rotational speed alone reveal subtle differences among fuel types, highlighting the importance of utilizing machine learning techniques. • It is hypothesized that the captured stationary and transient components of the data reflect the fuel properties and combustion characteristics. Predictive maintenance in internal combustion engines can be enhanced by accurately identifying the fuel type based on data collected from sensors or electronic control units (ECUs). This paper presents a study that aims to predict the fuel type (HVO100 or EN590) using machine learning techniques, specifically based only on the engine's rotational speed. The rotational speed data of a heavy-duty 6-cylinder diesel engine is measured and downsampled to frequencies of 100, 1000, and 10,000 Hz. To extract relevant features from the time series data, hundreds of features are extracted using hypothesis tests via the tsfresh library. Subsequently, selected features are trained using Databricks' automated machine learning (AutoML) platform. The study explores the relationships between the number of features, downsampling frequency, and the choice of machine learning models. The results indicate that, under the current configuration, the best test F1 score of 0.995 is achieved using logistic regression with 20 features and a downsampling frequency of 10,000 Hz. The analysis of SHAP values and p-values revealed that components of the Fourier transform and wavelet transform of the rotational speed play crucial roles in distinguishing between the fuel types. It is our hypothesis that the differences observed in the frequency domain are related to variations in fuel characteristics. Overall, this study presents a simple, interpretable, and computationally cost-efficient machine learning solution for predicting fuel type in industrial engines. The findings demonstrate the potential of applying this approach in real-world production environments.