Multiple-Class Classification of a Diesel Locomotive Engine’s Mechanical Condition Based on Wavelet Properties of Vibration Signals via Machine-Learning Techniques

Objective: to develop and experimentally validate a multiclass classification algorithm for assessing the technical condition of a diesel locomotive engine based on wavelet properties derived from vibration signals, employing machine learning methods to ensure the reliability and safety of railway transport. Methods: experimental measurements were performed on a D50 diesel engine at the Diesel Locomotive Laboratory named after Professor Ya. M. Gakkel, Emperor Alexander I St. Petersburg State Transport University. A total of 491 triaxial vibration recordings were obtained across seven diagnostic states (one of which represented normal operation and six of which simulated fuel supply faults by cutting off fuel delivery to specific cylinders). Wavelet packet decomposition to the eleventh level, using a fourth-order Daubechies wavelet, yielded 45,984 diagnostic features from the decomposition nodes. Principal Component Analysis (PCA) was employed to minimize the dimensionality of the extracted features, retaining components that together explained 95 % of the variation. Stratified 5-fold cross-validation was used to compare eight machine-learning algorithms in order to ensure an unbiased evaluation of the models’ generalization capabilities. Results: the highest classification performance of 99.32 % was obtained using logistic regression with L1 regularization (Accuracy = 0.9932, F1-macro = 0.9921). The stratified cross-validation confirmed the stability and reproducibility of this outcome, yielding an F1-macro score of 99.56 % ± 0.55 %. This method demonstrated the lowest variability across all tested algorithms, indicating strong robustness to data variability. Practical significance: the experimental findings have confirmed the feasibility of automatic classification of diesel engines into diagnostic categories based on wavelet properties of vibration signals. Logistic regression has been selected as the optimal method because it provides a favorable balance of predictive accuracy, stability of results, and model interpretability. The developed algorithm can serve as a foundation for real-time on- board diagnostic systems for diesel locomotives.