Classification of steady and unsteady diesel engine operating conditions based on relative power range

In actual diesel engine operation, the mixture of steady-state and transient data significantly reduces the accuracy of machine learning models. Therefore, a simple, reliable, and online-achievable method for condition classification is urgently needed. Current research on classifying steady- and unsteady-state data predominantly relies on single-feature metrics, lacking comparative evaluations across multiple indicators and a systematic analysis of threshold effects. To address this issue, this study acquires multi-condition operational parameters through bench tests and proposes the relative power range as a classification feature based on a comparative analysis of data variations across eight variables under different operating conditions. By analyzing the impact of varying the relative range threshold on classification performance, the optimal relative power range threshold is determined. The modeling accuracy of a back-propagation (BP) neural network is adopted as the evaluation metric to compare the proposed method with the K-means clustering algorithm. The results show that after screening steady-state data via the proposed relative power range classification method with the optimal threshold of 0.2%, the determination coefficient R 2 of the BP torque prediction model reaches 97.40%, which is 4.45 percentage points higher than the model trained on raw mixed steady/unsteady data, and 4.39 percentage points higher than the optimal result of K-means clustering ( R 2 = 93.01%). Among three relative-range classification indicators (power, rotational speed, torque), the relative power range achieves the highest classification accuracy of 95.53%, far exceeding the relative rotational speed range (90.64%) and relative torque range (64.27%).