A Prediction Method of Surface Wear in Internal Combustion Engine Cylinder Liners Using Multi-Scale Combined Recurrent Neural Networks

In the paper, wear on the surface of internal combustion engine cylinder liners is the focus for prediction. The wear condition of liners significantly affects the sealing performance between the piston and liner, directly affecting the operational efficiency, reliability, and service life of internal combustion engines. With the ongoing evolution of sensor technology and data acquisition systems, the monitoring capabilities of mechanical systems have been greatly enhanced. Parameters such as temperature, pressure, and vibration are continuously tracked, allowing for real-time insights into the system’s condition to be obtained. This real-time monitoring serves as the foundation for predictive maintenance, in which maintenance activities are scheduled based on the actual state of the equipment, effectively preventing failures and minimizing operational downtime. In recent years, machine learning, a subset of artificial intelligence, has seen rapid growth and has found applications across various domains, including mechanical engineering. Machine learning algorithms’ ability to process and analyze extensive datasets makes them valuable tools for predictive maintenance in the context of internal combustion engine cylinder liner wear prediction. In our study, a novel multiscale mixed Recurrent Neural Network (MIXRNN) model is introduced, characterized by the integration of multi-scale feature extraction techniques with a composite recurrent neural network (RNN) architecture, thereby enabling the adept capture and learning of the temporal characteristics and dynamics of wear on cylinder liners. The multi-scale feature extraction, a critical aspect of the MIXRNN model, is emphasized for its allowance of data analysis at varying scales or resolutions. This capability is deemed vital for discerning both short-term and long-term wear patterns in cylinder liners, which is essential for comprehending the multifaceted nature of mechanical wear that may emerge from diverse factors over varying time frames. Furthermore, the MIXRNN integration of various RNN architectures addresses traditional RNNs’ limitations, such as difficulty in learning long-range data dependencies. This synthesis enhances the model’s predictive accuracy and reliability. Wear in the real world typically follows nonlinear patterns and is influenced by various factors such as constantly changing operating conditions, material fatigue, and environmental influences. Nonlinear models may better capture these complexities than linear models. A nonlinear transformation of the actual operating data from the RTA38 internal combustion engine cylinder liner is presented, wherein the total amount of wear is converted into an increment of wear. The transformation of this data proves beneficial for simulating more realistic changes in wear under actual working conditions. It is shown that these data are relevant for model training and testing across different indicators, training data ratios, and lengths of time series. The effectiveness of the model component is further validated by the ablation experiment. The superior performance of the MIXRNN over traditional RNNs and their variants is demonstrated across key performance indicators, including mean absolute error, mean relative error, root mean square error, and coefficient of determination. The robustness and predictive accuracy of the MIXRNN, especially in tests involving small sample datasets and extended time series data, are particularly notable. In conclusion, the MIXRNN is represented as a robust and accurate tool for the prediction of wear on internal combustion engine cylinder liners. Its capabilities contribute to the improvement of the scientific and rational aspects of the overall design and maintenance strategies for these engines.