In the field of tribology, many studies now use machine learning (ML). However, ML models have not yet been used to evaluate the relationship between the friction coefficient and the elemental distribution of a tribofilm formed from multiple lubricant additives. This study proposed the possibility of using ML to evaluate that relationship. Friction tests revealed that, calcium tribofilms formed on the friction surface, with the friction coefficient increasing as a result of the addition of OBCS. Therefore, we investigated whether the convolutional neural network (CNN) model could recognize the tribofilms formed from OBCS and classify image data of the elemental distributions of these tribofilms into high and low friction-coefficient groups. The CNN model classifies only output values, and it’s difficult to see how the model has learned. Gradient-weighted class activation mapping (Grad-CAM) was performed using a CNN-based model, and this allowed the visualization of the areas important for classifying elemental distributions into friction coefficient groups. Furthermore, dimension reductions enabled the visualization of these distributions for classification into the groups. The results of this study suggested that the CNN model, the Grad-CAM, and the dimension reductions are useful for evaluating frictional features of tribofilms formed from multiple lubricant additives.
New methods to reduce carbon dioxide (CO) emissions, the principal driver of global warming, are urgently needed. One critical approach is to reduce friction loss in machinery, a major source of CO emissions. This is generally accomplished by tailoring the profiles of lubricant additives for a specific machine.
Continuously variable transmissions (CVTs) have been widely implemented in automobiles to improve their fuel efficiency. CVTs transfer power from the car engine to the wheels via the friction in the metal-to-metal contacts between the CVT belt and pulleys. To further enhance CVT transmission efficiency, it is thus necessary to increase friction at the metal-to-metal contact points between the belt and pulleys. To that end, lubricating oils known as CVT fluid (CVTF) play an important role. CVTFs usually consist of a base oil and various kinds of lubricant additives. Among these additives, phosphorus- and sulfur-based extreme pressure (EP) additives and calcium-based detergents influence the friction properties at the metal-to-metal contacts between the CVT belt and pulleys. By appropriately combining phosphorus-, sulfur-, and calcium-based additives, tribofilm formation can be controlled and a high friction coefficient can be achieved. However, it remains to be elucidated how tribofilm that is formed from CVTF, which contains such additives, provides high friction.
Many studies have investigated the chemical properties of tribofilms formed from lubricant additives to elucidate the friction mechanism. Some studies have utilized X-ray photoelectron spectroscopy (XPS) and time-of-flight secondary ion mass spectrometry (ToF-SIMS). In the present study, we conducted friction tests with a laboratory-made high-frequency reciprocating tribometer (HFRT) to measure the coefficients of friction on a tribofilm formed from sample oils formulated with combinations of tricresyl phosphate (TCP), dibenzyl disulfide (DBDS), and overbased calcium sulfonate (OBCS) as phosphorus-, sulfur-, and calcium-based additives, respectively. After the friction tests, the elemental distributions of the tribofilms were analyzed by electron probe micro-analysis (EPMA). The friction test results showed that, the calcium tribofilms formed on the friction surface and that the addition of OBCS increased the friction coefficient. This suggests that the formation of thick tribofilms is related to high friction through the addition of OBCS. However, it was suggested that the relationships between the chemical properties of the tribofilm and the friction coefficient are not one-to-one but rather intricately intertwined. Therefore, we considered that machine learning (ML) could be used to investigate the friction mechanism in a comprehensive manner.
ML has been applied across various academic fields and has contributed significantly to many studies. Within the field of tribology, ML models have been constructed to predict fault diagnosis, estimate life, determine lubrication regimes, and analyze wear properties from sensor datasets. In addition, ML models have been developed to predict wear properties, friction coefficients, and surface morphologies from laboratory-scale experimental datasets. Moreover, ML models have been built to develop the lubricating oils. The results of these studies suggest that ML models were able to express various relationships between input and output values. Thus, tribological application of ML-based models has been extensive, and their use in this field is expected to continue to expand.
The present study used a convolutional neural network (CNN) because the image data of the elemental distributions of tribofilms were used as the input values. The CNN, which is a kind of neural network, is used to predict the output value from image data. This is because neural network models predict only output values from per-pixel values of image data, whereas CNN models can recognize features from some parts of the image data. Many studies in tribology have used CNN models, but CNN models have not been constructed to investigate the relationship between the elemental distributions of tribofilms and their friction coefficients. The disadvantage of CNN models is that it is impossible to see how they learn. Therefore, a gradient-weighted class activation map (Grad-CAM) was used in this study. A Grad-CAM can visualize the areas in the image data that are important for predicting output values. Grad-CAMs have been used in many studies, though rarely in the field of tribology. Furthermore, dimension reductions were used to classify the image data of the elemental distributions of tribofilms in this study. Dimension reductions have been used not only to increase ML learning speed but also to visualize the distribution of high-dimensional data. However, dimension reductions have not yet visualized the distribution of tribofilm image data to investigate the relationship between the elemental distribution of a tribofilm and its friction coefficient.
Therefore, the purpose of this study is to investigate the possibility of using the CNN model, Grad-CAM, and dimension reductions to evaluate the relationship between the friction coefficient and the elemental distribution of a tribofilm that contains multiple lubricant additives. The friction tests revealed that adding OBCS led to the formation of calcium tribofilms and increased their friction coefficients. Therefore, we investigated whether the CNN model and dimension reductions were able to classify the image data of the elemental distributions of tribofilms into high and low friction-coefficient groups. We also used Grad-CAM to investigate whether the CNN model could recognize the tribofilms that formed from OBCS in order to perform the classification.
