Cylinder lubrication image dataset

This study focuses on the main engine cylinders of ships, and the data used is sourced from two large bulk carriers operated by WINNING Corporation, WINNING UNIVERSE and SUNNY KANKAN. The displacement of both ships is about 200000 tons, with a fixed route, shuttling between Yantai Port in China and Bokai Port in Guinea. The sea conditions during the voyage are complex and varied, making it highly representative. Both ships are equipped with MAN B&W 6S60MC low-speed two-stroke diesel engines, featuring a six cylinder structure, which are widely used in large bulk carriers and have typical characteristics. In terms of data collection, a digital camera is used as the image acquisition terminal, which can accurately capture the subtle wear and carbon deposition characteristics of the cylinder inner wall. The specific collection process is as follows: when the ship is in a stopped state, the operator enters the engine room and takes multi angle photos of the inner walls of each main engine cylinder through the scavenging port to ensure that the images can fully reflect the key state characteristics such as carbon deposition and wear on the cylinder surface, providing high-quality image data support for model training and oil injection strategy optimization. To enhance the representativeness and applicability of the data, the collection work covers typical working conditions such as ship parking and berthing, as well as various sea conditions. Since February 13, 2020, a total of 2838 image samples of the main engine cylinder have been obtained, and a data foundation with time span and operational status differences has been constructed. Given that there is currently no publicly available dataset specifically for retrieving images of ship engine cylinders, and it is not possible to capture scanning port images during normal ship navigation, the samples obtained in this study mainly come from bulk carriers WINNING UNIVERSE and SUNNY KANKAN. The number of original images is relatively limited. In order to enrich the structural diversity of cylinder image samples, improve the generalization ability of the model, and effectively suppress overfitting problems, this paper performed data augmentation on the original image dataset, using methods such as image rotation, random cropping, brightness adjustment, clarity changes, and random occlusion to simulate the effects of lighting conditions, equipment differences, and personnel operations in actual shooting. Some enhanced samples are shown in Figure 3. After enhancement, the total number of image samples increased from the original quantity to 31218. To meet the requirements of video memory resources during the training process of deep learning models, all images are uniformly adjusted to 224 × 224 pixels before input to the model. This preprocessing operation effectively reduces the computational complexity during the training process while fully preserving the key feature information of the cylinder, ensuring efficient operation of the model under limited hardware conditions.