A.K. Devarashetti, L. Forte III, W. Giegerich, Ph.J. Schneider
ACV Auctions,
United States
Keywords: engine oil, oil leak, oil quality, undercarriage, oil cap, dipstick
Summary:
Abstract: This work focuses on the development and implementation of machine-learning models that detects and quantifies the severity of oil defects within vehicles. These defects include oil leaks, low oil level, and bad oil condition. Our system utilizes a set of deep-learning models that assess images taken of vehicle undercarriages, engine oil dipsticks, and engine oil caps. The dataset is composed of vehicles at different stages of their life with varying maintenance histories. The images and labels used to train and test the models were obtained by hundreds of vehicle inspector experts as well as expert human annotators. We discuss the challenges and approaches to mitigating the label noise present in our dataset. In our experimentation, we investigate the accuracy of labels with different levels of detail and quantity. We show that pre-training on many weak labels and then fine-tuning on fewer strong labels yields higher accuracy than either method evaluated independently. Using a held-out test set, the model achieved a preliminary AURoC score of 0.84 for oil condition and 0.73 for oil leaks. Background: It is important to detect oil defects as they can lead to cascading issues within vital vehicle components such as the engine (Fig. 5) or transmission. Coolant mixing with engine oil could make the oil milky (Fig. 2) and cause head-gasket issues and abnormal exhaust smoke. Low oil level (Fig. 4) or bad oil could lead to engine failure. Leakage from engine, transmission or differential indicate problems with the respective components. Methods: Bad oil quality detection is done in two stages. In the first stage, we predict bounding boxes for oil cap and dipstick as seen in Fig. 3, then in the second stage, we predict 7 classes from the image-crops. For oil leaks, we also employ a two-stage system where the first stage detects components of the car from its undercarriage image taken by a proprietary Virtual Lift™ imaging system [1] (Fig. 3). The second stage determines if the component is leaking or not from the image crop. Experimentation: The training, validation and testing datasets are taken as the data collected in different time periods. These are then balanced by class. A fraction of these images are labeled with bounding boxes by the experts. We conducted experiments where we varied the size and resolution of the crops. We observe that the crop with native 999x999 maximum resolution, along with crop-padding of 150px yields the best score mentioned above. For oil leaks, we experimented with the raw VirtualLift images, a filtered subset of high quality images and with a text-filter of inspector notes. Each experiment outperformed the previous experiment. With data coming from recent months, we will double the training data for the future models.