Training and Generalization of a Fault Classification AI Model for Lift Door System Using Data Augmentation

Data augmentation is an effective method in machine learning to address data scarcity. When the cost of acquiring specific state data in professional fields are high, data augmentation can expand the limited datasets and improve model performance and generalization by creatively transforming existing data. In this study, by installing sensors on the lift door system, metrics such as door panels' 3-dimensional vibration accelerations, door motor current, distance of door panels, are continuously monitored in real time. Different deformation fault states of the lift door panels were simulated, and the original monitored metric data were collected under continuous operation. Data augmentation methods were used to transform part of the original data and create new data, thus expanding the dataset for machine learning. After preprocessing, the dataset formed a multi-dimensional feature dataset with fault labels, and a previously designed and validated high-accuracy AI model was used for training. The input layer of this model includes feature data of vibration acceleration and current, while the hidden layer consists of a combination of LSTM and fully connected layers. This study found that the use of data augmentation methods, such as random noise generation, random sub-sequence cropping, time warping, and Brownian motion simulation, had limited overall effect on the generalization ability of the AI model for classifying deformation faults of the lift door system. Among these methods, the model augmented using Brownian motion simulation achieved an accuracy of 67.2% in identifying "unknown" deformation levels of lift door faults. This indicates that the Brownian motion simulation data augmentation method provided the AI model with a limited level of generalization capability for identifying different types of lift door deformation faults.