In machine learning, the one-class classifier (OCC) has become a promising area where extensive research has been devoted to anomaly detection. Anomaly detection methods require high-quality features. Feature adaptation for anomaly detection often suffers feature deterioration termed catastrophic collapse, where all (including anomalous) samples are mapped to the same point. Existing deep learning-based anomaly detection methods often rely on self-supervised feature learning for learning strong features but lack to adapt to new domains. Thus, an effective solution for controlling the collapse of the original features is to adapt Pre-trained features in unsupervised deep learning models on external datasets to improve the detection performance trained on new domains.