Hierarchical Weighted LSTM with One-class Classifier for Preventive Protection of Cultural Heritage in Museums

Environmental monitoring and assessment of environmental risk levels are of great significance for the preventive protection of cultural relics in museums. Conventional methods focus on univariate statistical analysis, ignoring the interaction among variables and the time-related characteristics in environmental sequences. In addition, the environmental data is mostly unlabeled. Due to the strict requirements of museums on the environment, the monitoring data is mostly in the normal state, and there exists a serious problem of class imbalance. It increases the difficulty of establishing a supervised classification model for evaluating environmental risk levels. To solve the above problems, this article proposes a hierarchical weighted long-short-term memory (HWLSTM) with a one-class classifier for the preventive protection of cultural heritage in museums. Compared with the conventional methods, the proposed HWLSTM pays attention to the interaction among key variables and extracts dynamic features from the multivariate environmental time series. It hierarchically extracts short-term and long-term dynamic features to overcome the long-term dependency caused by long sequences. A dynamic weighting strategy is proposed to highlight the cumulative impact of continuous abnormal states. Furthermore, to address the issue of lacking labeled data, an unsupervised one-class classifier is proposed based on the one-class support vector machine (OCSVM) to achieve environmental monitoring and assessment of risk levels only using unlabeled data. Instead of constructing different models for each task, it provides a unified framework for the preventive protection of cultural relics. In this article, the proposed method is verified with the paper cultural heritage in Shanghai Museum. The experimental results validate the effectiveness of the proposed method.