A novel transformer with soft-cluster prior attention and multi-scale convolution for remaining useful life prediction in air conditioning systems under multi-operating conditions

To ensure the stable temperature and humidity environment required for the collection of the university library, accurately predicting remaining useful life for air conditioning system under various operating conditions is of great significance. Therefore, a novel Transformer with soft-cluster prior attention and multi-scale convolution is developed for multi-condition remaining useful life forecasting of air conditioning systems. Firstly, an enhanced multi-view fuzzy C-means classifier is used to obtain reliable operating condition labels and soft membership degrees. Moreover, a soft-cluster prior attention module in the Transformer network is added to separate the internal and cross-condition dependencies through a transition-aware gate. Besides, the multi-scale convolution is embedded to form a state pattern-aware encoder, and an LSTM decoder generates the remaining useful life prediction trajectory. The training focus is on the transition segments and a lightweight gating regularization, and inference uses sliding-window membership update to preserve causality. It is tested to significantly improve the prediction accuracy and stability, enhancing the adaptability to environmental temperature changes, and reducing the error of the near-mode switching.