A Mixture-of-Experts and Contrastive Learning-Based Method for Open-Set Diagnosis and Incremental Learning of Air Conditioner Fault Sound Signals

With the widespread use of air conditioners, the sound signals generated during their operation contain rich information about equipment status. Traditional fault diagnosis methods based on sound signals are usually developed under the closed-set assumption, which makes them insufficient to handle the recognition of new fault categories and the challenge of continuous learning in real-world scenarios. To address this issue, this paper proposes an open-set diagnosis and incremental learning method for air-conditioner fault sound signals based on a Mixture of Experts (MoE) model with contrastive learning. The proposed framework leverages multiple expert submodules to extract diverse features, while contrastive learning is employed to enhance feature discriminability and generalization. In open-set conditions, the learned representations enable effective detection of unknown fault types, and clustering with prototype construction is further applied to generate representations of new categories. An incremental learning strategy is designed to incorporate these new categories into training, thereby achieving continuous optimization. Experiments conducted on an air-conditioner sound signal dataset collected in the laboratory demonstrate that the proposed method effectively improves the robustness and adaptability of fault diagnosis in open-set recognition and incremental learning tasks.