Online adaptation via incremental clustering for cross-domain chiller fault diagnosis

The chiller is a critical component of the heating, ventilation and air conditioning systems, and its fault diagnosis is essential for reducing building energy consumption. However, frequent working condition variations degrade the performance of traditional parametric models. Recently, cross-domain fault diagnosis methods have been used to address condition drift in chillers. However, traditional algorithms face two main limitations: (1) inefficiency in real-time prediction due to the need for iterative training on full target domain data, and (2) inability to cover all possible operating conditions. This paper proposes a real-time adaptive fault diagnosis algorithm based on cross-condition information fusion. Firstly, sensor data are fused into a mutual information matrix, and a predictive uncertainty separation algorithm divides the data into low-entropy and high-entropy samples. Low-entropy samples are labeled based on the previous time model, while high-entropy samples are processed through a memory buffer-based feature fusion space. Additionally, a prototype-based representation consistency method aligns online and historical data to fuse information across working conditions, enabling adaptation to changing conditions. Simultaneously, real-time incremental clustering ensures feature alignment for the same fault categories across batches. Experimental results demonstrate the model’s excellent real-time fault diagnosis performance in chillers.