An attention-based knowledge fusion dual-stream network for intelligent machinery fault diagnosis

Machinery operating under complex conditions inevitably generates new fault modes over time. Traditional fault diagnosis methods cannot effectively identify these new fault modes. To address the above issue, this paper proposes an attention-based knowledge fusion dual-stream network (AKFDN) for class incremental learning. First, a gradient-decoupled dual-stream convolutional layer is constructed to achieve efficient fusion of old and new knowledge. The memory stream uses a parameter isolation mechanism to retain existing knowledge, while the learning stream adapts quickly to new fault modes through adaptive gradient updates. Secondly, an adaptive knowledge fusion strategy is introduced to enhance the balance between prior and newly acquired knowledge during the fusion process, thereby effectively addressing the data imbalance problem in incremental replay. Finally, an information-preserving attention mechanism is developed to capture correlations across dimensions in time series data. This enhances the model’s ability to represent sequential features. Experimental results show that the AKFDN method delivers excellent diagnostic performance across various benchmark datasets. This provides a new perspective for advancing research in class-incremental fault diagnosis of machinery components.