Abstract: The broad learning system (BLS) is a fast and effective neural network model recently proposed. However, the shallow structure of BLS limits its feature extraction capability, consequently affecting its classification performance. To address this, we introduce a broad learning system called CNN- Attention BLS (CA- BLS), which combines CNN and an attention mechanism. CA- BLS built upon BLS and leveraged the local connectivity and weight sharing properties of CNN to extract local feature information at various levels in the samples. Simultaneously, we incorporated the attention mechanism to dynamically adjust feature channel weights during the convolution process, enabling the model to focus on global features in the samples and thereby enhancing classification performance. Experimental results demonstrate that the classification accuracy of the CA- BLS model surpasses that of other comparative methods, and the effectiveness of this fusion approach is validated through ablation experiments.