With the rapid development of the Internet of Things and intelligent sensing technology, sensor array systems have been widely applied in various fields. However, traditional methods struggle to effectively process high-dimensional nonlinear signals, and existing deep learning models suffer from feature redundancy and excessive computational complexity. In this paper, we propose a sensor array signal classification algorithm fused with big data mining, achieving accurate and efficient classification through multistage optimization. Firstly, a hybrid feature engineering framework is constructed, combining signal time–frequency analysis with genetic-algorithm-rough-set-based feature selection to reduce feature dimensionality while retaining discriminative information. On this basis, an attention-mechanism-enhanced long short-term memory fully convolutional network classification model is designed, which dynamically focuses on key response segments through attention weights. Experimental results show that the proposed algorithm outperforms traditional methods on multiple public datasets. Moreover, the algorithm achieves good balance between parameter quantity and inference time, providing feasibility for deployment on resource-constrained edge sensor nodes. This algorithm relies on high-quality annotated data, suffers from insufficient model interpretability, and faces domain shift issues, with remaining room for optimization in edge deployment. These problems will be addressed by integrating few-shot learning, interpretable design, and lightweight technologies in future work.