GACB-Loc: A CSI Indoor Localization Method Based on Graph Convolutional Multichannel Attention Using CNN and BLSTM

Indoor localization remains challenging due to multipath propagation, dynamic obstacles, and environmental noise. Traditional methods based on geometric or probabilistic models often fail under such complex conditions. The core challenge lies in effectively modeling spatial, temporal, and multi-channel characteristics of noisy wireless signals. Channel State Information (CSI) has the potential to address these issues by providing more detailed spatial and frequency domain features, making it a promising candidate for robust indoor localization. To address these limitations, this paper proposes a unified indoor localization framework—GACB Loc—which integrates graph convolution-based multi-channel attention, convolutional neural networks (CNNs), and bidirectional long short-term memory (BLSTM) to jointly model spatial, temporal, and channel-wise dependencies in CSI data. Aiming at the multi-channel characteristics of CSI data, a Transformer-inspired graph convolution attention mechanism framework suitable for CSI data is proposed. First, the CSI phase data are preprocessed, and CNN is employed to extract advanced features and capture complex spatial and frequency domain patterns from the CSI phase data. Then, by utilizing the graph structure of CSI data and adaptively focusing on the most important channels, the model’s ability to prioritize relevant information is improved. Finally, BLSTM is proposed to capture temporal dependencies in the data. We conducted experiments on the proposed method using both publicly available datasets and real-world deployment environments. The results on two public datasets showed mean localization errors of 0.4945 m and 0.6546 m, while real-world tests achieved average errors of 0.1691 m and 0.8259 m, demonstrating our approach’s effectiveness and robustness. Compared to ten other representative methods—including ILCL, BLS, MLP, NN, Horus, MOR, RADAR, SWIM, Bayes, and DTE—our approach achieved average improvements of approximately 74.95% and 86.1%, respectively.