Gas-insulated switchgear (GIS) represents core power system equipment. Its maintenance can be complex and time-demanding, with outages sometimes impacting fault-free components, leading to considerable financial losses. Hence, this paper introduces a multi-modal risk analysis fusion network (BDA-FRAMN) utilizing bidirectional attention. First, a multi-modal dynamic enhancement module is outlined to identify risks within internal modalities, effectively diminishing the redundancy of auxiliary modalities. Additionally, a bidirectional attention module is proposed to detect detailed multi-modal risk hazards through an innovative bidirectional multimodal dynamic routing mechanism. Specifically, the bidirectional attention module initially emphasizes low-level multi-modal risk hazards. Subsequently, these low-level multimodal hazards undergo a bidirectional multimodal dynamic routing procedure. This allows for the dynamic refinement and examination of higher-level, more granular multi-modal risk exposures. Finally, when contrasted with non-attention-based multi-modal learning and attention-based multi-modal learning models, the Bidirectional Attention Multi-Modal Risk Analysis Fusion Network (BDA-FRAMN) model exhibits significant advantages over other multi-modal models.