Thematic analysis of power customer service data is increasingly suitable for value mining in large-scale composite multimodal data. However, multi-source cross-modal data fusion faces challenges due to frequently updated coupling constraints, leading to instability in coupling effectiveness. This results in low usability of the data foundation for power customer service thematic analysis. To address this issue, this paper proposes a multi-source cross-modal data coupling model-based data extraction method for power customer service thematic analysis. At the feature-level multimodal data coupling stage, a joint iterative feature attention mechanism is introduced to trace the root cause of feature alignment issues during the fusion phase through distributed recursive iteration operations. To quantify feature alignment errors and correct abnormal alignment results, the method integrates and captures multimodal fusion feature deviations, thereby improving the robustness of multimodal data fusion under anomalous data interference. To validate the model, two typical power customer service analysis scenarios—service risk assessment and electricity bill recovery analysis—are selected. By analyzing the semantics and structure of customer inquiries, the method designs a correlation matrix in a co-attention mechanism to capture vector relationships and interactions, constructing a feature vector model. A matching algorithm computes the similarity between input query vectors and existing problem vectors, generating a series of relevant corpora for feedback. Finally, the data coupling model foundation is evaluated using power customer service data, user feedback, and corpus completeness scoring to verify the effectiveness and applicability of the proposed method.