PDCC: Prediction-Driven Collaborative Compensation for Dynamic Multimodal Quality Evolution

With the rapid development of edge intelligence and autonomous unmanned systems, multimodal learning has been increasingly applied to resource-constrained distributed scenarios. However, dynamic environments often introduce modality missingness, quality degradation, and reliability fluctuations, which impair learning robustness. Existing methods mainly rely on complex fusion architectures or reactive compensation strategies based on current modality states, making it difficult to satisfy the real-time and proactive requirements of edge systems. To address this issue, this paper proposes a Prediction-Driven Collaborative Compensation method for Dynamic Multimodal Quality Evolution (PDCC). Instead of reconstructing missing modality features, PDCC introduces a lightweight prediction mechanism based on historical modality confidence sequences to estimate future modality reliability trends. Based on the predicted degradation risks, a prediction-aware confidence collaboration strategy is designed to adaptively adjust reliability estimation, collaborative decisions among neighboring nodes, and replay buffer management by estimating future compensation utility from predicted modality reliability under limited storage resources. Extensive experiments on the CREMA-D and AVE datasets demonstrate that PDCC achieves strong robustness under gradual degradation, abrupt degradation, and heterogeneous multi-node degradation scenarios. Compared with existing state-driven collaborative compensation methods, PDCC improves model convergence efficiency while maintaining low communication overhead through lightweight confidence-level interactions. The results validate that the proposed method improves learning robustness in resource-constrained edge intelligence scenarios under dynamic multimodal quality evolution.