The pumping unit is the core equipment of oilfield production, and the lack of its power data seriously affects the accuracy of oilfield equipment condition monitoring. To address the limitations of traditional imputation methods in temporal data imputation – limited modeling capacity, sample homogeneity in generative models, and computational inefficiency, this paper propose the Dynamic Sparse Activation Conditional Diffusion Model (DSAN-CSDI). This framework integrates conditional diffusion models’ probabilistic generation with dynamic sparse attention mechanisms through two innovations: (1) incorporating a dynamic sparse gating module that adaptively activates feature-critical neurons to minimize computational redundancy; (2) developing a validation-loss-driven adaptive sparsity strategy that dynamically optimizes accuracy-efficiency equilibrium. The methodology establishes an efficient framework for high-dimensional industrial time-series restoration tasks.