To address the issues of low instability sample proportion, missed detections caused by fixed thresholds, and reduced interpretability due to high-dimensional inputs in electric-drive tracked vehicles under complex conditions, an instability mode identification method based on staged adaptive threshold decision-making is proposed. First, optimal features are selected using random forest out-of-bag error ranking and Shapley interaction values to enhance efficiency and interpretability. Second, an XGBoost model with parameters optimized by particle swarm optimization is constructed. Finally, considering asymmetric misclassification costs, the multi-classification task is restructured into a two-stage process of "instability detection–instability mode identification," with an adaptive threshold function based on classification uncertainty. Experimental results show that compared to a fixed-threshold XGBoost model, the proposed method reduces the overall missed detection rate from 4.7% to 1.3%, increases accuracy from 95.80% to 96.94%, and improves the F1 score from 94.05% to 95.56%.