Joint-Gated Continual Learning for Roadheader-Mounted Millimeter-Wave Radar Range Prediction in a Coal Mine

Coal-mine excavation generates dust and water mist that impair optical sensing; millimeter-wave radar supports surrounding-range perception for assisted autonomous guidance, while changing conditions can cause predictor mismatch and unfavorable online updates. This paper proposes a joint-gated continual learning framework that integrates temporal-evolution context with observation–prediction consistency for selective adaptation. The temporal context characterizes local measurement dynamics and determines a state-conditioned residual tolerance; together with prediction innovation, it regulates parameter updates and online-memory insertion. Observation-stream evolution is decoupled from model adaptation, so measurements rejected for learning remain available to subsequent prediction, preserving information continuity while suppressing unstable updates. A lightweight regression predictor with bounded memory enables continual operation under limited resources. Experiments on field radar data and controlled disturbances show 1.35% lower MAE than a fixed predictor and 2.55% lower RMSE than independently source-tuned consistency-only adaptation, with a 0.09 mm MAE increase and updates decreasing from 94.83% to 85.51%. At matched learning rates, RMSE decreases by 0.63 mm versus consistency-only and 0.28 mm versus approximately rate-matched random thinning. RMSE was 206.76 mm versus 204.63 mm for Fixed, while the frozen GRU and Transformer achieved lower MAE and RMSE. At $k=8$, 99.11% of injected learning values are rejected. The framework supports efficient continuous sidewall-range prediction and model maintenance for assisted autonomous roadheader guidance in dust- and mist-affected excavation environments.