With the advancement of autonomous driving, the technical requirements for perception sensors such as light detection and ranging (LiDAR) continue to increase. As LiDAR relies on laser emission and active thermal control, its performance can degrade under environmental and physical stressors. In this study, we analyzed failure modes via accelerated degradation tests and predicted failures of laser-diode thermoelectric coolers (TEC) using data collected from real-world on-road tests. We developed a lightweight transfer-learning framework that is robust across temperature ranges to enable rapid, on-vehicle fine-tuning and reliable operation in the interpolation and extrapolation domains. Brief, spike-shaped outliers pose a notable practical obstacle that destabilize fine-tuning with small batches. To address this issue, we introduce an Archimedes spiral–weighted loss that detects and down-weights spike outliers while preserving informative trends. To improve reliability in unseen conditions, we estimated uncertainty using Monte Carlo (MC) dropout and report expected calibration error (ECE) and mean prediction interval width (MPIW). The resulting confidence intervals are used to gate predictions and trigger targeted fine-tuning when uncertainty is high. The results of experiments show that the proposed model achieved an average R-squared value of ≥ 0.97 across cases, including challenging extrapolation scenarios, using only 10% of the available data for fine-tuning. The evaluation of uncertainty further confirmed calibrated behavior, with ECE ≤ 0.05 at the 95% confidence level. These results indicate that the proposed spiral-weighted transfer learning method with calibrated uncertainty gating provides an efficient and reliable approach for LiDAR TEC failure prediction under realistic constraints.