Akademik Veri Yönetim Sistemi
MEASUREMENT: JOURNAL OF THE INTERNATIONAL MEASUREMENT CONFEDERATION, sa.271, ss.120824, 2026 (SCI-Expanded, Scopus)
Reliable thermal management is pivotal for the efficiency and lifetime of high-power LED lamps. This work presents a lightweight, time-series-only pipeline that forecasts the lamp PCB temperature (L-PCB) one step ahead at a 30 s cadence using four thermocouple channels (driver IC, input capacitor, inductor/coil, and PCB). Two interface conditions—boron-nitride thermal paste vs. dry contact—produce 206 and 48 timestamps, respectively. A compact 2 × 64 LSTM encodes 30 × 31 feature windows; a 32-D bridge is concatenated with instantaneous features and passed to a LightGBM regressor. Strict leakage controls exclude any forwardfilled indices from validation/testing, and chronological splits with blocked cross-validation are employed. On held-out real windows, the model attains macro-averaged MAE=𝟎.𝟎𝟔 ◦C and RMSE=𝟎.𝟎𝟕 ◦C, outperforming persistence and linear baselines. SHAP analysis indicates that recent L-PCBlevel and short-term gradients dominate attribution, followed by IC–PCB couplings—consistent with thermal-inertia intuition. For deployment, the LSTM is quantized (int8) and the tree head compiled to fixed-point C; on an STM32L562 (Cortex-M33) the split-inference path achieves 91ms end-to-end latency within 148 kB RAM, enabling on-device prognosis. Limitations include the short No-Paste horizon (48 timestamps) and a single product measured at 25 ◦C/50% RH. Overall, the results demonstrate sub-1 ◦C short-horizon prediction with microcontroller-class resources, aligning with embedded implementation relevance.