The stable operation of electromechanical equipment is vital for modern infrastructure. However, thermal faults remain a leading cause of unexpected shutdowns. Traditional single‑point monitoring and infrared (IR) inspections often fail to detect incipient, localized anomalies owing to limited timeliness and spatial resolution. Therefore, we developed a predictive system based on high‑density temperature sensor arrays and advanced spatiotemporal modeling. In the system, an optimized 64‑channel Class A PT100 resistance temperature detector topology with MAX31865 converters was deployed to capture dynamic thermal fields. A complete preprocessing pipeline, including denoising and feature engineering, ensured data integrity by suppressing stochastic hardware noise. The prediction model was established by integrating convolutional long short‑term memory (ConvLSTM) with a convolutional block attention module (CBAM), focusing on critical spatial regions. Quantile regression analysis was used for probabilistic fault estimation and a multi‑level warning system. The validation results of the system performance on a 75 kW pump for three months demonstrated that the ConvLSTM + CBAM model showed a mean absolute error of 0.50 °C, a 23% improvement over the ConvLSTM model. The model extended the average fault warning lead time to 52 min, representing a 150% increase compared with that of the LSTM model. These results show that dense sensor arrays combined with attention‑based deep learning deliver accurate thermal diagnostics comparable to IR imaging, offering a scalable solution for proactive industrial maintenance.