The conventional risk assessment of power supply equipment relies on manual inspections or single-sensor alarms, which cannot capture multi-dimensional, time-varying equipment states. To solve this problem, we developed a multi-sensor data fusion method combining entropy-based initialization with particle swarm optimization (PSO) for dynamic sensor data weighting. The method comprises three layers: field sensing (infrared temperature, vibration, partial discharge, gas composition, and electrical quantity sensors), edge computing, and central platform layers. Using 4560 samples from a 220 kV substation collected over 24 months, PSO increased the partial discharge weight from 0.22 to 0.28 to prioritize insulation fault detection. The method shows a 96.8% accuracy, a 96.2% precision, an F1-score of 0.960, an area under the curve of 0.978, and a root mean square error of 0.072, outperforming the analytic hierarchy process, fuzzy evaluation, backpropagation neural network (an 88.2% accuracy), and Dempster–Shafer evidence theory. Online inference required only 0.15 ms per sample. Under simulated disturbances, the accuracy was determined to be 85.4% at a 10 dB signal-to-noise ratio, above 89% with a failed sensor, and 85.9% with 30% missing data. Despite the promising results, performance validation is required with multiple sites and various conditions, since the method was designed for a single substation, simulation-only robustness testing, overall risk-level rather than fault-type diagnosis, and static batch training. The results of this study can be used for the development of a self-adaptive, fault-tolerant framework for reliable risk assessment in modern power systems.