Hardware-Resource-Constrained Neural Architecture Search for Edge-Side Fault Diagnosis of Wind-Turbine Gearboxes

Edge-side fault diagnosis requires lightweight models. They are, typically, empirically handcrafted by experts, a process that is time-consuming and labor-intensive. The configurable resource capacity for the edge hardware is not considered in manual lightweight models; therefore, they may not meet deployment requirements. Here, a method based on a hardware-resource-constrained neural architecture search is proposed. Fault-diagnosis models are automatically designed for edge hardware considering the hardware configurable resource capacity. First, an empirically inspired search space is designed to reduce the search difficulty in lightweight models. Subsequently, a hardware-resource-constrained reward function is modeled to guide an agent in considering the hardware configurable resource capacity during the search and for iteratively sampling lightweight diagnostic models meeting the deployment requirements. Finally, Pareto-optimal domination is used to obtain a competitive Pareto-optimal solution set, providing an optimal model for the edge hardware to achieve fault diagnosis of wind-turbine gearboxes. The feasibility and effectiveness of the method were verified on a gearbox dataset from a drivetrain diagnostics simulator and on measured wind-farm data. Particularly, in the wind-turbine gearbox dataset, HRC-NASNet-b achieved a 98.32% diagnostic accuracy with 0.28 M parameters and 0.25 GFLOPs — a 2.83% higher accuracy with 7.96× and 3.20× fewer parameters and FLOPs, respectively, than the first-performing edge-side-friendly model, MobileNet-v2.