HKT-FNAS: Search Robust Neural Architecture via Heterogeneous Knowledge Transfer for Defect Detection of PV Energy Storage Modules

Defect detection has become a key task in intelligent manufacturing of photovoltaic (PV) energy storage modules. Nevertheless, noise and unpredictable uncertainties, which adversely affects the effectiveness of Convolutional Neural Networks (CNNs), making it difficult to determine the optimal network architecture and thereby hindering industrial applicability. To overcome the aforementioned challenge, we propose a progressive fuzzy neural architecture search framework via heterogeneous knowledge transfer (HKT-FNAS), aiming to search for efficient fuzzy CNN models with fuzzy processing ability. First, we propose a fuzzy CNN search space and an architectural representation strategy, which integrate a series of fuzzy operations (e.g., fuzzy convolution, and fuzzy pooling) into CNN. Then, we develop a progressive evolutionary search framework, which utilizes the knowledge transfer strategy to assist the multi-stage search. Especially, the architectural insights learned from a smaller search space are utilized to guide the exploration of the search over a larger space. Next, we devise a predictor-based evolutionary search strategy, which learns an online predictor to directly evaluate the candidate architectures for efficient evolution. We carry out a series of comparative experiments on widely used fuzzy benchmark datasets. Results validate the effectiveness and efficiency of the HKT-FNAS framework, achieving the state-of-the-art performance compared to counterparts.