Optimization Strategy for Resource Allocation in Deep Learning Driven Computing Systems

Fine-grained recognition is a highly challenging task since the distinctions among sub-categories—such as different types of transformer faults (e.g., partial discharge, thermal faults, or insulation degradation)—are often quite subtle and slight. In transformer fault detection, existing approaches predominantly focus on the localization of discriminative parts (e.g., hot spots in infrared images or discharge patterns in partial discharge signals), while neglecting the semantic information hidden behind the chaotic distribution of parts, which may contain critical diagnostic clues. To address this limitation, this paper learns an interpretable and compact trie model, namely Co-Activation Trie (CA-Trie), which mines the semantic and meaningful co-activations in fault-related features. In the trie model, each node represents an activated part (e.g., a specific fault signature in vibration or thermal data), and each edge encodes the relationship of co-activated parts in the same feature map, capturing the underlying fault mechanisms. Further, we transfer the obtained co-activations to the task of unsupervised fine-grained fault recognition, enabling the identification of fault types without extensive labeled data—a crucial advantage in real-world power systems where labeled fault samples are scarce. We evaluate our method on two benchmarks (including a transformer fault dataset), and experiments demonstrate that the co-activations mined by CA-Trie consistently represent stable semantic relationships across different fault instances, improving the reliability of fault diagnosis.