A Case Study on TBM Disc Cutter Damage Recognition Using Fine-Tuning Strategies Under Limited Data

Research on damage recognition of Tunnel Boring Machine (TBM) disc cutter wear is vital for enhancing operational efficiency. In practice, however, collecting well-labeled fault data under real TBM operating conditions is challenging, which makes it difficult to design and validate robust diagnostic models. To explore how transfer learning can be used in such limited-data scenarios, this paper presents a case study on fine-tuning strategies for TBM disc cutter damage recognition using vibration signals acquired from a TBM cutter rock-breaking test bench. A multi-channel vibration dataset is constructed that includes normal wear and three representative abnormal damage types under two distinct cutter configurations, which emulates cross-machine transfer between source and target domains. Models are first pre-trained on an abundantly labeled source configuration and then fine-tuned on a target configuration with limited labeled samples. Within this setting, we compare eight mainstream fine-tuning strategies across five representative backbone models, including both convolutional and temporal architectures, with the aim of providing practical guidance for model-strategy selection for TBM disc cutter damage recognition under limited data. The Auto-RGN strategy demonstrates exceptional data efficiency, achieving 99.5% of full-data accuracy with merely ten target samples; Convolutional models consistently surpass temporal models, with a 10.3% higher average accuracy.