Evaluation of Few-Shot Learning with Vision Language Models for Needle Bearing Defect Detection

Needle bearing defect detection is crucial for industrial machinery maintenance, yet traditional methods often require extensive training data and expertise. This study explores the potential of Vision Language Models (VLMs) for few-shot learning in classifying needle bearing defects using vibration signal spectrograms. We evaluate the performance of leading closed-source VLMs, including Gemini-1.5 Pro, 1.5-Flash, OpenAI GPT-4o, GPT-4o-mini, Anthropic Claude-3-5-Sonnet and Claude-3-Opus, in this task. Our results demonstrate that VLMs can effectively process and reason with spectrograms, surpassing human expert performance in needle-bearing spectrogram classification. This approach not only achieves superior accuracy but also offers a resource-efficient alternative to conventional deployment methods. The findings contribute to the rapidly evolving research area of VLM applications in spectrogram analysis and highlight the potential of few-shot learning with VLMs for industrial diagnostic tasks. Our study underscores the significance of this innovative approach in advancing predictive maintenance strategies and improving the reliability of industrial machinery.