Enhancing aero-engine blade few-shot anomaly detection with visual-language multi-modal models under domain shift conditions

Abstract

The blade is a critical component of the aero-engines, and the regular inspection of blades is essential for the healthy operation of aero-engines. Data-driven deep learning (DL) methods for blade anomaly detection gain significant success. However, the nature of data-hungry hampers the further development of general DL-based methods. In light of the insights garnered from the intricate feature embedding of the recent large-scale models, we discern their aptitude for the extensive potential in blade anomaly detection (BAD), notably within contexts constrained by scant sample sizes. Additionally, we are aware of the real-world domain shift problems caused by complex circumstances in aero-engine BAD highly cripple the performance of large-scale models. Thus, we propose a domain semantic perception for blade anomaly detection (DSP4BAD), a few-shot BAD method based on the visual-language large-scale model (CLIP), for exploring the potential of large-scale models and mitigating the above existing challenges. In general, benefiting from the CLIP’s extensive pre-training knowledge in large-scale datasets, DSP4BAD can obtain excellent anomaly detection performance by embracing the tailored knowledge from few-shot samples. To this end, we develop the domain-state prompt augmentation template (DSPAT) and the attention-guided feature adaptation module (AGFAM) to facilitate the adaptation of CLIP’s general knowledge to the domain-specific ones of BAD. Meanwhile, given the redundancy of in-context information in CLIP, an in-context semantic refinement module (ISRM) is devised for purifying the context semantic about anomaly to further alleviate the domain shift issues. Extensive experiment results demonstrate that our DSP4BAD achieves the state-of-the-art performance of anomaly detection with few-shot samples, which provides a promising tack for applications of large-scale models toward real-world blade inspection.