YOLO V8-ANA for Defect Identification of Steel Surface Using Microstructure Images and Acceleration Vibration Signals

In the intelligent manufacturing domain, defects identification (e.g., microscopic cracks) of the steel surface is essential to reduce the propagation of failure, and improve the reliability of industrial systems. Existing deep learning methods (e.g., YOLO-V8) often fail to consider the detrimental impact of background noise on defect identification, models are highly dependent on the quality and diversity of training data. This paper proposes a novel approach named YOLO-V8 with attention network architecture (YOLOV8-ANA) for defects identification of steel surface. Specifically, the original datasets of steel surface are decomposed into a series of bidirectional intrinsic mode functions (BIMFs) using the bidimensional empirical mode decomposition (BEMD), in which each BIMF is represented by specific spatial-temporal characteristics of the defect data. Second, the proposed YOLOV8-ANA model is elaborated by integrating the spatial-shift multilayer perceptron version 2 (S2-MLPv2) into the channel-to-pixel (C2f) module of the classical YOLO V8 model. Meanwhile, the original spatial pyramid pooling fast (SPPF) in backbone and neck of the YOLO V8 is replaced by the spatial pyramid pooling-fully convolutional spatial pyramid convolution (S-CSPC) module. Eventually, two experimental cases, microstructure images (e.g., crazing, inclusion, patches, pitted surface) and acceleration vibration signals (e.g., intact plate, breach and crack) of the steel surface, are conducted for performance investigation. The experimental results demonstrate that the identification accuracy of the proposed approach is much higher than the state-of-the-art methods such as YOLO-V5 and YOLO-V8 models in terms of recognition precision and mean average precision (mAP).