Defect Detection Using Shuffle Net-CA-SSD Lightweight Network for Turbine Blades in IoT

Early detection of blade defects is crucial for turbines in Internet of Things (IoT) as it can prevent failures, minimize downtime, and enhance system reliability. Deep-learning-based methods have improved the accuracy of defect recognition and classification. However, there are challenges in balancing recognition efficiency, accuracy, and the ability to detect small targets. Therefore, a lightweight defect detection model of turbine blades using single-shot multibox detection based on ShuffleNetv2 and coordinate attention (SN-CA-SSD) in IoT is developed in this article. First, within the single-shot multibox detector framework, ShuffleNetv2 replaces the VGG-16 network as the fundamental feature extraction network. This substitution ensures detection accuracy while reducing model complexity. Second, the coordinate attention mechanism is introduced after the feature map, weighting important features to enhance the model’s ability to perceive key information. Finally, the loss function is redesigned using Efficient Intersection over Union (EIoU) loss to improve detected targets’ accuracy and localization precision. The proposed method is applied to defect detection images of turbine blades, and the results demonstrate that the algorithm significantly improves accuracy, detection efficiency, and the ability to detect small targets, striking a balance between precision and efficiency. By combining the model with an interpretable algorithm, the algorithm’s decision-making process is analyzed from the perspective of representation visualization, enhancing the algorithm’s interpretability.