Defect identification method of carbon fiber sucker rod based on GoogLeNet-based deep learning model and transfer learning

Abstract

Carbon fiber sucker rods are widely used in oil production site due to their light weight, high strength and corrosion resistance, but there is still a lack of effective internal defect detection methods during production and installation. Aiming at the characteristics of irregular interface in the sucker rod, a novel defect identification method of carbon fiber sucker rod based on multi-sensor information fusion and GoogLeNet-based deep learning model was proposed to identify online the internal defects of carbon fiber sucker rod. First, the full coverage scan of the sucker rod in the cross-section was performed by a water-immersed ultrasonic array containing 32 probes, and the corresponding ultrasonic reflection signals was obtained. Then, a multi-sensor information fusion method was proposed to integrate amplitude and flight time of received ultrasonic reflection signals with the spatial angle information of each probe into defect images. Time signal waveforms of ultrasonic signals with different defects were mapped into different defect images, so that we can rely on deep learning models in the field of image identification to identify those defects. Finally, A GoogLeNet-based deep learning model were trained to identify the image-based defect of the carbon fiber rod. The transfer learning method, which transferred weights of the pre-trained GoogLeNet model by ImageNet large database to the GoogLeNet-based defect identification model, was adopted to enhance the convergence speed and generalization ability of the model for insufficient training samples. The testing results show that the overall defect identification accuracy of the trained GoogLeNet-based deep learning model was 99.72%, which can identify effectively four typical defects and no defects of carbon fiber sucker rods.