NDTNet: An End-to-End Design of an Optical Nondestructive Evaluation System for Metallic Objects on FPGA

Borwankar, Raunak

Etd

NDTNet: An End-to-End Design of an Optical Nondestructive Evaluation System for Metallic Objects on FPGA

Public

The automated and rapid inspection of high-volume metallic components has been a key area of industrial research for many years. A main challenge is to nondestructively detect faulty components on a consistent basis without human intervention. Although a host of nondestructive evaluation (NDE) methods exhibit high resolution, they are difficult to implement on the factory floor; the approaches encompass acoustic resonance, magnetic particle, laser scanning, eddy current, and ultrasonics. Unfortunately, these inspection techniques are usually tailored to specific industrial components and flaw types. They furthermore require sophisticated test arrangements which are not scalable to high-speed inspection. As computer systems have evolved, research in computer vision based optical NDE methods have become feasible. These optical approaches can be broadly categorized into two parts, computer vision based and neural network based. While computer vision based optical NDE methods are highly efficient, they require human intervention in identifying critical flaw features. Moreover, computer vision algorithms are highly sensitive to the manufacturing environment, particularly fluctuating light conditions and background noise. As an alternative, convolutional neural networks (CNNs) are increasingly employed in nondestructive optical inspection. Even though state-of-the-art CNNs have proven efficient when coupled with transfer learning, they are generally not optimized for rapid testing of production samples on low-cost, dedicated hardware platforms. In this dissertation, we propose a general workflow to automatically construct and optimize a CNN architecture using existing computational frameworks. The proposed approach is tested with different production datasets for surface flaw inspection. Based on a novel sensing arrangement, we achieve precision of nearly 99% for both datasets. Out of other convolution acceleration techniques like fast finite impulse response (FIR) and fast Fourier transform (FFT), we use Winograd based accelerators to speed up the convolutions. State-of-the art Winograd based accelerators are usually designed to perform stride-1 convolutions. In this research, we developed a Winograd accelerator which can perform both stride-1 and stride-2 convolutions on the same hardware platform. The novel hardware implementation is 3.25 times more computationally efficient for stride-1 convolutions when compared to a standard approach, while it is 1.44 times more efficient for stride-2 convolutions. It therefore lends itself as a highly flexible, scalable and rapid inspection methodology suitable for many high-volume production environments.

Creator