Interview

Abstract

Mario A de Oliveira from Mato Grosso Federal Institute of Technology, Brazil, talks to Electronics Letters about his paper ‘Embedded application of convolutional neural networks on Raspberry Pi for SHM’, page 680. Mario A de Oliveira My research involves monitoring structural integrity through the development of new methods to detect and evaluate structural damage, with particular focus on intelligent systems to provide accurate and automatic diagnosis about the overall structural condition. The methods I work on can be applied to several types of structures and across several research fields, such as aerospace, aeronautical, bridges, civil infrastructure, rotating machineries, and wind turbines. I started working in this field during my PhD in 2009. At this time I was very excited to work in this fascinating field and since then I have been focusing on this area. The most interesting issue is developing new methods that can lead to an increase in safety and reduce maintenance costs in structural health monitoring systems. Over the years, structures suffer from degradation and can be damaged without prior warning. A strict preventive maintenance process can prevent major damage and ensure the smooth operation of the infrastructure. However, this process significantly increases operating costs. For this reason, structural health monitoring (SHM) techniques have been extensively studied to increase safety and reduce maintenance costs. For example, owing to the high level of safety required, the aerospace industry has demanded high investments in order to guarantee adequate operating conditions in aircraft. Within this field, surveys have shown that SHM systems could significantly reduce maintenance costs as damage could be detected at an early stage. Although there are several SHM methods in development, methods based on convolutional neural networks (CNNs) are very attractive, especially for those based on images and video processing. This is due to their higher performance in extracting significant features of the damaged area, which can increase the sensitivity of the method making the damage identification. Although many CNN applications that focus on monitoring structural faults have been proposed in the SHM field, the disadvantage of using CNNs is that they rely heavily on the availability of a GPU (Graphics Processing Unit). Furthermore, in several applications such as in aircraft, turbines, and machinery, due to size and weight issues, it becomes costly to install a typical computer within these places. To overcome that, we developed an application embedded in a low cost (about $40) and small hardware (Raspberry Pi). As a result, we demonstrate the power, performance, and feasibility of the application to identify structural failures with good reliability and accuracy. The embedded CNN is promising and may represent a new frontier for the development of low-cost SHM systems. The idea of embedding the CNN algorithm is attractive as the processing occurs locally and the cost of implantation reduces due to a fewer number of components being required. Running the CNN algorithm in a small size, low-cost, low-weight and power hardware can be advantageous from both financial and dimension points of view. Areas that require image processing in real-time could be the most beneficiated to take advantage of the proposed method. In the long term, structural damage detection and classification research is expected to grow rapidly. Methods based on CNNs have exploded in popularity and real world applications using CNNs are one of the most recent major breakthroughs in several areas, including computer vision, speech recognition, biomedical systems and natural language processing. The developing role of image and video processing in SHM systems is obvious in recent years and is expected to grow rapidly in the coming decades. In this context, the embedded CNN algorithm undoubtedly will play an important role. I have been focusing on developing small and compact acquisition hardware to obtain signals from a set of sensors attached to an aeronautical structure. This new hardware will be integrated to the developed CNN application to form a compact and complete SHM system. Of course, new improvements in the CNN-based method will be in course. Since I have been working in the SHM field, the quantity and quality of solutions have been widely exploited. Big companies have proposed embedded SHM systems that could be integrated into their aircrafts; however, this number is still insignificant. Over the next ten years I expect that companies will incorporate more SHM solutions into their aircrafts.