Automated Diagnosis of Material Condition in Hammering Test Using a Boosting Algorithm Hiromitsu Fujii 1 , Atsushi Yamashita 1 and Hajime Asama 1 Abstract— Automated diagnosis systems are necessary for the maintenance of superannuated social infrastructure. This paper presents a methodology for detecting material defects using acoustic signals in a hammering test. The approach comprises a feature extraction step using Short-Time Fourier Transform (STFT) and a classifier training step based on AdaBoost, an ensemble learning algorithm. Especially, we use weak learners based on a simple template matching method that can consider both the variable scale of amplitude and the variable frequency band. The experiments discriminate between defective and clean materials using different hammering test methods: rubbing and tapping. I. INTRODUCTION In recent years, superannuation of social infrastructure has become a major problem involving installations such as tun- nels and bridges built during Japan’s rapid economic growth era. Early detection of problems by continuous inspection of that infrastructure is indispensable. However, a huge amount of infrastructure needs inspection [1]. Moreover, the locations to be inspected, such as high and narrow places, are dangerous for workers in many cases. It is extremely difficult to inspect all of them manually. Therefore, development of an automated inspection system, such as one using robots, is strongly desired. At equipment inspection sites, visual diagnosis and percus- sion diagnosis (Fig. 1) have been widely used. Particularly percussion diagnosis is mainly adopted because of its high accuracy and ease of execution. However, manual diagnosis relies on personal skill. Much experience is necessary for accurate diagnosis. Furthermore, skilled inspectors are de- creasing in number because of their retirement age. Devel- opment of automated diagnosis methods that can be executed quickly, accurately and easily is urgently in demand. Although many studies of diagnostic systems of infras- tructure inspection have been made, such systems are not efficient enough because most of these are manual system depending on visual inspection by remote operators. How- ever, automated hammering test robots have been developed, such as one which detects cavities from inner walls of concrete tunnels [2] and one which detects tile exfoliations from outer walls of high-rise buildings [3]. However, these robots are difficult to install and use because they are large- scale systems and their diagnostic methods have not become automated enough. BETOSCAN [4], a sensor equipped in compact robotic systems, can detect corrosion in reinforced 1 H. Fujii, A. Yamashita and H. Asama are with the Department of Precision Engineering, Faculty of Engineering, The University of Tokyo, 7- 3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan. {fujii, yamashita, asama}@robot.t.u-tokyo.ac.jp Rubbing Tapping Percussion stick Fig. 1. Hammering test (rubbing and tapping). concrete decks. Nevertheless, automation of diagnostic pro- cesses is limited to specific problems when the diagnosis is based on detailed analysis of a material or a structure. This limitation is common among numerous inspection systems. A considerable number of proposed automated diagnostic techniques for infrastructure are based on image process- ing [5], [6] or machine learning methodologies such as Support Vector Machine [7] and Neural Networks [8]. As one example of acoustic diagnosis, a diagnostic decision- support system of concrete pipelines was developed by Iyer et al. [8]. The study used ultrasonic signals and presented a methodology based on Multi-Layer Neural Network to detect multi-modal defects such as holes and cracks of various sizes. Although these methodologies can support human work such as walk-around checks, they are insufficient from the perspective of automation of huge-scale inspections. These facts underscore the necessity of a methodology that can diagnose defects quickly, precisely, and automatically. As described in this paper, a proposed methodology can construct a classifier adaptively to detect defects for a diag- nosis. We specifically examine a hammering test, which pro- vides an accurate diagnosis with ease of execution. A method of extracting feature vectors from acoustic signals obtained in the test and a method of construction of classifiers based on a boosting algorithm are presented. Using crack detection experiments, we verify the proposed method. II. AUTOMATED DIAGNOSTIC METHOD A. Hammering Test Hammering tests using a special stick or a hammer, called a percussion stick, are widely used for inspection work. The diagnostic tests include several methods such as rubbing by the sound of stroking on the material surface or tapping by the sound of hitting, as shown in Fig. 1. Both of these