Smart-RAM: intelligent system for defect detection by acoustic resonance

Abstract

The project proposes the manufacture of an automated test station for the detection of defects in the production line, based on the resonant acoustic spectroscopy technique (ASTM-E2001), where the establishment of acceptance criteria for the classification of the parts is assisted by machine learning algorithms capable of efficiently detect micro defects in small parts and parts with complex geometry. The system to be developed can be used in the inspection of sintered metal parts, abrasives, brake pads, biomaterials, technical ceramics, porcelain, tiles and cast-iron parts. The resonant acoustic spectroscopy method is well established and is widely used in the industry for the classification of metallic and sintered parts. The existing equipment are based on the establishment of criteria (windows for the location of the frequency peaks, acceptable values for the relationship between the measured frequencies, etc.) that are normally established by a specialist from a behavioral study of the spectrum from the acoustic response of parts with and without defects. However, in small parts or parts with complex or very asymmetric geometry, the characteristic frequencies involved are high, which results in a low signal-to-noise ratio, and a spectrum with a complicated profile. In these cases, it is not possible to apply this methodology or, if applied, it results in very low efficiency, which in many cases leads to the application of visual inspection methods, which are extremely expensive and unreliable, mainly in production lines with a high flow of pieces per minute. Another disadvantage of the systems used today is that, due to the complexity of the knowledge involved in the detection process, fundamentally related to the application of digital signal processing techniques, the operation of these systems requires personnel with a high level of training, knowledge and familiarity with the software to establish criteria that lead to satisfactory classification results. The expectation lies in the manufacture of equipment of competitive cost and simple operation, which can solve the problem of automated classification, including for small parts and complex geometry, with an innovative design methodology, at the international level. For this, it is expected to use the latest advances in the area of machine learning (machine learning) in conjunction with connectivity and Cloud storage, which will allow a continuous improvement of the detection algorithms by incorporating the data collected during the operation of the equipment. Other features may also be incorporated into the equipment, depending on the client's interest, since, once the acoustic response is obtained, variations in other properties such as density or composition, for example, can be detected and used in production feedback. (AU)