ALUMINUM PLATE DEFECT DETECTION BASED ON MULTILEVEL LSTM

Aluminum plates are widely used in fields such as aviation, aerospace, and construction due to their excellent fatigue resistance and ductility. However, during the production process, various defects can occur due to external environmental factors, operating processes and so on. These defects can affect the mechanical properties of aluminum plates, such as reducing their strength, ductility, and toughness, which leads to a shortened service life. This article proposes a method of using a multilevel long short-term memory (LSTM) neural network for assisting the detection of inclusions (void defects) in aluminum plates under the condition of single transmission and reception of ultrasonic waves. The propagation of ultrasonic waves in an aluminum plate containing inclusion defects is simulated using the finite element software COMSOL Multiphysics. Waveform data containing defect information is then derived. By training the waveform data, a neural network model reflecting the relationship between waveform data and the size and location of inclusions is obtained. In addition, we adopt a hard voting method to alleviate the problem of complex parameter adjustment during network training and improve the reliability of the detection results. The results show that the accuracy of radius detection for inclusion defects exceeds 98%, the accuracy of depth detection for inclusion defects reaches 1, and the accuracy of horizontal position detection for inclusion defects exceeds 95%. It provides reference for the application of LSTM neural network in ultrasonic non-destructive testing.