Fatigue prediction of hybrid joints and perforated plates using neural network

Subject Areas : journal of Artificial Intelligence in Electrical Engineering

Ali  Yousefnezhad Oskooi ¹ (Computer Engineering University College of Nabi Akram, Tabriz,Iran)
vahid  Pourmohammad ² (Department of Mechatronics Engineering University College of Nabi Akram, Tabriz,Iran)
Karim  Samadzamini ³ (Department of Computer Engineering, University College of Nabi Akram, Tabriz, Iran)
Firooz  Esmaeili Goldarag ⁴ (Department of Computer Engineering University College of Nabi Akram, Tabriz,Iran)

Received: 2023-05-19 Accepted : 2023-08-10 Published : 2023-03-01

Keywords: Neural network, Artificial Intelligence, hybrid connections, bolt connections, perforated plates,

Abstract :

Hybrid connections (bolts, glue) and perforated plates are one of the most important topics in various industries, including aerospace. This type of process occurs due to the growth of small cracks in the metal structure as a result of cyclic or intermittent loading. Since failures occur suddenly, terrible accidents such as plane crashes, shipwrecks, bridge collapses, and toxic radioactive fallout can occur. To prevent these incidents, fatigue tests are performed on a sample of parts that is similar to the real part, so that the fatigue life can be obtained through this method. However, because fatigue tests are time-consuming and expensive, artificial intelligence methods have been used in this research to estimate the fatigue life of hybrid joints and perforated plates. In the experimental part of this research, plates made of aluminum alloy 2024-T3, which is one of the widely used materials in aerospace, the used materials are screws made of Hex head M5 and a special adhesive made of Loctite 3421 (Henkel ltd). Fatigue tests are extracted as input and output data from the related article. Out of a total of 71 fatigue tests, 35 tests were performed for perforated plates, 18 tests for hybrid joints, and 18 tests for bolted joints. Also, according to the number of data, the best result was when 80% of the data was considered for training the network and 20% was used as test data to evaluate the performance of the network. Finally, the predicted output was compared with the actual output and it was seen that the best performance of the neural network was after normalizing the data, that the error value was close to zero.