An application of BRANN and MFL methods: Determining crack type and physical properties on M5 steel sheets

Abstract

Magnetic flux leakage (MFL) method is used frequently in researches on cracks on ferromagnetic materials. Recent studies on this method focused on the improvement of low power consumption and sensor sensitivity. Generally, it was determined as to how the changing characteristics of MFL signals collected through magnetic sensor and the size of amplitude value of signal were changed based on the physical properties of crack on the ferromagnetic material (width of crack, depth of crack, etc.). As different from literature, primary purpose of our study was determining how this change occurred based on the type and physical properties of an artificial crack that were with known physical properties and that were formed with M5 directed steel layers; and the secondary purpose of our study was to develop an artificial neural network estimating the type and physical properties of a crack with unknown physical properties in the light of the data obtained. Accordingly in our study, firstly a magnetic measurement system was produced consisting of a mechanical scanning system to ensure three-dimensional movement of sensor, of a data collection module to process the data from sensor and to send the same to the computer, and of a computer software to evaluate the data from computer and to record the same. Then, artificial crack samples of different types and physical properties were prepared from M5 samples in the shape of a plate. These artificial crack samples were magnetized by placing on a core that was converted into electromagnets using 50 kHz AC signal; and the surface of material was scanned one-dimensionally with position controlled fluxgate sensors. The sensors created in fluxgate sensor based on the position were examined in terms of harmonics with DSP Lock-in Amplifier; and the amplitude values of harmonics showing the biggest change were included in the computer. Following the determination of changing graphics based on the scanning lengths of MFL signals, the mathematical curves and formulas best suiting to the characteristics of such change were determined. The changes of variables in such mathematical curve formulas were analyzed based on the type and physical properties of crack. Lastly, 4 different BRANNs (Bayesian regulated artificial neural networks) were developed estimating the type and physical properties of crack that were using the MFL signals of artificial crack samples with known physical properties and that were trained accordingly and that had unknown physical properties. First of those four was used to determine the type of crack and the other three were used to find the depth, width, lower and upper sound thickness values of crack based on the type of crack. Accuracy degrees were obtained from those BRANNs that were obtained during the training stage, respectively as R = 0.998, R = 0.959 and R = 0.964. The BRANNs trained provided results corresponding to the actual for artificial crack models with unknown crack types and physical properties.