METHOD FOR ASSESSING THE NATURE OF A METAL FRACTURE USING NEURAL NETWORK CLASSIFICATION AND FRACTAL ANALYSIS Russian patent published in 2022 - IPC G01N19/08 G06N20/00 

FIELD: material engineering.

SUBSTANCE: invention relates to the field of research methods for determining

the causes and mechanisms of destruction of products, i.e. the ratio in metal

fractures of the viscous and brittle component under various types of load over a

wide temperature range. Essence: preliminary tests are carried out for impact

bending of the metal under study in the temperature range from -100°C to +20°C.

Images of fractures are obtained for each test sample with high sharpness and

fractographic analysis is performed to identify the viscous and brittle component in

the fracture. Preliminary processing of the fracture image is carried out, while the

fracture zone is selected, image brightness equalized, wavelet filtering and

binarization of the selected zone are performed, after which the image is divided

into 100 equal-sized sections, and the fractal dimension of the image is evaluated.

The obtained data serve as input data for an artificial neural network (ANN)

classification, at the output of which 2 classes are obtained, namely a viscous and

brittle fracture in the fractions of each class, which in total give 100%. In the

process of training the neural network, the output parameters of the ANN are

compared with the results of fractographic analysis, and the ANN is considered

trained when an error is reached at the comparison stage for all the studied material

fractures of no more than 3%. After that, the ANN works in an automated mode and

receives at the output the percentage ratio of the viscous and brittle component in

the fracture for the studied fracture images obtained by pre-processing them,

splitting them into several zones and calculating the fractal dimension of the image.

EFFECT: improving the accuracy of the assessment of the nature of the metal

fracture.

1 cl, 2 dwg, 2 tbl