Classification of tensile test results of unidirectional carbon fiber-polysulfone composite material based on random forest, KNN and CNN methods

Creation and quality control of products from unidirectional composite materials based on polymer matrices reinforced with carbon fiber is one of the urgent tasks of modern aviation, automotive and power engineering. When solving this problem, it is necessary to have reliable classifiers that allow engineers to choose carbon fiber reinforcing polymer matrix based on its influence on mechanical properties. The purpose of our work is, based on the random forest method, the K nearest neighbors’ method and convolutional neural networks, to develop a reliable model for classifying the type and grade of carbon fiber based on the results of tensile tests of unidirectional carbon fiber-polysulfone composite material and to give an interpretation of the constructed models.Tensile tests of samples of unidirectional carbon fiber-polysulfone composite materials obtained by impregnation of dry carbon filaments with a solution of thermoplastic polymer-polysulfone in n-methylpyrrolidone with subsequent removal of the solvent have been carried out. The deformation behavior of samples based on high-strength and high-modulus carbon fibers under different test modes has been investigated. Based on the experimental data, six models of classification of the results of tensile tests of unidirectional carbon fiber-polysulfone composites by type of carbon fiber (high-strength or high-modulus) or by brand of carbon fiber were developed.It is shown that classifiers based on the random forest method have higher classification quality metrics than classifiers based on the KNN method and convolutional neural network. When classifying the test results by carbon fiber type (high-strength or high-modulus), the correlation coefficient of Matthews in the random forest model is 0.985, in the KNN model is 0.963 and, in the CNN, model is 0.880. In building multiclass classification models using principal component analysis, it was demonstrated that F1 metrics and Matthew's correlation coefficient are better for comparing classifiers. On the validation data, it was demonstrated that the best metrics are possessed by the random forest model with a median F1-score of 0.993, and the convolutional neural network model ranks second in terms of classification quality.Despite the high-quality metrics of the random forest-based classification model, when analyzing the contributions of explanatory variables to the classification results, it was found that the random forest model underestimates the effect of polymer concentration in binary classification. Whereas the convolutional neural network model underestimates the change in strain corresponding to the tensile strength.Interpretation of the binary classification models shows that reinforcement of polysulfone with high-modulus fiber provides high ductility of the composite material, but low strength relative to the composite material reinforced with high-strength fiber. Whereas the presence of high-strength fiber in the composite material provides high strength but low ductility relative to the material containing high-modulus carbon fibers. The results revealed in the interpretation of classification models agree with the previously conducted results of statistical analysis of experimental data.In multiclass classification, Toray M35J and Toray T700 fibers were found to have the highest probability of establishing the carbon fiber grade based on the tensile test results of unidirectional composites. The use of these grades in the composite material provides high strength over the entire range of polysulfone concentrations, compared to UMT400 and Toho Temax HTS40 fiber grades.As a result of the conducted research conducted, classification models with high quality metrics were obtained, which allow us to choose a brand or type of carbon fiber in the development of unidirectional carbon fiber-polysulfone composite material.