V.N. Klyachkin – Dr.Sc.(Eng.), Professor, Department «Applied Mathematics and Informatics», Ulyanovsk State Technical University
Yu.E. Kuvayskova – Ph.D.(Eng.), Associate Professor, Department «Applied Mathematics and Informatics», Ulyanovsk State Technical University
D.A. Zhukov – Post-graduate Student, Department «Applied Mathematics and Informatics», Ulyanovsk State Technical University
For the safe operation and reliability of the equipment important diagnosis of the current status and prediction of the condition in the fu-ture. At the same time, both standard statistical methods and methods of machine learning are actively used. Recognition of the state of the object is carried out according to the results of performance measurement. As the initial data is considered many precedents: objects with the specified performance and corresponding states. On the basis of these data, it is necessary to restore the relationship between performance and the state of the object. This is a special case of one of the tasks of machine learning is classification when learning on precedents. These are classical statistical models (logistic regression, discriminant analysis, Bayesian classifiers), and methods specially oriented to machine learning (support vector method, neural networks), compositional methods (bagging, boosting) and others. The quality of the classification depends on a number of factors: the volume and quality of the original sample, method of machine learning, source separation method of sampling for training and reference parts. It is of interest to share the various classification methods built on the training sample. In order to achieve the best result, it is necessary to solve questions about which methods of machine learning to use together, how to combine them, and how to make a single decision about the integrity of the object based on the decisions of indi-vidual methods? To create a single health solutions unit based on the decisions of individual classification methods, reviewed the results of aggregation by mean, median, and with the help of the voting procedure. Use the full search of all possible sets of base methods. For evaluating the quality of models in terms of predicting the original sample divided into two disjoint subsets: learning sample (with the help of which the task of learning) and reference (or test), not to be used for training. When using cross-validation sample is divided into k parts. (k − 1) is used for training, one is for control. All variants are successively sorted. For each split task learning and evaluate the function errors on the control sample. When unbalanced classes the percentage of errors cannot objectively evaluate the quality of the classification. Much more informative accuracy and completeness, depending on the number of correctly classified working conditions number of misclassified healthy conditions and number of misclassified defective States. Based on these two indicators can be formed a single criterion is the harmonic mean of precision and completeness (F measure): the closer the F value for the unit, the quality classifi-cation of the above. Application of aggregated classifiers in the examples showed higher accuracy of diagnosis of an object rather than using the basic techniques of machine learning (percentage of erroneous decisions on control sample decreased 3−11% comparison with the best of the base methods). To improve the efficiency of machine learning in diagnostics the state of equipment, it is necessary to de-velop a system for studying the influence of various factors on the quality of classification with the source data for a particular facility that would ensure the application of optimal approaches.
- Klyachkin V.N., Kuvajskova Yu.E., Buby’r’ D.S. Prognozirovanie sostoyaniya ob’‘ekta s ispol’zovaniem sistem vremenny’x ryadov // Radiotexnika. 2015. № 6. S. 45−47.
- Montgomery D.C. Introduction to Statistical Quality Control. New York: John Wiley and Sons. 2009. 754 r.
- Ryan T.P. Statistical Methods for Quality Improvement. New York: John Wiley and Sons. 2011. 687 r.
- Klyachkin V.N. Statisticheskie metody’ v upravlenii kachestvom: komp’yuterny’e texnologii. M.: Finansy’ i statistika, INFRA M. 2009. 304 s.
- Witten I.H., Frank E. Data Mining: Practical Machine Learning Tools and Techniques. San Francisco: Morgan Kaufmann Publishers. 2005. 525 r.
- Sokolov E.A. Mashinnoe obuchenie. URL = http://wiki.cs.hse.ru/ Mashinnoe _obuchenie. _1/2016_ 2017.
- Zhukov D.A., Klyachkin V.N. Diagnostika ispravnosti texnicheskogo ob’‘ekta s ispol’zovaniem paketa Matlab // Trudy’ Mezhdunar. nauchno-texnich. konf. «Perspektivny’e informaczionny’e texnologii». Samara: Izd-vo Samarskogo nauchnogo czentra RAN. 2018. S. 55−57.
- Klyachkin V.N., Kuvajskova Yu.E., Alekseeva V.A. Statisticheskie metody’ analiza danny’x. M.: Finansy’ i statistika. 2016. 240 s.
- Klyachkin V.N., Shunina Yu.S. Sistema oczenki kreditosposobnosti zaemshhika i prognozirovaniya vozvrata kreditov // Vestnik komp’yuterny’x i informaczionny’x texnologij. 2015. № 11 (137). S. 45−51.
- Klyachkin V.N., Kuvayskova Yu.E., Zhukov D.A. The Use of Aggregate Classifiers in Technical Diagnostics, Based on Machine Learning // CEUR Workshop Proceedings. V. 1903. Data Science. Information Technology and Nanotechnology. 2017. P. 32−35.
- Kuvayskova Y.E. The Prediction Algorithm of the Technical State of an Object by Means of Fuzzy Logic Inference Models // Procedia Engineering «3rd International Conference «Information Technology and Nanotechnology (ITNT 2017)». 2017. S. 767−772.