The article discusses the process of planning the repair of energy equipment. Using a decision support system is proposed because of the large number of rules of comparing flow charts of technical defects. Such a system can speed up the planning process and reduce economic costs. A conceptual model of the system has been built; further it will be presented as a multi-label classification of cross-cutting classes. The “one-vs-all” approach has been used: each flow chart can use its individual classifier. Metrics are proposed for evaluating classifiers: a portion of accurately classified objects, precision, fullness and F-measure. To summarize the evaluation results the concept of micro-average was chosen. A defect classification algorithm has been described. An experiment was conducted using different classification algorithms: decision trees, Bayes classifier and multilayer perceptron. The results of the experiment proved that 80-90% of the correctly classified objects were found (high values), but the average values of accuracy and fullness occurred low (3-7%). There were found sets of data, where different output data corresponded to similar input data. Thus, machine learning can be used to support decision-making, but in some cases information about the order is not complete. Defect classification can be combined with manual clarifying of results or with different algorithms.
Paper
Full text
USING MACHINE LEARNING METHODS IN POWER EQUIPMENT REPAIR PROGRAMS
Semantic Scholar · Engineering · 2019
Abstract
The article discusses the process of planning the repair of energy equipment. Using a decision support system is proposed because of the large number of rules of comparing flow charts of technical defects. Such a system can speed up the planning process and reduce economic costs. A conceptual model of the system has been built; further it will be presented as a multi-label classification of cross-cutting classes. The “one-vs-all” approach has been used: each flow chart can use its individual classifier. Metrics are proposed for evaluating classifiers: a portion of accurately classified objects, precision, fullness and F-measure. To summarize the evaluation results the concept of micro-average was chosen. A defect classification algorithm has been described. An experiment was conducted using different classification algorithms: decision trees, Bayes classifier and multilayer perceptron. The results of the experiment proved that 80-90% of the correctly classified objects were found (high values), but the average values of accuracy and fullness occurred low (3-7%). There were found sets of data, where different output data corresponded to similar input data. Thus, machine learning can be used to support decision-making, but in some cases information about the order is not complete. Defect classification can be combined with manual clarifying of results or with different algorithms.