The use of machine learning for various predictive models of the occurrence of pipe defects

The paper analyses and compares possible predictive models for predicting pipeline defects using machine learning. As an example of the models, the data from in-tube diagnostics for stress corrosion cracking are selected. Special attention was paid to the description of the learning process of models based on machine learning algorithms based on retrospective data. Extended metrics are presented for the prepared parameters of the defects themselves and related data from the survey reports. The performance metrics of the algorithms are given in comparison to the probabilities of a correct prognosis. Conclusions have been drawn on the correlation of defects of this type with data on the soil composition of the soil of the segment of the analysed pipe. The analysis of statistical data on the defects by the pipe orientation was carried out. An approach has been proposed and tested to improve the accuracy characteristics of the model based on the Random Forest algorithm using preliminary data selection, which made it possible to achieve a heuristic probability of more than 80%.