Нейронні мережі як засіб удосконалення метрологічних характеристик металоконструкцій з урахуванням міжфазних шарів

Abstract

Розглянуто проблему обстежень, контролю параметрів і діагностування стану поверхневих шарів металу підземних трубопроводів з урахуванням впливу корозійного середовища. Запропоновано методику контролю характеристик, що полягає у врахуванні основних інформативних параметрів за допомогою штучних нейронних мереж, а також визначено напрями застосування методології для контролю технічного стану трубопроводів (КТСТ) (товщина стінки, наявність дефектів, енергетичні характеристики міжфазних шарів, корозійні струми, процеси розвитку дефектів тощо). Метою методології КТСТ є удосконалення нормативних документів у сфері метрології.

The problem of inspection, control of parameters and diagnostics of the state of surface metal layers of underground pipelines with consideration of influence of corrosive environment is considered. The method of control of characteristics is proposed, which is to take into account the main informative parameters with the help of artificial neural networks, as well as the directions of application of the methodology for controlling the technical condition of the pipelines (CTCP) (wall thickness, defects, energy characteristics of phase layers, corrosion currents, defect development processes and etc). The purpose of the CTCP methodology is to improve the regulatory documents in the field of metrology. Analyzing statistical data, it is found that the most suitable for most tasks for the selection of parameter sets for non-destructive testing (NDT) and evaluation of the technical state are multilayer neural networks trained by the Levenberg-Marquardt error-back propagation algorithm. The main paradigm of learning in this case is learning with a teacher. The described form of learning “with the teacher” is nothing more than training on the basis of error correction – the reverse error distribution. This is a closed-loop feedback system that includes an environment. The productivity of such a system can be evaluated in terms of the mean square error or the sum of squares errors in the training sample, presented as a function of the free system parameters. The back propagation algorithm is the most popular among the algorithms for learning multilayer neural networks. That is, it is a gradient method, not an optimization method. To implement the described sequence of artificial neural network training, it is recommended to use the Neural Network Toolbox in the Matlab environment 16. The training error during the Neural Network Toolbox setting should be 5%. This is due to the fact that, as a rule, the total level of error of measurement of target and informative parameters, as well as stochastic components does not exceed 5%. It is recommended for each case of selected complexes of informative parameters to perform training 5–7 networks of the same architecture. Such a number of networks is arbitrary, but it avoids the occurrence of an ascent of the training algorithm in the local minimum and the effect of “retraining”, which will be accompanied by memorizing target values that are relevant to informative resources, rather than establishing a relationship between them. Training all neural networks for all possible combinations needs to be tested using pre-selected test datasets that were not used during training. The obtained results of calculating the values of target parameters are compared with the standard ones by means of absolute and relative error and the calculation of their average value. Among the outputs of neural networks, the smallest is selected. As criteria of optimality, in this case, choose the following: the minimum possible set of informative parameters; highest accuracy of target parameter determination. The set of informative parameters selected according to the above criteria can be considered optimal and acceptable.

Рассмотрена проблема обследований, контроля параметров и диагностирования состояния поверхностных слоев металла подземных трубопроводов с учетом влияния коррозионной среды. Предложена методика контроля характеристик, которая заключается в учете основных информативных параметров с помощью искусственных нейронных сетей, а также определены направления применения методологии для контроля технического состояния трубопроводов (КТСТ) (толщина стенки, наличие дефектов, энергетические характеристики межфазных слоев, коррозионные токи, процессы развития дефектов, т. под.). Целью методологии КТСТ является совершенствование нормативных документов в области метрологии.