Abstract:
The degradation of high energy piping systems is very complex to simulate due to the many variables that influence the useful lives of such systems. Estimation of the extent of degradation is however important in the maintenance planning process. In this research the use of data driven machine learning techniques to deal with this complex problem is investigated. A hybrid recurrent neural network is created that consists of a combined recurrent neural network and a feed forward neural network. The hybrid model is trained on historical data that has been captured over a six-year time period. The following variables related to piping system components are used as inputs to the machine learning model: operating temperature and pressure time history, distance to the closest anchor point, distances to neighbouring supports, elevation survey readings, as well as the last known creep damage measurements on the component. The model is created in Python using the Tensorflow library. A recurrent neural networks (RNN) is employed, namely the gated recurrent unit (GRU). The adaptive movement estimation optimization algorithm, called Adam, is used to optimize the machine learning model. The trained model is able to predict the degradation classification of a component with an accuracy of up to 92% on the training dataset and up to 55% on the validation data set. When using this model to predict components with high creep damage, more than 400 voids per a hit rate of 25% is achieved. The current system employed at operating power stations shows a historic hit rate of 14%. This is a significant increase in performance and could be used to compile more efficient inspection plans. The model is successful in recognising patterns within the data and offers an automated way to parse large data sets that consist of a temporal and static data mixture simultaneously. Conventional data driven models are only able to look at either temporal data or static data. This suggests a generic approach to make objective decisions on similar complex data driven problems and its application is not limited to this particular problem. The methods applied in this research are expected to perform even better on problems where the frequency of data collection is higher than what is used in this research.
