The goal of the thesis is implementation of a predictive system for detecting
failures in industrial machines. We tackle the problem by using different
machine learning approaches and methods. Initially, we transformed the
received data into a representation for supervised learning. In the next step
we trained the classifiers and evaluated their performance. We applied three
different approaches, as follows. In the first approach we trained memoryless
models without using historical data; in the second approach we extended the
data with additional historical attributes; in the third approach we trained
memory-retaining models (LSTM and GRU) with a non-historic dataset. On
the basis of our experimental results we discovered that the third approach
gives as the best results.
|
