Feasibility study for the application of a neural network for operating condition detection of a centrifugal pump

Abstract

Artificial intelligence (AI) technology is successfully used in many fields. Originally, AI was developed for image and voice recognition and has been spread out on other fields like detecting deceases, behaviour, traffic movement etc. AI is also applied in fluidmachinery, for example in order to detect defaults in pumps, fans, compressors and turbines. This work verifies the feasibility for the application of a neuronal network (NN) for operation point detection of an impeller pump by vibration signals. Once the NN is trained with the vibration signal, it can recognize the operating point by the vibration signal of the impeller pump. In order to test whether the neural network method is feasible, a pump loop system was built up, which contains a radial centrifugal pump with a replaceable impeller. The vibration signal of the impeller pump is taken for different operating points and used for training a NN. In the first step a simple self-programmed NN in C++ is applied in order to find out a convenient NN-topology for operating point detection. The congruence rate of this simple NN method already reaches about 90%. The research results in this paper show that an operating point can be detected by the application of a simple NN. It is carried out how large the influence of the topology of a NN is in regard to the congruence rate (CR). By further application of AI like convolutional layers, batch normalization etc. a better CR seems to be very likely. From this point of view this contribution reports about a starting point of operating point detection of impeller pumps by the application of a NN.