Development of diagnosis algorithm for the check valve with spectral estimations and neural network models using acoustic signals

We have developed a method for detecting and diagnosing a disk wear failure and a foreign object failure among the various failure modes of check valves. The method is based on the acoustic emission sensors which can detect the sound wave of the leakage flow and the estimation of the power spectral densities with an auto-regressive model. For validating the method, we implemented a hydraulic test loop with an artificially failed check valve. We have found that the frequency spectrums from the acoustic signals are strongly dependent on the failure modes of the check valve and that they are nearly independent of the failure size and operating pressure through an estimation of the power spectral density with an auto-regressive signal processing model. In addition, the root mean square values of the acoustic signal and the amplitudes of the power spectral density as well as the loop pressure have a strong dependency on the failure size in each failure mode of the check valve. We developed a diagnosis algorithm by using neural network models in order to identify the type and size of the failure in the check valve. The diagnosis algorithm consists of a hierarchical model composed of three back-propagation neural networks. The results of our research and the experiments show that the diagnosis algorithm is proven to be a good solution for identifying the failures of the check valves without any disassembling work.