Multiple Engine Faults Detection based on Variational Mode Decomposition and Echo State Network 2020-01-0418
As a major power source, diesel engines are widely used in a variety of fields. However, because of complex structure, some faults which cannot be detected by direct signals would occur on engines and even lead to accidents. Among all kinds of indirect signals, vibration signal is the most common choice for faults detection without disassemble because of its convenience and stability. This paper proposed a novel approach for detecting multiple engine faults based on engine block vibration signals using variational mode decomposition (VMD) and echo state network (ESN). Since the quadratic penalty has a great influence on adaptable VMD that might make expected component signals cannot be extracted exactly, this paper proposed a dynamic quadratic penalty value, which will change with decomposing levels. This paper selected a best dynamic quadratic penalty value by analyzing a large amount of data and results showed that this approach can obtain more exact decomposing results. Based on decomposing results, 8 characteristic parameters were extracted to describe faults features comprehensively, while high-dimensional data raised computational difficulty for classifier. To improve the recognition rate of ESN for high-dimensional data, this paper proposed an approach that using dimensionality reduction method, in which the kernel local Fisher discriminant analysis (KLFDA) based on Gauss kernel function was proved to be the better choice, to reduce the dimension of original data firstly. Then the ESN was used to recognize faults by analyzing the low-dimensional data. Results showed that this approach obtained ideal recognition rate and had the ability to detect multiple engine faults exactly.