Pipeline Network System Leak Fault Diagnosis Based On Data-driven Methods

With the development of science and technology, the status of the pipe-networktransportation becomes more important. It has even become an integral part of the materialsource, so almost every country pays more attention to apply the pipeline network safelyand environmentally. More than half of the pipe-network has come to service around theworld more than three decades, now the pipeline’s aging and corrosion wear and abrasionbecome a serious problem. Pipeline leakage and pollution incidents happened frequentlyin these days. Our country pays more and more attention to the pipe-network’s leakdetection and fault dignosis, it is crucial to extrat the characteristic values which canreveal the fault state effectively and feasibly.Through the basic principles analysis of the oil pipeline leak location, we build anexperiment platform of the pipeline leak fault diagnosis system, and we use Labviewsoftware to acquire the negative pressure wave and acoustic signals in real-time. Wepropose two methods for signal feature extraction based on the acquired data. We use thewavelet packet transform to eliminate the noise through the Matlab software. Throughfurther study on wavelet packet transform, at last I extract the energy value of the complexpipeline signal which can highlight the leak feature. The traditional signal processingmethods are based on the theory of the Fourier transform, so they cannot get rid of theshortcomings of the Fourier transform, such as the fixed basis function, poor adaptiveability and so on. So this paper applies the empirical mode decomposition method toanalyze the nonlinear transient signals. In order to avoid the modal aliasing in lowfrequency, we use an improved algorithm of EMD. Firstly, reconstructing the signalthrough second generation wavelet interpolation, and then making the reconstructed signaldecomposed by EMD, thus it can increase the accuracy of empirical mode decomposition.We choose extract feature vectors in the perspective of energy form like the waveletpacket, in other words, it means each IMF component energy after EMD decomposition.lay the foundation for the research of subsequent neural network fault diagnosis.This paper applies a three-layer BP neural network, including the input layer, hiddenlayer and output layer. We regard the extracted feature vectors as input, the normal andleak state as output, and then training the netwoek offline. After the network trainingcomplished, putting the test samples to test the established neural network, the results show that both methods can achieve the pipeline network fault diagnosis. Compared to thenetwork performance established through two feature vectors, we find that the improvedEMD method has higher accuracy and precision. The research provides a good theoreticalbasis for applicating the pipe network fault diagnosis system in practice.