Classification of Microseismic Events and Blasts Using Deep Belief Network

Identification and separation of microseismic (MS) events and blasting signals is typically a challenging issue. In particular, the extraction of features from large amounts of data using shallow neural networks is inefficient and not always complete. To address this, the use of a deep belief network model to identify MS events and blasting signals during construction processes is proposed. The aim is to take advantage of the strong learning and feature extraction abilities of such models. Using data recorded during the construction of the Jinping Hydropower Station II in China, nine typical source parameters are selected to use as characteristic parameters. This is accomplished by comparing and analyzing the differences and probability density distribution ranges of the different source parameters associated with MS and blasting events. The model uses the data collected for the source parameters to learn, train, and test. A classification accuracy of 94.4% was thus achieved, which is superior to that obtained using traditional support vector machine and Fisher classifiers. The results also show that the method has other advantages compared to the other methods: false positive rates are low, the characteristic parameters are easy to acquire, and the calculation speed is fast. It is thus an effective method of identifying MS and blasting events when using MS monitoring techniques.