Discrete-synapse recurrent neural network for nonlinear system modeling and seismic signal classification.

摘要

A passive seismic sensor is very useful to a perimeter protection system since it is cheap, easily deployable and concealed underground so as not to be detected by intruders. However, its limited frequency response is challenging enough to necessitate wider temporal analysis or a hybrid method which combines spectral analysis and cadence analysis including gait pattern analysis. Although they are mutually complementary and independent analysis methods, they are computationally expensive and cause trade-off issues during the simplification process for hardware implementation on a small chip. In terms of analysis method, although Dynamic Synapse Neural Network (DSNN) as a biologically-inspired model is a good candidate for more fundamental understanding of seismic signals capturing its dynamics with wider temporal information, it is also computationally expensive in the terms of hardware implementation and inefficient in training.;In this thesis, a simplified model of DSNN, Discrete-Synapse Recurrent Neural Network (DSRNN) is explored and developed so that the model not only does the seismic signal classification task with a more fundamental understanding of the signal, but also does the nonlinear system modeling task learning of a DSNN as a nonlinear system, establishing the DSRNN as a simplified and replaceable model of DSNN.;DSRNN I was designed based on the differential equations for DSNN's presynapse and nonlinear characteristics of the other part of DSNN. Also DSRNN II was designed with additional recurrent connections to DSRNN I so that it could capture wider temporal dynamics of the signal. DSRNN II's nonlinear system modeling task learning of a 2x2 DSNN compared to DSRNN I's learning of a single DSNN proved that DSRNN II outperforms DSRNN I by virtue of added recurrent connections. Also, seismic signal classification task produced the same conclusion by a much lower false recognition rate from DSRNN II and showed the possibility of a simple hardware implementation of a biologically-inspired model without losing its functionalities.