Time Series Feature Recognition Design Based On Liquid State Machine

Electromagnetic signal is a kind of time series,which will interfere with the normal operation of other devices.Accurate identification of electromagnetic signals is very helpful to the safe operation of power equipment.The data in this paper come from electromagnetic signals actually collected in the laboratory,and noise is added to electromagnetic signals in some comparative experiments.Before the experiment,the electromagnetic signal is preprocessed first.Considering that the electromagnetic signal itself is too small and its characteristics are not obvious enough,it is enlarged,and then reduced to the original size after the training is completed.Liquid State Machine(LSM)is a brain-like model,which is more consistent with biological structure.This paper selects this method for signal prediction and recognition.Based on the above pre-processing methods,prediction and recognition methods,this paper realizes the accurate recognition of the electromagnetic signal collected.The main contents of this paper are as follows:1)Build a liquid state machine model.In this paper,the structure of liquid state machine is described in detail.It is a kind of pulse neural network and a recursive neural network.Two types of liquid state machine models are constructed,which are prediction model and recognition model respectively.The prediction model realizes the basic structure of liquid state machine.While recognizing the basic structure of the model,the back propagation mechanism is added to make the model more concise and improve the effect.2)Design the prediction method based on liquid state machine.The electromagnetic signal is preprocessed and put into the network.The method of liquid state machine is used to construct three layers,including the input layer,the reserve pool and the output layer.The liquid state machine prediction model is constructed,and the final result is obtained through linear regression.Using pulse signals in the model not only achieves higher dimensional mapping,but also achieves prediction,which is better than the echo state network calculated by the same reserve pool.3)Design the recognition method based on liquid state machine.By using Binds Net framework,a basic liquid state machine model is built,and a backpropagation mechanism is added between the reservoir layer and the output layer in the basic model.Change the way predictive inputs are entered one by one,and enter them in groups.The recognition model is more complex than the prediction model.Experimental comparison between echo state network and deep echo state network similar to liquid state machine proves that the training accuracy of the liquid state machine recognition model is improved and the training time is relatively short.Through the recognition and prediction of electromagnetic signal,the liquid state machine method is verified to be effective,and pulse neurons can better identify and predict.Liquid state machine added back propagation mechanism,can better electromagnetic signal feature recognition.In the future,the mechanisms of synaptic plasticity could be altered to allow fluid states to better learn the characteristics of electromagnetic signals.