Design Of Binary Neural Network System For One-dimensional Time Series Signals

The one-dimensional time series signal has been increasingly utilized in both life and industry for its strong dependence on temporal attributes and large capacity of informative contents.For example,audio event detection and optical network failure prediction are the two kinds of one-dimensional time series signal tasks that are practically useful in applications such as multi-scene analysis and communication information transmission respectively.Deep learning algorithms have been widely concerned in academic research for they outperform traditional algorithms in the tasks of audio event detection and optical network failure prediction.Owing to substantial parameters and the iterative calculation process,the deep learning models are capable of fitting a large number of abstract data.However,large quantity of calculation resources and storage resources are required in the complex training process of the deep models.These factors result in the fact that the traditional floating-point deep learning models are mostly implemented on the GPUs suffering from high price and high power consumption.Consequently,the hardware implementation of deep learning based one-dimensional time series signal processing system remains a challenging problem in both academic and industrial fields.Binary Neutral Network(BNN)converts the weights and activation values in the original floating-point model to two values,which improves the training efficiency of the model as well as compresses the traditional deep learning model.Thus,adopting BNN can effectively reduce the requirements of computing resources and storage resources for deep model and can further promote the development and deployment of deep learning based one-dimensional time series signal processing in embedded system.This thesis devises BNN based one-dimensional time series signal processing systems,the capacities of which are verified in the tasks of audio event detection and optical network failure prediction.The main contributions and innovations are as follows.We study the design of BNN systems towards float-pointing one-dimensional time series signal processing networks.We first form the floating-point data based Deep Neural Network(DNN)and Convolutional Neural Network(CNN)as baseline models.We binaries these networks to construct Binary Deep Neural Network(BDNN)model and Binary Convolutional Neural Network(BCNN)model respectively.On the premise of little performance loss,these binary models are analyzed and testified in the task of audio event detection and optical network failure prediction.In terms of audio event detection,the experimental results on Dcase2018 birdsong detection task show that the BCNN model designed in this paper retains 82.9%of the CNN model performance,Binary Deep Neural Network(BDNN)model even outperforms the Deep Neural Network(DNN)model by 0.6%.Moreover,the binary network model has stronger generalization performance on the new data set and faster convergence speed in the training process.In terms of optical network failure prediction,we elaborate extensive experiments in the physical layer data of WDM network provided by Guangdong Telecom operators.The experimental results suggest that the BDNN model reaches 99.7%accuracy in the failure prediction task,which is equivalent to floating-point DNN model,while enjoys from faster model convergence speed and stronger generalization performance.Finally,this paper conducts a comparative analysis of the hardware computing resources between the binary neural model and the conventional floating-point model in the optical network failure prediction task.The resource occupancy of different bits in the multiplier,adder and comparator is compared on FPGA in our experiments.More detailed comparisons are made on theoretical resource occupancy before and after the binarization of a single layer neuron in the network.Whereas BDNN achieves comparable performance with DNN,it only takes 1.43%of the resources that required by the DNN in the calculation of single layer forward propagation process.Besides,the multiplication and addition operation of the original network can be simplified to exclusive or operation by two values in the BDNN.