Research On Deep Spiking Neural Networks And Its Applications

Deep neural networks(DNNs),in which the partitioning mechanism of biological visual cognitive systems are used for reference,representing data as a series of vectors for feature learning is a hot research topic in the field of machine learning(ML).Computer vision has achieved remarkable achievements with the use of DNNs.Spiking neural networks(SNNs),as the main tool of "brain-inspired computing",are biological plasticity neural networks,which make use of time-varying spike trains for information transmission between spiking neurons.SNNs can integrate temporal and spatial information better.In the thesis,the advantages of DNNs and SNNs are combined to analyze model characteristics of the existing deep spiking neural networks(DSNNs),then research on spiking coding and learning methods based on DSNNs are conducted.In addition,the fault diagnosis method of manipulator based on DSNNs is studied.The specific contents are as follows:First of all,the research background and significance of DSNNs are introduced in the thesis,the research status of DSNNs at home and abroad are reviewed,then the research content and technical route of the thesis are compounded.In addition,the development,model structure of deep convolutional neural networks(DCNNs),deep belief networks(DBNs),SNNs,and the realization method and learning algorithm of DSNNs model are introduced in the thesis,providing theoretical support for the follow-up research.Also,a fault classification method based on DCNNs for manipulator was proposed.The sensor data status and preprocessing technology of UCI manipulator were introduced.The ability of DCNNs to process one-dimensional time-series signals was also analyzed.The collected manipulator force and torque sensing data are combined in time and data dimensions,and experimental verification is performed on the CPU(Intel Core i5-7200U)and GPU(GFX NVIDIA Ge Force GTX1060 3G)using 1D and 2D convolution methods.The experimental results show that the processing method of the one-dimensional time-series signal data of the manipulator can well fit the DCNNs model,and the classification accuracy is better than the traditional classification method.What's more,the spiking mechanism is integrated into the DCNNs model.Integrated-and-Fire(IAF)neurons are used instead of traditional artificial neurons.Two types of training methods are adopted for experimental verification:(1)Supervised methods are used to propose deep spiking convolutional neural network based on time-latency of frequency coding for the classification of manipulator execution failures.The experimental results show that the proposed time-latency of frequency coding method has the characteristics of both latency coding and frequency coding,but also reduces the network's inference time and calculation amount.Then,the similarity principle is introduced in the model learning stage to improve the generalization ability of the model.(2)Using unsupervised method,STDP-based DSCNN for manipulator fault classification is proposed.Winner-Takes-All(WTA)and Lateral Inhibition Mechanism(LIM)competition mechanisms are integrated into the traditional convolutional layer.The experimental results prove that the introduction of a neuron competition mechanism in the model can improve the Anti-noise capability of the network.Besides,a fault classification method for the manipulator of the deep spiking belief network(DSBN)is proposed and verified by experiments.For DSBN model,Siegert neurons were used to replace the neurons in traditional RBM.The Intrinsic Plasticity(IP)learning rules were introduced into Siegert neurons.The experimental results show that compared with the previous three models,the model training time is reduced from 125 minutes of DCNNs model to 4.16 seconds.Finally,summarize the work done and look forward to the next step.