Design And Verification Of Musical Instrument Recognition Algorithm Architecture Based On Memristor

With the rise of artificial intelligence technology,the model of artificial neural networks has become more and more complex,and the requirements for computer computing performance have become higher and higher.In traditional Von Neumann computers,the physical separation of storage and reading and writing will result in high transmission costs and data delays,which greatly limits the computing speed;when using ultra-high-performance GPU processors to implement deep learning algorithms,its complex The hardware circuit also has the characteristics of high power consumption.In order to solve the above problems,there are currently two mainstream methods:1.Break the von Neumann structure and realize the integration of computing and storage;2.Use new low-power devices that meet the computing needs to replace the CMOS devices of traditional computers.Memristors have the characteristics of multi-value storage and low computational power consumption.The neuromorphic computing supported by memristors can be matched with deep learning algorithms,which provides the possibility to solve the above problems at the same time.Compared with the simple handwritten character recognition network and speech recognition network,this paper attempts to build a musical instrument recognition network that is rarely studied and finally combines the winding machine neural network with the memristor to realize the musical instrument recognition circuit system.The main research work includes:（1）Pt/HfO₂/ZrO₂/Ti memristor devices were successfully fabricated.The DC characteristics and pulse characteristics of the memristor were tested,the device mechanism was analyzed,and the electrothermal coupling model was established by using multiphysics CMOSOL to simulate the SET and RESET processes of the device,and the two-dimensional distribution of oxygen vacancy concentration,voltage and temperature physical variables were analyzed.The results show that the memristor devices prepared in this paper are oxygen-vacancy conductive filament-dominated devices.（2）Implement the convolutional neural network musical instrument recognition algorithm architecture with memristor as the carrier.Based on Tensorflow,a convolutional neural network training framework for musical instrument recognition was built.By adjusting the parameters of the model,it achieved 95.3%and 93.1%accuracy on the training set and test set,respectively.Optimize the trained musical instrument recognition model:prune the convolution channel,prune the convolution kernel parameters and normalize the data,which greatly reduces the scale of the network structure with the least loss of accuracy;Quantize the trained network model so that the weights of the convolution kernels are accurately mapped to the memristor cross array;design a parallel memristor calculation array,and use the mixed copy mapping operation for the convolution layer to improve the degree of parallelism.Reduce computation time.（3）Design the peripheral circuit to complete the input and arithmetic processing of the audio signal,and realize the musical instrument identification circuit system.The input module of the audio signal is designed with bit-wise coding,the circuit control scheme is designed according to the memristor cross-array,the output module is designed with the differential operation and analog-to-digital conversion,and the activation and pooling module is designed according to the integrity of the convolutional network.Finally,the workflow of the whole circuit is simulated,a circuit simulation platform based on C++language is designed,and a visual human-computer interaction interface is designed.Finally,the circuit simulation platform built has achieved 91%accuracy of musical instrument recognition in the experiment and supports simulation by changing the parameters of the memristor and the quantization bits of the network model.The design and verification of the algorithm architecture in this paper also provides a foundation for the realization of large-frame convolutional neural networks.