Research On Heterogeneous Acceleration Of Deep Learning Algorithms

In this thesis,the heterogeneous hardware acceleration research is carried out for the fault diagnosis task in deep learning algorithm.According to the structural characteristics of the convolutional neural network,combined with the computing advantages of FPGA,the convolutional neural network is hardened on the ZYNQ-XC7Z020 SOC heterogeneous hardware platform.The system adopts the software and hardware coordination method,which means the input,storage and display of bearing vibration signal data is completed in the programming system(PS),and the programmable logic(PL)part performs heterogeneous computing acceleration of CNN.Through the optimization of the network calculation process and the improvement of the data storage process,the acceleration of convolutional neural networks on heterogeneous hardware platforms is realized.The experimental results show that,in terms of recognition accuracy,the average recognition accuracy of the heterogeneous acceleration system is 95.25%,which only loses 1.03%compared with the 32bit floating-point implementation on the CPU platform.The power consumption of the CPU is 13 times that of the heterogeneous acceleration system,and the power consumption of the GPU is 22 times that of the heterogeneous system.In the acceleration performance,the processing time of the heterogeneous acceleration system realizes the processing of a set of complete data is only 0.37ms,while CPU and GPU need 5.78ms and 1.19ms respectively,so the heterogeneous system has high acceleration performance.Taking advantage of the high parallelism of FPGA,a reasonable design of the input and output parallelism between the layers of the neural network are adopted to realize the high efficiency of network operation under the condition of making full use of on-chip resources.The system hardened with convolutional neural network can quickly and accurately identify bearing vibration fault signals.Compared with general CPU and GPU hardware platforms,FPGA also has incomparable advantages in terms of portability and energy efficiency.