Research On Spectrum Monitoring Algorithm Based On Machine Learning And FPGA Implementation

With the development of communication technology,real-time and accurate monitoring of the external electromagnetic spectrum environment has become an increasingly urgent requirement.Spectrum monitoring technology has more and more application prospects in military or civilian fields.At present,machine learning technology has become the development trend of the new generation of information technology.Machine learning technology has been deeply integrated with many fields,and has achieved a series of development results,which greatly promoted the transformation and upgrading of traditional industries.The Convolutional Neutral Network(CNN)algorithm in machine learning has been widely used in image detection,speech recognition and radar signal recognition because of its excellent performance in extracting features from high-dimensional data.In this paper,the convolutional neural network algorithm is combined with the spectrum monitoring technology,and a spectrum monitoring system based on convolutional neural network is proposed.Finally,the system is implemented on the Pico Zed-SDR-Z7035-AD9361 hardware development platform.In this paper,the theoretical analysis of the forward propagation and backpropagation process of artificial neural network algorithm is carried out firstly.Then the structure and function of convolutional neural network are systematically discussed.For each layer in the network,such as convolutional layer In-depth study of the composition and calculation methods of the pooling layer,the activation function layer,and the fully connected layer.According to the forward propagation process of convolutional neural networks and the specific calculation method of each network layer,four parallel computing methods suitable for forward propagation of convolutional neural networks are proposed.After the theoretical analysis is completed,the logic design and hardware implementation of the spectrum monitoring system are completed based on the hardware development platform integrating the AD9361 chip and the Zynq-7035 chip.The input end of the AD9361 chip is connected to the receiving antenna,and is responsible for receiving the external wireless electromagnetic signal,and performing analog-to-digital conversion and high-speed sampling on the received signal inside the AD9361 chip.The PL part of the Zynq-7035 chip completes the logic design of the spectrum monitoring system,and performs a series of digital signal processing operations such as windowing,fast Fourier transform,spectrum estimation,and logarithmic operation on the time domain sampling signal output by the AD9361 chip.A spectral signal corresponding to the time domain signal is generated.According to a series of digital signal processing algorithms used in the spectrum monitoring system,the system simulation is completed in the MATLAB simulation platform,and the simulation results are compared with the hardware implementation results in Zynq to verify the correctness of the spectrum monitoring system.According to the four parallel computing methods applicable to convolutional neural networks,the logic design and hardware implementation of the convolutional neural network are completed in the PL part of the Zynq-7035 chip.Firstly,some common computing units in convolutional neural networks,such as convolution calculation,pooling calculation,and full connection calculation,are logically designed to generate a fixed IP core to facilitate subsequent development.Then the four parallel computing methods are reasonably matched to form a combined parallel computing method,and the system framework of the forward propagation of the whole convolutional neural network is designed.In the logic design,the FPGA pipeline architecture is fully utilized,and combined with the parallel computing method,the designed and implemented convolutional neural network can complete the forward propagation calculation of the convolutional neural network in the shortest clock cycle.The input of the convolutional neural network is the output of the spectrum monitoring system,that is,the spectral data of the received signal.The output of the convolutional network is a vector of length 100,indicating whether there is a spectrum in each sub-band of 300 k Hz in the spectrum of the 30 MHz bandwidth.Identification result.After the completion of the convolutional neural network on the FPGA platform,compared with the convolutional neural network implemented by the CPU platform,it is proved that the FPGA platform can greatly shorten the time of the forward propagation process of the convolutional neural network,and has good real-time performance.