Hardware System Design Of Pulse Feature Extraction Of Calorimeter Based On Convolutional Neural Network

The large-scale physics experiment device contains a variety of detectors.Digitally collecting the output of the detector and analyzing the particle information based on the data is an important part of the physics experiment.The calorimeter is mainly used to measure the energy of particles,and the output is usually in the form of pulse signals.Extracting time and amplitude information from the pulse signals is an important step in particle analysis.There are many ways of impulse feature extraction,such as curve fitting,population technology and so on.With the continuous increase of the amount of detector experiment data,the detector environment often contains random noise,long-term drift,short-term changes and other effects.Traditional digital processing methods have certain limitations in eliminating the effects of noise in the detector environment,and cannot guarantee the best accuracy when extracting pulse features.Due to its advantages of higher accuracy and faster processing speed,deep learning methods are gradually being used in the data analysis of physical experiments.At present,there are many researches on the use of deep learning to process impulse signals,such as the use of deep learning models to complete the recognition of stacked impulses,and the use of long and short-term memory network models for impulse feature extraction.However,most of these studies analyze the data based on the offline state.When the amount of calculation increases,these models cannot be implemented on the hardware platform.In order to reduce the pressure of data transmission on the system in the detector,it is often necessary to process data in real time,and the hardware system for online processing is more in line with the needs of the experimental environment.In order to facilitate the realization of online processing systems,the convolutional neural network in deep learning not only contains the advantages of deep learning to process impulse signals,but also has simple calculations and is easy to implement,making it a new impulse feature extraction method.In this paper,based on the zero-degree angle calorimeter of the low-temperature high-density nuclear matter measurement spectrometer,a hardware system for pulse feature extraction is designed on the FPGA platform based on convolutional neural network,which mainly extracts the time and amplitude information of the calorimeter pulse.The system consists of two parts:neural network calculator and controller.The Neural Network Calculator supports the convolution operation,deconvolution operation and matrix multiplication in the convolutional neural network using the acceleration method of parallel operation and data multiplexing,and undertakes the core calculation part of the system.The controller completes the configuration of the neural network calculator in the form of instructions,and undertakes the supervision and scheduling part of the system.When each layer is running,the controller sends configuration information and calculation data to the neural network calculator,and the neural network calculator transmits the results back to the controller after completing the calculation of one layer.The controller quantifies the data as the input of the next layer,and completes the operation of all layers in the convolutional neural network according to the cyclic process.After testing,the simulation is performed when the working frequency of the controller is 100MHz and the working frequency of the neural network calculator is 25MHz.The second layer of the fully connected layer in the system takes the longest time,which is 966.8 ?s.On the real experimental platform,the software processing in the Tensorflow environment based on the Python platform uses the floating-point number calculation method,and the system's hardware processing uses the 8bit fixed-point number calculation method.Under the same convolutional neural network architecture,the pulse width is 1?s,and the ADC chip sampling voltage is±5V,the time resolution of the hardware processing is 0.01500 ?s,the amplitude resolution is 0.02747v,and the time resolution of the software processing is 0.01329 ?s,The amplitude resolution is 0.02499v.There is little difference between the processing performance of the hardware system and the software,which can replace the software used in the detector device to process the pulse signal,and extract the time and amplitude information of the pulse in real time.