Application Of Explosion State Monitoring Based On Long And Short Term Memory Neural Network Hardware Acceleration

Nowadays,air drilling technology has become one of the important research directions in the field of oil and natural gas drilling.However,when an oil and gas layer is encountered during the drilling of air drilling,downhole explosions may occur and cause huge economic losses.In view of this,this paper designs a combustible gas explosion state monitoring system based on long and short-term memory neural network(LSTM)hardware acceleration.The system is divided into slave data acquisition terminal and host data processing terminal.The slave machine collects the concentration of methane,oxygen,carbon dioxide,temperature and humidity in the air through a variety of electrochemical gas sensors,and preprocesses the above environmental parameters and sends them to the host.The host communicates with multiple slaves,and uses software and hardware to cooperate with the LSTM neural network algorithm to realize the monitoring function of the explosion state of multiple nodes.The slave is equipped with a catalytic combustion gas sensor to detect the methane concentration,and combined with the pulse power supply bridge method to expand the methane concentration detection range of the catalytic combustion gas sensor.The test results from the machine show that the pulse power supply bridge method used in this design enables the catalytic combustion gas sensor to detect methane gas with a concentration of up to 80,000 ppm.At the same time,the pulse bridge uses short-period pulse power supply to increase the measurement speed of methane concentration to 2 seconds.The host uses the Cortex-M3 series 32-bit industrial control processor IP to form a system-on-chip(SOC)on a field programmable gate array(FPGA)with pure logic resources,and builds a long-and short-term memory neural network hardware accelerator on the systemon-chip.Multi-state classification model of gas explosion.The host test results show that the LSTM accelerator uses a parallel optimized deep pipeline to perform calculations,which improves the use efficiency and calculation speed of each module,so that the dense LSTM neural network calculated on Artix-xc7a100 t in this design can reach a throughput of 13.7GOP/S.7.41 GOP/S/W energy efficiency ratio.An experimental environment for combustible gas combustion and leakage was set up,and corresponding real machine tests were carried out on the slave data acquisition terminal and the host data processing terminal.Experimental results show that the system can normally detect the concentration of oxygen,carbon dioxide,and methane,quickly classify and judge the explosion state of combustible gas,and has a high accuracy rate.