| Artificial intelligence is developing rapidly,and there is a tendency to migrate to embedded systems.At present,the artificial neural network mainly depends on the operating system to realize,and no training and learning of the artificial neural network has been found on the microcontroller chip.The microcontroller chip has the characteristics of low power consumption,real time,stability,reliability,and low cost.Building a neural network on the microcontroller chip can move the data processing forward to achieve edge computing.This system builds an LSTM recurrent neural network on the microcontroller chip,and uses this to predict atmospheric haze concentration.Use the floating-point operation and DSP instructions of the microcontroller chip to write the complex data operation functions required by the neural network.Use these data operation functions to build the neural network and implement machine learning and prediction directly on the microcontroller.The matrix is updated.The Mist scattering particle concentration sensor measures the haze value as the supervision of the neural network,and the weight matrix is continuously updated by the gradient descent method to gradually reduce the haze prediction error.The main research work is as follows:(1)It is proposed and realized that the LSTM recurrent neural network is built directly on the microcontroller chip,which has the functions of machine learning and prediction.The forward propagation and back propagation of the recurrent neural network are built on the microcontroller chip.The sensor measurement value is used as the supervision of the neural network.The network can complete training and learning in use.(2)Write and encapsulate the interface required for the neural network built by the microcontroller chip.Since Cortex-M series chips have added digital signal processing unit and FPU floating point arithmetic unit after M4,it is possible to complete machine learning and prediction directly on the microcontroller chip.Through in-depth research and analysis of the microcontroller chip and neural network,the interface required for the microcontroller chip to build the neural network is written.(3)Design and implement machine learning on the microcontroller chip to predict the haze concentration.Recurrent neural networks are good at processing time-series signals and implemented on a general-purpose microcontroller chip to predict smog,use limited computing power to achieve intelligent prediction,and truly achieve instrument intelligence.(4)Encapsulate the neural network.One of the main uses of the microcontroller chip is signal processing.The long-term and short-term recurrent neural network is mainly used to process time-series signals,so that the long-and short-term recurrent neural network can be widely promoted on the microcontroller.Encapsulation of LSTMs gate,unit state,error term calculation,etc.,only need to call functions and change parameters when using,you can build a neural network for processing different timing signals.Implement a long-term and short-term recurrent neural network on a general-purpose microcontroller chip,first perform pre-learning,then place the device at a monitoring point,and supervise learning from the particle concentration sensor data to predict the haze concentration and predict the haze concentration after one day.Experimental data shows that it is feasible to build a neural network on a general-purpose microcontroller chip to predict the haze concentration. |
