Research On Key Techniques Of Neural Network Architecture Search And Optimization For Time Series Classification

With the progress of science and technology,the operation mode of industry,agriculture,military and other fields is gradually becoming automated.When the machine in each scene works,it will produce massive data,some of which are called time series data because of their temporal correlation.There is abundant information in time series.If we can use this information,we can gain great value.In the industrial scene,the sensor always collects the status data of the device,so these data can be considered as time series data.By classifying these time series data with high dimension and large amount,we can judge the operation status of the equipment,maintain or repair the equipment on times,so as to avoid the equipment failure caused by continuous work.Therefore,it can be considered as a time series classification problem to judge the sensor state data in the industrial scene.In the problem of time series classification,deep learning can save the complex data preprocessing and manual feature extraction process by modeling the time series end-to-end.Therefore,this paper uses deep learning neural network model to study the classification of time series(1)Based on previous engineering experience,a model composed of convolutional neural network(CNN)and bi-directional long and short term memory neural network(Bi LSTM)is designed.In view of the high dimension of industrial time series data,using CNN can make the data dimension rise first and then decline,enrich the data feature space,compress the series and extract the key information at the same time.Considering the temporal correlation of time series data,the method of sending back to Bi LSTM is adopted.Bi LSTM can not only save long-term memory,but also analyze the data in two directions to further improve the classification accuracy of the model.(2)Neural network search technology is used to enrich the structure of the neural network designed by CNN and Bi LSTM and select the internal parameters of the network.Firstly,this paper defines the search space of the model,uses the search strategy based on reinforcement learning,selects the sub network with good performance by using RNN as the controller,and feeds back the accuracy of the sub network on the test set to RNN as a reward function,updates its internal parameters,and then carries on the next search.Compared with the neural network designed before,a better network model is obtained.(3)Considering that the devices in the actual industrial scene have limited computing power,small memory capacity and belong to edge devices,those models with large scale and long reasoning time are not suitable for deployment in such a scene.This paper distills and quantifies the knowledge of the neural network search model in order to reduce the storage space and the running time of the model,and achieves good results.