Research On Time Series Classification Technology For Industrial Big Data Based On Deep Learning

In the era of digital industry,due to the intensive and close connection of production process,the occurrence of faults has caused considerable losses to industrial production.It is very important to judge the state of industrial equipment by classifying the time series data collected by industrial sensors,and to adjust and eliminate the faults in time.The high-frequency sampling of large number and variety of sensors cause the complexity of industrial big data,which has the characteristics of high spatial dimension,complex dependence relationship,changeable rules,massive volume and so on.These characteristics bring challenges to the accuracy,efficiency and adaptability of industrial time series classification.The current algorithms applied in industry and excellent algorithms in the field of time series classification are difficult to conquer these challenges.Meanwhile,the end-to-end deep learning algorithm has a good performance in many fields nowadays.This paper studies the application of deep learning in industrial big data time series classification.In view of the characteristics of industrial data and the challenges of industrial time series classification,it starts from three aspects: precise classification,efficient classification and incremental learning.In terms of precise classification,this paper designs an end-to-end deep learning neural network using LSTM(Long Short-Term Memory)and GRU(Gated Recurrent Unit),and proposes a network structure design method to select the layer number of neural network according to the integration order of series in the training set,which solves the adaption problem of the neural network model structure.In the experiment,the proposed algorithm achieves 97%-98% classification accuracy,which is 3.91% higher than other non-deep-learning algorithms.In terms of efficient classification,SRU(Simple Recurrent Unit)and CNN(Convolutional Neural Network)are used to design a highly parallel end-to-end deep learning neural network.At the same time,Summarizer-Attention mechanism is proposed to further improve the classification effect of CNN.Experiments show that,compared with the precision classification algorithm proposed in this paper,the classification efficiency of the proposed efficient classification algorithm model is increased by 1.39-5.27 times when the accuracy decreases by 0.5%-0.8%.In the aspect of incremental learning,this paper designs a model updating strategy based on measuring the difference of data distribution by KL divergence,and designs an incremental learning model of neural network based on data distribution mapping by referring to transfer learning,which solves the problem of adapting the model to the change of data distribution.Experiments show that the incremental learning strategy and model can update the model as low-frequent as possible,while maintaining the adaptability of the model to new data distribution and high classification accuracy of 91.84%-96.92%.The above three aspects of research together constitute an algorithm system for industrial big data time series classification,which meets the accuracy,efficiency and adaptability requirements of industrial time series classification.The innovations include: solving the structure adaptability of the neural network model by analyzing the integration order of series in the training set;Summarizer-Attention mechanism which summarizes the time-related features into time-independent features;designing incremental learning strategies and models based on the characteristics of industrial data,using the idea of data distribution differences and migration learning.