Research On Applying Deep Learning Method On Image And Time Series Analysis

Artificial intelligence(AI)has become the hottest research topicin now,and various related works have been carried out in various fields,including medical treatment,biology and Internet.Obviously,today,artificial intelligence is no longer a mystery to most people,and people's requirements for it have become higher and higher.It has changed from"can do it" to "can do it well".In order to better explore the effectiveness of artificial intelligence in various fields,my paper will carry out my work from two directions,including brain tumor image recognition and anomaly detection of multi-feature time series.The application of artificial intelligence to medical image analysis has become a hot research direction in the past few years.As the medical image problem is related to the treatment progress of patients,it directly affects the health of patients.This leads to medical image problems that require high speed and accuracy.Research on medical image analysis is already underway in the industry,and the proliferation of articles indicates that the field is getting more and more attention.The image segmentation direction of brain tumor I studied has been widely concerned at home and abroad.2.Since different types of tumors need to correspond to different treatment regimens,it is necessary to distinguish different types of tumor regions and define clear boundaries for them to facilitate treatment.In order to solve these two problems,I built an end-to-end semantic segmentation model m-unet based on neural network,which can quickly process the mri images of brain tumors,obtain the boundary division of the brain tumor region,and accurately identify multiple types of tumors and divide them according to the type.On the basis of the previous work,I designed two new network modules for my network structure-the double layer convolutional network module and the anti-pooling convolution module.The purpose of designing the two-layer convolution module is that because the proportion of non?tumor areas in the brain image is often very large,the network needs to quickly filter out the non-tumor areas,while the tumor areas also need to distinguish the tumor categories,which requires the network to elaborate the analysis of the image,so there is a contradiction between these two types of problems.Therefore,the two-layer convolutional network is used to solve this problem.By combining the convolutional layer and the empty convolutional layer,the focus area can be analyzed in detail while simultaneously filtering the non-focus area rapidly.In addition,because traditional neural network methods often use the method of transposed convolution or deconvolution in the upper method,and these methods have the problem of dilution parameters,so in order to solve this problem and improve the network effect,I designed a new up-sampling module to deconvolution convolution module.This method makes use of the idea of regional concern,and gives the real physical meaning to the convolution layer,which can improve the feature extraction ability of the network while restoring the feature layer to its original size.In addition,in order to restrain the influence of data imbalance on training,I designed a weighted Dice loss function to adjust the training proportion of various samples by the number of samples.Similarly,time series anomaly detection has gradually become a research direction with great potential in the field of data mining.There are many examples of successful migration of methods from other fields to the field of exception detection.One of the main problems I found in sorting out the previous methods was that they didn't make effective use of the information about the connections between multiple features.For the problem of anomaly detection,how to accurately locate the abnormal features depends on many factors,and the most important one is whether the features maintain a normal relationship.Many methods fail to take this into account and misjudge.In this paper,I propose a new unsupervised multivariate time series anomaly detection framework GARN to solve this problem.My model USES the attention map neural network,which is widely concerned,to model the correlation between the features and the time points in the time series.This method does not need any prior information about the correlation between features,so it can be used in more scenarios.The model treats each univariate time series as a separate feature and includes two graphical attention layers to learn the correlation of multivariate time series in time and feature dimensions.In addition,GARN jointly optimizes the prediction-based model and the refactor-based model,and obtains a better time series representation through the combination of single timestamp prediction and time series reconstruction.In this paper,a large number of experiments have been conducted to prove the effectiveness of GARN,which performs better than other existing models on three data sets.Further analysis shows that GARN has good interpretability and can be used for anomaly diagnosis.