Research On Time Series Similarity Measure And Fault Detection Based On Siamese Neural Network

A set of sequence data acquired by a dynamic event in timeline order is called a Time Series.Time series data mining is a challenging area of research.There are many applications in the fields of speech recognition,power load forecasting,fault diagnosis,bioinformatics,patient care and the like.However,due to the continuous high-dimensional,complex shape and dynamic amplitude changes of time series data,it is difficult to analyze its hidden information and laws effectively.In data mining tasks such as time series classification clustering,it is one of the cores of time series data mining to analyze the relationship or similarity between sequence data samples.The traditional similarity measures such as euclidean distance,cosine distance and dynamic time warping are only for the data itself to calculate the difference,ignoring the effect of the knowledge annotations contained in different data sets on the similarity measure,a suitable and effective "Dynamic "similarity metrics are critical.A time series similarity measure based on twin neural network(SNN)is proposed for this purpose,and the research results are applied to time series classification and fault diagnosis of rolling bearings.Firstly,this paper proposes a metric learning method that considers the similarity of time series.This method is based on twin neural network,Ieans the neighborhood relationship between data from the supervised information of sample tags,and mines the implicit feature representation of time series.The paper analyzes the variation of sample hypothesis interval under different similarity metrics,and uses t-SNE dimensionality reduction visualization to show the eigenvector spatial distribution of SNN output.Secondly,it combines the nearest neighbor algorithm(INN)to 22 sets of UCR time series data sets of different fields.Perform a classification experiment.The experimental results show that the proposed method can effectively measure the similarity of time series.Compared with the classification method based on Euclidean distance and dynamic time bending metric,the classification accuracy is improved significantly,and the classification of high-dimensional and complex time series data is good.which performed.Finally,the SNN-1NN algorithm based on feedforward neural network and the SCNN-1NN algorithm improved by one-dimensional convolution network are used to judge the running state of the bearing.Using the one-dimensional bearing vibration signal data collected by the bearing fault simulation experimental platform,the correct rate of bearing fault diagnosis under various different methods was tested.The results show that the SCNN-1NN method has higher accuracy and efficiency in bearing fault diagnosis experiments.Experiments are carried out to analyze the performance of neighbor classification algorithm under different similarity measure methods.The metric learning method based on twin neural network is verified by experiments for time series similarity calculation and rolling bearing fault diagnosis,which improves the robustness of the neighborhood classification algorithm.It is of great significance to find bearing faults to reduce production accidents and reduce economic losses.The paper method has a good application prospect.