Research On Time Series Prediction Applied To Information Equipment Fault Early Warning

The operation status perception and fault alarm of traditional information equipment mainly rely on manual experience,which have disadvantages such as high cost,low efficiency,and high false alarm rate.The rapid development of artificial intelligence is driving the intelligent transformation of the operation and maintenance field.Traditional artificial experience and static rules are gradually being replaced by algorithm models.This study fully excavated the pain points in the early warning of faults,and provided targeted solutions in several aspects such as difficulty in timing feature extraction,insufficient prediction accuracy,and difficulty in determining the alarm threshold.First of all,this thesis improves the firefly algorithm,which greatly improves the optimization accuracy and convergence speed of the algorithm.Secondly,this thesis proposes a dynamic threshold setting mechanism,which can calculate the dynamic threshold interval under the specified confidence based on the prediction result,so as to realize more intelligent fault early warning and greatly reduce the human and material cost of enterprises.Finally,this thesis proposes two time series forecasting models,and conducts comparative experiments of single-dimensional time series forecasting and multidimensional time series forecasting respectively to verify the effectiveness of the proposed models.For single-dimensional time series forecasting,this thesis proposes a DWAFE(discrete wavelet transform-ARIMA-EWFA-ELM)composite model.The model splits the original time series into multiple subsequences through discrete wavelet transform,and uses the ARIMA model and the Firefly Algorithm(FA)optimized extreme learning machine model(ELM)for processing according to the difference in stationarity.Finally,the prediction results of each subsequence are integrated by inverse wavelet transform.Experiments on the received traffic data of the core routers of State Grid Ningxia Electric Power Co.,Ltd.show that the method achieves better performance than benchmark models such as BiLSTM and GRU.For multi-dimensional time series prediction,this thesis applies the attention mechanism to the GRU model and proposes the TCAG(TDA+CNN+Attention GRU)model.The model uses TDA to extract the most compact and effective topological features in the data,uses CNN to extract the spatiotemporal features from the data,and creatively combines spatiotemporal features with topological features for time series prediction.The comparative experiments on the monitoring data of the core routers of State Grid Ningxia Electric Power Co.,Ltd.show that the accuracy and robustness of the TCAG model are better than the baseline model.34 Figures,9 Tables,109 References;...