Abstract: Traditional information equipment's operating status perception and fault alarms mainly rely on manual and traditional automated operation and maintenance, which have disadvantages such as high overhead, low efficiency, and high rate of false alarms. This paper proposed a dynamic threshold setting mechanism, which can calculate the dynamic threshold interval under the given confidence level based on the prediction results. In order to get the accurate prediction results, the Discrete wavelet transform-ARIMA(auto-regressive moving average model)- EWFA(exponential weighted firefly algorithm)-ELM(extreme learning machine) composite model was proposed. In this model, the original time series can be divided into several subsequences by discrete wavelet transform, and the ARIMA model and the extreme learning machine (ELM) optimized by firefly algorithm (FA) were used for processing according to different stationarity. Finally, the prediction results of each subsequence were integrated by inverse wavelet transform. In addition, the exponential weighted firefly algorithm (EWFA) was proposed to improve the optimization performance and convergence speed of the firefly algorithm. Experiments on the core router data of Ningxia electric power company show that the method achieves better performance than Bi-LSTM, GRU and other benchmark models, and can achieve accurate and efficient information for equipment fault early warning, thus greatly reducing the human and material costs of enterprises.