The Research On Data Mining Method Based On Ionospheric Abnormal Phenomenon

It has been proved by a great deal of statistical study and case analysis thationosphere precursor anomaly occurres before strong earthquakes. Observingionospheric precursors has been considered as an effective mean of the short-termand imminent earthquake prediction. In terms of the huge ionospheric data, thetraditional data processing methods are unable to solve this problem. Therefore, it isan inevitable trend to develop new technology which introduces the data mining toanalyze the eletromagnetic anomalies of the ionosphere.In this paper, we apply the French satellite data to study the prediction methodto the ionosphere precursory anomalies of the magnetic storm and earthquake basedon the neural network data mining methods. Besides, the impact on the trainingresults with different parameters is discussed. Overall we mainly stduy the followingtasks:Facing the uneven distribution problems of the satellite data, we apply theconcept of ‘space to time’ to pretreat the raw data of the DEMETER satellite,transforming the one-demensional time series data into two-dementional image atthe cost of sacrificing time resolution. This approach can effectively reduce theinfluence of the asymmetry of the data sample distribution on the neural networktraining results. We can achieve the effect of dinosing, smoothing and filling themissing values by means of averaging and interpolating. What’s more, the abnormalvalue has been highlighted by calculating the standard deveation and using thethreshold segmentaion to transform the original data into0-1map. Then we take thepretreated standard deviation, the sum of the standard deviation and the binaryimage as the training samples of neural network to inspect the impact of the sampleset to the training result.For the magnetic storm which is more common and significant phenomenon ofelectromagnetic radiation of the ionosphere, the sample set, which is obtained underthe condition of different resolutions and different data preprocessing methods, isstudied. And the influence of different network topology and network parameters onthe performance of neural network prediction is discussed. The results show that thegeomagnetic storm prediction accuracy rate can reach80%under the condition of taking the sum of the standard deviation as the input of BP neural network, twohidden layers and variable rate learning algorithm.Due to the complexity of the earthquake ionospheric disturbances, we use thepredictability advantages of neural network for unknown model to study theprediction method of earthquake precursors based on the neural network. By thecorresponding analysis of the data sample and the ionospheric anomalies data in theearthquake preparation area, we study the prediction of the earthquake and furtheranalyze the ‘correlation degree’ of the ionopheric parameters and the earthquakeusing BP neural network and RBF neural network. Finaly these two neural networkhave demonstrated consistent results: the ‘correlation degree’ of the earthquake andthe parameters of eletron density and ion density is higer than other parameters ofelectron temperature and ion temperature, which is consistent with earthquake casestudy in the literature and verifies the validity of data processing method used in thispaper. Moreover, the results also show the RBF neural network is better than BPneural network in terms of the forecast accuracy.