Thunder Signal Pattern Recognition Based On HHT Feature Extraction

With the fast development of society and economy, the danger generated by the thunderstorm is more and more serious; the protection for thunderstorm has long been closely concerned. Lightning not only may destroy the important equipment and facilities, but also may hurt people or livestock and cause the accidents such as fire, explosion, blackouts, bringing the serious economic loss to the society. The thunder often causes frequent false alarm of some monitoring equipment. Therefore, in order to prevent wrong operation caused by thunder interference, the pattern recognition research of thunder signal is very important.This thesis mainly studies the pattern recognition of thunder signal. Because the thunder signals collected in reality belong to non-stationary signals, even it is complicated signal with noise, and the resource of the noise differs with changed environment, season, weather and time, it is not effective by usual noise reduction. Firstly, this article carries on the singular value decomposition to get a set singular value in order. Then, we judge the mount of singular value which represents for thunderstorm signals according to the MDL criterion. Next, we can realize noise reduction of the thunder signal combining with SVD. The simulation experiment results show that using the method of SVD to reduce noise can achieve the desired effect.Then, in the process of thunder signal feature extraction based on the HHT transform, the article uses EMD decomposition for those thunder signals after pretreatment, extracts 6 dimensional feature vectors from IMF and extracts 25 dimensional feature vectors based on the marginal spectrum. So we get 31 dimensional feature vectors which are the input feature parameter of the neural network pattern recognition. This method extracts the intrinsic characteristics and more comprehensive energy characteristics of the thunder signal. It is more simple and quick and it can extract more characteristic parameters than ever before.Lastly, this thesis builds up the BP neural network to recognize the thunder signal, uses the improved Levenberg-Marquardt learning algorithm and selects 30 training samples and 15 samples of interference (acoustic signals) to train the network. After the network training success, we input the 10 groups of thunder test samples and 5 groups of disturbed samples to network, and the recognition rate reaches 100%. The simulation experimental results show that the recognition system has ideal recognition accuracy and stability in a certain range, and has certain auxiliary significance to improve the monitoring equipment wrong operation.