| The tanks are important storage facilities of the petrochemical enterprises. Due to theimpact of the external environment and storage material, the tanks are prone to suffer fromcorrosion and leaks, which are the major threat to the work safety. It is the hardest to detectthe tank bottom in the tank corrosion detection. When using the conventional testingmethods, it is required that the tank is not on work and the tank is empty. During theregular examination process, generally only few tanks do suffer from severe corrosion andleaks and need for maintenance. Many tanks in good condition are also emptied anddetected, which resulted in a waste of resources. The acoustic emission testing technologycan achieve online and real-time monitoring without the need for clearing the tank. It hasadvantages over other conventional methods on the testing for the large-scale normalatmospheric metal tank.Through consulting a large number of domestic and foreign literatures, the basicprinciples of acoustic emission detection technology and the Neural Networks, SVMmethods of acoustic emission signal recognition are deeply studied. Relevance vectormachine is a supervised learning method based on the Bayesian framework, which requiresa small number of training samples and tests quickly. Through learning the characteristicsparameters and the frequency-domain parameter of the acoustic emission signal, it canrealize the classification of acoustic emission signal. Through learning and researching thebasic principles of relevance vector machine (RVM) model to understand the performanceof the classification model is closely related to its kernel function parameters. Respectively,artificial bee colony algorithm (ABC), particle swarm optimization (PSO) and geneticalgorithm (GA) are applied to optimize the kernel function parameter of the relevancevector machine. The basic principle of artificial bee colony algorithm is deeply studied andthe performances of the three optimization algorithms are compared. Based on the binarytree structure and the one-against-all method, the binary-classification RVM model isextended to establish a four-classification model. The impact of using the different acousticemission parameters as the input feature vector on the classification performance is analyzed. The specimen of the Q235steel, which is the material of the tanks, is corrodedwith the electrochemical methods. The corrosion acoustic emission signal of differentphases is gathered by the acoustic emission detection system. After the acoustic emissionsignal is de-noised, the extracted frequency domain information and the acoustic emissionsignal parameters of the time domain are used as the input data of the classification modelfor training. The tank bottom corrosion acoustic emission evaluation model is established.Based on the evaluation model, two sets of acoustic emission signal of the tank bottomcorrosion are identified. The practicability of theABC-RVM model is proved. |
