Fault Prediction Of Carbide Anvil Based On Support Vector Data Description

Due to the endurance of alternating high temperature and pressure, the carbide anvil which is used in diamond manufacturing process is very prone to break. If the fractured carbide anvil continues to be used, an accident will possibly be induced. Furthermore, the accident will also lead to the failures of other five carbide anvils. These failures will jeopardize the safety operations when producing diamonds. Therefore, it is of highly importance to guarantee the safety production of diamonds when the possible fractures of these carbide anvils can be effectively recognized and diagnosed in time.This thesis proposes a fault detection method of carbide anvil based on the supporting vector data descriptors (SVDD), the specific researches contents are as following:1)Due to the problem of heavily contaminated background noise of the crack signals, after analyzing the characteristics of the collected signals, this thesis designs a high pass filter (HPF) to filter the background noises on the basis of the difference between crack signals and background noise and hence extracts the characteristic signals according to the energy threshold. The effectiveness of this method is validated by experiments.2)Extensive empirical research results revealed that the malfunctioned and normal signals of carbide anvils con not be precisely discriminated by methods of classical parametric analysis. Therefore, in order to extract the characteristic parameters, this thesis proposes three feature parameters, that is, the zero passing rate, the linear cepstrum and the power spectral density. It is experimentally validated that the discrimination of malfunctioned and normal signals can be successfully achieved with these three parameters. 3)By analyzing constraints of the recognition of malfunctions, a method of optimizing feature parameters is proposed. The method selects a feature via incremental gain after quantizing the continuous information. The experiment results show that the selected feature parameters with this method before constructing a classifier will improve the effectiveness.4)With the continuous data sampling process, the increased data set will enrich the training set. And the information of the increased data has potential values. Therefore, an incremental learning is introduced. Experimental results verified that the time for classification can be reduced and the correct rate can also be improved.