22 February 2022 in High Voltage
Dissolved Gas Analysis (DGA) is an important method for oil‐immersed transformer fault diagnosis. However, collecting labelled DGA data is difficult because the determination of the transformer fault is time‐consuming and expensive in the transformer substation, but DGA data without labels is easier to obtain. Therefore, the paper proposed a semi‐supervised two‐stage diagnostic system based DGA by using less labelled samples. The two‐stage system includes a novel semi‐supervised feature selection based Genetic Algorithm (GA) and Support Vector Machine (SVM) model (SSL‐FS‐GASVM) for selecting optimal features and a novel semi‐supervised transformer fault diagnosis model based improved Artificial Fish Swarm Algorithm (AFSA) and SVM (SSL‐IAFSA‐SVM) for optimising the SVM parameter. Finally, the performances of SSL‐FS‐GASVM and SSL‐IAFSA‐SVM models are tested and compared with traditional supervised diagnostic models combined with other optimisation methods, respectively. The results show that the proposed two‐stage system works in optimising features and parameters and has strong robustness in solving small sample classification problems.
1 April 2017 in Journal of Medical Imaging and Health Informatics
Study on retaining good classification performance when reducing training effort is valuable for the development of brain-computer interface (BCI). In this paper, an improved self-training algorithm using only a small number of labeled samples is adapted for motor imagery EEG classification. Unlike the traditional self-training algorithm that uses only one classifier, an additional classifier is introduced in the improved version for sample selection, which helps to reduce the error reinforcement, increases the classification accuracy, and improves the convergence in the iteration process. Experiment results show that the improved self-training algorithm outperforms its prototype in the two-class classifications. Even for the subjects who have performance degradation when using the traditional self-training algorithm, the improved algorithm still increases the classification performance iteration by iteration. We also compare the improved method with a supervised method which uses all the training data, and there are no significant differences in classification accuracy. It can achieve a little higher performance comparing with an existing semi-supervised method on the same data.