A Semi-supervised Knowledge Assessment Paradigm Based on T-CNN Algorithm for the Industrial Massage System

Abstract

with the popularity of the Industrial Internet and Information Technology, most data resources in the industrial production process could be exploited and utilized completely. But there are some fragmented knowledges, like product data, project documentation and multimedia materials, that need to be collected and organized. For instance, Industrial Massage System (IMS) is a platform for operating field professionals to exchange and share their experiences and knowledge containing significant potential value. In order to effectively utilize the fragmented information, a semi-supervised knowledge assessment paradigm is proposed for this system using Tri-training method based on Convolutional Neural Network (T-CNN). The method initially trains three CNN classifiers employing a small number of labeled examples, and a large pool of unlabeled examples are then assigned pseudo-labels under certain conditions. In further training, these classifiers are fine-tuned through original labeled examples and pseudo-examples. Specifically, in the iteration of the tri-training, a pseudo-example for a classifier is obtained from the reliable hypothesis if the other two classifiers have the same predictions on the labeling. Compared to previous training method, it does not require enormous original labeled data to initialize the model, nor does it need to depend immoderately on expert's domain to label examples. And there are also no demands of the instance space to be described with sufficient and redundant views, like previous co-training style algorithms. Experiments tested on two benchmarks indicate that the algorithm can effectively exploit unlabeled data to enhance the learning performance and achieves better classification capability.