LI Lei, MA Yulin, HU Gang, KONG Xuefeng, YANG Jun, XU Yanwei
To accurately identify the head defects of the GH159 bolt after hot upsetting, this paper proposes a defect recognition method based on transfer learning, where datasets under scenes with different brightness are set as the source domain and target domain in transfer learning, respectively. First, considering the multi clusters of the conditional distribution in the domain, this paper adopts the K-means algorithm to cluster samples with the same defect and determine the cluster centers in this defect, then a novel measurement of the distribution discrepancy can be constructed on the cluster centers. Second, based on the distances between cluster centers and the distances between each cluster center and the samples belonging to the cluster, a new intra-class discrepancy can be established for improving the computational efficiency of transfer learning. Finally, the optimization objective of the proposed method is built on minimizing the weighted sum of the constructed distribution discrepancy and intra-class discrepancy to effectively identify defects under scenes with different brightness. According to the requirement on partial parameters setting of the proposed method, the pseudo-accuracy is designed using the reverse verification strategy, then the parameters are set as the parameters’ combination with the highest pseudo-accuracy. Using the collected dataset on head defects of the GH159 bolt after hot upsetting, the analysis and application of the defect recognition are carried out to verify the effectiveness of the proposed method.