Abstract: In the prediction of the lead-acid battery state of health (SOH), the selection of representative feature set based on small sample plays an important role, considering the various factors resulting in the battery aging and the restriction of the battery aging experiment that the full charge and discharge time and the number of samples are limited. Therefore, based on the analysis of battery characteristics, an SOH feature selection algorithm based on unsupervised ACCA-FCM and supervised SVM-RFE is proposed. The algorithm, first, utilizes the improved ant colony clustering algorithm (ACCA) to select the effective eigenvalue clustering center from the global feature set, which overcame the clustering center sensitivity and local optimal problem of fuzzy C-means clustering algorithm (FCM), and removes the redundant features by the features correlation; second, according to the SVM-RFE feature sorting algorithm, rules out the non-critical interference (Low predictive) features; and finally, obtains the low-dimensional eigenvector with the largest correlation as well as the minimum redundancy of the test result, and avoids the process of complete discharge under the premise of ensuring the accuracy. The SOH model of the battery is verified by the support vector machine (SVM), which has been improved significant and accurate.

Key words: feature selection, ant colony clustering algorithm, fuzzy C-means clustering algorithm, SVM-RFE, state of health