In the topic of damage detection of composite structures based on lamb wave technology, damage index is commonly used for damage identification. However, its threshold is largely of expertise-dependence and poor performance at knowledge generalization. Therefore, a method based on the concept of least margin is proposed, which integrates even machine learning models and outputs the identification result by polling all models’ decision. The proposed method avoids the shortage that damage recognition relies on a single but incomprehensive model, and puts the confidence on a number of most qualified models instead. Significantly higher accuracy of damage identification for composite structures is manifested through test verification.