We propose an adversarial automatic data augmentation with hard sample suppression by incorrect class likelihood for time series data. Automatic data augmentation (ADA) is a practical framework when the number of training data is small. In particular, an ADA based on adversarial loss, such as TeachAugment, can improve classification accuracy in optimizing complex transformation parameters with class and instance dependencies. In this paper, we aim to extend adversarial ADA for image recognition to time series data. The critical challenge is that data augmentation for time series data is more difficult than that for images in designing a transformation function that does not cross the discriminant class boundary. When existing ADA frameworks are naively applied to time series data, the transformation function often generates augmented data similar to the incorrect class. As a result, this overly hard-to-identify data degrades the accuracy of a target classifier. In general, this hard-to-identify data near the class boundary has a high likelihood both in the correct and incorrect classes. The proposed method can reduce this hard-to-identify data by introducing a novel fidelity loss that suppresses only the likelihood of the incorrect class. Comprehensive experiments on two datasets demonstrate the effectiveness of the proposed method. The project code is included in the supplemental material.