In the construction industry, continuous monitoring of actions performed by construction machinery is a critical task in order to achieve improved productivity and efficiency. However, measuring and recording each individual construction machinery's actions is both time consuming and expensive if conducted manually by humans. Therefore, automatic action recognition of construction machinery is highly desirable. Inspired by the success of Deep Learning approaches for human action recognition, there has been an increased number of studies dealing with action recognition of construction machinery using Deep Learning. However, those approaches require large amounts of training data, which is difficult to obtain since construction machinery are usually located in the field. Therefore, this paper proposes a method for action recognition of construction machinery using only training data generated from a simulator, which is much easier to obtain than actual training data. In order to bridge the feature domain gap between simulator-generated data and actual field data, a video filter was used. Experiments using a model of an excavator, one of the most commonly used construction machinery, showed the potential of our proposed method.