Anomaly detection is a significant task in various fields, such as industry and daily life. One anomaly detection method for time-series data is reconstructive reservoir computing (RRC), which uses an echo state network (ESN). There, a single RRC module trained with one piece of normal data reconstructs similar normal data successfully. Reconstruction failure indicates anomalies in the fed data. However, we have to deal with diverse normal data in reality. Therefore, an ensemble system using multiple modules and various normal data for training will improve anomaly detection ability. The system utilizes multiple neural networks and integrates outputs from the networks. Researchers traditionally use average pooling for integration. In this paper, we propose an RRC ensemble with minimum pooling. Our proposed ensemble employs multiple RRC modules, each trained with different normal data. As the final reconstruction error of the proposed system, we use the minimum value among all the reconstruction errors obtained from the RRC modules. The minimum reconstruction error becomes small when one of the RRC modules in the system successfully reconstructs the fed data. Consequently, if the fed data is truly normal, the decline of the final reconstruction error is more certain than when using a single RRC module. We conduct experiments using a large amount of real machine sound data. The experiments show that the RRC ensemble with minimum pooling enhances robust reconstruction against a variety of normal data, resulting in better performance than averaging pooling.