The increasing and flexible use of autonomous systems in many domains -- from intelligent transportation systems, information systems, to business transaction management -- has led to challenges in understanding the "normal" and "abnormal" behaviors of those systems. As the systems may be composed of internal states and relationships among sub-systems, it requires not only warning users to anomalous situations but also provides "transparency" about how the anomalies deviate from normalcy for more appropriate intervention. We propose a unified anomaly discovery framework "DeepSphere" that simultaneously meet the above two requirements -- identifying the anomalous cases and further exploring the cases' anomalous structure localized in spatial and temporal context. DeepSphere leverages deep autoencoders and hypersphere learning methods, having the capability of isolating anomaly pollution and reconstructing normal behaviors. DeepSphere does not rely on human annotated samples and can generalize to unseen data. Extensive experiments on both synthetic and real datasets demonstrate the consistent and robust performance of the proposed method.