In the Internet of Things era, thousands of sensors are used to monitor machine health conditions in modern factories. Such sensing data are often analyzed to detect machinery failures and diagnose the suspicious causes to prevent loss. However, failures rarely happen and are highly diverse in the real world, which brings difficulties in collecting abnormal data for supervised models. Furthermore, large numbers of sensors would lead to the problem of high dimensionality that would decrease the accuracy of anomaly detection. To address these issues, this paper proposes the COrrelation-based SpatioTemporal Anomaly Detection (COSTAD) framework to analyze large-scale sensing data. Temporal anomalies are detected in an unsupervised manner using only normal data. Spatial anomalies are analyzed while reducing the dimensions of input features for machine learning models. Our experimental results indicate that the temporal detection approach outperforms some existing methods, and spatial detection can help analyze the suspicious causes of machinery failures.