Pumps are one of the most critical machines in the petrochemical process. Condition monitoring of such parts and detecting faults at an early stage are crucial for reducing downtime in the production line and improving plant safety, efficiency and reliability. This paper develops a fault detection and isolation scheme based on an unsupervised machine learning method, sparse autoencoder (SAE), and evaluates the model on industrial multivariate data. The Mahalanobis distance (MD) is employed to calculate the statistical difference of the residual outputs between monitoring and normal states and is used as a system-wide health indicator. Furthermore, fault isolation is achieved by a reconstruction-based two-dimensional contribution map, in which the variables with larger contributions are responsible for the detected fault. To demonstrate the effectiveness of the proposed scheme, two case studies are carried out based on a multivariate data set from a pump system in an oil and petrochemical factory. The classical principal component analysis (PCA) method is compared with the proposed method and results show that SAE performs better in terms of fault detection than PCA, and can effectively isolate the abnormal variables, which can hence help effectively trace the root cause of the detected fault.