The purpose of this study was to investigate how to detect abnormalities in electric submersible pumps (ESPs) in advance and how to classify the faults by monitoring the production data before pumps break down. Additionally, a new method based on the denoising autoencoder (DAE) and support vector machine (SVM) is proposed. Firstly, the ESP production data were processed and fault-related features were screened using the random forest (RF) algorithm. Secondly, input data were randomly damaged by the addition of noise, a DAE network structure was constructed, and the optimal learning rate, noise reduction coefficient, and other parameters were set. Thirdly, the real-time status of the production data of ESP was monitored with reconstruction errors to detect the point when an abnormality occurs signifying a pending fault. Finally, SVM was used to distinguish the type of fault. Compared with existing fault diagnosis methods, our method not only has the advantages of easy extraction of effective data features, higher accuracy, and strong generalization ability but can also detect an abnormal state indicating a coming fault and identify its type, hence enabling the preparation of an appropriate advance solution.