The petroleum industry is crucial for modern society, but the production process is complex and risky. During the production, accidents or failures, resulting from undesired production events, can cause severe environmental and economic damage. Previous studies have investigated machine learning (ML) methods for undesired event detection. However, the prediction of event probability in real-time was insufficiently addressed, which is essential since it is important to undertake early intervention when an event is expected to happen. This paper proposes two ML approaches, random forests and temporal convolutional networks, to detect undesired events in real-time. Results show that our approaches can effectively classify event types and predict the probability of their appearance, addressing the challenges uncovered in previous studies and providing a more effective solution for failure event management during the production.