In the era of the Internet of Things (IoT), everyday objects are equipped with sensors and actuators. They are also referred to as IoT devices that collect data and perform actions by communicating with each other. The resource-constrained IoT devices may be put under significant stress or external influences that may lead to their failures. In IoT environments where safety is a critical requirement, the failures need to be prevented before they can cause any harm. By implementing Predictive Maintenance (PdM) with Machine Learning (ML), different methods can be applied to predict the failures and provide enough time for maintenance. IoT devices collect data that can be used to train the ML algorithms to recognize the failure patterns of individual devices. This thesis presents the Failure Prediction Platform (FPP), a platform that enables simple integration and management of ML models trained for failure prediction of IoT devices. Furthermore, different ML algorithms are compared regarding their failure prediction performance and their training and prediction speed. Based on the evaluation results, the tree-based algorithms showed the best performance in predicting failures, while the linear classifiers had the worst results. Finally, a prototype of the FPP is implemented and presented by using Long Short-Term Memory (LSTM) as an online learning approach to make time series predictions.