<p>A fog-based IoT system can have dozens of thousands of heterogeneous devices with distinct computing constraints, presenting different unknown and expected behaviors. Furthermore, the fog nodes’ expected behaviors can change depending on spatiotemporal features. Some investigations use supervised machine learning (ML) techniques to deal with anomalies, but they do not learn new behaviors in execution time. Unsupervised ML can detect outliers as anomalies, but it does not consider the spatiotemporal features. Therefore, it is challenging to design a fog-based IoT system that identifies the behaviors of every fog node during execution and which considers the variability of the expected behavior. This paper proposes a system named IoT-Beeps that uses ML to group these nodes’ behaviors into clusters and identifies each fog node’s performance anomalies. Different behaviors were used to evaluate the proposed system, using traditional ML clustering algorithms such as Mean-Shift, BIRCH, and K-Means, in two fog nodes. The clustering data models created were also evaluate using Silhouette, Davies-Bouldin, and Calinski-Harabasz indexes. Finally, the proposed system performance in the simulated experiment achieved at least 99% accuracy for the first fog node when using the K-Means algorithm and at least 98% accuracy for the second fog node using the K-Means or BIRCH algorithms, using the Calinski-Harabasz index for clustering validation.</p>