IoT mesh networks typically consist of resource-constrained devices that communicate wirelessly. Since such networks are exposed to numerous attacks, designing intrusion detection systems is an important task and has attracted a lot of attention from the research community. Most existing work, however, has only considered a few network topologies, often also assuming a fixed number of nodes. In this paper, we generate a new large attack dataset, using Multi-Trace, a tool that we recently devised to generate traces to train machine learning algorithms. We show that using more and more diverse training data, the resulting intrusion detection models generalize better compared to those trained with less and less diverse training data. They even generalize well for larger topologies with more IoT devices. We also show that when we train different machine learning methods on our dataset, the resulting intrusion detection systems achieve very high performance. © 2022, The Author(s),