Mosaic-AM: multimodal sensors anomaly detection with inherent compensation for absent modalities

Large scale deployment of the Internet of Things (IoT) has produced a disruptive effect in many fields in the recent past. Continuous connectivity combined with relatively low physical implementation cost has produced a Big Data paradigm shift, especially in industrial contexts. Unfortunately, the ability to process and adequately make use of this data has not kept pace with deployment. Specifically, machine learning models in use today lack the ability to perform well with data from a variety of sources. For instance, many models are trained using only one type of data. Even models trained on multimodal data lack the ability to predict on different combinations of this data. Sensor deployment on identical machines is often different depending on context, leading to the need for multiple models created for the same machine. The data in question has the ability to radically shift how equipment failure is predicted and when maintenance is completed, when processed correctly. The cost savings on large industrial machines and potential saved downtime could be enormous. This research investigates a new semi-supervised, technology agnostic anomaly detection framework that can be utilized on any combination of data modalities collected from sensors deployed on an equipment in a factory setting. This framework is then tested on a real-world anomaly dataset. Through testing, we demonstrate that the model maintains state of the art performance even in the face of multiple sensor failures.