Model-based testing The system has access to a model of the manufacturing processes. The behavior of each of the processing steps is known for a variety of different failure modes, as well as for normal running. Causal modeling takes account of the long chain of effects caused by an upstream failure. So, given a hypothesis about the current state of the process, with or without faults, this model can be used to derive predictions that can be compared with the sensor data streams coming in from the actual process.

Situation awareness Our monitoring system maintains a working model corresponding to its belief about the current process state. Where the predictions of that model differ from the observations of the process, hypotheses are evoked that purport to explain the deviation. The amended models are simulated to discover which of them better matches the observations, and belief is changed accordingly. Because multiple faults will tend to cause multiple sets of deviations in the incoming sensor data, the system can diagnose multiple-fault situations.

Explanation Because our system uses the evidence – the sensor observations – coupled with the cause-to-effect reasoning from its compositional modeling, it can justify its conclusions to human operators, explaining its reasoning to them. This allows users to understand how the system works, and to supervise it in case of error.