• Last week, we spent some time on the concept of redundancy as a basis for automatic data validation. We also discussed four ways to achieve redundancy. This week, we will detail model-based redundancy and outline an approach to using it for data validation. The first step here is to take all the sensor data we wish to validate and build a model connecting them to each other.

    For example, let us assume we have four sensors – current, torque, speed and noise – that we wish to monitor for bad data. We build a model (our preference is to use a Bayesian network) relating current to torque, torque to speed and speed to noise (see Slide 1). The arrows connecting them represent physical processes captured using well understood physics or (and) data. 

    Once we have the model, we can start validating each sensor data against each other. In the example discussed here, we input the value of current into the model and see what the model says torque should be. If the model predicted torque agrees with the sensed torque, then belief in the torque sensor as well as the process relating current to torque is increased. If not, those beliefs are decreased. We do this for all combinations of sensors. For the model in the example, we end up with 6 different validations. After the 6 steps are completed, we add up the beliefs to determine potential sensor / process faults.
    Slide 2 tabulates the sensor and process fault beliefs after the 6 validation steps for the case when there is no sensor fault, when there is a faulty torque sensor, and when there is a process fault. Of course, the model used in this example is a simple representational model. The real model is a lot more complex and would involve combining various sub models. The model that is currently deployed in Sentinel RT is on Slide 4. Remember that the goal with these models is to maximize redundancy. More details can be found in SPE-181076-MS and SPE-166387-MS for those interested in further study.

    Slide 5 shows a fault caught and corrected by Sentinel RT in the hook load reading. The approach above is not without issues or challenges. Stay tuned to learn more.

    Click below for slides on the topic:

    Data Quality Method