Sensor Validation

Definition

It is said that you cannot manage what you cannot measure, therefore it is important that your measures exist, function, are reasonably accurate and are useful. Here we explore three perspectives of Sensor Validation, from the most basic to the most sophisticated, all three being very useful.

---

First Perspective: Making Sure a Sensor is Operational

Validating for Correctly Operating Sensors

Challenge:

Sensor values are important for the proper operation of most any device, whether that be your home's thermostat or the core temperature of a nuclear reactor. In this case we are validating that there is at least a signal from the sensor, the sensor is not "flat lining", the sensor is emitting the expected form of data and the sensor's values are within a proper range based on definition or on physics. This is implemented in real-time.

Solution:

Handling this is conceptually easy but surprisingly not well attended to in practice; one merely needs to check that the sensor's values exist, they are changing within an expected time period (a "live" sensor measuring a natural property such as temperature or pressure has an expected degree of noise), and are congruent with the data type emitted by the sensor. For example, some sensors are numeric such as a temperature or pressure and thus text amongst the numerics (e.g., "Bad", "I/O Error", etc.) would be inappropriate and indicate a problem along the information chain. On the other hand, some sensors emit text by their nature (e.g., "OPEN" or "RUNNING") and in that case null (empty) or numeric or particular text (e.g., "I/O Timeout") are clearly invalid sensor readings. The sensor's values should be rational, for example a negative Kelvin for a temperature is wrong, unless one is working on a new theory in physics... (smile). In Intellect implementations these types of checks are routinely implemented. Alerts and maintenance actions can be issued based on detected invalid data or errors. Metrics and reports on data validity and availability can be created.

---

Second Perspective: Detecting When a Sensor is Failing or Out of Calibration

Validating for Congruent Sensor Values

Challenge:

A sensor's readings should be related to the conditions it is measuring. If the sensor starts to drift or starts making irrational readings, the comparison between the virtual sensor and the actual physical sensor will indicate the problem. Once detected a repair or recalibration can be investigated.

Solution:

It is possible to build a multivariate regression model using surrounding conditions vs known-good sensor readings then make independent estimates of the sensor's values. We refer to these as "Virtual Sensors". We can then compare the virtual sensor's estimates to the values from the currently operating sensor in real-time. A threshold on the difference or computed metrics such as deviance or variance, instantaneous readings or moving through time, can be used to drive alerts and maintenance actions ("condition-based maintenance"). In some cases the divergence is predictable as a time series and thus this solution can move forward to become a "Predictive Maintenance" solution.

---

Third Perspective: Understanding the Utility of a Sensor

Validating that the Sensor Provides Unique Useful Information

Challenge:

In new operations there is a tendency to over-instrument processes in order to be assured of the necessary information to accomplish one or more results. However, over-instrumentation causes unnecessary capital and continuing operational expenses, sometimes substantial. The challenge is to identify the set of necessary and sufficient sensors to achieve one or more technical and business objectives.

Solution:

Our exclusive non-linear predictive modeling technology uses combinatorial "sensor selection" optimization. This combinatorial search identifies and ranks sensors by their information utility and sheds sensors that do not provide additional information or create more noise than additional unique information. We conduct off-line studies to model and analyze the relative utility of sensors to determine the key information-containing sensors and possible alternative sensor sets to achieve the desired technical and business results. You can use this information to decide which sensor(s) you need to carry forward in future process implementations, reducing your capital and operating expense.