Operationalizing a problem: Time-series Anomaly detection with multiple units and multiple sensors

Question

I have an IoT problem I am trying to operationalize.

I have multiple machines that should behave similarly over time (a good example is wind turbines nearby). They each have multiple sensors. And I have data over time.

Besides detecting absolute values that are dangerous (e.g. overheating), I am trying to operationalize how would I cross-reference the different sensors and different machines to detect other anomalies such as sensor failure.

Best I can think of at the moment are disjointed models within-machine/cross-sensors regression, and across-machines/across-sensors regression and then looking at then flagging large residuals.

HOWEVER, I am thinking there could be a way to maybe assemble this as some sort of covariance matrix or another, more computationally elegant solution. However, googling and reviewing literature has not yielded much of value yet.

I know this is broad, but hopefully allowed.

Thanks

Answer 1

Just a thought, you could use "boosting", e.g. monitor the relation of two similar sensors. To stick to the example, the ratio of temparature measurement of unit 1 and unit 2 should stay within a certain boundary.

I would consider combining sensors that either measure similar things, or use similar technologies, co-variances may be too computationally heavy.

Another technique, similar to the above, that is used in a different context is to compare one such measurement to the median and average of all such measurements. If the value of one sensor exceeds a threshold, you can do a relatively simple voting algorithm. This is how aeroplanes select altitudes from four sensors.