Two signals supposed to be identical (plant redundancy) differ due to weather effects. How design a model to correct for weather effects? Context
Imagine a plant where a piece of equipment is placed outside, subject to wind and temperature changes. To monitor a specific aspect of this equipment, two identical sensors were installed at different positions at the equipment. The sensors and produced by the same manufacturer, and they measure exactly the same thing. Two sensors were installed due to redundancy. When the two signals differ by more than 5%, control room personnel receive an alarm and an automatic work order is generated instructed technicians to inspect the sensors to see what is wrong.
However, after some time in operation, it is noticed that the alarms generated often are not due to actual different readings by the sensors, but because the local environment where the sensors are placed are subject to different weather effects. Changes in wind direction, wind strength and temperature correlate with the observed sensor drift. A data scientist is now asked to build a model that quantifies how much of the drift that can be attributed to weather effects. The goal is for the control room staff to only receive alarms when the signals differ by more than 5% if these 5% cannot be explained by the weather effects.
Data
We have access to 5 years of wind strength data, wind direction data, and temperature data at the plant (not necessarily exactly where the sensors are placed). In addition, we also have access to 5 years of sensor data from both transmitters. All data is updated every 10 seconds.
My question
What would be a simple way to attack this model request? The question pertains only to what kind of model to build and how to analyze the data. Deploying the model into production is not within the question scope.
My initial thoughts
Make a regression model with the drift (difference between the two signals) and wind strength, wind direction, and temperature as features. Select subset of data for training:

*

*subset where the two signals with confidence were not drifting

*capture a range of different weather effects (strong wind, weak wind, all directions, large temperature range)

But I do not have much experience with this type of analysis. Can anyone provide some tips and guidance? I would love to prototype this in Python, and eventually move this to a more suited enterprise tool. I put the self-study tag to indicate that tips and hints are welcome, but this is an actual case from the industry.
 A: I don't know about the 5 percent thing, but you could model the difference between the sensor readings using weather features. Then, when actual differences are "different enough" from predicted differences, you could sound the alarm.
If you have labels in your data that indicate when there actually was a problem, this is much easier. Then you could predict if there's a problem using sensor and weather data.
OP requested an example, so here goes. At any time t, suppose your sensor readings are
$$
s_1(t), s_2(t)
$$
and your weather readings are
$$
w_1(t),w_2(t), w_3(t)
$$
Let
$$
d(t) = s_1(t) - s_2(t)
$$
be the difference between sensor readings. Then you can fit a linear regression like
$$
d(t) = \beta_0 + \beta_1w_1(t) + \beta_2w_2(t) + \beta_3w_3(t)
$$
to the data. The idea here is your model "learns" the normal relationship between sensor differences and weather effects. Of course, it doesn't have to be a linear regression.
Once you've fit the model, you feed it weather data and monitor difference between the actual and predicted sensor differences:
$$
d(t) - \widehat{d}(t)
$$
If actual differences are "different enough" from predicted differences, then weather alone can't explain what's happening, so it must be something else. You'll have to define "different enough".
