Trusting the Mesh: When a Sensor Network Stops Telling the Truth
A sensing system divides a volume into three parts and reports two of them. There is what a sensor reached and read correctly, there is what no sensor reached, and there is what a sensor reached and misread. The third part is the dangerous one, because a drifted sensor does not go quiet: it keeps reporting on schedule with the same decimal places it had on installation day, and everything downstream counts it as coverage. This course builds the competency that catches it. Published drift rates are the starting data, the mesh's own readings are the diagnostic, and the deliverable is a disclosure a reader can check rather than a number they have to trust. Every figure is sourced and every bench runs on the data given.
▶ Start the course ← All coursesWhere this sits, and what moves it.
Binding constraint · Calibration drift measured against the difference a decision must resolve. A field-installed sensor loses accuracy at a published rate, and a reading is usable only while that error stays smaller than the difference it is being asked to resolve, which makes the usable life a property of the decision rather than of the device.
Calibration was a maintenance schedule inherited from a vendor datasheet, applied uniformly because there was no method for doing otherwise, and a drifted node was found when somebody noticed a number that looked wrong.
Drift rates for field soil sensors are published, so the interval can be computed per decision rather than assumed, and a mesh can be audited against itself: a real field event is spatially coherent and instrument drift stops at the enclosure, which separates them without a reference instrument and without a field visit.
The untrusted fraction becomes a reported quantity alongside resolved and unknown volume, because it is the one a reader can recover with a technician rather than a new sensor. What would move this is a data standard carrying measurement uncertainty and drift with the reading, since ISOBUS and ADAPT move the payload today and the payload has no uncertainty field.
Every hard thing was impossible until the constraint that made it impossible was named. How we read a frontier →
What a mesh actually claims
Separate volume no sensor reaches from volume a sensor reaches and misreports, and compute how long a reading stays usable against the decision it feeds.
- L2The wrongly known fractionForty nodes cover a field, six of them drifting. Which is the larger number?Distinguish volume no sensor reaches from volume a sensor reaches and misreports, and show why the second is the more expensive of the two.→
- L2How fast a sensor forgetsWhy does one sensor produce several different calibration intervals?Put measured drift rates against a decision threshold and compute how long a reading stays usable.→
Catching a mesh that lies
Establish that agreement between identical sensors tests only their unshared error, and use spatial coherence to separate instrument drift from real field change.
- L3Agreement is not accuracyTwo identical nodes agree to within 0.4 percent. What has been established?Show that redundant sensors can agree closely and be wrong together, and identify the condition under which agreement carries information.→
- L3Drift has a signatureSeparate instrument drift from real change in the field using the spatial and temporal structure each one has.→
Acting on it
Price a calibration interval against the decisions it protects, and write a disclosure that states resolved, unknown and untrusted volume.
- L3What recalibration is worthA cheap node feeds an expensive decision. What happens to its calibration interval?Price a calibration interval against the cost of the decisions it protects, and find the interval that minimises total cost.→
- L4Reporting what you cannot trustWhy report untrusted volume separately instead of folding it into unknown?Write a sensing disclosure that states resolved, unknown and untrusted volume, and defend it against the objection that admitting uncertainty weakens the claim.→