Downtime on an automated line is one of the most expensive failure modes in industry, running from hundreds of thousands to millions of dollars an hour depending on the process. A system that produces false positives stops production unnecessarily. A system that misses real defects ships bad product. Both are costly, and both usually come from the same root: a system built against conditions the line no longer matches. And where a system is closing a control loop on live equipment, the constraints get harder again: the deadline is real, the hardware cannot be taken out of service, and an incorrect actuation is not recoverable by retrying.
Failure in a serious system is rarely random. These are the shapes we look for first.
Line variance the system never saw
The development set captured the line on a good day. Production brings the full range: lighting changes across shifts, material lot variation, wear, and conditions clean data never contained. Performance degrades as the line drifts away from what the system was built against.
Signatures that shift with the operating regime
The same healthy machine produces different vibration and current signatures at different loads and speeds. Judged against a global baseline, normal operation flags as anomalous every time the regime changes, which is the single largest source of false alarms in condition monitoring.
The real-time and protocol layer
Adaptive control on live hardware runs over industrial protocols on equipment that cannot be stopped. The loop deadline is part of the specification, the protocol round-trip is inside it, and the controller behaves as it implements Modbus rather than as the standard describes.
False positives that erode trust
A system that stops the line too often gets overridden, then ignored, then switched off. An inspection or maintenance system nobody trusts is worse than none, because it carries cost without delivering protection, and the day it stopped being believed is never recorded anywhere.
The same method, in your language.
We architect and build floor systems designed for the variance of real production from the start, including the ones with no model in them: adaptive real-time control on live PLC hardware over Modbus, where the work is the control law, the timing, and the protocol layer rather than any inference.
We harden deployed systems with monitoring that catches drift as the line changes, and we set operating points against the plant's economics rather than against a metric.
When an inspection or maintenance system is underperforming, we establish whether the cause is line variance, a pipeline divergence, a sensor or signal problem, or a genuine limit, and we fix it in the order that holds.
Industrial programs are where the division of labour in this market breaks down most obviously. The work spans the sensor, the signal chain, the control law, the timing budget, the industrial protocol, the controller firmware behaviour, and sometimes a model, and almost every vendor owns one of those and depends on the rest being somebody else's problem. We delivered gain-scheduled adaptive control on live drilling equipment over Modbus to a running controller, inside a safety-instrumented system, on equipment that could not be taken out of service, with no model in it anywhere. We delivered condition monitoring across an 87-unit fleet where false alarms fell 41 percent, because three of the five deciding decisions were below the model. And we built an industrial camera from the sensor and FPGA up to a standard host, where the failure that decides the product is silent. One owner across the sensor, the control, and the protocol means the system can be changed when the line changes.