Topic
Vendor Skepticism
Category
Topics
3 issues matching filters
Vendor Skepticism
- Deep DiveJul 20, 2026
Predicting the failure before the anomaly: what remaining-useful-life forecasting needs that the reference stack does not have
Detection tells a plant that a machine is behaving abnormally now. Prediction claims to tell it how many days remain before the machine fails, and the vendor pitch treats that as the same problem solved one step earlier. It is not. Remaining-useful-life forecasting is a supervised problem that needs run-to-failure histories the reference stack has never recorded, and the published literature reaches its published accuracy on a benchmark of engines run deliberately to destruction. This issue separates the two questions the pitch fuses, models what a real plant can and cannot forecast from the history it actually keeps, and prices the two ways the forecast is wrong: retiring a healthy part early and missing the failure it promised to catch.
The prior issue put a small language model on the gated alert and enforced the one rule that kept it useful, that it may quote the manual and may never infer the cause. This issue takes the next and larger claim in the agentic-maintenance pitch, predictive maintenance, which promises to forecast the failure before the anomaly ever appears and to schedule the repair against a countdown rather than an alarm. The pitch presents prediction as detection solved one step earlier, the same competence pointed slightly further into the future, and that framing hides a change in the kind of problem being solved and the kind of data it requires. Anomaly detection, the capability the reference stack already has, is unsupervised. The Isolation Forest built across the early issues learns the shape of normal operation from healthy running data and flags departures from it, and healthy running data is the one thing every plant has in abundance. Remaining-useful-life forecasting is supervised and needs the opposite kind of data. To learn how many operating days remain before a failure, a model must be trained on examples of machines observed continuously from healthy through degradation to the failure itself, with the failure point labeled, and it needs many such examples to generalize, because a single run to failure teaches the model one machine's one death rather than the distribution of deaths. That run-to-failure history is precisely what a competently run plant does not keep, because competent maintenance intervenes before the failure and the record ends at the repair, not at the breakdown, so the endpoint the supervised model must learn from is the endpoint good operations is designed to prevent. The canonical public dataset that anchors most of the published RUL literature, NASA's C-MAPSS turbofan set, resolves the shortage the way a lab can and a plant cannot, by simulating engines run deliberately all the way to destruction under controlled fault injection, which is why the published accuracy figures describe forecasts on manufactured death and not on the interrupted, incomplete histories real plants accumulate. The issue separates the two questions the pitch fuses, models what a real stack can honestly forecast from the history it actually holds, which is narrower than a full RUL number and closer to a trend and a rate than a date, and then prices the asymmetry that governs the economics of any forecast: an estimate that retires a healthy component early converts unused remaining life into scrap and buys unnecessary labor, an estimate that overshoots and misses the failure it was purchased to prevent spends the unplanned downtime the whole program was justified by, and the two errors are not symmetric in cost, so the forecast's usefulness depends less on its average accuracy than on which direction it errs and what each direction costs on the specific asset. The numbers in this issue are modeled from published third-party benchmarks and the public prognostics literature, cited inline.
Predictive Maintenance·Remaining Useful Life·Rul·Prognostics·C Mapss·Run To Failure - Field NotesJul 13, 2026
Wiring a language model to the alert pipeline: the work order it can draft, and the diagnosis it still cannot
Issue 11 read the AI-in-manufacturing pitch from the floor and left one bounded test open: put a small local model where the pitch insists it belongs, on the alert the stack already produces, and measure what it adds. This issue runs that test on paper. When a gated flag fires, a small model turns the raw alert and the machine's manual into a serviceable first-line work order, the document task the FactoryBench results say it can actually do. Ask the same model in the same loop to name the fault from the signal and it walks straight into the 17.7 percent task it cannot. The value is real and the boundary is exact, and the design that earns the model its place is the one that never lets it cross the line.
Issue 11 set the franchise beat, the intersection of AI and manufacturing read from the floor, and closed by naming the first test: take the alert the reference condition-monitoring stack already produces and wire a small local language model to it, then model what the model adds rather than what a vendor demo claims. This issue runs that test. The finding is that the loop contains two tasks the agentic pitch keeps fused and that the model handles them at opposite ends of its competence. The first task is document retrieval. When a gated anomaly flag fires, the model receives the alert, the affected asset identifier, and the relevant section of the maintenance manual, and it drafts a first-line work order, naming the asset, quoting the applicable procedure, listing the parts, the tools, and the lockout and tagout steps, and leaving the priority and the assignment to a human. Published work on retrieval-augmented generation over maintenance and equipment documentation reports that small models do this reliably, because the task is retrieving and reformatting text the manual already contains rather than reasoning about the machine. The second task is diagnosis, deciding from the vibration signature what is actually wrong, and that is the exact task FactoryBench measured the best current models at 17.7 percent on, because reading an industrial time-series signal and retrieving a protocol document are separate competencies these models conflate. The same small model in the same alert loop can draft the paperwork and cannot name the cause, and the danger is that a confidently phrased wrong cause is indistinguishable in tone from a right one, so a model allowed to guess the fault will file plausible fiction into the maintenance record. The issue models the economics from the same edge-inference benchmarks issue 11 cited, a one-to-three-billion-parameter model returning single-digit to low-double-digit tokens a second on an owned edge box, which is fast enough to draft a work order in the seconds after a flag and far too slow to reason in a real-time loop, at a hardware cost that amortizes to a few dollars a month and adds nothing to the per-query bill because there is no per-query bill. The design conclusion is a single rule: the model may quote the manual and may never infer the cause, and the pipeline enforces the rule structurally by giving the model the alert and the document and withholding the raw signal it would need to hallucinate a diagnosis from. The numbers in this issue are modeled from published third-party benchmarks and vendor specifications, cited inline.
Retrieval Augmented Generation·Small Language Models·Work Orders·Cmms·Factorybench·Edge Inference - Deep DiveJul 6, 2026
The agentic-AI factory pitch meets the floor: the 95 percent that never reach the P&L, and the benchmark the best models fail
2026 is the year every industrial vendor sells agentic AI to manufacturing, and the adoption curve under the pitch is real. The trouble is the three numbers the pitch does not put on the slide: the MIT finding that 95 percent of generative-AI pilots return nothing measurable to the income statement, the benchmark where the best current model scores 17.7 percent at diagnosing a machine from its own sensor data, and the pilot that costs ten thousand dollars a month and five hundred thousand at production scale. This issue reads the wave from the floor rather than the keynote, and finds that the AI which actually pays is small, local, and narrowly scoped, with one asterisk the small-model camp does not advertise either: the eight-millisecond edge demo is a single-pass detection number, not a reasoning agent's, and conflating the two is the same trick one altitude down.
The series changes altitude and sets its beat for the issues ahead: the intersection of AI and manufacturing, read from the floor by a publication that has built the small version and can therefore read the large pitch. The first ten issues built a reference condition-monitoring stack, an Isolation Forest scoring vibration on one inexpensive VM, and that build is the credential for the question this issue turns to. The pitch is stated at full strength first, because the honest broker steelmans before it counters. Adoption is genuine: about a third of manufacturing operations are AI-augmented today by Rockwell's 2026 survey, Deloitte's 2026 Manufacturing Industry Outlook expects agentic adoption to roughly quadruple from about 6 to 24 percent even as only one in five reports being equipped to scale it, and the vendors at Hannover Messe and CES 2026 are selling agentic copilots, industrial foundation models trained on CAD and sensor data, and physical-AI robot cells. Then three numbers are set against it. MIT's NANDA study found that for 95 percent of companies, enterprise generative-AI pilots showed little to no measurable P&L return, and Gartner projects 60 percent of AI projects abandoned through 2026 for lack of the data foundation the first ten issues were spent building. FactoryBench, published in 2026, is the first benchmark to test whether a model can read industrial machine time-series rather than retrieve a manual, and the best current models, Claude Sonnet 4.6 and GPT-5.1, scored below 50 percent on structured reasoning and 17.7 percent on root-cause decision-making, direct evidence that the foundation models do not yet understand a machine from its signal. And pilot economics invert, a ten-thousand-dollar pilot becoming five hundred thousand at fifty times scale. The constructive turn is the small, task-specific model run locally, which Dell and Gartner are calling the real 2026 story, and which is the shape of the stack this series already built. But the issue adds the asterisk the small-model camp leaves off: the eight-millisecond, two-thousand-dollar edge figure describes a single-pass detection workload, not a generative reasoning agent, and on published edge-inference benchmarks a one-to-three-billion-parameter reasoning model returns single-digit to low-double-digit tokens a second, fine for drafting an offline work order and far short of a real-time agent. The honest synthesis is to match the model class to the task and to refuse both camps when they swap one spec sheet for the other. The numbers in this issue are modeled from published benchmarks and vendor specifications, cited inline.
Agentic Ai·Industrial Ai·Small Language Models·Foundation Models·Physical Ai·Factorybench