SRIJAN SANCHAR |
METHODOLOGY SERIES | IEF v3.0 | APPLIED FORESIGHT
Foresight on the
Research Landscape of
Industrial Machine Learning
Applying Integrated Epistemic Foresight v3.0
(Morphological Synthesis as Step III.0 within the Forward CCN) to map where
research on industrial machine learning has come from, where it is heading, and
where deliberate intervention can shape it. The subject is framed as presented
in Markus Geipel's post "Why Industrial Machine Learning is Hard".
|
Field |
Entry |
|
Subject (variable) |
Industrial machine learning: machine
learning applied to physical industrial assets and processes, where labels
are scarce, useful data is rare, exploration is costly and systems are only
partly observed. |
|
Landscape boundary |
In: ML for turbines, factory
machines, process plants, logistics and robotics; data-efficient and Bayesian
methods; symbolic AI (search, optimisation, planning); simulation and digital
twins; human-in-the-loop systems. Out: web and consumer ML (used only as the
contrast case), chip and accelerator design, enterprise back-office AI.
Geography: global, not region-specific. |
|
Prepared by |
Srijan Sanchar (working draft) |
|
Prepared for |
Not yet specified. Written for a
research institution or funder setting an industrial ML agenda; to be
confirmed. |
|
Date and version |
16 November 2026 | v0.1 |
|
Foresight window |
2026 to 2036; sequential link to
prior foresight: none (this is the baseline year). |
|
Method |
IEF v3.0 | Master equation: P(Sₙ₊₁ |
CCF, M, P(Tᵢ|Sₙ), Ω, H) × V(Sₙ₊₁) = Strategic Forecast |
Machine
learning succeeded first on the web, where data is abundant, labels come almost
free and a wrong recommendation costs little. Industry is a different setting.
In his post "Why Industrial Machine Learning is Hard" (11 November
2026), Markus Geipel argues that the difficulty is not mainly computational or
algorithmic. It comes from four properties of industrial problems: labels need
specialist knowledge, the most useful data (failures and anomalies) is rare,
exploration on a real machine is expensive and can be dangerous, and physical
systems are only partly measured.
The post then
describes five responses that work in practice: data-efficient algorithms,
algorithms that augment human operators, symbolic AI, simulations and digital
twins, and interdisciplinary teams.
This document
takes that practitioner account as its starting frame and asks a wider
question: where is research on industrial machine learning going over the next
ten years, and what should a research institution or funder do about it? It
answers by working through the IEF v3.0 Research Landscape template in order.
Phase 0 defines the phenomenon. Component I generates 25 theses across five
modes of thinking. Component II traces the field's history as six convergence
nodes and extracts the constraints they left behind. Component III arranges
research directions on a 4×4 matrix and forecasts how the landscape may move.
Component IV sets a target direction and four interventions.
|
How to read this draft What rests
on the post. The four challenges, the five strategies and the industrial
examples (Francis turbine cavitation, steam turbine rotors, gas turbine
combustion and control) come from Geipel's post. What is
the analyst's addition. The historical nodes, the two axes, the matrix,
the probabilities and the interventions are analytical judgements built on
the post and on general knowledge of the field. They are not the post
author's views. What is
missing. No bibliometric, patent or funding search was run, and no
experts were interviewed. Probabilities are judgements, not measurements. The
Human Node step (III.5) is left open for a real expert session. |
The five-component sequence of IEF v3.0, as applied to
the industrial machine learning research landscape
|
# |
Component |
Primary function |
Deliverable in this
report |
|
0 |
Phenomenon Definition |
Define in five dimensions; set
objective; name two provisional axes |
Profile of industrial ML research
and the decision it serves (Part I) |
|
I |
Taxonomy of Thinking |
Five-tier thesis generation with
axis, seed and quadrant flags |
25 theses on the field's drivers,
structure, paradigms and futures (Part II) |
|
II |
CCN / CCF |
Map genealogy as convergence nodes;
extract constraint field; derive Axis 1 |
Six nodes from process control to
digital twins; six constraints; Axis 1 = knowledge locus (Part III) |
|
III |
Forward CCN |
III.0 matrix; III.1 constraint test;
III.2 weighting; III.3 forecast state; III.4 resolution map; III.5 human
nodes |
4×4 matrix of research directions;
forecast state: Bifurcation (Part IV) |
|
IV |
Directional Vector |
Derive aim; P×V gap; spatially
grounded interventions |
Target cell and four interventions
for funders, institutions and researchers (Part V) |
This table
records what the foresight drew on, so it can be audited. Most streams were not
searched for this version. That is the main limit on the confidence of the
forecast.
|
Evidence stream |
What was captured for
industrial machine learning |
Source / date accessed |
|
Publications and citations |
One practitioner essay and the three
references it lists. No volume trend, venue analysis or citation data was
collected. |
Geipel, "Why Industrial Machine
Learning is Hard", accessed 16 November 2026. No database query run. |
|
Patents and standards |
Not consulted. |
None |
|
Funding and policy |
Not consulted. |
None |
|
Institutions, people, infrastructure |
Only what the post names: a Siemens
factory example and pretrained model hubs. Wider institutional picture is
from general knowledge. |
Post; analyst's general knowledge |
|
Industry and society |
The post's account of how factories
operate: systems designed to run without failure, parts replaced before
damage, sensors chosen by cost and benefit. |
Post |
|
Expert voices |
A single practitioner voice (the
post's author). No interviews, surveys or workshops. |
Post. Step III.5 is pending. |
Output Document 1: Phenomenon Definition and Objective
Statement
|
Dimension |
Defining question |
Profile |
|
Cognitive |
What is being thought differently as
research evolves? Which questions are opening, which closing? |
The field is moving from "which
algorithm is best?" to "what information does the plant actually
give us, and what do we already know?". Closing: the belief that
web-scale recipes transfer directly. Opening: how to combine prior knowledge
with small data, how to learn safely, and how to decide whether a new sensor
is worth its cost. |
|
Economic |
How are funding, commercialisation,
talent and tools being restructured? Who captures value? |
Data and assets sit with equipment
makers and plant owners, so they set the terms. Each failed experiment has a
real cost in wear, energy and material. Value tends to go to whoever holds
fleet data and simulators. Specialists who can label machine data are scarce
and expensive. |
|
Social |
How are research communities,
institutions, trust and public engagement being reorganised? |
Work is shifting from separate data
science groups to teams where domain experts sit with ML engineers for the
whole project. Operator trust decides whether a system is used. Operators and
engineers are stakeholders, not only users. |
|
Epistemic |
What counts as evidence, method and
expertise, and how is that changing? |
Benchmark accuracy counts for less.
Field performance, calibrated uncertainty and physical plausibility count for
more. Simulation is becoming an accepted source of training data. Expertise
is defined as joint: domain knowledge plus ML. |
|
Civilisational |
What long-run human capacities are
implicated? What would be irreversibly lost or gained? |
At stake is who understands the
machines that supply energy, water and goods. Possible gain: safer plants
that use less energy and material. Possible loss: operating knowledge that is
never written down and leaves with a retiring workforce, or passes into
closed models that no one can inspect. |
|
# |
Element |
Statement |
|
1 |
Decision Horizon |
Assumed, to be confirmed: a research
institution or funder must set a three-to-five-year research agenda and
partnership plan for industrial ML by mid-2027. |
|
2 |
Strategic Question |
Which research directions in
industrial ML will become dominant by 2036, which valuable directions are
neglected, and where can a modest, deliberate investment change the path? |
|
3 |
Acceptable Uncertainty Range |
Act on forecast states at or above
0.4. Treat states between 0.25 and 0.4 as options to keep open through small,
reversible commitments. |
|
4 |
Provisional Axis Candidates |
Candidate A: Knowledge-led (explicit
models and rules) → Data-led (learned from data) Candidate B: Human decides (analysis
and advice) → Machine acts (closed-loop autonomy) |
☑ The research phenomenon is describable in all
five dimensions.
☑ The objective names a real decision horizon
(assumed; to be confirmed with the commissioning body).
☑ At least three pressures on the current state
are identified: (1) pretrained foundation models and assistants arriving from
the web setting; (2) loss of expert operating knowledge as experienced staff
leave; (3) pressure to cut energy and material use; (4) falling sensor cost,
which shifts what is observable. Pressures 2 to 4 are analyst additions, not
from the post.
☑ Two provisional axis candidates are named.
Output Document 2: Thesis Inventory (25 theses with
morphological flags)
Theses are
numbered T1 to T25. Flags mark the raw material for Step III.0. Theses that
restate the post are T1, T8, T9, T10, T12 and T19; the rest are
analyst-generated.
What drives the field, how it is organised, what it
resembles and how its approaches can be ordered.
|
Mode |
Generating question |
Theses |
Flag |
|
Causal |
What drives the growth, direction
and pace of research? |
T1: Research direction is set less by
algorithmic novelty than by four properties of industrial problems: scarce
labels, adversely skewed data, costly exploration and partial observability.
Pace is set by asset owners' economics, not by compute. |
- |
|
Structural |
How is the field organised? |
T2: The field is split between the ML
community and engineering communities (control, maintenance, process
systems). Industrial firms hold the assets and data, so public benchmarks are
few and unrepresentative. |
- |
|
Comparative |
Which other fields share its deep
structure, and what does that predict? |
T3: Industrial ML shares deep structure
with medical ML: only experts can label, critical events are rare,
experiments are costly and safety is regulated. This predicts slow,
assurance-led adoption, with decision support arriving before autonomy. |
- |
|
Classificatory |
What categories of research make up
the field? Can they be ranked? |
T4: Approaches can be ranked by where
knowledge comes from: first-principles and symbolic methods, hybrid
physics-plus-data methods, pretrained and transfer methods, and models
trained from scratch on local data. T5: Approaches can also be ranked by
who acts: offline analysis, advice to an operator, supervised autonomy within
limits, and full closed-loop autonomy. |
AXIS CANDIDATE (both) |
Explanations, analogies and combinations. Each
combination names the matrix position it implies.
|
Mode |
Generating question |
Theses |
Flag |
|
Abductive |
What explanation makes the field's
anomalies intelligible? |
T6: Many pilots and few scaled
deployments make sense if the binding limit is the information in the data
(rare failures, unmeasured states), not model capacity. A larger model cannot
recover what was never recorded. |
- |
|
Analogical |
What is this research most deeply
analogous to? |
T7: Industrial ML is structurally like
learning to fly in a simulator: experience is gained where failure is free,
then transferred under supervision to the real asset. |
- |
|
Combinatorial |
What happens when the field combines
with adjacent methods and tools? |
T8: Simulation combined with
reinforcement learning gives control policies trained on a twin and deployed
under supervision, as Schaefer et al. (2007) did for gas turbines. T9: Bayesian methods combined with
human-in-the-loop design give recommendations that state their own
uncertainty, so the operator knows when to trust them. T10: Pretrained models combined with
operator workflows give assistants and few-shot diagnostics that need little
local data. |
MATRIX SEED (all three) |
|
Hypothetical |
What if the dominant assumption is
wrong? |
T11: What if the scarcity assumption is
wrong? If data is pooled across fleets and models are pretrained on it,
data-led methods could win in industry as they did on the web. |
- |
These theses set the value dimension, V(Sₙ₊₁), for each
cell.
|
Mode |
Generating question |
Theses |
Flag |
|
Normative |
What should this research achieve? |
T12: Research should make industrial
systems safer and less wasteful, and should strengthen operator and engineer
judgement instead of displacing it. |
- |
|
Skeptical |
Who benefits from the dominant
agenda? |
T13: The dominant agenda favours those
who own data, simulators and compute: large equipment makers and platform
vendors. Smaller firms and operators supply knowledge and labels but capture
little. A shared data commons is the alternative this reveals. |
- |
|
Deconstructive |
What is the field really doing
beneath its stated aims? |
T14: Beneath its stated aims, much of
the field produces papers on a few public benchmark datasets and supports
vendors' AI positioning. Evidence from real deployments is under-reported. |
- |
|
Evaluative |
Which possible future is better, by
stated criteria? |
T15: The better future is the one judged
by verified field performance per unit of data and risk, by auditability and
by retained human expertise. On those criteria, assured hybrid systems beat
black-box autonomy. |
- |
Loops, interfaces, thresholds and selection pressures.
|
Mode |
Generating question |
Theses |
Flag |
|
Feedback |
Which reinforcing and balancing
loops drive the trajectory? |
T16: Reinforcing loop: deployment builds
trust, trust brings sensors and data, data improves models. Balancing loop:
one visible failure or a stream of false alarms removes operator trust and
data access. In safety-critical assets the balancing loop dominates today. |
- |
|
Emergent |
What appears at the interfaces that
no single group produces? |
T17: At the interface of ML, control and
simulation, physics-informed ML has emerged as a sub-field, and so has the
role of the ML engineer embedded with domain experts. No single community
produced either. |
- |
|
Threshold |
Where are the discontinuities? |
T18: Likely discontinuities: cheap
retrofit sensing; simulation accurate enough for policies to transfer; a
certification route for learned components; one flagship autonomy result on a
major asset class. |
- |
|
Evolutionary |
What is the landscape selecting for? |
T19: The landscape selects for methods
that are data-efficient, interpretable and compatible with human operators.
Data-hungry methods survive only where a simulator or a fleet supplies the
data. |
- |
History, default trajectory, constraints and distinct
futures.
|
Mode |
Generating question |
Theses |
Flag |
|
Historical |
How did the field arrive at its
current form? |
T20: The field arrived here through
process control, model-based control, expert systems, condition monitoring
and the deep-learning wave. The recurring pattern: each data-led wave is
later re-anchored in domain knowledge. |
- |
|
Prospective |
What are the trajectories if nothing
deliberate is done? |
T21: If nothing deliberate is done: to
2028, assistants and advisory tools multiply; 2028 to 2032, hybrid twins
appear only on high-value assets; 2032 to 2036, autonomy concentrates in a
few vendor platforms while smaller firms stay at advisory level. |
- |
|
Constraint |
What must remain true however
research evolves? |
T22: Whatever happens, safety limits on
exploration, a clear line of accountability and the physical limits of what
is measured must hold. These forbid fully autonomous systems that learn by
live trial and error, or that cannot be verified. |
FORBIDDEN ZONE |
|
Scenario |
What are the structurally distinct
futures? |
T23: Scenario "Assured
hybrid": knowledge-anchored models trained partly in simulation run
under supervised autonomy on high-value assets. T24: Scenario "Foundation
scaling": cross-fleet pretrained models move from advice to bounded
action, led by a few platform owners. T25: Scenario "Advisory
plateau": the field stays at data-led alerts and dashboards, and pilots
continue without scaling. |
MATRIX QUADRANT (all three) |
Matrix cells
are written as (Axis 1 position, Axis 2 position). Axis 1 is knowledge locus;
Axis 2 is agency locus. Both are defined in Part IV.
|
ID |
Thesis (short form) |
Tier / mode |
Flag |
Matrix cell |
|
T1 |
Research direction is set less by
algorithmic novelty than by four properties of industrial problems: scarce
labels, adversely skewed data, costly exploration and partial observability. |
1 / Causal |
- |
(Mid-low, Mid-low) |
|
T2 |
The field is split between the ML
community and engineering communities (control, maintenance, process
systems). |
1 / Structural |
- |
(High, Mid-low) |
|
T3 |
Industrial ML shares deep structure
with medical ML: only experts can label, critical events are rare,
experiments are costly and safety is regulated. |
1 / Comparative |
- |
(Mid-low, Mid-low) |
|
T4 |
Approaches can be ranked by where
knowledge comes from: first-principles and symbolic methods, hybrid
physics-plus-data methods, pretrained and transfer methods, and models
trained from scratch on local data. |
1 / Classificatory |
AXIS CANDIDATE |
Axis 1 |
|
T5 |
Approaches can also be ranked by who
acts: offline analysis, advice to an operator, supervised autonomy within
limits, and full closed-loop autonomy. |
1 / Classificatory |
AXIS CANDIDATE |
Axis 2 |
|
T6 |
Many pilots and few scaled
deployments make sense if the binding limit is the information in the data
(rare failures, unmeasured states), not model capacity. |
2 / Abductive |
- |
(High, Mid-low) |
|
T7 |
Industrial ML is structurally like
learning to fly in a simulator: experience is gained where failure is free,
then transferred under supervision to the real asset. |
2 / Analogical |
- |
(Mid-low, Mid-high) |
|
T8 |
Simulation combined with
reinforcement learning gives control policies trained on a twin and deployed
under supervision, as Schaefer et al. (2007) did for gas turbines. |
2 / Combinatorial |
MATRIX SEED |
(Mid-low, Mid-high) |
|
T9 |
Bayesian methods combined with
human-in-the-loop design give recommendations that state their own
uncertainty, so the operator knows when to trust them. |
2 / Combinatorial |
MATRIX SEED |
(Mid-low, Mid-low) |
|
T10 |
Pretrained models combined with
operator workflows give assistants and few-shot diagnostics that need little
local data. |
2 / Combinatorial |
MATRIX SEED |
(Mid-high, Mid-low) |
|
T11 |
What if the scarcity assumption is
wrong? If data is pooled across fleets and models are pretrained on it,
data-led methods could win in industry as they did on the web. |
2 / Hypothetical |
- |
(Mid-high, Mid-high) |
|
T12 |
Research should make industrial
systems safer and less wasteful, and should strengthen operator and engineer
judgement instead of displacing it. |
3 / Normative |
- |
(Mid-low, Mid-high) |
|
T13 |
The dominant agenda favours those
who own data, simulators and compute: large equipment makers and platform
vendors. |
3 / Skeptical |
- |
(Mid-high, Mid-low) |
|
T14 |
Beneath its stated aims, much of the
field produces papers on a few public benchmark datasets and supports
vendors' AI positioning. |
3 / Deconstructive |
- |
(High, Mid-low) |
|
T15 |
The better future is the one judged
by verified field performance per unit of data and risk, by auditability and
by retained human expertise. |
3 / Evaluative |
- |
(Mid-low, Mid-high) |
|
T16 |
Reinforcing loop: deployment builds
trust, trust brings sensors and data, data improves models. |
4 / Feedback |
- |
(High, Mid-low) |
|
T17 |
At the interface of ML, control and
simulation, physics-informed ML has emerged as a sub-field, and so has the
role of the ML engineer embedded with domain experts. |
4 / Emergent |
- |
(Mid-low, Mid-low) |
|
T18 |
Likely discontinuities: cheap
retrofit sensing; simulation accurate enough for policies to transfer; a
certification route for learned components; one flagship autonomy result on a
major asset class. |
4 / Threshold |
- |
(Mid-low, Mid-high) |
|
T19 |
The landscape selects for methods
that are data-efficient, interpretable and compatible with human operators. |
4 / Evolutionary |
- |
(Mid-low, Mid-low) |
|
T20 |
The field arrived here through
process control, model-based control, expert systems, condition monitoring
and the deep-learning wave. |
5 / Historical |
- |
Rows Low to Mid-low |
|
T21 |
If nothing deliberate is done: to
2028, assistants and advisory tools multiply; 2028 to 2032, hybrid twins
appear only on high-value assets; 2032 to 2036, autonomy concentrates in a
few vendor platforms while smaller firms stay at advisory level. |
5 / Prospective |
- |
(Mid-high, Mid-low) |
|
T22 |
Whatever happens, safety limits on
exploration, a clear line of accountability and the physical limits of what
is measured must hold. |
5 / Constraint |
FORBIDDEN ZONE |
(High, High), (Mid-high, High) |
|
T23 |
Scenario "Assured hybrid":
knowledge-anchored models trained partly in simulation run under supervised
autonomy on high-value assets. |
5 / Scenario |
MATRIX QUADRANT |
(Mid-low, Mid-high) |
|
T24 |
Scenario "Foundation
scaling": cross-fleet pretrained models move from advice to bounded
action, led by a few platform owners. |
5 / Scenario |
MATRIX QUADRANT |
(Mid-high, Mid-high) |
|
T25 |
Scenario "Advisory
plateau": the field stays at data-led alerts and dashboards, and pilots
continue without scaling. |
5 / Scenario |
MATRIX QUADRANT |
(High, Mid-low) |
|
Taxonomy Completeness Check (v3.0) All eight
checks pass. |
☑ Minimum three theses in each tier (5, 6, 4,
4, 6).
☑ No thesis duplicated across tiers.
☑ At least one normative thesis: T12.
☑ At least one feedback thesis: T16.
☑ At least one constraint thesis: T22.
☑ At least two AXIS CANDIDATE flags: T4, T5.
☑ At least three MATRIX SEED theses: T8, T9,
T10.
☑ At least two MATRIX QUADRANT flags: T23, T24,
T25.
Output Document 3: CCN Map and CCF Document (with Dead
End Register)
The history of
industrial ML is mapped as six convergence nodes. Each is a moment when a
tension made the earlier view untenable and a synthesis resolved it, leaving a
constraint. The post supplies the content of Nodes 4 to 6. Nodes 1 to 3 are the
analyst's reconstruction from general knowledge of control engineering and AI
history, and the dates are approximate.
|
NODE 1: Measurement replaces craft judgement |
|
|
Node ID and era |
1; 1920s to 1950s |
|
Prior worldview |
A skilled operator's senses and
experience were the control system. This seemed complete because no
instrument could match an expert's feel for a machine. |
|
The polarity |
Tacit craft judgement versus formal
measurement: can quality and stability be held by numbers instead of by feel? |
|
Competing theses |
Craft position: only experience
captures the whole machine. Measurement position: variation can be measured,
charted and corrected by rule. |
|
Failed attempts |
Attempts to codify craft by written
procedure alone failed because procedures could not react to variation in
real time. |
|
The synthesis |
Statistical process control and
feedback control. Measured signals drive the routine corrections; the
operator handles the exceptions. |
|
CCF constraint established |
Any future synthesis must rest on
measured, traceable quantities and must say plainly what is not measured. |
|
Axis Polarity Score (1-5) |
4: the question of what the human
still decides returns in Nodes 3 and 5. |
|
Political filter (who gained, who
was excluded) |
Plant owners and engineers gained
control and consistency. Craft workers lost standing, and part of their
knowledge was never recorded. |
|
Dead ends to note |
The idea that procedure can fully
replace judgement. |
|
NODE 2: Model-based estimation and control |
|
|
Node ID and era |
2; 1960s to 1980s |
|
Prior worldview |
Controllers were tuned by trial on
the plant. This seemed sufficient for single loops. |
|
The polarity |
First-principles models versus
empirical tuning: should control come from physical equations or from what
works on the plant? |
|
Competing theses |
Modellers: physics gives guarantees
and can estimate states that no sensor reads. Empiricists: models are never
accurate enough, and tuning is cheaper. |
|
Failed attempts |
Purely theoretical controllers
failed on model mismatch. Purely empirical tuning failed on large, coupled
processes. |
|
The synthesis |
State estimation and model
predictive control: a physical model corrected continuously by measurement.
Unmeasured states are inferred, not ignored. |
|
CCF constraint established |
Any future synthesis must stay
consistent with established physical models and must treat unobserved states
explicitly. |
|
Axis Polarity Score (1-5) |
5: models versus data is the tension
that returns in every later node. |
|
Political filter (who gained, who
was excluded) |
Process industries with resources to
build models gained most. Plants without modelling capacity were left with
manual tuning. |
|
Dead ends to note |
The belief that a model, once built,
stays valid as the asset ages. |
|
NODE 3: The rise and fall of expert systems |
|
|
Node ID and era |
3; 1980s to 1990s |
|
Prior worldview |
Expert knowledge was thought to be
fully expressible as rules, so a rule base could stand in for the expert. |
|
The polarity |
Encoded knowledge versus learned
knowledge: write the rules down, or let the system learn them? |
|
Competing theses |
Knowledge engineers: rules are
transparent and need no data. Learning advocates: rules are brittle and
experts cannot articulate all they know. |
|
Failed attempts |
Large monolithic rule bases became
too costly to build and maintain. Early neural networks lacked data and
compute. |
|
The synthesis |
Explicit knowledge survived as
narrow, verifiable components: heuristic search, optimisation and planning.
These are the methods the post calls symbolic AI. |
|
CCF constraint established |
Any future synthesis must not depend
on exhaustive hand-encoding, and its knowledge must be maintainable by the
people who run the process. |
|
Axis Polarity Score (1-5) |
5: the same polarity as Node 2,
restated for AI. |
|
Political filter (who gained, who
was excluded) |
Operations research and planning
groups gained durable tools. Many experts gave their knowledge and saw the
systems abandoned. |
|
Dead ends to note |
The universal rule-based reasoner.
It is being revived in new form as knowledge graphs paired with language
models. |
|
NODE 4: Condition-based maintenance under rare failures |
|
|
Node ID and era |
4; 1990s to 2010s |
|
Prior worldview |
Parts were replaced on a fixed
schedule. This seemed safe and complete because it prevented most failures. |
|
The polarity |
Scheduled replacement versus
prediction from condition: replace by the calendar, or by what the machine's
signals show? |
|
Competing theses |
Schedulers: early replacement is
cheap insurance. Predictors: it wastes useful life and still misses
unexpected faults. |
|
Failed attempts |
Supervised failure classifiers
failed for lack of failure examples. Systems are built to run without failure
and parts are replaced before damage, so the data that matters is the rarest. |
|
The synthesis |
Condition monitoring built on
anomaly detection and on models of normal behaviour, with experts
interpreting the alerts. |
|
CCF constraint established |
Any future synthesis must give
valid, uncertainty-qualified outputs when failures are rare and labels are
few. |
|
Axis Polarity Score (1-5) |
3: the scarcity it addresses
persists, but the polarity itself is largely settled. |
|
Political filter (who gained, who
was excluded) |
Equipment makers gained service
revenue from monitoring. Independent maintainers often lost access to the
data. |
|
Dead ends to note |
Running assets to failure to collect
labels. |
|
NODE 5: The deep-learning wave meets the factory |
|
|
Node ID and era |
5; about 2011 to 2020 |
|
Prior worldview |
Methods that worked on web data were
expected to work anywhere given enough data and compute. |
|
The polarity |
Web-scale data-led methods versus
industrial conditions: abundance, cheap labels and cheap errors on one side;
scarcity, expert labels and costly errors on the other. |
|
Competing theses |
Transfer view: collect more data and
the methods will work. Industrial view: the data cannot be collected at
acceptable cost, and crowdsourcing fails because labelling needs specialists. |
|
Failed attempts |
Web recipes transplanted directly
produced many pilots and few scaled systems. Full automation was resisted by
the operators who had to live with it. |
|
The synthesis |
The position Geipel's post sets out:
data-efficient methods (augmentation, transfer learning, classical
statistics, Gaussian processes) and systems that advise the operator instead
of replacing them. |
|
CCF constraint established |
Any future synthesis must keep a
named, competent human or a certified function answerable for the outcome,
and must preserve operator understanding. |
|
Axis Polarity Score (1-5) |
4: the question of human versus
machine agency from Node 1, now sharper. |
|
Political filter (who gained, who
was excluded) |
ML platform vendors and data
scientists gained standing. Domain experts were often treated as label
suppliers, not as partners. |
|
Dead ends to note |
End-to-end models trained from
scratch on small plant datasets. |
|
NODE 6: Simulation and digital twins as the place to
learn |
|
|
Node ID and era |
6; 2007 to present |
|
Prior worldview |
Learning required real experience on
the real asset, so learning to control was limited by what the asset could
safely endure. |
|
The polarity |
Explore on the real asset versus
explore in a model of it: exploration is how reinforcement learning improves,
but on a machine it causes wear, cost and risk. |
|
Competing theses |
Real-world learners: only the real
asset is accurate. Simulation advocates: a good enough model allows free
exploration and makes data-hungry methods usable again. |
|
Failed attempts |
Cautious exploration on live
machines yielded too little data. Policies trained on inaccurate simulators
failed when transferred. |
|
The synthesis |
Train on a simulation or twin,
validate offline, then deploy under supervision. Schaefer et al. (2007) built
a gas turbine simulation with ML, trained control policies on it and later
deployed them on real turbines. |
|
CCF constraint established |
Any future synthesis must not
explore without bounds on live assets; learned control must be validated
offline before it acts. |
|
Axis Polarity Score (1-5) |
3: recent, and partly resolved
wherever a good simulator exists. |
|
Political filter (who gained, who
was excluded) |
Firms with simulators and fleet data
gained a strong advantage. Operators of older or unusual assets with no twin
were left out. |
|
Dead ends to note |
Unsupervised live exploration on
production equipment. |
|
# |
Source node |
CCF constraint
statement |
Status |
|
C1 |
Node 1 |
Any future synthesis must rest on
measured, traceable quantities and state what is not measured. |
Non-negotiable |
|
C2 |
Node 2 |
Any future synthesis must stay
consistent with established physical models and treat unobserved states
explicitly. |
Non-negotiable |
|
C3 |
Node 3 |
Any future synthesis must not depend
on exhaustive hand-encoding; its knowledge must be maintainable by those who
run the process. |
Provisional |
|
C4 |
Node 4 |
Any future synthesis must give
valid, uncertainty-qualified outputs when failures are rare and labels few. |
Non-negotiable |
|
C5 |
Node 5 |
Any future synthesis must keep a
named human or certified function answerable, and preserve operator
understanding. |
Non-negotiable |
|
C6 |
Node 6 |
Any future synthesis must not
explore without bounds on live assets; learned control must be validated
offline first. |
Non-negotiable |
|
Dead-end idea |
Why it failed |
Being revived now? |
|
Universal rule-based expert system |
Too costly to build and maintain;
experts cannot state all they know. |
Yes, as knowledge graphs paired with
language models. Worth re-examining (see M3). |
|
Fully unmanned plant |
Could not handle exceptions; removed
the people who caught problems. |
Yes, in autonomous-plant narratives.
C5 and C6 still apply. |
|
Web recipes with crowdsourced labels |
Machine data needs specialist
labellers; the general crowd cannot do it. |
Partly, through self-supervised
pretraining that avoids labels. |
|
Run-to-failure data collection |
Too costly and unsafe. |
No, except inside simulators. |
|
Item |
Entry |
|
Highest-scoring polarity |
Nodes 2 and 3 (score 5), recurring
in Nodes 5 and 6: explicit models and encoded knowledge versus knowledge
learned from data. |
|
Axis 1 as a continuum |
Knowledge-led (first principles,
rules, symbolic methods) → Data-led (learned from local data alone) |
|
Compared with Taxonomy AXIS
CANDIDATE flags |
Agrees with T4. The second
candidate, T5 (who acts), is carried forward as Axis 2. It also scored 4 in
Nodes 1 and 5, so it has support from the history as well as from the
Taxonomy. |
|
The Deepest Polarity Rule, applied Remove the
tension between explicit knowledge and learned knowledge and the history
above stops making sense: model-based control, expert systems, the
deep-learning wave and digital twins are each an answer to it. All five of
the post's strategies are also ways of bringing prior knowledge back in when
data is short. That is why it is Axis 1. |
Step III.0 opens the Forward CCN: morphological
organisation before probability weighting
Output Document 4: Morphological Matrix
|
# |
Element |
Entry |
|
1 |
Axis 1 (from Step II.3) |
Knowledge locus: Knowledge-led
(first principles, rules, symbolic) → Data-led (learned from local data
alone) |
|
2 |
Axis 2 (from Taxonomy flags) |
Agency locus: Human decides (offline
analysis) → Machine acts (full closed-loop autonomy). From T5. |
|
3 |
Orthogonality test |
High Axis 1 with low Axis 2 is
possible: a purely data-driven dashboard that only informs an analyst. Low
Axis 1 with high Axis 2 is possible: a classical controller built from
physical equations that runs with no human in the loop. Both exist today, so the
axes are independent. |
|
4 |
Granularity decision |
4×4. The interesting movement is in
the two middle positions of each axis (hybrid and pretrained; advice and
supervised autonomy), which a 2×2 would hide. |
Each cell is a
research direction. Rows are Axis 1 (knowledge locus); columns are Axis 2
(agency locus).
|
AXIS 1 ↓ / AXIS 2 → |
Low: human analyses
offline |
Mid-low: system advises,
human acts |
Mid-high: system acts
within limits, human supervises |
High: full closed-loop
autonomy |
|
High: learned from local data alone |
Plant data mining Descriptive analytics on historian
data POPULATED |
Black-box alerts T2, T6, T14, T16, T25 CURRENT STATE |
Black-box bounded control Works only where the system is
richly instrumented and errors are cheap POPULATED (narrow) |
Live trial-and-error autonomy T22 FORBIDDEN (C6, C5) |
|
Mid-high: pretrained and transferred |
Pretrained models for engineering
analysis No theses placed EMPTY |
Assistants and few-shot diagnostics T10, T13, T21 POPULATED; waystation shared by both
paths |
Fleet-pretrained bounded control T11, T24, M1 EMPTY (candidate M1) |
Unverifiable pretrained autonomy T22 FORBIDDEN (C5, C4) |
|
Mid-low: hybrid of physics and data |
Physics-informed surrogate models Grey-box models for design and
analysis POPULATED |
Uncertainty-aware hybrid advice T1, T3, T9, T17, T19 THINLY POPULATED; waystation |
Assured hybrid autonomy T7, T8, T12, T15, T18, T23 TARGET |
Self-optimising plant M2 EMPTY (candidate M2) |
|
Low: first principles, rules,
symbolic |
Simulation and engineering analysis T20 POPULATED (mature) |
Rule-based advisory M3 LARGELY VACATED (candidate M3) |
Planning and optimisation under
supervision Scheduling, routing, robot motion
planning POPULATED (mature) |
Classical automatic control T20 POPULATED (mature incumbent) |
|
Cell status |
Shown as |
Meaning |
|
CURRENT STATE |
Yellow, gold border |
Where the dominant research
programme sits now: data-driven alerts and predictions evaluated on
benchmarks. |
|
TARGET |
Green, green border |
The cell the Directional Vector aims
at: highest V(Sₙ₊₁). |
|
FORBIDDEN |
Pink, red border |
Inadmissible under a non-negotiable
constraint (named in the cell). |
|
Waystation |
Light blue |
Cell on the path between current
state and target. |
|
EMPTY |
White |
Reachable but not yet articulated:
white space. |
|
POPULATED |
Light grey |
Established research direction that
is neither the current centre nor the target. (Added to the template's five
statuses for clarity.) |
|
Reading |
Question |
Finding for industrial
machine learning |
|
1. Cluster pattern |
Where do theses concentrate? |
Two forward-looking clusters with a
third at the starting point. Cluster A, hybrid row: 11 theses around
(Mid-low, Mid-low) and (Mid-low, Mid-high). Cluster B, pretrained row: 5
theses plus M1. Cluster C, the current cell: 5 theses. Pre-classification: Bifurcation,
with a small gap because A and B share a waystation. |
|
2. Empty cell inventory |
Which empty, non-forbidden cells
exist, and which are novel? |
M1 (Mid-high, Mid-high): models pretrained across a
fleet take bounded control actions on a single asset. M2 (Mid-low, High): a plant that re-optimises itself
using hybrid models, with accountability held by a certified function instead
of a person at the controls. M3 (Low, Mid-low): rule-based advice rebuilt with
language models and knowledge graphs, used to capture the knowledge of
experienced operators before they leave. (Mid-high, Low): pretrained models
used as tools for engineering analysis. Not novel; no candidate raised. |
|
3. Forbidden zone geometry |
Open feasible set, narrow corridor,
or zones approaching each other? |
A large open feasible set. The two
forbidden cells sit together in the top-right corner (most data-led, most
autonomous). No corridor and no sign of constraint collision. The corner
blocks the shortest path from the current state to autonomy, so progress must
first move down Axis 1. |
|
4. Intervention trajectory |
What path leads from the current
cell to the target? |
(High, Mid-low) → (Mid-high,
Mid-low) → (Mid-low, Mid-low) → (Mid-low, Mid-high). First add prior
knowledge while staying advisory, then earn the right to act. Knowledge
before agency. |
☑ Matrix-generated thesis candidates: M1, M2,
M3.
☑ Forecast state pre-classification:
Bifurcation.
☑ Morphological trajectory: (High, Mid-low) →
(Mid-high, Mid-low) → (Mid-low, Mid-low) → (Mid-low, Mid-high).
☑ Morphological Matrix document produced (this
section).
Output Document 5 begins here: Forward CCN Forecast
Report (Steps III.1 to III.5)
Every thesis
and matrix-generated candidate is tested against the six constraints. P = pass,
C = conditional, F = fail. No thesis fails outright, because T22 had already
placed the inadmissible directions in forbidden cells.
|
Thesis |
Source |
Matrix cell |
C1 |
C2 |
C3 |
C4 |
C5 |
C6 |
CCF result |
Weight |
|
T1 |
Taxonomy |
(Mid-low, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T2 |
Taxonomy |
(High, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T3 |
Taxonomy |
(Mid-low, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T4 |
Taxonomy |
Axis 1 |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T5 |
Taxonomy |
Axis 2 |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T6 |
Taxonomy |
(High, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T7 |
Taxonomy |
(Mid-low, Mid-high) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T8 |
Taxonomy |
(Mid-low, Mid-high) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T9 |
Taxonomy |
(Mid-low, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T10 |
Taxonomy |
(Mid-high, Mid-low) |
P |
C |
P |
C |
P |
P |
CONDITIONAL |
H |
|
T11 |
Taxonomy |
(Mid-high, Mid-high) |
C |
C |
P |
C |
P |
C |
CONDITIONAL |
M |
|
T12 |
Taxonomy |
(Mid-low, Mid-high) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T13 |
Taxonomy |
(Mid-high, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T14 |
Taxonomy |
(High, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T15 |
Taxonomy |
(Mid-low, Mid-high) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T16 |
Taxonomy |
(High, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T17 |
Taxonomy |
(Mid-low, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T18 |
Taxonomy |
(Mid-low, Mid-high) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T19 |
Taxonomy |
(Mid-low, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T20 |
Taxonomy |
Rows Low to Mid-low |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T21 |
Taxonomy |
(Mid-high, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
T22 |
Taxonomy |
(High, High), (Mid-high, High) |
P |
P |
P |
P |
P |
P |
PASS (defines the forbidden cells) |
H |
|
T23 |
Taxonomy |
(Mid-low, Mid-high) |
P |
P |
P |
P |
P |
P |
PASS |
H |
|
T24 |
Taxonomy |
(Mid-high, Mid-high) |
P |
C |
P |
C |
C |
C |
CONDITIONAL |
M |
|
T25 |
Taxonomy |
(High, Mid-low) |
P |
P |
P |
P |
P |
P |
PASS |
M |
|
M1 |
III.0 matrix |
(Mid-high, Mid-high) |
P |
C |
P |
C |
C |
C |
CONDITIONAL |
L (needs validation) |
|
M2 |
III.0 matrix |
(Mid-low, High) |
P |
P |
P |
C |
C |
C |
CONDITIONAL |
L (needs validation) |
|
M3 |
III.0 matrix |
(Low, Mid-low) |
P |
P |
C |
P |
P |
P |
CONDITIONAL |
L (needs validation) |
Weighting is
done by cluster. Scores are analyst judgements on the template's four criteria;
they have not been checked against publication or deployment data.
|
Criterion |
Cluster A: Assured
hybrid |
Cluster B: Foundation
scaling |
Cluster C: Advisory
plateau |
|
Internal coherence |
High. Each step follows from the
constraints. |
Medium. Assumes fleet data can be
pooled, which data owners may refuse. |
High, but it describes a stall, not
a programme. |
|
Anomaly-resolution capacity |
Resolves three: stalled pilots, the
rare-failure problem and unsafe exploration. |
Resolves one or two: label scarcity,
and possibly rare failures across a large fleet. |
Resolves none. |
|
Constraint-compatibility depth |
Reinforces C2, C4 and C6. |
Satisfies the constraints
conditionally; in tension with C2 and C5. |
Satisfies all, by attempting little. |
|
Maturation |
Worked examples exist (Schaefer et
al. 2007; Gaussian process practice) but are few. |
Assistants are spreading quickly at
the advisory level; bounded control is undemonstrated. |
Fully mature. |
Weighted
result, P(Tᵢ|Sₙ) for 2036: Cluster A (T1, T3, T7, T8, T9, T12, T15, T17,
T18, T19, T23) = 0.45. Cluster B (T10, T11, T13, T21, T24, M1) = 0.30. Cluster
C (T2, T6, T14, T16, T25) = 0.25. Only Cluster A clears the 0.4 action
threshold set in Step 0.2.
|
State |
Test |
Mark and explain |
|
A: Forced Move |
One cluster dominates and the matrix
confirms it. |
Not marked. Cluster A leads but does
not dominate. |
|
B: Bifurcation |
Two clusters of comparable weight;
the matrix shows the size of the gap. |
Marked. A (0.45) and B (0.30) are both
live. The gap is small: they share the waystation (Mid-high, Mid-low) and sit
in adjacent cells at the supervised-autonomy level. They may converge if
pretrained models are given physical structure. |
|
C: Constraint-Collision |
No thesis satisfies the full CCF
without relaxation. |
Not marked. Cluster A passes all six
constraints. |
Agreement
with III.0 pre-classification: Yes. One refinement: there is a third
outcome, the advisory plateau (0.25), which is what happens if neither path
reaches its tipping condition.
|
# |
Element |
Entry for industrial
machine learning |
|
1 |
Fork description |
Path A moves down Axis 1 to hybrid
methods, then right: (Mid-high, Mid-low) → (Mid-low, Mid-low) → (Mid-low,
Mid-high). Path B stays in the pretrained row and moves right: (Mid-high,
Mid-low) → (Mid-high, Mid-high). The fork is at the shared waystation. |
|
2 |
Tipping condition |
Path A tips when (Mid-low, Mid-low)
becomes well populated: several independent, published field deployments of
uncertainty-aware hybrid advice that operators keep using. Path B tips when
(Mid-high, Mid-high) is populated by one credible demonstration of a
fleet-pretrained model controlling an asset within limits. |
|
3 |
Intervention design |
In sequence: (a) open,
expert-labelled benchmarks that include rare events; (b) shared twin testbeds
for safe exploration; (c) an assurance route for learned components; (d)
embedded teams that keep operator knowledge in the loop. |
|
4 |
Timeline indicator |
The trust loop in T16 is the clock.
The current wave of assistants is likely to reach saturation of attention
around 2028 to 2029 (judgement). If hybrid advice has not shown field results
by then, the plateau becomes more likely. |
|
Status: pending No domain
experts, funders or practitioners have been consulted. The column below holds
the analyst's provisional hypotheses, offered as prompts for that session.
They are not expert responses and should be replaced. |
|
Human node |
Matrix-specific
question |
Provisional hypothesis
(to be replaced by named responses) |
|
H1 Abductive injection |
Which empty cell, if populated,
would most change how you understand industrial ML? |
Probably M1. If a fleet-pretrained
model can act safely, the scarcity premise of the whole field weakens. |
|
H2 Values reassignment |
Is the target cell the highest-value
synthesis, or should it be another cell on ethical grounds? |
Some practitioners may argue the
target should stay at (Mid-low, Mid-low), keeping the human as the actor, at
least for safety-critical assets. |
|
H3 Early collision sensing |
Is any forbidden cell actually
feasible? Is the constraint logical or political? |
(Mid-high, High) may be forbidden by
present verification methods, not by logic. It could open if verification
improves. (High, High) looks forbidden by physics and cost. |
|
H4 Suppressed thesis recovery |
Should any Dead End Register idea be
reactivated? |
The expert system, in the form of
M3, may be worth reviving for knowledge capture. |
|
H5 Cross-domain isomorphism |
Does the matrix resemble a topology
in another domain? |
It resembles levels of driving
automation and clinical decision support: advice first, bounded action next,
with full autonomy blocked by assurance. |
Output Documents 6 and 7: Directional Vector and P×V
Matrix; Strategic Intervention Map
|
# |
Element |
Entry |
|
1 |
Irreducible purpose |
To help people run physical systems
more safely and with less waste, by turning scarce measurements and hard-won
expertise into better decisions. |
|
2 |
Preservation conditions |
Operators and engineers remain able
to understand and override the system. Knowledge of the process stays with
those who run it, not only inside a vendor's model. Evidence comes from field
performance, not benchmark scores alone. Smaller firms and public
institutions have access to data and testbeds. These are the criteria for
V(Sₙ₊₁). |
|
3 |
Directional aim |
This foresight is aimed at the
future in which safer, less wasteful operation of physical systems is
preserved and expanded, under the condition that human understanding,
accountability and broad access are kept. Target cell check: Yes. (Mid-low,
Mid-high) anchors models in physics and keeps a human supervising. |
|
Synthesis state / cell |
P(Sₙ₊₁) |
V(Sₙ₊₁) |
P×V implication |
Strategic role |
|
Current state: (High, Mid-low)
Black-box alerts |
HIGH as a starting point; 0.25 that
it persists as the plateau |
Low to medium. Useful, but leaves
the main problems unsolved. |
The probable baseline |
Redirect |
|
Target: (Mid-low, Mid-high) Assured
hybrid autonomy |
0.45 |
HIGHEST |
The normative aim; probable enough
to act on |
Build toward |
|
Waystations: (Mid-high, Mid-low) and
(Mid-low, Mid-low) |
High for the first; medium for the
second |
Medium; high for the second |
The second waystation is the tipping
point |
Populate through intervention |
|
Rival: (Mid-high, Mid-high)
Fleet-pretrained bounded control |
0.30 |
Medium. High capability; weaker on
access and physical grounding. |
A live alternative that may converge
with the target |
Keep open with small, reversible
work |
|
Forbidden: (High, High) and
(Mid-high, High) |
ELIMINATED |
N/A |
Outside the feasible set |
Design against |
|
INTERVENTION 1: Open rare-event benchmarks to populate
(Mid-low, Mid-low) |
|
|
Matrix cell target |
(Mid-low, Mid-low): hybrid of
physics and data; system advises, human acts |
|
Trajectory position |
Early |
|
P×V gap addressed |
High value, thinly populated.
Without shared evidence, hybrid methods cannot be compared and Cluster C
(benchmark papers on unrepresentative data, T14) persists. |
|
Mechanism |
A consortium of universities and
asset owners releases expert-labelled machine datasets that include failures
and anomalies, with the physical model of each asset and a scoring rule that
rewards calibrated uncertainty. Expert labelling is funded as research work
in its own right. |
|
Resolution condition produced |
The second waystation fills.
Observable as papers reporting calibrated results on the shared sets and as
methods reused across asset types. |
|
Timeline |
Start 2027; first datasets 2028;
visible effect by 2029, before the assistant wave saturates. |
|
Primary actor |
A university consortium, with two or
three equipment makers or plant operators persuaded to contribute data. |
|
INTERVENTION 2: Shared digital-twin testbeds to reach
(Mid-low, Mid-high) |
|
|
Matrix cell target |
(Mid-low, Mid-high): hybrid of
physics and data; system acts within limits, human supervises |
|
Trajectory position |
Mid |
|
P×V gap addressed |
High value, medium probability. Safe
exploration is available today only to firms that own simulators. |
|
Mechanism |
Open twin testbeds for a few asset
classes (for example a turbine, a pump system and a production cell), each
paired with a physical rig. A common protocol measures how well a policy
trained on the twin transfers to the rig. The same testbeds are used to study
when an added sensor is worth its cost, the question the post calls a tough
call. |
|
Resolution condition produced |
The target cell gains independent
demonstrations outside vendor laboratories. Observable as published
twin-to-rig transfer results and reuse of the protocol. |
|
Timeline |
Start 2028; results 2030 to 2032. |
|
Primary actor |
Public research funders and
engineering departments, with an equipment maker supplying asset models. |
|
INTERVENTION 3: An assurance route for learned components
at (Mid-low, Mid-high) |
|
|
Matrix cell target |
(Mid-low, Mid-high), and the
boundary with the forbidden cells |
|
Trajectory position |
Late |
|
P×V gap addressed |
The target cannot scale until
someone can certify that a learned controller stays within its limits (C5,
C6). |
|
Mechanism |
A working group of standards bodies,
insurers, operators and researchers defines what evidence is needed before a
learned component may act: an operating envelope, an offline validation
record, uncertainty monitoring and a defined handover to the human. |
|
Resolution condition produced |
Supervised autonomy becomes
insurable and repeatable. Observable as a published guideline and its first
use on a real asset. |
|
Timeline |
Start 2029; guideline by 2032;
adoption 2032 to 2036. |
|
Primary actor |
Standards and safety bodies,
convened with industry associations. |
|
INTERVENTION 4: Embedded teams and operator-knowledge
capture at (Mid-low, Mid-low) and M3 |
|
|
Matrix cell target |
(Mid-low, Mid-low) and the candidate
cell (Low, Mid-low) |
|
Trajectory position |
Early, and continuing |
|
P×V gap addressed |
The hybrid path depends on domain
knowledge that is leaving the workforce. This also keeps the value condition
on human understanding. |
|
Mechanism |
Joint programmes in which domain
experts work inside ML teams for the whole project, as the post recommends,
plus a research strand on recording operator knowledge in maintainable form
(candidate M3). Curricula that teach engineers ML and teach ML students
control and physical modelling. |
|
Resolution condition produced |
A supply of people who can work in
the hybrid row, and a test of whether M3 is viable. Observable as joint
appointments, shared courses and operator-authored knowledge bases in use. |
|
Timeline |
Start 2027; first cohort and pilots
by 2029. |
|
Primary actor |
Universities and plant operators;
operators and engineers treated as partners. |
|
Dharma gate (Srijan Sanchar addition, not part of the
IEF v3.0 template) Each
intervention was screened on intent and on execution before inclusion. Intent.
All four aim at safer operation, less waste and stronger human judgement.
None is aimed at removing people from the work. Guardrails
to build before misuse. For 1 and 4: experts and operators who supply
labels and knowledge must be credited and must share in the benefit, so that
knowledge capture does not become knowledge extraction. For 2: testbed
results must not be used to justify autonomy on assets the testbed does not
represent. For 3: the assurance route must not become a way to shift
liability onto operators. This box can
be removed if the report is issued in strict template form. |
|
# |
Check |
Status |
Notes |
|
1 |
Thesis-to-Node: every
historical-claim thesis is consistent with the CCN nodes |
Pass |
T20 matches Nodes 1 to 6. |
|
2 |
Constraint-to-Thesis traceability:
every constraint traces to a node and conditions at least one thesis |
Pass |
C1 to C6 trace to Nodes 1 to 6; each
conditions at least one of T10, T11, T24, M1 to M3. |
|
3 |
Forecast-to-Taxonomy completeness:
every cluster contains theses from at least three tiers |
Pass |
A: five tiers. B: tiers 2, 3, 5. C:
tiers 1 to 5. |
|
4 |
Vector-to-CCF coherence: the
directional aim contradicts no constraint |
Pass |
|
|
5 |
P×V actionability: at least one
intervention is executable within the decision horizon |
Pass |
Interventions 1 and 4 can start in
2027. |
|
6 |
Matrix-to-CCF consistency: every
forbidden cell maps to a constraint; a constraint that forbids nothing is too
weak or the axes too narrow |
Partial |
Both forbidden cells map to
constraints (C4, C5, C6). C1, C2 and C3 forbid no cell on these axes; they
act only as conditions. This suggests a third axis (for example, how
observable the asset is) may be needed in v0.2. |
|
7 |
Trajectory feasibility: the path
passes only through non-forbidden cells |
Pass |
|
|
Component |
Minimum standard |
Status in this report |
|
Taxonomy |
15-25 theses; all five tiers; 2+
AXIS CANDIDATE, 3+ MATRIX SEED, 2+ MATRIX QUADRANT flags |
Met: 25 theses; 2, 3 and 3 flags. |
|
CCN / CCF |
4-8 nodes; Axis Polarity Scores;
political filter applied; Axis 1 derived |
Met in form: 6 nodes. Nodes 1 to 3
rest on general knowledge and need source checking. |
|
Step III.0 |
Two orthogonal axes; all theses
placed; four readings; 2+ matrix-generated candidates; trajectory mapped |
Met: 3 candidates. |
|
Steps III.1-5 |
All theses constraint-tested;
forecast state confirmed against III.0 |
Partly met: III.5 Human Nodes is
pending. |
|
Directional Vector |
Target cell on matrix; 3-5
interventions with cell target and trajectory position |
Met: 4 interventions. |
|
# |
Output document |
Section in this report |
Complete? |
|
1 |
Phenomenon Definition and Objective
Statement |
Part I |
Yes (decision horizon assumed) |
|
2 |
Thesis Inventory |
Part II |
Yes |
|
3 |
CCN Map and CCF Document |
Part III |
Yes (Nodes 1 to 3 to be
source-checked) |
|
4 |
Morphological Matrix (Step III.0) |
Part IV, III.0 |
Yes |
|
5 |
Forward CCN Forecast Report |
Part IV, III.1-III.5 |
Draft (III.5 pending) |
|
6 |
Directional Vector and P×V Matrix |
Part V, IV.1-IV.2 |
Yes |
|
7 |
Strategic Intervention Map |
Part V, IV.3 |
Yes |
This is the
first foresight in the sequence, so nothing is carried forward from a prior
year. It sets the baseline for a 2027 review. That review should re-test three
assumptions: that failures stay rare in the available data even across pooled
fleets (T11); that assistants remain advisory and do not move into bounded
control (M1); and that no certification route for learned components appears
earlier than expected (T18). Each added year widens the uncertainty, even where
the story becomes more plausible.
|
Executive Summary Research on
industrial machine learning sits today mostly in data-driven alerts and
predictions that advise an operator. Many pilots do not scale, because the
limit is the information in the data: failures are rare, labels need experts,
experiments are costly and much of the machine is unmeasured. The deepest
tension in the field's history is between explicit knowledge (physics, rules,
models) and knowledge learned from data. A second tension is who acts, the
human or the machine. Arranged on these two axes, the landscape shows a
Bifurcation. One path anchors learning in physics and simulation, then moves
to supervised autonomy (probability 0.45). The other scales pretrained models
across fleets (0.30). If neither tips, the field plateaus at advice (0.25).
These figures are analyst judgements. The target is
assured hybrid autonomy: physics-anchored models, trained partly on digital
twins, acting within limits under human supervision. It scores highest on
safety, understanding and access. Three
interventions matter most: open rare-event benchmarks with expert labels
(from 2027); shared digital-twin testbeds (from 2028); and an assurance route
for learned components (guideline by 2032). Embedded interdisciplinary teams
support all three. Watch for:
field results from uncertainty-aware hybrid advice by 2029, and any credible
demonstration of a fleet-pretrained model controlling a real asset. |
Markus Geipel, "Why
Industrial Machine Learning is Hard", Markus' Blog, 11 November 2026. mgeipel.github.io/posts/industrial-machine-learning
(accessed 16 November 2026).
A. M. Schaefer,
D. Schneegass, V. Sterzing and S. Udluft, "A Neural Reinforcement Learning
Approach to Gas Turbine Control", IJCNN 2007, pp. 1691-1696. Cited as
listed in the post; not read directly.
Historical
material in Nodes 1 to 3 and the pressures marked as analyst additions draw on
general knowledge and have not been checked against sources for this version.