Predictive maintenance AI with accountable human handover

Learn in your browser · Chapter 01 of 06

An alert requests examination of evidence. It does not diagnose a fault, predict a remaining life or approve a maintenance change.

Tern Reliability Learning Desk uses three invented pump identifiers, P1, P2 and P3, to practise a review process. They are rows in a local dataset, not machines you can operate. The packet has 30 numbered daily observation slots per asset, except one deliberately missing reading. Each day means one equal teaching interval. There is no site, person, plant, customer, model service or equipment connection. The condition index is an invented unitless value; it is not vibration, temperature or a transferable physical limit. Scores are supplied ordinal inputs, not outputs of a model trained in this course.

Start by defining the decision. Tern wants to know whether a proposed alert queue is worth further controlled investigation, and what a person must check before changing an existing plan. The educational output is a reproducible evaluation and a draft handover. It is not a work instruction. The exercise cannot tell you how to inspect, isolate, operate, service or repair any pump. A score of 80 does not mean an 80 percent chance of failure. A later event label does not identify a defective part. Those meanings require evidence this packet does not contain.

Scroll sideways to see every column.

The target is deliberately narrow: at an eligible anchor day t, was at least one supplied event for the same asset recorded to occur strictly after t and no later than t+3? Day t itself is excluded from this future-window target. An event exactly three days later is included. The rule is a case definition for retrospective learning, not a universal maintenance rule. A positive label means the target condition occurred in that window; it does not say the score caused it or that an intervention would have prevented it. The same event can label several overlapping windows positive.

Keep observation time, availability time, event time and review time separate. A reading may describe yesterday but arrive tomorrow. An event may occur on day 19 and become known on day 21. The eventual record can support a later evaluation without having been available to a planner on day 19. This distinction applies even when a spreadsheet contains all columns in one file. File proximity is not evidence of historical availability. Every result should name the information cutoff and the source edition used.

A poor proposal says, 'The model predicts bearing failure in three days, so postpone the scheduled task and close the alert.' The packet supports none of those three actions or diagnoses. A defensible question is, 'Under the declared three-day target, which supplied alerts anticipated which invented events, what was withheld for data quality, and can a reviewer inspect the evidence?' A planner can choose whether further authorised investigation is needed under the real organisation's own process. Course completion creates no such authority.

1. Name the decision and audience before selecting metrics.
2. Write a target definition with asset identity, interval endpoints and label meaning.
3. State which fields were observable at the intended decision time.
4. List the numerical assumptions, excluded uses and responsible review role.
5. Define the paper output and keep any equipment action outside this exercise.

NIST SP 800-82 Revision 3 describes operational technology as interacting with the physical environment and recognises its performance, reliability and safety requirements. This course uses that distinction only to keep a desk-study analysis separate from equipment control. It does not implement the guide or claim a security assessment. The cited 2023 final edition and the dated source-review record identify the reference being used; a later draft is not silently substituted for it.

Learn in your browser · Chapter 02 of 06

A clean-looking average can combine observations that should never share one feature window. Expose coverage loss before calculating performance.

The feature rule uses the current reading and the two preceding scheduled days for the same asset. It requires all three rows, correctly aligned measurement days, availability by the anchor, and one common epoch. The checker reports the three-day mean condition index, but the alert decision uses the supplied score on the anchor row. It does not learn a relationship between the mean and future events. This separation prevents a worked data-quality calculation from being mistaken for a validated predictive model.

The raw file contains 89 rows from a nominal 90 daily asset slots. P3 day 16 is absent. P2's day-20 row was measured on day 18, so it is stale under this teaching rule even though the row arrived on day 20. P1 changes epoch at day 14; P2 changes at day 21. These are deliberately supplied reset markers. Preserve them in the source copy. Do not infer a reset from a convenient value change or insert a missing reading just to make an average calculable.

Scroll sideways to see every column.

P1 anchor 16 has mean (24+25+26)/3 = 25, with source days 14, 15 and 16 in epoch 1. Anchor 15 needs day 13 in epoch 0, so its superficially plausible arithmetic is irrelevant: the window violates the stated comparison rule. P2 day 20 can contaminate three consecutive features. Counting bad source rows alone therefore understates how much evaluation coverage is lost. One absent P3 row likewise removes three anchor windows, not just one.

The evaluation proposes anchors 15 through 27 inclusive for three assets: 13 times 3 = 39 nominal windows. Seven are held: P1/15; P2/20,21,22; P3/16,17,18. Thirty-two remain eligible, or 32/39 = 82.05 percent coverage. The held fraction is 7/39 = 17.95 percent. These are window-coverage measures, not uptime, reliability or safety percentages. A held window is not a negative prediction, a correctly avoided false alarm or evidence that no event happened.

Keep a reason list even when a window has more than one defect. The simple checker reports the first applicable reason in a declared order: missing row, stale measurement, unavailable evidence, then mixed epochs. P2 anchors 21 and 22 report stale evidence even though they also cross a reset. That reporting convention does not make the second problem disappear. A reviewer should inspect the supplied source rows when diagnosing a hold. If a real system needs every cause, its specification must say so and its implementation must retain all causes.

1. Preserve the original CSV and its declared schedule.
2. Check unique asset/day identity and the meaning of measured versus available time.
3. Construct the exact preceding-three-day set by asset.
4. Hold missing, stale, not-yet-available or cross-epoch windows.
5. Report nominal, eligible and held counts with asset/day reasons.
6. Only then calculate features and retain their source days.

The program rejects duplicate asset/day rows rather than choosing the last arrival. If two records conflict, arrival order cannot establish authenticity. It also rejects boolean or out-of-range numerical inputs, future measurement times and impossible availability ordering. These checks constrain the exercise format. They cannot detect a sensor that is consistently wrong while producing valid-looking numbers, a forged epoch label or an omitted event. Test such limits with known synthetic mutations, not with real machinery.

Learn in your browser · Chapter 03 of 06

A time-ordered split still leaks if a development label was not yet available when the decision was frozen.

Tern freezes its illustrative threshold and evaluation policy at the end of day 14. A development anchor needs a three-day future target and the declared two-day label reporting delay. To know the complete label at freeze, anchor + 3 + 2 must be no later than 14. Therefore the latest permitted development anchor is day 9. Days 3 through 9 are the declared development range. The two earlier days lack a full three-reading history. The score values remain supplied inputs; the course does not fit a detector or claim a learned threshold.

A tempting split uses development anchors through day 11 because day 11+3 reaches day 14. It overlooks label availability. That window's complete target is only available by day 16 under the declared delay, after the freeze. The checker rejects this change. A later file export cannot repair the chronology by putting the outcome column beside old features. A source may also become available later than its nominal event date. Reconstruct the information that could actually have been used at the time.

Scroll sideways to see every column.

Evaluation uses anchor days 15 through 27. Outcomes are declared complete through day 30 and available at day 32. Those declarations are fixture assumptions, not evidence that a real event register is complete. A final anchor at day 28 would need outcomes through day 31 and labels by day 33. Excluding that immature window is different from labelling it negative. At the evaluation cutoff, some real studies may have censored observations whose future outcome is not yet known. An explicit eligibility rule is necessary before choosing a method to handle them.

The five anchor days 10 through 14 are not used as development or evaluation anchors. This is not a generic instruction to discard five rows. There are three assets, unequal observed row counts and a missing slot. A gap measured in rows can represent different elapsed times after pooling assets or losing observations. Use asset identity and the declared calendar logic. The scikit-learn TimeSeriesSplit documentation describes time-ordered folds, notes equal spacing for comparable fold durations and defines gap in samples. This local exercise implements its own explicit day rules and does not depend on scikit-learn.

A further source of leakage is preparing a transformation using the evaluation period. If you later experiment with a fitted normaliser, imputer or threshold, estimate it from permitted development evidence only, then freeze its parameters before assessing later data. The present mean is an arithmetic feature with no fitted population parameter. That does not make every possible transformation safe. A centred mean using tomorrow's reading would violate the prospective feature definition even if the training and evaluation files were separate.

1. Write the freeze, development anchors, evaluation anchors and outcome-availability cutoff.
2. For every development target, check horizon plus reporting delay against freeze.
3. For every feature, check source measurement and availability times against its anchor.
4. Preserve asset identity and reset epochs when forming windows.
5. Freeze threshold and any learned preprocessing before later evaluation.
6. Record rejected immature or leaking rows instead of silently scoring them as negatives.

The exercise tests chronology, not generalisation to a new plant. All three identifiers occur in the synthetic history, and the score rule is not learned. A time holdout on familiar assets would answer a different question from transfer to unseen equipment. Before a real proposal, its evaluation design would need to match the intended future use, sources and maintenance context. Do not claim that a clean time split proves cross-site performance or removes all dependence between overlapping windows.

Learn in your browser · Chapter 04 of 06

Window recall, event coverage and review workload use different units. Report all three before interpreting an apparently excellent result.

Apply the fixed threshold 70 to the32 eligible windows. A score equal to 70 alerts; this is a greater-than-or-equal comparison. The later labels identify10 positive windows and 22 negative windows. Eleven eligible windows alert. Seven are positive and four are negative, so TP=7, FP=4, FN=3 and TN=18. Check7+4+3+18=32 and 7+3=10. Missing or cross-reset windows are absent from this matrix and remain visible in the coverage report. They are not true negatives.

Scroll sideways to see every column.

Repeated alerts for one event can make the queue seem busier without detecting another event. Tern therefore groups adjacent alerted days on the same asset and within one epoch into a review episode. A gap in alert days starts another episode; different assets never merge. The grouping is an invented workload convention. It must not conceal individual source rows or imply that any outstanding work has been completed. Keep both row and episode counts in the report.

Scroll sideways to see every column.

Eleven alert rows become seven episodes, of which three have no supplied target event. The false-episode fraction is 3/7=42.86 percent. It is not a false-alarm rate per operating hour: the fixture contains no verified operating-hour exposure. Four episodes each anticipate one of four declared evaluation events. Event coverage is 4/4=100 percent, even though window recall is 70 percent. There is no contradiction. Several missed positive windows still relate to an event that another earlier alert covered.

Warning time uses the first qualifying alerted anchor inside the three-day target window for each event. E1 occurs on19 and first alerts on16, giving3 days. E2 occurs28 and first qualifies25, also3. E3 occurs22 and first qualifies19, also3. E4 occurs25 and first qualifies23, giving2. Mean warning time across these four detected events is 11/4=2.75 days; median3 and minimum2. These are retrospective intervals, not promised intervention time. They exclude undetected events from the lead-time average, so report the event denominator and misses alongside them.

Raising the illustrative threshold to 90 yields four alert rows: P1/16, P1/26, P2/23 and P3/23. The matrix becomes TP3, FP1, FN7, TN21. Window precision rises to 75 percent, recall falls to 30 percent, and only three of four events are covered. E3 is missed. The detected lead times are 3,2 and 2 days, averaging7/3, about 2.33 days. This is a sensitivity comparison, not permission to choose a threshold on the evaluation set and call its performance independently tested. Any choice informed by these outcomes needs a new evaluation design.

1. Show the nominal and eligible window denominators.
2. Reproduce all four confusion-matrix cells.
3. Group alerts using a stated episode rule and retain the original days.
4. Count every declared evaluation event, including those with no eligible warning.
5. Calculate first qualifying warning time only for detected events and report misses separately.
6. Discuss workload and uncertainty before making a bounded recommendation.

Learn in your browser · Chapter 05 of 06

A review request must preserve what is known now, what is unknown and who owns the separate maintenance decision.

Use P2 at day 19 as the prospective handover case. Days 17,18,19 contain condition indices37,38,39, giving mean38 within epoch 0. The current supplied score is 80, above threshold 70. At that moment the later E3 event on day 22 is not known; its label becomes available on day 24. A retrospective reviewer can later count a three-day warning, but the original handover cannot cite that future result as contemporaneous evidence. Store the two records with their different purposes and times.

The structured draft records packet version tern-v1, P2, anchor 19, source days 17-19, score 80, threshold 70, draft-for-human-review status and responsible-person-review as the next step. It states changes_existing_plan=false. This does not authenticate the author or reviewer. It simply lets another learner compare a structured draft with the supplied source. The actual human must read the complete free text, ask for missing evidence and make any real decision through the responsible organisation's process.

Scroll sideways to see every column.

The useful handover asks the responsible person to review the observations and determine whether further authorised action is appropriate. It names the source edition, known quality holds, intended audience, unresolved questions, existing-plan boundary and response owner. It does not prescribe an inspection technique. The HSE maintenance guidance emphasises competent people, suitable information and planned arrangements. That supports keeping real maintenance decisions with the responsible competent process; it does not turn this workbook into a maintenance method or an assessment of any real site.

Human review needs time. Seven review episodes at an invented18 minutes each require126 minutes. The available 90-minute analyst window fits five complete reviews, leaving two episodes unserved and 36 minutes of demand. Eleven separate alert rows would consume198 minutes if reviewed independently under the same assumption. Grouping saves72 review minutes in this arithmetic, but that does not prove a real productivity gain or make two pending episodes disappear. A queue can be precise and still exceed its review capacity.

Do not allocate scarce review slots by using future labels from the evaluation. On the day of the alert, the desk does not know which episodes will later match an event. A real priority rule would require authorised, current context and its own evidence. Here the learner reports the capacity gap and requests a human decision; the exercise deliberately supplies no live prioritisation rule. A missed response or absent reviewer must not be converted into automatic permission, an automatic negative label or cancellation of existing maintenance.

Test 32 changes only the free_text field to a false diagnosis and a fabricated work-completion claim. The structured checker still returns a match because it does not understand arbitrary prose. This is a designed negative capability demonstration, not an approval. Read that output together with its limitation. A field-level pass can coexist with a dangerous sentence. The learning evidence is your identification and repair of that sentence, supported by the source, not the fact that you clicked a practice control or obtained exit code0.

Learn in your browser · Chapter 06 of 06

A metric can improve because difficult evidence disappeared. A defensible recommendation keeps excluded data, missed events and unresolved human work visible.

Repeat the analysis after removing every P3 reading from a copy. Do not remove E4 from the declared evaluation event list. The nominal39 windows remain, but only 22 are eligible: P1 contributes12 and P2 contributes10. Coverage falls from 32/39=82.05 percent to 22/39=56.41 percent. Seventeen windows are now held, or43.59 percent. The missing asset has not become reliable, silent or safe. Its evidence is unavailable, so the report must retain its identity and missing coverage.

The remaining eligible matrix is TP6, FP3, FN1, TN12. Window recall rises from 70 percent to 6/7, about 85.71 percent. Yet event coverage falls from 4/4 to 3/4 because E4 has no eligible warning. This is the course's central monitoring trap: a better conditional metric can accompany worse coverage of the intended population. Neither result alone can justify expansion. Report the missing asset and both denominators before discussing the model. Do not silently remove an undetectable event from the event denominator.

Scroll sideways to see every column.

A changing result may arise from a new score distribution, sensor availability, operating context, maintenance reset conventions, label delay or the outcome population. Do not call every change model drift without examining the source. Track asset-level coverage, stale-data holds, missing events, input versions, review latency and pending work. A frozen threshold can still produce a different queue when the input process changes. Conversely, editing a threshold without recording its version makes comparisons uninterpretable even when the headline percentage stays the same.

The fallback in this educational design is a human-controlled review route under the existing maintenance arrangements. Missing AI or sensor output supplies no permission to extend an interval, restart equipment, omit a task or assume normal condition. The responsible organisation must define its own response to unavailable evidence. The course provides no operational response sequence. In the paper packet, record the hold, affected asset, last trustworthy source edition, pending review owner and condition for resuming the comparison. Do not substitute a copied last value for a fresh observation merely to keep the dashboard green.

Feedback is also time-bound evidence. A person marking a review as complete does not establish an event outcome. A replacement record does not automatically mean a predicted failure was prevented. Actual interventions can change what is observed, making counterfactual claims harder. The synthetic packet has no intervention trial, control group or causal model. It supports reproducible counting under stated assumptions, not an estimate of avoided failures, life extension, reduced downtime or financial return. Keep those benefits as untested hypotheses if a future proposal needs them.

1. Freeze the input, policy, outcome-completeness declaration and report versions.
2. Compare coverage by asset before comparing conditional accuracy.
3. Retain every declared evaluation event, including unobservable ones.
4. Separate data-quality changes, model changes and human queue changes.
5. Record fallback ownership and unresolved review without inventing an operating procedure.
6. Use a new, suitable evaluation to assess any revised policy; retain the old evidence.

The final recommendation should be modest and specific. Tern's threshold 70 finds all four supplied events, but excludes seven windows and creates126 minutes of review demand against 90 available. The labels and scores are invented, the event count is tiny and no model is trained. Continue only as an educational desk study. Repair the data-quality and capacity assumptions before designing another controlled evaluation. The packet cannot establish site readiness, professional competence, engineering fitness or statutory compliance.

Mickai Training Academy is not an accredited training provider. This course has 40 scored knowledge questions and six ungraded written practices. A knowledge sitting draws 20 questions with a 45-minute limit and requires strictly more than 80 percent: 16 correct fails and 17 passes. The raw score is not rounded into a pass. Completing the workbook, running tests or writing a capstone does not itself pass the knowledge assessment or verify practical competence. Check the published course and assessment pages for current examination and certificate availability. This workbook supplies no equipment service.