“The peak press force was within specification, yet assembly later revealed deformation or a loose fit.” When this happens, a single peak value cannot explain how force developed throughout the stroke. A manufacturer considering AI anomaly detection for press force waveforms should first decide which signal to capture, how to link each stroke to its part, die and material lot, and who can release a held product. This guide covers measurement, model evaluation, the quality decision, RFP requirements and FAT/SAT for press fitting, crimping and forming lines in Thailand.
Scope: a single stroke, not a die-life forecast
The unit of analysis here is one press stroke and its force-versus-displacement or force-versus-time curve. An unusual early contact, steep rise, mid-stroke drop or unstable end point becomes a candidate for inspection. A score does not prove that the part is defective or name the root cause. Quality assurance confirms disposition using the actual part, dimensions, other inspections and process history.
Traceability records which machine and die made a part and when. That is essential context, but a history record alone does not evaluate a waveform. Die-life prediction addresses wear and replacement over many strokes; see our guide to die-life prediction AI. This article evaluates the behavior of one stroke and does not predict remaining tool life. If the equipment itself has not yet been selected, start with our servo press and press-fit selection guide. A waveform model cannot replace press rating, fixture stiffness, existing inspection or machine safety functions.
Why peak force cannot describe the whole press process
Two strokes can reach the same maximum force by different paths. In press fitting, the first contact position, the section where friction rises, and final seating may differ. In crimping, tool contact, material deformation and the final squeeze can vary. Kistler describes acquiring force and displacement with sensors and evaluating their relationship as a process curve. A burster equipment manual shows a window criterion for monitoring, including an example around initial contact. These are vendor-specific examples of what a curve can reveal; they do not show that AI will detect every fault.
Imagine two strokes that each peak at 20 kN. One rises smoothly and ends at the intended position. The other reaches 17 kN too early, drops, and rises again. Misalignment, debris or fixture movement could be investigated, but material variation and a shifted displacement zero can produce similar signals. The interface and work instruction must say plainly that an anomaly score is neither a defect verdict nor a causal diagnosis.

Displacement axis, time axis, or both?
Force versus displacement makes the load at a physical position visible. It is useful where contact position or final seating matters. Force versus time can also be necessary where dwell, speed or hydraulic response matters. If speed changes while only a time trace is available, the same physical event appears at a different time. Before a time-aligned model flags it, check speed, cycle definition and trigger timing.
If the equipment permits it, retain force, position, time and start/end events under the same stroke ID. Discarding one raw axis makes later investigation of speed variation versus position shift harder. Choose the sampling rate from the equipment response, the shortest event that matters and the capacity to transmit and store the signal. No universal hertz value fits every press; specify the measurement requirement in the RFP instead of copying an arbitrary number.
Establish the measurement system first
Check force sensor range, resolution and calibration; displacement mounting; frame and fixture deflection; and trigger repeatability. If repeated runs of known good parts vary greatly under the same conditions, AI may learn instrument noise as a process anomaly. AIAG’s Measurement Systems Analysis errata warns that a bias study can mislead if repeatability is unacceptable. Compare measurement variation with both the process tolerance and the waveform difference you intend to detect.
Treat sensor replacement, die reinstallation and fixture adjustment as changes to the data condition. Mixing their curves into a single unlabelled baseline can create false alarms or hide real shifts. The first operational dashboard should show missing traces, context join rate, sensor saturation, trigger delays and zero shifts alongside any AI scores.
Why part, die and material-lot context matters
A common “normal” curve becomes too wide when part A and part B have different force requirements. Even within one part number, die ID and repair history, material lot, thickness, heat treatment, lubrication, fixture, machine, recipe and shift can change the curve. NIST’s industrial AI work discusses equipment, design, execution, part quality and human feedback as connected manufacturing data. Applied to a press line, the practical rule is to compare strokes under conditions that are physically comparable.
The minimal join fields are stroke ID, timestamp, machine ID, part number, recipe version, die ID, part or tray/lot ID and final quality disposition. A multi-cavity operation may produce several parts per stroke. In that case, map cavities to parts or trays explicitly. A waveform with the previous die ID after a changeover looks complete in a database but is not trustworthy evidence.
Define join rate as the share of all strokes that can be linked to required context and the quality result within the agreed time. Do not quietly remove missing strokes from the denominator. Separate trials, rework, setup, communication loss and delayed quality labels. When a destructive test returns days later, retain both the “unknown at inference time” state and the later adjudication timestamp. Overwriting labels destroys the audit trail.

Avoid both excessive splitting and excessive pooling
A dedicated model for every part × die × lot × shift cell may have too little data. Pooling everything can bury the very difference that matters. Begin with physically significant distinctions such as part, operation and recipe. Use die and lot either as monitoring strata or separate baselines after looking at real variation. Record when a new baseline is needed and what happens for an unseen combination.
Do not train on data later than the test period. Strokes from the same lot are correlated; a random stroke split often makes results look better than deployment. Include later lots, another shift, die replacement and restart after a stop in independent evaluation. If data are limited, a documented “hold on unknown condition” is more defensible than a sweeping accuracy claim.
Compare explicit process rules with AI on the same data
Keep established peak limits, position windows, final force limits and slope rules. If a process monitoring device already catches the relevant defects with acceptable workload, the case for AI is weak. First list defects or process deviations missed by current rules but verified by additional inspection. Then compare three approaches on identical data: (1) explicit peak/window rules, (2) distance to a reference curve or section features, and (3) an anomaly model.
Use the same chronological split and quality labels. Report missed critical defects and excessive holds by condition, not just overall accuracy. More complex models add maintenance and explanation costs. Adopt AI when it brings a practical gain on later lots and dies without an unacceptable burden. Features such as initial contact position, section slope, a local force drop, area under the curve, final force and dwell stability are useful starting points because process engineers can interpret them. Yet none is automatically a root cause. Show the suspect section, reference population, model/threshold version and missing-input state with the score.
A percentage score does not measure the workload
Ten alerts in a row and ten alerts spread across a month have different effects on production. Measure escapes, false holds, alerts per shift, handling time and work-in-process under hold. Break results down by defect, part, die, lot and shift where possible. “99% accuracy” in a mostly-good population does not describe detection of a rare critical fault.
For a workload illustration, 1,000 strokes per day at a 2% hold rate means 20 reviews per day. At five minutes each, that is 100 minutes of review before isolation, remeasurement or investigation. This is arithmetic, not a forecast of benefit. Set an acceptable hold load and staffing capacity first. Do not lower a threshold to suppress alerts at the cost of critical escapes without quality approval.
Who decides what happens to a part after an AI alert?
Define at least three states: normal candidate, review required, and measurement unavailable. A normal candidate still follows mandatory inspection and customer requirements. A review state identifies and holds the affected stroke, part and container to an agreed boundary. An unavailable measurement must never silently turn into pass. Depending on product risk and the existing control plan, fallback may be conventional inspection, full hold or process stop.
AI proposes that a curve section differs from its reference group. Quality assurance checks the part, dimensions, images, further tests and die/material history, then decides release, retest, rework, scrap or wider containment. Production checks the operating condition; maintenance checks sensors, fixtures and the die. A low score alone cannot release a held item, and a high score alone should not be recorded as a confirmed defect.
The NIST AI Risk Management Framework calls for defined human-AI roles, oversight and operational monitoring. Put approvers, substitutes, escalation, release conditions and audit logs into the interface specification. Store the item ID, curve comparison, AI version, hypothesis, QA decision, decision maker, timestamp and supporting evidence. Keep overrides for label review rather than automatically feeding them into retraining. NIST’s 2026 report on deployed AI monitoring also identifies the difficulty of ongoing monitoring and human-AI feedback.

Keep quality decisions separate from machine safety
A protective stop for dangerous force belongs to the applicable machine safety design and validation. A quality anomaly score is not a substitute for an emergency stop or safety interlock. ISO/TR 22100-5:2021 addresses safety implications of AI/ML used within specified limits in machinery, but explicitly excludes AI-based safety systems from its scope. This guide describes product hold and investigation. If a stop output is proposed, the machine safety owner must separately assess applicable standards, risk, timing and fault behavior.
What to put in an RFP for press-force waveform AI
“Detect anomalies with AI” cannot be priced or accepted. Define the process boundary, measurement, joins, decision rights and evidence. Agree numerical limits from plant measurements and risk rather than copying a supplier default.
| Requirement | What to specify | Acceptance evidence |
|---|---|---|
| Scope | Machines, operations, parts, dies, lots and stroke volumes | Scope list and production logs |
| Measurement | Force, position, time, range, resolution, calibration, triggers and missing data | Specifications, calibration and sample traces |
| Join | Stroke to part, die, lot and quality disposition | Traceable examples and join rate |
| Decision | Existing rules, AI output, hold, unavailable and unknown conditions | State transitions and screen captures |
| Evaluation | Misses by defect, false holds, alerts/shift and latency | Stratified results and timestamp logs |
| Authority | QA release, retest, override, change and rollback | Roles, access tests and audit trail |
| Integration | PLC/MES/QMS, outage, retry, duplicate and clock sync | Injected-fault tests and reconciliation |
| Operations | Retention, export, model version and retraining approval | Procedures and restore test |
Request export of raw and processed curves, decisions, final disposition and model/threshold versions in an associated format. Proprietary-only files can make later quality investigations and migration difficult. Where a Thai plant and Japanese headquarters share data, settle access, location, retention and customer-contract restrictions at quotation time.
Avoid an acceptance clause that says only “95% model accuracy.” It omits defect type, sample count, condition and allowed misses. If rare-defect evidence is insufficient, state that limit and define hold operations and further sampling. Every threshold change during tuning requires evaluation on subsequent independent data.
FAT: reproduce acquisition, decision and hold
At factory acceptance testing, use the intended sensors, acquisition device, edge computer and PLC/MES/QMS interfaces as closely as practical. Run known good parts and safely obtained abnormal samples. Trace acquisition, joining, inference, hold instruction, operator display and logging. Agree on simulated waveforms or safe test pieces when deliberately reproducing a fault would damage a die or part.
| FAT case | Input | Evidence of pass |
|---|---|---|
| F01 | Resend the same stroke ID | No duplicate decision or hold; retry logged |
| F02 | Lose or delay the start trigger | Mark measurement unavailable; never pass |
| F03 | Saturate the force sensor | Do not use the trace as valid quality evidence |
| F04 | Omit part or die ID | Do not choose the wrong reference group; hold |
| F05 | Supply a curve with a known abnormal section | Show section and model version |
| F06 | Disconnect and resend after restoration | Reconcile order and duplicate dispositions |
| F07 | Try release with operator rather than QA rights | Deny and log unauthorized action |
| F08 | Change model/threshold version | Demonstrate approval, regression and rollback |
Record instruments, sample counts, clock synchronization, test traces, expected and actual results, and open items. Vibration, real material lots, changeovers and night-shift instructions that FAT cannot reproduce belong in SAT with an owner and due date. Passing FAT does not prove production quality performance.
SAT: test the actual Thai production conditions
Run shadow mode across real parts, dies, material lots and shifts while current inspection determines disposition. Count missing and unjoined strokes in the denominator and report them separately. Where a suspect part exists, inspect it and classify whether the cause is a confirmed defect, a normal process difference or a measurement issue.
Test first-off trials after changeover, stop/restart, sensor replacement, die cleaning and material-lot change. Check that an unknown part does not inherit a known part’s “normal” result. Observe whether Thai alert text and instructions tell night-shift personnel exactly what to hold and whom to call. Include offline trace buffering, duplicate removal after recovery, clock order and restoration of the unresolved-hold list.
SAT completion combines data availability, agreed evaluation of critical anomalies, manageable hold load, working fallback for unavailable measurement, and QA control of releases. If a rare defect never occurs during SAT, do not claim that it was validated. Hand over a documented in-scope and untested-condition list.
Cost, return and staged investment
The budget includes more than an AI licence: load and position sensing, acquisition, PLC interface, stroke IDs, MES/QMS integration, raw-trace storage, labels, part review, training, monitoring and change control. Existing monitors may be reusable, but verify data granularity, timestamps and licence rights. Include line downtime and sample preparation.
Separate possible benefits into avoided escapes, reduced sorting and rework, shorter investigation and better setup control. An alert itself does not save money; the subsequent action must change an outcome. If extra review takes 40 hours per month and sorting falls by 20, the direct labor balance is negative. A critical escape may still justify investment, but probability and loss estimates need evidence. Do not print unsupported claims such as “AI improves quality by 30%.”
A staged order can start with one machine, one part and one defect mode, joining traces to actual QA disposition and comparing rules with AI. Expand to other dies, lots and shifts only after validating hold workload and update procedures. A successful small proof of concept should not be extrapolated mechanically to every line.
Monitoring and change control after deployment
Watch missing traces, failed joins, unresolved-hold time, QA overrides, confirmed defects and changes to dies, material and recipes alongside alert rates. Do not automatically retrain at the first data shift. Document the reason, training scope, exclusions, comparison with the old version, approval and rollback. Regression-test old known defects when adding a new normal condition.
Review signals with maintenance and QA. An operator’s observation that “this alert appears after setup” or “the trace changed after die cleaning” is a hypothesis to test, not noise to delete. Before suppressing alerts, replay past defects under the proposed setting. Preserve representative curves, borderline cases and disposition reasons so that new staff can understand the boundary.
When a Japanese headquarters approves model changes for a Thai plant, both sides need to see who inspected the local evidence, accepted the quality risk and deployed which version when. Restore a prior version and reconcile any holds created during the transition as part of the deliverable.
Define trace and quality-label data before training
Agree on what one record means. A “stroke” might be a single downward movement, a complete pressing sequence for one part, or a cycle producing several parts. If a PLC sequence number differs from the MES production ID, define the mapping and what happens when the PLC counter resets after a restart. Document shift-boundary dates, PLC/server clock differences and when a manually entered part number becomes valid. Joining a curve to an inspection result merely because their timestamps are close is especially risky on short-cycle lines.
Classify trace states as valid, missing, truncated, sensor-saturated, unknown trigger and unknown zero. Do not merely exclude invalid traces from training: define when each becomes “measurement unavailable” in production. A trace that ends halfway cannot demonstrate a normal final section. If a die has two stages, cross-check start/end events and stage count before marking a stroke complete. FAT must reject any display or API response that interprets missing data as “no anomaly.”
Quality labels need more than OK/NG. Separate confirmed physical defects, inspection pending, dimensions passed but appearance not yet checked, accepted after rework and customer-approved concessions. QA should define which states are positive or negative training examples. If a reworked part’s final pass is attached to its original press trace as a normal example, evaluation becomes distorted. A part scrapped in process may never appear in the final-inspection database; do not treat that absence as “no final defect.” Retain its stop reason and physical disposition.
Keep defect observations separate from causal hypotheses. “Insufficient press-fit depth” is an observed result; “poor lubrication” is a possible cause. Die wear, dimensional variation, debris or sensor shift can lead to a similar result. Automatically writing “die wear” into a cause field when AI first alerts can bias later investigation. Store what was known at inference time, what inspection later found, and what a root-cause review confirmed, each with its timestamp and owner. When a label changes, retain the previous value and the evidence for the revision.
Collect boundary cases as well as ordinary good parts
Include legitimate edge conditions: material lots at the normal range boundary, temperature changes, first-off production, fixture cleaning and die service. For historical rejects, check that trace, inspection record and equipment state truly describe the same part. A historical curve with no surviving sample can illustrate an unusual shape, but it is weaker evidence that the system detects a defect. Do not damage tools to manufacture rare failures. Simulated traces can validate interfaces and screens, while detection performance on real defects remains untested.
Review cases jointly with QA and production: good parts with extreme curves, defective parts whose curves look ordinary, and sensor faults that appear anomalous. A defective part with a normal-looking force trace may be invisible to that sensor placement and measured quantity. A larger model might not solve it; vision, dimensional or electrical testing may be needed. The RFP should name defects that will not be decided from the force curve as clearly as those in scope.
Stage gates from one machine to production
First, limit the pilot to one machine and part number and establish reliable acquisition and context. The exit criterion at this stage is not AI accuracy: stroke IDs must be unique, links to part/die/lot and disposition must work, and missing traces must be countable. Locate the gap at the sensor, PLC, MES or inspection database rather than predicting away unavailable information. Confirm that acquiring the trace does not alter takt time or the PLC safety logic.
Second, freeze the current threshold/window baseline and run the AI candidate in shadow mode. For every stroke, compare the existing rule, AI suggestion and physical inspection. Investigate AI alerts that current rules passed and the reverse. If inspection cannot resolve a case, label it undecided instead of counting it conveniently as a success or failure. Keep tuning and final evaluation periods separate and report untested conditions with the result.
Third, connect AI proposals to a real hold only for the scope QA approves. Verify that the notice arrives before the affected item leaves the controllable area and that container changes or rework do not break ID tracking. A warning on a screen is not containment if the part cannot be located. Require inspection evidence and QA approval to release a hold, and hand unresolved holds to the next shift. Machine-stop and product-hold permissions may need different owners.
Finally, add another die, material lot, shift and post-maintenance condition one at a time. Show the difference against the previous baseline. If a new condition requires a looser threshold, replay known old defects. Even an identical model of press may differ in sensor position, fixture stiffness, control period and recipe. Do not copy a model or average trace across lines without calibration and comparison tests. At each gate, record “proceed,” “proceed with restricted scope” or “return to measurement.”
Set ownership before acceptance. Maintenance owns calibration and replacement; IT owns connectivity and storage; production owns setup and physical isolation; QA owns inspection results and final disposition. A workflow routed only to an AI specialist will stall at night or on holidays. Demonstrate deputies, review frequency and escalation of unresolved holds during the trial. Monthly review should examine potential escapes, unavailable measurements, reasons for release and changed conditions as well as alert counts. If a defect proves invisible to the signal, revisit sensor placement or inspection method rather than only moving the AI threshold.
FAQ
How is press-force waveform AI different from peak-force monitoring?
Peak monitoring compares a small set of values with limits. Curve monitoring considers initial contact, the rise, local force loss and the final section. AI can assist with comparisons across conditions, but its added value should be tested against existing windows and thresholds on the same data.
Can we deploy with force-versus-time only?
Yes, within a defined scope. Speed and trigger changes move the same physical event to a different time. If contact position matters, add displacement; otherwise retain speed conditions and limit the validated operating envelope. Do not diagnose a defect from trace shape alone.
What if we have mostly good parts and very few defects?
An anomaly model can flag departures from known good traces. It cannot substantiate a detection-rate guarantee for rare defects that were never tested. Use safe simulations, historical traces and part inspection where appropriate; disclose untested conditions and keep the hold workflow.
What determines the cost of press-force waveform AI?
Machine and part scope, available sensors and raw traces, linkage to quality results, storage volume, review workload and change-control requirements. Compare supplier offers against the same measurement, integration and operations scope.
Should an AI anomaly score automatically stop the press?
Quality holds and machine safety stops are different designs. A safety owner must specify the standards, signal timing, fault behavior and validation for any stop function. A quality score alone does not replace a protective function.
Conclusion
The useful starting point is repeatable force and displacement measurement, a trustworthy join from stroke to part/die/lot, a fair comparison with existing rules, a QA-owned hold decision and reevaluation when conditions change. Defining this as a single-stroke anomaly use case keeps the RFP and FAT/SAT distinct from traceability-only projects and long-term die-life prediction.
Even at the stage of selecting a Thai pilot line, contact TOMAS TECH to review the sensors, traces, quality records and current hold procedure you already have. A machine, part number and troublesome defect mode are enough to begin scoping the work.
Sources
- Kistler, Force-displacement measurement — vendor description of sensing and curve evaluation.
- burster, DIGIFORCE 9311 operating manual — a specific window-monitoring example.
- AIAG, Measurement Systems Analysis errata — repeatability and bias-study caveat.
- NIST, Industrial Artificial Intelligence Management and Metrology — manufacturing data and application context.
- NIST, AI Risk Management Framework Core — human oversight and ongoing monitoring.
- NIST, Challenges to the Monitoring of Deployed AI Systems (2026) — post-deployment monitoring challenges.
- ISO, ISO/TR 22100-5:2021 — scope of AI/ML safety considerations in machinery.