Screw Tightening Torque Curve AI Anomaly Detection: Implementation Guide

A tightening controller can report OK while a later inspection finds a seating issue or loose joint. To implement AI anomaly detection for screw tightening torque curves, connect the full torque–angle or torque–time trace to the product, joint position and tightening program. This guide explains how to add AI process monitoring without removing the controller’s existing OK/NG decision. It covers data capture, verified labels, false accepts and false rejects, intervention routing, an RFP and FAT/SAT. For equipment choice, see our nutrunner selection guide for Thailand; for screw feeding and cell design, see our screw fastening automation guide.

Define the quality problem before choosing a torque curve AI model

The aim is not to declare every defect automatically. A controller checks configured torque, angle and process limits. AI can add value by ranking traces that look unusual even when those configured limits pass: a changed rundown, seating point or slope; a pause and restart; or a possible retightening. Start with one or two observed failure modes that could escape the normal controller decision. Without a defined quality question, there is no defensible ground truth for an AI score.

Cross-threading, screw misalignment, debris under a seating face, floating parts and retightening can affect traces in different ways depending on the joint and strategy. A curve alone seldom proves the physical cause. In the first release, use “review this joint” rather than a definitive defect name. Route the product to inspection, document the physical finding and disposition, and use that evidence to improve the model. An AI score starts an investigation; it does not conclude quality assurance.

Scope the process as carefully as the model. Product family, fastener, joint stiffness, tightening sequence, tool type and program revision can create legitimate differences. A suitable pilot station already exports traces, links each tightening to a product and position, and can hold suspected products. A station that stores attractive charts but cannot locate the corresponding workpiece cannot deliver a usable containment process.

Controller OK/NG and AI have different jobs

Keep the manufacturer’s controller program responsible for its normal control. Controller NG follows the existing stop or quarantine rule. For controller OK plus high AI risk, quality decides whether to stop, hold or sample a product. Low risk still needs a stored result and model version for later tracing. Record the suspected pattern separately from the physical inspection and final disposition. This makes it possible to evaluate whether a changed intervention rule actually improved the process.

Screw Tightening Torque Curve AI Anomaly Detection: Implementation Guide - figure 1

What a torque–angle trace can and cannot tell you

A trace may show rundown, contact or snug, a load-building phase and a final point. Its shape depends on joint stiffness, fastener, lubrication, speed, tool strategy and measurement location. Two joints can finish at the same torque but reach seating at different angles. That difference can guide an investigation into initial position, stack-up or interference; it cannot by itself prove a root cause.

The OPC Foundation’s Joining Systems specification defines result values with measured quantity, evaluation, engineering units, sensor identification, limits and tags such as FINAL and SNUG/SEATING. Its conformance units also describe traces linked to results. Consequently, exporting only final torque values is not the same as obtaining the sample series needed for curve analysis. Check the actual tool’s supported fields and sampling settings; “OPC UA compatible” is not a promise that every optional result and trace feature is implemented. See ResultValueDataType and Conformance Units.

Atlas Copco explains why the curve carries information beyond a final value. Yet torque is not a direct measurement of clamp force. Friction at the thread and bearing surface, material and joint geometry can change preload at the same torque. AI that finds an unusual trace does not thereby certify clamp force. Critical joints still need independent audit or other tests selected by design and quality engineers. Atlas Copco also discusses why a torque wrench reading alone does not guarantee joint quality.

Select features from a process hypothesis

Candidate signals include final torque and angle, estimated seating point, segment slopes, time to target, trace peaks, interruption and restart, and a possible second pass over the same joint. These are not universal defect formulas. A steep slope, for example, could be normal for a hard joint. Gather verified good traces by condition and compare them with physically reviewed abnormal cases. Keep raw traces alongside calculated features so that disputed warnings and future model revisions can be investigated.

Sampling matters. If a short seating event is lost by low-rate acquisition, a more complex model cannot restore it. Ask suppliers to specify trace axes, units, sample count and interval, missing-value convention, clock synchronization, export method, retention and firmware changes that alter the data format.

Design torque curve capture around one tightening event

The unit of analysis is a tightening event, not an image of a graph. Link result ID, product or workpiece ID, joint position, tool and controller IDs, program ID and revision, timestamps, controller OK/NG, final torque and angle, trace and units. Retightening and repair must create linked new events, not overwrite the first event. Otherwise, an investigation can confuse the original assembly with a corrective attempt.

The OPC UA Joining Systems model discusses result metadata, associated tool and controller identifiers, external product or joint identifiers, traces and intervention results. It is a useful interface reference, but confirm the required fields on the actual equipment. The specification separates capabilities into conformance units; supported subsets differ. Put a field-level list in the RFP and inspect sample messages and a live connection. See the OPC Foundation conformance units.

For a path from tool to controller, edge gateway, MES or quality database, and analytics, preserve event and trace IDs at every boundary. Buffer during network outages, then deduplicate resent events. Synchronize clocks and record time zones rather than trying to repair timestamps later. Missing traces must yield “undecidable,” not an assumed normal result.

Screw Tightening Torque Curve AI Anomaly Detection: Implementation Guide - figure 2

Label what was physically checked

Do not copy controller OK/NG into the training label and call it ground truth. That would train the system mainly to reproduce the controller while missing the stated objective: finding questionable cases inside controller OK. Use evidence from product inspection, independent measurement, teardown, repair or part replacement to distinguish verified normal, review needed, confirmed anomaly and undecidable. Do not force a held product with unknown cause into a specific defect class.

Keep reviewer, date, inspection method, references to photos or measurements, and relevant conditions with each label. If two reviewers disagree, settle the inspection rule before training. True defects may be rare; do not assume abundant examples in every class. Ranking departures from a verified normal population, followed by quality review, may be an appropriate starting point. A claim to detect a named defect still needs independently verified examples and evaluation.

Split evaluation data to prevent leakage

Randomly splitting adjacent operations from one product, lot or tool between training and evaluation can make performance look better than deployment. Reserve a later time period and, where feasible, separate tool or product-family tests. A new program, bit or lubricant needs fresh evaluation before the old model is trusted. Record exclusions and split rules so FAT and SAT findings can be reproduced.

Route AI findings into a workable quality process

A useful system is defined by action, not just a score display. A basic route is: controller NG to immediate hold; missing communication or trace to undecidable; controller OK with high AI risk to a quality review queue; controller OK with low risk to normal flow. If a high-risk AI result automatically stops a line, approve the stop, restart authority and false-stop recovery process. Shadow mode can store AI results without changing product flow while the factory measures warning burden.

Decide who inspects, where held products go and how shipping is blocked until disposition. Record release, retightening, teardown, replacement or scrap as distinct outcomes. This feedback becomes evaluation evidence. Risk levels may differ between critical and ordinary joints, but product design and customer requirements govern the actual quality rule.

Measure false accepts and false rejects separately

One “accuracy” number is inadequate. A false accept is a later-confirmed abnormal joint that remained in controller OK and AI low risk; it threatens escape. A false reject is a verified normal joint that AI sent to unnecessary review; it consumes inspection and line capacity. Measure detection by target failure mode, warnings per thousand good tightenings, undecidable rate, trace-to-product linking rate and time from warning to completed review. Agree numeric targets according to failure severity and available staff rather than promising a generic “99% AI accuracy.”

A lower alert threshold usually increases candidates and inspection load; a higher one may miss more. In acceptance testing, show several threshold options with associated escapes and warning counts. Select an operating point with quality and production owners. Compare future model versions on the same frozen evaluation set, including threshold settings.

Keep four assurance domains separate

First, process monitoring observes controller results and traces; AI is primarily here. Second, independent joint audit checks preload, loosening, function or durability by methods chosen for the joint requirement. A trace is not a direct certificate for these properties. Third, tool performance and calibration maintain confidence in the measuring system. ISO 5393:2017 specifies a performance test method for rotary tools used with threaded fasteners; it is not an AI certification for each assembled joint. Fourth, AI operational quality covers detection, false alerts, missing data, model versions and retraining triggers.

When a defect is found, inspect tool condition and program, then joint material, assembly and lubrication. A high AI score alone does not order tool replacement; a low one does not waive physical inspection. Quality, manufacturing engineering and maintenance should own change approvals and preserve before-and-after traces.

Run a small, evidence-driven pilot

Choose one joint family and one tool system with a clear quality question. First collect data and audit missing events, incorrect product links, clock shifts and inconsistent units. Link controller results to physical inspection and write the labeling rule. Compare simple limits and segment features with more complex models; choose the method that repeats reliably and leads to a usable action.

In shadow mode, save AI decisions for quality review without issuing an immediate operator pass/fail. Measure warning count and review time, then tune priority and thresholds. Next, connect only important warnings to a hold queue. Enable automatic containment only after factory approval. Set a pilot period in advance, but judge completeness by coverage of normal product, lot and tool conditions and relevant abnormal examples, rather than elapsed weeks alone.

The pilot report should state what the factory can do with every type of result. An alert that cannot locate and hold its product has limited benefit. A modest classifier may still be valuable if it improves review order and creates reliable evidence. Quality, production, engineering, IT and maintenance should jointly decide whether to continue, change or stop the project.

Screw Tightening Torque Curve AI Anomaly Detection: Implementation Guide - figure 3

Specify data, operations and security in the RFP

List target tools and controllers, export interface, torque/angle/time series, sampling, final values, controller result, event ID, product ID and program revision. Request actual sample records. Ask which firmware and licenses determine available fields. For outputs, specify risk score, explanation shown to reviewers, undecidable result, model and threshold version, hold queue and decision history.

Define MES and quality-system integration, buffering and deduplication during outages, retention, backup, access rights and audit logs. If traces leave the factory for cloud analysis, settle data ownership, deletion and permitted retraining use before contracting. Acceptance must cover record completeness, alert-to-hold delay, false alerts during normal production, protection of controller NG if the AI service fails, undecidable behavior on missing data and rollback to an approved model version. Assign responsibility for line modifications, tool programs and quality rules.

FAT: test replay and failure behavior

Replay verified normal, abnormal and undecidable traces. Check expected results and resulting hold events. Test duplicate result IDs, missing samples, reordered traces, an unknown program revision and invalid units. Confirm that an AI server outage does not bypass the controller’s own quality or safety behavior. Record the source and verification of test curves, model version and thresholds. A reviewer must be able to identify the exact product and joint from an alert; visual appeal of the dashboard is secondary. Test permissions, correction history and audit-log export. If verified abnormal examples are insufficient, say so and carry an explicit additional collection condition into SAT.

SAT: test the actual line and its people

Verify live tool sampling, equipment clocks, product changeover, program revisions, network outage, retightening and operator hold actions. Measure whether good-part warning load is manageable and whether warned products are physically held. Ensure both original and rework events survive. Introduce representative faults only under factory safety rules and in a test setup that prevents shipment. SAT completion should include quality-approved rules, maintenance-approved recovery, IT-approved connection and backup, and an operator-approved standard procedure. A future tool, material or product change must trigger reassessment.

Estimate total cost and value

Separate trace-export license or tool upgrades, edge capture, MES connection, storage, physical labeling, model development, operator screens, training, support and model updates. Check trace availability early; no software can analyze samples a controller never exports. Calculate storage from actual sampling and tightening volume with the intended retention period.

For benefits, establish current inspection, rework, scrap, customer escape response and line-stop costs. Identify which of those the proposed intervention can affect. More warnings may temporarily raise inspection hours. Update the investment case with pilot evidence on both quality and workload. Critical joints may require a risk decision beyond a simple payback calculation.

Make the system usable across shifts and languages

A translated display alone does not define action. Write local work instructions for where an operator puts a held part, what a line leader receives and what quality checks. Make a displayed defect name explicitly provisional until physical review. Define who can stop or restart a line and what happens on night shifts when quality staff are unavailable.

Quality owns joint requirements, containment and independent audit. Manufacturing engineering owns program and product revisions. Maintenance owns tool and recovery. IT owns IDs, retention, access and backup. The AI owner tracks model versions and evaluation. Put the contact sequence and decision authority on one practical flowchart.

For mixed-model production, allow only validated combinations of product, joint, program revision and AI model. If the combination is unknown, return undecidable. Record reason, approver, start time and affected lot for emergency program changes. Review new-condition traces before reusing the old model decision.

Track daily warning count alongside time to disposition, verified causes, rework, trace loss, undecidable events and broken product links. A warning surge could come from fasteners, seating, tool bits, lubrication, supply or program change as well as model drift. A sudden fall can indicate collection failure, a threshold change or a product mix shift. Review useful versus unproductive warnings with the production team.

Preserve evidence for measurement and model updates

A data-quality report should give the period, tools, total tightening events, events with traces, events linked to product and joint, and duplicate, missing or unknown-unit counts. Performance measured on a clean subset alone can hide a weak data pipeline. Missingness may cluster by station or shift and deserves its own corrective action.

An evaluation report should specify training and test periods, product and joint scope, labeling evidence, targeted failures and cases left unverified. Give numerators and denominators with detection percentages; small classes have large uncertainty. Avoid accepting one average over distinct failure modes. Split false alerts by tool, product and shift to find concentrated workload.

Before approving an update, test old and new versions on the same frozen set and compare them in shadow mode on recent production. Quality should inspect cases found by only one version. Version the feature definition, threshold, eligible products, release time and rollback instruction. Retain every event’s model version even after rollback so a later quality investigation can identify affected lots.

Practical checks when connecting an existing tightening system

Identify which device holds the authoritative trace: the tool, controller, vendor software or supervisory system. A curve visible on a screen may not be exportable as samples. Before quoting, extract one real event and verify that its result ID matches the display. For periodic file export, check whether warning delay fits cycle time and the last possible containment point. Real-time interfaces still need tested behavior during network congestion and retries.

Define angle origin, direction and unit, torque unit, trace start and end, and sampling rules in a data dictionary. “Angle” might mean movement from the beginning or only after seating. The same column name can change meaning after a tool or firmware replacement. Shared models across plants particularly need this definition before pooling data.

Choose retention from quality traceability requirements, network capacity and search needs. Long-term raw traces, later compression and longer retention for suspicious events are possible choices. During a pilot, retaining only AI alerts prevents analysis of missed cases among controller-OK products. Keep an evaluable population of normal traces for at least the assessment period, and get quality approval before changing retention.

Where products have no serial number, a workpiece or pallet ID may support traceability. Time alone is unsafe when multiple products run concurrently. Test whether a reviewer can reach the exact physical product from an alert. If not, improve identification before enabling automatic containment. For products whose tightening sequence matters, preserve the order of events too. If a simple controller rule can reliably detect a sequence violation, use that rule and focus AI on less explicit curve changes.

An operator alert should show product, joint, hold location and next approver. Group multiple alerts on one product for handling, while preserving each event and disposition in the database. Retain original entries, corrections, author, reason and time so shipping decisions can be audited. Contracts should also separate the vendor’s model maintenance from the factory’s authority to change quality criteria or enable an automatic stop. A tool-program change should trigger an out-of-scope state until the model is reassessed. Assign incident support, interim inspection, retesting and deletion requests before go-live.

FAQ: torque curves, AI quality control and acceptance

How is torque–angle anomaly detection different from final torque limits?

Limits check configured final or step values. Curve analysis examines the path to those values, including unusual controller-OK cases. Detectable conditions depend on trace quality, the joint and verified examples. Keep existing controls in service.

Does tightening AI directly guarantee clamp force?

No. Torque and angle are process observations, not necessarily direct preload measurements. Friction and joint design matter. Use independent tests required by design and quality for critical joints; use AI to prioritize investigation.

What should be checked first when collecting torque curves?

Confirm the actual tool exports a trace and each trace links uniquely to a tightening event, product and joint. Check units, sampling, missing values, retry behavior and program revisions. A CSV of final torque alone lacks the series for curve analysis.

How do the RFP, FAT and SAT differ?

The RFP specifies fields, operations, owners and metrics. FAT checks replay and failure behavior in a controlled setting. SAT verifies live tools, product linking, alerts, containment, rework and recovery on the real line. Measure both detection and false-alert burden.

Can a pilot start with few abnormal labels?

Yes, as a ranked review of departures from verified normal behavior. Claiming reliable detection of a named failure requires verified examples and independent evaluation. Do not hide uncertainty behind a broad accuracy claim.

Conclusion

Torque curve AI is useful when full traces are linked to a specific product and joint, controller OK/NG remains intact, and suspicious parts can be inspected and contained. Separate process monitoring, independent joint audit, tool calibration and model assurance. Set acceptance around verified labels, false accepts, false rejects, missing data and actual operator action.

If you are assessing what your current tools can export, how MES can link each curve, or which failure to pilot first, contact TOMAS TECH. We can help define a focused pilot and its acceptance criteria from your current joint and inspection process.