Imagine your torque tool has been delivering readings that slowly but steadily creep toward the upper tolerance limit for the past three weeks. Every individual value still passes. No alarm. No intervention. Until the process tips over - and an entire batch has to be reworked.
That scenario is exactly the blind spot of the classic tolerance window. And it's the starting point for an approach that is gaining serious traction in manufacturing IT: data-driven anomaly detection in torque data.
The Problem with Fixed Limits
The logic of a rigid tolerance window is straightforward: if the measured value falls between the lower and upper limit, the fastening is OK. If it falls outside, it's not OK. That binary logic has its place - it's robust, standards-compliant, and sufficient for simple fastening applications.
But it has structural limitations that regularly cause problems in practice:
Gradual drift stays invisible. A process can shift continuously over hundreds of fastenings without ever violating a limit. The most common approach evaluates pass/fail solely based on torque and angle thresholds - the curve shape is not assessed. A plateau in the torque rise that indicates embedment relaxation goes undetected as long as the final torque value looks right.
False alarms from natural variation. Setting limits tight to catch problems early inevitably produces false alarms from process-compliant scatter. The clamp force resulting from the tightening torque is significantly influenced by fluctuating friction coefficients and the torque scatter of the tool itself - the combined effect of these varying friction influences can produce clamp force variation of 50% or more, even with high torque repeatability. Setting limits wide means missing real faults.
Context blindness. Fixed windows make no distinction between a material change, a batch change, or the onset of tool wear. They react to the final value - not to the pattern that precedes it.
The core problem: A tightening process can shift systematically over weeks without triggering a single limit violation. Fixed tolerance windows detect states — not trends.
How This Differs from Classical SPC
It's worth drawing a clear distinction here. Statistical Process Control (SPC) supports in-process monitoring by combining continuous measurement with statistical analysis methods, enabling early detection of critical parameter shifts and timely corrective action. SPC control charts such as Shewhart charts or CUSUM charts are an important step beyond pure tolerance windows - they detect trends and mean shifts.
But classical SPC also has its limits: it typically works with a single characteristic (e.g., the final shutoff torque), assumes a stationary process distribution, and requires manual configuration of control limits. Choosing the right rules carefully matters - too many rules and you get too many false alarms. Every rule carries its own false-alarm risk: given enough time, even a perfectly random process will eventually trigger a rule violation.
Data-driven anomaly detection goes further: it analyzes not just individual final values, but the entire torque curve - and it learns what is normal for this process, this tool, and this batch of components.
What Data-Driven Anomaly Detection Does Differently
Classical methods often look at only individual measurement points or final values. AI-assisted anomaly detection, by contrast, analyzes the complete torque curve - making subtle patterns, trends, and irregularities visible that would otherwise go unnoticed.
In practical terms for torque data, this means not just "Did the fastening reach 25 Nm?" but rather "What does the torque-angle curve look like compared to the last 500 fastenings? Has the slope changed? Is there an unusual plateau? Is the curve's center of gravity shifting systematically?"
Statistical Methods: The First Step
The simplest entry point into data-driven monitoring is statistical methods. Statistical methods form the foundation of anomaly detection. The Z-score measures how many standard deviations a data point lies from the mean - for normally distributed data, values beyond ±3 standard deviations are commonly flagged as anomalies.
For time-series data such as torque curves, drift and trend detection methods are also well suited: Change Point Detection identifies moments of significant distributional change using statistical tests such as CUSUM or Bayesian Change Point Detection. These methods are interpretable, computationally efficient, and can be integrated directly into existing analysis systems.
Unsupervised Learning: Modeling Normal Behavior
The next step is unsupervised machine learning. The key advantage: in industrial settings, you typically have many measurements from a correctly functioning system but very little fault data - because faults are rare. The algorithm learns a model of normal behavior and flags deviations from it.
Methods relevant to torque data:
| Method | Strength | Typical Use Case |
|---|---|---|
| Isolation Forest | Efficient, scales well, no distributional assumption | Final-value outliers, batch anomalies |
| CUSUM / Change Point Detection | Detects gradual shifts | Tool drift, mean offset |
| Autoencoder (LSTM) | Learns complex curve shapes | Full curve analysis, multivariate data |
| One-Class SVM | Robust boundary modeling | Small datasets, well-defined normal range |
Control limits are variable and can be adjusted by self-learning algorithms in the event of a process or product change. Process differentiation is based on process segmentation using pattern recognition methods.
AI-generated imagePractical Value: What Anomaly Detection Actually Makes Visible
The following overview shows typical process changes in torque assembly that a rigid tolerance window regularly misses - and that a data-driven model can flag early:
Tool drift: Trend analysis of curve slope over the production run can serve as an early warning for tool wear, before the first limit is violated. Tracking slope changes over production provides an early warning signal for tool wear.
Batch and material changes: Altered friction characteristics from a new fastener supplier or a different surface coating shift the entire curve profile. Friction coefficient analysis provides insight into the friction factors of your fasteners and enables optimization of fastening quality. An anomaly model trained on prior normal behavior detects this shift immediately - even if all final values still fall within the tolerance band.
Changed components: Differential wear across cavities in casting molds produces threaded holes with significant dimensional variation relative to each other - yet all still within permissible tolerances. One component may have its fastening point at the upper tolerance limit, while the same component from a different cavity sits at the lower limit. A trained model can identify these systematic differences between component batches.
Friction changes: Temperature fluctuations on the shop floor, lubricant application, or corrosion on threaded inserts alter the curve profile in characteristic ways - long before any final value raises a flag.
AI-based torque process monitoring can reduce analysis time in the event of a product recall by more than 90 percent, because every curve is linked to a serial number and patterns can be evaluated retrospectively.
Prerequisites: What You Actually Need
Honesty matters more than enthusiasm here. Anomaly detection is only as good as the data it's built on.
Not just final values — the entire torque-angle profile for each tightening must be recorded. Spot checks are insufficient for trend models. Data must be linked to tool ID, timestamp, and component reference.
The model learns what 'normal' looks like — so the training dataset must reflect a stable, validated process. Data from ramp-up phases, tool changes, or known disturbances must be flagged or excluded.
Rule of thumb: the more complex the tightening case (multi-stage tightening, soft joint partners, thread forming), the more reference tightenings are needed. For simple cases, a few hundred tightenings may suffice; for multivariate curve models, several thousand are advisable.
Raw data must be exportable from the tool or controller — ideally in real time or with low latency. Proprietary data silos block every subsequent step.
No model replaces the judgment of the process engineer. Alerts must be assessed, confirmed, or dismissed — this feedback iteratively improves the model and prevents alert fatigue caused by false positives.
Naming the Limitations Clearly
Data-driven anomaly detection is not a silver bullet. Keep the following constraints in mind when planning:
- False alarms are unavoidable. Every statistical model has a false-positive rate. If the model is too broad, real anomalies get classified as normal - conversely, if it's set too sensitive, it generates false alarms that erode operator trust.
- Context is not automatically known. The model sees a curve deviation - but not whether it's caused by a material change, an operator error, or a genuine process fault. Labeling and contextual data from the MES are essential.
- No substitute for process knowledge. Anomaly classification using self-learning methods combined with formalized domain knowledge enables the system to issue actionable recommendations to the user or machine operator. But that domain knowledge has to exist and be formalized in the first place.
- Distinction from Predictive Maintenance: This article focuses on quality and process anomalies - specifically, whether the joint was correctly assembled. The question of when a tool needs to be replaced is a related but separate topic.
The Data Foundation: OPERATOR® and QUANTEC MCS® as the Basis
Anomaly detection starts with data acquisition - and here, the quality of the tool determines the quality of the analysis.
The OPERATOR® captures torque and angle continuously over Wi-Fi and transmits data in real time to EasyWin® or QuanLab Pro®. The OPERATOR® EST01 communicates directly with the production system via PLC and Open Protocol - the prerequisite for seamless, serial-number-based traceability.
The QUANTEC MCS® delivers precise curve data through its fixed-point-free angle measurement with an accuracy of ±1% between 10 and 100% of the nominal range - free from the measurement errors that reaction torques introduce in fixed-point-based systems. This precision is the fundamental requirement for an anomaly model to distinguish genuine process changes from measurement noise.
Both systems connect to MES and analytics platforms via open interfaces. The data delivered by OPERATOR® and QUANTEC MCS® is therefore not only usable for real-time pass/fail evaluation - it forms the training basis for statistical models and the inference basis for ongoing anomaly detection.
Want to know whether your current tightening data already provides the foundation for data-driven anomaly detection? We'll analyze your data landscape and show you which methods make sense for your process.
Request a free screw process analysisInteractive: Which Method Fits Your Process?
The right anomaly detection method depends on data volume, process complexity, and available IT infrastructure. The following tool helps you find your starting point:
Conclusion: No Hype - But a Real Step Forward
Data-driven anomaly detection in torque data is not an AI promise for the future. The methods exist, the tools are available, and the data foundation is already in place in many manufacturing environments - it's just not being used yet.
The transition from a rigid tolerance window to adaptive process monitoring is not an either/or choice. Fixed limits remain indispensable for real-time pass/fail evaluation. Anomaly detection adds a layer on top that makes gradual changes visible before they lead to scrap or recalls.
The prerequisite is clean, complete, and traceable data acquisition - and the willingness not just to receive alerts, but to evaluate them and convert them into process knowledge.
Accuracy by GWK.
