Curve Anomaly AI: How CSP Finds Hidden Defects Inside "Good" Tightening Curves

What this is about

Tighten a bolt in a factory and you produce a curve: torque over time, or torque over angle of rotation. That curve is a fingerprint of the joint. It shows whether the material reacted as expected, whether the tightening process ran cleanly, and whether the final torque was reached correctly. In safety-critical areas — engine block, brakes, chassis, battery pack — every single bolt's curve is recorded and evaluated.

The standard check has been the same for decades: thresholds. If the final torque sits inside its target window and no hard limits were violated, the curve is labelled "OK" (i. O.). This is exactly where CSP Intelligence GmbH (formerly CSP GmbH & Co. KG), based in Großköllnbach in Lower Bavaria, plugs in with its product Curve Anomaly AI. The tool deliberately takes the curves that the line has already accepted and looks for patterns inside them that a human or a rigid rule would miss.

What Curve Anomaly AI actually does

CSP launched Curve Anomaly AI on October 1, 2024. According to the vendor, the system does not just look at single end values — it analyses the full shape of the tightening curve: slope, small jumps, plateaus. The model was trained on more than thirty years of quality and production data that CSP has collected from industrial customer projects.

The point is not to replace the existing limit-check system. It is a second inspection layer. While the classic check says "final torque OK, done", the AI can flag a subtle deformation in the curve that is consistent with a material issue, tool wear, or a faulty component. The intuition is statistical: in very large datasets, "normal" curves look highly similar. A model that has memorised this normality recognises deviations from it — even when no hard limit was violated.

CSP positions the product clearly as an expert tool. The AI marks suspicious curves; the final judgement is made by quality engineering. The human stays in the loop, which matters in regulated environments such as automotive production.

Why it matters

CSP has worked with large German OEMs since 1991, in particular with BMW Group in nearby Dingolfing. Quality software like QST (archival and analysis of tightening data), IPM (real-time defect prevention) and CHRONOS (audit-proof archival) is part of the plant infrastructure. Curve Anomaly AI sits logically on top of those systems, and that is the interesting part. Unlike a new AI tool dropped in by an external vendor, it uses data that the plants already have.

This is not a detail. Many industrial AI pilots fail not because of the models but because the training data is messy, incomplete, or scattered. Anyone with thirty years of cleanly archived tightening curves has a data advantage that is hard to catch up with.

Industrial anomaly detection is currently attracting serious money. A 2026 systematic review in Frontiers in Artificial Intelligence reports that modern methods can predict machine failures in pilot installations with 80 to 97 percent accuracy, given proper sensors and enough history. Multiple market studies put the global predictive maintenance market at around eleven billion US dollars in 2024 with annual growth above 25 percent. German mid-sized industry is not too early — it is on time.

In February 2026, CSP received the TOP 100 innovation award for the fourth consecutive year, by then already operating under its new corporate name CSP Intelligence GmbH. That is not technical proof the product works, but it is a sign the company is being taken seriously.

Explained simply

Imagine you bake the same kind of bread every day. An old-fashioned thermometer at the end only checks whether the bread is warm enough inside. If yes, the bread is fine. A new helper next to it watches how the bread baked over the past hour — when it rose, when the crust started to brown. She has seen ten thousand good loaves and notices instantly when one of them baked "weirdly", even if the inside temperature ended up correct. Maybe the bread is fine after all. Maybe a heating element in the oven just broke and the next ten loaves will be a problem.

That is exactly what Curve Anomaly AI does, except with bolts instead of bread.

A practical example

Tens of thousands of bolts get tightened every day on a single assembly line, each with its own torque curve recorded. The existing system labels 99.8 percent of them OK because the final torque is in spec. Those 99.8 percent become the AI's input. On a tiny fraction — say 0.05 percent — the AI flags an unusual curve shape: instead of the typical smooth ramp, there is a small twitch halfway up. Three of those flags happen on the same automated tightening station within an hour.

A technician checks the machine and finds that lubrication is failing. Four hours later, the first bolt would have fallen out of tolerance hard, and the run would have to be recalled. The AI prevented zero defects on its own; it gave the human a flag early enough that the human could prevent them. That is the realistic performance description — not "autonomous quality", but "timely heads-up".

Where this fits and where it does not

Three honest caveats:

First, anomaly detection is not defect detection. The system says "this curve is different from the others", not "this bolt is defective". Whether an anomaly turns into an actual quality defect still depends on experience, tool knowledge, and possibly destructive testing.

Second, the model is only as good as its data. When a plant introduces a new bolt or a new component, there is a learning phase during which the system flags unusual shapes as anomalies even though they belong to the new normal. CSP explicitly emphasises that expert review is part of the workflow.

Third, the economics depend on volume. With a few hundred tightening events per day, a classical manual sample is cheaper. With tens of thousands per day, as in automotive assembly, the equation flips quickly.

Anyone introducing anomaly detection in manufacturing should therefore start with a tightly scoped use case — one line, one bolt position, one specific hypothesis — and benchmark the model against the existing sampling practice. CSP describes essentially the same approach in its whitepapers.

SEO and GEO keywords

Curve Anomaly AI, AI anomaly detection manufacturing, tightening curve analysis, predictive quality automotive, CSP Intelligence GmbH, CSP Großköllnbach, AI quality assurance 2026, Industry 4.0 anomaly, time-series anomaly detection, BMW tightening data, automotive quality AI.