Abstract
This paper addresses the ability to enable machines to automatically detect failures on machine tool components as well as estimating the severity of the failures, which is a critical step towards autonomous production machines. Extracting information about the severity of failures has been a substantial part of classical, as well as Machine Learning based machine vision systems. Efforts have been undertaken to automatically predict the severity of failures on machine tool components for predictive maintenance purposes. Though, most approaches only partly cover a completely automatic system from detecting failures to the prognosis of their future severity. To the best of the authors' knowledge, this is the first time a vision-based system for defect detection and prognosis of failures on metallic surfaces in general and on Ball Screw Drives in specific has been proposed. The authors show that they can do both, detect and prognose the evolution of a failure on the surface of a Ball Screw Drive.
Introduction: The Challenge of Predictive Maintenance
In modern manufacturing, unexpected machine failures are costly, causing massive supply chain delays. While many "Condition Monitoring" systems can detect a fault once it happens, few can perform true "Predictive Maintenance"—forecasting when a component will fail before it actually does.
The authors focus on Ball Screw Drives (BSDs), critical components in machine tools that often fail due to "pitting" (small material breakouts caused by fatigue). Current vision-based approaches are mostly limited to simple classification (e.g., "defect" vs. "no defect"). This paper bridges the gap by quantifying the size of these defects over time and using that history to predict future wear severity.
Methodology: From Detection to Prognosis
The researchers developed a multi-stage pipeline to move from raw images to a future forecast. To capture high-quality data, a specialized test bench was used to artificially wear down BSD components. A camera system was integrated into the housing to monitor the spindle surface during operation.
The processing pipeline consists of four main stages:
- Detection: The EfficientDet D0 object detection model locates pittings in images of the spindle surface.
- Quantification: To measure the exact area of a defect, a CNN predicts an optimal threshold value for binarizing the image, which is more robust than standard methods like Otsu’s algorithm in varying lighting conditions.
- Expert System: This layer applies domain knowledge—specifically, the physical reality that a surface defect can grow but never shrink. This helps filter out noise caused by oil or dirt on the spindle.
- Forecasting: The system uses Linear Regression to project the growth of the defect area into the future.
Key Insights and Results
A critical finding of the study was that purely data-driven models can be "fooled" by surface pollution (like oil droplets) that makes a defect look smaller or larger than it is. By encoding the rule that "defects only grow," the expert system significantly smoothed the prediction curve and removed outliers.
Furthermore, while one might expect complex, high-order polynomials to model wear better, the authors found that Linear Regression provided the most stable and accurate long-term forecasts. Higher-degree models tended to overfit and oscillate wildly, making them useless for practical maintenance planning.
Performance Summary
- Detection Accuracy: The system achieved a 92% validation accuracy in correctly identifying and bounding pittings.
- Forecasting Performance: The linear regression model achieved the lowest Root Mean Square Error (RMSE = 0.125) compared to 2nd-degree (0.233) or 3rd-degree (0.553) polynomials.
- Predictive Capability: The system could accurately predict the end-of-life of a component after observing only the first 40% of its lifetime.
Conclusion and Future Work
Vision-based systems allow for "inspection by eye," making the AI's decisions more interpretable and verifiable by human technicians than traditional vibration sensors. However, limitations remain, such as resolution constraints for very early-stage failures and heavy surface pollution. Future research could combine these visual forecasts with time-domain signals like motor current or acoustic emissions to create an even more robust "tandem" prediction system.
Summary updated