Application of machine learning for fleet-based condition monitoring of ball screw drives in machine tools

Berend Denkena, Marc-André Dittrich, H. D. Nöske, D. Lange, Carolin Benjamins, Marius LindauerThe International Journal of Advanced Manufacturing Technology 2023 · DOI

TL;DR

This paper demonstrates how to move condition monitoring for ball screw drives out of the lab and into the factory. By applying Automated Machine Learning (AutoML) and a "unified outlier score" to a real-world fleet of 13 industrial machines, the researchers achieved significantly higher fault detection accuracy than traditional industry methods and previous academic baselines.

What Problem Does This Paper Address?

Ball screw drives are the "muscles" of machine tools, responsible for high-precision positioning. When they fail, production lines grind to a halt, leading to massive downtime costs. While researchers have tried to use machine learning (ML) to predict these failures, most studies rely on "artificial" data—faults created intentionally in a controlled lab setting.

The real world is messier. In a factory "fleet," even identical machines have slightly different vibration signatures, and the transition from "healthy" to "failing" happens gradually over months, not minutes. This paper addresses the gap between lab-grown models and industrial reality by using data from a real automotive production line to detect both sudden and insidious wear.

Methods

The authors explored two primary ML paths:

  1. Supervised Anomaly Detection (Auto-Sklearn): When you have labeled data (knowing exactly when a fault occurred), you can train a classifier. Instead of manually picking an algorithm, the authors used Auto-Sklearn, an AutoML tool that automatically searches through dozens of classifiers (like Random Forests or SVMs) and preprocessing techniques to find the optimal "pipeline."
  2. Semi-supervised Anomaly Detection (Unified Outlier Scores): In many factories, fault data is rare. The authors used a K-Nearest Neighbor (KNN) approach to establish a "baseline" of normal behavior. They then calculated a Unified Outlier Score, which normalizes different sensor readings (using Gaussian or Gamma scaling) into a single health index. This allows the system to flag "weird" behavior even if it hasn't seen that specific type of failure before.

For feature extraction, they used the tsfresh library to generate over 700 time-series features (like mean, variance, and Fourier coefficients) from the raw sensor data.

Datasets

The study used a unique industrial dataset:

  • The Fleet: 13 identical five-axis machine tools (MAG Specht 600) used in automotive series production.
  • Duration: Data collected over 8 months.
  • Scale: 1,540 test cycles recorded.
  • Ground Truth: During the study, 4 ball screws actually failed and were replaced, providing rare "real-world" transition data from healthy to faulty states.
  • Sensors: Internal motor torque ($M_{BSD}$) and external 3-axis acceleration sensors ($Acc_{1-3}$).

Key Insights and Ideas

  • Torque is Not Enough: While many industrial systems rely solely on motor torque, the authors found it insufficient for early-stage detection. External acceleration (vibration) sensors were much more sensitive to the "insidious" loss of preload and early raceway damage.
  • The "Individual" Matters in the Fleet: Even though the machines were identical, their baseline signals varied. The authors discovered that a model trained on the entire fleet performed worse than a model trained on the first few cycles of an individual machine. This suggests that "separate training mode" is essential for high-precision monitoring.
  • AutoML Outperforms Humans: The Auto-Sklearn tool consistently found better model configurations than the manually selected models used in previous literature, proving that the complexity of sensor data is often best handled by automated optimization.
  • Gamma Scaling for Outliers: When calculating outlier scores, Gamma scaling proved more robust than simple linear or Gaussian scaling, better capturing the "long tail" of sensor noise in a factory environment.

Key Results

  • Supervised Performance: Auto-Sklearn achieved an f1-score of 0.9576, significantly outperforming standard baselines like Support Vector Machines (~0.88) or Decision Trees (~0.82).
  • Early Detection: The semi-supervised approach detected failures in the "transition phase" that traditional threshold-based industry tools missed entirely.
  • Economic Impact: The authors note that if a ball screw is repaired based on these early warnings rather than replaced after failure, 30–50% of replacement costs can be saved.

Limitations and Future Work

The primary limitation is the small number of actual failure events (only 4). While 8 months of data is extensive, more failures are needed to fully map the "transition phase" between healthy and faulty states.

Future work could involve Transfer Learning, where a model learned on one machine fleet is adapted to a different type of machine with minimal new data. The authors also suggest exploring deep learning architectures, though they emphasize that "simpler" ML models optimized via AutoML are currently more practical for industrial deployment.

Summary updated

Abstract

Abstract Ball screws are frequently used as drive elements in the feed axes of machine tools. The failure of ball screw drives is associated with high downtimes and costs for manufacturing companies, which harm competitiveness. Data-based monitoring approaches derive the ball screw condition based on sensor data in cases where no knowledge is available to derive a physical model-based approach. An essential criterion for selecting the condition assessment method is the availability of fault data. In the literature, fault patterns are often artificially created in an experimental test bench scenario. This paper presents ball screw drive monitoring approaches for machine tool fleets based on machine learning. First, the potentials of automated machine learning for supervised anomaly detection are investigated. It is shown that the AutoML tool Auto-Sklearn achieves a higher monitoring quality compared to literature approaches. However, fault data are often not available. Therefore, unified outlier scores are applied in a semi-supervised anomaly detection mode. The unified outlier score approach outperforms threshold-based approaches commonly used in industry. The considered data set originates from a machine tool fleet used in series production in the automotive industry collected over 8 months. Within the observation period, multiple ball screw failures are observed so that sensor data about the transient phases between normal and fault conditions is available.
BibTeX
@article{denkena2023application,
  author = {Berend Denkena and Marc-André Dittrich and H. D. Nöske and D. Lange and Carolin Benjamins and Marius Lindauer},
  title = {{Application of machine learning for fleet-based condition monitoring of ball screw drives in machine tools}},
  journal = {The International Journal of Advanced Manufacturing Technology},
  year = {2023},
  doi = {10.1007/s00170-023-11524-9},
}
Figure