A Lightweight Hybrid Approach for Milling Force Reconstruction from Internal Machine Signals

Eric Hirsch, Christian FriedricharXiv 2026 · arXiv

Monitoring milling forces is the "holy grail" of process optimization and tool condition monitoring. However, in a real-world factory, you can't just bolt a \$20,000 dynamometer to every machine. While internal machine signals (like spindle current) are "free" data, they are often too slow or too heavily filtered to capture the high-frequency "heartbeat" of a cutting tool.

In their latest paper, Eric Hirsch and Christian Friedrich propose a clever workaround: a lightweight hybrid approach that uses machine learning to predict the "big picture" and physics to fill in the high-frequency details.

TL;DR

This paper presents a hybrid method to reconstruct high-frequency milling forces using only low-frequency internal machine signals and process metadata. By using Ridge regression to predict average forces and a mechanistic model to reconstruct the instantaneous profile, the authors achieve over 96% accuracy without requiring expensive external sensors during production.

What Problem Does This Paper Address?

To optimize a milling process, you need to know exactly how the tool is hitting the material. Direct measurement requires piezoelectric dynamometers, which are costly, intrusive, and reduce the available workspace.

The alternative—using internal signals from feed drives and spindles—is attractive but technically difficult. These signals have low sampling rates and are influenced by machine inertia and friction. They can tell you the average load, but they usually can't "see" the rapid force variations as individual tool flutes engage and disengage the workpiece, especially at high RPMs.

Methods

The authors' core innovation is a two-stage "cascade" architecture that avoids the pitfalls of trying to learn a complex, high-frequency signal directly from noisy, low-frequency data.

The overall hybrid workflow combining data-driven estimation and physics-based reconstruction.
  1. Data-Driven Estimator: Instead of predicting the force at every millisecond, a Ridge regression model predicts the mean in-plane forces ($\bar{F}_x$ and $\bar{F}_y$) for an entire cutting run. It uses 11 features extracted from internal signals (like spindle power and torque) and 14 metadata descriptors (like tool diameter, feed rate, and material type).
  2. Physics-Based Reconstructor: These predicted mean forces are then used to calculate cutting force coefficients. These coefficients are plugged into a static mechanistic milling model (based on the well-known Altintas formulation). This model "unfolds" the average force into a high-resolution reconstruction of the force progression over a single spindle revolution.

To ensure the results make physical sense, the authors include physics-guided corrections, such as enforcing specific ratios between tangential and radial force coefficients to prevent the model from producing physically impossible results.

Datasets

The approach was validated using a substantial experimental dataset:

  • Machine: A five-axis Hermle C40 machining center.
  • Data: 1,391 labeled cutting runs.
  • Materials: 42CrMo4+QT steel and AlMg4.5Mn aluminium.
  • Tools: Eight different solid-carbide end mills (varying diameters, flute counts, and coatings).
  • Ground Truth: A custom platform using four Kistler 9317B force sensors was used to provide reference measurements for training and validation only.
The experimental setup featuring the reference-force measurement platform.

Key Insights and Ideas

  • Feature Selection over Complexity: The authors found that simple signal features like the global mean, Higuchi fractal dimension (HFD), and SVD entropy were highly effective when combined with machining metadata.
  • The "Average" is Enough: The most significant "aha" moment is that internal signals do contain enough information to accurately estimate the average load. Once you have the average load and the geometry of the cut (from metadata), physics can handle the high-frequency reconstruction better than a pure ML model could.
  • Robustness through Hybridization: Purely data-driven models often struggle when moving between different materials (e.g., steel vs. aluminum). By anchoring the ML output in a physical model, the system becomes much more robust to changes in the machining environment.

Key Results

The results demonstrate that "lightweight" doesn't mean "low accuracy":

  • Mean Force Prediction: The Ridge regression model achieved an $R^2$ of 0.984 for $F_x$ and 0.991 for $F_y$.
  • Force Reconstruction: When comparing the reconstructed force profiles to the sensor-based baseline, the model achieved an $R^2$ of 0.962 ($F_x$) and 0.982 ($F_y$).
Comparison of predicted vs. measured mean forces across different materials.

Qualitatively, the model captures the principal force directions, load levels, and dominant engagement peaks very well, particularly in steel where the signal-to-noise ratio is higher.

Representative force profiles showing the high fidelity of the reconstruction compared to filtered measurements.

Limitations and Future Work

While impressive, the authors note a few key limitations:

  • Static vs. Dynamic: Because the reconstruction uses a static model, it cannot reproduce dynamic effects like vibrations, chatter, or local force oscillations.
  • Tool Wear: This study focused on fresh, unworn tools. Future work will need to investigate how tool wear—which changes the cutting coefficients—affects the reconstruction accuracy.
  • Real-time Implementation: The current workflow is designed for offline analysis. Moving this to an online, real-time monitoring system is the next logical step for industrial deployment.

Summary updated

Abstract

Data-driven monitoring based on milling force can enable advanced manufacturing services including process optimization and condition monitoring. However, industrial implementation is often hindered by costly and intrusive hardware and large training data requirements. This paper therefore presents a hybrid approach using machine-internal signals and machining metadata to enable force-based monitoring on brownfield machines without additional sensors. Commonly available machine signals have relatively low sampling rates and cannot directly represent high-frequency milling-force variations, especially at higher spindle speeds. Instead of learning the complete force signal, which would increase model complexity, data requirements, and overfitting risk, a supervised machine-learning model predicts only the average milling force from selected machine-signal features and machining parameters. This low-frequency estimation parameterizes a static milling force model that reconstructs the force behavior over one spindle revolution. Although the static model does not reproduce all dynamic effects, it captures the force distribution, load levels, and angular regions of highest tool engagement. The approach is evaluated on one milling machine across varying process parameters, engagement conditions, cutting tools, and two workpiece materials. Dynamometer measurements are required only for development and validation. Ridge regression achieves an $R^2$ of $0.984$ and $0.991$ for average-force prediction on the test data ($\bar{F}_x$ and $\bar{F}_y$). The reconstructed force behavior over one spindle revolution achieves an $R^2$ of $0.962$ and $0.982$ ($F_x$ and $F_y$) compared with the reconstruction based on measured force features. The method provides a cost and computationally efficient basis for process monitoring, tool condition assessment, and process optimization on existing machines.
BibTeX
@misc{hirsch2026lightweight,
  author = {Eric Hirsch and Christian Friedrich},
  title = {{A Lightweight Hybrid Approach for Milling Force Reconstruction from Internal Machine Signals}},
  year = {2026},
  eprint = {2611.25921},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.25921},
}
Figure