⚡ electrical engineering

Tool Wear Monitoring via the Fusion of Cutting Force Coefficients and Fractal Features of Cutting Force Signal

This paper proposes a novel online milling tool wear monitoring method that fuses cutting force coefficients and fractal features within a Bayesian inference framework, demonstrating superior accuracy compared to standalone approaches.

Original authors: Chenbo Li, Xuyu Zhang, Feng Xia, Zixu Chen, Tongshun Liu

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chenbo Li, Xuyu Zhang, Feng Xia, Zixu Chen, Tongshun Liu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to tell how dull a knife has become just by listening to the sound of it slicing through a carrot. You know that a sharp knife makes a clean, crisp sound, while a dull knife makes a ragged, uneven noise. But what if you could also measure exactly how hard you are pushing the knife?

This paper is about doing exactly that, but for industrial machines cutting metal. The researchers wanted to create a smarter way to tell when a milling tool (a high-tech drill or cutter) is wearing out before it breaks or ruins the metal part it's working on.

Here is the simple breakdown of their approach:

The Two Clues: The "Rule" and the "Roughness"

The researchers realized that the force a machine exerts while cutting metal tells two different stories at the same time. They decided to listen to both stories instead of just one.

1. The "Rule" (Cutting Force Coefficients)
Think of this as the steady rhythm of the machine.
When a tool cuts, there is a predictable pattern to how much force is needed based on the shape of the cut. The researchers call these "cutting force coefficients."

The Analogy: Imagine a drummer keeping a perfect beat. As the tool gets dull, the "beat" changes in a specific, predictable way. The force required to push the tool through the metal gets heavier and more consistent. This part of the signal is like the regular, predictable part of a song.

2. The "Roughness" (Fractal Features)
Think of this as the chaos or static in the signal.
Real-world cutting isn't perfect. There are vibrations, heat, and tiny imperfections that make the force signal "jittery" or irregular. The researchers used a math concept called "fractal analysis" to measure this irregularity.

The Analogy: Imagine the same drummer, but now they are hitting the drums with a bit of a tremor in their hand. A fresh tool has a lot of this "tremor" or "jitter" because the sharp edge is biting into the metal in a chaotic way. As the tool gets dull, the edge becomes smoother (like a worn-down stone), and the signal becomes smoother and less chaotic.

The Secret Sauce: Mixing the Clues

The paper argues that looking at just the "Rule" (how hard it pushes) or just the "Roughness" (how jittery it is) isn't enough.

If you only look at the Rule, you might miss subtle changes.
If you only look at the Roughness, you might get confused by noise.

The Solution: They combined both clues into a single "super-signal."
They took the 6 numbers that describe the "Rule" (how the force is distributed) and the 3 numbers that describe the "Roughness" (how chaotic the signal is). They fed all 9 of these numbers into a smart computer brain called a Bayesian Network.

The Analogy: Imagine a detective trying to solve a mystery. One witness (the Rule) says, "The suspect is getting heavier." Another witness (the Roughness) says, "The suspect is moving more smoothly." If you only listen to one, you might be wrong. But if you listen to both, you get a much clearer picture of who the suspect is and how dangerous they are.

The Experiment: The "Tough Cookie" Test

To prove their idea worked, they didn't just guess; they used real data from a famous competition (PHM2010).

The Setup: They used a machine to cut a very tough metal (Inconel 718) with a ball-nose cutter. The tool made 315 cuts in a row, getting duller and duller with every pass.
The Test: They trained their computer brain on some of the cuts and then asked it to guess the tool's condition on new cuts it hadn't seen before.

The Results: Smarter than the Sum of Its Parts

The results showed that their "Super-Signal" method was the winner:

Using just the "Rule" (coefficients) had a certain error rate.
Using just the "Roughness" (fractals) had a slightly higher error rate.
Using both together was the most accurate.

The Bottom Line:
By combining the predictable changes in force with the changes in chaos, the researchers improved the accuracy of predicting tool wear by about 5.6% compared to using just the force rules, and by 9.8% compared to using just the chaos analysis.

In simple terms: They found that listening to both the steady beat and the background static of a cutting machine gives you a much better idea of when the tool is about to fail than listening to either one alone.

Technical Summary: Tool Wear Monitoring via the Fusion of Cutting Force Coefficients and Fractal Features

Problem Statement
Tool wear in milling operations significantly compromises machining quality and efficiency, often leading to tool fracture and substantial economic losses. While cutting force signals are rich in wear-related information, their high sampling rates generate large, redundant datasets that are unsuitable for direct estimation. Effective monitoring requires extracting low-dimensional, discriminative features. Existing research has largely treated two distinct aspects of the cutting force signal in isolation: the regular patterns captured by cutting force coefficients (derived from mechanistic models) and the irregular fluctuations characterized by fractal analysis. The paper posits that relying on only one of these perspectives limits monitoring precision, as the cutting process inherently contains both regular physical interactions and irregular dynamic behaviors.

Methodology
The study proposes a novel monitoring framework that fuses regular and irregular features extracted from triaxial cutting force signals, processed through a Bayesian inference system.

Feature Extraction:
- Cutting Force Coefficients (Regular Features): A six-parameter mechanistic cutting force model is employed. This model relates instantaneous uncut chip thickness to cutting forces via shear and friction coefficients in tangential, radial, and axial directions. Six coefficients ( $K_{t,s}, K_{r,s}, K_{a,s}, K_{t,f}, K_{r,f}, K_{a,f}$ ) are identified using a least-squares algorithm applied to the raw force signals. These coefficients reflect the systematic distribution of forces across the tool-chip and tool-workpiece contact zones.
- Fractal Features (Irregular Features): To capture signal irregularity, the study utilizes the correlation dimension ( $D_c$ ). Unlike box or regularization dimensions, the correlation dimension is deemed more effective for characterizing signal irregularity. The authors reconstruct the phase space of the cutting force signals and calculate the correlation dimension for the three force components ( $F_x, F_y, F_z$ ). An embedding dimension of 80 is selected based on stability analysis, and the dimension is computed by linearly fitting the logarithm of the correlation integral within a specific scale range.
Fusion and Estimation:
- The nine extracted features (6 coefficients + 3 correlation dimensions) are integrated using a Naïve Bayesian Network.
- The network treats the tool wear state (flank wear, $VB$) as a hidden node and the nine features as observable nodes.
- A sequential estimation approach is implemented where the prior probability of the current wear state is updated based on the posterior probability of the previous state. This allows the system to incorporate both instantaneous feature data and the temporal evolution of tool wear, enhancing noise robustness.
- The final wear estimate is determined via Maximum A Posteriori (MAP) probability.

Experimental Validation
The methodology was validated using the PHM2010 Prognostic Health Monitoring competition dataset.

Setup: High-speed CNC milling of Inconel 718 (and stainless steel in some contexts) using 6mm ball-nose carbide tools.
Data: Three-channel cutting force signals (50 kHz sampling) from experiments C1, C4, and C6.
Protocol: A cross-validation strategy was used where two experiments served as training data to learn network parameters, and the third served as testing data. This cycle was repeated three times.

Key Results

Feature Behavior: As tool wear progressed, the cutting force coefficients (both shear and friction) consistently increased, reflecting higher friction and force distribution changes. Conversely, the correlation dimensions decreased, indicating that the cutting force waveform became smoother and more regular as the tool tip blunted.
Monitoring Accuracy: The fusion method demonstrated superior performance compared to standalone approaches.
- The fusion method achieved relative accuracy improvements of 5.6% over the coefficient-based method and 9.8% over the fractal-based method.
- Specific error rates ($Er$) for the fusion method across test cases (C1, C4, C6) were 5.09%, 5.49%, and 5.59%, respectively, which were lower than the errors produced by using either feature set alone.

Significance and Claims
The paper claims that the primary contribution lies in the complementary integration of regular and irregular signal characteristics. By acknowledging that tool wear induces both systematic changes in force distribution (captured by coefficients) and changes in signal irregularity (captured by fractal dimensions), the proposed hybrid approach provides a more comprehensive characterization of the wear state. The study concludes that this fusion, processed through a Bayesian framework that accounts for wear evolution, significantly enhances the precision of online tool wear monitoring compared to methods relying on a single feature type.

The Two Clues: The "Rule" and the "Roughness"

The Secret Sauce: Mixing the Clues

The Experiment: The "Tough Cookie" Test

The Results: Smarter than the Sum of Its Parts

More like this