⚡ electrical engineering

Natural visibility graph-based multiscale dispersion entropy for chatter identification in thin-walled parts finishing milling

This paper proposes an intelligent chatter identification method for thin-walled parts finishing milling that combines second-order generalized synchrosqueezing transform for signal reconstruction with natural visibility graph-based multiscale dispersion entropy for feature extraction, achieving high recognition accuracy through a crayfish optimization algorithm-optimized support vector machine.

Original authors: ZeRui Bai, Junxue Ren, Weijun Tian, Yizhuo Wang

Published 2026-07-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: ZeRui Bai, Junxue Ren, Weijun Tian, Yizhuo Wang

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 trying to listen to a faint, whispering secret (the "chatter") in a room that is constantly being shaken by a loud, rhythmic drumbeat (the "spindle rotation"). This is the challenge engineers face when machining thin-walled parts (like delicate airplane engine components). If the machine vibrates too much (chatters), the part gets ruined. But because the part is so thin and flexible, the "whisper" of the vibration is weak and easily drowned out by the "drumbeat" of the machine itself.

This paper presents a new, clever three-step method to catch that whisper before it ruins the part. Here is how it works, broken down into simple analogies:

Step 1: The "Noise-Canceling" Filter (GSST2)

The Problem: The raw signal from the machine is a messy soup of the machine's own rhythmic noise and the subtle, dangerous vibrations we want to find.
The Solution: The authors use a tool called GSST2. Think of this as a super-smart noise-canceling headphone. Instead of just blocking all sound, it specifically identifies the "drumbeat" (the machine's rotation) and subtracts it out.
The Result: What's left is a clean, reconstructed signal that contains only the "whisper" of the chatter. The paper shows this method is better at cleaning up the signal than older methods, leaving behind a much clearer picture of what's actually happening.

Step 2: Turning Sound into a Map (NVG-MDE)

The Problem: Once you have the clean signal, how do you measure if it's "bad"? Standard math tools often miss the subtle, complex patterns of these vibrations because they look at the data as just a line on a graph.
The Solution: The authors use a technique called Natural Visibility Graph (NVG). Imagine the vibration signal is a landscape of hills and valleys. The NVG method draws "lines of sight" between the peaks and valleys. If two points can "see" each other without a taller hill blocking the view, they are connected. This turns the wiggly line into a map or a social network of connections.
The Twist: They then apply a "multiscale dispersion entropy" (MDE) to this map. Think of this as a "complexity meter" that walks around the map at different zoom levels. It counts how chaotic or organized the connections are.
The Result: This map-based approach is much better at spotting the difference between a "stable" cut and a "chattering" cut than traditional methods. It's like how a detective might look at a crime scene not just for footprints, but for the pattern of how the footprints connect to each other.

Step 3: The "Smart Coach" (COA-SVM)

The Problem: Now that we have a "complexity score" from our map, we need a computer to decide: "Is this safe or is this dangerous?"
The Solution: They use a machine learning classifier called SVM (Support Vector Machine). However, SVMs are like students who need to be taught the right study habits (parameters) to pass the test. If the habits are wrong, they fail.
The Twist: To find the perfect study habits, they use a new algorithm called Crayfish Optimization Algorithm (COA). Imagine a group of crayfish searching for food in a cave. They have three modes:

Heat Avoidance: Hiding from the sun (exploring new areas).
Competition: Fighting for the best spot (refining the search).
Foraging: Eating the best food found (locking in the solution).
The "crayfish" swarm works together to find the absolute best settings for the SVM "student."
The Result: The "crayfish-coached" SVM becomes a very sharp detective.

The Final Score

The team tested this entire system on real thin-walled titanium parts.

The Outcome: Their new method correctly identified whether the machine was stable, slightly chattering, or severely chattering 90.08% of the time.
Comparison: This was better than using older "map" methods, older "entropy" methods, or other types of "animal-swarm" coaches (like shark or snake algorithms).

Summary

In short, the paper says: "To hear the faint warning signs of a machine failing, first filter out the machine's own noise, then turn the remaining signal into a topological map to measure its hidden complexity, and finally use a crayfish-swarm coach to teach a computer how to read that map perfectly."

This allows manufacturers to catch problems early and ensure their delicate parts come out perfect.

Technical Summary: Natural Visibility Graph-based Multiscale Dispersion Entropy for Chatter Identification in Thin-Walled Parts Finishing Milling

Problem Statement
Thin-walled parts are highly susceptible to chatter during finishing milling due to their weak, time-varying rigidity and complex structures. This phenomenon severely degrades machining precision and surface quality. While stability lobe diagrams (SLD) offer theoretical predictions, they often fail to prevent chatter in actual finishing operations due to model simplifications and environmental complexities. Consequently, effective real-time detection is necessary. However, identifying chatter in thin-walled parts is particularly challenging because the signals exhibit strong nonlinearity and non-stationarity. Early-stage chatter components are often weak and easily masked by periodic components induced by spindle rotation. Existing signal processing methods, such as time-frequency analysis (STFT, CWT) and decomposition techniques (VMD, EEMD), often suffer from parameter sensitivity, the Heisenberg uncertainty principle, or bias in instantaneous frequency estimation, making them insufficient for accurately extracting weak, time-varying chatter signals. Furthermore, traditional entropy-based feature extraction methods often fail to capture the implicit topological relationships embedded within these weak signals.

Methodology
The paper proposes an intelligent chatter identification framework comprising three main stages: signal preprocessing, feature extraction, and classification.

Signal Preprocessing (GSST2): To isolate the weak chatter component from dominant periodic spindle signals, the authors employ the Second-Order Generalized Synchrosqueezing Transform (GSST2). Unlike first-order methods (e.g., MSST), GSST2 utilizes high-order instantaneous frequency estimation to achieve unbiased tracking of nonlinear frequency trajectories. The method filters out coefficients corresponding to the spindle rotation frequency and its harmonics in the time-frequency matrix, allowing for the high-fidelity reconstruction of the chatter signal while preserving its nonlinear characteristics.
Feature Extraction (NVG-MDE): The core innovation is the Natural Visibility Graph-based Multiscale Dispersion Entropy (NVG-MDE).
- Natural Visibility Graph (NVG): The reconstructed time series is mapped into a graph signal where data points become vertices and visibility conditions define edges. Unlike Horizontal Visibility Graphs (HVG), NVG considers both amplitude values and time interval weighting, preserving more local fluctuation information.
- Multiscale Dispersion Entropy (MDE): Dispersion entropy is calculated on the graph structure rather than the raw time series. This involves constructing an embedding matrix from the weighted adjacency matrix, mapping amplitudes to categories via a normal cumulative distribution function, and calculating the Shannon entropy of the resulting dispersion patterns.
- Multiscale Analysis: To capture signal complexity at different time scales, a maximum coarse-graining process is applied to the time series before graph transformation, generating entropy values across multiple scales.
Classification (COA-SVM): The extracted NVG-MDE features are fed into a Support Vector Machine (SVM) classifier. To overcome the difficulty of manually selecting optimal SVM parameters (penalty factor $c$ and kernel parameter $\sigma$ ) for complex, non-stationary data, the Crayfish Optimization Algorithm (COA) is introduced. COA, inspired by the summer behaviors of crayfish (heat avoidance, competition, and foraging), optimizes the SVM parameters to maximize classification accuracy and avoid local optima.

Key Contributions
The paper outlines three primary contributions:

Preprocessing: The application of GSST2 for denoising and extracting chatter components from thin-walled part finishing milling signals, effectively removing periodic interference.
Feature Representation: The proposal of NVG-MDE, which differs from classical entropy measures by capturing implicit topological relationships and amplitude distributions within the signal, creating a more informative feature set for distinguishing subtle chatter conditions.
Optimization: The development of a COA-SVM model that automatically optimizes classifier parameters, enabling accurate recognition of stable, slight chatter, and severe chatter conditions.

Results
The proposed method was validated through finishing milling experiments on TC4 thin-walled parts using a five-axis machine tool. Acceleration signals were collected under various cutting parameters (spindle speeds of 5000–8000 rpm and depths of cut of 0.2–0.4 mm).

Preprocessing Validation: Comparisons of FFT spectra demonstrated that GSST2 effectively removed periodic components (e.g., spindle harmonics) while preserving chatter frequencies, outperforming the MSST-based method in signal reconstruction fidelity.
Feature Discrimination: Entropy curves showed that NVG-MDE could clearly differentiate between stable, slight chatter, and severe chatter conditions across various scale factors. In contrast, HVG-MDE and classical MDE failed to clearly distinguish between slight and severe chatter.
Classification Performance: The COA-SVM model utilizing NVG-MDE features achieved an average recognition accuracy of 90.08% over 10 experimental runs. This outperformed other optimization algorithms (SCSO-SVM, SO-SVM, WSO-SVM) and the standard SVM. The T-SNE visualization confirmed that NVG-MDE features provided better separability between classes compared to HVG-MDE and MDE.

Significance and Claims
The authors claim that the proposed method addresses the specific challenges of finishing thin-walled parts where chatter features are subtle and difficult to extract. By defining entropy on graph signals, the method successfully captures topological relationships that are invisible in the time or frequency domains. The study validates that defining entropy on graph signals is a feasible and effective strategy for milling chatter identification. The authors modestly note that while the method is effective, the computational efficiency of graph signal construction requires improvement, and future work will explore the method's robustness under varying cutting parameters and further optimize graph entropy settings.

Step 1: The "Noise-Canceling" Filter (GSST2)

Step 2: Turning Sound into a Map (NVG-MDE)

Step 3: The "Smart Coach" (COA-SVM)

The Final Score

Summary

More like this