Data Acquisition and Chatter Recognition Based on Multi-Sensor Signals for Blade Whirling Milling

Xinyu Li, Riliang Liu, Zhiying ZhuMachines 2025 · DOI

TL;DR

This paper introduces a novel approach for detecting chatter, a detrimental vibration phenomenon, during blade whirling milling using multi-sensor data. By fusing vibration and acoustic emission signals with a deep learning model (MLGRU-SAM), they achieve high accuracy and real-time detection, crucial for improving machining efficiency and quality of complex bladed components.

What Problem Does This Paper Address?

Bladed components are vital for engines and propulsion systems, but their intricate shapes, thin structures, and the demanding material removal rates during machining make them notoriously difficult to manufacture. Whirling milling offers a promising solution to boost machining efficiency for these parts. However, this process is highly susceptible to chatter, a self-excited vibration that degrades surface quality, increases tool wear, and can even damage the workpiece. Existing chatter detection methods often struggle to adapt to the unique challenges of whirling milling, especially for complex geometries like blades. This paper aims to bridge this gap by developing an effective and timely chatter recognition system specifically for blade whirling milling.

Methods

The researchers propose a multi-sensor data acquisition and chatter recognition framework. The core of their approach involves:

  • Multi-Sensor Data Acquisition: They collect data from accelerometers (measuring vibrations in X, Y, and Z directions) and an acoustic emission sensor during both roughing and finishing stages of blade whirling milling.
  • Feature Engineering: A comprehensive set of features is extracted from the raw sensor signals, including time-domain (peak value, standard deviation, kurtosis, margin factor), frequency-domain (centroid frequency, mean square frequency), and time-frequency domain features derived from wavelet packet decomposition.
  • Feature Selection and Fusion: To manage the high dimensionality and potential redundancy of features, they employ Principal Component Analysis (PCA). PCA is used to reduce the dimensionality of the feature sets while retaining the most informative components that capture chatter-related information. A threshold of 60% accuracy is used to filter out underperforming features before PCA.
  • Deep Learning Model (MLGRU-SAM): The fused features are fed into a custom deep learning model. This model consists of a Multi-Layer Gated Recurrent Unit (MLGRU) for capturing temporal dependencies and a Self-Attention Mechanism (SAM) layer to further refine feature extraction and focus on the most relevant information for chatter detection. This architecture is compared against traditional machine learning and other deep learning models.

Datasets

The study utilizes experimental data collected from a custom blade whirling milling setup. The workpiece material is 5052-H112 aluminum alloy. Data was acquired during two distinct machining stages: roughing and finishing. For each stage, 60 data samples were collected, with each sample comprising four-channel sensor signals (X, Y, Z vibration and acoustic emission). These samples were segmented into equal lengths (2 seconds, 40,000 points) and then further subsampled to 3000 points for model input to reduce computational load. The paper does not introduce a new benchmark dataset but rather uses its own experimental data for validation.

Key Insights and Ideas

  • Blade Edge Vulnerability: The study confirms that the blade edge regions (leading and trailing edges) are significantly more prone to chatter, especially during rough machining. This is attributed to their lower rigidity and greater sensitivity to external excitations.
  • Roughing vs. Finishing Chatter: Chatter appears at lower feed rates during roughing compared to finishing. This is linked to the substantial material removal in roughing, which alters the workpiece's rigidity and makes it more susceptible to vibrations.
  • Multi-Sensor Synergy: Combining vibration and acoustic emission signals provides a more comprehensive understanding of the machining state than using a single sensor type. Different sensors capture distinct aspects of chatter, and their fusion enhances detection accuracy.
  • The Power of Attention: The integration of a Self-Attention Mechanism (SAM) within the MLGRU model proved crucial for optimizing feature selection and improving the model's ability to discern subtle chatter patterns.

Key Results

The proposed MLGRU-SAM model demonstrated strong performance:

  • High Recognition Accuracy: An average recognition accuracy of 89.16% was achieved across both roughing (86.66%) and finishing (91.67%) machining datasets.
  • Real-Time Detection: The model achieved an impressive average response time of less than 0.4 seconds, making it suitable for real-time chatter monitoring and potential suppression during machining.
  • Feature Fusion Effectiveness: PCA-based feature fusion significantly improved computational efficiency and model performance compared to using raw or unfused features.

Limitations and Future Work

  • Limitations: While the model shows high accuracy, the paper acknowledges that the data acquisition and processing steps are still somewhat time-consuming. The current study focuses on detection, not active suppression.
  • Future Work: The authors suggest future research should focus on developing real-time chatter suppression strategies to further enhance machining stability and product quality. They also aim to explore more advanced deep learning architectures and potentially integrate more sensor modalities for even more robust detection.

Summary updated

Abstract

Bladed components are essential in engines and propulsion systems, but their thin structures, complex geometries, and significant material removal rates during machining make them challenging to manufacture. This study investigates the chatter phenomenon in blade whirling milling, a promising method for improving machining efficiency. Multi-sensor signals, including vibration and acoustic emission signals, are collected during roughing and finishing machining. Time-domain, frequency-domain, and time-frequency features are extracted, filtered, and fused using principal component analysis (PCA) to retain relevant information while ensuring computational efficiency. The features are then input into an MLGRU-based chatter recognition model, incorporating a self-attention mechanism (SAM) for enhanced performance. The experimental results show that the proposed model achieves an average recognition accuracy of 89.16%, with a response time within 0.4 s, reflecting its effectiveness and timeliness in chatter detection. The findings also validate that the blade edge regions are more prone to chatter, especially during rough machining, due to their lower rigidity and greater sensitivity to external excitations.
BibTeX
@inproceedings{li2025data,
  author = {Xinyu Li and Riliang Liu and Zhiying Zhu},
  title = {{Data Acquisition and Chatter Recognition Based on Multi-Sensor Signals for Blade Whirling Milling}},
  booktitle = {Machines},
  year = {2025},
  doi = {10.3390/machines13030206},
}
Figure