TL;DR
This paper introduces a new, high-quality multi-sensor dataset for milling processes, encompassing varied parameters and materials. This dataset is crucial for advancing physics-informed machine learning approaches in manufacturing by providing the necessary data to accurately model milling forces and improve real-time monitoring.
What Problem Does This Paper Address?
Real-time monitoring of milling processes is vital for enhancing machining efficiency and product quality, especially for complex geometries. A key challenge is accurately modeling the relationship between milling parameters (like cutting depth, feed rate, etc.) and the sensor data generated during the process. While physics-informed machine learning offers a promising avenue by integrating physical laws, its effectiveness is hampered by the lack of comprehensive, high-quality datasets. Existing datasets are often limited in sensor types, parameter variations, or material diversity, making it difficult to build robust and generalizable models. This paper aims to bridge this gap by creating a rich, multi-sensor dataset that captures a wide range of milling conditions.
Methods
The core contribution of this paper is the creation and release of a new dataset. The experimental setup involved a 3-axis CNC milling machine. Data was collected using multiple sensors: a Kistler 9129AA dynamometer for force, a Kistler 9129AA dynamometer for vibration (though the text seems to have a typo here, likely meaning a separate vibration sensor), a GRAS 46AE noise sensor, and a CPL8100B low-frequency current probe.
The data collection involved two main experimental designs:
- Full Factorial Design: 100 samples were collected using an "all factors" method, varying cutting depth, cutting width, spindle speed, feed rate, and workpiece material (aluminum alloy 7030).
- Slot Milling Experiments: 15 samples were collected specifically for slot milling on two different workpiece materials (aluminum alloy 7030 and CK45 steel). These were used to calibrate mechanical milling force coefficients.
Signal processing involved removing unstable cutting-in and cutting-out stages, retaining 50,000 data points for stable cutting. Coordinate systems for force and vibration signals were unified for consistency.
Datasets
The paper introduces the JUST_CAS_milling dataset.
- Total Samples: 115
- Sensor Data: Force (X, Y, Z directions), Vibration (X, Y, Z directions), Noise, and Spindle Current (U, V, W phases).
- Milling Parameters: Cutting depth, cutting width, feed rate, spindle speed, and workpiece material.
- Workpiece Materials: Aluminum alloy 7030 and CK45 steel.
- Data Format: .xlsx (Microsoft Office EXCEL File).
- Data Availability: Links are provided to Zenodo, OPENICPSR, and Mendeley Data repositories.
This dataset is presented as a significant improvement over existing public datasets (like NASA, PHM, SDU-QIT, NUAA) by offering a broader range of signals and parameters, particularly for calibrating milling force coefficients.
Key Insights and Ideas
- Bridging the Data Gap: The most significant insight is the creation of a comprehensive dataset that directly addresses the lack of high-quality data for physics-informed machine learning in milling. This is crucial for moving beyond purely data-driven or purely physics-driven models.
- Multi-Sensor Fusion Potential: The dataset integrates force, vibration, noise, and current signals. This allows for exploring how fusing information from these diverse sensors can lead to more robust and accurate milling process monitoring, especially for conditions where direct force measurement is difficult or expensive.
- Calibrating Physical Models: The inclusion of specific slot milling experiments designed for calibrating milling force coefficients is a key methodological insight. This targeted approach is essential for improving the accuracy of physics-based models.
- Material Influence: The use of two distinct materials (aluminum alloy 7030 and CK45 steel) allows for studying the impact of material properties on sensor signals and milling forces, which is important for developing adaptable monitoring systems.
Key Results
The paper's primary "result" is the dataset itself. While quantitative performance metrics of models trained on this dataset are not presented in this "Data in Brief" article, the paper highlights the dataset's utility:
- It contains 115 samples with detailed multi-sensor data under varied milling conditions.
- The data is described as having high accuracy due to the use of high-precision sensors and high sampling frequencies (50 kHz).
- The dataset is suitable for training predictive models for residual life, life prediction under multiple conditions, and monitoring specific parameters like cutting depth and width.
- It is particularly valuable for calibrating milling force coefficients, a task often hindered by insufficient or imprecise public data.
Limitations and Future Work
- Tool Wear: The authors acknowledge that all experiments were conducted with the same tool. This means the recorded data might be influenced by tool wear, which could affect the accuracy of models trained on this dataset, especially for longer-term predictions. Future work could involve collecting data with tools at different wear stages.
- Noise in CK45 Steel: The high strength of CK45 steel induced stronger noises in the signals during data collection. While the dataset is provided, the authors note that this increases the workload for noise reduction, implying that further signal processing might be needed for optimal use of the noise data.
- Future Directions: The paper implicitly suggests future work by detailing the dataset's applicability. Researchers can use this dataset to develop and validate physics-informed machine learning algorithms for milling process monitoring, tool wear prediction, and fault diagnosis. The comprehensive nature of the data also opens avenues for exploring advanced signal processing and feature extraction techniques.
Summary updated