Research on Real-time Control of Machining Surface Quality Stability Based on Wear Monitoring

Liao Xiaoping, Kai Chen, Lu JuanJournal of Mechanical Engineering 2020 · DOI

TL;DR

This paper introduces a novel real-time control system for maintaining stable machining surface quality by monitoring tool wear. By leveraging historical data and a generalized regression neural network (GRNN), the system predicts tool wear and optimizes machining parameters on-the-fly, ensuring product quality remains close to customer demands.

What Problem Does This Paper Address?

In precision manufacturing, the quality of machined surfaces, particularly their roughness, is critical for a part's performance, durability, and proper assembly. However, during the cutting process, tool wear is an inevitable phenomenon that accumulates over time. This wear leads to fluctuations in cutting forces, vibrations, and ultimately, a degradation of surface quality, making it difficult to consistently meet customer specifications. Existing methods often struggle to balance control accuracy with the rapid response times needed for real-time adjustments, relying heavily on manual experience or offline analysis. This paper aims to bridge this gap by developing an intelligent, real-time system that actively monitors tool wear and dynamically adjusts machining parameters to maintain consistent surface quality.

Methods

The proposed method consists of two main components: wear monitoring and parameter optimization.

  1. Wear Monitoring:
    • Data Acquisition: Cutting force signals (X, Y, Z directions) are continuously collected.
  2. Feature Extraction: Time-domain (TD), frequency-domain (FD), and wavelet packet analysis (WP) are used to extract 28 features from the cutting force signals. This significantly reduces data dimensionality.
  3. Feature Selection: A Genetic Algorithm (GA) is employed to select the most relevant features from the initial 84 (28 features per direction) to improve computational efficiency and model performance.
  4. Wear Prediction Model: A Generalized Regression Neural Network (GRNN) is used to establish a mapping between the selected cutting force features and the actual tool wear. GRNN is chosen for its fast learning speed and strong approximation capabilities, crucial for real-time applications.
  5. Parameter Optimization:
    • Quality Prediction: The GRNN model is also used to predict surface roughness (Ra) based on tool wear, machining time, and other parameters.
  6. Error Minimization: The difference between the predicted roughness (Ra) and the target customer requirement (RGoal) is defined as an error (Z). The objective is to minimize the square of this error (Z²).
  7. Non-linear Programming (NLP): This minimization problem is formulated as a non-linear programming problem. The optimization aims to adjust machining parameters (feed rate, spindle speed, radial depth of cut) while keeping axial depth of cut constant and respecting parameter bounds.
  8. Adaptive Control: The system uses the predicted tool wear from the previous cutting pass to optimize parameters for the current pass, allowing for continuous, real-time adjustments without interrupting the machining process. The NLP problem is solved using a modified gradient descent method to avoid local optima.

Datasets

The study utilized data collected from milling experiments on TC18 titanium alloy.

  • Experimental Setup: A VDL-600A machining center was used with APMT1135 milling inserts. A YDCB-III05 three-directional force sensing system measured cutting forces. Surface roughness was measured using a surface roughness tester (Mitutoyo SJ310), and tool wear was measured using a metallographic microscope (AXIO SCOPE A1).
  • Data Generation: A total of 765 data points were generated from 9 sets of cutting tools, each undergoing 85 passes. The data was split into a training set (720 data points) and a testing set (45 data points). The training set was used to build the GRNN models for wear and roughness prediction, and the optimization model. The testing set was used to evaluate the performance of the proposed method.

Key Insights and Ideas

  • Wear as a Predictor: The research strongly validates the idea that cutting force signals, when properly processed, are highly indicative of tool wear and can be used for reliable wear monitoring.
  • GRNN for Real-Time: The choice of GRNN is a key enabler for real-time control. Its ability to learn from historical data without explicit weight training and its fast response make it suitable for dynamic adjustments.
  • Bridging Prediction and Control: The paper effectively links predictive modeling (wear and roughness) with an optimization framework (NLP) to create a closed-loop control system. This is crucial for proactive quality management.
  • Addressing the Accuracy-Speed Trade-off: The research directly tackles the common challenge in real-time control where high accuracy often comes at the cost of slow response. By using GRNN and an optimized NLP solver, the system achieves both.
  • Adaptive Parameter Adjustment: The system's ability to adjust machining parameters based on the predicted state of the tool wear in the next pass is a significant advancement for maintaining consistent quality throughout the tool's life.

Key Results

  • Wear Prediction Accuracy: The GRNN model achieved a high correlation (R=0.983) between predicted and actual tool wear.
  • Roughness Prediction Accuracy: The roughness prediction model also demonstrated excellent performance with a correlation of R=0.988.
  • Quality Stability: The optimized machining parameters consistently kept the surface roughness close to the target value (RGoal), with a maximum deviation of only 0.078 µm, significantly outperforming the raw predicted and actual roughness values which showed larger fluctuations.
  • Response Time: The system demonstrated rapid decision-making, with control decisions being made within a minimum of 12 seconds, allowing for continuous machining without pauses.
  • Comparative Performance: In comparative analysis, GRNN outperformed other models (RBF, SVM, GPR) in terms of Mean Absolute Percentage Error (MAPE) and Coefficient of Determination (R²) for both wear and roughness prediction. The proposed NLP optimization method also showed significantly faster optimization times compared to PSO and GA.

Limitations and Future Work

  • Generalizability: The authors acknowledge that the current method was validated on TC18 titanium alloy. Future work will focus on extending the control method to different machining processes, materials, and cutting conditions to improve its universality.
  • Integration: The next step is to integrate this intelligent control system into actual production machinery for practical implementation.
  • Abnormal Wear: The current model assumes gradual wear. Further research may be needed to handle sudden or abnormal tool wear scenarios.
  • Optimization Algorithm Refinement: While the modified NLP approach is efficient, further optimization of the algorithm or exploring alternative real-time optimization techniques could potentially reduce response times even further.

Summary updated

Abstract

摘要: 在切削加工过程中,由于磨损会随时间的变化不断积累,导致加工质量产生波动。针对这个问题,提出了在大数据背景下,基于磨损监测对加工表面质量进行实时控制的方法。通过历史数据库建立切削力信号与磨损之间的映射关系,刀具磨损反映当前加工状态,将当前加工状态的加工质量与客户需求进行比较,从而优化加工参数使加工质量尽可能接近客户需求。优化模型是利用广义回归神经网络的建模原理进行建模,使得优化问题能通过非线性规划求解,并能快速做出调控决策。研究还对TC18材料进行大量铣削试验,试验结果验证了方法的可靠性,也证明了此方法能做到对加工状态变化的快速响应。此研究解决了现有研究中不能同时保证控制精度和响应时间的问题,并为在线智能控制切削加工表面质量提供了新思路。
BibTeX
@article{xiaoping2020research,
  author = {Liao Xiaoping and Kai Chen and Lu Juan},
  title = {{Research on Real-time Control of Machining Surface Quality Stability Based on Wear Monitoring}},
  journal = {Journal of Mechanical Engineering},
  year = {2020},
  doi = {10.3901/jme.2020.11.240},
}
Figure