Elsevier

Journal of Manufacturing Systems

Volume 57, October 2020, Pages 298-310
Journal of Manufacturing Systems

Physics guided neural network for machining tool wear prediction

https://doi.org/10.1016/j.jmsy.2020.09.005Get rights and content

Highlights

  • A new physics guided neural network model is proposed for tool wear prediction.

  • The physics guided loss function eliminates the physical inconsistency.

  • The model effectiveness is experimentally validated via performance comparison.

Abstract

Tool wear prediction is of significance to improve the safety and reliability of machining tools, given their widespread applications in nearly every branch of manufacturing. Mathematical modelling, including data driven modelling and physics-based modelling, is an important tool to predict the degree of tool wear. Howerver, the performance of conventional data driven models is restricted by the absent representation of physical inconsistency. The physics-based models usually fail to consider the complex tool cutting conditions and dynamic changes of physical parameters in practice. To address these issues, a novel physics guided neural network model is presented for tool wear prediction. Firstly, a cross physics-data fusion (CPDF) scheme is proposed as the modelling strategy to fuse the hidden information explored by a physics-based model and a data driven model. Secondly, the information hidden in the unlabelled sample is explored by the physics-based model of tool cutting, inspired by semi-supervised learning. Thirdly, a novel loss function which takes the physical discipline into account is proposed to evaluate the physical inconsistency quantitatively. The advantage of the developed method is that it explores sufficient information from both physics and data domains to eliminate the physical inconsistency existing in conventional data driven models.

Introduction

In the manufacturing industry, machine cutting is a common part forming method and plays an important role in advanced manufacturing. Statistics shows that about 70 %–80 % of the parts are formed by cutting, and the breakdown of cutting tools attributes up to 20 % of machine downtime [1]. Thus, tool wear prediction takes an essential role in the manufacturing industry.

Much effort has been put on constructing the mathematical modelling of tool wear prediction, including data driven models and physics-based models. Backtracking in the last century and early period of 21 st century, the sea of physics-based tool cutting models was proposed to describe the tool cutting process. A novel framework was established for constructing the relationship between the cutting force and the clearance wear [2]. Jawahir et al. [3] proposed a parametric method utilizing chip-grooved features for assessing the tool-wear. Altintas et al. [4] investigated the mathematical relationship among the cutting force, velocity and acceleration in the cutting process. Bhuiyan et al. [5] demonstrated the positive relationship between the amplitude of vibration in the directions of x, y, z, and cutting depth and speed. The finite-element analysis was employed for evaluating the distributions of all the physical variables in the cutting process [6]. Qin et al. proposed a physics-based model for tool cutting force of ultrasonic-vibration-assisted grinding (UVAG), which provided the novel method to estimate other physical parameters in the UVAG [7].

Recently, data explosion results in the wide use of data science as an indispensable tool for knowledge discovery. Specially, research focusing on tool wear prediction has gradually turned to classic machine learning models, especially deep learning models. Ko et al. [8] combined unsupervised learning and adaptive time series modelling algorithms to monitor the dynamic parameters in the cutting process, and predicted the tool wear. Karandikar et al. [9] presented an iterative method to update the tool life curve continuously based on Bayesian inference. Corne et al. [10] conducted tool breakage prediction by means of neural network using the spindle power data. Wang et al. [11] utilized particle filter to bridge the gap between vibration samples and tool wear. The particle filter method was then combined with a mathematical model of tool cutting to evaluate the degree of the tool wear based on sensor signals [12]. Moreover, Wang et al. [13] presented an augmented particle filter on the basis of virtual sensing with support vector regression (SVR) to investigate uncertainties associated with the tool degradation process. The individual use of a data driven model or a physics-based model usually encounters challenges when dealing with a physical problem. Conventional data driven models are short of physical consistency due to the absent representation of physical issues. In addition, the performance of a data driven model is limited by the volume of the training sample. For pure physics-based models, enormous cost may be introduced by manually adjusting parameters in practical application due to the dynamic and complex working conditions.

The physics guided data driven (PGDD) method proposed in recent years makes full use of data and physical knowledge. Thus, it can explore sufficient information from both data and physics domains with improved generalization and robustness. A novel framework combining the numerical approach and finite element analysis was proposed to capture the complexity of thermal transformations, and a reliable software was also developed [14]. An automatic method of nondestructive testing and evaluation in real thermography was developed to discriminate the various defects with high resolution and detectability on the basis of the PPMM [15]. The further study was extended into the fusion of the information in multiple physical domains [16]. The physics-data system for smart grid integrating the physics-based model and the ECD algorithm was proposed to prevent the attack from the fake data, and the performance of the proposed method of the anomaly detection was justified [17]. Another interesting subject discussed the cyber-physical system based on digital twin to solve the automatic configuration in the smart workshop and interpreted the interconnection between the physics and digit [18,19]. Loizou et al. [20] developed a broaching tool wear characterization system using automated image cropping and digital imaging processing tools without any manual intervention required. By combining a physics-based model with a data-driven model, a novel physics-informed machine learning approach was proposed to predict material removal rate in chemical mechanical planarization [21]. However, prior research on PGDD has several potential concerns: 1) When dealing with physical problems, a profound understanding of the physical process is required as the training guidance in PGDD. 2) The PGDD model performs unsatisfactorily in the absence of labeled samples. 3) There exists no standard method to refer to when designing the structure of the model that integrates data and physics effectively.

Based on the above discussion, a novel physics guided GRU model (PGGM) integrating a data driven model and physical knowledge is proposed for tool wear prediction. It includes a novel fusion method, cross physics-data fusion (CPDF), proposed as the modelling strategy of PGGM, inspired by the feature level-based data fusion [22]. As illustrated in Fig. 1, CPDF is utilized to fuse information extracted from the both domains of physics and data. The principle of CPDF is to map the sensor signals into the target space through both channels of physics-based model and data driven model, and the mapping results are physical prediction and data prediction. The data prediction and physical prediction are fused using regression analysis to obtain the final prediction.

In summary, the physic-based models are integrated with the data-driven models following three aspects: 1) feature fusion, 2) data augmentation, and 3) novel constraint loss term. This paper sheds some light on the effective combination of physic-based models and data-driven models to help the research community to embrace these two different methodologies. The intellectual merits of this paper rest on the following:

  • a)

    A novel PGGM is constructed for continuous prediction of tool wear depending on CPDF, which integrates data mining and physical principles appropriately. Moreover, the hidden relationship between the input and output is fully explored.

  • b)

    Inspired by semi-supervised learning, the information hidden in the unlabelled samples is explored by the physics-based model of tool cutting. On the one hand, the restriction of the lack of labeled samples is handled by the usage of unlabelled samples. On the other hand, the information hidden in the unlabelled samples is explored and thus improves the generalization and robustness of PGGM.

  • c)

    A novel loss function termed physics-guided loss function is proposed on the basis of the common physics knowledge that tool wear increases with the number of tool cuts. Without profound understanding of tool cutting, the physics-guided loss function can evaluate the physical inconsistency quantitatively to improve the predicting accuracy of PGGM.

The rest of the paper is composed as follows. The methodology of the physics-guided data driven model and the employed data driven model is introduced in Section 2. The detailed model construction is illustrated in Section 3 including the modelling process of the physical based model and the component of the deep learning model. The practicality of the proposed method is experimentally proved on the basis of the dataset collected from the CNC milling machine in Section 4. The contribution of the proposed model is discussed in Section 5. Finally, the conclusion is illustrated in Section 6.

Section snippets

Physics-guided data analytics

In scientific research, the physics-based model and data driven model which depend on only one of the two available information sources represent the two types of knowledge discovery. The unique strength and effect are exerted in different types of applications. One the one hand, physics-based models utilize simulation software and physics knowledge to simulate physical processes, which are well understood conceptually by following scientific principles. On the other hand, data-driven models

Overview of PGGM

An overview of PGGM for tool wear prediction is illustrated in Fig. 4. It is comprised of five steps. Firstly, local features, which are regarded as the input of the Bi-directional GRU model, are extracted from labeled samples of cutting force and cutting velocity. Secondly, based on empirical equation of tool cutting, the tool cutting physical model is fitted using the cutting forces and actual wear values. The unlabeled samples of tool cutting force are pushed into the tool cutting physical

Experimental setup

The high-speed CNC machine tool experimental platform as shown in Fig. 9 is used for verifying the effectiveness of the proposed method [33]. The data collected from muti-sensors is utilized as testing and training samples. The data consists of three sets of labeled samples (C1, C4, and C6) and three sets of unlabeled samples (C2, C3, and C5). Each labeled sample contains 315 data tables, each of which has a corresponding flank wear covering the life cycle of the tool. In addition, each data

Comparion of contribution factors

The advantages brought by the two main contributions, CPDF and physics-guided loss function, are discussed emphatically in the section. Four models are modified from PGGM to illustrate the two contributions. The detail of the four models is described as follows:

  • a

    Model A: The Bi-directional GRU, as mentioned in Section 4.

  • b

    Model B: The structure setting of model B is consistent to the PGGM except that the loss function is mean square error rather than the physics-guided loss function.

  • c

    Model C:

Conclusions

In this paper, PGGM is proposed to address the insufficiency of the physics-based model and data driven model for tool wear under complex operating conditions in the high-speed CNC machine. PGGM innovatively solves the difficulty of tool wear prediction owing to the complexion of the wear condition and the dynamic changes of physical parameters. The model can be easily extended to any specific neural network structures. The following conclusions can be drawn:

  • 1)

    The modelling strategy, CPDF, is

Declaration of Competing Interest

The authors report no declarations of interest.

Acknowledgments

This research acknowledges the financial support partially provided by Natural Science Foundation of China (No.U1862104), National Intelligent Manufacturing Comprehensive Standardization Project: remote operation and maintenance standard development for CNC machine tools and verification platform construction and Science Foundation of China University of Petroleum, Beijing (No. ZX20180008 and No.2462020YXZZ052).

References (36)

  • J. Loizou et al.

    Automated wear characterization for broaching tools based on machine vision systems

    J Manuf Syst

    (2015)
  • T. Yu et al.

    Predictive modeling of material removal rate in chemical mechanical planarization with physics-informed machine learning

    Wear

    (2019)
  • Z. Wang et al.

    Crack characterization in ferromagnetic steels by pulsed eddy current technique based on GA-BP neural network model

    J Magn Magn Mater

    (2020)
  • S.K. Choudhury et al.

    Tool wear prediction in turning

    J Mater Process Technol

    (2004)
  • C. Hu et al.

    Three-dimensional analytical modeling of cutting temperature for round insert considering semi-infinite boundary and non-uniform heat partition

    Int J Mech Sci

    (2019)
  • J. Weng et al.

    A novel approach to thermal modeling based on three-dimensional analysis in turning Inconel 718 with round insert

    J Mater Process Technol

    (2019)
  • T.R. Ko et al.

    Cutting force model for tool wear estimation

    (1989)
  • M. Kuttolamadom et al.

    High performance computing simulations to identify process parameter designs for profitable titanium machining

    J Manuf Syst

    (2017)
  • Cited by (135)

    • A New Semi-supervised Tool-wear Monitoring Method using Unreliable Pseudo-Labels

      2024, Measurement: Journal of the International Measurement Confederation
    View all citing articles on Scopus
    View full text