A cyber physical system for tool condition monitoring using electrical power and a mechanistic model

Highlights

• Tool wear increased cutting power in experiments by 102–213 %.

• Mechanistic model provides a reasonable indicator for tool wear (R² = 0.801).

• Requires prior testing to establish coefficients and knowledge of cutting parameters.

• Cyber Physical milling model outlined to determine cutting parameters from NC code.

Abstract

Tool Condition Monitoring (TCM) systems, which aim to identify tool wear automatically to avoid damage to the machined part, are often not suitable for industrial applications due to: (i) sensing methods which can be expensive and invasive to install and (ii) testing regimes that only evaluate a narrow operating window under controlled conditions. To combat these issues, the research outlined in this paper explores the feasibility of TCM using electrical power consumption of an industrial Computer Numeric Control (CNC) cutting machine in combination with a mechanistic model for end milling operations. End milling of aluminium 6082 was performed over a range of cutting parameters (i.e. spindle speed, feed rate, tool diameter) and repeated with increasing levels of tool wear to ascertain the suitability of this approach. Reasonable correlation between the predicted and observed tool wear was found using the mechanistic model (R² = 0.801), however variance between the cutting parameters highlights the limitations of the predictions. To combat this variance, an averaging window is taken over the course of a cutting program to reduce the overall error. A concept TCM system is then proposed, utilising cyber-physical models of the milling process to determine the width and depth of cuts for complex geometries which change in real time, with the challenges of implementing such a system discussed.

Introduction

Computer Numeric Control (CNC) has enabled the automation of machining processes, such as milling, turning and drilling, by programming machining operations and parameters such as paths, spindle speed and feed rates. However, CNC machines generally have limited or no knowledge of their working environment, meaning that they cannot detect, adapt or deviate from a cutting program to avoid issues such as chatter, tool wear and breakage, tool deflection or collisions without the intervention a human operator (Cao et al., 2016).

Cutting tools are consumable components of the CNC machine which degrade and wear with use through mechanisms such as abrasion, adhesion, diffusion, fatigue and chemical wear (Altintas, 2000). Tool wear can significantly impact the machining process of a part, reducing the quality of the surface finish and potentially damaging the part, resulting in either additional rework or the need to scrap the part (Pimenov et al., 2018). Research into Tool Condition Monitoring (TCM) systems, where sensing, decision making and control strategies are introduced to detect the onset of tool wear, has therefore attracted a significant amount of interest from the academic community with a number of review articles listed (Cao et al., 2016; Teti et al., 2010; Siddhpura and Paurobally, 2013; Ambhore et al., 2015; Abellan-Nebot and Romero Subirón, 2010; Visariya et al., 2018). However, key challenges remain, including the development of TCMs which can be cost effective to install and maintain and are robust and reliable in the presence of the variability and discontinuities of real-world machining.

Several sensing methods have been evaluated within the literature to determine tool wear indirectly including force (Nouri et al., 2015; Choi et al., 2004), vibration (Rubio and Teti, 2010), spindle power (Shao et al., 2004; Al-Sulaiman et al., 2005; Drouillet et al., 2016), acoustic emission (Patra, 2011), eddy current (Abbass and Al-Habaibeh, 2015) or combinations of these (via sensor fusion) (Chattopadhyay et al., 2006; Cuka and Kim, 2016; Zhang et al., 2016). However, no research has been published focused on investigating the feasibility of monitoring the electrical power consumption of the entire CNC machine as a method of detecting tool wear. Electrical power monitoring has advantages over some of the aforementioned approaches due to its relative low cost (Nouri et al., 2015), non-intrusive installation (Abellan-Nebot and Romero Subirón, 2010) and non-reliance upon open machine architectures to allow access and integration with existing sensor systems (e.g. spindle current sensors) (Cao et al., 2016), which may also cause interoperability and data veracity issues when deployed across different machines.

Developing robust methods of interpreting sensor data to infer tool wear levels causes additional challenges due to the variability and discontinuous nature of cutting processes such as end milling. These variabilities are caused by wide operating and transient cutting parameters, such as spindle speed, feed rate, depth and width of cut, tool type, workpiece geometry and material, discontinuous tool paths and noise from background operations influencing the measured signal (e.g. coolant pump, lighting, computers). Several black-box approaches, such as Artificial Intelligence (AI), have been used to identify tool wear with good accuracy (Chattopadhyay et al., 2006; Zhang et al., 2016; Corne et al., 2017). However, these are usually only evaluated within limited operating conditions; therefore when changes occur to the system (i.e. different machines, tools, workpiece materials) retraining of the algorithms is needed which requires further controlled testing and data collection. These approaches reduce the robustness and scalability of the system and increase the operating cost which will limit the relevance of the research to industry. On the other hand, white box approaches, such as mechanistic models, use scientific first principles to represent the real-world complexities through mathematical relationships, with examples developed to determine tool wear based upon changes in electrical spindle power (Shao et al., 2004) and force measurements (Nouri et al., 2015). Mechanistic models are useful as they can be reused over a number of scenarios by changing the model parameters to match the changes within the real world. However, it is difficult to model and measure all of the real-world complexities that may impact the tool wear measurements. This leads to an inevitable trade-off between accuracy and usability/robustness when developing models in real applications where not all parameters can accurately be predetermined.

Several studies relevant to tool wear condition monitoring are listed in Table 1. Shao et al. (Shao et al., 2004) develop a mechanistic model to predict tool wear using the electrical power of the spindle and known cutting parameters. Whilst experimental results from this study appear to show good correlation with the mechanistic model, only limited validation of the technique is demonstrated over a small range of cutting parameters and tool wear is only presented as new or worn. Drouillet et al. (Drouillet et al., 2016) present an interesting approach using neural networks to update and correct Remaining Useful Life (RUL) prediction of an end milling tool, using historical tool degradation patterns from the electrical spindle power. The approach is used to correct errors in RUL prediction which is caused by stochastic tool wear degradation. However, all predictions are based upon nominal spindle power (i.e. cutting power of a new tool), so at present the approach is unable to accommodate changing cutting parameters. Nouri el al. (Nouri et al., 2015) develop a real-time tool condition monitoring method to track changes within cutting force coefficients which are independent of cutting conditions and correlated with tool wear within a mechanistic model. A Cumulative Sum (CUSUM) control chart is used to determine the transition of the cutting coefficient from gradual wear to the failure region. Wang et al. (Wang et al., 2019) use a combination of cutting force measurement and electrical power sensors to predict tool wear in drilling. A mechanistic model is used to predict the cutting force, whilst idle and auxiliary power demand of the motor is predicted based upon a regression model using empirical data.

This research investigates the use of electrical power monitoring of the entire CNC machine (rather than just the spindle) in combination with a mechanistic model as a low cost and process robust method to detect and determine tool wear in CNC end milling.

There are two novel aspects of this work: (i) a comprehensive assessment of the mechanistic model using only the electrical power under variable conditions (i.e. different spindle speed and feed rates), which has only been assessed under a limited set of experiments by Shao et al. (Shao et al., 2004), and (ii) the proposal of a cyber-physical system to support parameter calculation of the mechanistic model from the digital part program to enable real time implementation of a TCM system.

The structure of the paper is as follows. In section 2 the mechanistic model used to determine cutting power and tool wear is discussed. Section 3 details the test methodology used to assess the mechanistic model under variable cutting conditions. In Section 4 the results of the experiments are presented and discussed, whilst within section 5 an approach to tool conditioning monitoring combining a mechanistic model and cyber-physical system is presented and discussed. Conclusions to the research and future work are presented in Section 6.