Machining is a supervised material-removal technique that cuts a material (or a work piece) to a specific final form and size. Parts made by removing raw material using cutting tools are still used to build manufactured objects ranging from nails to planes. Several machines like Lathes, Drill Presses and Milling Machines etc. are used to machine a material. In these machines, the most important and vulnerable component is the cutting tool. Given the steady environment, the quality of machining is heavily reliant on consumables such as cutting tools. Hence, it is of utmost importance that there exists a tool condition monitoring (TCM) system that detects worn out tools in time and replaces them immediately [1].
There are two ways to measure tool wear: directly and indirectly. Direct TCM techniques mostly use optical instruments, which may include a camera, a microscope (such as a tool maker's microscope), artificial lighting, and image processing algorithms. This method, however, is subject to environmental factors like lighting, and also contributes to higher cost of production owing to the fact that it requires expensive equipment, not to mention the huge time wasted to shut down the machines and inspect the tool wear. Modern sophisticated machining systems must be able to automatically detect tools that have been subjected to wear or damage. This can improve machining precision while also lowering production costs. Indirect TCM techniques use one or more sensors to acquire signals associated with tool wear conditions, such as acoustic emission (AE), sound sensor, cutting force, vibration, and current signals, and then use artificial intelligence (AI) or neural network (NN) classifiers to estimate tool wear condition [2]. In our study, we particularly focus on machining of Nimonic C-263 super alloy which is widely used in situations that demand high temperature resistance. Also, due to its ability to withstand high load applications, it’s being used in components used in both aircraft turbine engine as well as applications involving land-based turbines. While machining, Nimonic alloys undergo work and age hardening due to high force and temperature. Because of their high heat resistance, high operating temperatures, toughness, hardness, low thermal conductivity, strength to weight ratio, chemical property to react with tool materials, and creep resistance, nickel-based super alloys are often recognized to be difficult to process materials. This results in rapid deterioration of most cutting tools which can cause poor dimensional accuracies, residual stress etc. [3].
The demand for Nickel-based alloys is increasing due to their superior mechanical and chemical properties, across many industries. At low weight, super alloys exhibits good yield strength and tensile strengths and reacts resistive against corrosion in acidic and hot, humid environments. These properties made super alloys become popular in machine shops.
Nickel based alloys have a gummy machining behaviour due to high ductility and work hardening because of an austenitic structure. Moreover, Super alloys designed with a focus on high temperature applications and maintain their strength properties even at chip formation temperatures during machining, and have low thermal conductivity in comparison with steel and other materials. In addition to these strength properties of super alloys, age hardening behaviour make super alloys more difficult to machine than steel. Work hardening and high heat generation during machining of Nickel based alloys, causes tools to fail quickly. This may ended up with scrapped parts and broken tools.
Table 1
Grouping of Nickel based alloys beads on machinability
Super Alloys Group | Characteristics of Group | Alloy | UNS # | Ni | Cu | Fe | Cr | Nb | Other |
A | Alloys containing essentially nickel for caustic alkali chemical and electrical applications. Lowest strength and work hardening of the nickel alloys. Exhibits gummy behaviour in the annealed condition; hardenable only by cold working which provides the best condition for machining. | Alloy 200 Alloy 201 Alloy 205 Alloy 212 Alloy 222 | N02200 N02201 N02205 | 99.6 99.6 99.6 97.0 99.5 | | | | | C 0 08 C 0.01 Mg 0.04, C 0.04 Mn 2.0, C 0.05 Mg 0.075 |
B | Nickel-copper and nickel-iron alloys for sulfuric acid and corrosion, and electrical applications. Have higher strength work hardened than Group A alloys. Most alloys cannot be hardened by heat treatment and best machining is obtained in cold-drawn or cold-drawn and stress-relieved conditions. | Alloy 400 Alloy 401 Alloy 450 Alloy 36 Alloy K Alloy MS 250 | N04400 N04401 C71500 K93600 K94610 | 66.5 42.5 30.0 36.0 29.5 19.0 | 31.5 55.5 68.0 | 0.3 0.7 64.0 53.0 76.0 | | | |
C | Mainly nickel-chromium and nickel-iron-chromium alloys for acid and high temperature corrosion applications. Similar in mechanical properties to austenitic stainless steels except for greater high temperature strength. These alloys are best machined in the cold-drawn or cold- drawn and stress-relieved conditions, | Alloy 600 Alloy 690 Alloy 601 Alloy 825 Alloy DS Alloy 330 Alloy 20 Alloy 800 Alloy 800HT Alloy 802 Alloy 270 Alloy K-500 (unaged) Alloy 75 Alloy 86 | N06600 N06690 N06601 N08825 N08330 N08020 N08800 N08811 N08802 N02270 N05500 N06075 | 76.0 61.0 60.5 42.0 37.0 35.5 35.0 32.5 32.5 32.5 99.98 65.5 80.0 64.0 | 2.2 3.5 29.5 | 8.0 9.0 14.0 30.0 41.0 44.0 37.0 46.0 46.0 46.0 1.0 | 15.5 29.0 23.0 21.5 18.0 18.5 20.0 21.0 21.0 21.0 19.5 25.0 | | |
D-1 | This group consists of a limited number of age-hardened alloys in the solution annealed condition. These alloys are relatively easily machined. | Alloy 301 Alloy 925 Alloy 902 | N03301 N09925 N09902 | 94.0 42.0 42.5 | 22 | 32.0 49.0 | 21.0 5.3 | | |
D-2 | This group consists of Group D-1 alloys in the age-hardened condition, most other age-hardenable alloys in both the solution annealed and hardened conditions, and some highly solution strengthened alloys. They contain strong solution straighteners and hard abrasive precipitates which make machining difficult. These alloys should be rough machined in the solution annealed condition and then finish machined after aging. A size contraction up to about 0.07% takes place upon aging which must be allowed for in rough machining. | Alloy 301 (aged) Alloy K-500 (aged) Alloy 902 (aged) Alloy 81 Alloy G-3 Alloy HX Alloy 625 Alloy 925 (aged) Alloy 716 Alloy 725 Alloy MA 754 Alloy 80A Alloy 718 Alloy PE11 Alloy 706 Alloy PE16 Alloy C-276 Alloy 751 Alloy X-750 Alloy 901 Alloy 617 Alloy 263 Alloy 105 Alloy 90 Alloy PK50 Alloy 115 Alloy B-2 Alloy 903 Alloy 907 Alloy 909 | N05500 N09902 N06985 N06002 N06625 N90925 N07716 N07725 N07754 N07080 N07718 N09706 N10276 N07751 N07750 N09901 N06617 N07263 N07090 N10665 N19903 N19907 N19909 | 94.0 65.5 42.5 67.0 44.0 47.5 61.0 42.0 57.5 57.0 77.5 76.0 54.0 39.0 42.0 43.5 | 29.5 2.0 2.2 | 1.0 49.0 19.5 18.5 2.5 32.0 9.0 1.0 18.5 34.0 36.5 5.5 7.0 7.0 34.0 1.5 36.0 1.0 41.5 42.0 42.0 1.2 | | | |
E | A special Alloy 400 designed to be free machining for high production on automatic screw machines. | Alloy R-405 | | 66.5 | 31.5 | 1.2 | | | Mn 1.1, S 0.04 |
Due to heat generation during machining of nickel alloys, stringy chips causes built up edge formation round the tool cutting edge, which results in increased friction, further increases work hardening. An example of BUE is seen in the image below.
On the above tool shown in Fig. 1, chips from the workpiece (Inconel 718) have welded onto the cutting edge, severely decreasing the tool’s effectiveness.
It is important to understand how accurate and reliable each signal is in being used to predict the current wear state. Usually, a TCM system consists of several components including the sensors that detect the signals as aforesaid. It can either be a single sensor or a combination of multiple sensors. These sensors give out relevant signals which may or may not need further amplification. Signal processing techniques are then implemented to obtain the signal features from the signals received. These features are then sent to a prediction tool which, based on its architecture, techniques and training, determine whether the tool is in a usable state [4].
It is also important to note that not burdening the prediction tool with too much information is almost as important as the reliability of the sensors. On the flip side, finding an effective prediction system that can accurately output the tool wear, despite discrepancies in signal features is also worth the effort. Hence, it is important to keep modernizing the techniques and systems involved in TCM as technology is improving day by day. In this study, we aimed at creating such a system that can tackle the aforementioned issues commonly faced by existing TCM systems.
1.1 Literature Review
Being a relatively new field, TCM hadn’t been studied significantly before 2000. Having said that, there were in fact a few studies in the 90s that have immensely influenced current studies in that area. In 1990, S. Rangwala et al. conducted a study on extracting relevant features from noisy signals and from AE sensors and found that the major contributor to AE signal changes were crater wear and not flank wear [5]. Later in 1995, a study conducted by Aitchison and David Robert implemented laser backscatter technique, specifically to monitor cutting tool condition [6]. The research required the creation of a system that could manage a variety of cutting tools and deliver them to the laser scanning head in the correct position. Initially studies were mostly focused on turning operations. But later on, the focus was shifted to other machining tools like milling, drilling etc.
The efficiency of several parameters for tool condition monitoring (TCM) during milling processes was explored by Dong Jianfei in 2004. The system used two unique neural network approaches, Bayesian interpretations for support vector machines (BSVM) and automated relevance determination (ARD), to extract features from the cutting force signals [7]. To establish an optimized structure for the ANN tool condition monitoring system, a large variety of network architectures were tested systematically by Soo-Yen Lee in 2006 [8]. Pan Fu et al. successfully implemented a TCM procedure using advanced B-spline neurofuzzy networks. When compared to standard neural networks, such as BP type ANNs, this model has the benefits of swift convergence and localized learning capabilities [9].
Choudhury et al. explains in a study conducted in 2010 that the frequency range of the AE signal is much greater than that of machine vibrations and ambient sounds, and it does not interfere with cutting operations, which is a major advantage of using Acoustic Emissions to monitor tool condition. [10]. Later on in 2011, Tian Ran Lin et al. created a complete diagnostic tool for rotating machinery condition monitoring and systematic analysis [11]. F Čuš and U Župerl in the same year worked on improving sensor utilization factor (SUF) by minimizing the number of sensors. They presumed that cutting force signals provided the most valuable information for tool wear monitoring and also in the study, flank wear was found to be most sensitive to the thrust cutting force component [12.
Huang Sheng in 2012 investigated methods of condition monitoring for ball-nose end milling, with a focus on sculptured surface machining applications [13]. In the same year, Krzysztof Jemielniak et al. proposed a methodology that involved low-pass filtering of signal features to model their dependence on the used-up portion of tool life, removed comparable signal features through SF usability validation utilizing a determination coefficient between the feature and its low-pass-filtered estimate and SF repeatability [14]. In the following year, K.Venkata Rao et al. conducted research to determine the impact of cutting parameters on workpiece vibration, machined surface roughness, and metal removal volume in boring steel (AISI1040). For online data collecting, a laser Doppler vibrometer (LDV) was employed. ANOVA analysis was also performed [15]. Subsequently, Wennian Yu et al. in 2014 discovered that adding additional characteristics that share a lot of data does not improve prediction performance, but it does raise the strain on decision-making algorithms. Moreover, using characteristics with poor fitness values may worsen the prediction [16]. Fitness values of parameters were determined through correlation coefficient and statistical overlap factor (SOF) to determine this. Also, Lu Chen et al. trained a hidden markov model (HMM) using table vibration and AE, as well as spindles [17]. Yongfeng Hou et al. conducted a study with a particular approach where the friction force and the cutting force of the milling tool were treated separately. A calibration experiment was used to calibrate the model coefficients, and the influence of tool flank wear on milling force vectors was studied [18]. W. Li et al. worked on end milling with PVD coated tools to investigate the influence of tool wear on surface integrity and its impact on fatigue performance. Surprisingly, he found out that surface roughness was reduced as tool wear increased [19]. In a paper by Nitin Ambhore et al. the stages involved in TCM, as well as feasible experimental setup, sensors, and equipment necessary, were discussed in depth. Details on the efficacy of various factors are also provided [20].
Later in 2016, Danko Brezak et al. introduced the Radial Basis Function NN, because of its capacity to learn in one step and adjust the hidden layer structure in a simple and rapid manner. As an alternative to cutting force signals, vertical or Z-axis feed drive current (IZ) and main spindle current (IMS) were used as input signals [21]. A year later, Virginija Gylienė and Valdas Eidukynas utilized other numerical techniques in FEM, including Lagrangian, Eulerian, and Arbitrary Lagrangian-Eulerian [22]. Doriana M et al. used vibration signals and a cognitive model like ANNs to provide a method for assessing the wear state of a dresser in their research. Its purpose is to measure grinding wheel wear during dressing operations and correlate it with measured sensor data like as force, current, acoustic emission (AE), vibration and other variables. Back Propagation Neural Networks were chosen to predict both the coded output and the dresser wear values [23]. Sebastian Bombiski et al. also suggested methods in his research, for fully automatic identification of actual cutting, removal of air cutting, selection of typically stable signal segments representative of the tool state, and reduction of signal data excess. The steady state components of the sensor signals are utilized to calculate signal characteristics for tool wear monitoring. Without the involvement of the user, threshold values were determined [24].
In 2017, Karali Patra et al. employed a tri-axial accelerometer, a data collection and signal processing module, and an ANN to develop a TCM system for micro-drilling [25]. The ANN was created using the MATLAB neural fitting tool to forecast the drilled hole number by fusing the RMS values of vibration in three mutually perpendicular directions, as well as the spindle speed and feed parameters. Marinela Inţă et al. employed neural networks to assess tool wear in metal cutting on annealed C45 steel in another study [26]. For preliminary training of the NN, the values of force, temperature, and tool wear prediction derived from the simulation with FEM were used. The research demonstrated that, in addition to stresses, noise, and vibrations, including tool temperature as an input component in neural network training improves the accuracy and stability of the findings. J.Barreiro et al. examined acoustic emission and vibration signals to see if they could link the two to tool wear [27]. The spectrum was statistically analysed by separating it into four equal bands and calculating the kurtosis and skewness under various wear situations and spindle speeds.
Subsequently in 2018, Ari Setiawan et al. constructed a model of the relationship between temperature, vibration, and power consumption and the cutting tool lifetime, taking into account wear on the cutting tools [28]. The correlation coefficient was used to indicate the relationship between input and output variables. Sandra Sovilj-Nikić et al. found that covering a cutting tool with a suitable material extends its tool life by 60.6 percent when compared to hob milling tools with inserted TiN coated combs [29]. Huan Xu et al. in their research, presented a real-world gun drilling experimental dataset [30]. As a case study, the Machining Centre was employed. This research proposed a GRU-based TCM technique for estimating tool wear based on temporal aspects of historical sensory data. Force signals were used to determine flank tool wear. A multi-stage TCM system based on Multi-GRU was suggested by Huan Xu et al. in 2019 [31]. Under varied cutting circumstances, a gun-drilling experiment was carried out. A piecewise linear assumption is applied in this situation to interpolate the wear value between samples.
In 2020, several researches were conducted on tool condition monitoring. Łukasz Huchel et al. carried out an experiment using carbide end mills and aluminum as the stock material [32]. Yuqing Zhou and Weifang Sun carried out an experiment in which motor current sensors were used because they are cheap, easy to install, and have no effect on the milling process [33]. Clayton Cooper et al. developed a statistical model for milling acoustic signals (AS) as well as a ML based classification model based on a 2D CNN [34]. The signs included chatter, chip formation that differed from that of a fresh tool, and "squealing" while in contact with the workpiece. Ritesh Upase and Nitin Ambhore conducted research to study the impact of cutting parameters on cutting tool vibration during hard turning using a PVD coated carbide tool [35]. The Dino Lite Digital microscope was also used to measure tool wear. Berend Denkena et al. found a method for evaluating an initial set of features online with the goal of detecting a single non-sensitive feature if it exists by analyzing long-term patterns in repetitive machining operations [36]. The method's performance was assessed by comparing its results to the humanly established ground truth.
Later in 2021, Jiaqi Che et al. used a 3D laser scanner to create a 3D scanning model of the milling tool for finite element analysis, as well as a simulation model based on reverse engineering. [37]. The weight on the bit and the rotation speed of milling tools has an effect on and interact with one another, determining the milling tool's working performance. The Johnson-Cook model was employed as the material's model of constitution. Qinsong Zhu et al. used the Johnson-Cook model to acquire data for training [2]. A generative adversarial network (GAN) is used to create new samples that are similar to both the simulated and empirically measured samples. The problem of a reduced sample size is claimed to be solved as a consequence. To monitor the dynamic milling process, Kai Guo et al. developed a wireless vibration detecting tool holder with built - in wireless vibration sensing [38]. A tool condition monitoring technique was created based on the HE properties deduced from the vibration data acquired by the embedded wireless vibration sensor tool holder system.
1.2 Research Blank Spots
-
After extensive literature survey, the following blank spots were identified.
-
Very few studies done on determining the best sensors to be used to gain signals for a particular environment. Most studies follow previous studies by making use of a large number of input parameters. Though some have studied the influence and correlation between those input parameters and tool wear, almost no study has been done to find out correlation between the input parameters. This is important as increasing the number of input parameters doesn’t necessarily mean a more accurate result.
-
Another area that requires research is selection of optimal prediction tools. Many use AI models like linear regression. Very few have actually tried to compare prediction accuracies of different machine learning tools.
-
Most studies are still being conducted on materials like mild steel, aluminium etc. even though these have been studied for a long time.
1.3 Objectives
The major focus of this work is
-
To find out the most commonly used, but sensitive parameters towards tool wear in an environment like a high precision machine shop.
-
To determine the influence of each individual parameters on another.
-
To eliminate input parameters that have high correlations with each other in order to minimize the burden on the prediction tool (Neural Network).
-
To find out more modern Neural Network architectures are in fact more effective than conventional prediction tools.
-
To create a TCM systems to compare prediction results using harder materials like super alloys.
-
To create a GUI that lets users input data into the neural network and to display the output to the users.