The analysis of tool vibration signals by spectral kurtosis and ICEEMDAN modes energy for insert wear monitoring in turning operation

Surface finish quality is becoming even more critical in modern manufacturing industry. In machining processes, surface roughness is directly linked to the cutting tool condition; a worn tool generally produces low-quality surfaces, incurring additional costs in material and time. Therefore, tool wear monitoring is critical for a cost-effective production line. In this paper, the feasibility of a vibration-based approach for tool wear monitoring has been checked for turning process. AISI 1045 unalloyed carbon steel has been machined with TNMG carbide insert twenty-one times for a total of 27 min of machining, which was a necessary amount of time to exceed (300 μm) as a flank wear threshold. Vibration signals have been acquired during the operation and then processed in order to extract a correlation between the surface roughness, tool wear level, and vibration comportment. First, spectral kurtosis has been calculated for the twenty-one performed runs signals; this step has allowed the locating of the optimal frequency band that contains the machining vibration signature, yet it did not give significant information about wear evolution. The signals have then been decomposed with ICEEMDAN and the energy of the high-frequency modes has been calculated. It has been found that the energy of the optimal frequency ICEEMDAN modes has increased in proportion to the increase of surface roughness degradation and thus, to tool wear increase. Therefore, IMF’s energy can be used for tool wear condition monitoring.


Introduction
Machining processes are the core of the majority of manufacturing industries. The principle of machining is simple yet effective for high-quality surface finish. Generally, a contact is made between a cutting tool and a workpiece in order to obtain a particular shape by material removal. However, this contact must be with specific parameters in order to get the desired results, namely, a surface roughness that is in accordance with the detailed workpiece specifications. One of the parameters that may effect changes on workpiece surface condition is tool wear level. For example, the inserts used in turning process gradually lose their sharpness over time of machining, which creates thermomechanical phenomena that alters the surface finish of the workpiece [1]. Hence, the monitoring of the tool wear and the surface roughness is important. Signal processing has been used a lot for this purpose where several parameters including cutting force [2][3][4], temperature [5][6][7][8][9], current consumption [10], image processing [11], and sound and vibration signals [1] [12][13][14][15][16][17][18][19][20] have been exploited for tool wear surveillance. Vibration analyses are often favored due to their effectiveness when taking into account the small number of required instruments, the fast response time to wear level changes, and the diversity of available tools [1]. In addition, frequency domain features are usually preferred over time domain features because the latter may be sensitive to phenomena other than tool wear. However, frequency domain features are difficult to extract because they are based on the right choice of the optimal frequency band. Therefore, a perfect approach would exploit the benefits of the two types, namely, a time-frequency approach.
One of the mostly proposed combinations between statistical indicators and time-frequency tools in the previous works is with the wavelet analysis; in [21], the authors have used discrete wavelet decomposition for grinding wheel wear monitoring, the RMS value and variance of the decomposed signals have been selected as features vector for feeding a support vector machine, and the classification accuracy has reached 99%. In another work [22], RMS and linear regression of wavelet transformed signals have been used to estimate tool wear in ramp cuts in end milling. In [23], high SNR frequency band has been extracted using derived wavelet frames, and the spectrum of the chosen band has been then injected in a 2D convolutional neural network that recognizes wear signs of milling tool during cutting operation. In [24], the authors predicted flank wear of a drilling tool with acceptable accuracy using wavelet packet transform.
Wavelet transform usually needs to be preconfigured to extract a particular feature, specifically, an appropriate wavelet mother must be preselected before the application, which is a major drawback for this type of analysis because the basic functions are fixed while the signals features may vary from case to case. Alternatively, the empirical mode decomposition is a data-driven tool, able to do the same work of the wavelet analysis, which is the separation of the high-frequency components from the lower ones, without the need to set a basis decomposition function. In many works, EMD has been proven to be effective for tool wear monitoring; in [25], the authors have used the EMD for improving the sensitivity of the measured signals energy and mean power, and predict the tool wear for turning operation. In [26], EMD has been used in order to eliminate the milling tool teeth frequency, the spindle speed components, and isolate the milling process vibration signature for cutting conditions monitoring. In [27], turning acoustic signals have been decomposed with EMD and then analyzed by Hilbert-Huang transform, and a correlation between the sound pressure amplitude of the IMFs and the tool wear has been found. EMD has been combined with wavelet analysis in [28] in order to identify the different wear modes of a turning carbide tool. In another work [29], a support vector machine has been fed with energy and energy entropy of IMFs obtained by EEMD and found to be effective for milling tool wear estimation; EEMD is an improved version of the original EMD decomposition proven to give better results in other applications [30][31][32][33][34].
In this paper, a vibration-based tool wear monitoring approach has been tested for turning operation (Fig. 1), and the enhancement of condition-based maintenance systems to collect accurate information has been set as a goal for this study.
The workpiece has been machined twenty-one times for a total of 27 min in order to reach a wear level of more than 300 μm which is the recommended threshold by ISO. The vibration signals have been captured during the operation. The insert wear and the surface roughness have been quantified after each run in order to check their correlation with the tool's vibration comportment.
Spectral kurtosis and signal energy have been used for this purpose for future application in the industrial field.

Experimental setup
The tests have been conducted on a conventional KNUTH TURNADO lathe (Fig. 3); a workpiece of AISI 1045 unalloyed carbon steel that has the dimensions of (210 mm × 60 mm) has been machined with TNMG carbide insert. The machining conditions have been maintained as follows: V c = 285m/min, f = 0.034 mm/Rev, a p = 0.15 mm. Twenty-one runs have been carried out in order to reach the wear threshold (300 μm).
During each run, the cutting tool vibration signals have been acquired using an OPTOMET scanning laser Doppler vibrometer. The laser beam has been focused on the tool holder end which has been covered with a reflective tape to enhance the laser signal. The vibrometer scanning speed has Fig. 2 The experimental setup been set equal to the feeding speed so that the laser beam follows the tool holder (Fig. 2). Therefore, sixty acceleration signals have been captured in each run at different sections of the workpiece with a sampling frequency of 51200 Hz in the frequency band of 0-20,000 Hz. The signals in question are presented in Fig. 4.
At the end of each run, three values of surface roughness are measured with a PCE-RT 1200 2D roughness tester. The cutting insert is then dismounted and placed in an OPTIKA metallographic microscope in order to measure the flank wear. The experimental data are shown in Table 1.

Signal processing
For engine lathes, the raw vibration signal may contain gearbox meshing components, spindle rotational speed harmonics, cutting tool and workpiece natural frequencies, components of other rotating parts such as the feed rod and the bearings, plus some random components often described as noise. The isolation of the machining process vibration signature is necessary for a correct correlation between tool wear, tool vibration, and surface roughness. For this step, the study proposes the use of ICEEMDAN as a decomposition algorithm that extracts the different vibration modes embedded in the whole captured signal.
ICEEMDAN is the improved version of the previous CEEMDAN algorithm, proven in many works to give better results when used in signal processing field [34,35]. For the decomposition of a given signal x(t) with CEEMDAN, the below steps are followed: obtain the first IMF 1 by averaging: where ε 0 is the signal-to-noise ratio and n i is the Gaussian white noise. 2. Calculate the first residue as: ) until their first EMD mode and calculate the second mode: where E j is an operator that produces the jth mode obtained by EMD.
Steps from 4 to 6 are repeated until the obtained residue in no longer feasible to be decomposed and satisfies: with K as the total number of modes. The original signal x(t) can be expressed in the end as: The ICEEMDAN algorithm follows instead these steps: 1-Calculate the upper and lower envelopes mean of I realizations x i (t) = x + ε 0 E 1 n i (t) using the EMD algorithm. 2-Obtain the first residue r 1 by averaging the results of the first step: 3-Calculate the first mode: 4-Estimate the second residue r 2 as the average of upper and lower envelopes of the realizations r 1 + ε 1 E 2 n i (t) and then calculate the second mode: 5-For k = 3…. K, calculate the kth residue: 6-For k = 3…. K, define the kth mode as: 7-Go to step 4 for the next k.
CEEMDAN allows the extraction of the vibration modes embedded in the measured signal, with less number of iterations when compared to EMD and EEMD.   Spectral kurtosis (SK) is a statistical tool that can locate nonstationary or non-Gaussian comportment in the frequency spectrum of a raw signal. SK can be used as a preprocessing tool that provides useful information such as filtering parameters. The spectral kurtosis K(f), of a signal s(t), is calculated as follow: S(t, f) is the short-time Fourier transform of the signal s(t).

Wear quantification
In order to measure the flank wear, the cutting insert is dismounted after each run and then put under a metallographic microscope that enables the quantification of wear through the counting of the corresponding pixels in the image taken by the microscope high-resolution digital camera. Figure 5 shows the processed insert images of the twenty-one runs with their corresponding flank wear value.

Results and Discussion
Cutting inserts generally wear in a progressive manner. The speed of wear depends on several factors such as cutting parameters, workpiece material, and the material of the cutting insert itself. The wear level is considered acceptable if the cutting edge keeps providing satisfactory results taking into account the specific parameters of surface roughness. After a certain wear level, which is generally above 300 μm, the surface roughness drastically degrades affecting the vibratory signature of the whole machining process. Therefore, surface condition and insert's wear level could be revealed by the surveillance of the vibration signals spectral composition. Figure 6 shows the FFT spectrums of the twenty-one captured signals. The examination of the spectral composition shows that the frequency band of 0-5000 Hz has almost had the same frequency components over the whole twenty-one runs. Hence, the frequency peaks in this band cannot be linked to the surface condition or to the insert wear level. This has been confirmed with a contactless test, where we have turned on the lathe, activated the feeding movement with no contact between the cutting tool and the workpiece, and then captured the vibration signature. Figure 7 represents the signal spectrum of the band 0-8000 Hz, where we clearly notice the same frequency composition.
On the other hand, the components of the 5000-20,000 Hz band have behaved differently over the twenty-one tests. In the first three runs, equidistant peaks have clearly appeared between 10,000 and 20,000 Hz; however, in the 4th, 5th, 6th, 7th, 8th, and 9th runs, these peaks have been faded by the rising neighbor frequencies. The peaks reappear again in the runs [10th-17th] with less clearness when compared to the first three runs. Finally, the followed frequency components are covered again in the last four runs spectrums. These fluctuations in the peaks level have been linked to the surface roughness variations.
In order to confirm this hypothesis, spectral kurtosis has been calculated for each run. Figure 8 shows the SK variation of the first three runs. A high SK value is noticed around 16 KHz, a synonym of the presence of non-Gaussian behavior at the mentioned frequency. Notice that 16 KHz is included in the suspected frequency band. In addition, the SK value has   , and (2.09), respectively; this might be related to insert wear increase. However, for the next runs, SK has stopped giving meaningful information; Figure 9 illustrates the SK nonregular variation of the fourth and fifth runs, from which it is very difficult to know where the non-Gaussian components are concentrated in the frequency spectrum as we did previously. We mention that this behavior has been observed over eight runs, from the 4th to the 11th run, after that the highest SK value of the twelfth run has been around 16 KHz again (Fig. 10), yet its level has decreased to the value of (1.1). Once more, the SK has been irregular over the rest of runs, and the twenty-first run is illustrated in Fig. 11 as an example. In another procedure, ICEEMDAN has been used to filter out the insignificant frequencies, namely the low-frequency modes. The first modes produced by the ICEEMDAN decomposition contain generally the highest frequencies imbedded in the original signal, which is, in our case, the modes that perfectly cover the 5000-20,000 Hz band. Therefore, the energy of the first four IMFs of each captured signal has been calculated in order to follow the variations of the band magnitude. Figure 12 illustrates how the energy of the first four modes clearly increased after the 9th run with the same manner.
Furthermore, the investigation of the surface roughness variation during the twenty-one tests shows that Ra mean values have obviously increased after the 9th test as illustrated in Fig. 13. Thus, the augmentation of the IMFs energy could be correlated with the surface roughness increase.
Finally, the analysis of the tool wear behavior could explain the variations of the surface condition throughout the performed runs. The insert wear level has been gradually increasing as shown in Fig. 14; the changes in the insert's edge shape are the main cause of surface condition degradations in our case, since machining parameters have been fixed.

Conclusion
This paper checked the feasibility of tool wear and surface roughness monitoring based on vibration analysis. A laser Doppler vibrometer has been used in order to acquire tool's vibration signals during the turning of unalloyed carbon steel. First, a contactless test has been performed to identify the lathes fundamental frequencies. Spectral kurtosis has been then calculated for twenty-one runs vibration signals, which allowed to locate irregular behavior on the frequency spectrum. An empirical decomposition has been then performed using ICEEMDAN in order to isolate the significant vibration components from the lathe's frequencies. The analysis of the raw signals and their isolated modes with energy and spectral kurtosis variation has highlighted a correlation between the surface roughness, insert wear, and tool's vibration comportment: & SK has allowed to pinpoint the right frequency band linked to machining process, yet SK level variation is difficult to exploit for tool wear monitoring. & The high-frequency modes energy has increased in proportion to the increase of surface roughness. & Insert wear has gradually increased and been the main cause of surface degradation in our case, where machining parameters have been fixed. & High-frequency vibration components energy can be used for tool wear and surface roughness monitoring. & On the base of the SK analysis, only one IMF's energy has to be calculated in future, namely the IMF that covers the band that has the highest SK value.
Acknowledgment The machines, devices, and tools used in this work are the property of the Mechanics Research Center of Constantine; the latter is financed by the Algerian Ministry of Higher education and Scientific Research.
Funding The work is financed by the Algerian Ministry of Higher education and Scientific Research.

Declarations
Ethics approval I certify that the paper follows the ethical rules of good scientific practice mentioned in the "Ethical Responsibilities of Authors" of the journal.

Consent for publication
The author hereby grants to Springer International Publishing the right to publish this work.
Conflicts of interest I, Doctor Moahamed Lamine BOUHALAIS, the corresponding author of the paper, declare that we have no competing interests.