The tool-wear is a gradual process, where the worn rate depends on the workpiece and tool materials, the cutting fluids, and the cutting parameters, among others. Although, only flank wear and crater wear [3,21] are traditionally considered, there are also other kinds of tool-wear, i.e., nose wear, oxidation wear, primary notch, etc., as shown in Figure 1a [3].

The crater wear (Figure 1a, marked 2) appears at the attack face of the tool, and generally it is described as a diffusion mechanism caused by high temperatures in the tool-chip interface, apart from the chemical affinity between the workpiece and tool materials. Even when the crater wear affects the cutting process, this is determinant only under very high-speed conditions, and its measurement implies the utilization of a surface instrument to determine the maximum depth of the crater or the crater volume. In this way, most of low-speed industrial applications utilize the flank wear as main indicator in the tool change determination [21].

The flank wear (Figure 1a, marked 1) affects the tool incidence area and it is attributed to excessive rubbing with the machining surface at high temperatures [3]. In the same way, in Figure 1b, a lateral view of the tool-wear is presented, where the nose wear (VC) in region C, the width mean of flank wear, known as allowable wear land (VB) in region B, the width of maximum flank wear (VB_max), and the notch wear (VN) in region N are shown, Boothroyd and Knight [21]. Most of reported investigations have been limited to the tool-wear estimation through the measurement of the central portion (VB) exclusively. However, this region represents a portion of the tool-wear only. As it is described in literature, other two kinds of tool-wear in the flank exist: the nose wear and the notch wear.

The measurement of flank-wear area A_f (Figure 1b) gives an improved tool-wear estimator, as shown in the works of Scheffer et al. [10] and Sumit and Mingyuan [22]. The flank-wear area measurement has the advantage of considering the width of the flank wear, the nose wear and the notch wear, increasing the effectiveness on the estimation of tool-wear.

Different research works have been focused on the tool-wear monitoring or failure detection, in a qualitative way. The current sensors (Hall-effect sensors) and vibration sensors (accelerometers) are commonly utilized. For instance, on the one hand, Li [23] has utilized the current signal as estimator employing a Hall-effect sensor to qualitatively monitor the tool-wear in a turning process, utilizing a prediction model. Li and Tso [24], monitor the current in a drilling process, applying signal processing and fuzzy techniques to estimate the tool-wear, also qualitatively. Reñones et al. [25] utilize the electric power consumption through current sensors, as well as the work of Romero et al. [26] where a qualitative classification of the tool-wear applied to failure detection is presented. In a similar approach, Franco et al. [27] utilize spindle driver current to monitoring the cutting tool condition in a drilling process. On the other hand, vibrations are utilized by Dimla and Lister [7] to make a qualitative classification of tool-wear by means of neural networks. Salgado and Alonso [28] monitor the vibration signals through an accelerometer to determine the flank wear with spectral singular analysis and neural networks. In these works, a single primary sensor is utilized, but the signal processing is performed in a computer, separately. For the present research, a fused sensor approach is proposed for improving the tool-wear estimation when compared against a single-sensor approach.

Figure 2 shows the block diagram of the proposed smart-sensor for tool-wear area monitoring. The vibration signals A_x,y,z, are obtained from a standard 3-axial accelerometer and the feed-motor current I is directly taken from the current sensor at the servoamplifier. The smart sensor has two DAS (Data Acquisition System): one for acquiring vibrations and another one for the current acquisition. The signal processing is performed in the FPGA-based HSP unit. Besides, from the CNC machine, signals MS (Machining Start), depth of cut a_p and feed rate f are inputs to the smart sensor. Finally, the quantitative estimation of the flank-wear area is supplied as the result.

The block diagram of the HSP-unit internal structure to determinate the flank-wear area is shown in Figure 3. The vibration signals A_x,y,z and the feed motor current I, are input signals in the vibration and current DAS driver, respectively. Then, the three vibration signals A_x, A_y and A_z, acquired with the 3-axial accelerometer, are evaluated through Equation (1), obtaining in this way the resultant A_r that contains the information from the sensed axes: (1) A r = A x 2 + A y 2 + A z 2

Then, this signal is time windowed for only taking into account the time when the cutting is made during the machining process. The time windowing initializes its processing when the CNC machine indicates it through a logical signal MS (Machining Start). Later, according to Equation (2), the RMS (Root-Mean Square) value of the produced signal by time windowing is obtained, being A_ri the i-th sample and m the length of windowed samples. Simultaneously, the current signal is processed too, with an FIR low-pass filter (32^nd order and cutoff frequency f_c = 120 Hz) and its corresponding time windowing. Afterwards, the I_rms value of the current signal according to Equation (3) is obtained, where I_i represents the i-th sample and m the length of windowed samples: (2) A rms = ∑ i = 1 m ( A ri ) 2 m (3) I rms = ∑ i = 1 m ( I i ) 2 m

Starting from results in Equations (2) and (3), the fusion of this information is proposed as a simple weighting function [29] U = W(A_rms, I_rms, a_p, f) that takes the effects of the machining parameters into account: depth of cut a_p, and feed rate f. In a next step, the flank-wear area A_f is obtained by a polynomial of approximation A_f(U). The polynomial approximation is shown in Equation (4), where K_C is a scaling factor that depends on the depth of cut a_p; n is the polynomial degree, coefficients a_i and K_C are fitted under Matlab from the experimental data. Finally, the obtained value is sent to the output interface for display, and optionally, results are sent to a personal computer through a USB interface for storage and further processing purposes: (4) A f   ( U ) = K C ∑ i = 0 n a i U i

Figure 4 shows the vibration and current signals processing flow up to the evaluation in the polynomial approximation for obtaining the quantitative estimation of the flank-wear area A_f. On the one hand, the processing flow starts with the vibration signal acquisition; then, it computes the vibration resultant, and after a time windowing the RMS value A_rms is obtained. On the other hand, the current signal is acquired and filtered, then a time windowing is applied to calculate the RMS value I_rms later on. The next step is the signal fusion through the weighting function that combines acceleration, current, and machining parameters. Finally, the tool-wear area is estimated with a polynomial approximation.

In order to find a simple function that improves the flank-wear area estimation, the exploration of weighting functions through addition, subtraction, product and quotient operations between A_rms and I_rms parameters, is made. Ten experiments with ten different inserts are utilized for this purpose: the results are shown in Figure 6. In Figures 6a and 6b, the behavior of I_rms/A_rms and A_rms/I_rms are not monotonic, for this reason, these operations are not considered. Figures 6c, 6d and 6e, show better responses; then, these data are processed to obtain a polynomial approximation under Matlab, and the corresponding absolute errors are calculated as depicted in Figure 6f. The weighting function that relates A_rms and I_rms parameters with product shows lower error. Then the weighting function is proposed to contain the product of current and vibration, to further include the machining parameters.

The weighting function must include adjusting factors that consider the effects of the machining conditions. Equation (5) is proposed as the full weighting function, including the product of current and vibrations, alongside two adjusting factors; a scaling factor K_S and a shifting factor K_D. By considering these adjusting factors, the smart sensor is insensitive to changes in cutting conditions: (5) U = K S   A rms I rms + K D

In order to fit the weighting function with the adjusting factors, several tests were performed. These tests consist on ten turning cycles of ten inserts with different wear area for each of five depths of cut a_p and five feed rates f as summarized in Table 1. The cutting tools are coated carbide inserts CNMG433MA machining a conic surface because in this way the three components of acceleration can be analyzed (oblique cutting) over at medium-carbon steel (AISI 1,045), stock material diameter D = 25.4 mm, without coolant liquid. The inserts were classified according to their degree of tool-wear, in a range that goes from new, medium tool-wear to broken inserts. Then, their flank-wear area was estimated for calibration purposes using a NIKON EPIPHOT 200 microscope and further image processing techniques. The inserts were provided by a local metal-mechanic industry with the aim of having a representative sample of an industrial process. The cutting conditions were selected according to the general recommendations for turning operations of medium and high-C steels with coated carbide tools, as suggested in Kalpakjian and Smith [3]. The capabilities of the lathe were also considered.

The current signals are processed with the polynomial fitting methodology presented by Scheffer et al. [10], for the machining parameters specified in Experiment 1 from Table 1. The flank-wear area A_f as a function of current signal I_rms is estimated and then compared against the measured flank-wear area as shown in Figure 7. This figure also shows the cutting tool micrograph and its quantitative flank-wear areas for selected inserts (marked a, b, c and d). The 3^rd degree polynomial coefficients of A_f(I_rms) are fitted under Matlab from the experimental results to obtain (6). The absolute error for this estimation reports a mean of μ = 0.0207 mm² and a standard deviation of σ = 0.0118 mm². The estimated relative error for this approach is 12.6%. (6) A f   ( I rms ) = 213.8785   I rms 3 − 175.1362   I rms 2 + 47.5507   I rms − 4.1563

The polynomial fitting shown in Equation (6) is valid for the specified machining parameters. If machining conditions, such as feed rate or depth of cut are changed, to recalculating the fitting coefficients is necessary.

This experiment utilizes the RMS value of vibration signals A_rms for the estimation with the polynomial fitting methodology presented by Scheffer et al. [10] for the machining parameters specified in Experiment 1 from Table 1. In Figure 8, results for this approach as estimator of flank-wear area can be seen. In the same way that the RMS values of current signal in order to obtain the equation of flank-wear area A_f as function of A_rms values, a polynomial of 3^rd degree was fitted under Matlab. The fitted polynomial of A_f(A_rms) is shown in Equation (7). In this case, the mean of the absolute error from the estimation is μ = 0.0139 mm², with a standard deviation of σ = 0.0105 mm² and a relative error of 11.99%. (7) A f   ( A rms ) = 6.3008   A rms 3 − 16.7350   A rms 2 + 14.9711   A rms − 4.3642

Similar to the current approach, if machining conditions are changed in vibration monitoring, it is necessary to recalculate the fitting coefficients of the polynomial.

The fused smart-sensor proposed methodology in this work consists on the fusion of current, vibration and machining parameters as stated in Equation (5). Equation (8) shows the resultant weighting function from experimental data, considering the whole set of machining conditions stated in Table 1, where K_S = 1.3860 and K_D = 0.5836. A 3^rd degree polynomial approximation under Matlab is fitted to obtain Equation (9) from the weighting function: (8) U = 1.3860   A rms I rms + 0.5836 (9) A f   ( U ) = 30.1716   U 3 − 24.8065   U 2 + 6.9746   U − 0.5220

In order to make a fair comparison between the proposed fused methodology results against the reported single-parameter approaches, the flank-wear area is estimated with the same machining parameters from Experiment 1 described in Table 1. Figure 9 shows the tool-wear area estimation with the smart sensor methodology. The relative error for the fused approach is 3.7% with a mean μ = 0.0053 mm², and a standard deviation of σ = 0.0036 mm².

Table 2 summarizes results from the three methodologies under the same machining conditions as stated in Experiment 1 from Table 1.

The fitted polynomial in the proposed smart sensor methodology is capable of changing the machining parameters without requiring the coefficient recalculation different from the single parameter approaches that are fitted for a single machining condition. In order to demonstrate this feature, Table 3 shows the tool-wear area estimation errors for the whole set of experiments described in Table 1.

The fused smart-sensor methodology developed in this work improves the tool-wear area estimation, achieving less than four percent of relative error when compared against over ten percent obtained with single-parameter estimations. This comparison was performed considering the same machining conditions for the three approaches. The single-parameter approaches are limited to a single machining condition for their tool-wear area estimation, whereas the fused methodology is not limited because the weighting function includes the effects of machining parameters.

When machining conditions are changed in the proposed smart sensor, the weighting function adjusts the tool-wear area estimation to consider these conditions. Results from Table 3 show that for five different feed rates and five different depths of cut the behavior of the estimation has repeatability, provided by the mean and the standard deviation of the absolute error. Additionally, the relative error on the tool-wear area estimation for the set of experiments is kept below 10%, average.