*2.1.3. Sensor signal feature extraction and selection in the time domain and neural network pattern recognition for tool wear estimation*

Sensor signal feature extraction can also be carried out in the time domain with the aim to find correlations between the extracted sensor signal features and the tool wear state.

In [22], sensor signal feature extraction in the time domain was applied to CFRP/CFRP stack drilling performed with a traditional twist drill bit, usually employed in the aircraft industry

**Figure 5.** Reconstructed tool wear curve obtained by the ANN trained with the three features extracted from the thrust force signal (6000 rpm–0.11 mm/rev).

Drilling of Fiber-Reinforced Composite Materials for Aeronautical Assembly Processes http://dx.doi.org/10.5772/intechopen.80466 75

The ANN performance was evaluated in terms of mean square error (MSE), that is, the variance of the differences between the VB values predicted by the ANN and the target VB values. The very low MSE values, between 9.82E-07 and 1.57E-05, confirmed the robust correlation between the sensor signal features extracted in the frequency domain and the tool wear, due

Through pattern recognition based on the frequency domain signal features extracted from the multiple sensor monitoring signals, the ANN accurately reconstructed the tool wear curve. **Figure 5** presents the reconstructed tool wear curve for the drilling test performed at 6000 rpm and 0.11 mm/rev, obtained by applying the leave-k-out method with *k* = 1 to the ANN with three hidden layer nodes trained with the three features extracted from the thrust force signal (MSE = 1.57E-05). **Figure 6** presents the reconstructed tool wear curve for the same drilling test obtained by applying the leave-k-out method with *k* = 1 to the ANN with 5 hidden layer nodes trained with thefive features extracted from the thrust force and the torque signal (MSE= 3.79E-06).

features coming from both thrust force and torque signals, provided better results than FPV<sup>1</sup>

The accurate tool wear curve reconstruction achieved by the ANN can be effectively utilized to determine the end of useful tool life through on-line diagnosis during CFRP/CFRP stack drilling, identifying the transition of the tool flank wear between the second and the third

Sensor signal feature extraction can also be carried out in the time domain with the aim to find

In [22], sensor signal feature extraction in the time domain was applied to CFRP/CFRP stack drilling performed with a traditional twist drill bit, usually employed in the aircraft industry

**Figure 5.** Reconstructed tool wear curve obtained by the ANN trained with the three features extracted from the thrust

*2.1.3. Sensor signal feature extraction and selection in the time domain and neural network* 

correlations between the extracted sensor signal features and the tool wear state.

, including frequency domain

.

to the interaction between the drill bit and the anisotropic CFRP material.

For all drilling conditions, the sensor fusion pattern vector FPV<sup>2</sup>

phase of the wear curve.

74 Characterizations of Some Composite Materials

force signal (6000 rpm–0.11 mm/rev).

*pattern recognition for tool wear estimation*

**Figure 6.** Reconstructed tool wear curve obtained by the ANN trained with the five features extracted from the thrust force and the torque signal (6000 rpm–0.11 mm/rev).

and an innovative geometry step drill bit. The traditional tool was a 2-flute 6.35-mm diameter twist drill with 125° point angle and 30° helix angle tungsten carbide (**Figure 7a**), while the innovative step drill bit was a 2-flute 6.35-mm diameter drill bit with 120° point angle and 20° helix angle made of tungsten carbide (WC) (**Figure 7b**). Different cutting parameters were used for the experimental drilling tests: three feed values (0.11, 0.15, and 0.20 mm/rev) and three spindle speeds (2700, 6000, and 7500 rpm).

A statistical approach in the time domain was applied for feature extraction from the thrust force and torque sensor signals acquired during the drilling tests. The following five statistical signal features were extracted: arithmetic mean, variance, skewness, kurtosis, and signal energy. For each extracted feature, the Spearman correlation coefficient, *r<sup>s</sup>* , was calculated to assess its correlation with tool wear (0 < *r<sup>s</sup>* < 0.3, weak correlation, 0.3 < *r<sup>s</sup>* < 0.7, moderate correlation, and 0.7 < *r<sup>s</sup>* < 1, strong correlation). Based on the *r<sup>s</sup>* coefficient, five features showed a strong correlation with tool wear: thrust force average (Fz\_av), torque average (Tav), thrust force variance (Fz\_var), torque variance (Tvar), and thrust force kurtosis (Fz\_kurt).

Every 10 holes, a magnified picture of the tool was acquired through an optical measuring machine (Tesa Visio V-200) to measure the flank wear [21]. A 3rd order polynomial interpolation of the VB values was applied to reconstruct the tool wear curves for each drill bit under the tested drilling conditions (**Figures 8** and **9**).

The graphical analysis of the selected sensor signal features, with particular reference to the thrust force average (**Figures 10** and **11**), which is the most correlated feature, shows that the behavior of the features and the behavior of tool wear (**Figure 8** and **9**) display the same increasing trend with increasing number of holes. The average thrust force is higher for the innovative drill bits.

The selected statistical features were employed to construct, for each drilling condition, 60 sensor fusion feature pattern vectors (SFPVs), that is, one for each drilled hole. Each SFPV was built by combining the selected statistical signal features and the corresponding hole number, *n*. Each SFPV was associated to its matching flank wear value (VB) to create input-output

**Figure 7.** (a) Traditional twist drill bit; (b) innovative step drill bit [22].

**Figure 8.** Tool wear values and interpolated curves. Traditional drill bits.

vectors for ANN learning. Three-layer cascade-forward backpropagation artificial neural networks (ANN) using the Levenberg-Marquardt training algorithm were setup [20]. For each drilling condition, 60 input-output vectors (i.e., one for each drilled hole) were built.

To assess the performance of the ANN in properly forecasting the upcoming tool wear values based on a restricted number of initial training input-output vectors and define the smallest number of vectors required to get a reliable tool wear forecast, the number, *m*, of input-output vectors utilized for ANN training was gradually increased by steps of 10 from *m* = 20 to *m* = 50.

The pattern recognition performance was measured by the root mean squared error (RMSE), that is, the sample standard deviation of the differences between the ANN predicted tool wear values and the experimental tool wear values.

In general, the prediction performance improved by increasing *m* since, by enlarging the number of input-output vectors used for training the ANN, the latter was able to enhance the forecast of future tool wear values. Also in the case of the experimental tests carried out

Drilling of Fiber-Reinforced Composite Materials for Aeronautical Assembly Processes

http://dx.doi.org/10.5772/intechopen.80466

77

**Figure 9.** Tool wear values and interpolated curves. Innovative step drill bits.

**Figure 10.** Thrust force average vs. hole number. Traditional drill bits.

**Figure 9.** Tool wear values and interpolated curves. Innovative step drill bits.

**Figure 10.** Thrust force average vs. hole number. Traditional drill bits.

vectors for ANN learning. Three-layer cascade-forward backpropagation artificial neural networks (ANN) using the Levenberg-Marquardt training algorithm were setup [20]. For each

To assess the performance of the ANN in properly forecasting the upcoming tool wear values based on a restricted number of initial training input-output vectors and define the smallest number of vectors required to get a reliable tool wear forecast, the number, *m*, of input-output vectors utilized for ANN training was gradually increased by steps of 10 from *m* = 20 to *m* = 50. The pattern recognition performance was measured by the root mean squared error (RMSE), that is, the sample standard deviation of the differences between the ANN predicted tool

drilling condition, 60 input-output vectors (i.e., one for each drilled hole) were built.

wear values and the experimental tool wear values.

**Figure 8.** Tool wear values and interpolated curves. Traditional drill bits.

**Figure 7.** (a) Traditional twist drill bit; (b) innovative step drill bit [22].

76 Characterizations of Some Composite Materials

In general, the prediction performance improved by increasing *m* since, by enlarging the number of input-output vectors used for training the ANN, the latter was able to enhance the forecast of future tool wear values. Also in the case of the experimental tests carried out

**Figure 11.** Thrust force average vs. hole number. Innovative drill bits.

with the innovative drill bits, the prediction performance notably improved by enlarging the number of input-output vectors utilized for training the ANN.

The best and the worst ANN prediction performances were found for the innovative drill bits. In the worst case (2700 rpm and 0.11 mm/rev), a maximum RMSE of 0.0164 for *m* = 20 and of 0.0056 for *m* = 50 were obtained due to some outliers (**Figure 12**). In the best case (6000 rpm and 0.20 mm/rev), a minimum RMSE of 0.00023 for *m* = 50 was obtained. **Figure 13** shows the tool wear curves predicted by the ANN for this case. The final curve relative to *m* = 50 is essentially overlaid to the measured tool wear curve, meaning that the ANN is capable to accurately forecast the future 10 tool wear values.

Even though the ANN prediction can be considered as more demanding in the case of the innovative drill bits, which have a more complex geometry and tool wear development, the ANN performance was still suitable, showing very low prediction errors (minimum RMSE = 0.00023 and maximum RMSE = 0.0164).

The decision making procedure for identifying the end of the tool life based on the ANN prediction can operate in a safe manner, allowing the multiple sensor monitoring procedure to be valuably utilized for on-line tool condition monitoring aimed at the implementation of a condition-based tool substitution strategy instead of a time-based strategy.

In [24], the sensor signal feature extraction methodology in the time domain was applied to drilling of CFRP/CFRP stacks performed using a traditional twist drill bit with different drilling conditions (rotational speed: 2700, 6000, and 9000 rpm and feed: 0.11, 0.15, 0.20 mm/rev).

about the state of the tool via artificial neural network-based pattern recognition paradigms. Based on the calculation of the Pearson's correlation coefficient, five best correlated features were identified (thrust force average, thrust force variance, thrust force skewness, torque average, and acoustic emission average), and they were utilized to form the input sensor fusion pattern vector for ANN pattern recognition. The ANN prediction of tool wear was highly

**Figure 13.** Best case ANN forecast: experimental tests at 6000 rpm–0.20 mm/rev with innovative drill bit [22].

**Figure 12.** Worst case ANN forecast: experimental tests at 2700 rpm–0.11 mm/rev with innovative drill bit [22].

Drilling of Fiber-Reinforced Composite Materials for Aeronautical Assembly Processes

http://dx.doi.org/10.5772/intechopen.80466

79

In this case, thrust force, torque, and acoustic emission sensor signals were acquired during the experimental drilling tests. On these signals, statistical feature extraction and data fusion were implemented to construct sensor fusion pattern vectors in order to make predictions Drilling of Fiber-Reinforced Composite Materials for Aeronautical Assembly Processes http://dx.doi.org/10.5772/intechopen.80466 79

**Figure 12.** Worst case ANN forecast: experimental tests at 2700 rpm–0.11 mm/rev with innovative drill bit [22].

with the innovative drill bits, the prediction performance notably improved by enlarging the

The best and the worst ANN prediction performances were found for the innovative drill bits. In the worst case (2700 rpm and 0.11 mm/rev), a maximum RMSE of 0.0164 for *m* = 20 and of 0.0056 for *m* = 50 were obtained due to some outliers (**Figure 12**). In the best case (6000 rpm and 0.20 mm/rev), a minimum RMSE of 0.00023 for *m* = 50 was obtained. **Figure 13** shows the tool wear curves predicted by the ANN for this case. The final curve relative to *m* = 50 is essentially overlaid to the measured tool wear curve, meaning that the ANN is capable to

Even though the ANN prediction can be considered as more demanding in the case of the innovative drill bits, which have a more complex geometry and tool wear development, the ANN performance was still suitable, showing very low prediction errors (minimum

The decision making procedure for identifying the end of the tool life based on the ANN prediction can operate in a safe manner, allowing the multiple sensor monitoring procedure to be valuably utilized for on-line tool condition monitoring aimed at the implementation of a

In [24], the sensor signal feature extraction methodology in the time domain was applied to drilling of CFRP/CFRP stacks performed using a traditional twist drill bit with different drilling conditions (rotational speed: 2700, 6000, and 9000 rpm and feed: 0.11, 0.15, 0.20 mm/rev). In this case, thrust force, torque, and acoustic emission sensor signals were acquired during the experimental drilling tests. On these signals, statistical feature extraction and data fusion were implemented to construct sensor fusion pattern vectors in order to make predictions

condition-based tool substitution strategy instead of a time-based strategy.

number of input-output vectors utilized for training the ANN.

**Figure 11.** Thrust force average vs. hole number. Innovative drill bits.

78 Characterizations of Some Composite Materials

accurately forecast the future 10 tool wear values.

RMSE = 0.00023 and maximum RMSE = 0.0164).

**Figure 13.** Best case ANN forecast: experimental tests at 6000 rpm–0.20 mm/rev with innovative drill bit [22].

about the state of the tool via artificial neural network-based pattern recognition paradigms. Based on the calculation of the Pearson's correlation coefficient, five best correlated features were identified (thrust force average, thrust force variance, thrust force skewness, torque average, and acoustic emission average), and they were utilized to form the input sensor fusion pattern vector for ANN pattern recognition. The ANN prediction of tool wear was highly

**3. Conclusions**

laminates was investigated.

based pattern recognition paradigms.

which unacceptable hole quality is generated.

drilling tests.

This chapter provided an overview of the main challenges related to drilling of fiberreinforced plastic composite materials which are extensively employed in the aeronautical industry. Rapid tool wear is generated due to the abrasiveness of the reinforcing fibers, and different types of damages affecting material integrity and surface quality, with particular

Drilling of Fiber-Reinforced Composite Materials for Aeronautical Assembly Processes

http://dx.doi.org/10.5772/intechopen.80466

81

With reference to aeronautical industry applications, where the assembly of CFRP components requires "one-shot" drilling processes so as to allow for easier subsequent riveting avoiding misalignment issues, drilling of CFRP/CFRP stacks made of two superimposed

Based on a wide experimental drilling campaign, the case studies analyzed the influence of drilling parameters, tool type and geometry on tool wear development, hole quality and surface integrity, and the opportunity to implement advanced sensor monitoring procedures for tool condition monitoring based on the acquisition and processing of thrust force and torque signals. Diverse multiple sensor process monitoring procedures were implemented in the drilling of CFRP/CFRP stacks for the assembly of aircraft components, with the aim to support on-line decision making on tool replacement time through cognitive tool wear estimation and hole quality assessment. The monitoring procedures were based on the acquisition and processing of thrust force, torque, and acoustic emission sensor signals during the experimental

With the purpose to explore the complex frequency content of the thrust force and torque sensor signals acquired in the multidirectional CFRP/CFRP stack drilling experimental tests,

The sensor signal processing techniques, comprising signal conditioning, feature extraction in the time and frequency domain and data fusion, were implemented to construct sensor fusion feature pattern vectors—made of sensor signal features coming from sensors of different nature—with the aim to find correlations with tool state via artificial neural network-

The ANN performance results achieved in the case studies indicated that, for all CFRP/CFRP stack drilling conditions, by using sensor fusion pattern vectors made of selected features extracted from force and torque sensor signals, a very accurate ANN prediction of tool wear is achieved. As a matter of fact, these procedures demonstrated reliable correlations between sensor signal features and tool wear level both in the case of the features extracted from the

The prediction of tool wear can be functionally utilized to forecast the quality of the drilled holes. As a matter of fact, a correspondence between exit delamination factor and tool wear transition between the second and third phase of the wear curve was observed. In this transition, an exit delamination factor value, Df = 1.4, was identified and set as threshold beyond

time domain and in the case of the features extracted from the frequency domain.

advanced signal processing was also carried out in the frequency domain.

reference to delamination damage generation, are often produced by drilling.

**Figure 14.** Exit diameter error, tool flank wear, and exit delamination (6000 rpm–0.20 mm/rev).

accurate, with an RMSE <4e-3, showing reliable correlations between fused signal features and tool wear level. The reliable predictions of tool wear development can be used to support on-line decision making on drill bit replacement need.

Moreover, the prediction of tool wear can be functionally utilized to forecast the quality of the drilled holes. The latter was assessed considering dimensional accuracy and entry/exit delamination, which both have an effect on the performance of the CFRP assembly. Dimensional accuracy was measured with reference to hole diameter error, that is, the difference between actual and nominal hole diameter divided by nominal hole diameter [30]. The entry/exit delamination was assessed with reference to the delamination factor, Fd, that is, the ratio between the diameter of the circle encompassing the damaged area and the nominal diameter of the hole [25].

The drilled hole quality evaluations were utilized to set up a criterion for tool replacement need, which is required when the tool wear is responsible for a drilled hole, the quality of which is no longer acceptable. As the lower limit of the tolerance range corresponds to the nominal diameter of the hole, any negative hole diameter error is unacceptable. For each drilling condition, the occurrence of negative hole diameter errors was detected and associated with the corresponding flank wear value and the exit delamination factor.

Under all drilling conditions, the hole diameter error became negative when the flank wear, VB, reached the typical value of 0.04 mm (see **Figure 14**). The latter can be used as a threshold to determine the need for tool change due to an undersized hole diameter. For all drilling conditions, the exit delamination factor grew with increasing number of holes and reached a value between 1.3 and 1.4 when the flank wear reached 0.04 mm. This suggests that the flank wear threshold could also be associated with the second hole quality parameter represented by the exit delamination factor.

Hence, a correspondence between hole diameter error and exit delamination factor with tool wear level was observed. As a result, via on-line prediction of tool wear during drilling, taking into account the identified flank wear threshold, the cognitive sensor monitoring paradigm can provide diagnosis and prognosis services to support decision making on tool replacement need, which is essential for drilling automation.
