1. Introduction
Carbon fiber-reinforced plastic (CFRP) has been widely used in aviation parts to reduce their weight, due to its excellent material properties, such as high specific strength and specific stiffness. Meantime, traditional metallic materials, such as titanium alloys, are still irreplaceable for key structural components due to their excellence in bearing multi-directional loads. As such, CFRP/Ti stacks structures are formed, and tens of thousands of high accuracy holes need be drilled for bolted or riveted assembling [
1]. The reliability and safety of equipment will be directly affected by the drilling quality [
2,
3]. However, CFRP and Ti titanium are both well-known as hard-to-cut materials and have distinguished cutting performance. These would lead to severe and rapid tool wear, consequently increasing the risk of drilling defects. As most of these stack holes are produced by drilling in the assembly site, even a slight defect would be a risk of huge losses of the whole assembled parts; therefore, the cutting tool condition needs to be precisely monitored for further constant control of drilling process.
For a given tool, the standard rule of presetting a conservative tool life is still utilized in many industrial scenes, which is not realistic and relies on operator skills. In this case, premature failure occurs occasionally, which causes unexpected material losses, machine downtime and even equipment damage, while another tool may exceed the preset life [
4]. Normally, only 50–80% of a cutting tool’s life is actually used in order to reduce the risk of scrapped parts and unexpected downtime [
5,
6]. Automatic monitoring technology is the key component of unmanned machining, as it can help to avoid damages and reduce the production costs in industries [
4]. However, it is not feasible to measure tool-wear conditions directly in the continuous machining process, due to the harsh machining environment. Data-driven intelligence has achieved good performance in enabling in-process tool-condition monitoring (TCM) and tool-condition forecasting (TCF) with the rapid application and development of digital technology in the current era of Industry 4.0, such as smart sensors, signal acquisition and data analytics [
7,
8]. TCM is used to predict the current tool conditions based on measuring signals, while TCF forecasts tool conditions with the cutting tool in continued use. Compared with current states’ monitoring, TCF is more meaningful for intelligent manufacturing [
9].
The measured signals consist of cutting force [
10,
11], vibration [
12,
13,
14], acoustic emission [
15], and current [
16,
17,
18], which are widely adopted as the input data in the literature. Cutting force is the most stable and reliable signal; it is very sensitive to tool wear [
19], but difficult to apply due to the high cost and limited size of measuring instruments [
20]. It remains difficult for the industrial application of acoustic emission sensors because the signals are easily affected by various factors, such as sensor location and environmental noise [
21]. Compared to cutting force and acoustic emission, cutting vibration and current/power signals are easy and economical to acquire. In addition, multiple sensors and fusion can be utilized for improving the deficiency of single sensor [
22,
23], such as low signal-to-noise ratio and sensitivity to the mechanical system characteristics, and increase the reliability of TCM and TCF systems. Thus, in this paper, the cutting spindle power and vibration for drilling CFRP/Ti stacks are used for tool-wear forecasting in order to facilitate its application in the production site.
The cutting signal changes are not only affected by the tool conditions, but are also related to the machining variables, such as cutting parameters, clamping method, and cooling/lubrication conditions. The relationships between signals and tool conditions are nonlinear and complex, and classification and regression models are essential to map from measured signals to cutting tool conditions. Machine learning plays an important role in classification and regression, and has been extensively used in different engineering applications, such as laser welding, robotics optimization, etc., having achieved good results [
24,
25,
26,
27]. For many tool condition classification and regression methods, the models have been presented using support vector machine [
16,
17], hidden Markov model [
28], and fuzzy logic [
29]. However, for tool-wear forecasting, deep learning models have been verified to outperform the traditional techniques [
30] regarding the capture of the non-linear dependencies. For example, Kothuru et al. presented a TCM approach for end mill based on the deep convolutional neural network (CNN) [
31]. Huang et al. used deep CNN to propose a tool-wear-predicting model for milling operation of a three-flute cutter [
32]. The CNN model can combine multi-domain feature fusion, overcoming the deficiencies of manual feature fusion, which possesses the advantage of extracting high correlated signal features. Domínguez-Monferrer et al. proposed a tool wear evolution monitoring method on the basis of random forest machining learning for drilling CFRP automatically and emphasized it is effective as a process control indicator [
33]. Marani et al. proposed a LSTM model for tool flank wear prediction during the turning of a steel alloy [
34]. The LSTM model contained two layers and eight hidden units, using spindle current signals, and demonstrated the highest accuracy. However, the turning condition for predicting tool wear is simple and stable. LSTM is an appropriate method for learning the process features of different time periods. Liu et al. presented a novel method for tool wear monitoring using a parallel ResNet to extract multi-domain signals’ features and using a stacked bidirectional LSTM to obtain time series features [
35]. The raw signals of cutting force, vibration and acoustic emission for milling single material are utilized as inputs to overcome the low efficiency of manually selecting signal features. However, data are still the cornerstone for these approaches, and the accuracy and robustness are directly related to the quality of signal features. Unfortunately, for CFRP/Ti stacks, the cutting spindle power and vibration signals change are complex, influenced by more factors, compared to single material stack, due to the different materials properties. The sensitive features are difficult to extract. The effectiveness of deep learning models for tool-wear forecasting of drilling CFRP/Ti stacks remains questionable and need to be studied, as it is rarely reported in the literature.
In this paper, an in-process TCF method of drilling CFRP/Ti stacks based on a ResNet and LSTM network has been proposed. The cutting spindle power and vibration signals are measured and preprocessed. The ResNet-based model is utilized for predicting the current tool wear. Then, an LSTM network is used to estimate tool wear of the next drilling holes using the predicted latest tool-wear values as inputs. The identification accuracy is studied through an experiment, and the results using raw measured signals and preprocessed datasets are tested for comparison. The proposed data preprocessing method for drilling CFRP/Ti stacks can improve the tool-wear-forecasting accuracy, which meets the demand for online TCF. The proposed TCF model is effective in estimating tool wear of the next drilling holes.
The structure of this paper is as follows.
Section 2 presents the proposed methodology of the TCF model for drilling CFRP/Ti stacks, including the signal preprocessing method, the ResNet-based TCM model, and the LSTM network. The experimental details for drilling CFRP/Ti stacks are presented in
Section 3.
Section 4 presents the verifications and discussions of the methods using measured cutting signals and tool-wear values, followed by the conclusions, which are drawn in
Section 5.
2. Methodologies
In this paper, an approach for in-process TCM and TCF is proposed. It estimates the current tool conditions using a ResNet-based model. In addition, an LSTM network is used for estimating the nearest future tool-wear values. Otherwise, a signal process method for cutting vibration and spindle power is proposed to increase the model accuracy for the drilling process of CFRP/Ti stacks. The framework of TCF is shown in
Figure 1.
2.1. Signal Preprocess
Cutting spindle power and vibration signals with a different sampling frequency are both used as the input raw signals. The vibration signals for three directions are collected by the sensor installed on the spindle end face, where the sampling frequency is 2048 Hz. The cutting spindle power signal is acquired through OPI profibus communication connected to NCU of 840D PL siemens system, and the sampling frequency is 50 Hz.
The measured cutting spindle power and vibration signals for drilling eight CFRP/Ti stacks holes are shown in
Figure 2. The
x,
y,
z directions illustrated in
Figure 2 are perpendicular to each other. Meanwhile, the
z direction is parallel to the tool axis and the
y direction is parallel to the length direction of workpiece. The power curve ascends as the drilling tool moving down in
z direction with the feed speed, as shown in
Figure 2a, while it is hard to find the same correlation for raw vibration signal curves in
Figure 2b, due to there being many interference signals with complex factors contained within. Therefore, a signal preprocessing method is needed to extract high correlated features, aiming to improve the accuracy of TCF model.
The statistical parameter, root mean square, has shown high sensitivity for tool condition [
18]. In this paper, moving average root mean square (MARMS) is used to preprocess raw vibration signals, which can eliminate the effect of noise. It is calculated as follows.
The number of sampled signal data for each tool revolution is calculated using Equation (1) [
36].
where
n is the spindle speed and
fs is the sampling frequency.
The
RMSk can be obtained for the
kth revolution using Equation (2) [
36].
where
xi is the
ith collected vibration data.
Then, the MARMS values is calculated using Equation (3) [
36].
where
j is the successive revolution number for moving average.
The raw cutting vibration can be preprocessed to MARMS values, and the sampling frequency for MARMS values is 60
n Hz, which is shown in
Figure 3. The RMS values ascend with the drilling of the holes, and the change caused by the drilling process is more obvious on the RMS curve, that is, it is hard to capture in the vibration acceleration curve shown in
Figure 2b. This means that the effect of tool condition on the RMS values is more significant. Otherwise, the RMS curve can provide the basis for eliminating the signal offset, using the method described below.
The cutting spindle power is also preprocessed to the same sampling frequency with cutting vibration MARMS values using Equation (4) and the moving average filter.
It can be seen in
Figure 2a and
Figure 3 that there is a signal offset with the increase in sampling time, even when drilling holes using the same parameters. The offset is influenced by the sensor’s structural parameters and the temperature, which is difficult to avoid. The offset will significantly affect the measured value for the long sampling time. It is the interference noise independent of tool condition, and the offset should be eliminated. A data preprocessing correction method is proposed to increase the TCF model accuracy.
The no-cutting phase signals are acquired through the segmentation of sampled signals using Equation (5).
where
Pmean is the mean value of cutting vibration MARMS values or moving average power values, and the number
M can be calculated using Equation (6).
The t(s) is the cutting time from the position where the drill tip attacks the CFRP top interface to the position where the drill tip penetrates into the exit boundary of the Ti phase.
Then, the linear regression is utilized, due to the outstanding fitting accuracy in predicting the deviation caused by the signal offset, for fitting the no-cutting phase signals and for predicting the signal offset values of cutting phase. The predicted values are used to compensate for the raw preprocessed signals, and the data is also modified by subtracting the minimum value. As illustrated in
Figure 4, the moving average power values are modified using the data correction method to eliminate the signal noise and the same method is utilized for cutting vibration MARMS.
2.2. ResNet-Based In-Process TCM Model
The TCM model is needed to predict current tool wear using measured real-time signals because it is nearly impossible to measure tool-wear values in the CFRP/Ti drilling process. The ResNet can use multi-channel signals as input, with the ability to extract multi-domain sensitive signals features, which helps to overcome the low efficiency of manually selecting signal features [
35]. Otherwise, ResNet increases the performance of the network with a large number of layers by handling the vanishing gradient problem [
5]. Therefore, it is chosen in this paper due to its ability to improve accuracy and robustness.
The framework of the ResNet-based model is shown in
Figure 5. The inputs are four-dimensional time series signal segments as
Xn = [
pn,
vn1,
vn2,
vn3], including moving average spindle power,
X cutting vibration MARMS,
Y cutting vibration MARMS and
Z cutting vibration MARMS signals, which are acquired by the proposed signal preprocessing method.
As depicted in
Figure 5, this TCM model is used to learn a mapping from the input time series signal segments
Xn to an output value
, the tool-wear values at drilling hole numbers.
The inputs are preprocessed by a de-noising block into a post-activation residual block followed by the pre-activation residual blocks. The structure of post-activation residual block starts with a convolutional layer (Conv), followed by batch normalization layer (BN), rectifier linear unit (ReLu), dropout layer in sequence, and ends with Conv. Otherwise, a max-out pooling layer is utilized as the shortcut of residual learning framework to reduce and eliminate redundant features.
Every pre-activation residual block starts with a BN, followed by Relu, Dropout, Conv, repeating in sequence. In addition, a number of pre-activation residual blocks are used in TCM model. Dense is a fully-connected layer, which compiles the extracted data of previous layers to form the final output. It is used, finally, to produce a tool-wear value.
2.3. LSTM Network for In-Process TCF
The LSTM network is well-suited for processing and predicting long time series, having been designed to remove the vanishing gradient problem [
37]. An LSTM network is used in this paper, which estimates tool-wear values of the next drilling holes using several latest past tool-wear values, which are estimated by the ResNet-based model.
The network is shown in
Figure 6. It consists of an encoder, a decoder and a context tensor. The same LSTM units with the dimensions depicted in
Figure 6 are used for both the encoder and decoder.
The estimated n values of the TCM model are input to the encoder. Then, a context tensor is calculated by the encoder and a full connection layer, which is input to the decoder by a duplicated vector layer. The decoder estimates m sequential output values, which are the future tool-wear values.
However, the input tool-wear values obtained using TCM may wave violently during the machining process, and the LSTM network may amplify the errors caused by the TCM model. The estimated values should not be directly output.
A data correction method based on mean values is used to calculate the output from the original estimated values of the decoder. The tool-wear value regarding time
t is calculated as follows for the estimated values (
,
, …,
).
Usually, the tool-wear values ascend gradually, which is different from what occurs with tool tipping. Thus, the final tool-wear value is obtained using Equation (8) to further reduce the impact of data fluctuations.
3. Experimental Details
A CNC machining center RAMMATIC 1201G is used to conduct the experiment. The experimental setup is depicted in
Figure 7. The uncoated K44UF tungsten carbide drill bit is used, and its geometric parameters are illustrated in
Table 1. The CFRP/Ti stacks composed of CFRP and Ti6Al4V are used in the experiment, with their thickness being depicted in
Figure 7. The used CFRP composite is quasi-isotropic laminates, which are composed of the T800 grade carbon fiber in an epoxy matrix. The CFRP/Ti stacks workpiece’s size is 300 mm × 200 mm × 5.5 mm, and it is fixed on the fixture with avoiding holes. The drilling sequence is from CFRP to Ti when drilling CFRP/Ti stacks. The through holes are drilled with cutting parameters as follow: cutting speed
vc = 60 m/min, feed speed
F = 318 mm/min.
One PCB 356A02 tri-axial vibration acceleration sensor and a NI PXI-4496 acquisition system are used to acquire the drilling vibration signals, with a sampling frequency of 2048 Hz. Additionally, the vibration acceleration sensor is installed on the spindle end face. The cutting spindle power signal is acquired, with a sampling frequency of 50 Hz. The cutting edge wear values are measured using a Dino-lite AM7013MZT4 optical profiler, as illustrated in
Figure 8.
Two of the same drills are used for the experiment with the same cutting parameters and experimental set-up. In addition, the curves of measured tool-wear values vs. drilling hole numbers for two drill bits are shown in
Figure 9. The drilling tool-wear values both ascend with hole numbers.
The cutting vibration and spindle power signals of the two drills are used to build and validate the proposed TCF model. The cutting signal datasets of drill No. 1 are used for building the model, including training and testing to acquire the optimal model parameters. The shuffle of datasets and cross validation are utilized to avoid over fitting or under fitting, which acquire high convergence speed for the training loss. Moreover, those of drill No. 2 are used for validating the effectiveness of the model in forecasting tool wear. The validating datasets never appeared for training and testing.
5. Conclusions
In this paper, aiming to monitor and forecast tool wear of drilling CFRP/Ti stacks, an in-process TCF method based on deep learning has been proposed. Using the cutting spindle power and vibration signals preprocessed using the proposed method, the LSTM network with the ResNet-based model integrated can forecast tool-wear values of the next drilling holes. A case study shows the effectiveness of the proposed method. In addition, the results using raw measured signals and preprocessed datasets are tested for comparison. The proposed method can extract the sensitive signal features of tool wear for drilling CFRP/Ti stacks and improve the accuracy of TCF. Based on the work conducted, the following conclusions can be drawn.
(1) The LSTM network can estimate tool-wear values of the next drilling holes, using the predicted values of the ResNet-based model as inputs. The results fit nicely with the measured tool-wear curve. Using historical cutting datasets, it is feasible to predict future cutting tool-wear values for drilling CFRP/Ti stacks.
(2) For the LSTM network, the small errors accumulate to greater errors because the input-monitored tool-wear values may wave violently during the machining process. The proposed data correction method based on the preprocessed mean values can reduce the MAE to 20.43 μm, decreasing the impact of data fluctuations.
(3) Compared to using raw measured signals, the results of the TCF model using preprocessed data are better accuracy. The MAE of LSTM network results using raw signals are 45.01 μm, which is 2.20 times bigger than that using preprocessed signals. The proposed data preprocessing method for drilling CFRP/Ti stacks can improve the tool-wear-forecasting accuracy, which meets the demand for online TCF.
However, there are some limitations for this approach, which call for further research. More experiments for more drill bits should be conducted to test the model’s robustness and reliability. Otherwise, the impact of the drilling workpiece’s structure should be studied with various stiffness and different drilling parameters on the accuracy.