A multimodal intelligent monitoring system for turning processes

Journal of Manufacturing Processes 35 (2018) 547–558

Contents lists available at ScienceDirect

Journal of Manufacturing Processes journal homepage: www.elsevier.com/locate/manpro

A multimodal intelligent monitoring system for turning processes Bin Zhang, Yung C. Shin

⁎

T

School of Mechanical Engineering, Purdue University, West Lafayette, IN 47907, United States

A R T I C LE I N FO

A B S T R A C T

Keywords: Tool condition monitoring Chatter detection Sensor fusion Feature selection Type-2 fuzzy basis function network Support vector machine

Online process condition monitoring is an essential component of closed-loop process-level automation of machining operations. This paper describes the development of an intelligent monitoring system for turning processes, which consists of three units: a tool wear predictor, a chatter detector and a tool chipping detector. Features are extracted from the signals of multiple low-cost and low-intrusive sensors, and then normalized using a novel scheme to eliminate their dependence on cutting conditions, workpiece materials and cutting tools. A systematic feature selection procedure, coupled with automated signal preprocessing parameter selection, is presented to select the optimal feature set for each unit. The tool wear unit is built with type-2 fuzzy basis function networks to predict tool wear with uncertainty bounds, while the chatter unit and tool chipping unit are built with support vector machines to maximize the classification fidelity. Experimental results show that the monitoring system achieved high accuracy, generalized applicability and satisfactory robustness for all the three process conditions, by using two affordable sensors: a power meter and an accelerometer. The three monitoring schemes are integrated into a monitoring software so that they can be implemented in different environments with minimal calibration efforts.

1. Introduction Smart factories in next-generation manufacturing call for automation at the process level for machining operations and machine tools [1], in pursuit of faster response to the rapidly-changing product design, boosted productivity, consistent product quality and lower labor cost in shop-floor operations. The closed-loop machining system aims to optimize the process performance and suppress adverse conditions by automatically adjusting operating parameters (e.g., feed, speed) based on the monitoring of relevant process variables. However, in-process direct measurements are not feasible for many process conditions, such as tool conditions and surface integrity. Therefore, indirectly inferring the unmeasurable process conditions by fusing multi-sensor data with artificial intelligence models is a promising paradigm that has been pursued in academia and industry for decades [2]. By sensing measurable process variables, such as cutting force [3,5], spindle power [8,9], vibration [5,6] and acoustic emission [6,7], and extracting descriptive features from collected signals, the process conditions of interest can be estimated using intelligent models, such as artificial neural networks [8–10], fuzzy systems [11], support vector machines [12,13], and random forests [14]. In this paper, to develop a comprehensive monitoring system for turning processes, three fault conditions are considered: tool wear,

⁎

chatter and tool chipping. Tool wear is one of the most widely studied process conditions for monitoring, because tool wear always happens during machining and negatively impacts work quality and productivity [1,15]. Hence, estimating the tool wear and scheduling timely tool replacement before it exceeds a certain limit are essential for process-level automation. Chatter is a kind of strong self-excited vibration during machining mainly because of the regenerative effect [16], which results in poor surface finish and could damage the cutting tools or machine tools. Tool chipping refers to the breaking away of small pieces of tool material from the cutting edge [17], which leads to degraded surface finish. Chatter and tool chipping are regarded as catastrophic failures with which the machining cycle must be stopped immediately. Though they can be avoided mostly by choosing proper cutting parameters in process planning, chatter or tool chipping could still happen unexpectedly in actual operations due to the variability of machining environment. Therefore, in-process detection of chatter and tool chipping is desired to prevent defective products and damage to the machines. By reviewing the literature, a significant amount of research can be found in the fields of tool wear monitoring [3–6,10–12], chatter detection [18–21] and tool chipping detection [22–24], in which the theoretical fundamentals were studied and exploratory applications were presented. Industrial implementation of process monitoring

Corresponding author. E-mail addresses: [email protected] (B. Zhang), [email protected] (Y.C. Shin).

https://doi.org/10.1016/j.jmapro.2018.08.021 Received 23 May 2018; Received in revised form 3 August 2018; Accepted 24 August 2018 1526-6125/ © 2018 Published by Elsevier Ltd on behalf of The Society of Manufacturing Engineers.

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

condition monitoring scheme that can be generalized to different operating conditions hasn’t been established.

systems has also been explored: Altintas presented a process monitoring module integrated with a virtual machining system for tool breakage detection by sensing feed and spindle drive motor current, which has been implemented on a CNC machining center for use in production [25]. Zhu demonstrated a cyber-physical framework for tool condition monitoring with three case studies of implementations [26]. A microphone-based chatter monitoring device that could work with various CNC controllers is developed in [27]. However, the number of successful industrial implementation, especially for the tool wear monitoring systems, reported in the literature is limited. In the authors’ opinion, the gap between laboratorial research and industrial implementation can be attributed to the missing of following enabling factors:

In this paper, an exemplary intelligent multisensor monitoring system is developed for turning processes, in which the above missing enabling factors are brought about. The tool wear monitoring scheme is adopted from [33]. The contribution of this paper is that the methodologies established in [33] are further extended to build the chatter and tool chipping monitoring schemes, which validated the effectiveness and versatility of the proposed feature normalization and systematic feature selection schemes, and an integrated monitoring system that is suitable for industrial implementation is developed. The monitoring system is highlighted with the followings innovations:

1) The sufficient generalization capacity [10]. To be used in a machining environment with high variability, the monitoring system should be generalizable to different machine tool set-ups, toolworkpiece combinations and cutting conditions. This requires the signal features and predictive models used in the system be independent of these factors. Also, for different machining setups, the signals and features that best describe the process conditions might be different. Therefore, a systematic feature selection procedure that combines the model-independent methods, which are computationally fast for large candidate pools, and model-dependent methods, which yield the best performance of predictive models [10,28], is highly desired. In addition, the signal preprocessing parameters, which will significantly influence a feature’s correlation to process conditions, need to be considered during feature selection. However, such a generalizable scheme and a systematic feature selection procedure are not readily available in the literature, to the best of the authors’ knowledge. 2) The practicability of instrumentation setup. For massive industrial applications, the sensors used for instrumentation must be affordable and not intrusive to original machine tool set-ups. A counterexample is the piezoelectric dynamometer used for cutting force measurement [25], which is expensive and interferes with typical machine tool set-ups, though the cutting force might be the most informative variable to indicate process conditions. 3) The consideration of process uncertainties. There are considerable uncertainties embedded in the machining processes [17], such as the variations of material properties, impact of cutting fluid, noise from the environment and uncertainties in tool wear measurements. Hence, even with the same operating conditions and process conditions, reproducing the same signal data is impossible. However consideration of process uncertainties in monitoring models has been addressed only in recent years: enhanced particle filter [29], extended Kalman filter [30], particle learning [31] and random forest with interval output [32] have been adopted to predict tool wear with uncertainty bounds. However, these models are mostly trained for a certain set of operating conditions and a robust process

1) The signal features are normalized using a novel scheme to minimize their sensitivity to cutting tools, workpiece materials and cutting conditions. The normalization parameters can be calibrated with minimal experimental efforts and the calibration procedure is also provided. 2) The optimal feature set for each monitoring scheme and the optimal signal preprocessing parameters for each feature are selected using a hybrid and systematic procedure, which is efficient for large candidate feature/parameter pools and could maximize the performance of predictive models in each scheme. 3) The predictive models are built by considering process uncertainties. Interval type-2 fuzzy basis function networks [36] are adopted to quantify the uncertainty bounds associated with tool wear, while support vector machines (SVMs) with selected dimensionless features are constructed to maximize the classification margins for chatter and tool chipping. Experimental results show that the developed tool wear predictor, chatter detector and tool chipping detector all possess high accuracy, generalized applicability and satisfactory robustness. These three schemes are integrated in a unified framework to compose a comprehensive process monitoring system and a software is developed to implement such a system.

2. Methodologies The proposed process monitoring system extracts a set of selected features from properly processed sensor signals, normalizes the features to minimize their sensitivity to varying cutting conditions, and then feeds the normalized features into predictive models to forecast process conditions. The architecture of the proposed monitoring system is shown in Fig. 1. To construct such a system, three offline training modules are needed: normalization parameter calibration, optimal feature set selection, and predictive model training. This section will elaborate the methodologies used in the three modules.

Fig. 1. Architecture of the process monitoring system. 548

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

preprocessed with different parameters before extracting the candidate features. The optimal preprocessing parameters for each feature are selected as those maximizing the feature’s correlation. Then the size of candidate feature pool is reduced by eliminating the features with low correlations. The procedure is presented in Fig. 2. The correlation criteria adopted in this paper include the followings:

2.1. Feature normalization A signal feature that correlates to process conditions may also be influenced by many other factors, such as machine tool dynamics, sensor mounting positions, cutting conditions, cutting tools, and workpiece materials, which makes it difficult to build a generalizable monitoring system. The purpose of feature normalization is to scale a feature based on the operating conditions to minimize its dependence on nontarget factors while preserving its sensitivity to the target process conditions. Assuming that the raw signal feature before normalization is sf and the normalized feature is sfN, the normalization scheme is proposed as Eq. (1):

sfN =

• The coefficient of determination (R ) of a regression model for tool 2

wear prediction. To avoid the assumption that the feature has a linear correlation to tool wear, nonlinear models, such as quadratic or exponential models, could be used [28]. The R2 measures how well the feature can predict tool wear through the regression model, as defined in Eq. (3):

sf (V , f , ap ) V α V0

β

( )( )( )

sf0 *

f f0

ap

R2 = 1 −

γ

ap,0

(1)

where V, f and ap are the operating cutting speed, feedrate and depth of cut at which sf is extracted, sf0 is the feature value at the nominal cutting condition V0, f0 and ap,0, and α, β, γ are parameters to be identified for each feature. The three ratio terms on the denominator are used to compensate for the influence of varying cutting conditions. The parameters (sf0, α, β, γ) in the proposed scheme can be experimentally calibrated in two steps:

•

1) On a specific machine tool setup and for a specific cutting toolworkpiece combination, the nominal feature value sf0 can be extracted from a cutting test at the nominal condition of V0, f0 and ap,0. The test should be conducted using a fresh tool without chatter or chipping. 2) By varying cutting conditions around the normal condition and extracting features for each condition, the parameters (α, β, γ) can be found through the optimization in Eq. (2):

findα, β, γ to minimize

1 n

n

∑i = 1 [sfN , i (α, β , γ ) − 1]2

∑i (yi − y )2

(3)

where yi is the measured tool wear, y is the mean of yi, and yˆi is the predicted tool wear. The F-statistic from an ANOVA test. The chatter detection and tool chipping detection are binary classification problems. The one-way ANOVA test can be used to assess a feature’s fitness for classification. The F-stat indicates a feature’s statistical significance with respect to the class variables, i.e., the labels of chatter or tool chipping status, as defined in Eq. (4):

F=

•

∑i (yi − yˆi )2

variance of class means mean of within class variances

(4)

Calculations of the numerator and denominator can be found in [34]. A high F-stat indicates that the feature has distinct means between classes and thus is suitable for classification. The classification accuracy of a single-input linear SVM classifier. This criterion measures a feature’s linear separability, i.e., how well the samples with and without chatter, or before and after chipping, can be separated by setting a linear threshold for the feature.

(2) Although the procedure in Fig. 2 is exhaustive, since the training of complicated predictive models is not involved, it can be executed efficiently in an automated computer program, which is scalable to large candidate feature pools and signal preprocessing parameter pools.

where n is the total number of feature samples extracted from all cutting conditions. Once calibrated, the parameters (sf0, α, β, γ) can be stored in a database and retrieved for online normalization whenever the corresponding cutting tool-workpiece combination is being used. The normalized features will always be close to 1 for fresh tools, regardless of cutting conditions, and only deviate from 1 because of the growth of tool wear or the onset of chatter or tool chipping.

2.2.2. Frequency band analysis If a feature’s correlation to a process condition depends on frequency bands, the procedure in Fig. 2 only gives the single best passband for that feature. In this step, the effect of frequency band is analyzed more systematically with the following three manipulations:

2.2. Feature selection Even in a sensor fusion paradigm for process monitoring, feeding a redundant set of features into a machine learning model might not improve, or could even degrade the model’s performance. Instead, a compact feature set that best predicts the target process conditions should be selected. To carry out this task, the systematic feature selection procedure in [33], which is initially proposed for tool wear monitoring, is extended to chatter and tool chipping monitoring in this paper. 2.2.1. Preliminary feature selection By reviewing the literature, a considerable amount of signal features that are potentially useful for process monitoring can be found. The feature’s correlation to a certain process condition could be dependent on the signal preprocessing parameters. Different features combined with different parameters yield a large amount of choices, which need to be downselected by a computationally efficient method. Therefore, the feature’s correlation to a target process condition is evaluated based on some model-independent criteria in this step. A sensor signal is first

Fig. 2. Flow chart of the preliminary feature selection procedure. 549

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

models, and will not be excessively time-consuming since the size of downselected candidate feature pool N is small.

2.2.2.1. Merging frequency bands. If a feature has good correlations in several frequency bands, merging these separate bands into a larger one gives a more compact feature set and might enhance the feature’s performance. The analysis of variance (ANOVA) test is used to compare the feature’s statistical significance in merged and original bands. If a higher F-statistic and a lower p-value are achieved by the feature in the merged band, then merging the bands could be beneficial. For tool wear prediction, an ANOVA test is applied to the single-input regression model between feature and tool wear. For chatter detection and tool chipping detection, one-way ANOVA test is conducted with the class variable: stable (-1) and chatter (1), or before chipping (-1) and after chipping (1).

2.3. Predictive model training The tool wear prediction is a regression problem to track the progressive growth of tool wear, while the chatter and tool chipping detections are binary classification problems that map the signal features into chatter or chipping status, and thus they are dealt with different machine learning models: 2.3.1. Type-2 FBFN for tool wear prediction The machine learning models for tool wear protection are usually trained to minimize the error between measured and predicted tool wear. However, these error-based learning models cannot estimate the uncertainties in the predictions. Hence, the type-2 fuzzy basis function network (type-2 FBNF) proposed in [36] is adopted in this work to predict the tool wear with uncertainty bounds. Type-2 FBFN is a two-layer network constructed using a fuzzy inference system, with interval type-2 fuzzy sets at the output layer. A type-2 FBFN with n input variables and J fuzzy rules can be formulated to approximate a nonlinear function in an interval form, as shown in Eq. (8):

2.2.2.2. Multi-band analysis. Instead of merging bands, another strategy is to use the same type of feature in multiple frequency bands to predict process conditions. Taking the tool wear prediction as an example, the multi-band regression can be formulated as Eq. (5):

x i = log (Feature@band#i) n y = a0 + ∑i = 1 ai x i ⎧ ⎨ ⎩ y is the measured tool wear

(5)

Using multiple bands (n > 1) might give better predictions than using a single band (n=1). For chatter and tool chipping detection, the regression model can be replaced with a SVM classifier.

J J y͠ = [yl , yu ] = [pwl , pwu] = ⎡∑ j = 1 wl j pj (x), ∑ j = 1 wuj pj (x) ⎤ ⎦ ⎣

2.2.2.3. Band feature fusion. Different features in a same frequency band may contain information complementary to each other, and thus fusing these features may better describe the system behavior at that band. Assume that three features are selected for fusion, a permutation pool of multiplication and/or division of the features can be created as in Eq. (6):

x i = log (Feature#i@the band) V=

{x i , x i2 ,

pj (x) =

(6)

Taking the tool wear prediction as an example, the feature fusion can be assessed using linear regressions that choose terms from the permutation pool, as shown in Eq. (7):

vi ∈ V n y = a0 + ∑i = 1 ai vi ⎧ ⎨ ⎩ yis the measured tool wear

(8)

where x1, x2, …, xn are the input variables, pj (x) is the fuzzy basis function in the j-th fuzzy rule, μi j (x i ) is the fuzzy membership function, [yl , yu ] are the lower and upper bounds of the output, and [wl , wu] are the lower- and upper-bound weighting factors in the sense of neural network, or the end points of interval type-2 fuzzy sets in the sense of fuzzy logic [33]. The type-2 FBFN is constructively trained using the generic algorithm (GA) [35] and quadratic programming [36] to capture the uncertainty bounds in both tool wear measurements and inputoutput (feature-wear) mapping so that its prediction interval will always enclose the tool wear measurement bounds, and hence the actual tool wear value. For more details about the type-2 FBFN and its training, see [36].

i = 1, 2, 3

x i x j , x i / x j , x i x j xk , x i x j /xk } i, j, k ∈ {1, 2, 3}, i ≠ j ≠ k

∏in= 1 μi j (xi ) j ∑Jj = 1 (∏in= 1 μi (xi ))

(7) 2

A potential fusion of permutation terms is identified if R of the regression is higher than those of single-term regressions. ANOVA test is also applied to ensure each term in the regression model is above a significance level (p < 0.01). Since a compact feature set is desired, n in Eq. (5) and Eq. (7), which controls the number of bands and features, should be small (n≤3).

2.3.2. SVMs for chatter and tool chipping detections The support vector machine (SVM) developed by Vapnik is a learning machine for two-class classification problems [37]. A SVM classifier constructs an optimal hyperplane wT x + b = 0 with the maximum margin to separate data samples from two classes (y=+1 or y=‒1), as in Eq. (9):

2.2.3. Optimal feature set selection Assuming that N candidate features in their most suitable frequency bands have been selected after the first two steps, the optimal feature set, which yields the best performance of a predictive model, can be chosen from the N candidates using a backward feature elimination method:

n

Minimize

1 2

Subject to

T ⎧ yi (w x i + b) ≥ 1 − ξi ⎨ ξi ≥ 0 ⎩

|| w ||2 + C ∑i = 1 ξi

(9)

where n is the number of training samples, ξi is the slack variable to allow misclassifications, and C is a parameter to control over-fitting. A SVM predicts which class the input sample x belongs to as y=+1 if wT x + b ≥ 1, and y=‒1 if wT x + b ≤ 1. If the data are not linearly separable, a nonlinear transformation ϕ (x) can be used to construct the hyperplane wT ϕ (x) + b = 0 in a high-dimensional space. The performance of an SVM classifier is primarily measured by the classification accuracy. Another useful performance index is the mean classification margin, as defined in Eq. (10). An SVM with larger margin would be more robust against noises and uncertainties.

(1) Exclude one feature from the N candidates, use the remaining N-1 features to train a model and record the performance index of the model. (2) Repeat (1) for all the N features and remove the feature whose N-1 complementary set has the best performance index. (3) Set N = N-1 and repeat (1) ∼ (2) until the minimum number of features left. For tool wear models, the performance index could be the coefficient of determination (R2) or mean square error (MSE). For chatter and tool chipping classifiers, the performance index could be the classification accuracy and margin. This procedure can be applied to any

n

m= 550

∑i = 1 2yi (wT ϕ (x i) + b) n

(10)

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

Table 2 Cutting conditions of normalization tests.

3. Instrumentation, experiment design and feature extraction 3.1. Instrumentation setup The instrumentation setup is the same as the one used in [33]. As shown in Fig. 3, a CNC lathe was instrumented with five sensors: A Halleffect power meter was installed on the spindle motor drive to measure the input power to the spindle unit; two accelerometers, denoted as AccY and AccZ, were mounted on the tool holder to measure the cutting tool vibrations in Y-(cutting) and Z-(feed) directions; two AE sensors, denoted as AEL and AEH, were also attached to the tool holder to measure the acoustic emissions in low-frequency (50–200 kHz) and high-frequency (100–900 kHz) ranges. The signals of power meter and accelerometers were sampled at 10 kHz, while the signals of AE sensors were sampled at 2 MHz.

3.2.1. Tool wear and tool chipping experiments The tool wear experiments were the same as those designed in [33]. As shown in Table 1, three workpiece materials were used: hardened 4140 steel (HRC35), Inconel 718 and Ti-6Al-4V. The 4140 steel and Ti6Al-4V were machined with uncoated carbide inserts (Kennametal SPGN 422, grade K68), while the Inconel 718 was machined with ceramic inserts (Greenleaf SNGN 452, grade WG-300). Coolant was not used during the experiments. The depths of cut were selected to ensure that all tests were free of chatter, even with large tool wear. The length of each tool path was 2.5 inch. After each tool path, tool wear was measured with five repetitions. The mean of five measurements was used as the nominal tool wear, and the minimum and maximum were used as the measurement uncertainty bounds. An insert would be replaced when flank wear reached 200 μm or when tool chipping was observed.

Test Test Test Test Test Test Test Test Test Test Test

#1 #2 #3 #4 #5 #6 #7 #8 #9 #10 #11

Depth of Cut (mm)

Usage

Wear Type

4140 Steel 4140 Steel 4140 Steel 4140 Steel Inconel 718 Inconel 718 Inconel 718 Inconel 718 Ti-6Al-4V Ti-6Al-4V Ti-6Al-4V

90 90 90 120 120 120 100 120 100 125 100

0.20 0.10 0.20 0.20 0.075 0.075 0.075 0.10 0.075 0.075 0.10

0.25 0.25 0.15 0.25 0.25 0.15 0.25 0.25 0.75 0.75 0.75

Training Training Training Training Validation Validation Validation Validation Training Training Validation

Flank Flank Flank Flank Flank Flank Flank Flank Crater Crater Crater

Depth of Cut (mm)

4140 Steel

80, 90, 100

0.15, 0.25, 0.35

Inconel 718

80, 100, 120

Ti-6Al-4V

75, 100, 125

0.100, 0.150, 0.200, 0.250 0.050, 0.075, 0.100, 0.125 0.050, 0.075, 0.100, 0.125

0.15, 0.25, 0.35 0.25, 0.50, 0.75, 1.00

3.3. Feature extraction To leverage the process-relevant information embedded in collected signals, different signal processing schemes are applied, and a variety of signal features are extracted in the time-domain, frequency-domain and time-frequency domain, as discussed below for each physical sensor: 3.3.1. Power meter The power meter outputs an averaged power signal with only lowfrequency content, and thus only time-domain features are extracted, which include the mean, RMS, standard deviation (STD), skewness, kurtosis, peak-to-peak amplitude and crest factor. The mean describes the overall trend of spindle power; thus, the power signal is smoothed with a moving average filter. All the other features describe the spindle power fluctuation at different orders; thus, lowpass filters are used to remove high-frequency noise while preserving the fluctuation below cut-off frequencies.

Table 1 Design of tool wear experiments. Feed (mm/ rev)

Feedrate (mm/rev)

3.2.2. Chatter experiments To prevent the testbed being damaged by the strong vibration and cutting force during chatter, chatter experiments were conducted using annealed 4140 steel workpieces (20 HRC) and the depth of cut was increased carefully by 0.1 mm each tool path, until chatter was observed. The workpiece was 1.5 in. in diameter and 3.5 in. in overhang length. The spindle speed was fixed at 1000 rpm, and the feedrate was fixed at 0.1 mm/rev. The surface roughness of machined surface was measured after each tool path to check the occurrence of chatter, since the chatter marks would significantly degrade the surface finish. The chatter experiments used for classifier training are summarized in Table 3. To validate the chatter classifier, the experiment data of two tool passes in [38], as shown in Table 4, were used, which were collected from a different machine tool setup.

3.2. Experiment design

Cutting Speed (m/min)

Cutting Speed (m/ min)

The crater wear in Ti-6Al-4 V machining weakens the cutting edge as it leads to larger positive rake angle and smaller included angle. Hence, the tool chips after a certain amount of crater wear. Chipping was observed in Test#9 ∼ Test#11, and the signal data around chipping onset were used to build the tool chipping classifier, without designing specialized tool chipping experiments. The normalization parameters were calibrated for each candidate feature using the procedure in Section 2.1. The nominal conditions for the three materials were those in Test #1, #5 and #9, and the cutting conditions to determine the coefficients α, β, γ were listed in Table 2.

Fig. 3. Diagram of the instrumentation setup.

Material

Material

Table 3 Design of chatter experiments for training.

Test#1 Test#2 Test#3 Test#4

551

Depth of Cut (mm)

Surface Roughness (μm)

Status

0.1 0.2 0.3 0.4

1.8 1.8 1.4 5.0

Stable Stable Stable Unstable

± ± ± ±

0.2 0.1 0.1 0.5

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

Table 4 Operating conditions of the chatter data set for validation [38]. Cutting Tool Workpiece Spindle Speed Feed Stable cut Unstable cut

Kennametal SPGN 422, uncoated insert AISI 4140 steel, 40.19 mm in diameter, tailstock used 1250 rpm 0.0965 mm/rev Depth of cut:1.424 mm, surface roughness: 2.56 μm Depth of cut: 2.136 mm, surface roughness: 9.54 μm

3.3.2. Accelerometers Time-domain features are the same as those of the power meter. Frequency-domain features are extracted from the power spectral density (PSD) of signals, which include the PSD mean, peak amplitude, frequency centroid, PSD moment of inertia and PSD entropy. Timefrequency-domain features are extracted as the powers of wavelet coefficients in a 4-level wavelet packet decomposition (WPD), which reflect the signal’s strength at each WPD node. Accelerometer signals are preprocessed with bandpass filters to inspect the features’ banddependent correlation, except for wavelet features, since a WPD node inherently corresponds to a frequency band. Some specialized chatter detection features found in the literature are also extracted, including the wavelet parameter γ [19], waveform irregularity coefficient [20], permutation entropy [21] and spectral coherence of two crossed accelerations [18]. In addition, to depict the sudden increase of acceleration signal amplitude caused by tool chipping, the jerk, which is the time derivative of acceleration, is calculated and time-domain features were extracted from the jerk signal.

Fig. 4. Comparison of the mean power feature before and after normalization.

spindle power as an example, as shown in Fig. 4, feature values in different cutting conditions were collapsed into a uniform range via normalization, so that the feature’s correlation to tool wear could be assessed using a same regression model. The preliminary feature selection was applied to the normalized features to determine the best signal preprocessing parameters for each feature and downselect the features. After eliminating the features with R2 lower than 0.5 and excluding the sensors that only contributed features with moderate correlations (0.5 < R2 < 0.7), the features from the power meter and ydirectional accelerometer (AccY), both of which provided highly correlated features, were further investigated. The frequency band analysis was applied to the AccY features (RMS, PSD mean, frequency centroid and PSD moment of inertia, and wavelet coefficient power). As can be seen from Table 6, the multi-band regression of RMS, the band feature fusion of PSD mean, frequency centroid and PSD moment of inertia, and the multi-node regression of wavelet coefficient power, all with n = 3, provided considerably better tool wear predictions, in terms of R2 and root mean square error (RMSE), than those with single band or single feature (n = 1). The 9 AccY features in the n = 3 cases in Table 6 and the mean spindle power composed the 10 candidates for optimal feature set selection. The features of 4140 steel and Inconel 718 might have different slopes with respect to flank wear, due to the difference of material properties. To use them in a same flank wear model, the difference of slopes was compensated by Eq. (11):

3.3.3. AE sensors Time-domain features are the same as those of the power meter. Since AE signals are sampled at 2 MHz, applying any complicated transforms online to such high-frequency signals will not be computationally practical. Hence, only some basic frequency-domain features are extracted, which include the frequency centroid, peak amplitude and peak frequency. The candidate signal preprocessing parameters to be assessed for candidate features are listed in Table 5. Signals are downsampled with different rates, preceded with proper digital anti-aliasing filters, to investigate the effect of sampling rate.

sfn,718 = (sfn,718 − 1)*k 4140/ k 718 + 1

4. Results and discussion

(11)

where sfn,718 is the feature of Inconel 718, k4140 is the baseline slope of 4140 steel, and k718 is the original slope of Inconel 718. Through a backward feature elimination procedure, 4 features were selected to form the optimal feature set, which included the power mean, the AccY RMS in 375–475 Hz and 500–700 Hz bands and the frequency centroid in 75–775 Hz band. The first three features all increase with tool wear since the tool wear tends to increase the spindle power consumption and tool vibration amplitude. The frequency centroid also increases slightly with tool wear, due to the shift of spectrum center caused by tool wear. With the optimal feature set, a flank wear model was trained using the 4140 steel test data (815 samples) and validated using the Inconel 718 test data (376 samples), and a crater wear model was trained using the Ti-6Al-4 V test data (354 samples). Since the cutting tool is never

To develop a comprehensive monitoring system, the methodologies discussed in Section 2 are applied to all the three tasks: tool wear prediction, chatter detection, and tool chipping detention. The results of the three tasks are presented respectively, and then integrated in a unified framework for implementation. The development of tool wear prediction scheme has been elaborated in [33], but for integrity and clarity of this paper, some key results are briefly presented below. 4.1. Tool wear prediction Before selecting features for tool wear prediction, all candidate features were normalized using the scheme in Eq (1). Taking the mean Table 5 Candidate signal preprocessing parameters for candidate features. Sensor

Power Power AccY & AccZ AEL AEH

Features

Mean Others All except WPD features All All

Candidate Preprocessing Parameters Down-sampling rate (Hz)

Filter Type

Filter Parameter

500, 200, 100, 50, 25 500, 200, 100, 50, 25 5000, 3333, 2500 500k, 333k 1000k, 500k

Moving average Low-pass Band-pass Band-pass Band-pass

Window size (s): 0.1, 0.2, 0.5, 1, 2, 5 Cut-off frequency (Hz): 5, 10, 25, 50, 100, 200 Passband (Hz): 0–100, 25–125, …, 900–1000, …, 2400–2500 Passband (Hz): 50k–100k, 100k–150k, 150k–200k Passband (Hz): 100k–150k, 150k–200k, …, 850k–900k

552

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

Table 6 Results of frequency-band analysis.

n=1 n=3 n=1 n=3 n=1 n=3

Features

R2

RMSE (μm)

RMS 400-700 Hz RMS 25-125 Hz; RMS 375-475 Hz; RMS 500-700 Hz WPD Node [3,2] WPD Node [2,1]; WPD Node [4,1]; WPD Node [3,5] PSD Mean 500-700 Hz PSD Mean; Freq. Centroid; PSD Moment of Inertia 75775 Hz

0.89 0.92 0.86 0.90 0.88 0.92

16.29 13.46 16.86 15.08 16.40 13.71

Fig. 6. Comparison of model prediction upper bounds and tool wear measurement upper bounds, and tool change margin when tool change alert issued.

self-healing, a latching mechanism was applied to the model outputs so that once the predicted tool wear reached a certain level, it would never drop below that level afterward. As shown in Fig. 5, the nominal predictions of type-2 FBFN models tracked the nominal tool wear measurements closely with an R2 value of 0.97, and more importantly, the prediction bounds always enclosed the uncertainty bounds of tool wear measurements. Therefore, when the prediction upper bound had reached a preset tool life limit, the actual tool wear was still slightly below that limit, which enables conservative but reliable decision making regarding tool changes. The outputs of type-2 FBFN models were postprocessed to report the tool wear information to users in a more concise form: 1) The tool life span was divided into N quantized classes, among which class N indicated the highest degree of tool wear; 2) The nominal prediction was assigned into corresponding classes to visualize the tool wear level; 3) When the nominal prediction got into class N, a yellow alert would be issued; 4) When the upper-bound prediction got into class N, an orange alert would be issued; 5) Finally, once the upper-bound prediction reached the tool life limit (200 μm in this work), a red alert would be issued to request a tool change. As shown in Fig. 6, the type-2 FBFN model issued the red alert with a margin of 6.2% on average and 11.6% at the maximum. The margin indicated that the average level of

uncertainty was about 6% in this work. Based on this uncertainty level, the number of classes was selected as N = 10, as discussed in [33], with which a classification accuracy of 88.8% was achieved by the nominal model predictions. 4.2. Chatter detection The chatter detection was studied by following the same procedure as tool wear prediction. By extracting features from signals every 0.1 s, 220 stable samples and 60 chatter samples were obtained from the chatter experiments. In preliminary feature selection, features with pvalue larger than 0.01 in ANOVA test (poor statistical significance) and classification accuracy less than 90% with single-input linear SVMs (poor linear separability) were removed. The features in the reduced candidate pool, and their F-stat with the optimal signal preprocessing parameters, were grouped by sensors and listed in Table 7. The two accelerometers, AccY and AccZ, contributed most of the features, and AccY features had higher F-stat than AccZ features. Therefore, it is reasonable to conclude that AccY is the most useful sensor for chatter detection in this work. The selected AccY features could be classified

Fig. 5. Tool wear prediction for selected tool wear tests. 553

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

Table 7 Candidate chatter detection features after preliminary feature selection. Features

F

Features

F

AccY permutation entropy AccY PSD entropy AccY RMS/STD AccY PSD moment of inertia AccY peak to peak AccY frequency centroid AccY PSD mean AccY PSD peak amplitude AccY waveform irregularity AccY wavelet coefficient power AccY wavelet γ AEL PSD moment of inertia AEL RMS/STD

15951.3 13091.8 7274.5 5729.6 4969.9 3433.3 3375.4 3028.3 2213.7 1047.6 756.3 437.9 382.2

AccZ permutation entropy AccZ PSD entropy AccZ RMS/STD AccZ PSD moment of inertia AccZ peak to peak AccZ frequency centroid AccZ PSD mean AccZ PSD peak amplitude AccZ waveform irregularity AccZ wavelet coefficient power AccZ wavelet γ AccY & AccZ coherence AEH frequency centroid

4875.2 4883.1 810.2 4742.3 912.5 3748.5 817.1 872.5 996.2 454.7 397.5 886.3 433.7

Fig. 8. Power spectral density (PSD) of the AccY signal. Table 8 Dimensionless candidate features for chatter detection.

into two groups: 1) the magnitude-based features used to describe the change of signal amplitude caused by chatter, including the RMS, standard deviation, peak to peak amplitude, PSD mean, PSD peak amplitude and wavelet coefficient power; 2) the non-magnitude-based features used to depict the signal’s energy distribution patterns and waveform characteristics, including frequency centroid, PSD moment of inertia, PSD entropy, wavelet parameter γ, waveform irregularity coefficient, permutation entropy and spectral coherence of AccY and AccZ. The two groups of features were fed into a frequency band analysis, in which the features’ F-stat with different band sizes were compared, as shown in Fig. 7. It was observed that the features in the first group had the highest F-stat in a narrow band around the chatter frequency, because the increase of vibration magnitude caused by chatter was most significant at the chatter frequency. In contrast, the features in the second group had higher F-stat in wider bands, since the signal’s overall energy distribution pattern was better characterized. As shown in Fig. 8, with chatter, the signal had intensively increased magnitude (note the difference of y-axis scales) at the chatter frequency (around 1000 Hz) and the spectrum exhibited a singleton pattern distinct from the chatter-free case. Using the magnitude-based features in the first group for chatter detection is feasible, but with the following issues: The features’ dependence on cutting conditions need to be removed by feature normalization; the features’ preference of narrow band around chatter frequency should be satisfied by chatter-frequency identification and bandpass filtering of signals; also, the features’ correlation to tool wear may lead to coupling between the chatter classifier and tool wear model. Although these issues are resolvable, using the non-magnitudebased features in the second group is much preferable, because these issues can be avoided rather than to be resolved. In the second group, the frequency centroid is chatter-frequency dependent, and the spectral

Feature

Interpretation and Trend

Normalized PSD moment inertia Normalized PSD entropy

Depict the signal power distribution about the frequency centroid, decrease with chatter, see Fig. 8 Measure the complexity of a signal’s spectrum, decrease with chatter, see Fig. 8 Ratio of a signal’s zero-crossing count to peak-valley count, increase to 1 with chatter Measure a signal’s complexity as a time series, decrease with chatter Depict the signal power distribution among wavelet scales, increase with chatter [19]

Waveform irregularity coefficient Permutation entropy Wavelet parameter γ

coherence requires using two sensors, but its performance is not superior to the others. Therefore, only the five features listed in Table 8 would be considered, all of which are inherently independent of cutting conditions, chatter frequency and tool wear. The moment of inertia and entropy extracted from normalized PSD (Pxx n = Pxx / ∑ Pxx , where Pxx is the raw PSD) decrease with chatter, because the spectral power is concentrated to the chatter frequency. The permutation entropy decreases with chatter, while the waveform irregularity coefficient rises to 1 at the onset of chatter, because the time-domain signal is dominated by the periodic vibration at the chatter frequency. The chatter classifier was built by choosing inputs from the features in Table 8 to train SVMs with Gaussian kernel. 280 samples obtained from the experiments in Table 3 were used for training, while 32 samples obtained from [38] (16 stable samples and 16 chatter samples) were used for validation. The optimal feature sets with one, two and three features, in terms of the classification accuracy for the validation data set, are listed in Table 9. Since chatter induced distinct aberrances in sensor signals, a 100% accurate classification could be achieved for the training data set with one feature (the normalized PSD entropy). Introducing more features would increase the classification margin, which made the classifier more robust against noise and uncertainties. Table 9 Optimal feature sets for chatter detection (Classification accuracy for training set are 100% for all feature sets).

Fig. 7. Statistical significance of AccY features for chatter detection. 554

Feature set

Mean classification margin of support vector samples

Classification accuracy for validation set

Normalized PSD entropy Normalized PSD entropy Permutation entropy Normalized PSD entropy Permutation entropy Wavelet parameter γ

1.46

93.75%

1.92

93.75%

1.96

96.88%

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

Fig. 9. Distribution of features selected for chatter detection.

Fig. 11. Distribution of features selected for tool chipping detection.

these three features: normalized PSD entropy, permutation entropy and wavelet parameter γ. The data distribution of selected features set is shown in Fig. 9. Considering that the validation data set was collected from a different machine tool setup with different chatter frequency and cutting conditions [38], the 96.88% accuracy implied that the classifier had a generalized applicability. 4.3. Tool chipping detection The tool chipping data were obtained from Test#9 ∼ Test#11 in Table 2. As shown in Fig. 10, the stripping off of tool materials dulled the cutting edge and increased the magnitudes of tool vibration. To describe the increase of vibration magnitude, the jerk signal was calculated by taking derivative of the acceleration signal. By extracting features every 0.1 s from the 5 s signals before chipping and 5 s signals after chipping, 300 samples were obtained. The candidate features with p-value larger than 0.01 in ANOVA test and classification accuracy less than 75% with single-input linear SVMs were eliminated by the preliminary feature selection program. Features in the reduced candidate pool were all magnitude-based features from AccY and AccZ. The best feature was found to be the wavelet coefficient power of AccZ, which achieved a F-stat of 598.4 and a classification accuracy of 92.0%. By

Fig. 10. Observations about tool chipping. (a) Top-view image of rake face; (b) Tool vibration signal.

Since the feature set with three features provided the highest accuracy for the validation set (96.88%, only 1 out of the 32 samples was misclassified) and largest margin, it was decided to build the classifier with 555

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

Fig. 12. Diagram of the integrated monitoring system for implementation.

Fig. 13. User Interface of the software and monitoring results of the tool wear example.

with a single sensor. To pursue tool chipping detection with only one sensor, a normalization scheme with a moving average window, in addition to the scheme in Eq. (1), was applied. Taking RMS as an example, the RMS feature at instant k was divided by its mean value in a preceding

choosing feature sets from the pool to train SVMs, it was found that a 100% accurate classification could only be achieved by using features from both AccY and AccZ, mainly because the variance of cutting-edge topographies in different tests; for example, the different crater wear and volumes of materials chipped off made the features less separable 556

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

and to 1.79 by using three. As a trade-off between computational cost and classifier robustness, the feature set of two features was selected, which consisted of the AccY RMS and AccY PSD mean. The distributions of selected features are shown in Fig. 11, with the suffix ‘Norm’ indicating the use of normalization in Eq. (12). As a comparison, the optimal feature set without normalization, which consisted of the AccY jerk RMS and AccZ jerk RMS, involved two sensors rather than one, but only attained a margin of 1.68 (smaller than the margin of 1.76 with normalized features). This validated the effectiveness of the normalization scheme introduced in Eq. (12). 4.4. Integration for implementation Fig. 14. Tool wear measurement of 4140 steel machining.

During the development of tool wear prediction unit, chatter detection unit and tool chipping detection unit, considerable efforts have been made to maximize their performance, generalize their applicability, ensure their robustness against uncertainties and minimize their computational cost. The three units were integrated into an implementable framework as shown in Fig. 12. Two sensors are required for implementation: the power meter and y-directional accelerometer (AccY). The tool wear prediction unit preprocesses the power signal and AccY signal with moving average filter and bandpass filter, respectively. Four features are extracted from filtered signals and then normalized to minimize their sensitivities to cutting conditions. The normalization parameters are calibrated from experiments, stored in a database and can be retrieved for use based on the user-specified tool-work combination. Predictions from the type-2 FBFN model are postprocessed to be assigned into quantized tool wear levels, and to decide whether a tool change alert should be issued. The progressive growth of tool wear can be predicted every 1 s, or with a larger interval. In contrast, the fault detections of chatter or tool chipping are more time-sensitive since actions need to be taken immediately to protect the workpiece and machine tool. Therefore, the chatter detection unit and tool chipping detection unit are executed every 0.1 s. These two units only use one sensor (AccY) and don’t require signal preprocessing with bandpass filters, since they don’t have any preference on specified frequency bands, which saves considerable computation time for the time-sensitive applications. The chatter unit extracts three dimensionless features without requiring any normalization, while the tool chipping unit normalizes two magnitude-based features using Eq. (12) with N = 10 (1 s) to capture the sudden change of vibration magnitude. A monitoring software was developed to implement the framework discussed above. The user interface and data acquisition interface were built in LabVIEW. The algorithms associated with the three units were compiled and integrated into the program as DLL (dynamic-link library) files. The user interface of this software is shown in Fig. 13. The software inquires the workpiece material, cutting tool, and cutting conditions being used, and displays the diagnostic information regarding tool wear, chatter and tool chipping, with the options of showing the processed signals, features or model predictions that the user is interested to see. To demonstrate the effectiveness of this software, two implementation examples are provided below. The first example is the tool wear monitoring of 4140 steel machining. The cutting conditions in this example are cutting speed of 90 m/min, feed of 0.2 mm/rev and depth of cut of 0.25 mm. As shown in Fig. 13, the RMS of vibration signal increases with time due to the progressive growth of tool wear, and the software displays monotonically increasing tool wear predictions with uncertainty bounds. A red tool change alert was issued after 8 tool passes, when the upperbound prediction exceeded the tool life limit (200 μm). The tool wear after the 8th tool pass was measured as 190 μm, as shown in Fig. 14. Thus, the monitoring software requested tool change with a 5% margin, which is reasonably conservative to prevent prematured tool failure. The second example is the detection of interrupted chatter. During one tool pass of 4140 steel machining, chatter occurred interruptedly.

Fig. 15. Monitoring results of the chatter example. (Red arrow: unstable cutting; Green arrow: stable cutting).

window, as in Eq. (12):

RMSnorm [k ] =

RMS [k ] 1 N

N

∑i = 1 RMS [k − i]

(12)

where N was the size of window. By introducing this normalization, the RMS, or other magnitude-based features, would become insensitive to the progressive changes of cutting-edge geometry and other uncertainties, but only responsive to the sudden changes like tool chipping, i.e., it stayed close to 1 regardless of crater wear and rose above 1 at the onset of chipping. Hence, the feature’ statistical significance with respect to tool chipping was substantially improved. The best normalized feature (with N = 10) was the RMS of AccY, which achieved the Fstat of 1521.4, the classification accuracy of 100%, and a mean margin of 1.69. The margin could be improved to 1.76 by using two features, 557

Journal of Manufacturing Processes 35 (2018) 547–558

B. Zhang, Y.C. Shin

By inspecting the machined surface, chatter marks were observed at the beginning of the tool path, disappeared shortly in the middle, then occurred again before finally disappearing. As shown in Fig. 15, the chatter status (0=stable, 1=chatter) predicted by the software coincides with the interrupted chatter marks on the machined surface, and thus the effectiveness of the chatter detection unit is validated.

2014;33:476–87. [7] Lee DE, Hwang I, Valente CMO, Oliveira JFG, Dornfeld DA. Precision manufacturing process monitoring with acoustic emission. Int J Mach Tools Manuf 2006;46:176–88. [8] Corne R, Nath C, El Mansori M, Kurfess T. Study of spindle power data with neural network for predicting real-time tool wear/breakage during Inconel drilling. J Manuf Syst 2017;43:287–95. [9] Drouillet C, Karandikar J, Nath C, Journeaux AC, El Mansori M, Kurfess T. Tool life predictions in milling using spindle power with the neural network technique. J Manuf Process 2016;22:161–8. [10] Sick B. On-line and indirect tool wear monitoring in turning with artificial neural networks: a review of more than a decade of research. Mech Syst Signal Process 2002;16:487–546. [11] Ren Q, Balazinski M, Baron L, Jemielniak K, Botez R, Achiche S. Type-2 fuzzy tool condition monitoring system based on acoustic emission in micromilling. Inf Sci (Ny) 2014;255:121–34. [12] Shi D, Gindy NN. Tool wear predictive model based on least squares support vector machines. Mech Syst Signal Process 2007;21:1799–814. [13] Subrahmanya N, Shin YC. Automated sensor selection and fusion for monitoring and diagnostics of plunge grinding. J Manuf Sci Eng 2008;130:031014. [14] Wu D, Jennings C, Terpenny J, Gao RX, Kumara S. A comparative study on machine learning algorithms for smart manufacturing: tool wear prediction using random forests. J Manuf Sci Eng 2017;139:071018. [15] Akhavan Niaki F, Mears L. A comprehensive study on the effects of tool wear on surface roughness, dimensional integrity and residual stress in turning IN718 hardto-machine alloy. J Manuf Process 2017;30:268–80. [16] Quintana G, Ciurana J. Chatter in machining processes: a review. Int J Mach Tools Manuf 2011;51:363–76. [17] Kalpakjian S. Manufacturing processes for engineering materials. 2nd ed. Reading, Mass.: Addison-Wesley; 1991. [18] Li XQ, Wong YS, Nee AYC. Tool wear and chatter detection using the coherence function of two crossed accelerations. Int J Mach Tools Manuf 1997;37:425–35. [19] Choi T, Shin YC. On-line chatter detection using wavelet-based parameter estimation. J Manuf Sci Eng 2003;125:21. [20] Yang F, Zhang B, Yu J. Chatter suppression with multiple time-varying parameters in turning. J Mater Process Technol 2003;141:431–8. [21] Nair U, Krishna BM, Namboothiri VNN, Nampoori VPN. Permutation entropy based real-time chatter detection using audio signal in turning process. Int J Adv Manuf Technol 2010;46:61–8. [22] Kwak JS. Application of wavelet transform technique to detect tool failure in turning operations. Int J Adv Manuf Technol 2006;28:1078–83. [23] Neslušan M, Mičieta B, Mičietová A, Čilliková M, Mrkvica I. Detection of tool breakage during hard turning through acoustic emission at low removal rates. Meas J Int Meas Confed 2015;70:1–13. [24] Balsamo V, Caggiano A, Jemielniak K, Kossakowska J, Nejman M, Teti R. Multi sensor signal processing for catastrophic tool failure detection in turning. Procedia Cirp 2016;41:939–44. [25] Altintas Y, Aslan D. Integration of virtual and on-line machining process control and monitoring. CIRP Ann Manuf Technol 2017;66:349–52. [26] Zhu K, Zhang Y. A cyber-physical production system framework of smart CNC machining monitoring system. IEEE ASME Trans Mechatron 2018:2834622https:// doi.org/10.1109/TMECH.2018. [27] Lee WC, Cheng HC, Wei CC. Development of a machining monitoring and chatter suppression device. 2018 IEEE Industrial Cyber-Physical Systems (ICPS). 2018. p. 404–8. [28] Guyon I, Elisseeff A. An introduction to variable and feature selection. J Mach Learn Res 2003;3:1157–82. [29] Wang J, Wang P, Gao RX. Enhanced particle filter for tool wear prediction. J Manuf Syst 2015;36:35–45. [30] Akhavan Niaki F, Michel M, Mears L. State of health monitoring in machining: extended Kalman filter for tool wear assessment in turning of IN718 hard-to-machine alloy. J Manuf Process 2016;24:361–9. [31] Zhang J, Starly B, Cai Y, Cohen PH, Lee YS. Particle learning in online tool wear diagnosis and prognosis. J Manuf Process 2017;28:457–63. [32] Wu D, Jennings C, Terpenny J, Kumara S, Gao RX. Cloud-based parallel machine learning for tool wear prediction. J Manuf Sci Eng 2018;140:041005. [33] Zhang B, Katinas C, Shin YC. Robust tool wear monitoring using systematic feature selection in turning processes with consideration of uncertainties. J Manuf Sci Eng 2018;140:081010. [34] Neter J, Kutner MH, Nachtsheim CJ, Li W. Applied linear statistical models. Chicago: Irwin; 1996. [35] Lee CW, Shin YC. Construction of fuzzy systems using least-squares method and genetic algorithm. Fuzzy Sets Syst 2003;137:297–323. [36] Ngo PD, Shin YC. Modeling of unstructured uncertainties and robust controlling of nonlinear dynamic systems based on type-2 fuzzy basis function networks. Eng Appl Artif Intell 2016;53:74–85. [37] Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995;20:273–97. [38] Rao BC, Shin YC. A Comprehensive dynamic cutting force model for chatter prediction in turning. Int J Mach Tools Manuf 1999;39:1631–54.

5. Conclusions An intelligent multisensor monitoring system for turning processes has been developed, which is composed of three units: a tool wear predictor, a chatter detector and a tool chipping detector. To construct these units, a set of methodologies are proposed, which includes a practical normalization scheme to eliminate the feature’s dependence on cutting conditions, a systematic feature selection procedure to select the optimal feature sets for each unit, and the training methods of intelligent models with consideration of uncertainties. The effectiveness of the developed units and proposed methodologies has been validated through designed experiments: 1) The tool wear prediction unit built with four systematically selected features works for three types of materials (4140 steel, Inconel 718 and Ti-6Al-4 V), two types of tools (carbide and ceramic), and various cutting conditions. The type-2 FBFN model always encloses tool wear measurements within its prediction interval, and requests tool changes conservatively but reliably with a margin of 6.2% on average and 11.6% at a maximum, according to the tests evaluated in this work; 2) The chatter detection unit built with three dimensionless features that maximize the margin of SVM classifier could detect chatter with an accuracy of 96.88% even when used on a different machine tool setup; 3) The tool chipping detection unit built with two magnitude-based features normalized using a moving average window can always detect tool chipping timely by capturing the increase of vibration magnitude caused by chipping. The three units are implemented in an integrated framework by using two low-cost sensors (a power meter and an accelerometer) and the developed software. This integrated monitoring system exhibits high prediction accuracy for all the three process conditions, generalized applicability in a wide range of operating conditions and satisfactory robustness against process and measurement uncertainties. In addition, it is affordable in instrumentation, efficient in computation and easy to calibrate. Therefore, it is a promising solution for implementation in an industrial environment, as a stand-alone monitoring system or as a component of a closed-loop process control system. References [1] Liang SY, Hecker RL, Landers RG. Machining process monitoring and control: the state-of-the-Art. J Manuf Sci Eng 2004;126:297–310. [2] Teti R, Jemielniak K, O’Donnell G, Dornfeld D. Advanced monitoring of machining operations. CIRP Ann Manuf Technol 2010;59:717–39. [3] Bhattacharyya P, Sengupta D, Mukhopadhyay S. Cutting force-based real-time estimation of tool wear in face milling using a combination of signal processing techniques. Mech Syst Signal Process 2007;21:2665–83. [4] Nouri M, Fussell BK, Ziniti BL, Linder E. Real-time tool wear monitoring in milling using a cutting condition independent method. Int J Mach Tools Manuf 2015;89:1–13. [5] Dimla DE, Lister PM. On-line metal cutting tool condition monitoring. I: force and vibration analyses. Int J Mach Tools Manuf 2000;40:739–68. [6] Bhuiyan MSH, Choudhury IA, Dahari M. Monitoring the tool wear, surface roughness and chip formation occurrences using multiple sensors in turning. J Manuf Syst

558