- Email: [email protected]
Drill wear monitoring in cortical bone drilling
ARTICLE IN PRESS
JID: JJBE
[m5G;April 16, 2015;19:21]
Medical Engineering and Physics 000 (2015) 1–7
Contents lists available at ScienceDirect
Medical Engineering and Physics journal homepage: www.elsevier.com/locate/medengphy
Drill wear monitoring in cortical bone drilling Tomislav Staroveski a,∗, Danko Brezak b, Toma Udiljak a a
Department of Technology, Faculty of Mechanical Engineering and Naval Architecture, University of Zagreb, Ivana Lucica 5, Zagreb, Croatia Department of Robotics and Production System Automation, Faculty of Mechanical Engineering and Naval Architecture, University of Zagreb, Ivana Lucica 5, Zagreb, Croatia
b
a r t i c l e
i n f o
Article history: Received 27 October 2014 Revised 9 February 2015 Accepted 27 March 2015 Available online xxx Keywords: Medical drill wear Thermal osteonecrosis Neural networks Computational modelling Medical devices
a b s t r a c t Medical drills are subject to intensive wear due to mechanical factors which occur during the bone drilling process, and potential thermal and chemical factors related to the sterilisation process. Intensive wear increases friction between the drill and the surrounding bone tissue, resulting in higher drilling temperatures and cutting forces. Therefore, the goal of this experimental research was to develop a drill wear classification model based on multi-sensor approach and artificial neural network algorithm. A required set of tool wear features were extracted from the following three types of signals: cutting forces, servomotor drive currents and acoustic emission. Their capacity to classify precisely one of three predefined drill wear levels has been established using a pattern recognition type of the Radial Basis Function Neural Network algorithm. Experiments were performed on a custom-made test bed system using fresh bovine bones and standard medical drills. Results have shown high classification success rate, together with the model robustness and insensitivity to variations of bone mechanical properties. Features extracted from acoustic emission and servomotor drive signals achieved the highest precision in drill wear level classification (92.8%), thus indicating their potential in the design of a new type of medical drilling machine with process monitoring capabilities. © 2015 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction Along with the development of new surgical approaches and techniques, bone drilling has become mostly routine and continuously increasing medical intervention. Bone tissue is formed by organic and mineral phases whose interactions result in complex mechanical and thermal properties. All bone drilling interventions include heat generation due to the friction between tool, bone and chips, which can significantly influence the post-operative recovery. Hence, novel drilling techniques have been used to study the interaction between the drill and bone in order to reduce drilling forces and improve chip removal from the drilling site [1–3]. Although mechanisms of thermal bone damages are still not thoroughly explained, it is a known fact that temperature rise in the drilling zone can cause thermal osteonecrosis, which prevents quality tissue healing and bone regeneration [4,5]. There are several factors influencing drilling temperature variations and potential occurrence of thermal osteonecrosis: mechanical characteristics of the bone and its cortical thickness [6,7], drill design and geometry [8–11], inefficient external cooling effect due to low bone thermal conductivity [12], inadequate cutting speeds and feed rates [13–15], improper tool position/trajectory which increases fric-
∗
Corresponding author. Tel.: +385 1 6168341; fax: +385 1 6156940. E-mail address: [email protected], [email protected] (T. Staroveski).
tion between the drill body and the hole surface, and drill wear. Drill wear is among those factors with the highest impact on heat generation during bone drilling. It is an unavoidable and irreversible process which increases friction in the cutting zone. Besides negative thermal impact, it causes higher cutting forces and tool vibrations, which can result in the cutting edge breakage, or complete drill breakage in the flutes or shank zone. These situations cause mechanical bone damages, which also have negative impact on the post-operative therapy progress. Several studies of drill wear effect on temperature variations conducted in the field of oral and maxillofacial surgery point to the strong and proportional relationship between tool wear dynamic and bone drilling temperature rise [16–19]. At the same time, some researchers emphasised the frequent usage of worn drills in medical interventions, as well as absence of hospital standards and activities focused on drill wear identification and tool labelling [20]. It is therefore logical to presume that precise and reliable drill wear models could become important elements in the reduction of the occurrence of bone tissue damages. Their role would be particularly important in the development of semi- or completely automated advanced medical drilling systems. In the past 25 years, numerous studies on drill wear monitoring have been published, mainly focused on industrial applications [21]. Drilling process dynamic in industrial applications usually differs from that related to the bone drilling in terms of different types
http://dx.doi.org/10.1016/j.medengphy.2015.03.014 1350-4533/© 2015 IPEM. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
JID: JJBE 2
ARTICLE IN PRESS
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
of drills, bone vs. industrial workpiece material characteristics, machining parameters, cooling methods, etc. Nevertheless, methodology used in designing industrial tool wear monitoring systems based on multi-sensor approach and nonlinear modelling could also be potentially applicable in medical drill wear identification. However, to our knowledge, there is still a lack of tool wear monitoring solutions specifically designed for medical applications, or experimental studies involving the implementation of some already designed and tested industrial monitoring systems in bone drilling applications. Although several models for temperature estimation in bone drilling have been proposed [22,23], none of them includes drill wear identification. There are two major obstacles which prevent precise tool wear level quantification. Direct measurement of tool wear during the cutting process is not possible due to constant contact between bone and drill cutting edges. It can only be estimated using tool wear features extracted from different types of process signals and other known machining parameters (cutting speed, feed rate, drill characteristics). Additionally, industrial applications have shown that tool wear is usually highly nonlinear and sometimes even a partially stochastic process. Similar characteristics can be expected in bone drilling due to the complex interrelation of all aforementioned process parameters. The complexity of wear process in industrial applications motivated many researchers to experiment with different types of computational intelligence algorithms, primarily artificial neural networks, to build reliable and accurate wear models. Artificial neural networks gained such popularity because of their nonlinear modelling capabilities based on parallel processing and integration of system/process data (in this case, drill wear features and machining parameters). Taking all this into consideration, the aim of this experimental study was to analyse the performances of the multi-sensor based medical drill wear classification model, and its potential for application in the next-generation medical drilling machines. Drill wear features were extracted from force, current and acoustic emission signals, and then processed using Radial Basis Function Neural Network (RBF NN), which is known for short learning procedure, as well as simple and quick hidden layer structure adaptation. 2. Methods 2.1. Experimental setup and parameters Experimental work has been performed using the custom-made 3-axis bench-top mini milling machine adjusted for the purpose of the bone drilling research (Fig. 1). The machine has been retrofitted with the 0.4 kW (1.27 N m) permanent magnet synchronous motors with integrated incremental encoders (type Mecapion SB04A), corresponding servomotor drives (DPCANIE-030A400 and DPCANIE-060A400), ball screw assemblies and LinuxCNC open architecture control (OAC) system. Apart from the control loop referent currents taken from the main spindle and feed drives, cutting forces have been measured using triaxial Kistler piezoelectric dynamometer 9257B coupled with 5017B charge amplifier, and acoustic emission (AE) signals using Kistler industrial sensor type 8152B1 coupled with 5125B interface module. Direct observations of drill cutting edges were done by industrial camera type DMK41AF02 equipped with the telecentric lenses type TC2309. The experiment was further characterised with the following features: (1) Drill type and tool wear levels Drilling has been performed using Komet Medical 4.5 mm standard medical drills (type S2727.098), which were not subjected to sterilisation conditions. Three tool wear levels (TWL) were analysed – sharp drill (SD), medium worn drill (MWD) and worn drill (WD). Measurements were first taken while drilling
Fig. 1. Experimental setup.
with completely sharp drill. After completing the experiment for the first TWL, the drill was used to bore an additional 100 holes, at 900 rev/min and 0.3 mm/s feed rate, until it was worn out to the second TWL. The same procedure was repeated for the second and the third TWL. After approximately 560 holes altogether, drill has worn out to its final condition related to the WD level. Flank wear was observed as a dominant type of drill wear. This can be seen in Fig. 2, where flank wear zones on one cutting edge are presented at the beginning and at the end of experiment for every analysed TWL. Practically identical flank wear areas have been noticed on both cutting edges. (2) Cutting speeds and feed rates Twelve combinations of feed rates (0.01, 0.03, 0.05, 0.1 mm/rev) and cutting speeds (10, 30, 50 m/min) have been analysed. These cutting speeds correspond respectively to 707.4 rev/min, 2122.1 rev/min and 3536.8 rev/min, while feed rates fall within the interval (0.12 mm/s–5.9 mm/s) as presented in Table 1. The combinations of machining parameters have been chosen with regard to existing clinical practice based on hand-held drilling machines, as well as potential development of completely robotised high-speed drilling systems. Measurements for each combination of machining parameters and every tool wear level were randomly repeated 10 times.
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
3
Fig. 2. Cutting edge images at the beginning and at the end of the drilling experiment with the sharp drill (SD), medium worn drill (MWD) and worn drill (WD). Drill wear is observable as a dark area along the cutting edge on the drill flank. Fig. 4. Filtered force (FZ , FR ) and current (IZ , IMS ) signals. Table 1 Machining parameters.
2.2. Data acquisition and signal processing Cutting speed (vC )
Feed rate (f) (mm/rev)
(m/min)
(rev/min)
0.01
0.03 0.05 (mm/s)
0.1
10 30 50
707.4 2122.1 3536.8
0.12 0.35 0.59
0.35 1.06 1.77
1.18 3.54 5.90
0.59 1.77 2.95
Fig. 3. Position of temperature probe with thermocouple for bone temperature measurement.
(3) Bone specimens Bovine tibia samples were purchased fresh and then frozen to below –10 °C in order to preserve their thermo-mechanical characteristics [1,24]. Bones belonged to animals of unknown age and their mechanical properties were not measured. Specimens were formed from tibia diaphysis with average cortical thickness or drilling depth of 8.5 mm (from 4.7 to 18.4 mm). They were warmed up to 37 °C in the microwave oven before drilling. (4) Temperature measurements In order to analyse the influence of drill wear on drilling temperature alterations, thermocouple was installed 0.5 mm from the hole edge and 3 mm below the bone surface (Fig. 3). Also, several thermocouples were installed in the area around the cutting zone to measure surrounding air temperature maintained at approximately 37 °C.
Four types of servomotor drive referent current signals were analysed: horizontal (IX , IY ) and vertical (IZ ) feed drive currents, and main spindle current (IMS ). Since horizontal feed drive currents (IX , IY ) were low (usually below 0.1 A), drill wear effect was analysed using only IZ and IMS signals. On the other hand, the main intention of measuring forces was to compare performances of features extracted from force and related current signals in TWL classification. Potential replacement of force signals with motor drive currents would simplify the design of medical drilling system with TWL monitoring capabilities, and reduce its cost. Similarly to the feed drive currents, components FX and FY were significantly smaller (approx. 10 times) than the vertical component FZ , so IZ and IMS were compared with FZ and resultant cutting force FR determined on the basis of all three force components:
FR = FX2 + FY2 + FZ2 . The first analyses based on graphical comparisons of the shapes of filtered current and force signals have expectedly revealed high analogy between IZ and FZ , as well as IMS and FR signals (Fig. 4.). Both types of signals were continuously sampled at 1 ms rate and then analysed in the frequency domain using Fast Fourier Transform (FFT) algorithm. After this analysis, signals were filtered with Butterworth low-pass filter (cutoff frequency of 2 Hz), and two types of features from time domain were extracted from each type of signal (Fig. 5). The first type was “maximum” signal value calculated as an average of the 10% of the highest signal values (features F1 , F2 , I1 and I2 – see Table 2). Extreme signal values were averaged in order to neutralise the influence of higher local maximum. The second type was mean value of filtered current and force signals measured within the machining (cortical bone drilling) time (F3 , F4 , I3 and I4 ). Machining time has been estimated from the IMS signal. As for the features from the frequency domain, power in frequency components related to the rotational (RF) and cutting edges frequency (CEF) was determined using the FFT algorithm [25]. Since the drill has two cutting edges, CEF was twice as high as RF. Subsequent analyses of the FFT power spectrum of current signals have shown dominant spectral components on RF and CEF in mostly main spindle current signals. This was usually not the case with the IZ signals (Fig. 6). Therefore, dominant spectral components of the IMS signals have been chosen in drill wear classification (I5 , I6 ). They were compared with the related spectral components of the FR signals (F5 , F6 ). The power spectrum of the FR and FZ signals was very similar, although dominant components on RF and CEF were more distinguished in FR power
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE 4
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
Fig. 5. Extreme and mean values of filtered FR and IMS signals between times T1 (estimated start of cortical bone drilling) and T2 (estimated end of cortical bone drilling). In this example T1 = 3 s and T2 = 9.2 s.
spectrum (Fig. 6). In most samples, the power of the RF component was higher than that related to the CEF component. The third type of measured signal, i.e., AE signals, was sampled at the frequency of 2 MHz. One sample with duration of 0.1 s was taken per hole. Signals were filtered using the Butterworth band-pass filter with the frequency bandwidth from 40 to 500 kHz. This was in accordance with the specified measurement range of the utilised sensor. All features were extracted from the frequency domain. Measured frequency bandwidth was divided into seven equally wide ranges (50–100 kHz; 100–150 kHz; . . . ; 350–400 kHz), and the energy of every range was taken as a drill wear feature (AE1–AE7). Energy is calculated from the expression
ψ = 2
fU
fL
Sy df ,
(1)
where Sy is one-sided PSD function of the AE signal, while fL and fU are lower and upper frequency values chosen to reflect the energy in the range of interest [26].
Fig. 6. Power spectrum of FZ , FR , IZ and IMS signals. In this example vC = 30 m/min (n = 2122.1 rpm) → RF = 35.4 Hz; CEF = 2RF = 70.8 Hz.
tion. A type of RBF NN algorithm for dealing with classification types of problems was implemented. This type of neural network has been highly utilised in the past two decades, and the model applied in this research has already been successfully tested in industrial tool wear monitoring application [27]. The algorithm is based upon a feedforward three-layered RBF NN architecture, where the matrix/vector of synaptic weights c is calculated in the learning phase using the expression
c = H + y,
(2)
where y stands for the matrix/vector of desired output values, and H+ is Moore–Penrose pseudoinverse of the matrix of hidden layer neuron RBF outputs or activation function outputs (H). The pseudoinverse is defined as follows
H + = (H T H )−1 H T .
2.3. Tool wear features Altogether, 19 features have been extracted from all three types of signals (Table 2).
(3)
The matrix/vector of the desired output values y is obtained in the testing phase from the expression
y = Hc.
(4)
2.4. Neural network classification model
Elements of matrix H are determined according to the expression [13]
Extracted tool wear features were processed using the neural network model in order to determine their capacity for TWL classifica-
1 Hij = exp − rij 2 , 2
i = 1, . . . , N, j = 1, . . . , K,
(5)
Table 2 List of tool wear features. Type of signal
Feature
Description
Cutting forces and servomotor drive currents
F1 , I1 F2 , I2 F3 , I3 F4 , I4 F5 , I5 F6 , I6
Average of the group of the 10% highest FZ and IZ values Average of the group of the 10% highest FR and IMS values Mean value of the FZ and IZ signals Mean value of the FR and IMS signals Power of rotational frequency component of the FR and IMS signals Power of cutting edges frequency component of the FR and IMS signals
Acoustic emission
AE1 AE2 AE3 AE4 AE5 AE6 AE7
Energy of the AE signal in the frequency range 50–100 (kHz) Energy of the AE signal in the frequency range 100–150 (kHz) Energy of the AE signal in the frequency range 150–200 (kHz) Energy of the AE signal in the frequency range 200–250 (kHz) Energy of the AE signal in the frequency range 250–300 (kHz) Energy of the AE signal in the frequency range 300–350 (kHz) Energy of the AE signal in the frequency range 350–400 (kHz)
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
5
where rij is the Mahalanobis distance between vectors composed from the ith element of all input vectors (tool wear features) and the jth hidden layer neuron. Squared Mahalanobis distance is calculated using the expression
T −1 xi − t j , rij 2 = xi − t j j
(6)
where j is the covariance matrix belonging to the group of learning samples that are connected to the jth hidden layer neuron, xi is the L-dimensional vector composed of the ith element of all L input vectors and tj is the L-dimensional vector of the jth hidden layer neuron centre. Covariance matrix is quadratic matrix with non-zero elements (squared σ vector components) on the main diagonal and zeros elsewhere,
⎡
σ1 2
j = ⎢ ⎣ 0 0
0 .. . 0
0
⎤
⎥ 0 ⎦.
(7)
σL 2
Vector σ is composed of the maximal Euclidian distances between learning samples belonging to the analysed group and the centre of that group, regarding all (L) dimensions separately,
σg |j = max {zpg − tg , p = 1, . . . , LKG } |j , g = 1, . . . , L,
(8)
where zpg is the gth component of the pth sample of the jth group which is defined with the LKG numbers of samples, and tg is the gth component of the jth group centre vector (jth hidden layer neuron centre vector). Hidden layer neuron centres are defined using a configuration method which helps a teacher to determine quickly network structure regarding the nature of the learning problem and desirable generalisation characteristics. Grouping of tool wear features (network input elements) and centres calculations are based on the parameter β C . Higher β C reduces the number of hidden layer neurons and vice versa (for β C = 0 every input vector forms one hidden layer neuron or its centre). The method is explained in [27] in detail.
Fig. 7. Maximum drilling temperatures for all tool wear levels and machining parameter combinations. Temperatures were averaged over 10 randomly measured samples belonging to a certain combination of machining parameters.
3. Results and discussion With 12 combinations of machining parameters, three drill wear levels, and 10 randomly performed measurements for each combination of parameters and drill wear level, 360 sets of tool wear features have been extracted from all types of signals. Half of them, or five random measurements belonging to each combination, were used in the learning phase of the RBF NN classifier, while the remaining data participated in the formation of test sets (five tests for each combination). At first, all measured data were used to analyse the influence of drill wear on drilling temperature and resultant forces. Strong and proportional relationship between wear intensity and temperature rise during bone drilling have been confirmed, as expected (Fig. 7). The same relationship was also expected in the case of cutting forces. However, as can be seen in Fig. 8 and Fig. 4, this was not the case in the present study. Unexpected force reduction when drilling with the worn drill (WD), compared with drilling using the medium worn drill (MWD) and in some cases even the sharp drill (SD), was presumably caused by different mechanical properties (lower hardness) of the set of bone specimens used in the experiment with the worn drill. This hypothesis is further supported by the fact that higher cutting forces have been observed in all measurements when drilling with the medium worn drill compared with those measured using the sharp drill. Considering that this situation is expected in medical practice, and that similar process signals measured at two different tool wear levels can reduce classification precision, it is interesting to establish tool wear classification capacity of chosen features in this experiment. The results presented in Figs. 7 and 8 have shown that changes in mechanical properties of the bone specimens did not reduce drilling
Fig. 8. Maximum resultant forces for all tool wear levels and machining parameter combinations. Forces were averaged over 10 randomly measured samples belonging to a certain combination of machining parameters.
temperature, thus indicating its dependency on the feed rate, cortical thickness and drill wear level. It should also be noted that the lowest drilling temperatures were achieved during high speed drilling (f = 0.1 mm/rev; vC = 50 m/min), while, at the same time, this machining regime resulted in the appearance of the highest cutting forces (above 400 N). Potential implementation of high speed drilling in medical purposes would thus necessarily imply utilisation of fully robotised drilling systems. After these first analyses, drill wear classification performance of the chosen tool wear features and their combinations were thoroughly tested using RBF NN. In order to find a model which provides the best classification performance, learning/testing procedure was divided into several steps. In the first step, every feature was analysed separately, and the results are presented in Table 3 in the form of the percentage of accurately classified samples (tests T1–T5 + average classification success rates).
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE 6
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7 Table 3 Results of individually tested features with β C = 0 (accurately classified samples in %). Feature
Table 5 Average results of selected feature combinations for each tool wear level (accurately classified samples in %).
Test T1
T2
T3
T4
T5
Average
F1 F2 F3 F4 F5 F6
69.4 66.7 63.9 66.7 30.6 36.1
75.0 72.2 69.4 69.4 44.4 38.9
66.7 66.7 66.7 63.9 36.1 25.0
58.3 61.1 63.9 55.6 38.9 19.4
58.3 61.1 52.8 50.0 38.9 50.0
65.4 65.6 63.3 61.1 37.8 33.9
I1 I2 I3 I4 I5 I6
61.1 55.6 72.2 33.3 41.7 30.6
61.1 69.4 63.9 58.3 41.7 47.2
61.1 52.8 66.7 44.4 44.4 33.3
66.7 50.0 52.8 44.4 36.1 33.3
58.3 52.8 41.7 55.6 44.4 38.9
61.7 56.1 59.4 47.2 41.7 36.7
AE1 AE2 AE3 AE4 AE5 AE6 AE7
69.4 47.2 44.4 66.7 50.0 61.1 30.6
58.3 55.6 63.9 83.3 58.3 61.1 44.4
58.3 69.4 55.6 61.1 58.3 58.3 63.9
58.3 58.3 58.3 66.7 44.4 55.6 47.2
77.8 61.1 50.0 75.0 47.2 66.7 69.4
64.4 58.3 54.4 70.6 51.7 60.6 51.1
Features
F1 –F6 I1 –I6 AE1–AE7
These results were obtained using full RBF NN hidden layer structure of 180 neurons, i.e., with β C = 0 (the number of hidden layer neurons were equal to the number of learning samples). It should also be mentioned that all results were achieved using cutting speed and feed rate as two additional RBF NN inputs, and that all data were normalised between 0 and 1. Average indidvidual performances of features extracted from force and current signals are very comparable, although force features from the time domain achieved somewhat higher classification accuracy than similar current features. Generally, force and current features from the time domain (F1 , . . . , F4 , I1 , . . . , I4 ) have individually achieved higher classification precision than those from the frequency domain (F5 , F6 , I5 , I6 ). The difference is particularly noticeable in the case of force features. On the other hand, all AE features have individually correctly classified tool wear level in more than 50% of cases (average result), and their result can be compared with the force and current time domain features. The energy of the AE signal in the frequency range from 200 to 250 Hz (AE4) achieved the best individual result among all features (70.6%). Based on these first results, further analyses of different feature combinations have been performed, again using full hidden
Average SD
MWD
WD
68.3 78.3 91.7
81.6 80 86.7
43.3 61.7 91.7
layer structure (β C = 0, 180 hidden layer neurons). Feature combinations have generally and expectedly achieved higher classification accuracy (Table 4) compared with their individual performance. At first, combinations of features belonging to every type of signal were tested. In this case, combination of time domain current features (I1 –I4 ) and all current features (I1 –I6 ) achieved better results than the comparable combinations of force features (F1 –F4 , F1 – F6 ), thus indicating the implementation potential of current signals in the medical drill wear monitoring process. The highest average classification precision of 90% was accomplished by the combination of all AE features, which was expected based on their individual performances. At the end, selected combinations of features from different types of signals were tested (F + AE, I + AE). These tests confirmed the high average classification accuracy of all combinations. The best overall performance was observed in the case of current and AE features combination: I1 –I4 + AE1–AE7, which accomplished slightly better average result (92.8%) than the combination of AE features alone (AE1–AE7). For the best combination of features, additional tests were performed in order to analyse classification (generalisation) characteristics of different RBF NN structures. These structures were characterised by several hidden layer configurations, where the number of hidden layer neurons was smaller than the number of learning samples (β C > 0). The chosen neural network algorithm has shown satisfactory drill wear classification capacity and application potential by achieving average classification accuracy of up to 85% (Table 4) with almost four times less hidden layer neurons (43) compared with the full hidden layer structure (180). In order to establish the influence of bone mechanical properties on the performance of tool wear features of all three types of signals, average classification success rates achieved using feature combinations F1 –F6 (64.4%), I1 –I6 (73.3%) and AE1–AE7 (90%) have been presented separately for each tool wear level in Table 5.
Table 4 Results of selected feature combinations (accurately classified samples in %). Features
βC
Test
RBF NN structure
T1
T2
T3
T4
T5
Average
F1 –F4 F5 , F6 F1 –F6
0
63.9 41.7 66.7
61.1 44.4 61.1
66.7 44.4 69.4
58.3 52.8 66.7
63.9 38.9 58.3
62.8 44.4 64.4
6-180-3 4-180-3 8-180-3
I1 –I4 I5 , I6 I1 –I6
0
61.1 41.7 66.7
77.8 41.7 88.9
77.8 44.4 72.2
66.7 33.3 66.7
80.6 33.3 72.2
72.8 38.9 73.3
6-180-3 4-180-3 8-180-3
AE1–AE7 F1 –F4 + AE1–AE7 F1 –F6 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I6 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7
0 0
86.1 91.7 88.9 97.2 94.4 91.7 86.1 83.3 88.9 80.6
91.7 97.2 97.2 97.2 94.4 94.4 86.1 86.1 86.1 80.6
94.4 94.4 94.4 91.7 88.9 91.7 83.3 83.3 83.3 83.3
86.1 86.1 86.1 83.3 80.6 80.6 80.6 83.3 80.6 72.2
91.7 88.9 88.9 94.4 94.4 97.2 91.7 83.3 83.3 80.6
90.0 91.7 91.1 92.8 90.6 91.1 85.6 83.9 84.4 79.4
9-180-3 13-180-3 15-180-3 13-180-3 15-180-3 13-161-3 13-120-3 13-75-3 13-43-3 13-33-3
>0
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
JID: JJBE
ARTICLE IN PRESS
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
When comparing these results, it can be noted that tool wear features from cutting force signals have the highest, and those extracted from AE signals the lowest, sensitivity to changes in bone hardness. Only 5 of 60 samples (8.3%) belonging to the WD classification group were misclassified in the case of AE features, and all of them were classified in the first nearest tool wear level (MWD). 4. Conclusion After analysing all chosen drill wear features individually and in combinations, it can be concluded that the combination of features extracted from the AE signal and two servomotor drive signals (IZ and IMS ) have the highest potential for medical drill wear classification. The application of multi-sensor data in medical drill wear monitoring is in line with the numerous research studies in the field of industrial tool wear monitoring, which have confirmed that this approach is necessary to build robust and applicable tool wear monitoring systems. Application of AE sensors seems particularly important, since features extracted from AE signals have shown the highest individual quality, as well as precision when used in combinations, and are practically insensitive to changes in bone mechanical properties. Future research activities will include analyses of several new variables on drill wear dynamic, such as different drill geometries, influence of the sterilisation process parameters and application of ultrasonically-assisted drilling technique, which has high benefits in reducing forces (torques) and surface roughness. Ethical approval In the experimental part of this research bovine tibia bone specimens were used. The bones belonged to animals slaughtered solely for the purpose of further processing in the food industry. Therefore, no ethical approval is required. Conflicts of interest None. Acknowledgement This work was financially supported by the Ministry of Science, Education and Sport of the Republic of Croatia Grant No. 120-12019481938 through the fund for national scientific projects. References [1] Alam K, Mitrofanov AV, Silberschmidt VV. Experimental investigations of forces and torque in conventional and ultrasonically-assisted drilling of cortical bone. Med Eng Phys 2011;33:234–9. [2] Alam K, Mitrofanov AV, Bäker M, Silberschmidt VV. Stresses in ultrasonicallyassisted bone cutting. J Phys Conf Ser 2009;181:012014. [3] Alam K, Mitrofanov AV, Silberschmidt VV. Measurements of surface roughness in conventional and ultrasonically-assisted bone drilling. Am J Biomed Sci 2009;1:312–20.
7
[4] Augustin G, Zigman T, Davila S, Udiljak T, Staroveski T, Brezak D, et al. Cortical bone drilling and thermal osteonecrosis. Clin Biomech 2012;27:313–25. [5] Karmani S. The thermal properties of bone and the effects of surgical intervention. Curr Orthop 2006;20:52–8. [6] Cordioli G, Majzoub Z. Heat generation during implant site preparation: an in vitro study. Int J Oral Maxillofac Impl 1997;12:186–93. [7] Eriksson RA, Albrektsson T. Temperature threshold levels for heat induced bone tissue injury: a vital microscopic study in rabbit. J Prosthet Dent 1983;50:101–7. [8] Natali C, Ingle P, Dowell J. Orthopaedic bone drills—can they be improved? J Bone Joint Surg 1996;78-B:357–62. [9] Augustin G, Davila S, Mihoci K, Udiljak T, Vedrina DS, Antabak A. Thermal osteonecrosis and bone drilling parameters revisited. Arch Orthop Trauma Surg 2007;128:71–7. [10] Paszenda Z, Basiaga M. FEM analysis of drills used in bone surgery. Arch Mater Sci Eng 2009;36:103–9. [11] Kalidindi V. Optimization of drill design and coolant systems during dental implant surgery (Ph.D. thesis, 2004). University of Kentucky, USA. URL: http://uknowledge.uky.edu/cgi/viewcontent.cgi?article=1317&context= gradschool_theses (last accessed 10.14). [12] Davidson SRH, James DF. Measurement of thermal conductivity of bovine cortical bone. Med Eng Phys 2000;22:741–7. [13] Reingewirtz Y, Szmukler-Moncler S, Senger B. Influence of different parameters on bone heating and drilling time in implantology. Clin Oral Implants Res 1997;8:189–97. [14] Bachus KN, Rondina MT, Hutchinson DT. The effects of drilling force on cortical temperatures and their duration: an in vitro study. Med Eng Phys 2000;22:685– 91. [15] Nam O, Yu W, Choi MY, Kyung HM. Monitoring of bone temperature during osseous preparation for orthodontic micro-screw implants: effect of motor speed and pressure. Key Eng Mater 2006;321–323:1044–7. [16] Allan W, Williams ED, Kerawala CJ. Effects of repeated drill use on temperature of bone during preparation for osteosynthesis self-tapping screws. Br J Oral Maxillofac Surg 2005;43:314–19. [17] Chacon GE, Bower DL, Larsen PE, McGlumphy EA, Beck FM. Heat production by 3 Implant Drill Systems after repeated drilling and sterilization. J Oral Maxillofac Surg 2006;64:265–9. [18] Queiroz TP, Souza FÁ, Okamoto R, Margonar R, Pereira-Filho VA, Garcia IR, et al. Evaluation of immediate bone-cell viability and of drill wear after implant osteotomies: immunohistochemistry and scanning electron microscopy analysis. J Oral Maxillofac Surg 2008;66:1233–40. [19] Jochum RM, Reichart PA. Influence of multiple use of Timedur – titanium cannon drills: thermal response and scanning electron microscopic findings. Clin Oral Implants Res 2000;11:139–43. [20] Singh J, Davenport JH, Pegg DJ. A national survey of instrument sharpening guidelines. Surg 2010;8:136–9. [21] Jantunen E. A summary of methods applied to tool condition monitoring in drilling. Int J Mach Tools Manuf 2002;42:997–1010. [22] Davidson S.R.H. Heat transfer in bone during drilling (M.S. Thesis, 1999). Institute of Biomaterials and Biomedical Engineering, University of Toronto, Canada. URL: https://tspace.library.utoronto.ca/bitstream/1807/14728/1/MQ46070.pdf (last accessed 10.14). [23] Lee J, Rabin Y, Ozdoganlar OB. A new thermal model for bone drilling with applications to orthopaedic surgery. Med Eng Phys 2011;33:1234–44. [24] Laurito D, Lamazza L, Garreffa G, De Biase A. An alternative method to record rising temperatures during dental implant site preparation: a preliminary study using bovine bone. Ann dell’Istituto Super Sanita 2010;46:405– 10. [25] Prickett PW, Johns C. An overview of approaches to end milling tool monitoring. Int J Mach Tools Manuf 1999;39:105–22. [26] Scheffer C, Heyns PS, Klocke F. Development of a tool wear-monitoring system for hard turning. Int J Mach Tools Manuf 2003;43:973–85. [27] Brezak D, Udiljak T, Mihoci K, Majetic D, Novakovic B, Kasac J. Tool wear monitoring using radial basis function neural network. In: International joint conference on neural networks – IJCNN, Budapest, Hungary, vol. 3; 2004. p. 1859–63. R
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
JID: JJBE
[m5G;April 16, 2015;19:21]
Medical Engineering and Physics 000 (2015) 1–7
Contents lists available at ScienceDirect
Medical Engineering and Physics journal homepage: www.elsevier.com/locate/medengphy
Drill wear monitoring in cortical bone drilling Tomislav Staroveski a,∗, Danko Brezak b, Toma Udiljak a a
Department of Technology, Faculty of Mechanical Engineering and Naval Architecture, University of Zagreb, Ivana Lucica 5, Zagreb, Croatia Department of Robotics and Production System Automation, Faculty of Mechanical Engineering and Naval Architecture, University of Zagreb, Ivana Lucica 5, Zagreb, Croatia
b
a r t i c l e
i n f o
Article history: Received 27 October 2014 Revised 9 February 2015 Accepted 27 March 2015 Available online xxx Keywords: Medical drill wear Thermal osteonecrosis Neural networks Computational modelling Medical devices
a b s t r a c t Medical drills are subject to intensive wear due to mechanical factors which occur during the bone drilling process, and potential thermal and chemical factors related to the sterilisation process. Intensive wear increases friction between the drill and the surrounding bone tissue, resulting in higher drilling temperatures and cutting forces. Therefore, the goal of this experimental research was to develop a drill wear classification model based on multi-sensor approach and artificial neural network algorithm. A required set of tool wear features were extracted from the following three types of signals: cutting forces, servomotor drive currents and acoustic emission. Their capacity to classify precisely one of three predefined drill wear levels has been established using a pattern recognition type of the Radial Basis Function Neural Network algorithm. Experiments were performed on a custom-made test bed system using fresh bovine bones and standard medical drills. Results have shown high classification success rate, together with the model robustness and insensitivity to variations of bone mechanical properties. Features extracted from acoustic emission and servomotor drive signals achieved the highest precision in drill wear level classification (92.8%), thus indicating their potential in the design of a new type of medical drilling machine with process monitoring capabilities. © 2015 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction Along with the development of new surgical approaches and techniques, bone drilling has become mostly routine and continuously increasing medical intervention. Bone tissue is formed by organic and mineral phases whose interactions result in complex mechanical and thermal properties. All bone drilling interventions include heat generation due to the friction between tool, bone and chips, which can significantly influence the post-operative recovery. Hence, novel drilling techniques have been used to study the interaction between the drill and bone in order to reduce drilling forces and improve chip removal from the drilling site [1–3]. Although mechanisms of thermal bone damages are still not thoroughly explained, it is a known fact that temperature rise in the drilling zone can cause thermal osteonecrosis, which prevents quality tissue healing and bone regeneration [4,5]. There are several factors influencing drilling temperature variations and potential occurrence of thermal osteonecrosis: mechanical characteristics of the bone and its cortical thickness [6,7], drill design and geometry [8–11], inefficient external cooling effect due to low bone thermal conductivity [12], inadequate cutting speeds and feed rates [13–15], improper tool position/trajectory which increases fric-
∗
Corresponding author. Tel.: +385 1 6168341; fax: +385 1 6156940. E-mail address: [email protected], [email protected] (T. Staroveski).
tion between the drill body and the hole surface, and drill wear. Drill wear is among those factors with the highest impact on heat generation during bone drilling. It is an unavoidable and irreversible process which increases friction in the cutting zone. Besides negative thermal impact, it causes higher cutting forces and tool vibrations, which can result in the cutting edge breakage, or complete drill breakage in the flutes or shank zone. These situations cause mechanical bone damages, which also have negative impact on the post-operative therapy progress. Several studies of drill wear effect on temperature variations conducted in the field of oral and maxillofacial surgery point to the strong and proportional relationship between tool wear dynamic and bone drilling temperature rise [16–19]. At the same time, some researchers emphasised the frequent usage of worn drills in medical interventions, as well as absence of hospital standards and activities focused on drill wear identification and tool labelling [20]. It is therefore logical to presume that precise and reliable drill wear models could become important elements in the reduction of the occurrence of bone tissue damages. Their role would be particularly important in the development of semi- or completely automated advanced medical drilling systems. In the past 25 years, numerous studies on drill wear monitoring have been published, mainly focused on industrial applications [21]. Drilling process dynamic in industrial applications usually differs from that related to the bone drilling in terms of different types
http://dx.doi.org/10.1016/j.medengphy.2015.03.014 1350-4533/© 2015 IPEM. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
JID: JJBE 2
ARTICLE IN PRESS
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
of drills, bone vs. industrial workpiece material characteristics, machining parameters, cooling methods, etc. Nevertheless, methodology used in designing industrial tool wear monitoring systems based on multi-sensor approach and nonlinear modelling could also be potentially applicable in medical drill wear identification. However, to our knowledge, there is still a lack of tool wear monitoring solutions specifically designed for medical applications, or experimental studies involving the implementation of some already designed and tested industrial monitoring systems in bone drilling applications. Although several models for temperature estimation in bone drilling have been proposed [22,23], none of them includes drill wear identification. There are two major obstacles which prevent precise tool wear level quantification. Direct measurement of tool wear during the cutting process is not possible due to constant contact between bone and drill cutting edges. It can only be estimated using tool wear features extracted from different types of process signals and other known machining parameters (cutting speed, feed rate, drill characteristics). Additionally, industrial applications have shown that tool wear is usually highly nonlinear and sometimes even a partially stochastic process. Similar characteristics can be expected in bone drilling due to the complex interrelation of all aforementioned process parameters. The complexity of wear process in industrial applications motivated many researchers to experiment with different types of computational intelligence algorithms, primarily artificial neural networks, to build reliable and accurate wear models. Artificial neural networks gained such popularity because of their nonlinear modelling capabilities based on parallel processing and integration of system/process data (in this case, drill wear features and machining parameters). Taking all this into consideration, the aim of this experimental study was to analyse the performances of the multi-sensor based medical drill wear classification model, and its potential for application in the next-generation medical drilling machines. Drill wear features were extracted from force, current and acoustic emission signals, and then processed using Radial Basis Function Neural Network (RBF NN), which is known for short learning procedure, as well as simple and quick hidden layer structure adaptation. 2. Methods 2.1. Experimental setup and parameters Experimental work has been performed using the custom-made 3-axis bench-top mini milling machine adjusted for the purpose of the bone drilling research (Fig. 1). The machine has been retrofitted with the 0.4 kW (1.27 N m) permanent magnet synchronous motors with integrated incremental encoders (type Mecapion SB04A), corresponding servomotor drives (DPCANIE-030A400 and DPCANIE-060A400), ball screw assemblies and LinuxCNC open architecture control (OAC) system. Apart from the control loop referent currents taken from the main spindle and feed drives, cutting forces have been measured using triaxial Kistler piezoelectric dynamometer 9257B coupled with 5017B charge amplifier, and acoustic emission (AE) signals using Kistler industrial sensor type 8152B1 coupled with 5125B interface module. Direct observations of drill cutting edges were done by industrial camera type DMK41AF02 equipped with the telecentric lenses type TC2309. The experiment was further characterised with the following features: (1) Drill type and tool wear levels Drilling has been performed using Komet Medical 4.5 mm standard medical drills (type S2727.098), which were not subjected to sterilisation conditions. Three tool wear levels (TWL) were analysed – sharp drill (SD), medium worn drill (MWD) and worn drill (WD). Measurements were first taken while drilling
Fig. 1. Experimental setup.
with completely sharp drill. After completing the experiment for the first TWL, the drill was used to bore an additional 100 holes, at 900 rev/min and 0.3 mm/s feed rate, until it was worn out to the second TWL. The same procedure was repeated for the second and the third TWL. After approximately 560 holes altogether, drill has worn out to its final condition related to the WD level. Flank wear was observed as a dominant type of drill wear. This can be seen in Fig. 2, where flank wear zones on one cutting edge are presented at the beginning and at the end of experiment for every analysed TWL. Practically identical flank wear areas have been noticed on both cutting edges. (2) Cutting speeds and feed rates Twelve combinations of feed rates (0.01, 0.03, 0.05, 0.1 mm/rev) and cutting speeds (10, 30, 50 m/min) have been analysed. These cutting speeds correspond respectively to 707.4 rev/min, 2122.1 rev/min and 3536.8 rev/min, while feed rates fall within the interval (0.12 mm/s–5.9 mm/s) as presented in Table 1. The combinations of machining parameters have been chosen with regard to existing clinical practice based on hand-held drilling machines, as well as potential development of completely robotised high-speed drilling systems. Measurements for each combination of machining parameters and every tool wear level were randomly repeated 10 times.
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
3
Fig. 2. Cutting edge images at the beginning and at the end of the drilling experiment with the sharp drill (SD), medium worn drill (MWD) and worn drill (WD). Drill wear is observable as a dark area along the cutting edge on the drill flank. Fig. 4. Filtered force (FZ , FR ) and current (IZ , IMS ) signals. Table 1 Machining parameters.
2.2. Data acquisition and signal processing Cutting speed (vC )
Feed rate (f) (mm/rev)
(m/min)
(rev/min)
0.01
0.03 0.05 (mm/s)
0.1
10 30 50
707.4 2122.1 3536.8
0.12 0.35 0.59
0.35 1.06 1.77
1.18 3.54 5.90
0.59 1.77 2.95
Fig. 3. Position of temperature probe with thermocouple for bone temperature measurement.
(3) Bone specimens Bovine tibia samples were purchased fresh and then frozen to below –10 °C in order to preserve their thermo-mechanical characteristics [1,24]. Bones belonged to animals of unknown age and their mechanical properties were not measured. Specimens were formed from tibia diaphysis with average cortical thickness or drilling depth of 8.5 mm (from 4.7 to 18.4 mm). They were warmed up to 37 °C in the microwave oven before drilling. (4) Temperature measurements In order to analyse the influence of drill wear on drilling temperature alterations, thermocouple was installed 0.5 mm from the hole edge and 3 mm below the bone surface (Fig. 3). Also, several thermocouples were installed in the area around the cutting zone to measure surrounding air temperature maintained at approximately 37 °C.
Four types of servomotor drive referent current signals were analysed: horizontal (IX , IY ) and vertical (IZ ) feed drive currents, and main spindle current (IMS ). Since horizontal feed drive currents (IX , IY ) were low (usually below 0.1 A), drill wear effect was analysed using only IZ and IMS signals. On the other hand, the main intention of measuring forces was to compare performances of features extracted from force and related current signals in TWL classification. Potential replacement of force signals with motor drive currents would simplify the design of medical drilling system with TWL monitoring capabilities, and reduce its cost. Similarly to the feed drive currents, components FX and FY were significantly smaller (approx. 10 times) than the vertical component FZ , so IZ and IMS were compared with FZ and resultant cutting force FR determined on the basis of all three force components:
FR = FX2 + FY2 + FZ2 . The first analyses based on graphical comparisons of the shapes of filtered current and force signals have expectedly revealed high analogy between IZ and FZ , as well as IMS and FR signals (Fig. 4.). Both types of signals were continuously sampled at 1 ms rate and then analysed in the frequency domain using Fast Fourier Transform (FFT) algorithm. After this analysis, signals were filtered with Butterworth low-pass filter (cutoff frequency of 2 Hz), and two types of features from time domain were extracted from each type of signal (Fig. 5). The first type was “maximum” signal value calculated as an average of the 10% of the highest signal values (features F1 , F2 , I1 and I2 – see Table 2). Extreme signal values were averaged in order to neutralise the influence of higher local maximum. The second type was mean value of filtered current and force signals measured within the machining (cortical bone drilling) time (F3 , F4 , I3 and I4 ). Machining time has been estimated from the IMS signal. As for the features from the frequency domain, power in frequency components related to the rotational (RF) and cutting edges frequency (CEF) was determined using the FFT algorithm [25]. Since the drill has two cutting edges, CEF was twice as high as RF. Subsequent analyses of the FFT power spectrum of current signals have shown dominant spectral components on RF and CEF in mostly main spindle current signals. This was usually not the case with the IZ signals (Fig. 6). Therefore, dominant spectral components of the IMS signals have been chosen in drill wear classification (I5 , I6 ). They were compared with the related spectral components of the FR signals (F5 , F6 ). The power spectrum of the FR and FZ signals was very similar, although dominant components on RF and CEF were more distinguished in FR power
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE 4
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
Fig. 5. Extreme and mean values of filtered FR and IMS signals between times T1 (estimated start of cortical bone drilling) and T2 (estimated end of cortical bone drilling). In this example T1 = 3 s and T2 = 9.2 s.
spectrum (Fig. 6). In most samples, the power of the RF component was higher than that related to the CEF component. The third type of measured signal, i.e., AE signals, was sampled at the frequency of 2 MHz. One sample with duration of 0.1 s was taken per hole. Signals were filtered using the Butterworth band-pass filter with the frequency bandwidth from 40 to 500 kHz. This was in accordance with the specified measurement range of the utilised sensor. All features were extracted from the frequency domain. Measured frequency bandwidth was divided into seven equally wide ranges (50–100 kHz; 100–150 kHz; . . . ; 350–400 kHz), and the energy of every range was taken as a drill wear feature (AE1–AE7). Energy is calculated from the expression
ψ = 2
fU
fL
Sy df ,
(1)
where Sy is one-sided PSD function of the AE signal, while fL and fU are lower and upper frequency values chosen to reflect the energy in the range of interest [26].
Fig. 6. Power spectrum of FZ , FR , IZ and IMS signals. In this example vC = 30 m/min (n = 2122.1 rpm) → RF = 35.4 Hz; CEF = 2RF = 70.8 Hz.
tion. A type of RBF NN algorithm for dealing with classification types of problems was implemented. This type of neural network has been highly utilised in the past two decades, and the model applied in this research has already been successfully tested in industrial tool wear monitoring application [27]. The algorithm is based upon a feedforward three-layered RBF NN architecture, where the matrix/vector of synaptic weights c is calculated in the learning phase using the expression
c = H + y,
(2)
where y stands for the matrix/vector of desired output values, and H+ is Moore–Penrose pseudoinverse of the matrix of hidden layer neuron RBF outputs or activation function outputs (H). The pseudoinverse is defined as follows
H + = (H T H )−1 H T .
2.3. Tool wear features Altogether, 19 features have been extracted from all three types of signals (Table 2).
(3)
The matrix/vector of the desired output values y is obtained in the testing phase from the expression
y = Hc.
(4)
2.4. Neural network classification model
Elements of matrix H are determined according to the expression [13]
Extracted tool wear features were processed using the neural network model in order to determine their capacity for TWL classifica-
1 Hij = exp − rij 2 , 2
i = 1, . . . , N, j = 1, . . . , K,
(5)
Table 2 List of tool wear features. Type of signal
Feature
Description
Cutting forces and servomotor drive currents
F1 , I1 F2 , I2 F3 , I3 F4 , I4 F5 , I5 F6 , I6
Average of the group of the 10% highest FZ and IZ values Average of the group of the 10% highest FR and IMS values Mean value of the FZ and IZ signals Mean value of the FR and IMS signals Power of rotational frequency component of the FR and IMS signals Power of cutting edges frequency component of the FR and IMS signals
Acoustic emission
AE1 AE2 AE3 AE4 AE5 AE6 AE7
Energy of the AE signal in the frequency range 50–100 (kHz) Energy of the AE signal in the frequency range 100–150 (kHz) Energy of the AE signal in the frequency range 150–200 (kHz) Energy of the AE signal in the frequency range 200–250 (kHz) Energy of the AE signal in the frequency range 250–300 (kHz) Energy of the AE signal in the frequency range 300–350 (kHz) Energy of the AE signal in the frequency range 350–400 (kHz)
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
5
where rij is the Mahalanobis distance between vectors composed from the ith element of all input vectors (tool wear features) and the jth hidden layer neuron. Squared Mahalanobis distance is calculated using the expression
T −1 xi − t j , rij 2 = xi − t j j
(6)
where j is the covariance matrix belonging to the group of learning samples that are connected to the jth hidden layer neuron, xi is the L-dimensional vector composed of the ith element of all L input vectors and tj is the L-dimensional vector of the jth hidden layer neuron centre. Covariance matrix is quadratic matrix with non-zero elements (squared σ vector components) on the main diagonal and zeros elsewhere,
⎡
σ1 2
j = ⎢ ⎣ 0 0
0 .. . 0
0
⎤
⎥ 0 ⎦.
(7)
σL 2
Vector σ is composed of the maximal Euclidian distances between learning samples belonging to the analysed group and the centre of that group, regarding all (L) dimensions separately,
σg |j = max {zpg − tg , p = 1, . . . , LKG } |j , g = 1, . . . , L,
(8)
where zpg is the gth component of the pth sample of the jth group which is defined with the LKG numbers of samples, and tg is the gth component of the jth group centre vector (jth hidden layer neuron centre vector). Hidden layer neuron centres are defined using a configuration method which helps a teacher to determine quickly network structure regarding the nature of the learning problem and desirable generalisation characteristics. Grouping of tool wear features (network input elements) and centres calculations are based on the parameter β C . Higher β C reduces the number of hidden layer neurons and vice versa (for β C = 0 every input vector forms one hidden layer neuron or its centre). The method is explained in [27] in detail.
Fig. 7. Maximum drilling temperatures for all tool wear levels and machining parameter combinations. Temperatures were averaged over 10 randomly measured samples belonging to a certain combination of machining parameters.
3. Results and discussion With 12 combinations of machining parameters, three drill wear levels, and 10 randomly performed measurements for each combination of parameters and drill wear level, 360 sets of tool wear features have been extracted from all types of signals. Half of them, or five random measurements belonging to each combination, were used in the learning phase of the RBF NN classifier, while the remaining data participated in the formation of test sets (five tests for each combination). At first, all measured data were used to analyse the influence of drill wear on drilling temperature and resultant forces. Strong and proportional relationship between wear intensity and temperature rise during bone drilling have been confirmed, as expected (Fig. 7). The same relationship was also expected in the case of cutting forces. However, as can be seen in Fig. 8 and Fig. 4, this was not the case in the present study. Unexpected force reduction when drilling with the worn drill (WD), compared with drilling using the medium worn drill (MWD) and in some cases even the sharp drill (SD), was presumably caused by different mechanical properties (lower hardness) of the set of bone specimens used in the experiment with the worn drill. This hypothesis is further supported by the fact that higher cutting forces have been observed in all measurements when drilling with the medium worn drill compared with those measured using the sharp drill. Considering that this situation is expected in medical practice, and that similar process signals measured at two different tool wear levels can reduce classification precision, it is interesting to establish tool wear classification capacity of chosen features in this experiment. The results presented in Figs. 7 and 8 have shown that changes in mechanical properties of the bone specimens did not reduce drilling
Fig. 8. Maximum resultant forces for all tool wear levels and machining parameter combinations. Forces were averaged over 10 randomly measured samples belonging to a certain combination of machining parameters.
temperature, thus indicating its dependency on the feed rate, cortical thickness and drill wear level. It should also be noted that the lowest drilling temperatures were achieved during high speed drilling (f = 0.1 mm/rev; vC = 50 m/min), while, at the same time, this machining regime resulted in the appearance of the highest cutting forces (above 400 N). Potential implementation of high speed drilling in medical purposes would thus necessarily imply utilisation of fully robotised drilling systems. After these first analyses, drill wear classification performance of the chosen tool wear features and their combinations were thoroughly tested using RBF NN. In order to find a model which provides the best classification performance, learning/testing procedure was divided into several steps. In the first step, every feature was analysed separately, and the results are presented in Table 3 in the form of the percentage of accurately classified samples (tests T1–T5 + average classification success rates).
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
ARTICLE IN PRESS
JID: JJBE 6
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7 Table 3 Results of individually tested features with β C = 0 (accurately classified samples in %). Feature
Table 5 Average results of selected feature combinations for each tool wear level (accurately classified samples in %).
Test T1
T2
T3
T4
T5
Average
F1 F2 F3 F4 F5 F6
69.4 66.7 63.9 66.7 30.6 36.1
75.0 72.2 69.4 69.4 44.4 38.9
66.7 66.7 66.7 63.9 36.1 25.0
58.3 61.1 63.9 55.6 38.9 19.4
58.3 61.1 52.8 50.0 38.9 50.0
65.4 65.6 63.3 61.1 37.8 33.9
I1 I2 I3 I4 I5 I6
61.1 55.6 72.2 33.3 41.7 30.6
61.1 69.4 63.9 58.3 41.7 47.2
61.1 52.8 66.7 44.4 44.4 33.3
66.7 50.0 52.8 44.4 36.1 33.3
58.3 52.8 41.7 55.6 44.4 38.9
61.7 56.1 59.4 47.2 41.7 36.7
AE1 AE2 AE3 AE4 AE5 AE6 AE7
69.4 47.2 44.4 66.7 50.0 61.1 30.6
58.3 55.6 63.9 83.3 58.3 61.1 44.4
58.3 69.4 55.6 61.1 58.3 58.3 63.9
58.3 58.3 58.3 66.7 44.4 55.6 47.2
77.8 61.1 50.0 75.0 47.2 66.7 69.4
64.4 58.3 54.4 70.6 51.7 60.6 51.1
Features
F1 –F6 I1 –I6 AE1–AE7
These results were obtained using full RBF NN hidden layer structure of 180 neurons, i.e., with β C = 0 (the number of hidden layer neurons were equal to the number of learning samples). It should also be mentioned that all results were achieved using cutting speed and feed rate as two additional RBF NN inputs, and that all data were normalised between 0 and 1. Average indidvidual performances of features extracted from force and current signals are very comparable, although force features from the time domain achieved somewhat higher classification accuracy than similar current features. Generally, force and current features from the time domain (F1 , . . . , F4 , I1 , . . . , I4 ) have individually achieved higher classification precision than those from the frequency domain (F5 , F6 , I5 , I6 ). The difference is particularly noticeable in the case of force features. On the other hand, all AE features have individually correctly classified tool wear level in more than 50% of cases (average result), and their result can be compared with the force and current time domain features. The energy of the AE signal in the frequency range from 200 to 250 Hz (AE4) achieved the best individual result among all features (70.6%). Based on these first results, further analyses of different feature combinations have been performed, again using full hidden
Average SD
MWD
WD
68.3 78.3 91.7
81.6 80 86.7
43.3 61.7 91.7
layer structure (β C = 0, 180 hidden layer neurons). Feature combinations have generally and expectedly achieved higher classification accuracy (Table 4) compared with their individual performance. At first, combinations of features belonging to every type of signal were tested. In this case, combination of time domain current features (I1 –I4 ) and all current features (I1 –I6 ) achieved better results than the comparable combinations of force features (F1 –F4 , F1 – F6 ), thus indicating the implementation potential of current signals in the medical drill wear monitoring process. The highest average classification precision of 90% was accomplished by the combination of all AE features, which was expected based on their individual performances. At the end, selected combinations of features from different types of signals were tested (F + AE, I + AE). These tests confirmed the high average classification accuracy of all combinations. The best overall performance was observed in the case of current and AE features combination: I1 –I4 + AE1–AE7, which accomplished slightly better average result (92.8%) than the combination of AE features alone (AE1–AE7). For the best combination of features, additional tests were performed in order to analyse classification (generalisation) characteristics of different RBF NN structures. These structures were characterised by several hidden layer configurations, where the number of hidden layer neurons was smaller than the number of learning samples (β C > 0). The chosen neural network algorithm has shown satisfactory drill wear classification capacity and application potential by achieving average classification accuracy of up to 85% (Table 4) with almost four times less hidden layer neurons (43) compared with the full hidden layer structure (180). In order to establish the influence of bone mechanical properties on the performance of tool wear features of all three types of signals, average classification success rates achieved using feature combinations F1 –F6 (64.4%), I1 –I6 (73.3%) and AE1–AE7 (90%) have been presented separately for each tool wear level in Table 5.
Table 4 Results of selected feature combinations (accurately classified samples in %). Features
βC
Test
RBF NN structure
T1
T2
T3
T4
T5
Average
F1 –F4 F5 , F6 F1 –F6
0
63.9 41.7 66.7
61.1 44.4 61.1
66.7 44.4 69.4
58.3 52.8 66.7
63.9 38.9 58.3
62.8 44.4 64.4
6-180-3 4-180-3 8-180-3
I1 –I4 I5 , I6 I1 –I6
0
61.1 41.7 66.7
77.8 41.7 88.9
77.8 44.4 72.2
66.7 33.3 66.7
80.6 33.3 72.2
72.8 38.9 73.3
6-180-3 4-180-3 8-180-3
AE1–AE7 F1 –F4 + AE1–AE7 F1 –F6 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I6 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7 I1 –I4 + AE1–AE7
0 0
86.1 91.7 88.9 97.2 94.4 91.7 86.1 83.3 88.9 80.6
91.7 97.2 97.2 97.2 94.4 94.4 86.1 86.1 86.1 80.6
94.4 94.4 94.4 91.7 88.9 91.7 83.3 83.3 83.3 83.3
86.1 86.1 86.1 83.3 80.6 80.6 80.6 83.3 80.6 72.2
91.7 88.9 88.9 94.4 94.4 97.2 91.7 83.3 83.3 80.6
90.0 91.7 91.1 92.8 90.6 91.1 85.6 83.9 84.4 79.4
9-180-3 13-180-3 15-180-3 13-180-3 15-180-3 13-161-3 13-120-3 13-75-3 13-43-3 13-33-3
>0
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014
JID: JJBE
ARTICLE IN PRESS
[m5G;April 16, 2015;19:21]
T. Staroveski et al. / Medical Engineering and Physics 000 (2015) 1–7
When comparing these results, it can be noted that tool wear features from cutting force signals have the highest, and those extracted from AE signals the lowest, sensitivity to changes in bone hardness. Only 5 of 60 samples (8.3%) belonging to the WD classification group were misclassified in the case of AE features, and all of them were classified in the first nearest tool wear level (MWD). 4. Conclusion After analysing all chosen drill wear features individually and in combinations, it can be concluded that the combination of features extracted from the AE signal and two servomotor drive signals (IZ and IMS ) have the highest potential for medical drill wear classification. The application of multi-sensor data in medical drill wear monitoring is in line with the numerous research studies in the field of industrial tool wear monitoring, which have confirmed that this approach is necessary to build robust and applicable tool wear monitoring systems. Application of AE sensors seems particularly important, since features extracted from AE signals have shown the highest individual quality, as well as precision when used in combinations, and are practically insensitive to changes in bone mechanical properties. Future research activities will include analyses of several new variables on drill wear dynamic, such as different drill geometries, influence of the sterilisation process parameters and application of ultrasonically-assisted drilling technique, which has high benefits in reducing forces (torques) and surface roughness. Ethical approval In the experimental part of this research bovine tibia bone specimens were used. The bones belonged to animals slaughtered solely for the purpose of further processing in the food industry. Therefore, no ethical approval is required. Conflicts of interest None. Acknowledgement This work was financially supported by the Ministry of Science, Education and Sport of the Republic of Croatia Grant No. 120-12019481938 through the fund for national scientific projects. References [1] Alam K, Mitrofanov AV, Silberschmidt VV. Experimental investigations of forces and torque in conventional and ultrasonically-assisted drilling of cortical bone. Med Eng Phys 2011;33:234–9. [2] Alam K, Mitrofanov AV, Bäker M, Silberschmidt VV. Stresses in ultrasonicallyassisted bone cutting. J Phys Conf Ser 2009;181:012014. [3] Alam K, Mitrofanov AV, Silberschmidt VV. Measurements of surface roughness in conventional and ultrasonically-assisted bone drilling. Am J Biomed Sci 2009;1:312–20.
7
[4] Augustin G, Zigman T, Davila S, Udiljak T, Staroveski T, Brezak D, et al. Cortical bone drilling and thermal osteonecrosis. Clin Biomech 2012;27:313–25. [5] Karmani S. The thermal properties of bone and the effects of surgical intervention. Curr Orthop 2006;20:52–8. [6] Cordioli G, Majzoub Z. Heat generation during implant site preparation: an in vitro study. Int J Oral Maxillofac Impl 1997;12:186–93. [7] Eriksson RA, Albrektsson T. Temperature threshold levels for heat induced bone tissue injury: a vital microscopic study in rabbit. J Prosthet Dent 1983;50:101–7. [8] Natali C, Ingle P, Dowell J. Orthopaedic bone drills—can they be improved? J Bone Joint Surg 1996;78-B:357–62. [9] Augustin G, Davila S, Mihoci K, Udiljak T, Vedrina DS, Antabak A. Thermal osteonecrosis and bone drilling parameters revisited. Arch Orthop Trauma Surg 2007;128:71–7. [10] Paszenda Z, Basiaga M. FEM analysis of drills used in bone surgery. Arch Mater Sci Eng 2009;36:103–9. [11] Kalidindi V. Optimization of drill design and coolant systems during dental implant surgery (Ph.D. thesis, 2004). University of Kentucky, USA. URL: http://uknowledge.uky.edu/cgi/viewcontent.cgi?article=1317&context= gradschool_theses (last accessed 10.14). [12] Davidson SRH, James DF. Measurement of thermal conductivity of bovine cortical bone. Med Eng Phys 2000;22:741–7. [13] Reingewirtz Y, Szmukler-Moncler S, Senger B. Influence of different parameters on bone heating and drilling time in implantology. Clin Oral Implants Res 1997;8:189–97. [14] Bachus KN, Rondina MT, Hutchinson DT. The effects of drilling force on cortical temperatures and their duration: an in vitro study. Med Eng Phys 2000;22:685– 91. [15] Nam O, Yu W, Choi MY, Kyung HM. Monitoring of bone temperature during osseous preparation for orthodontic micro-screw implants: effect of motor speed and pressure. Key Eng Mater 2006;321–323:1044–7. [16] Allan W, Williams ED, Kerawala CJ. Effects of repeated drill use on temperature of bone during preparation for osteosynthesis self-tapping screws. Br J Oral Maxillofac Surg 2005;43:314–19. [17] Chacon GE, Bower DL, Larsen PE, McGlumphy EA, Beck FM. Heat production by 3 Implant Drill Systems after repeated drilling and sterilization. J Oral Maxillofac Surg 2006;64:265–9. [18] Queiroz TP, Souza FÁ, Okamoto R, Margonar R, Pereira-Filho VA, Garcia IR, et al. Evaluation of immediate bone-cell viability and of drill wear after implant osteotomies: immunohistochemistry and scanning electron microscopy analysis. J Oral Maxillofac Surg 2008;66:1233–40. [19] Jochum RM, Reichart PA. Influence of multiple use of Timedur – titanium cannon drills: thermal response and scanning electron microscopic findings. Clin Oral Implants Res 2000;11:139–43. [20] Singh J, Davenport JH, Pegg DJ. A national survey of instrument sharpening guidelines. Surg 2010;8:136–9. [21] Jantunen E. A summary of methods applied to tool condition monitoring in drilling. Int J Mach Tools Manuf 2002;42:997–1010. [22] Davidson S.R.H. Heat transfer in bone during drilling (M.S. Thesis, 1999). Institute of Biomaterials and Biomedical Engineering, University of Toronto, Canada. URL: https://tspace.library.utoronto.ca/bitstream/1807/14728/1/MQ46070.pdf (last accessed 10.14). [23] Lee J, Rabin Y, Ozdoganlar OB. A new thermal model for bone drilling with applications to orthopaedic surgery. Med Eng Phys 2011;33:1234–44. [24] Laurito D, Lamazza L, Garreffa G, De Biase A. An alternative method to record rising temperatures during dental implant site preparation: a preliminary study using bovine bone. Ann dell’Istituto Super Sanita 2010;46:405– 10. [25] Prickett PW, Johns C. An overview of approaches to end milling tool monitoring. Int J Mach Tools Manuf 1999;39:105–22. [26] Scheffer C, Heyns PS, Klocke F. Development of a tool wear-monitoring system for hard turning. Int J Mach Tools Manuf 2003;43:973–85. [27] Brezak D, Udiljak T, Mihoci K, Majetic D, Novakovic B, Kasac J. Tool wear monitoring using radial basis function neural network. In: International joint conference on neural networks – IJCNN, Budapest, Hungary, vol. 3; 2004. p. 1859–63. R
Please cite this article as: T. Staroveski et al., Drill wear monitoring in cortical bone drilling, Medical Engineering and Physics (2015), http://dx.doi.org/10.1016/j.medengphy.2015.03.014