You are on page 1of 4

CIRP Journal of Manufacturing Science and Technology 25 (2019) 22–25

Contents lists available at ScienceDirect

CIRP Journal of Manufacturing Science and Technology


journal homepage: www.elsevier.com/locate/cirpj

Adaptive thermal displacement compensation method based on deep


learning
Makoto Fujishima(3)a , Koichiro Narimatsua,* , Naruhiro Irinoa , Masahiko Mori(1)a ,
Soichi Ibaraki(2)b
a
DMG MORI Co., Ltd., Meieki 2-35-16, Nakamura, Nagoya 450-0002, Japan
b
Department of Mechanical Systems Engineering, Hiroshima University, Kagamiyama 1-4-1, Higashi-Hiroshima 739-8527, Japan

A R T I C L E I N F O A B S T R A C T

Article history: Temperature variation of the machine tool structure due to machine internal heat generation or heat
Available online 16 May 2019 exchange with the ambient environment causes thermal deformation. To compensate the deformed
machines, various methods have been proposed. However, these methods failed to compensate machines
in severe situations such as unexpected temperature change and sensor failure. In this paper, a novel
Keywords: thermal displacement compensation method using deep learning is proposed. In the proposed algorithm,
Machine tool
reliability in thermal displacement prediction is evaluated based on deep learning to change the
Thermal error
compensation weight adaptively. High-performance thermal displacement compensation result on
Compensation
turning center is presented.
© 2019 CIRP.

Introduction displacement prediction from the temperature under the ambient


temperature change [12]. On the other hand, in order to apply deep
In recent years, due to lack of labor force in the manufacturing learning technology to the industrial field and to compensate
industry, automation systems and robotized workpiece transpor- thermal displacement, it is important to clarify the reliability of
tation systems have been required for unmanned or reduced prediction. In case of low prediction reliability, such as the
personnel factory. For continuous machining operation by unlearned operating condition and sensor failure, it is necessary to
automation, it is important to suppress the deterioration of suppress the deterioration of accuracy.
accuracy. Thermal deformation has significant effects on the In this paper, the reliability evaluation method based on
machining accuracy under machining operation. Internal and Bayesian dropout [13] is presented. Adaptive compensation
external heat sources of machine tools and their changes are the method which adjusts the compensation value based on the
main factor that deteriorates the machining accuracy. To achieve evaluated reliability value is proposed. This method can prevent
high quality and high-performance automated manufacturing the deterioration of accuracy due to extreme compensation values
system, the techniques to successfully suppress thermal displace- caused by sensor failure with extreme fluctuation of measured
ment have been required. Much research has been carried out on value due to disconnection. This method also adapts to cutting heat
this topic. However, not many good results were gained in practice. generation, operating heat generation, such as spindle rotation and
Against this issue, various thermal displacement control techni- axes movement, and ambient temperature change. The results
ques by controlling temperature and compensation techniques have show that the deterioration of accuracy can be adaptively
been developed [1,2]. For example, method using thermal sensors prevented according to the mode of sensor failure.
[3–6], prediction from NC unit internal data [7] and prediction by
neural networks [8] were developed. In recent years, in the artificial Calculating prediction uncertainty in deep learning
intelligence technologies including deep learning, the performance
of image recognition has become more accurate [9–11]. A deep learning regression predicts some value from input data.
The authors presented that convolutional neural network In the prediction of thermal displacement, displacement value is
model using deep learning had high performance of thermal predicted from data such as values of temperature sensors. In the
deep learning classification and the distinction, reliability of
prediction is evaluated by using ‘softmax’ as an output function
* Corresponding author. [13]. However, it has been difficult to easily calculate reliability in
E-mail address: ko-narimatsu@dmgmori.co.jp (K. Narimatsu). the deep learning regression.

https://doi.org/10.1016/j.cirpj.2019.04.002
1755-5817/© 2019 CIRP.
M. Fujishima et al. / CIRP Journal of Manufacturing Science and Technology 25 (2019) 22–25 23

Fig. 3. Machine structure of turning center.

Fig. 1. Schematic diagram of dropout.

The ‘Bayesian dropout’ approach uses dropout to evaluate


reliability by calculating uncertainty. Dropout technique has been
widely used to prevent overfitting of deep learning models [14,15].
The concept of dropout for neural networks used in deep learning is
shown in Fig. 1. The dropout technique is normally used to train
models, but Bayesian dropout uses dropout technique not only
during training, but also during prediction. This method can be easily
implemented because trained model can be used without modifica-
tion and no additional processing is required for input data.
The results of the displacement prediction from temperature
Fig. 4. Displacement sensors and directions.
using Bayesian dropout are shown in Fig. 2. The blue line is the
measured thermal displacement of a turning center in X-axis temperature values are measured by thermocouples attached on the
direction, which is the direction of change in diameter, in the spindle machine structure of the turning center shown in Fig. 3, and the
operating condition. In order to calculate the prediction uncertainty displacement values are obtained by measuring the relative movement
range, dropout was applied to the trained deep learning model for between workpiece and turret. As shown in Fig. 4, there are five
displacement prediction, and the calculations were repeated 100 displacement sensors: two are for measuring displacement in machine
times while switching the node to be dropped. From the calculated X-axis direction, two are for Y-axis direction, and one is for Z-axis
values, the distribution of the predicted values shown in green color direction.
density. The thin red lines are the prediction uncertainty range Table 1 shows the variants in the experimental conditions. On
(mean  2s ). The tick red line is the predicted displacement. and off of coolant heating to emulate the cutting heat, spindle
operation, axes movement, and ambient temperature change are the
Thermal displacement prediction from temperature under variants. Thermal displacement was measured with single operating
various operating conditions condition turned on by itself and combination of two operating
conditions. The movable temperature variable booth shown in Fig. 5
Thermal displacement under various operating conditions was used for the ambient temperature change condition.

Thermal displacement and temperature data of the turning center Thermal displacement prediction by deep learning and evaluation of
were collected to investigate the accuracy and uncertainty of thermal prediction uncertainty
displacement prediction under various operating conditions. The
Using the collected data, thermal displacement prediction and
calculation of prediction uncertainty were carried out by deep
learning. The neural network model shown in Fig. 6 was used for
training and prediction. Temperature values for 30 min of 12
temperature sensors were given as input data. 5 displacement
values were also given as supervised data. From the collected data,
41 data from 3 machines were used as training data, 8 data from
other 3 machines were used as evaluation data.
The results of the prediction and the prediction uncertainty are
shown in Fig. 7. The blue line is the measured thermal
displacement of the X-axis direction. The tick red line is the

Table 1
Experimental variants.

Name Operating condition


Condition 1 Coolant heated at the equivalent heat of cutting
Condition 2 Spindle running for 8 h at 4,000 min 1
Condition 3 X and Z axes moving at 30,000 mm/min for 8 h
Condition 4 Ambient temperature increasing and decreasing by 8  C in 9 h
Fig. 2. Calculation results of prediction uncertainty.
24 M. Fujishima et al. / CIRP Journal of Manufacturing Science and Technology 25 (2019) 22–25

Fig. 5. Movable temperature variable booth and air conditioner.

Fig. 7. Prediction results of displacement in the X-axis.

Table 2
Sensor failure simulation modes.

Failure mode Value


Mode 1. Resistive sensor contact failure NV/OR switch randomly
Mode 2. Voltage sensor disconnection 0
Mode 3. Resistive sensor disconnection OR
Mode 4. Voltage sensor contact failure NV/0 switch randomly

NV: normal value, OR: over range ( 128  C).

of disconnection. Voltage sensors (modes 2 and 4) output zero value


in case of disconnection. Assuming that one of the input temperature
sensors failed, displacement was predicted by inputting the wave-
forms shown in Fig. 8 to the trained model.
Fig. 9 shows the prediction results of X-axis thermal displace-
ment and the prediction uncertainty. In the failure modes 1 and 3,
the values of the predicted thermal displacement greatly change at
around 10 h when sensor signal fails and the prediction
uncertainty also increase simultaneously. On the other hand, in
Fig. 6. Schematic diagram of neural network model. the failure modes 2 and 4, the influences on the prediction
accuracy are small, and a change in uncertainty is hardly observed.
predicted displacement. The thin red lines are the prediction As described above, in a situation where the prediction accuracy
uncertainty. The (a1) and (a2) charts are evaluation data, the (b1) deteriorates due to a failure of the sensor, it can be determined that
and (b2) charts are training data, the (a1) and (b1) charts are
ambient temperature change conditions, and the (a2) and (b2)
charts are main spindle operating conditions. Despite the training
data and the evaluation data are measured in the same condition
but on different machines, the prediction accuracy and the
uncertainty are similar degrees. Therefore, it can be appeared
that the trained model has robustness for the unlearned data.

Prediction uncertainty variation and adaptive adjusting in


thermal displacement compensation value

Prediction uncertainty variation due to sensor failure simulation

In order to investigate how much the displacement prediction


accuracy and the prediction uncertainty change when an unintended
fluctuation occurs in the input temperature data, thermal displace-
ment was predicted with intentionally tampering temperature
sensor data to emulate sensor failure. Table 2 shows the sensor
failure modes. Two types of temperature sensors were considered to
simulate values when the sensor is disconnected. Resistive sensors
(modes 1 and 3) output an extreme abnormal value ( 128  C) in case Fig. 8. Simulated waveforms of sensor failure.
M. Fujishima et al. / CIRP Journal of Manufacturing Science and Technology 25 (2019) 22–25 25

By adjusting the compensation value based on the uncertainty, the


deterioration of accuracy can be adaptively prevented when a failure of
the sensor or an unexpected machine operating condition occurs.

Conclusion

In this paper, the prediction reliability evaluation method for


thermal displacement prediction based on Bayesian dropout was
presented. The method adapted to not only ambient temperature
change but also cutting heat generation and operating heat
generation by spindle rotation or axes movement. The results
showed that the trained deep learning model had good prediction
performance and the method could evaluate prediction reliability
as uncertainty. By adjusting the compensation values based on the
uncertainty, the method can prevent the deterioration of accuracy
due to extreme prediction values caused by sensor failure.

Conflict of interest

Fig. 9. Simulation results of sensor failure. Makoto Fujishima and Masahiko Mori are board menbers of
DMG MORI CO., LTD. Koichiro Narimatsu and Naruhiro Irino are
employees of DMG MORI CO., LTD. Soichi Ibaraki is an academic
supervisor of Koichiro Narimatsu.

References

[1] Mori, M., Mizuguchi, H., Fujishima, M., Ido, Y., Mingkai, N., Konishi, K.,
2009, Design Optimization and Development of CNC Lathe Headstock to
Minimize Thermal Deformation. CIRP Annals — Manufacturing Technology,
58/1: 331–334.
[2] Mayr, J., Jedrzejewski, J., Uhlmann, E., Donmez, M.A., Knapp, W., Härtig, F.,
Brecher, C., 2012, Thermal Issues in Machine Tools. CIRP Annals — Manufactur-
ing Technology, 61/2: 771–791.
[3] Moriwaki, T., 1988, Thermal Deformation and Its On-line Compensation of
Hydrostatically Supported Precision Spindle. CIRP Annals — Manufacturing
Technology, 37/1: 393–396.
[4] Moriwaki, T., Shamoto, E., 1998, Analysis of Thermal Deformation of an
Ultraprecision Air Spindle System. CIRP Annals — Manufacturing Technology,
47/1: 315–319.
[5] Moriwaki, T., Shamoto, E., Kawano, M., 1995, Estimation of Thermal Deforma-
tion of Machine Tool by Applying Neural Network (Improvement of Estimation
Accuracy by Utilizing Time-series Data of Temperature on Machine Surfaces).
Transactions of the Japan Society of Mechanical Engineers Part C, 61/584:
1691–1696. (in Japanese).
[6] Mayr, J., Blaser, P., Ryser, A., Hernandez-Becerroa, P., 2018, An Adaptive Self-
learning Compensation Approach for Thermal Errors on 5-axis Machine Tools
Fig. 10. Simulation result of compensation value adjust at sensor failure. Handling an Arbitrary Set of Sample Rates. CIRP Annals — Manufacturing
Technology, 67/1: 551–554.
[7] Brecher, C., Hirsch, P., Weck, M., 2004, Compensation of Thermo-elastic
the prediction reliability has deteriorated from the increased Machine Tool Deformation Based on Control Internal Data. CIRP Annals —
uncertainty. If a threshold value is set for the uncertainty, it is Manufacturing Technology, 53/1: 299–304.
possible to detect that prediction accuracy deteriorates. [8] Kang, Y., Chang, C.W., Huang, Y., Hsu, C.L., Nieh, I.F., 2007, Modification of a
Neural Network Utilizing Hybrid Filters for the Compensation of Thermal
Deformation in Machine Tools. International Journal of Machine Tools and
Adaptive adjusting of compensation value based on prediction Manufacture, 47/2: 376–387.
uncertainty [9] Schroff, F., Kalenichenko, D., Philbin, J., 2015, Facenet: A Unified Embedding for
Face Recognition and Clustering. in: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 815–823.
If the prediction uncertainty is large, prediction reliability is [10] He, K., Zhang, X., Ren, S., Sun, J., 2015, Delving Deep into Rectifiers: Surpassing
deteriorated. To suppress the deterioration of the compensation Human-level Performance on Imagenet Classification. in: Proceedings of the
IEEE International Conference on Computer Vision, 1026–1034.
accuracy, the compensation values were overridden using a [11] He, K., Zhang, X., Ren, S., Sun, J., 2016, Deep Residual Learning for Image
function of uncertainty value. The moving average filter param- Recognition. in: Proceedings of the IEEE Conference on Computer Vision and
eters were also adjusted to obtain a smooth fluctuation of Pattern Recognition, 770–778.
[12] Fujishima, M., Narimatsu, K., Irino, N., Ido, Y., 2018, Thermal Displacement
compensation value. Fig. 10 shows the simulation results with
Reduction and Compensation of a Turning Center. CIRP Journal of Manufactur-
adjusted compensation values based on the prediction uncertainty. ing Science and Technology, 22:111–115.
The blue line is the measured thermal displacement of the X-axis [13] Gal, Y., Ghahramani, Z., 2016, Dropout as a Bayesian Approximation: Repre-
direction. The red line is the predicted displacement. The light blue senting Model Uncertainty in Deep Learning. International Conference on
Machine Learning, 1050–1059.
line is the adjusted compensation values. The uncertainty values [14] Hinton, G.E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.R.,
greatly affected by the sensor failure simulation in failure modes 1 2012, Improving Neural Networks by Preventing Co-adaptation of Feature
and 3. The deterioration of accuracy can be suppressed by adjusting Detectors.. https://arxiv.org/abs/1207.0580.
[15] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014,
the compensation values. On the other hand, in the failure modes 2 Dropout: a Simple Way to Prevent Neural Networks from Overfitting. The
and 4, which have small influence on the uncertainty due to the Journal of Machine Learning Research, 15/1: 1929–1958.
sensor failure, the compensation values are not adjusted, and good
accuracy is maintained.

You might also like