Self-learning control strategy with application to milling system.
Cus, Franc ; Zuperl, Uros ; Gecevska, Valentina 等
1. INTRODUCTION
In the operation of computer numerically controlled (CNC) machine
tools, the use of large parameters such as the cutting depth and the
feed will significantly increase the production rate. However, as the
cutting force increases because of the increase of the cutting depth and
feed, the thermal expansion of the tool tip, tool and workpiece deflection, and machine chatter will rise such that the workpiece
precision is reduced. Consequently, maintaining the cutting force on the
tool tip at the appropriate value, despite variations in depth of the
cut, is one way of guaranteeing that the dimension error is permissible.
Machining processes are difficult to control because of system
nonlinearities and time-varying parameters, owing to variations in
cutting depth. Therefore, the use of an adaptive control systems to
increase productivity is achieved by an automatic control of feedrate to
maintain a constant spindle load (Liu & Wang, 1999). Due to problems
such as instability and large transient overshoots when the machining
conditions deviated from those for which the controller was designed
(Balic, 2000), the use of new modern adaptive control algorithms has
been proposed. The model reference adaptive control (MRAC) scheme
(Tomizuka et al., 1983), self-tuning control strategies (Huang &
Lin, 2002), and other adaptive control approaches (Zuperl & Cus,
2003) have been experimentally tested.
[FIGURE 1 OMITTED]
To design an adaptive controller requires modelling of the
machining process. It is very difficult to model the milling process
dynamics due to highly nonlinear and time-varying cutting
characteristics.
Therefore, an adaptive learning control of milling processes is
developed in this paper. The adaptive learning control system adaptively
acquires the knowledge of the controlled process through on-line
learning. It consists of a feedforward neural network and fuzzy feedback
mechanism.
The inputs to the controller are the error in the cutting force and
the change of the error of the cutting force. The neural network
predicts the inverse-dynamics model of the controlled process and fuzzy
feedback mechanism is used to guide an adaptive modification of
connection weights of neural network. By these two elements, the inverse
dynamics model of the controlled plant can be adaptively modified in
response to the variations in cutting conditions so as to obtain an
adjustable feedrate with a constant milling force automatically.
Experimental cutting tests are performed to verify the efficiency of
this adaptive learning control system.
2. ARCHITECTURE OF ADAPTIVE LEARNING CONTROL
Figure 1 shows the block diagram of the adaptive learning control
system which is used in milling to achieve an automatic on-line
adjustment of feedrate with a constant milling force [F.sub.ref].
[FIGURE 2 OMITTED]
The milling force will increase when the depth of cut increases in
the milling process. The control system immediately decreases feedrates
to avoid tool breakage. When the depth of cut decreases the system
generates a larger feedrate to maintain high cutting efficiency.
The developed system controls the peak milling force F in a tooth
period. The measured milling force F passes through a tapped delay line
(TDL) filter whose output vector contains the delayed values of the
measured milling force. Then, the delayed values of the milling force
are fed into the multi-layer feedforward neural network.
The neural network has 4 neurons in input layer, 6 neurons in
hidden layer and 1 neuron in output layer. The learning rate and the
momentum parameter are set to be 0.01 and 0.5, respectively. A limiter constrains the command signal to avoid any damage due to the rapid
feedrate. By extensive testing and simulations it is found that good
control performances can be achieved by using a 3-layer feedforward
neural network of a 4-6-1 type (Figure 2). Many different neural
networks have been tested and simulated during this research
(perceptrons, Hebbian, backpropagation networks). In the current work,
two supervised neural networks for modelling are compared. The first one
is a back propagation neural network (BP) with sigmoid transfer
functions in hidden layers and linear transfer function in the output
layer; the second is a radial basis network (RBN) with Gaussian
activation functions.
The fuzzy feedback mechanism (Figure 3) consists of a fuzzifier, a
knowledge base, a fuzzy inference engine and a defuzzifier. For the
fuzzy feedback mechanism, the input scaling factors (0.516/0.0331) and
output (5) were chosen.
[FIGURE 3 OMITTED]
3. EXPERIMENTAL TESTING AND DISCUSSION
In the experiments, a ball-end milling cutter (R218-16B20-030) with
two cutting edges was mounted on a Heller Bea 01 CNC machining center
equipped with a Fagor CNC controller. The cutting inserts R218-16 03 M-M
with 12[degrees] rake angle were selected. Cutting conditions are:
milling width [R.sub.D] = 16 mm, milling depth [A.sub.D] = 4 mm and
cutting speed v = 95 m/min.
The milling force signal was measured by using a table type
dynamometer (Kistler 9255B) mounted between the workpiece and the
machining table and recorded on a PC workstation through a data
acquisition board (PC-MIO-16E-4). To use the developed system on Figure
1 and to adjust the feedrate, the desired cutting force is [[F.sub.ref]]
= 280 N and pre-programed feed is 0.11 mm/teeth. The developed adaptive
control algorithm could not be directly implemented on Fagor controller.
Feedrate override panel provided by the CNC controller was connected to
the PC workstation.
Communication between the control system and the CNC machine
controller is accomplished over RS-232 protocol. The adaptive learning
control algorithm was then installed on the PC to adjust the feedrate
command. Test workpiece contains step changes of the axial depths of
cut.
The summation of the inverse-dynamics model [f.sub.m], and the
fuzzy feedback mechanism ff is actual feedrate which is sent to the CNC
controller. The experiment demonstrate that when the end mill starts to
cut the workpiece with a step increase of axial depth of cut (4 mm), the
cutting force immediately increases and even exceeds the reference
cutting force, 300 N. Then the fuzzy feedback mechanism modifies the
connection weights of the neural network.
As a result, the output signal of the inverse-dynamics model
decreases immediately and so does the feedrate.
4. CONCLUSION
In this paper, an adaptive learning control system is proposed to
control the milling process of milling with both constant cutting force
and fixed metal removal rate. The proposed control system consists of
two parts. A feedforward neural network is first used to acquire the
inverse-dynamics model of the controlled process.
Then, a fuzzy feedback mechanism is designed to perform an adaptive
modification of neural network connection weights. By the use of
adaptive learning control system the machining time is reduced for 24 %.
The experiments are performed to show that the adaptive learning
controller has the intelligence to maintain a constant metal removal
rate and milling load under varying cutting conditions. Further research
will be needed to determine the stability of the developed control
system.
5. REFERENCES
Balic, J. (2000). A new NC machine tool controller for step-by-step
milling. Int. j. adv. manuf. Technol., Vol. 8, 399-403
Huang, S.J. & Lin, C.C. (2002). A self-organising fuzzy logic controller for a coordinate machine. Int. J. Adv. Manuf. Technol., Vol.
19, 736-742
Liu, Y.; Zuo, L. & Wang, C. (1999). Intelligent adaptive
control in milling process. International Journal of Computer Integrated
Manufacturing, Vol. 12, 453-460
Tomizuka, M.; Oh, J.H. & Dornfeld, D.A. (1983). Model Reference
Adaptive Control of the Milling Process. Proceedings of the Symposium on
Manufacturing on Manufacturing Process and Robotic Systems, New York,
55-63
Zuperl, U. & Cus, F. (2003). Optimization of cutting conditions
during cutting by using neural networks. Robot. comput. integr. manuf.,
Vol. 19, 189-199
