Load control in high speed milling.

Zuperl, U. ; Kiker, E. ; Cus, F. 等

Abstract: In this paper, a neural controller with optimisation for the ball end milling process is described. Architecture with two different kinds of neural networks is proposed, and is used for the on-line optimal control of the milling process. A BP neural network is used to identify the milling state and to learn the appropriate mappings between the input and output variables of the machining process. The feedrate is selected as the optimised variable, and the milling state is estimated by the measured cutting forces. The goal is also to obtain an improvement of the milling process productivity by the use of an automatic regulation of the cutting force. Numerous simulations are conducted to confirm the efficiency of this architecture.

Key words: Force control, neural controller, end-milling, machining.

1. INTRODUCTION

A remaining drawback of modern CNC systems is that the machining parameters, such as feedrate, speed and depth of cut, are programmed off-line. The machining parameters are usually selected before machining according to programmer's experience and machining handbooks. To prevent damage and to avoid machining failure the operating conditions are usually set extremely conservative. As a result, many CNC systems are inefficient and run under the operating conditions that are far from optimal criteria. Even if the machining parameters are optimised off-line by an optimisation algorithm (Cus & Balic, 2003) they cannot be adjusted during the machining process. To ensure the quality of machining products, to reduce the machining costs and increase the machining efficiency, it is necessary to adjust the machining parameters in real-time, to satisfy the optimal machining criteria. For this reason, adaptive control, which provides on-line adjustment of the operating conditions, is being studied with interest (Liu & Wang, 1999).

2. ADAPTIVE CONTROL WITH OPTIMIZATION IN MILLING

To realise real-time optimal control of the machining process, an ALM neural network is proposed. It is used to determine the optimal inputs (feedrate), so we shall refer to it as a neural network (NN) for optimization (Fig. 1). This architecture that uses the Lagrange multiplier (ALM) method converges more quickly than other penalty methods. Detail information about this type of network can be found in (Balic, 2000). This method was first introduced in turning by (Balic, 2000). Combining both neural networks, an adaptive controller for the milling process is designed. Sequence of steps for on-line optimization of the milling process is presented below:

* NN for optimization determines the optimal feedrate and sends it to the milling machine and network for modelling,

* the measured output of the milling machine are used to train the NN for modelling,

* NN for optimization uses the newly upgraded neural model to find the optimal feedrate and sends it to the machine and neural model,

* steps 2 and 3 are repeated until termination of machining.

[FIGURE 1 OMITTED]

To realise the on-line modelling of the milling system, a standard BP NN is proposed based on the popular back propagation leering rule (Milfelner, 2002; Lee & Lin, 2000). During preliminary experiments it proved to be sufficiently capable of extracting the force and surface roughness model directly from experimental machining data. It is used to describe the cutting process. The NN for modelling needs four input neurons for milling federate (f), cutting speed ([v.sub.c]), axial depth of cut ([A.sub.d]) and radial depth of cut (Rd). The output from the NN is cutting force components and surface roughness.

3. COMPUTER SIMULATIONS AND EXPERIMENTAL PREPARATION

To examine the stability of the adaptive control, the simulations for variable milling depth have been carried out. The simulations and experiments with the ball-end milling cutter were prepared on the NC milling machine (type HELLER BEA1).

The objective of neural control is keeping the metal removal rate (MRR) as high as possible and maintaining cutting force as close as possible to a given reference value. In simulations the real end milling process was replaced with trained neural model (Cus & Balic, 2000).

[FIGURE 2 OMITTED]

Since an untrained modelling network has no knowledge about the system the initial training was applied to network. The adaptive controller is operated on PC and the adjusted feedrates are sent to CNC. For this purpose we carry out some tests. These tests used CK 45 and Ck 45 (XM) steel workpiece with an irregular profile. To optimise the feedrate, the constraints are [F]=280 N, pre-programmed feed is 0.08 mm/teeth and its allowable adjusting rate is from 0 to 150%.

4. DISCUSSION OF RESULTS

In the first simulation using constant feed rates (conventional cutting-Fig. 2a) the MRR reaches its proper value only in the last step. However, in second test (Fig. 2b), machining the same piece but using neural control, the average MRR achieved is much more close to the proper MRR. Comparing the Fig. 2a to Fig. 2b, the cutting force for the neural control milling system is maintained at about 240N, and the feedrate of the adaptive milling system is close to that of the traditional CNC milling system from point C to point D. From point A to point C the feedrate of the adaptive milling system is higher than for the classical CNC system, so the milling efficiency of the adaptive milling system is improved. The simulation results show that the milling process with the designed neural controller has a high robustness, stability, and also higher machining efficiency than standard controllers. Neural network adaptive control ensures continuous optimising feedrate control that is automatically adjusted to each particular cutting situation. When spindle loads are low, the system increases cutting feeds above and beyond pre-programmed feedrates, resulting in considerable reductions in cycle times and production costs. When spindle loads are high the feedrates are lowered, safeguarding machine tools and workpieces from damage and tool breakage. When system detects extreme load conditions, it automatically stops the machine to protect the cutting tool. The proposed system is still in experimental phase but it can be easily extended with minimal costs to real industrial use. It reduces the need for constant operator supervision.

5. CONCLUSION

The purpose of this contribution is to present a reliable, robust neural force controller aimed at adaptively adjusting feedrate to prevent excessive tool wear, tool breakage and maintain a high chip removal rate. The approach was successfully applied to an experimental milling centre Heller. The proposed architecture for on-line determining of optimal cutting conditions is applied to ball-end milling in this paper, but it is obvious that the system can be extended to other machines to improve cutting efficiency.

6. REFERENCES

Balic, J. (2000). A new NC machine tool controller for step-by-step milling. Int. j. adv. manuf. technol., Vol. 2, pp. 399-403.

Cus, F. & Balic, J. (2000). Selection of cutting conditions and tool flow in flexible manufacturing system. The international journal for manufacturing science & technology, Vol. 2, pp. 101-106.

Cus, F. & Balic, J. (2003). Optimization of cutting process by GA approach. Robot. comput. integr. manuf., Vol. 19, pp. 113-121.

Lee, T.S. & Lin, Y.J. (2000). A 3D Predictive Cutting-Force Model for End Milling of Parts Having Sculptured Surfaces, International Journal of Advanced Manufacturing Technology, Vol. 16, pp. 773-783.

Liu, Y. & Wang, C. (1999). Neural network based adaptive control and optimisation in the milling process, International Journal of Advanced Manufacturing Technology, Vol. 15, pp. 791-795.

Milfelner, M. (2002). Cutting Forces in End Milling. Acta Mech. Slovaca, pp. 107-112.