Artificial Neural Networks for Controlling the Temperature of Internally Cooled Turning Tools ()
1. Introduction
A closed loop cooling system integrated with a novel design of tool insert, Figure 1, has many advantages: it eliminates a potential health hazard for machine operators; it is environmentally friendly as there is no coolant wastage; it consumes less power; swarf is easier to dispose of and it has the potential to reduce machine operating costs, [1]. However, confining coolant to internal channels built into the tool insert, presents new challenges: first of all, it will inevitably change the way in which the heat produced by the cutting operation is distributed and disposed of and secondly it will change the state of lubrication at the tool—work interface. This paper is concerned with the first of these challenges where, with the new technology, there is a risk that the local temperature at the cutting edge of the tool, may, under some machining conditions, exceed a permissible threshold and therefore reduce or at least offset any benefit in tool life.
As for internally cooled tools, coolant is completely under control, the measurement of tool temperature becomes a more meaningful and practical proposition. It is not subject to large fluctuations by small adjustments of coolant flow and direction as is the case for externally applied coolant and it therefore relates, to a much greater extent, to the machining conditions, work piece material, tool design and state of wear. Thus, routinely monitoring the temperature of internally cooled tools is one way of observing and hence avoiding the conditions that may create excessively high temperatures. Furthermore, there is now a wealth of evidence to show [2,3] that for many materials there is an optimum cutting temperature which gives maximum tool life and good surface integrity so to be able to control tool temperature has real practical advantages. However this requires tool temperature measurement to be combined with an intelligent control system that can interpret the measurements and signify some form of corrective action. For example, simply knowing temperature allows a judgment to be made about
whether it is too high for the tool or work piece material, however knowing the machining conditions along with the temperature also enables the state of tool wear to be assessed when the temperature is well within material limits. Control systems based on artificial neural networks (ANN) are well suited to this form of decision making. They need to be trained with experimental data to generate a structured relationship between a number of input variables, in this case, machining conditions and an output variable-tool temperature. This relationship provides a reference for interpreting any new tool temperature measurements taken for any machining conditions within an allowable range. Furthermore, the ANN may be retrained any number of times enabling the interpretation tool temperature to be adapted to different work piece materials and tool designs.
Evaluating the benefit of using an ANN based system for controlling the temperature of internally cooled turning tools is the main objective of this paper. In this work, the tool is part of a system that incorporates a variable speed coolant delivery pump so that for a given tool design and work piece material the dominant input variables are coolant flow rate, cutting speed, cutting depth and feed/rev. Essential to the use of an ANN for controlling tool temperature in this system is how accurately it can predict tool temperature given these four input variables. But also of importance is how the prediction accuracy depends on the amount of experimental data and the time required for training.
Past work on ANN’s applied to process/machine control is extensive with more than 1000 papers published since 1990. Whilst there are many types of ANN that can be used for machine tool or industrial process control those based on the Multi Layer Perceptron (MLP) have been used in more than 60% of investigations and have met with most success, [4]. They have generally employed an activation function attached to hidden neurons to improve the accuracy of prediction for the continuous and nonlinear target data characteristic of real control systems. Training has generally been performed off line utilizing back propagation or feed forward algorithms i.e. [5-14]. However what isn’t clear from the published work is a strategy for determining the optimum ANN design for a given application. ANN designs vary widely and in many publications there is no indication as to how a particular design has been arrived at whilst in others, trial and error techniques or systematic variation of parameters have been used to determine an optimum design based on application specific, experimental data. Even in publications where systematic optimization of the ANN design has been undertaken, there have been apparently contradictory findings. For example, [15] found the best results were obtained with a low number of neurons in the hidden layer whilst [16] concluded that increasing the number of neurons improved the results. Alternatively [6] concluded that best results were achieved with an optimum number of neurons. Ref [17] presents the results of a wide ranging investigation into the effect of ANN design on prediction accuracy for data associated with wound toroidal cores. In this case, accuracy improved with the total number of neurons in the hidden layers but it didn’t matter whether they were arranged in a single layer or several layers.
The lack of clear guide lines for ANN design and apparently contradictory findings of past work, at least in relation to process control, suggests ANN design and hence performance is strongly dependent on the experimental data it is required to predict. For the case of internally cooled turning tools the data may be characterized by an extremely wide dynamic range, significant experimental scatter and a non linear relationship between the output parameter—tool temperature and the machining parameters that constitute the input variables.
Turning is a universal process performed on a wide range of materials with an equally wide range of tool designs, machining speeds may vary from a few mm/s to tens of m/s, similarly depths of cut may range from a fraction of a mm to more than ten mm and tool feed/rev from tens of microns to a several mm. To determine whether ANN’s can be of practical use to industry it is necessary to know how ANN design and prediction accuracy depends on data range; whether or not a single ANN design can cover a useful range of machining conditions and if so how much experimental data will be required for its training?
For a given tool and work piece material, tool temperature is primarily dependent on the machining conditions—speed, depth of cut and feed/rev. But it is also affected to a minor extent by many other factors that may collectively have a significant influence [18,19]. Tolerances on material specifications, tool geometries, the accuracy to which the work piece can be located, its initial geometry and how well it is clamped are just a few examples of variables that may influence tool temperature in an apparently random manner. How well do ANN’s cope with this type of variation, is it necessary to smooth or precondition experimental data before training the ANN in order to predict the required dominant trends or can the ANN filter out these random effects?
The dependence of tool temperature on machining conditions is well known to be non linear with temperature generally approaching some limiting value as cutting speed, depth and feed/rev continue to increase in severity. ANN’s with activation functions attached to hidden layer neurons are reported to cope well with non linear data but what is the best activation function to use, how does prediction accuracy depend on the degree of non linearity and how much experimental data is required for training?
In an attempt to answer the above questions and hence assess the suitability of an ANN based system for controlling the temperature of internally cooled tools, machining trials have been performed on CNC lathes fitted with conventional tools and with internally cooled tools and variable flow rate cooling systems. Conventional tools were used to generate dry machining data with which to assess the effect of ANN design on prediction accuracy; dry machining being a reference condition for internally cooled tools. The results of this study were then used to optimise ANN design for experimental data generated by internally cooled tools.
2. Dry Machining Experiments
A small desk top lathe, a Weiss WM280V-F was used for dry machining experiments. It was capable of handling work pieces of up to 250 mm diameter and 500 mm long. Spindle speed was infinitely variable from 50 to 2000 rpm and the tool feed/rev was incrementally variable between 0.07 and 0.42 mm. Motor power was limited to 2 kW peak and was the limitation to the depth of cut that could be used. The tool comprised of a tool holder with a tungsten carbide insert. The insert, DCMT060202, was diamond drilled from its underside as shown in Figure 2. A 0.6 mm hole positioned 1 mm from the cutting tip, was drilled to reach to within 0.8 mm of the cutting surface.
Figure 2. Diamond drilled tungsten carbide insert, type DCMT060202.
A slot was ground into the tool holder to allow a J type thermocouple, 0.5 mm diameter to access the hole in the insert and to be spring loaded against the end of the hole, 0.8 mm from the cutting surface. The lathe and tool were set up to dry machine 6082-T6 aluminium bar 75 mm in diameter and 300 mm long with a tailstock to provide rigid support to the work piece for all tests.
Experiments consisted of setting a depth of cut, the feed/rev and a surface speed and then maintaining the cut until the tool temperature reached a stable value. Figure 3 shows a typical tool temperature—time recording. In this case the stable cutting temperature was 119.6˚C. A series of 96 different cutting trials were performed covering the range of conditions: Cutting speeds, 0.2 - 3.3 m/s; Feed/rev, 0.07 - 0.42 mm; Depth of cut, 0.1 - 1.5 mm. For these conditions, recorded tool temperatures varied from 24˚C to 172˚C. Figures 4-6 show the experimental data obtained for feeds of 0.07, 0.28 and 0.42 mm respectively.
Figure 3. Typical tool temperature-time recording.
Figure 4. Experiments performed for a 0.07 mm feed/rev and 1.5, 0.8, 0.4, 0.2 and 0.1 mm depths of cut.
Figure 5. Experiments performed for a 0.28 mm feed/rev and 1.5, 0.8, 0.4, 0.2, 0.1 and 0.05 mm depths of cut.
Figure 6. Experiments performed for a 0.42 mm feed/rev and 0.8, 0.4, 0.2 and 0.1 mm depths of cut.
The data exhibited significant scatter. Generally data points were within +/−5˚C of a best fit curve but isolated points, i.e. at 0.28 mm feed/rev, 0.8 mm depth and 1.8 and 2.1 m/s were found to be as much as 16˚C from the curve. The reason for the scatter was attributed mainly to inconsistent swarf clearance. In these experiments a chip breaker was not used and the continuous stream of swarf varied somewhat randomly in path and direction. Occasionally swarf would wrap around the work piece or tool and produce obvious fluctuations in temperature, these experiments are not included in the results. Small temperature measurement errors (estimated at less than 2˚C) also occurred as in practice the tool temperature never reached a perfectly stable value. The bulk temperature of the tool and tool holder continued to rise after the initial rapid transient increase causing the recorded tool tip temperature to continue to increase slowly. The effect of this was minimized by waiting for the rate of increase of tool temperature rate to be less than 1˚C per minute.
Experiments were performed with 4 different work pieces and two tools. Selected experiments, i.e. 0.28 mm feed/rev, 2.3 m/s and 0.8mm and 0.2 mm depth were repeated with the second tool on different work pieces to determine whether tool wear or variation in material properties affected temperature measurements. Within the limits of experimental scatter, agreement with the originnal data suggested neither tool wear nor work piece material was a significant factor in these experiments.
Figures 4-6 clearly show the relationship between tool temperature and cutting speed to be nonlinear, the degree of non linearity increasing with depth of cut. For example at a feed/rev of 0.07 mm and a depth of 1.5 mm tool temperature increases rapidly as speed is increased but reaches a maximum at 1.2 m/s. The same trend is evident at lower depths of cut but a maximum temperature is not reached within the speed range examined.
Figures 7 and 8 are extracted from Figures 4 to 6 and show the dependence of tool temperature on depth of cut and feed rate respectively. The 3 curves in Figure 7 are for cutting speeds of 0.5, 1.0 and 1.5 m/s whilst the 3 curves in Figure 8 are for cut depths of 0.1, 0.2 and 0.4
Figure 7. The effect of cut depth on tool temperature for cutting speeds of 1.5, 1.0 and 0.5 m/s.
Figure 8. Effect of feed rate on tool temperature for 0.4, 0.2 and 0.1 mm depths of cut.
mm. Both figures show that the relationship between tool temperature and depth of cut or feed rate to be non linear.
3. Optimization of ANN Design
The experimental data in Figures 4-6 was used as a basis for determining an optimum ANN design. The best fit curves to the data shows clearly defined non linear relationships between tool temperature and the machining parameters cutting speed, cut depth and feed rate. But the experimental data points show that superimposed upon the underlying trends there are small but significant random temperature measurement errors. This is broadly attributed to experimental scatter and is a characteristic of machining data. As part of a tool temperature control system, the ANN is required to provide a relationship between tool temperature and machining parameters against which real time tool temperature measurements can be compared and decisions made about the condition of the machining process and what adjustments may be required, if any. Thus, in this application the ANN is required to predict the underlying trends shown in Figures 4-6 and filter out the random temperature measurement errors.
3.1. Application to Data Containing Experimental Scatter
The design algorithm described in [15] was used to optimize the ANN design for this application. A network with a single hidden layer and a TanH activation function was used as a starting point for the work. For the 96 experimental data points of Figures 4-6 the effect of the number of neurons in the hidden layer upon convergence accuracy was determined by systematically varying the number of neurons and finding the minimum convergence error within 100,000 iterations. This was repeated 4 times to assess the repeatability of the convergence process. The results are shown in Figure 9, the curve is the average RMS convergence error and the black lines indicate maximum and minimum values. The convergence process exhibited considerable scatter and the minimum convergence error at 12.8% is considered high. Variation in the convergence process was attributed to the selection of random initial values with which to start the iteration process and the random selection of data points for testing and training. However the extent of the variation and the relatively high convergence error was attributed, at least in part, to experimental scatter within the data points used for training the network. The experimental data in Figures 4-6 show an average error of 7% relative to best fit curves and this must limit the accuracy that the convergence process can achieve.
Figure 10 shows an example of a comparison between ANN predictions and experimental data. This is for a network with 18 neurons in the hidden layer giving a convergence error of 18%. ANN predictions do show the underlying trends in the data but careful inspection of the