Journal of Intelligent Manufacturing (1997) 8, 235 ± 237
Predicting grinding burn using arti®cial neural networks H O N G X I N G L I U , * T A O C H E N and L I A N G S H E N G Q U Institute of Machinery Diagnostics and Cybernetics, Xi'an Jiaotong University, Xi'an 710049, P. R. China.
Received September 1996 and accepted January 1997
This paper introduces a method for predicting grinding burn using arti®cial neural networks (ANN). First, the way to model grinding burn via ANN is presented. Then, as an example, the prediction of grinding burn of ultra-strength steel 300 M via ANN is given. Very promising results were obtained. Keywords: Grinding burn, prediction, arti®cial neural networks (ANN), ultra-strength steel 300 M
1. Introduction
2. Modelling grinding burn using ANN
Many important industrial materials, such as bearing steels, ultra-strength steels and superalloys, often suer from burning in the grinding processes. Because grinding is often the last process of manufacturing, and a grinding-burnt machine part has very bad performance, we must solve two problems: how to identify a grinding-burnt machine part to prevent it from being assembled into a machine, and how to prevent burning from taking place in the grinding processes. Obviously, the latter seems more important and more necessary. In essence, in order to predict and prevent grinding burn, we must set up a mathematical model that relates grinding burn to selected grinding parameters. Because the factors resulting in grinding burn are numerous, and the burn mechanism is complex, traditional modelling methods are ineective. In particular, the methods are incapable of accounting for some qualitative factors that should not be neglected in modelling, such as grinding type, grinding wheel, and grinding ¯uid. There are dierences between grinding conditions in the factory and in the laboratory, and the initial model from limited laboratory data must be revised in factory practice. In other words, the modelling method must be adaptable. Arti®cial neural networks (ANN) provide a powerful method for setting up a complex non-linear model. The adaptability of ANN is ideally suited to the modelling of grinding burn. This paper will show that the qualitative factors that are often neglected in modelling can be taken into consideration via ANN.
We can choose a feed forward network to set up the grinding burn model (Fig. 1). The output layer is a single neuron to indicate burnt or non-burnt. Considering that the burn model is non-linear and very complex, we select two hidden layers in the ANN; the number of neurons in each hidden layer can be determined through tests. The number of neurons in the input layer is determined by the dimension of the input vector. The input vector, which is very dicult to determine, depends on the grinding parameters. None of the main parameters related to the burn must be neglected. However, the input vector should be convenient for modelling. With a workpiece of a given material to be ground, limited types of grinding wheel and grinding ¯uid are usually recommended. The grinding types ± plane grinding, external grinding and internal grinding ± are also limited. The worn condition of the grinding wheel can be measured by the accumulated grinding duration after a time of sharpening of the grinding wheel. Taking into consideration all the factors that result in burn, we determine the input vector as follows: 2 3 x
1 grinding type number 6 x
2 7 grinding wheel type number 6 7 6 x
3 7 grinding fluid type number 6 7 6 x
4 7 total grinding duration (min) 0 6 7
1 X 6 7 sÿ1 6 x
5 7 grinding wheel velocity Vs ;
m ÿ1 6 x
6 7 workpiece velocity Vw ;
m min 6 7 4 x
7 5 grinding depth ap ;
mm x
8 feed speed fa
mm rÿ1
mm=1st:
*Author to whom all correspondence should be addressed.
0956-5515 Ó 1997 Chapman & Hall
236
Fig.1. A feedforward network.
After determining the structure of the ANN, we can select the sigmoid function as the neuron feature function, and then take a large quantity of prepared samples to train the ANN. A sample consists of an input vector and the corresponding output vector, so its dimension is 9. Basically speaking, training an ANN is an optimization process. Some common optimization algorithms, such as the BP algorithm, the gene algorithm and the scaled conjugate gradient algorithm (SCGA), can be used. SCGA seems to be the best one amongst them. It is fast and of high accuracy. In this paper example, we used the SCGA technique. 3. A modelling and predicting example using ANN Ultra-strength steel 300 M, with high strength and good comprehensive performance, is regarded as the ideal material for aeroplane landing gears, but it suers easily from burn in grinding, which has greatly constrained its applications. So we chose the grinding burn of 300 M steel to model, and expected to make some progresses in solving this problem. We chose seven types of grinding wheel and three types of grinding ¯uid regarded as suitable for the steel, and did detailed burn tests within plane grinding, external grinding and internal grinding in our laboratory. The three types of grinding ± plane grinding, external grinding and internal grinding ± were numbered 0, 1, 2 respectively. The grinding wheels ± WA69KV, SA60KV, WA46KV, SA46KV, WA60JV, WA60LV (sparse) and CBN were numbered 0, 1, 2, 3, 4, 5, 6. The three types of grinding ¯uid ± HQ-1 waterbased grinding ¯uid, XGD-4 water-based grinding ¯uid and GF-2 grinding oil ± were numbered 1, 2, 3. Grinding without ¯uids was numbered as 0. It was now possible to express the ANN input vectors as in Equation 1. With nonburnt or burnt numbered as 0, 1, the output vector can be easily expressed. A sample consists of both an input vector and its corresponding output vector. A sample [0, 3, 1, 5, 19, 3, 14, 0.02, 8, 1] means plane grinding, SA46KV grinding wheel, HQ-1 grinding ¯uid, total grinding duration 5 min, grinding wheel velocity Vs 19.3 m s)1, workpiece
Liu, Chen and Qu velocity Vw 14 m min)1, grinding depth ap 0.02 mm, feed speed fa 8mm/1st, and resultant burn under these grinding conditions. Table 1 lists some of 87 samples from laboratory tests. (Acid etching was used to identify the presence or absence of grinding burn.) With the structure of the ANN determined as 8-9-5-1, and all samples normalized to values between 0.05 and 0.95 in every column, we trained the ANN with the samples by means of SCGA. After 497 iterations of training (about 2 min on a 486 computer) the sum of the squared ANN output errors for all samples was less than 0.0001, which indicated that the ANN had been trained well and the modelling using ANN was completed. Table 2 lists the predicted results under six sets of hypothetical grinding parameters using the above well-trained ANN. The ®rst three ANN outputs were all about 0.05, i.e. non-burnt. With the three input vectors compared with sample no. 78 in Table 1, we could see that these outputs were reasonable. The second three results were all about 0.95, which was also reasonable in comparison with sample no. 79 in Table 1. We also tested the ANN model with practical data from a factory. Table 3 lists the predicted results and actual reTable 1. Samples from testing No.
Samples
1 2 3 4 5 6
0 0 0 0 0 0
0 0 0 3 2 0
1 1 0 1 1 1
0 0 0 5 0 20
19 19 33 19.3 19.3 18
12 12 14 14 20 13
0.01 0.03 0.03 0.02 0.01 0.01
8 8 8 8 8 1
0 1 1 1 0 1
77 78 79 80 81 82 83 84 85 86 87
0 0 0 0 1 1 1 2 2 2 2
6 0 0 0 0 0 0 0 0 0 0
3 2 2 2 2 2 2 2 2 2 2
0 0 5 0 1 1 0 1 1 1 1
18 18 18 18 18 18 18 22 20.3 20.3 20.3
14 14 14 14 6.2 6.2 6.2 47.1 21.9 21.9 21.9
0.01 0.01 0.02 0.03 0.01 0.005 0.02 0.005 0.01 0.005 0.03
1 1 1 1 5.6 5.6 5.6 0.5 9 9 9
0 0 1 1 0 0 0 0 0 0 1
Table 2. Six sets of hypothetical grinding parameters and the predicted results via the ANN No.
Input vectors
1 2 3 4 5 6
0 0 0 0 0 0
0 0 0 0 0 0
2 2 2 2 2 2
Outputs 0 0 0 5 5 5
18 18 18 18 18 18
14 14 14 14 14 14
0.003 0.005 0.008 0.023 0.025 0.028
1 1 1 1 1 1
0.047 0.048 0.050 0.948 0.948 0.949
789 866 559 829 759 143
Predicting grinding burn
237
Table 3. Predicted and actual results for 300 M steel practical grinding data No. Input vectors 1 2 3 4 5 6 7 8 9 10 11 12
0 0 0 0 1 1 1 1 1 2 2 2
0 0 0 0 0 0 0 0 0 0 0 0
2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1
20 20 20 20 35 35 35 35 35 20.9 20.9 20.9
15 15 15 15 19.8 19.8 19.8 19.8 19.8 47 47 47
sults for 12 sets of practical data for 300 M steel . Among the 12 sets of data, only the ANN output no. 7 was incorrect. The prediction accuracy approached 92%, and this shows that the ANN model was feasible. With the practical data in Table 3 added into the training sets in Table 1, the ANN was retrained. We found that the new well-trained ANN could reach a higher accuracy in prediction. This indicated that modelling by means of ANN possesses good adaptability.
0.005 0.01 0.02 0.03 0.005 0.01 0.02 0.03 0.05 0.005 0.01 0.03
2 2 2 2 0.285 0.285 0.285 0.285 0.285 0.5 0.5 0.5
Outputs
Actual results
0.049 0.049 0.958 0.948 0.050 0.051 0.958 0.949 0.950 0.049 0.052 0.950
No-burn No-burn Burn Burn No-burn No-burn No-burn Burn Burn No-burn No-burn Burn
773 653 552 944 118 260 238 393 483 410 123 613
rately via ANN, and so it can be prevented eectively. Considering that ANN possesses a number of advantages over traditional methods, we believe that this method of predicting grinding burn will be widespread in practice. Acknowledgments Project supported by the National Science foundation of the People's Republic of China.
4. Conclusions ANN provides a powerful means for setting up a complex non-linear model of grinding burn. The qualitative factors often neglected in traditional modellings can be taken into consideration via ANN. Because of the adaptability of ANN, the ANN model of grinding burn can easily be revised in practice. Grinding burn can be predicted accu-
References Hongxing Liu (1994) The study on grinding burn of ultra-strength steel 300 M, Master's Thesis, Northwestern Polytechnical University, Xi'an, P.R. China. Zhenguo Cai (1993) The study on ANN based prediction, PhD Thesis, Xi'an Jiaotong University, Xi'an, P.R. China.
