Analysis Based on Generalized Regression Neural Network to Oil ...
Analysis Based on Generalized Regression Neural Network to Oil Atomic Emission Spectrum Data of a Type Diesel Engine ChunHui Zhang, HongXiang Tian, and Tao Liu College of Naval Architecture and Power, Naval Univ. of Engineering, Wuhan Hubei 430033, China
Abstract. In order to deeply mine the information of Oil Atomic Emission Spectrum Data, a simulation model and a prediction model of Cu concentration of a type of six- cylinder diesel engine were established by applying Generalized Regression Neural Network. Seven different working conditions had been set up and sixty-nine oil samples had been taken from engine. The results show that the absolute errors of the simulation value of the 69 samples are within the acceptable accuracy indices and the absolute errors of the prediction value of the 19 samples are lower than the acceptable accuracy indices. It has been proved effective that Cu concentration can be predicted via Generalized Regression Neural Network algorithm. Keywords: Generalized Regression Neural Network (GRNN), Diesel Engine, Atomic Emission Spectrum, Wear elements.
1 Introduction Modern mechanical equipment and power plant commonly runs oil film of several microns, lubricating oil carries abundant wear information. The analysis of physics & chemistry performance target, grain and concentration of elements can reveal wear conditions, the oil quality decay and contamination of machine[1]. Oil monitor technique is a technique of obtaining information of lubricant and wear, predicting faults and ascertaining reasons, types and parts of faults through analyzing performance change of lubricant and wear debris carried of checked equipment[2]. There are many common methods of deeply mining information of oil atomic emission spectrum data, including the grey system theory[3], the Factor Analysis[4], the maximum entropy principle analysis[5], the correlation coefficient analysis[6], the regression analysis[7] and the support vector machines[8], etc. This paper aims at setting up a relation between the concentration of wearing elements of diesel engine and it’s loads, cylinders’ clearances and run-time oil renewed by applying Generalized Regression Neural Network (GRNN) to dealing with the oil atomic emission spectrum data of a 6-cylinder diesel engine. S. Lin and X. Huang (Eds.): CSEE 2011, Part I, CCIS 214, pp. 574–580, 2011. © Springer-Verlag Berlin Heidelberg 2011
Analysis Based on GRNN to Oil Atomic Emission Spectrum Data
575
2 General Regression Neural Network Theory General Regression Neural Network was developed by Donald F. Specht in 1991, called GRNN for short[9]. GRNN have a strong ability of the nonlinear mapping, flexible network structure and a high degree of fault tolerance, be employed for solving nonlinear problems. GRNN have a higher advantage in approximation of function and learning speed than RBF and be convergence to optimization regression plane finally. What’s more, GRNN can deal with unstable date too. The GRNN topology consists of input layer, hidden layer (radial basis layer) and output layer, the internal structure of a GRNN is depicted in figure1[10].
Output layer
Hidden layer
Input layer
W 1 S1 × R
x1 x2
W2
dist n1
…
S1 × R
P
P
P
S1 × 1
P
xR
f1 ( x ) P
A1 S1 × 1
P
b1
Nprod P
S 2 × S1
P
n2 S 2 ×1
f 2 (x )
A2 P
S 2 ×1
P
P
P
P
S1 × 1 P
Fig. 1. Internal structure of GRNN
As show in the figure 1: X means input vector, R means dimension of input vector, S1 , S 2 means the numbers of neuron of each network. The number of neuron hidden layer equals to number of the training samples. The weight functions of hidden layer is Euclidean distance which calculates the distance from the input vector of length R to each of the Q training vectors of length R, expressed by
di st . b1 means
threshold value of the hidden layer. In general, the delivery function of the hidden layer is Gaussian function, the equation is as following:
⎛ 1 R i ( x ) = exp ⎜⎜ x − di 2 ⎝ − 2δ i
2
⎞ ⎟⎟ ⎠
(1)
x is the input sample; d i is center of the hidden layer node; δ i is smooth factor which decided basis function shape of the location of i hidden In the equation (1),
1
layer. The output of hidden layer A is as following. The output layer is a pure linear layer, the weight function of it is a normalization dot product the weight function, expressed by npord , which calculate
576
C.H. Zhang, H.X. Tian, and T. Liu
A
the 1
A
1
= exp
⎡ ⎢− ⎢ ⎣
( di
st b i1 2δ
)
2
2
⎤ ⎥ ⎥ ⎦
(2)
n2 of network. The elements of n2 are quotients that dot products of which vector and each row element of weight matrix
1
W 2 and the sum of elements of vector A .
( )
The value of output of network A = purelin n2 . The trait that the learning of network depends on the samples of dates decides to extremely avoid effects of results from one’s supposition. 2
3 Date Analysis and Discussion 3.1 Experimental Date and Instrument The SPECTROIL M spectrum instrument used in experiments is a special emission spectrometer designed for analyzing lubricating oil elements. The instrument has higher precision and well repeatability and can analysis 21 elements once. This emission spectrometer is calibrated by complete normalization and daily normalization before experiments. Matlab2010a is employed in analysis Emission Spectrum Data. Emission spectrum data analyzed in this paper comes from a reference[11]. Diesel engine has seven different operating conditions arranged by updating pistons of diesel engine and altering the clearance between cylinders and new pistons. 69 oil samples taken from the engine with different loads and operating time under different operating conditions has been analyzed by the SPECTROIL M spectrum instrument. The number, the clearance Between cylinder-piston and the code of cylinder-piston is shown in Table 1 [12]: Table 1. Seven Different Conditions of Diesel Engine No. 1 2 3 4 5 6 7
Clearance Between Cylinder-piston 0.7mm between the second cylinder-piston, 0.6mm between the fifth cylinder-piston, 0.87mm between the sixth cylinder-piston and other cylinder-pistons are normal. 0.7mm between the second cylinder-piston, 0.87mm between the sixth cylinder-piston and other cylinder-pistons are normal. 0.7mm between the second cylinder-piston and other cylinder-pistons are normal. 0.87mm between the sixth cylinder-piston and other cylinder-pistons are normal. 0.7mm between the second cylinder-piston, 0.6mm between the fifth cylinder-piston and other cylinder-pistons are normal. 0.6mm between the fifth cylinder-piston and other cylinder-pistons are normal. Normal clearance between every cylinder-piston.
Code of Cylinder-piston 0000001
0000010 0000100 0001000 0010000 0100000 1000000
Analysis Based on GRNN to Oil Atomic Emission Spectrum Data
577
3.2 Establishing the GRNN Model According to a trait of GRNN, numbers of neuron of input layer equals to dimension of input vector, numbers of neuron of output layer equals to dimension of output vector. There are 21 elements in the oil atomic emission spectrum data. Among 21 elements, Fe, Cr, Pb, Cu and Al are related with wear of diesel engine; Na, Ca, Mg, Ba ,P, B and Zn are related with additive of lubricating oil; Mg, B, Na, Si, Ca and P are related with exotic contamination; elements of which concentration are under sensitivity of instrument are disturbing elements, such as Ag, Ti and V. component elements of lubricating oil are not commonly analyzed in experiment, such as C,H. concentration of Cu which can reflect adequately wear condition of engine is chosen as output dates of neural network, which means the node of output layer being 1. Input vector is represented by matrix composed of the Clearance between Cylinder-piston and the Code of Cylinder-piston and operating time, the node of input layer is 9. Then, a GRNN model is established by a sentence of MATLAB newgrnn(P, T , SPREAD ) .Input vector is represented by P , output vector is represented by T , the spread rate of radial basis function is represented by SPREAD . The delivery function of the hidden layer and output layer are respectively Gaussian function and pure linear function, so parameter of artificially factors of GRNN is an only SPREAD . A great deal of experiments show that approximating process from network to sample date will be smoothly and the prediction error of sample date will be minor when SPREAD equals to 10. Concentration of Cu is simulated by trained GRNN. The output of simulated concentration of Cu is show in figure 2:
Fig. 2. The predicting result of Cu by GRNN method
578
C.H. Zhang, H.X. Tian, and T. Liu
3.3 Error Analysis Figure 2 shows the most of samples are simulated successfully. The SPECTROIL M spectrum instrument has prescribed the acceptable accuracy indices of Cu when the standard densities of Cu are 0, 5, 10, 30, 50, 100 and 300ppm, as shown in Table 2. Cubic function of MATLAB interp1(x, y, x i , ' cubic ') is employed for estimating accuracy indices of Cu of 69 samples. Standard concentration accuracy Indices of Table 2 are used as input vector x , y among the function. Table 2. Acceptable Accuracy Indices for Wear Metals-mean of Cu Standard concentration /ppm
0
5
10
30
50
100
300
Accuracy Indices/ppm
0.92
1.61
2.44
5.91
9.43
18.2
53.5
The absolute errors of the simulation value of Cu concentration have been obtained by comparing the simulation value with the observation value. Comparing the absolute errors with the acceptable accuracy indices obtained from cubic function can reveal the stimulating result. The result shows that all the 69 samples’ absolute errors are lower than the acceptable accuracy indices, as shown in Table 3. Table 3. Simulation Efficiency of Cu Concentration
No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Absolute Accuracy Error(ppm) Indices(ppm) -0.01 1.91 -0.08 1.9 0.36 1.83 0.48 1.9 0.1 1.96 0.54 1.91 -0.43 2.08 0.5 1.95 0.31 1.98 0.41 1.95 -0.28 2.06 0.22 1.98 -0.17 2.05 0.19 2.01 -0.4 2.12 -0.79 2.18 0.05 2.03 -0.23 2.08 0.15 2.01 -0.63 2.15 0.18 2.01 -0.15 2.12 -0.13 2.12
Overrun (Y/N) N N N N N N N N N N N N N N N N N N N N N N N
No. 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
Absolute Error(ppm) 0.05 0.06 0.04 0.04 0.06 0.01 0 0 0.03 0.34 0.26 0.03 0.05 0 0.03 0.02 0.06 0.05 0.1 0.09 0.02 0.02 0.02
Accuracy Indices(ppm) 1.4 1.4 1.42 1.42 1.43 1.43 1.39 1.43 1.43 1.43 1.69 1.66 1.66 1.51 1.52 1.54 1.55 1.51 1.11 1.14 1.18 1.18 1.19
Overrun (Y/N) N N N N N N N N N N N N N N N N N N N N N N N
Analysis Based on GRNN to Oil Atomic Emission Spectrum Data
579
Table 3. (continued) 24 25 26 27 28 29 30 31 32 33 34 35
0 0 -0.04 -0.05 -0.05 0.25 -0.14 -0.04 0.07 -0.04 0.01 -0.05
1.77 1.73 1.75 1.75 1.75 1.7 1.81 1.8 1.78 1.8 1.36 1.42
N N N N N N N N N N N N
59 60 61 62 63 64 65 66 67 68 69
0.02 0.02 0.07 0.13 0.22 0.13 0.06 0.15 0.14 0 0
1.21 1.21 1.22 1.25 1.03 1.05 1.07 1.09 1.09 1.09 1.09
N N N N N N N N N N N
3.4 Predicting and Analysis According to constantly changing parameter of operating condition (including running time and load) and machine parameter (clearance between cylinder-piston), an appropriate model is set up in terms of GRNN algorithm for predicting Cu concentration. 50 random samples were taken as the training sets. The remaining 19 samples have been predicted, results of prediction have been shown in Table 4: Table 4. Predicting Efficiency of Cu Concentration
No. 11 12 13 14 15 28 29 30 38 39 40 46 47 53 59 60 61 68 69
Observation value(ppm) 7.8 7.3 7.7 7.5 8.1 5.9 5.6 6.3 3.7 3.7 3.8 5.5 5.3 4.3 2.2 2.2 2.3 1.3 1.3
Prediction value(ppm) 7.29 7.3 7.31 7.65 7.66 5.95 5.95 6.18 3.62 3.7 3.7 5.02 5.29 4.6 2.28 2.28 2.5 1.17 1.18
Absolute Error(ppm) -0.51 0 -0.39 0.15 -0.44 0.05 0.35 -0.12 -0.08 0 -0.1 -0.48 -0.01 0.3 0.08 0.08 0.2 -0.13 -0.12
Accuracy Indices(ppm) 2.06 1.98 2.05 2.01 2.12 1.75 1.7 1.81 1.42 1.42 1.43 1.69 1.66 1.51 1.21 1.21 1.22 1.09 1.09
Overrun (Y/N) N N N N N N N N N N N N N N N N N N N
The data shown in Table 4 suggests that 19 samples’ absolute errors are lower than the acceptable accuracy indices.
580
C.H. Zhang, H.X. Tian, and T. Liu
4 Conclusions Based on the Generalized Regression Neural Network algorithm, the simulation model has been set up for the relation between Cu concentration of lubricating oil of diesel engine and its loads, cylinders’ clearances and running time oil renewed. The simulation model has proved to be effective, for the absolute errors of the simulation value of the 69 samples are lower than the acceptable accuracy indices. As regards to seven working conditions, the Cu concentrations of the random 19 samples were predicted accurately through the predicting model based on the Generalized Regression Neural Network analysis.
References 1. Toms, L.A., Toms, A.M.: Machinery Oil Analysis. Society of Tribologists & Lubrication Engineers, 1–20 (2008) 2. Yan, X.: Development and think Of oil monitor technique. Lubrication Engineering 14(7), 6–8 (1999) 3. Wang, L., Li, L.: The Sampling Time Prediction of Oil Analysis for Power-shift Steering Transmission. Lubrication Engineering 35(8), 84–87 (2010) 4. Liu, T., Tian, H.X., Guo, W.Y.: Application of Factor Analysis to a Type Diesel Engine SOA. In: 2010 International Conference on Measuring Technology and Mechatronics Automation, Changsha, China, March 13-14, pp. 612–615 (2010) 5. Huo, H., Li, Z., Xia, Y.: Application of maximum entropy probability concentration estimationapproach to constituting oil monitoring diagnostic criterions. Tribology International 39, 528–532 (2006) 6. Tian, H.X., Ming, T.F., Liu, Y.: Comparison among oil SPECTRAL date of six types of marine engine. In: International Conference on Transportation Engineering, Chengdu, pp. 43–48 (2009) 7. Zhou, P., Liu, D.F., Shi, X., Li, G.: Threshold Setting of Oil Spectral Analysis Based on Robust Regression. Lubrication Engineering 35(5), 85–88 (2010) 8. Fan, H.B., Zhang, Y.T., Ren, G.Q., Luo, H.F.: Study on Prediction Model of Oil Spectrum Based on Support Vector Machines. Lubrication Engineering 183(11), 148–150 (2006) 9. Shi, F., Wang, X.C., Yu, L.: Analysis of 30 cases of MATLAB neural network, pp. 73–80. Press of Beijing University of Aeronautics and Astronautics, Beijing (2010) 10. Yan, W.J.: Research on Discrimination of Tongue Diseases with Near Infrared Spectroscopy. Infrared Technology 32(8), 487–490 (2010) 11. Liu, T.: Study in the Field of Oil Atomic Emission Spectrum Data Mining, pp. 18–20. Naval University of Engineering, Wuhan (2009) 12. Liu, T., Tian, H.X., Guo, W.Y.: Application of Factor Analysis to a Type Diesel Engine SOA. In: 2010 International Conference on Measuring Technology and Mechatronics Automation, Changsha, China, March 13-14, pp. 612–615 (2010)
Abstract. In order to deeply mine the information of Oil Atomic Emission Spectrum Data, a simulation model and a prediction model of Cu concentration of a type of six- cylinder diesel engine were established by applying Generalized Regression Neural Network. Seven different working conditions had been set up and sixty-nine oil samples had been taken from engine. The results show that the absolute errors of the simulation value of the 69 samples are within the acceptable accuracy indices and the absolute errors of the prediction value of the 19 samples are lower than the acceptable accuracy indices. It has been proved effective that Cu concentration can be predicted via Generalized Regression Neural Network algorithm. Keywords: Generalized Regression Neural Network (GRNN), Diesel Engine, Atomic Emission Spectrum, Wear elements.
1 Introduction Modern mechanical equipment and power plant commonly runs oil film of several microns, lubricating oil carries abundant wear information. The analysis of physics & chemistry performance target, grain and concentration of elements can reveal wear conditions, the oil quality decay and contamination of machine[1]. Oil monitor technique is a technique of obtaining information of lubricant and wear, predicting faults and ascertaining reasons, types and parts of faults through analyzing performance change of lubricant and wear debris carried of checked equipment[2]. There are many common methods of deeply mining information of oil atomic emission spectrum data, including the grey system theory[3], the Factor Analysis[4], the maximum entropy principle analysis[5], the correlation coefficient analysis[6], the regression analysis[7] and the support vector machines[8], etc. This paper aims at setting up a relation between the concentration of wearing elements of diesel engine and it’s loads, cylinders’ clearances and run-time oil renewed by applying Generalized Regression Neural Network (GRNN) to dealing with the oil atomic emission spectrum data of a 6-cylinder diesel engine. S. Lin and X. Huang (Eds.): CSEE 2011, Part I, CCIS 214, pp. 574–580, 2011. © Springer-Verlag Berlin Heidelberg 2011
Analysis Based on GRNN to Oil Atomic Emission Spectrum Data
575
2 General Regression Neural Network Theory General Regression Neural Network was developed by Donald F. Specht in 1991, called GRNN for short[9]. GRNN have a strong ability of the nonlinear mapping, flexible network structure and a high degree of fault tolerance, be employed for solving nonlinear problems. GRNN have a higher advantage in approximation of function and learning speed than RBF and be convergence to optimization regression plane finally. What’s more, GRNN can deal with unstable date too. The GRNN topology consists of input layer, hidden layer (radial basis layer) and output layer, the internal structure of a GRNN is depicted in figure1[10].
Output layer
Hidden layer
Input layer
W 1 S1 × R
x1 x2
W2
dist n1
…
S1 × R
P
P
P
S1 × 1
P
xR
f1 ( x ) P
A1 S1 × 1
P
b1
Nprod P
S 2 × S1
P
n2 S 2 ×1
f 2 (x )
A2 P
S 2 ×1
P
P
P
P
S1 × 1 P
Fig. 1. Internal structure of GRNN
As show in the figure 1: X means input vector, R means dimension of input vector, S1 , S 2 means the numbers of neuron of each network. The number of neuron hidden layer equals to number of the training samples. The weight functions of hidden layer is Euclidean distance which calculates the distance from the input vector of length R to each of the Q training vectors of length R, expressed by
di st . b1 means
threshold value of the hidden layer. In general, the delivery function of the hidden layer is Gaussian function, the equation is as following:
⎛ 1 R i ( x ) = exp ⎜⎜ x − di 2 ⎝ − 2δ i
2
⎞ ⎟⎟ ⎠
(1)
x is the input sample; d i is center of the hidden layer node; δ i is smooth factor which decided basis function shape of the location of i hidden In the equation (1),
1
layer. The output of hidden layer A is as following. The output layer is a pure linear layer, the weight function of it is a normalization dot product the weight function, expressed by npord , which calculate
576
C.H. Zhang, H.X. Tian, and T. Liu
A
the 1
A
1
= exp
⎡ ⎢− ⎢ ⎣
( di
st b i1 2δ
)
2
2
⎤ ⎥ ⎥ ⎦
(2)
n2 of network. The elements of n2 are quotients that dot products of which vector and each row element of weight matrix
1
W 2 and the sum of elements of vector A .
( )
The value of output of network A = purelin n2 . The trait that the learning of network depends on the samples of dates decides to extremely avoid effects of results from one’s supposition. 2
3 Date Analysis and Discussion 3.1 Experimental Date and Instrument The SPECTROIL M spectrum instrument used in experiments is a special emission spectrometer designed for analyzing lubricating oil elements. The instrument has higher precision and well repeatability and can analysis 21 elements once. This emission spectrometer is calibrated by complete normalization and daily normalization before experiments. Matlab2010a is employed in analysis Emission Spectrum Data. Emission spectrum data analyzed in this paper comes from a reference[11]. Diesel engine has seven different operating conditions arranged by updating pistons of diesel engine and altering the clearance between cylinders and new pistons. 69 oil samples taken from the engine with different loads and operating time under different operating conditions has been analyzed by the SPECTROIL M spectrum instrument. The number, the clearance Between cylinder-piston and the code of cylinder-piston is shown in Table 1 [12]: Table 1. Seven Different Conditions of Diesel Engine No. 1 2 3 4 5 6 7
Clearance Between Cylinder-piston 0.7mm between the second cylinder-piston, 0.6mm between the fifth cylinder-piston, 0.87mm between the sixth cylinder-piston and other cylinder-pistons are normal. 0.7mm between the second cylinder-piston, 0.87mm between the sixth cylinder-piston and other cylinder-pistons are normal. 0.7mm between the second cylinder-piston and other cylinder-pistons are normal. 0.87mm between the sixth cylinder-piston and other cylinder-pistons are normal. 0.7mm between the second cylinder-piston, 0.6mm between the fifth cylinder-piston and other cylinder-pistons are normal. 0.6mm between the fifth cylinder-piston and other cylinder-pistons are normal. Normal clearance between every cylinder-piston.
Code of Cylinder-piston 0000001
0000010 0000100 0001000 0010000 0100000 1000000
Analysis Based on GRNN to Oil Atomic Emission Spectrum Data
577
3.2 Establishing the GRNN Model According to a trait of GRNN, numbers of neuron of input layer equals to dimension of input vector, numbers of neuron of output layer equals to dimension of output vector. There are 21 elements in the oil atomic emission spectrum data. Among 21 elements, Fe, Cr, Pb, Cu and Al are related with wear of diesel engine; Na, Ca, Mg, Ba ,P, B and Zn are related with additive of lubricating oil; Mg, B, Na, Si, Ca and P are related with exotic contamination; elements of which concentration are under sensitivity of instrument are disturbing elements, such as Ag, Ti and V. component elements of lubricating oil are not commonly analyzed in experiment, such as C,H. concentration of Cu which can reflect adequately wear condition of engine is chosen as output dates of neural network, which means the node of output layer being 1. Input vector is represented by matrix composed of the Clearance between Cylinder-piston and the Code of Cylinder-piston and operating time, the node of input layer is 9. Then, a GRNN model is established by a sentence of MATLAB newgrnn(P, T , SPREAD ) .Input vector is represented by P , output vector is represented by T , the spread rate of radial basis function is represented by SPREAD . The delivery function of the hidden layer and output layer are respectively Gaussian function and pure linear function, so parameter of artificially factors of GRNN is an only SPREAD . A great deal of experiments show that approximating process from network to sample date will be smoothly and the prediction error of sample date will be minor when SPREAD equals to 10. Concentration of Cu is simulated by trained GRNN. The output of simulated concentration of Cu is show in figure 2:
Fig. 2. The predicting result of Cu by GRNN method
578
C.H. Zhang, H.X. Tian, and T. Liu
3.3 Error Analysis Figure 2 shows the most of samples are simulated successfully. The SPECTROIL M spectrum instrument has prescribed the acceptable accuracy indices of Cu when the standard densities of Cu are 0, 5, 10, 30, 50, 100 and 300ppm, as shown in Table 2. Cubic function of MATLAB interp1(x, y, x i , ' cubic ') is employed for estimating accuracy indices of Cu of 69 samples. Standard concentration accuracy Indices of Table 2 are used as input vector x , y among the function. Table 2. Acceptable Accuracy Indices for Wear Metals-mean of Cu Standard concentration /ppm
0
5
10
30
50
100
300
Accuracy Indices/ppm
0.92
1.61
2.44
5.91
9.43
18.2
53.5
The absolute errors of the simulation value of Cu concentration have been obtained by comparing the simulation value with the observation value. Comparing the absolute errors with the acceptable accuracy indices obtained from cubic function can reveal the stimulating result. The result shows that all the 69 samples’ absolute errors are lower than the acceptable accuracy indices, as shown in Table 3. Table 3. Simulation Efficiency of Cu Concentration
No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Absolute Accuracy Error(ppm) Indices(ppm) -0.01 1.91 -0.08 1.9 0.36 1.83 0.48 1.9 0.1 1.96 0.54 1.91 -0.43 2.08 0.5 1.95 0.31 1.98 0.41 1.95 -0.28 2.06 0.22 1.98 -0.17 2.05 0.19 2.01 -0.4 2.12 -0.79 2.18 0.05 2.03 -0.23 2.08 0.15 2.01 -0.63 2.15 0.18 2.01 -0.15 2.12 -0.13 2.12
Overrun (Y/N) N N N N N N N N N N N N N N N N N N N N N N N
No. 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
Absolute Error(ppm) 0.05 0.06 0.04 0.04 0.06 0.01 0 0 0.03 0.34 0.26 0.03 0.05 0 0.03 0.02 0.06 0.05 0.1 0.09 0.02 0.02 0.02
Accuracy Indices(ppm) 1.4 1.4 1.42 1.42 1.43 1.43 1.39 1.43 1.43 1.43 1.69 1.66 1.66 1.51 1.52 1.54 1.55 1.51 1.11 1.14 1.18 1.18 1.19
Overrun (Y/N) N N N N N N N N N N N N N N N N N N N N N N N
Analysis Based on GRNN to Oil Atomic Emission Spectrum Data
579
Table 3. (continued) 24 25 26 27 28 29 30 31 32 33 34 35
0 0 -0.04 -0.05 -0.05 0.25 -0.14 -0.04 0.07 -0.04 0.01 -0.05
1.77 1.73 1.75 1.75 1.75 1.7 1.81 1.8 1.78 1.8 1.36 1.42
N N N N N N N N N N N N
59 60 61 62 63 64 65 66 67 68 69
0.02 0.02 0.07 0.13 0.22 0.13 0.06 0.15 0.14 0 0
1.21 1.21 1.22 1.25 1.03 1.05 1.07 1.09 1.09 1.09 1.09
N N N N N N N N N N N
3.4 Predicting and Analysis According to constantly changing parameter of operating condition (including running time and load) and machine parameter (clearance between cylinder-piston), an appropriate model is set up in terms of GRNN algorithm for predicting Cu concentration. 50 random samples were taken as the training sets. The remaining 19 samples have been predicted, results of prediction have been shown in Table 4: Table 4. Predicting Efficiency of Cu Concentration
No. 11 12 13 14 15 28 29 30 38 39 40 46 47 53 59 60 61 68 69
Observation value(ppm) 7.8 7.3 7.7 7.5 8.1 5.9 5.6 6.3 3.7 3.7 3.8 5.5 5.3 4.3 2.2 2.2 2.3 1.3 1.3
Prediction value(ppm) 7.29 7.3 7.31 7.65 7.66 5.95 5.95 6.18 3.62 3.7 3.7 5.02 5.29 4.6 2.28 2.28 2.5 1.17 1.18
Absolute Error(ppm) -0.51 0 -0.39 0.15 -0.44 0.05 0.35 -0.12 -0.08 0 -0.1 -0.48 -0.01 0.3 0.08 0.08 0.2 -0.13 -0.12
Accuracy Indices(ppm) 2.06 1.98 2.05 2.01 2.12 1.75 1.7 1.81 1.42 1.42 1.43 1.69 1.66 1.51 1.21 1.21 1.22 1.09 1.09
Overrun (Y/N) N N N N N N N N N N N N N N N N N N N
The data shown in Table 4 suggests that 19 samples’ absolute errors are lower than the acceptable accuracy indices.
580
C.H. Zhang, H.X. Tian, and T. Liu
4 Conclusions Based on the Generalized Regression Neural Network algorithm, the simulation model has been set up for the relation between Cu concentration of lubricating oil of diesel engine and its loads, cylinders’ clearances and running time oil renewed. The simulation model has proved to be effective, for the absolute errors of the simulation value of the 69 samples are lower than the acceptable accuracy indices. As regards to seven working conditions, the Cu concentrations of the random 19 samples were predicted accurately through the predicting model based on the Generalized Regression Neural Network analysis.
References 1. Toms, L.A., Toms, A.M.: Machinery Oil Analysis. Society of Tribologists & Lubrication Engineers, 1–20 (2008) 2. Yan, X.: Development and think Of oil monitor technique. Lubrication Engineering 14(7), 6–8 (1999) 3. Wang, L., Li, L.: The Sampling Time Prediction of Oil Analysis for Power-shift Steering Transmission. Lubrication Engineering 35(8), 84–87 (2010) 4. Liu, T., Tian, H.X., Guo, W.Y.: Application of Factor Analysis to a Type Diesel Engine SOA. In: 2010 International Conference on Measuring Technology and Mechatronics Automation, Changsha, China, March 13-14, pp. 612–615 (2010) 5. Huo, H., Li, Z., Xia, Y.: Application of maximum entropy probability concentration estimationapproach to constituting oil monitoring diagnostic criterions. Tribology International 39, 528–532 (2006) 6. Tian, H.X., Ming, T.F., Liu, Y.: Comparison among oil SPECTRAL date of six types of marine engine. In: International Conference on Transportation Engineering, Chengdu, pp. 43–48 (2009) 7. Zhou, P., Liu, D.F., Shi, X., Li, G.: Threshold Setting of Oil Spectral Analysis Based on Robust Regression. Lubrication Engineering 35(5), 85–88 (2010) 8. Fan, H.B., Zhang, Y.T., Ren, G.Q., Luo, H.F.: Study on Prediction Model of Oil Spectrum Based on Support Vector Machines. Lubrication Engineering 183(11), 148–150 (2006) 9. Shi, F., Wang, X.C., Yu, L.: Analysis of 30 cases of MATLAB neural network, pp. 73–80. Press of Beijing University of Aeronautics and Astronautics, Beijing (2010) 10. Yan, W.J.: Research on Discrimination of Tongue Diseases with Near Infrared Spectroscopy. Infrared Technology 32(8), 487–490 (2010) 11. Liu, T.: Study in the Field of Oil Atomic Emission Spectrum Data Mining, pp. 18–20. Naval University of Engineering, Wuhan (2009) 12. Liu, T., Tian, H.X., Guo, W.Y.: Application of Factor Analysis to a Type Diesel Engine SOA. In: 2010 International Conference on Measuring Technology and Mechatronics Automation, Changsha, China, March 13-14, pp. 612–615 (2010)
