PREDICTION OF WEAR AND FRICTION COEFFICIENT OF BRAKE PADS DEVELOPED FROM PALM KERNEL FIBRES USING ARTIFICIAL NEURAL NETWORK

Artificial neural network prediction of wear rate and friction coefficient of brake pads developed from palm kernel fibres (PKF) was carried out in this study. Major input parameters including materials formulation, manufacturing conditions, and operating conditions were introduced into the neural model while wear rate and friction coefficient were the outputs. The network architecture of LM12 [4-3]2 2 was selected for predicting the wear rate and coefficient of friction. The predicted wear rate and friction coefficient using ANN models were compared with the measured values using some statistical indicators. Results showed that the ANN predicted accurately the wear rate and friction coefficient of the developed automotive brake pads.


INTRODUCTION
Investigation and modeling of tribo-system (disc, brake pads) are key points for solving the friction and wear behavior of automotive brakes.Development of disc brake system is always a big challenge for car manufacturers and suppliers [1].The phenomenon of material transfer during sliding is important from both scientific and practical considerations.Wear performance are influenced by how wear particles act inside the control as a third body and how they are ejected from the contact [2].Friction and wear performance is two kinds of responses from one tribo-system.Wear in general relies on many factors such as temperature, applied load, sliding velocity, properties of mating materials, and durability of the transfer layer.The transfer film affects the friction and wear behavior of friction pairs [3].
Wear mechanisms operating in brake lining materials are found to be a combination of abrasion, adhesion, fatigue, delamination and thermal decomposition.That is why stochastic nature of wear of brake friction materials is a result of this wear mechanism and their transition from one combination to another.The dominant wear mechanism of brake friction materials is influenced by braking times, applied loads, and materials characteristics.Accordingly, prediction of wear performance, especially of friction materials, is a very complex task.A transfer film generating in the contact of friction pair further complicates it.The role of transfer film in poly-metal sliding contacts has long been realized as being responsible for a gradual transition from transient to steady state wear behavior [3].
Before now, many efforts have been made to develop wear/friction models and prediction of wear properties of different engineering materials.According to [4], Archard was one of the early researchers to develop a linear wear model for metals.In his model, the wear volume per sliding distance was in terms of the wear coefficient which was interpreted in various ways in the literature.On the other hand, [5] proposed a nonlinear wear model for friction materials in a disc brake assembly and his predicted wear compare well with the experimental data.[6] studied frictional and wear behavior at local contacts in an automotive brake system based on computer simulation by movable cellular automata method.[7], also, undertook a Finite Element (FE) model of a brake pad with particular respect to uncertainties.An inverse fuzzy arithmetical method was presented and applied to the modeling of a brake pad.Methodology for the modeling of the variability of brake linings was investigated by [1].
Modern and traditional techniques related to the modeling of friction and wear have not been applied to automotive brake pads developed from agricultural wastes.Also, the traditional mathematical techniques related to the modeling of friction and wear characteristics of brake pads were not able to provide the wear and friction models with inherent abilities to generalize the complex synergy of different influences of friction materials formulations, manufacturing, and brake operating conditions.Artificial neural networks (ANN) are ideally suited for simulating complex composite problems because it can be learned and trained to find solution [7].According to [9], artificial neural network can be applied to overcome the limitation of empirical formulas, and the advantage is that it is very useful for learning complex relationships between multi-dimensional data [3].
In the present study, an ANN approach in modeling the friction and wear characteristics of automotive brake pads developed from palm kernel fibres (PKF) using epoxy resin as a binder was employed.The artificial neural network prediction of friction coefficient and wear was based on functional approximation between (i) friction material formulation, (ii) its manufacturing condition, and (iii) operation condition as inputs against friction coefficient and wear rate as outputs.Results have shown that the developed ANN model predicted the wear rate and friction coefficient of the developed pads accurately.

Development of the brake pad
Palm kernel fibers (PKF) were collected and suspended in a solution of caustic soda (sodium hydroxide) for twenty-four hours to remove the remnant of red oil left after extraction.They were then washed with water to remove the caustic soda and sun-dried for one week.The dried PKF was the ground into powder using a Hammer mill.Thereafter, it was sieved using sieve size of less than or equal to100 µm aperture.The sieved palm kernel fibres (PKF) powder in varying percentages was added together with varied percentages of additives which included aluminum oxide, graphite, and calcium carbonate and epoxy-resin based on 176g weight of commercial brake pad.These combinations are presented in Table 1.The combinations were separately drymixed using a blender in order to achieve a homogeneous state ready for molding.The mixtures for each of the formulations were separately transferred to a mould designed for the purpose and pressed at 100 KN force for 2 minutes at room temperature.The produced brake pads were cured at 250 o C for 90 minutes after removal from the hydraulic press.The samples were then finished by polishing them using polisher-grinder with various grinding paper of various sizes to obtain the final products.The produced samples are presented in Figure 1.

Testing of the brake pads
An Inertia dynamotor (Figure 2) with specifications presented in Table 2 (located at Anambra State Motor Manufacturing Company (ANAMMCO) Nigeria) was used for the evaluation of the wear rate and friction coefficient of the developed brake pads.Each pair of developed brake pads was separately mounted on the brake assembly unit of the inertia dynamometer.The rotating speed of the driving motor was initially set at 5.56 m/s using variable speed drive with tachometer.Brake was applied after attaining the set speed with the measured values of wear rate, coefficient of friction through a computer connected to it.This procedure was repeated for speeds of 8.33 m/s, 11.11 m/s, 16.67 m/s, 22.22 m/s and 27.78 m/s to obtain the corresponding values of wear rate, coefficient of friction respectively.The wear rate and friction coefficient at 22.78 m/s (100 km) were selected for the artificial neural network modeling.

ANN model 2.3.1. Input data/output
Three groups of inputs data; materials formulation, manufacturing conditions; and brake operating condition captured were captured in the formulation of the neural model.Tables 3a and 3b shows details of the input parameters.Figure 3 shows the structure of the artificial neural network.The work done by brake application was calculated by integrating the braking power and stopping time using the relation in equation ( 1).The experimental wear rate and coefficient of friction of the developed brake pads were the outputs of the network.These are represented in Table 4.

Data pre-processing
The five (5) input parameters from the material formulation were presented to the network in percent by volume, while input parameters from manufacturing and operating conditions were scaled within the range of 0-1 using the relation given in equation (2).The output parameters (wear rate and coefficient of friction) were linearized using the relation in equation ( 3) [3].
( ) 1 (I I ) whereI Curr is current input value, I Max -maximum input value and I Min -minimum input value, O Curr is the current output, O Min is the minimum input value, O Max is the maximum output.4) was used between the input and the hidden layers while linear function (equation ( 5)) was used between the hidden and output layer.
where x is the value of weight used.

Neural network prediction
After series of repeated trials using the different neural network architectures, it was discovered that the LM 12 [4-3] 2 2 gave the best trade-off error and cost.It was, therefore, selected for the prediction.It also showed very good correlation for both the training and testing compared with the other architecture models.The training performance indicated values of correlation coefficients of 0.9967 at epoch 6 and overall performance R of 0.9965 for training, validation and testing.

Model validation
Three statistical criteria such as correlation coefficient (R), mean absolute error (MAE), and Nash-Scutcliffe efficiency (NSE) given by the equations ( 6), ( 7) and ( 8) respectively were used for the validation of the neural model [11][12].(E E) (P P) (E E) (P P) where: E is the sample of the experimental value, p is the sample of predicted value by an ANN model Ē and P are the mean value of E and P respectively, N is the number of samples.

RESULTS AND DISCUSSION
The predicted and experimental values of wear rate and coefficient of friction along with statistical analysis for validating the ANN model are presented in Tables 5 and 6 respectively.with values of R = 1.0000, 1.0000, 1.0000, 1.0000, MAE = 0.0000, 0.0000, 0.0000, 0.0000, and NSE = 1.0000, 1.0000, 1.0000, 1.0000.The values of the correlation coefficient compared better with the values of 0.99997, 0.99997 and 0.99998 by [10].The slight disparity in some of the predicted and experimented values could be attributed errors in data arising from miscalculation.
The statistical validation of ANN model for the prediction of coefficient of friction as presented in This was confirmed by the regression of the predicted and experimental wear rates as shown in Figure 6.There was a high correlation between the predicted and experimental wear rates with R 2 of 0.9990.This confirms that the ANN model was successful predicting wear rates.Similarly, the predicted coefficient of friction against work done by brake and interface temperature, (Figures 7 and 8) showed that almost all predicted coefficient of friction values by ANN model matched with the experimental values.
This was also confirmed by the regression of the predicted and experimental coefficient of friction shown in Figure 9.There was a high correlation between the predicted and experimental wear rates with R 2 of 0.9910.Thus confirming that the ANN model was successful in predicting coefficient of friction.

Fig. 3 .
Fig. 3. Neural network architecture showing influence of input parameters on wear and coefficient of friction.

Fig. 4 .
Fig. 4. Comparison of experimental and predicted wear rates against work done by brake.
Modeling of wear and coefficient of friction of produced brake pads from palm kernel fibers considering influence of relevant factors such as friction materials formulation, its manufacturing and brake operating conditions was successful.It was shown that the computer-based model provided by ANN predicted values of wear rates and coefficient of friction well compared to the experimental values as indicated by the statistical indicators used for validation of the model.

Table 1 .
Additives used for brake pads production from palm kernel fibre (PKF).

Table 3a .
The input parameters used for the model.

Table 5 .
Predicted and experimental wear rate.

Table 6 .
Predicted and experimental coefficient of friction.The values of statistical indicators R, MAE, and NSE for validation of the neural network model for prediction of wear rate as shown in Table5varied in the range of 0.9889 to 1.0000, 0.000 to 0.0367 and 0.9968 to 1.0000 respectively.The statistical indicators show that the prediction of the wear rate employing the ANN model was very satisfactorily with the values of, MAE, reasonably low, and R, NSE values reasonably high at almost all measured values.The predicted values of 3.62 mg/m, 3.64 mg/m, 4.0 mg/m and 3.98 mg/m were more accurate t Table 6 was observed to exhibit good agreement with the experimental values.The minimal values of MAE values, higher values of R and NSE were observed.The predicted values of coefficient of friction 0.19, 0.16, 0.33, exhibited the highest accuracy compared to other values, with lower MAE of 0.0000, 0.0000 and higher R values of 1.0000, 1.0000, 1.0000, and NSE of 1.0000, 1.0000, 1.0000.
Figures 4 -5 show the experimental and ANN predicted wear rate against work done by brake application and interface temperature.It was observed that almost all the values wear rate predicted by ANN model corresponded with the experimental values.