Prediction of Rattle Noise in Steering Gear System and Analysis of Contributing Factors through Neural Networks Using Design and

Manufacturing Execution Data

Hyeon-Cheol Jo 1 , Jae-Yong Seo 1 , Kyung-Hwan Park 1 1 Applied NVH Technology cell, Hyundai Mobis, Republic of Korea

ABSTRACT In a vehicle, rattle noise may occur when driving on bumpy roads due to clearance between parts of steering gear systems. The rattle noise is perceived by the drivers’ ears and it can be the cause of a repair campaign. In order to reduce rattle noise, the dimensions of clearance between parts are optimally designed. Also, in the development stage, general auto parts companies manage rattle noise below a certain level. However, it is impossible to measure noise for all products at the production stage due to cost and time issues. In this paper, the noise prediction model using the neural network model for all products based on design factors and the data of manufacturing execution system was introduced. In addition, the main factors contributing to rattle noise were analyzed in the process of creating the model. By referring to the main factors contributing to noise, it helps in the design of the steering gear to reduce the rattle noise value.

Keywords : Steering gear system, Rattle noise, Neural network

1. INTRODUCTION

The Steering system required to steer in a vehicle is divided into three main parts: the universal joint, the steering gear and the electric power steering system (1). Among them, the steering gear that changes the force in the rotational direction to the force in the linear direction causes a lot of noise problems. Typical noise types generated by steering gear are clanking noise and rattle noise. In general, these noises cause emotional quality degradation to customers who drive the vehicle, which incurs steady field claim costs to auto parts companies.

Therefore, auto parts companies have been studying the structure of the steering gear for a long time to reduce noise. (2) In addition, while maintaining the durability of each part, materials that are strong against noise were found and applied. Based on this research, the steering gear is designed in consideration of the vehicle layout, and then the design is verified. Engineers measure rattle noise and clanking noise and regulate it to a level that is not a problem for customers. (3)

In order to produce this developed steering gear, the correlation between the noise measured in the development process and the vibration value measured in the production line is carried out. However, it is very difficult to reproduce rattle noise on a production line. As shown in figure 1, when measuring rattle noise in the development stage, a silent exciter is used. It is a system that measures

1 clfclfdlsp@mobis.co.kr 2 jaeyong.seo@mobis.co.kr 3 khpark26@mobis.co.kr

the force applied to the steering gear in an actual vehicle and excites it, but considering the environment and productivity of the production line, it is very difficult to apply it to the line. In the end, the correlation result is only used to verify clanking noise. In a word, auto parts companies are mass-producing steering gears without verification on the production line for rattle noise.

Figure 1 – Rattle noise test equipment of steering gear system

The purpose of this study is to predict rattle noise without directly measuring rattle noise in a production line by making a neural network model using design factors and the data of manufacturing execution systems (MES). MES data are values such as dimensions, indentation load, and torsion measured in the production line. In addition, weights for factors affecting rattle noise were analyzed in the process of creating the prediction model.

2. Rattle Noise Measurement and Input Data

2.1 Rattle Noise Data Measurement was conducted on the entire type of steering gear being developed by our company. Rattle noise is generally caused by a strike between gear teeth when an impacted external force is repeatedly applied. (4) Noise generated by excitation with an impact external force lowers the repeatability of measurement, which generally causes a test error of 1.5% in the same sample. Therefore, in order to reflect these test errors, each sample was evaluated 5 times. In other words, 5 noise data per sample were created. The point to be noted here is that the noise values with the same design data and MES data as inputs are not assigned to the train set and the test set at the same time when learning. This is because, when some of the five data measured from the same sample are used as a train set and the rest are used as a test set, the error rate of the model comes out as low as the test error level.

And since two large peaks occur during excitation, the noise level consists of two peaks. Figure2 is a graph expressing rattle noise values by box plot by steering gear type. The same design applied steering gear also shows a maximum deviation of 19.2% and a minimum of 6.6% depending on the difference in MES data values.

ee |

Figure 2 – Box plot for Rattle Noise Data (Y- axis : Noise level)

2.2 Design Data

The weight, dimensions, inter-axial distance, number of teeth, etc. of each detailed part considered

PEAK

gear is largely composed of a rack housing, a rack bar, a rack bush, a pinion shaft, a yoke, and a rack bush.

Figure 3 – Structure of Steering Gear System

Typical design factors for rack housing parts are length between mountings, mounting angle, mounting height, and weight. In addition, the representative design elements of the rack bar are the outer diameter, length, rack lead angle, tooth height, and tooth center distance. And representative design elements of the yoke include side O-ring hardness, yoke liner type, yoke spring load, and rear O-ring hardness. Lastly, the main design factors of the pinion shaft include helical angle, tooth height, number of teeth, heat treatment depth, and weight. There are a total of 63 such design factors in 9 sectors, and the same value was used as input for all steering gears of the same type.

2.3 MES Data

The MES data section is largely divided into needle bearing assembly process, rack bar assembly process, pinion shaft assembly process, and yoke assembly process. In the needle bearing assembly process, the bearing pressing force and pressing depth are measured, and in the rack bar assembly process, the rack bar pressing force is measured. In the pinion shaft assembly process, the bearing press-in force, press-in depth, lock ring press-in force, press-in depth, caulking height, etc. are measured. Finally, in the yoke assembly process, the yoke clearance, yoke press-in depth, yoke press- in load, and rack bar torsional torque are measured. There are a total of 16 such MES factors in 6 sectors

3. Rattle Noise Prediction Model

3.1 Input Data Processing

MES data consists only of numerical data. And design data is divided into presence/absence data, type data, numeric data, and range data according data.

A simple processing procedure was performed to use this data as input data. First, in the case of numerical data, a standard scaler was applied. Second, the case of range data, both ends of the range were used as input. And in the case of presence/absence data, the input was set to a value of 1 in the case of presence and a value of 0 in the case of absence. Lastly, in the case of type data, when there are types of a, b, and c, 0, 1, 0 is input as input in case of b.

So the shape of the input data is (N, 88). If the MES data and design data are combined, there are a total of 79, but the input data is increased to 88 through processing. And since the output data consists of two peaks of rattle noise, the shape of the output data is (N, 2).

3.2 Creating the Rattle Noise Prediction Model

Machine learning was used to create a steering gear rattle noise prediction model based on the prepared data. The machine learning model used for predicting steering gear rattle noise is Linear Regression (LR), Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), and 1D Convolutional Neural Network (CNN). (5~8) As shown in Figure 4, MES data and design data are set as inputs to predict rattle noise, and gain values for variables are updated to increase the accuracy of the prediction model.

[Desien Prediction Model gy es —! Fadtor (Model Weigh) —ssiogeear— fatienowe FLD yD

Figure 4 – Diagram of the Steering Gear Rattle Noise Prediction

To select one of the machine learning models, Mean Absolute Percentage Error (MAPE) was used as an evaluation index of the model. MAPE can be defined as follows. (9)

y−y ̂ 𝑖

1 𝑛 ∑ |

𝑦 | 𝑛 𝑖=1 * 100 , (1)

This is because we thought that MAPE could represent the predicted level of noise most meaningfully. As mentioned in the introduction, datasets with the same input data were not divided into train set and test set in one learning cycle. And the predictive model performance was verified with 5-fold cross-validation. (10) As shown in figure 5, the best performance prediction model was created when using 1D CNN. The MAPE for the first peak level came out rather high. Referring to Figure 2, it is considered that the dispersion of the measured rattle noise is larger at the firs peak level.

Ocnn Gir Gor Gre Gsvm 50 49 30 20 10 oy ae ===

Figure 5– MAPE for Each Machine Learning Model (5-fold)(1 st Peak, 2 nd Peak)(Y-axis : MAPE)

And we made a learning model with the information of only the design data and only the information of the MES data. The reason for such learning was that MES data was not known at the development stage, so it would be helpful to have a model that can predict rattle noise only with design data.

Table 1 shows the average value of the performance prediction index of each model when verified by 5-fold. No matter which performance prediction indicator is used, the performance is good in the order of all data, MES data, and design data. Looking at the average value, a model trained only with MES data and Design data can be seen as a reliable model, but in reality the dispersion of 5 folds is larger

Ocnn Gir Gor Orr Gsvm 50 40 30 20 10 09 oa Be qnss

- All Data MES Data Design Data

- 1 st Peak 2 nd Peak 1 st Peak 2 nd Peak 1 st Peak 2 nd Peak

MAPE 0.88 0.80 1.43 1.32 2.45 2.19

MAE 0.47 0.43 0.76 0.70 1.30 1.17

MSE 0.38 0.31 1.17 0.99 2.77 2.32

RMSE 0.61 0.55 0.98 0.91 1.66 1.51

Table 1. Model Performance Results According to Input Data (5-fold Average.)

As shown in figure 6, when learning only with MES data, the dispersion of MAPE was large, and when learning only with design data, the dispersion was small but the MAPE was larger. This result shows that the best performance is achieved when both MES data and design data are input data.

Figure 6 – MAPE Box Plot for Each Machine Learning Model (5-fold) (1 st Peak, 2 nd Peak)

Additionally, a model based on 12 types of steering data among 13 types of steering gear data was created, and the prediction model was verified by putting the remaining one type of data in the created model. The reason for this analysis is that in reality, it is not easy to update the model whenever a new type of steering gear is released. So, through this analysis, it is possible to estimate how well it will predict a new type of steering gear when a model is created with data of an existing type of steering gear.

As a result, the MAPE verified by the model created except for one type was higher than the MAPE based on the entire data. And, as shown in Figure 7, the error rate of the type of steering gear with high dispersion of the measured data was higher. In order to predict rattle noise in actual mass production line, it was determined that it can be used after updating the model with data including the type of steering gear to be predicted.

leaial as

Figure 7 – The result of training with 12 types of steering gear data and

validating the model with the remaining data (1 st Peak, 2 nd Peak)

a 2 Pe |e ee ie a 2 -

3.3 Factor Analysis

For factor analysis, class activation map (CAM) was used. (11) In general, the structure of CNN is composed of Input – Convolution Layers – Fully Connected (FC) Layers. Instead of flattening the convolution layer with FC layer, new weights are created through Global Average Pooling (GAP). This weight was targeted to the last convolution layer that best describes the data structure and can be viewed as a probabilistic expression of the importance of the feature to the predicted value.

Because 5-fold validation was performed when creating the predictive model, we were able to obtain the weighting priority of 5 cases and calculated the average for 5 cases. As a result, as shown in figure 8, the weights were highest in the order of T-direction span 2, L-direction span 2, H-direction span, T-direction span 1, the weight of the rack bar and indentation depth of the pinion shaft bearing.

Figure 8 – Factor Weight of Steering Gear Rattle Noise

As shown in Figure 9, the T-direction span is the vertical distance from the rack bar axis to the mounting, and the L-direction span is the horizontal distance from the pinion center to the mounting rack bar. And the H-direction span is the vertical distance from the mounting floor to the axis of the rack bar. When considering the design factors related to the dimensions of the housing and the weight of the rack bar as the main factors, it can be assumed that the behavior of the housing against external force is the main factor in rattle noise. Although it is a different kind of external force, there is also a case of studying the relationship between housing behavior change and noise according to reversing steering. (12) Also, among the MES data, the pinion shaft bearing press fit depth was the most important factor. Since the rattle occurs around pinion shaft, it is estimated that the bearing press-in depth is the main factor.

Figure 9 – Illustration of Rack housing Span

Wintec

4. CONCLUSIONS

1) Among the steering gear prediction models, the prediction model using 1D CNN has an error

rate of about 1%, which is the level of test dispersion.

2) When learning only with MES data and design data, the error rate increases, and the error rate

dispersion is also large depending on the data set. The error rate distribution is proportional to the noise of steering gear measurement distribution for each type.

3) When verifying the untrained steering gear rattle noise, the error rate increases and is

proportional to the rattle noise distribution. When predicting rattle noise in the production line, it is reasonable to verify the learned steering gear type unless the performance of the existing model is improved.

4) The factors affecting the steering gear rattle noise are the dimensions of the rack housing and

the weight of the rack bar, and it is assumed that the behavior of the steering gear and the rattle noise are related.

5) The reliability of the model may have been lowered due to the non-reflection of the difference

between the design data and the actual steering gear’s various dimensional values due to production errors.

5. ACKNOWLEDGEMENTS

This research was supported by Hyundai Mobis.

6. REFERENCES

1. Y. Kozaki, G. Sekiya, and Y. Miyaura. Electric power steering (EPS), Motion & Control , vol.6,

pp. 9-15 (1999) 2. da Silva J, Zanini J. Design of Experiments Application (DOE) to Prevent Mechanical Noise in

Power Rack & Pinion Steering Systems, SAE Technical Paper , 2004-01-3377. (2004) 3. Fernholz C. A Simplified Approach to Quantifying Gear Rattle Noise Using Envelope Analysis,

SAE Technical Paper , 2011-01-1584. (2011) 4. Ognjanović1 M, Kostić SĆ. Gear Unit Housing Effect on the Noise Generation Caused by Gear

Teeth Impacts, Stronjniski Vestnik-Journal of Mechanical Engineering , 58, 327-337. (2012) 5. Loh, w. Classification and regression trees. Wiley interdisciplinary Reviews: Data Mining and

Knowledge Discovery, 1 (2011) 6. Brereton R.G., & Lloyd, G.R. Support vector machines for classification and regression. The

Analysis, 135 2, 230-67 (2010). 7. Krizhevsky, A., Sutskever, I., & Hinton, G.E. ImageNet classification with deep convolutional

neural networks. Communications of the ACM, 60, 84 – 90 (2012) 8. Babu, G.S., Zhao, P., & Li, X. Deep Convolution Neural Network Based Regression Approach

for Estimation of Remaining Useful Life. Database Systems for Advanced Applications, 214-228 (2016) 9. Myttenaere, A.D., Golden, B., Grand, B.L., & Rossi, F. Mean Absolute Percentage Error for

regression models. Neurocomputing , 192, 38-48. (2016) 10. Kohavi, R. A study of Cross-Validation and Bootstrap for Accuracy Estimation and Model

selection. International Joint Conference on Artificial Intelligence Organization(IJCAI), vol. 14, No. 2, pp. 1137-1145. (1995)

11. Zhou, B., khosla, A., Lapedriza, A., & Torralba, A. Learning Deep Features for discriminative

Localization. 2016 IEEE conference on computer vision and pattern recognition (CVPR) , 2921- 2929. (2016) 12. J. Kim, J.W. Lee, S.M. Lee and W. Kim. Study on Mechanism of Impact Noise on Steering Gear

While Turning Steering Wheel in Opposite Directions. Inter-noise 2016 , pp. 1095-1105.