International Journal of Automotive Technology, Vol. 16, No. 1, pp. 153−166 (2015) DOI 10.1007/s12239−015−0017−2
Copyright © 2015 KSAE/ 082−17 pISSN 1229−9138/ eISSN 1976−3832
HUMAN RIDE COMFORT PREDICTION OF DRIVE TRAIN USING MODELING METHOD BASED ON ARTIFICIAL NEURAL NETWORKS S. LERSPALUNGSANTI1)*, A. ALBERS2), S. OTT2) and T. DÜSER3) 1)
National Metal and Materials Technology Center (MTEC), Pathumthani, 12120, Thailand IPEK - Institute of Product Engineering, Department of Mechanical Engineering, Karlsruhe Institute of Technology (KIT), Karlsruhe, 76131, Germany 3) AVL Zöllner GmbH, Berliner Ring 95, Bensheim, 64625, Germany
2)
(Received 30 October 2013; Revised 26 February 2014; Accepted 4 March 2014) ABSTRACT−This article demontrates a systematic approach for human ride comfort objectification. The main objective is to integrate the prediction of the subjective comfort evaluation into the concept phase of product development process. By means of the human sensation model based on the Artificial Neural Networks (ANN), the subjectively sensed convenience of each passenger is estimated. In this paper, two examples of the implementation of the proposed methods are discussed. The first example represents an investigation of comfort-relevant design parameters of the drive train, such as the friction coefficient gradient of the clutch friction pair, the mass of inertia and the damping of the dual mass flywheel. The consequent vibrational properties and the subjective assessments during the start-up situation are predicted. The second example considers the determination of gear rattle tendency of a 5-speed manual transmission. To predict the gear rattle presence and to evaluate the resulted annoying level, the ANN-based models representing the NVH experts are elaborated. As a result, a good correlation of the subjective ratings and the predicted evaluations is attained. Consequently, the proposed method can be effectively applied to compare different gearbox solutions for new product concept. Hence, the development time and costs might be significantly reduced. KEY WORDS : Human ride comfort modeling, NVH, Vehicle start-up, Gear rattle, Artificial neural networks
1. INTRODUCTION
Neural Networks (ANNs) and the network topology optimization are explained in Section 2. Section 3 explains the procedure of the proposed approach to model the human ride comfort evaluation of drive train NVH phenomena by means of two examples. The first example refers to a study of ride comfort performance of occupant during the vehicle start-up. The second example considers the prediction of the gear rattle tendency and the evaluation of resulted annoying level of gear noise. Based on the proposed method, an ANN-based model represents the individual rating behavior of each passenger with a reference to ride comfort during each specific driving situation. Regarding the virtual product development, Section 4 shows the developed drive train model used to simulate the vehicle start-up situation. Considering the elasticity of driveshafts, the virtual drive train is applied to generate significant objective data for human ride comfort, such as the longitudinal seat vibration. In Section 5, the applications of the ANN-based human model to investigate NVH phenomena resulted from the variation of design parameters, which influence ride comfort in different ways, are demonstrated. The results are discussed, followed by the conclusion in Section 6.
Today’s vehicle developers are confronted with shortened development and product life cycles. While there are no legal requirements for the interior Noise, Vibration and Harshness (NVH) level in the passenger cabin, their effects are highly customer relevant. In automotive industries, the evaluation of NVH properties from the passenger point of view is generally performed by test persons using evaluation schemes (Schoeggl and Ramschak, 2000). To reach a high demand of ride comfort, a reliable method to predict the customer satisfaction of the developed vehicle is required. In this case, the term “human ride comfort modeling” is defined as reproduction of subjectively sensed convenience of a passenger through objectively measurable values. In this article, the systematic approach to objectify the drive train NVH characteristics, e.g. the seat vibration or gear noise, during driving maneuvers in an early stage of the product development process is proposed. The principles of human ride comfort modeling based on the Artificial *Corresponding author. e-mail:
[email protected]
153
154
S. LERSPALUNGSANTI, A. ALBERS, S. OTT and T. DÜSER
2. APPROACH OF HUMAN RIDE COMFORT MODELING BASED ON ARTIFICIAL NEURAL NETWORKS (ANNS) The determination of dynamic features of vehicle drive train commonly requires drive tests with passengers. To improve the NVH properties of the developed drive train, the corresponding modification on prototypes consequently leads to additional costs and time. To consider a customer demand in terms of ride comfort at the beginning of the drive train development process, a systematic approach of human ride comfort modeling based on ANNs, which is applied to evaluate the entire drive train properties based on each design concept, is developed, as illustrated in Figure 1. From the situation analysis, the development concepts and the target customer group are to be determined. Consequently, the concepts or ideas of potential solutions are to be concretized in form of simulation models or real prototypes. This process is followed by the generation of the objective data, which are inputs for the elaborated human ride comfort model for the prediction of ride comfort evaluation. Based on the predicted subjective comfort rating and the customer acceptance, the suggestions for one or more alternatives for a solution pave the way for a decision to be taken by developers. Further procedures, such as the repetition of the overall process or the implementation of the final concept, will be carried out successively. 2.1. ANN-based Human Ride Comfort Modeling An Artificial Neural Networks (ANNs) are defined as a computational tool mostly used either to find a relationship between inputs and outputs or to recognize a data pattern (Brause, 1995). In this study, a feed-forward network, which is one of widely-used class of neural networks (Gurney, 1997), is applied. The feed-forward network can be described with a simplified mathematical function z(w,x) defined by the nonlinear weight sum according to:
Figure 1. Approach of customer ride comfort modeling in the drive train design process.
z ( w, x ) = f ⎛⎝ ∑ wk ⋅ xk⎞⎠
(1)
k
z denotes the output vector consisting of the data to be approximated. x represents the input vector consisting of k input nodes, xk. w denotes the vector of the connection weights wk between the input node xk and the output node z. The type of the function z is defined by the applied activation function f. This can be either linear or nonlinear function, such as the log-sigmoid function. The structure of ANNs is composed of several (at least two) layers of neurons, connections, weights, biases and activation functions. The determination of the network topology, such as the number of layers or the number of neurons in each layer, generally depends on the problem to be solved. The training algorithm used during the modeling is defined as the procedure to minimize the error by modification of weights and biases in each network connection (Brause, 1995). Due to the capability of the ANNs to model the complex nonlinear relationships between the inputs and the outputs, the ANN-based methods are developed in this study to support the customer-oriented drive train design in the early phase of product development process. As illustrated in Figure 2, the operating principle of ANNs for the current research is separated into a training stage and a model application stage. Similarly to the way a passenger makes his evaluation, the ANN-based human model is introduced to find a correlation between the objective data and a subjective comfort evaluation from the individual customer point of view by “trained” weighted network connections at the training stage. At the application stage, the prediction of ride comfort evaluation is carried out by means of the trained ANN-based model. The input data for the ANNbased model can be both measured data from test vehicles or test benches and the generated data from simulation models. Earlier studies (Karnin, 1990; Albers and Albrecht, 2004; Kannan, 2013; Ortega and Silva, 2008) show that the
Figure 2. Operating principle of ANNs at the training phase and application phase.
HUMAN RIDE COMFORT PREDICTION OF DRIVE TRAIN USING MODELING METHOD BASED
155
increases for a specified number of iterations, the training is aborted. To represent the capability of the elaborated ANN-based model to predict the target value, the assessment method is developed in this study. The prediction accuracy; G, which ranges from 0 to 1, is determined due to Equation (3) Figure 3. Example of network topology for driver ride comfort modeling. network topology, the type of activation function, the training algorithm and the training parameters, which can be varied in accordance with each algorithm, particularly have different influences on the performance of ANNs. An example of network structure for the driver ride comfort modeling used in current study is illustrated in Figure 3. In this example, the 30 × 1 input matrix is composed of 30 objective parameters, which are defined as significant factors influencing the ride comfort. The 1 × 1 output matrix consists of a neuron representing the comfort evaluation. Depending of the number of training data set NT , the number of hidden neuron NH can be determined according to Equation (2) (Brause, 1995). NT – 1NH = -----------NI + 2
(2)
NI represents the number of input neuron. In this example, a hidden layer of 3 hidden neurons is set. The neurons of the input layer and the output layer, which are employed for this example use the linear activation function. The hidden neurons use the logarithmic sigmoid function. Due to this, the error function of the ANNs becomes continuously and universally differentiable, which is required for searching of a minimum of error. To avoid “memorising” the input-output pairs instead of discovering the rule of connection, the “early stopping” procedure (Brause, 1995) is applied. By applying this method, the available training data is divided into two subsets, as shown in Figure 4. The first subset is the learning set used for computing the gradient and updating the network weights and biases. The second subset is the validation set, used to approve the network performance during training. The error in the validation set is monitored during the training process. Due to the random partition between the learning and validation sets, each pair consisting of an input and a target participates an equal number of times to both of these sets. If the validation error
Figure 4. Partition procedure of data set and process of neural network training.
⎧ ( a + R2 – Δ M ) if 0 ≤ a ≤ 1 ⎪ ---------------------------2 ⎪ 2 G = ⎨ ( 2 – a + R – ΔM ) - if 1 < a ≤ 2 ⎪ ----------------------------------2 ⎪ 0 if a < 0, a > 2 ⎩
(3)
a and R2 denote the regression coefficient and coefficient of determination which are derived from the correlation of (actual) output and (desired) target values. Δm represents the average deviation of output and target values, respectively. From the results of previous study (Albers and Albrecht, 2004), increasing number of traing data set can generally improve the modeling performance. The more independent input-target sets are presented to the ANNs in the training stage, the better approximation in the application stage due to lower effect of outliers can be attained. After the training stage of network, the structure and the combination of weights and biases, which fulfils the abort condition of the training algorithm, are considered as the “knowledge” of the ANN-based model. Consequently, the trained ANNbased model, which represents the individual rating behavior of each passenger, is applied to predict the target data representing the human ride comfort evaluation. 2.2. Optimization of Network Topology The generalization of the network, i.e. the capability to achieve high accuracy of subjective evaluation prediction, generally depends on its structure. In this study, an approach of topology optimization is proposed. The optimization method is developed based on the neuron networks pruning. As described in (Prechelt, 1995) and (Karnin, 1990), the principle of neuron pruing involves the identification of significant and redundant neurons. To achieve the optimal topology of networks, the irrelevant or redundant neurons are to be removed from the initial networks. The relevancy estimation for each neuron is carried out by calculating the sensitivity of the global error function to the output/input of each connection. The sensitivity value Sij of each input connection can be defined according to Equation (4) (Karnin, 1990): N–1 wijf Sij = ∑ [ Δwij( n )]2 ⋅ -------------------------f iη ⋅ ( wij – wij ) n=0
(4)
Sij denotes the sensitivity of input neuron j to hidden neuron i. N represents the number of training iterations for each initialization. Unit j is connected to unit i via a weight connection at the beginning ωjfi and at the end ωjff of training. Δωji denotes the deviation of connection weights between input j and hidden neuron i after each presentation of a set of training patterns. η represents the used learning
156
S. LERSPALUNGSANTI, A. ALBERS, S. OTT and T. DÜSER
Figure 5. Methodology for the development of the ANNbased human sensation model. rate during network training. If each input neuron is connected with more than one hidden node, such as m nodes, the “Local Relative Sensitivity Index” (LRSI) (Karnin, 1990) representing the total sensitivity value for each input neuron is determined due to Equation (5): Sij LRSIij = ---------------ΣNm = 1 Sim
(5)
3. HUMAN RIDE COMFORT MODELING To objectify the human sensation according to each NVH phenomenon during specific maneuver, the comfortrelevant characteristic values as well as the subjective assessment scheme, such as the 10-digit scale (Aigner, 1982), are to be predefined. After obtaining all data, both sets of objective measurable data representing the input data and the subjective evaluation denoting the target data are combined into a pattern. This is used as an inertial training data set for the human sensation modeling, as illustrated in Figure 5. This section presents the procedure of the proposed methods to model the human ride comfort evaluation of NVH phenomena by means of two examples. The first is the ride comfort modeling during vehicle start-up and the second is the modeling of gear rattle noise evaluation. 3.1. Modeling of Vehicle Start-up Evaluation One of the driving situation which mostly influences the passenger ride comfort is the start-up procedure, i.e. the process of starting to drive from standstill with releasing of the brake and reaching of constant travel speed or decelerating. In current study, a front-drive, intermediateclass car with an automated clutch system has been set up to examine the effects of vibration during clutch engangement. By means of the integrated clutch system controller, different start-up characteristics defined by developers can be investigated. In addition, vibrational phenomena, which might decrease the degree of comfortability, like the chatter and the jerking (Albers and Herbst, 1998), are generated on purpose to verify the prediction capability of the developed ANN-based models. In this case, the data characterizing the start-up process can be expressed as vibrational vehicle response and driver demand. According to the vehicle response, the transverse accelerations in longitudinal, lateral and vertical directions
Figure 6. Examples of characteristic values captured during the vehicle start-up with standard adjustment.
are measured at the seat rail and at the neck rest. To consider the driver demand, i.e. his requirement of the start-up character, the gas pedal position over the course of time during the process is captured. Examples of some characteristics, which are investigated in this study, are illustrated in the following figures. The corresponding subjective assessments and the human ride comfort modeling results are presented in Section 5. Figure 6 demonstrates the curves corresponding to the alternative with standard clutch adjustment of production vehicle. During start-up process, all predefined characteristic values are captured. These are the resulting longitudinal acceleration; aseatrail, which is measured at the seat rail, the engine speed; nengine, the transmission input speed; ntransmission, the standardized clutch actuator way; sactuator, the corresponding standardized courses of the brake; sbrake, and the standardized courses of gas pedal; sgas. Figure 7 represents a start-up process with deliberately generated longitudinal oscillations in the range of the typical vehicle jerking frequencies (Albers and Herbst, 1998). In comparison to the start-up with stardard adjustment, this alternative is made with increasing the intensity of the pulsing in higher level which may affect the comfort of passengers. The network structure applied to predict the vehicle start-up evaluation in this study is presented in Figure 3. The ANN-based model is developed based on drive tests on proving ground, performed by twenty drivers
Figure 7. Examples of characteristic values measured during the vehicle start-up with deliberately generated longitudinal oscillations.
HUMAN RIDE COMFORT PREDICTION OF DRIVE TRAIN USING MODELING METHOD BASED
157
Figure 8. Measurement of vehicle speed and accelerations at neck rest and at seat rail of test vehicle. representing average customers. The network topology consists of thirty input neurons, three hidden neurons determined according to Equation (2) and one output neuron. The 30 × 1 input matrix consists of six objective parameters describing the driver demand, four objective parameters of the longitudinal acceleration signals, captured at the seat rail and the neck rest, as shown in Figure 8, nineteen modified power spectal density values of the longitudinal acceleration signals, and a value of startup duration. The examples of objective parameters which represent the driver demand are throttle pedal travel distance and the corresponding pedal speed. The modified power spectal density values of the longitudinal acceleration signals Φa,mod (ω) are calculated according to Equation (6) for a constant and sufficiently small filter frequency range Δω. 1 Φa, mod( ω ) = -----T
Figure 10. Comparison of sensitivity values of thirty input neurons.
T
∫ a (Δω)dt 2
(6)
0
The 1 × 1 output matrix consists of a neuron representing the comfort evaluation based on the 10-digit-scale (Aigner, 1982). As shown in Figure 9, the left endpoint marked as “unacceptable” stands for 0 and the right endpoint “none” means 1. As described in Section 2.2, the modeling process is then followed by the model optimization to achieve higher prediction performance. Figure 10 illustrates the calculated average sensitivity values of all input nodes over twenty drivers, due to Equation (4) and Equation (5). These are compared with the values of some driver models which achieve high performances, such as driver No. 2, No. 12 and No. 15. Considering the average LRSI over all models, the sensitivity values of input neuron No. 23 and No. 24 are the lowest from all 30 nodes. According to (Karnin, 1990), it can be assumed that both neurons are irrelevant for the
Figure 9. Comfort evaluation scheme based on the 10-digit scale (Aigner, 1982).
Figure 11. Comparison of performance values before and after network topology optimization based on neuron pruning.
modeling. Consequently, they can be removed from the initial network topology. Thus, the new network topology for the modeling of subjective comfort will be set to 28-3-1. After the training of twenty ANN-based models using the LevenbergMarquardt algorithm; trainlm, the higher prediction performances calculated according to Equation (3) are accomplished, except those of some drivers, such as driver No. 5, No. 12 and No. 17, as demonstrated in Figur 11. However, those differences ΔG of about 2% before and after model optimization are acceptable. The following example shows the capability of the elaborated ANN-based model to predict the subjective rating. In this case, the test data set derived from different experimental start-up conditions is applied. As shown in Figure 12, two dash lines represent a deviation from the exact value of ± 10%. In case of exact approximation, all
158
S. LERSPALUNGSANTI, A. ALBERS, S. OTT and T. DÜSER
Figure 14. 2-layer-Scale for gear rattle evaluation.
Figure 12. Comparison of subjective rating and ANN output, driver No. 12. points would lie on the first bisecting line. The prediction results for ANN-based model with the prediction accuracy; G of 95.7% show that a reasonably accurate approximation is possible. 3.2. Modeling of Gear Rattle Noise Evaluation This section presents the modeling results of gear rattle evaluation, which is one of the dominant noises from the drive train. The gear rattle phenomenon from transmission is caused by torsional vibration of transmission components, which are not under load and moving within their functional clearances and knocking against their limits. The gear rattle noise generated by low frequency vibration is perceived as particularly unpleasant, not because of its high airborne sound pressure level, but because of its unpleasant characteristic (Dogan, 1999). In this case, experimental investigations of gear rattle are carried out on the IPEK universal transmission test bench of the Insitute of Product Engineering, Karlsruhe, Germany. This experimental set-up was developed to generate rattle noise in a commercial vehicle transmission by means of the simulation the non-uniform rotational speed pattern of a vehicle internal combustion engine. As shown in Figure 13, the test bench consists of an input drive with an electric motor with a total output of 112 kW from two electrical motors, with a summarizing tooth belt drive to transmit the output to the drive shaft. The irregular rotational speed patterns can be generated with defined oscillating torque excitation, up to an angular acceleration of 1,200 rad/s2. To measure the air borne sound pressure level, an artificial head for binaural recording of gear rattle noise is used. By
Figure 13. Experimental investigation of gear rattle on the IPEK universal transmission test bench.
means of the measurement system for data acquisition and data analysis, the predefined characteristic psychoacoustic values, such as the A-weighted sound pressure level, the loudness or the sharpness are generated. The gear housing vibrations are measured at different positions using piezoelectric accelerometers. In addition, the gear shaft speed, the input and output torque and the transmission oil temperature are captured. In this study, the subjective assessment has been performed on a basis of the 2-layer-scale, which is developed in such a way that the appearance and the evaluation of gear rattle noise from the expert point of view are attained. This allows the allocation of the gear rattle noise appearance and the rating of disturbance level. As illustrated in Figure 14, the left box marked as “intolerable” stands for 1 and the right box with rating of 5 means “no disturbance”. In this case, a relation of gear rattle evaluation and significant objective data, consisting of measurable data and calculated psychoacoustic parameters, is to be determined. As a result from the modeling, the trained ANN-based model with the prediction accuracy; G of 84.34% is attained. As shown in Figure 15, the topology of network for gear rattle is consisting of twenty six objective parameters. These are: the A-weighted air borne sound pressure level, the structure-borne sound pressure at different gear housing positions, the captured duration of exposure and the calculated psychoacoustic parameters like the loudness and the sharpness (Zwicker, 1999). In addition, the network structure consists of three hidden neurons and an output neuron representing predicted comfort assessment by using the Levenberg Marquardt algorithm. At the application stage, this ANN-based model will be applied to predict the gear rattle tendency and its effect on human comfort sensation.
4. VIRTUAL DRIVE TRAIN To partially substitute real test drive with virtual one, the
Figure 15. Example of network topology for gear rattle prediction.
HUMAN RIDE COMFORT PREDICTION OF DRIVE TRAIN USING MODELING METHOD BASED
159
Figure 16. Vehicle model for start-up simulation. vehicle model is deliberately developed to simulate the start-up process of a middle-class car with a front-wheel drive. Nevertheless, the main goal of the vehicle modeling of the current study is not to simulate the real system dynamic property precisely, but to generate the objective data depending on applied drive train parameters. This should be accomplished in such a way that the ANN-based model provides the ride comfort rating corresponding to a real human behavior. As illustrated in Figure 16. Vehicle model for start-up simulation., the vehicle model consists of sub-models of a driver, a drive train and a sub-model for vehicle simulation. In addition to the output torque generated from the entire vehicle simulation, the time progression of the clutch actuation, the brake pedal and the throttle value are given by the driver model as input signals to the successive drive train model. Consequently, the rotational speed and the resulted torque of the driving wheels are generated as output. Using these two signals to simulate the entire vehicle together with the peripheral factors like rolling resistance or aerodynamic drag, the vehicle longitudinal acceleration and the output torque are simulated. The driver model presented in Figure 17 is developed to consider the driver demand on the vehicle start-up. In this case, the time progression of an acceleration pedal value, actuation of the brake pedal and the clutch pedal during the vehicle start-up can be either the measured signals from drive tests or the signals manually provided by the developer. In view of the simulation of the combustion engine, the engine map is used to generate the engine torque signal depending on the current engine speed and the throttle
Figure 18. Structure of the simplified drive train model of the test vehicle with the engaged clutch.
value. As shown in Figure 18, the engine torque signal containing an irregularity of a rotational speed is the input signal of a drive train unit. This enables the simulation of the longitudinal dynamic property of the entire vehicle based on each setup drive train configuration. The simplified drive train model consists of the combustion engine, the dual mass flywheel (DMF), the engaged clutch, the front-wheel drive transmission and tires. The rotary motion of each assembly can be described according to Equation (7) (Isermann, 2002): Jϕ·· = ΔT = Tdrive – Tload
(7)
ΔT denotes the difference of the driving torque; Tdrive and the loading torque; Tload of a rotational body. For instance, T1 and T2 represent the driving and loading torque of the body No.1. J and ϕ·· eclare the mass of inertia and the angular acceleration respectively. According to Equation (7), the rotational speed ϕ· as a function of time t is calculated by: T-----dt ϕ· ( t ) = ∫ Δ J
(8)
To consider the drive shaft elasticity of drive train components, the mass-spring-damper system is implemented according to the principle of angular momentum (Isermann, 2002), as followed: T = ϕ· ⋅ d + ϕ ⋅ c = Δϕ· ⋅ d + ∫ Δϕ· ⋅ c ⋅ dt
Figure 17. Driver model for simulation of driver demand on the vehicle start-up.
(9)
d and c represent damping and stiffness of the considered spring-damper system, respectively. Δϕ· denotes the difference of the rotational speed between the two components locating next to each other. For example, T2 is
160
S. LERSPALUNGSANTI, A. ALBERS, S. OTT and T. DÜSER
based on 20 laymen. Three groups of potential customer are determined on the basis of the style of driving. Each group of sporty and comfort-oriented drivers consists of 6 members and 8 members for a group of ordinary drivers. The 10-digit scale, as shown in Figure 9, is used for the comfort evaluation of each customer group. Depending on each drive train configuration, the rating score from these three customer groups is used to support developers by making a decision. On the basis of the applied rating scale, the acceptable limit is set to 5 points.
Figure 19. Comparison of measured and simulated signals during the vehicle start-up at full throttle. calculated using Equation (9) by applying Δϕ· 12 . This describes the difference between the rotational speed of the DMF-primary mass n1 and the gear input shaft n2. The simulation of this vehicle during the start-up process is carried out by considering the significant resistive forces consisting of the longitudinal resistant forces due to the rolling friction, air resistance and the acceleration resistance. Consequently, the output signals are generated based on the refered parameters of the drive train. As an example, Figure 19 shows comparisons between the measured signals and the simulated signals of the engine torque, the engine speed, the gear input shaft speed and the longitudinal acceleration during the start-up with full throttle. As shown in Figure 19, the engine speed signal measured from the ECU does not contain the rotational irregularity of the crankshaft, as the time progression of the measured engine torque presents. To consider this rotational irregularity, the engine model is enhanced. Consequenty, the time progression of the measured and the simulated signal are relatively similar. To prepare the input data for the ANN-based model, the predefined objective values are generated based on the simulated longitudinal acceleration signal. These values represent the measured data from the drive test, such as the maximum acceleration, the start-up duration, and the values calculated by the analysis software, like the power spectral density.
5.1.1. Variation of clutch actuation To investigate the human responses to different start-up characteristics, the clutch actuation signals are varied in the clutch model, as illustrated in Figure 20. In addition to the standard adjustment, the second alternative is generated with additional long slip phase. The other two alternatives are generated by verifying the intensity of the pulsing in low and high levels, which differently affect the rotary vibration of the entire vehicle and the comfort of passengers. The resulted comfort ratings from this variation are to be considered. Figure 21 represents the simulated signals of a rotational speed of both the engine and the transmission input shaft during the clutch engagement of all four
Figure 20. Clutch actuation variants during start-up process.
5. DESIGN PARAMETER INVESTIGATION AND PREDICTION OF NVHCHARACTERISTICS This section present the capability of the human ride comfort modeling methods to evaluate the NVH-characteristics by means of two examples: the vibrational vehicle response during vehicle start-up situation and the gear rattle noise, respectively. 5.1. Prediction of Start-up Comfort Evaluation based on Variation of Drive Train Design Parameters The human sensation models applied in this study are
Figure 21. Simulated time courses start-up by different clutch actuation variants, at throttle of 70%.
HUMAN RIDE COMFORT PREDICTION OF DRIVE TRAIN USING MODELING METHOD BASED
variants. In this case, the throttle value of 70% is set in the driver model for the start-up simulation. The start-up duration of the second alternative is longer than the one with the standard clutch adjustment due to the long slip phase. Consequently, the possibility of the occurrence of chatter is higher than the other alternatives. Nevertheless, this tendency is also depending on the friction property of the clutch disk. Considering the occurrence of the jerk, the stimuli are generated by varying the longitudinal oscillation levels of the clutch pressure plate. The low and the high intensities of the oscillation lead to different vibrational levels of the transmission and the entire vehicle during the start-up process. Consequently, the time progression of the longitudinal acceleration and the estimated comfort ratings using ANNbased models of each clutch actuation variant are demonstrated in Figure 22 and Figure 23, respectively. In accordance with the comfort evaluations made by the real drivers, the ANN-based model of the sporty drivers give
Figure 22. Simulated longitudinal acceleration signals by different clutch actuation variants, at throttle of 70%.
Figure 23. Estimated customer comfort rating using ANNbased models by different clutch actuation variants, at throttle of 70%.
161
the highest scores in nearly all simulated situations than the other models. On the other hand, the comfort ratings attained from the ANN-based models of the comfortoriented customers are the lowest. The ratings from the ANN-based model of the ordinary drivers mostly range around the mean of those from the other groups. In addition, the models representing the comfort-oriented customers evaluate all modified start-up processes with the scores under 5 points, which is interpreted as unacceptable. As a result from the model modification, the start-up with a long slip phase is evaluated by all representative drivers as less comfortable than the standard one due to the higher potential of the chatter. This is caused by both the longer start-up period and the large amplitudes of the longitudinal acceleration signal. As expected, the estimated scores of the start-up with a high level oscillation are lower than those from the start-up with a low level oscillation. With an exception of the start-up with a high level oscillation, all created situations reach the required level of the customer satisfaction. This result is in accordance with the subjective comfort assessment from the drive test. 5.1.2. Variation of the component mass of inertia One objective of the application of the dual-mass flywheel (DMF) is to eliminate the excessive rotational vibration caused by the irregularity of rotational combustion engine speed (Reik, 1987). To attain a high performance of the vibration damping, the increase in the primary J1 and the secondary mass of inertia J2a is one of the commonly applied solutions. Apart from the body of the secondary mass, other assemblies with a direct contact to the secondary mass like a clutch housing are parts of the mass of inertia J2a. The simplified schematic positions of J1 and J2a are illustrated in Figure 24. Another option to damp the vibration from combustion engine is to increase the mass of inertia of the whole assembly at the gear side without raising the sum of entire mass to be engaged. Practically, the mass of inertia of the clutch disk and the gear input shaft J2b can be increased for this purpose. As an example, this section demonstrates different damping effects of the entire drive train and the consequent
Figure 24. Schematic position of drive train assemblies: crank shaft, DMF-primary mass, DMF secondary mass, clutch and gear input shaft.
162
S. LERSPALUNGSANTI, A. ALBERS, S. OTT and T. DÜSER
Table 1. Variants of vehicle from the variation of the DMF masses of inertia. Vehicle Assembly
A
B
C
D
E
Mass of inertia [kgym ] 2
J2a
0.0450
0.0200
0.0900
0.0450
0.0450
J2b
0.0027
0.0027
0.0027
0.0012
0.0100
comfort evaluations due to the variation of J2a and J2b. The variation of the mass of inertia is carried out on the basis of the original vehicle parameters used in the simulation model: J2a = 0.045 kg·m2 and J2b of 0.0027 kg·m2. In this case, a total of five vehicles are generated by varying the J2a and the J2b as presented in Table 1. To generate the start-up processes at full throttle with the standard clutch adjustment, the vehicle A represents the test vehicle set up with all original parameters. Based on these parameters, the vehicle B and C are set with about the half and the double inertia of the original J2a , respectively. The vehicle D is created using the original J2a and about a
Figure 25. Simulated time progression of start-up and longitudinal acceleration signals by variation of J2a , at full throttle.
Figure 26. Simulated time progression of start-up and longitudinal acceleration signals by variation of J2b, at full throttle.
half of the original J2b. To investigate the effect of the increased mass of inertia at the transmission side, the vehicle E is built by increasing the J2b to 0.01 kg·m2. Figure 25 and Figure 26 demonstrate the time progressions of the simulated engine speed, the transmission speed, and the consequent longitudinal acceleration of all vehicles. On one hand, the damping effects of the gear input shaft speed and the longitudinal acceleration are obviously attained by increasing either the J2a (vehicle C) or the J2b (vehicle E). On the other hand, the damping effect is reduced by decreasing either the J2a (vehicle B) or the J2b, (vehicle D). By analyzing the simulated acceleration signal of all vehicles, the estimation of the subjective comfort ratings from all types of potential customer using ANNbased models is carried out. From Figure 27, the tendency of higher customer satisfaction is clearly achieved by increasing the J2b (the vehicle A and the vehicle E in comparison with the vehicle D). This progress is gained especially from the ANN-based model of the comfort-oriented driver. This is caused by the high comfort sensitivity of the comfort-oriented driver according to the change of objective values. On the contrary, the apparent raise of comfort rating appears only from the ANN-based model with the comfort-oriented style by increasing the J2a (the vehicle A and the vehicle C, in comparison with the vehicle B). Considering the evaluation from the ordinary and sporty drivers, the startup properties of the vehicle A meet their comfort expectation better than those from the vehicle B and C. These results indicate that the change of J2a is not worthwhile in this case. Due to the fact that the additional damping mass generally leads to more weights and higher costs of the entire system, the compromise among the customer comfortability, the economical and the energy-efficient aspects is to be found. Furthermore, the benefits of the damping performance and the improvement of customer satisfaction have to be considered. In this example, the
Figure 27. Estimated customer comfort rating by variation of J2a and J2b.
HUMAN RIDE COMFORT PREDICTION OF DRIVE TRAIN USING MODELING METHOD BASED
results show that only the increase of the mass of inertia of the clutch disk/gear shaft is profitable, and not the one of the secondary mass. 5.1.3. Variation of clutch friction coefficient gradient Many today studies related to the testing of different clutch disk materials have a goal to find alternatives providing the friction characteristic leading to the reduction of the vibration during the clutch engagement. One significant factor for this purpose is the gradient of clutch friction coefficient μ’. This value is defined as a change of the friction coefficient of the friction pair over the increasing slip speed of the clutch disk and the pressure plate (Albers and Herbst, 1998). The increasing of friction coefficient during the clutch engagement commonly causes the selfinduced chatter, i.e. the friction vibration, which affects the ride comfort of passengers (Albers and Herbst, 1998). This section presents the responses of the human ride comfort according to the variation of the friction coefficient
Figure 28. Simulated time progression of start-up by variation of clutch friction coefficient, at throttle of 50%.
Figure 29. Simulated longitudinal acceleration signals by variation of clutch friction coefficient, at throttle of 50%.
163
gradient. This value can be varied in the simulation model of the clutch. In this case, the standard clutch actuation signal and the throttle value of 50% are inputs for the drive train model. The generated signals are presented in Figrue 28 and Figure 29. To examine different levels of the friction vibration due to the chatter, the friction coefficient μ’ is varied in the original clutch model based on the original value μ’ of −0.003 s/m. As shown at the bottom right picture of both figures, the chatter is clearly induced in case of μ ’= −0.004 s/m. In addition to the large amplitudes of the transmission speed signal and the corresponding longitudinal acceleration appearing during the clutch engagement, the maximum of the longitudinal acceleration of approximately 4.4 m/s2 is reached. By scaling up of the friction coefficient with increasing of slip speed to μ’ = −0.001 s/m and −0.002 s/m, the friction contact is damped and behaves more stable. The friction vibration is not intense in comparison to those simulated by setting the slip speed; μ’ of −0.003 s/m and −0.004 s/m. Consequently, the start-up situation with μ’ = −0.004 s/m is rated as the worst from all models. As shown in Figure 30, the ANN-based model of the ordinary and the comfort-oriented drivers rate this situation with scores under a half point. In spite of the score of about 4 points from the sporty driver, this drive train feature is defined as unacceptable. The consequence of the enhancement of the clutch friction coefficient gradient μ’ to −0.001 s/m and −0.002 s/m is that the scores over 6 points are achieved from all three ANN-based models. The drive train with the original μ’ of −0.003 is assessed as acceptable only from the ordinary and the sporty drivers. Concerning the relatively low ratings from target customer groups, the enhancement of customer satisfaction should be achieved. One of potential solutions is to increase the clutch friction coefficient gradient μ’.
Figure 30. Estimated customer comfort rating by variation of the gradient of clutch friction coefficient, at throttle of 50%.
164
S. LERSPALUNGSANTI, A. ALBERS, S. OTT and T. DÜSER
Figure 31. Simulated time progression of start-up and longitudinal acceleration signals by variation of d34, at throttle of 50%.
5.1.4. Variation of drive train damping As mentioned in an earlier section, the reduction of the friction vibration due to the chatter can be achieved; for example by increasing the friction coefficient gradient of the friction pair. Other than this, some procedures described in the literature (Albers and Herbst, 1998), can be applied like the installation of the additional mass damper in the drive train to reach the higher damping. In this study, the drive train parameter d34 can be increased based on the original value to examine the effect of different damping. d34 denotes the damping of the driveshaft system No.3 of the simplified drive train model presented in Figure 18. This driveshaft system consists of drive train assemblies, such as the both drive axle shafts, the suspensions and the wheels. In this case, driving situations with various damping values during the appearance of the chatter are simulated based on similar conditions. These are the standard clutch adjustment, the throttle value of 50% and the friction coefficient gradient; μ’ of −0.004 s/m. As shown in Figure 31, the increase of d34 from the original value can effectively reduce large amplitudes of the transmission input shaft speed and the
longitudinal acceleration. Furthermore, the resulted maximum acceleration of around 4.8 m/s2 and 2.7 m/s2 appear by using the original d34 of 10.5 Nm·s/rad and 20 Nm·s/rad, respectively. Consequently, the obviously higher customer ratings due to the increase of the drive train damping from all ANN-based models are attained as presented in Figure 32. By setting d34 of 20 Nm·s/rad, the highest rating of around 7.4 points is reached from the ANN-based model of the comfort-oriented driver. In addition, the apparent progression of the comfort ratings estimated by both ANNbased models of the sporty and the ordinary drivers are achieved by increasing the d34. From this result, the modification of the damping of the drive train is reasonable. Due to the fact that aspects of the additional costs, the more weights and the successive fuel consumption are the consequences of this improvement, the overall consideration of the entire vehicle has to be carried out to find the best solution. 5.2. Prediction of Gear Rattle Noise Characteristic Examples in this section present the capability of the objectification tools and the elaborated ANN-based model to predict the evaluation of gear rattle noise generating in different testing conditions. The variation of testing condition is carried out by removing different parts of the tested transmission, as illustrated in Figrue 33, which
Figure 33. Concept of the tested 5-speed manual transmission. Table 2. Test plan for removed parts. Test Test precedure mode A B Figure 32. Estimated customer comfort ratings by variation of d34, at throttle of 50%.
C
idle
Removed parts
Possible rattling gears
output shaft
Gear No.1, No.2, No.5 and No.6
intermediate shaft
Gear No.3 and No.4
output shaft, Gear No. 2, Gear No.1 No.5 and No.6
Oil temp.
30oC
HUMAN RIDE COMFORT PREDICTION OF DRIVE TRAIN USING MODELING METHOD BASED
Figure 34. Structure-borne pressure level curves of transmission housing.
Figure 35. Objective parameters based on different removed parts.
causes different number of possible rattling gears. The test plan for this example is presented in Table 2. As an example of parameter investigation of different removed parts, Figure 34 shows a comparison of the structure-borne pressure level curves and the objective parameters from the measurements. Figure 35 illustrates the predefined objective parameters: the maximum loudness, the maximum air borne sound pressure, the maximum air borne sound pressure level, the maximum structure borne sound level, maximum roughness, maximum sound level, and maximum sharpness. They are applied as input data for the trained ANN-based model. By comparing of test A and B, the effect of removing the intermediate part (test B) causes the increase of the structure-borne pressure level by up to 8 dB and the reduction of the maximum air borne sound pressure level. In case of a possible rattling gear (test C), the resulted structure-borne pressure level curve is nearly similar to the one of test A. Considering the other objective values like psychoacoustic parameters, the differences of almost
165
Figure 36. Application of the driver model and virtual drive train to estimate the subjective evaluation.
objective data are relatively small in spite of different numbers of removed parts. Figure 36 demonstrates the comfort ratings resulted from each test based on the 2-layer-Scale for gear rattle evaluation, as shown in Figure 14, with the subjective comfort ratings on the abscissa and the calculated output values from ANN-based models on the ordinate. If exact approximation was possible, all points would lie on the first bisecting line. By using the trained ANN-based model with prediction accuracy; G of 84%, a relatively good correlation of the subjective evaluation and the calculated values is achieved. Consequently, the appearance of gear rattle noise of all testing conditions can be predicted. Concerning the disturbance level, the same trends of comfort rating of rattle noises with “small disturbance” from test A and C are attained. On the other hand, the same ANN-based model rated the generated noise from the test B as “strongly annoying”. The results indicate that the presented approach can be effectively applied to predict the gear rattle evaluation based on generated data from the test bench.
6. CONCLUSION This article presents the approach of human ride comfort modeling on a basis of the Artificial Neural Networks (ANN). The method for network optimization is proposed to provide reliable results and to reach higher performances of comfort prediction. In this study, the ANN-based models are elaborated to predict the individual subjective comfort ratings of passenger due to NVH-phenomena in vehicle drive train. The capability of the proposed approach to support the drive train design is presented by means of two practical examples. The first example denotes different design parameters
166
S. LERSPALUNGSANTI, A. ALBERS, S. OTT and T. DÜSER
investigations of the drive train assemblies, including the dual-mass flywheel or the clutch, to enhance the passenger ride comfort satisfaction of the developed drive train during a driving situation, e.g. the start-up process. For this purpose, the simulation model of the test front-wheel drive vehicle is applied to generate the predefined objective data. The presented vehicle model allows the consideration of driver demand in views of the vehicle start-up property. By means of the driver model, the actuation of the acceleration pedal, the brake pedal and the clutch can be given as the driver demand. In this study, three different types of driver regarding the driving style and ride comfort expectation are considered. Based on the variation of drive train design parameters, simulated driving processes are evaluated from the opinion of the sporty, the ordinary and the comfortoriented drivers. Consequently, the customer-oriented analysis, the decision of new concepts, and the corresponding improvement of developed drive train component can be achieved. The results of the human sensation modeling emphasizes that the selection of a suitable topology in views of input neuron type and training algorithm has a significant effect on the prediction performance. The experimental results indicate the different behavior of each type of driver to evaluate the same driving situation. In addition, examples of the parameter adjustment to achieve higher comfort ratings are demonstrated. In this study, the higher comfort ratings of all driver groups are attained by increasing the drive train damping. This approach can successfully reduce the rotary vibration according to the chatter during the clutch engagement. The second example considers investigations of the gear rattle noise on the IPEK universal transmission test bench. The human ride comfort modeling is carried out to assess the NVH property of a 5-speed manual transmission in view of gear rattle tendency and its disturbance level perceived by an expert. An example of model application is demonstrated based on tests with different removed parts of the tested transmission. As a result, the good correlation of the subjective rating and the predicted evaluation of all testing conditions is attained. By means of the presented examples, the potential of the developed methodology to support the gear rattle investigation is shown. The results from both examples indicate the potential of the developed methodology to support a developer by making a decision during the search of solution. To improve the NVH characteristic of the entire drive train and to reach a higher customer satisfaction, new concepts emerged from influencing drive train parameters can be examined. In many cases, a discussion of developers is required to find a compromise due to the conflict of goals. For example, the increase of drive train damping by installing the mass damper leads to more weight of an
entire system. In the long run, the optimization of the entire drive train system concerning the customer ride comfort can be integrated into an early phase of the product development process. This enables the quality gates and process assurance between design, testing and production of new product.
REFERENCES Aigner, J. (1982). Zur zuverlassigen Beurteilung von Fahrzeugen. ATZ Automobil-technische Zeitschrift 84, 9, 447−450. Albers, A. and Albrecht, A. (2004). Driver start-up comfort rating of automated clutch systems – an objectification using artificial neural networks. FISITA (Int. Federation of Automotive Engineering Societies): Proc. World Automotive Cong. Barcelona. Albers, A. and Herbst, D. (1998). Chatter – Causes and Solutions. LuK: Proc. 6th LuK Int. Symp., Baden-Baden. Brause, R. (1995). Neuronale Netze. 1st edn. B.G. Teubner. Stuttgart. Dogan, S. N. (1999). Loose part vibration in vehicle transmissions - Gear rattle. J. Engineering and Environmental Science 23, 1, 439−454. Gurney, K. (1997). An Introduction to Neural Networks. 1st edn. Routledge. London. Isermann, R. (2002). Mechanic Systems – Fundamentals. Springer. Berlin, Heidelberg, New York. Kannan, G. R. (2013). Artificial neural network approach to study the effect of injection pressure and timing on diesel engine performance fueled with biodiesel. Int. J. Automotive Technology 14, 4, 507−519. Karnin, E. D. (1990). A simple procedure for pruning backpropagation trained neural networks. IEEE Trans. Neural Networks 1, 2, 239−242. Ortega, A. V. and Silva, I. N. (2008). Neural network model for designing automotive devices using SMD LE. Int. J. Automotive Technology 9, 2, 203−210. Prechelt, L. (1995). Adaptive Parameter Pruning in Neural Networks, Technical Report No. 95-009, Int. Computer Science Institute, Berkeley, CA. Reik, W. (1987). Das Zweimassenschwungrad. RWTH Aachen. Proc. 1st Aachener Kolloquium Fahrzeug- und Motorentechnik, Aachen, Germany. Schoeggl, P. and Ramschak, E. (2000). Vehicle driveability assessment using neural networks for development, calibration and quality tests. The Engineering Society for Advancing Mobility Land Sea Air and Space. Proc. SAE 2000 World Congress. Detroit. Zwicker, E. (1999). Psychoacoustics : Facts and Models. 1st edn. Springer Verlag. Berlin, Heidelberg, Stuttgart.