Towards Sound Quality Engineering of the Virtual Car

The use of sound quality engineering tools in automotive development is well accepted - but their application is mainly restricted to the refinement and troubleshooting of physical prototypes. Currently, designers have few tools to help them listen to their ‘virtual’ models. Imagine what could be achieved if an awareness of sound quality could be implemented upfront. The concept team could use a joystick to listen to the engine note as it goes through an arbitrary speed change. Program engineers could modify an engine mount thickness to tune the car for improved sportiness. To do it all accurately, reliably, and interactively would dramatically change the whole approach to NVH engineering. Over the last few years LMS has been working on new hybrid technologies to enable just that...

As its name suggests, the scope of the recently completed international research project OBELICS (OBjective EvaLuation of Interior Car Sounds), was to develop an understanding of what exactly is involved in getting the sound of the car “right”. This is not as simple as it sounds: the driver’s perception of sound is governed not only by what is heard, but by many other prominent or subtle factors, including the current road speed v throttle position, the load on the engine, and even visual stimuli. One of the fruits of OBELICS was the development of a prototype sound quality tool that would enable the designer to change a parameter or moderate the throttle/rpm profile and listen to the results. Breakthroughs in not just one but three areas are required: time domain transfer path analysis, virtual sound modeling, and ‘driver-in-the-loop’ jury testing. This article focuses on the first two aspects.

Time Domain Transfer Path Analysis

Frequency Domain Transfer Path Analysis (TPA) is accepted as a tool to model or troubleshoot the lower engine orders for in-vehicle noise. The contribution of each excitation force coherent with the engine can be evaluated in terms of both the amplitude and phase of sound field partial pressures at the driver’s ears. This allows ranking of the individual paths and the identification of critical elements in the design. Usually, such analysis is made on a stationary or slowly varying speed change, where the goal is to accurately characterize the pathways, rather than the sound quality perceived by the driver. However, as is generally well known, there is little difference between a diesel and an ordinary engine in spectral content, but the perception of sound quality is very different. A formulation in the time domain is required.

Previous attempts to develop Time Domain TPA using the time domain formulation directly all suffered from the fact that the impulse responses of all noise transfer paths were needed. These functions are typically only accurate up to about 500Hz and have a very complex amplitude and phase behavior, especially at the higher frequencies. Even high order FIR and IIR filters are not able to reproduce such complex shapes without important phase distortion - and even relatively small phase distortions between orders can result in a different perception of the sound quality, especially the roughness. Long FFT approaches were considered to perform the convolutions. The computation times that are needed are considerably lower than those for time domain filters.

The main problem of the use of FFT is the fact that the method assumes that the signal is periodic, which is, by definition, not the case for a pseudo-stationary test condition such as an engine run-up. While listening to the filtered result, amplitude oscillations could occasionally be heard just before important increases in amplitude.

For the majority of cases, a good perception of the in-vehicle sound not only requires information of the lower orders, but also the high frequency noise (wind, road- and engine induced), and the other non-harmonic components. Therefore, a method was investigated that would not disturb the overall sound while substituting the existing order content by the desired order function. Keeping amplitude, phase, rpm and crankshaft angle synchronized in such a process is a major challenge. Next, this method had to be compatible with the partial pressure results of a frequency domain based Transfer Path Analysis tool. This way, the more advanced inverse force identification methods remain available for the time domain TPA process. The order substitution technique has the following steps:

A series of difficulties needed to be tackled to make this ‘simple’ procedure operational.

Frequency Domain TPA requires correct ‘relative phase’ for each order between the paths. Time Domain TPA requires exact phase between paths and orders: ‘absolute phase’ is required. Of course, from the absolute phase of the reference signal, all absolute phases can be calculated. However, the extraction of orders with absolute phase from the reference signal proved to be non-trivial, and several approaches were tried and rejected. Although FFT is very adequate for phase differences between signals (even for non-stationary conditions), the results cannot be used for the determination of absolute phase - especially between orders. Adaptive resampling into the angular domain in combination with triggered DFT was investigated with success. However, it requires a non-negligible computation effort since the over-sampling has to be extremely high to limit phase distortion and was ultimately set aside. Currently triggered tracked heterodyning filtering is used. Another important benefit, the filtered results are obtained in seconds rather than minutes.

The approach we adopted was to use specially developed double-precision tracking filters with an 0.01 order width, an attenuation exceeding 20dB, and an out-of-order phase distortion less than 101. Of course, ultra-sharp filters are typically not able to follow fast changes in amplitude of the component that is being filtered, but this is mainly a positive aspect, since the side-bands resulting from modulation on the orders are not affected in the filtering process. This is useful when trying to reproduce Roughness and Fluctuation Strength adequately, and allows the modulation character to be influenced or even be removed. The Kalman filter with the negligible phase distortion is an alternative, but the computational effort is substantially higher.

The partial pressure Frequency Domain TPA results have relative phase with respect to the reference signal. Their phase is corrected with the absolute phase of the reference signal and a sine sweep is generated for each order with the exact rpm, amplitude- and phase evolution. Finally, the synthesized sweeps are inserted back into the (filtered) background signal. Re-substituting the new orders in a phase accurate way is more complex than it seems: any timing errors with respect to the remaining order content would be easily heard as audible artifacts such as chirps, clicks and burrs, which would lead to an artificial impression of ‘roughness’. The trigger signal used to obtain the absolute phase of the reference signal in the angular domain determines the possible start trigger points. Then, the insertion has to be exactly at the trigger point that corresponds to the start rpm of the filtered target signal.
Now both original and synthesized signals are available for evaluation and direct comparison.

The next stage is to develop an off-line tool that would allow the evaluation of sound quality for a virtual car. Again, this is not as simple as it would seem: that the vehicle is accelerating, decelerating, or coasting clearly has a major impact on the engine note, and currently such work lies outside the capability of ‘traditional’ sound quality systems. The support of arbitrary load and speed variations requires a very sophisticated model comprising the vehicle dynamics and the load and rpm parameters as inputs.

The option was taken to describe the complete throttle domain using engine accelerations from idle and decelerations from full speed run-ups for a limited number of fixed throttle positions. Currently, only run-ups in a fixed gear are considered. This Design of Experiments approach results in 3D surfaces for each order (amplitude and phase). Next, the same strategy was applied for calculation of the background noise - following elimination of the order content. Interpolation is used for both throttle and rpm to obtain the instantaneous order amplitudes and Bark noise spectra.
In our experience, interior automotive sounds can be characterized by approximately 32 orders, and a pseudo-stationary background (mainly aerodynamic noise and road noise) parameterized in 23 Bark bands (a better subjective model than one third octaves) as a function of rpm.

A simplified General Vehicle Model was built that calculates the rpm evolution starting from the current speed taking into account the torque of the engine (as a function of engine speed); the wind resistance (as a function of speed); the roll resistance (proportional to the vehicle mass); and the “climbing force” due to the slope of the road.

The engine torque and wind resistance is determined from known vehicle characteristics and fine-tuned with the acceleration and deceleration experiments, using the simplified dependencies of these parameters on rpm, speed and mass. The model showed to be sufficiently accurate for the experiment.

Starting from the synthetic rpm/time function obtained from the General Vehicle Model, the time waveforms of each target order are synthesized taking into account the phase-, amplitude-, and frequency-profiles as a function of time. Similarly, the background noise is synthesized by filtering white noise according the background model and the target rpm/time function. The time histories for the 32 orders and 23 random backgrounds are then summed and prepared for binaural replay.

Tests on a 4-cylinder mid-segment family car were made to validate the prototype sound synthesis model. The car was tested in third gear, on a roller bench in a semi-anechoic chamber, for a limited number of fixed throttle positions: 10, 20, 30, ... 90, 100% throttle acceleration; 10, 20, 30% deceleration; and idle rundown. The driver then performed a number of free accelerations and decelerations. Next to the sound measured with a binaural head placed on the passenger seat, rpm and throttle position were recorded.

The processing of the raw time data of the different conditions into a sound synthesis model was performed and the vehicle parameters for the General Vehicle Model determined. Finally, one of the non-stationary throttle runs was selected for sound reproduction.

The comparison of the actual rpm evolution with the outcome of the GVM showed that the model was sufficiently accurate: differences between real- and computed rpm could be kept within reasonable limits. The second step was to create the sound from the given throttle and the computed rpm curves. All orders and Bark bands are synthesized individually and summed. The resulting sound was then compared to the original sound. Both sounds sound very similar, only small differences can be heard, mainly in the composition of the noise, especially below 1500rpm. The calculation of acoustic and psycho-acoustic metrics supports this first subjective analysis.

The results of the experiment are very promising and a major step forward has been achieved. Early sound synthesis prototypes, used by several car manufacturers, have received very positive reviews. The Time Domain TPA module will be released as part of LMS CADA-X 3.5.C. The next steps are: the development of a real-time version of the sound synthesis model; an improved understanding of road- and wind noise in order to disconnect rpm from vehicle speed and consequently support gear shifts and automatic gearboxes; and the integration of source/ transfer/receiver models to come to the ultimate goal - Virtual Car Sound Modeling.