Perception-Based Engineering: Integrating Human Responses into Product and System Design

Product-design engineers should be educated in measuring and analyzing how people perceive noise.

Engineered systems produce both wanted and unwanted products. For example, a computer printer produces text and images but also produces noise in the process. Yet, in the printer-design process, people’s perceptions of the quality of the noise and the quality of the images are often left out of the definition of what constitutes a good printer. In the past, designers have striven to design printers that produced exact replications of original images. In fact, it would be more effective to focus on replicating the characteristics of the image that are important to the observer (Wu et al., 2001).

Once the imaging problem has been solved, designers must address the noise problem. Merely reducing the level of noise may not be sufficient, because even soft sounds may be annoying if they fluctuate (changing sounds tend to attract attention). However, people might like to hear that the printer is operating, so eliminating noise altogether is not desirable either. The sound must meet people’s expectations.

In general, noise control should be focused on reducing the sound attributes that are most annoying or that cause acoustical discomfort and enhancing the sound attributes that people wish to hear. This can be difficult, however, if several outputs or different audiences compete rather than complement each other. In the case of the printer, for example, increased speed may mean increased noise and lower print quality. Or desirable sound for the person awaiting the printed page may be annoying or worse to the person working close to the printer. How should designers make these trade-offs?

People-Sensitive Criteria
Clearly, “people-sensitive” criteria must be included in the design process. However, the development of these criteria is problematic for most designers, because they are rarely educated in measuring and analyzing how people perceive the outputs of their products. Thus, if end users’ opinions are considered, they are usually tested with ad hoc methodologies developed without the help of psychologists, who are trained in making these types of measurements. In addition, testing is usually done so late in the design process that significant changes are no longer feasible. Whereas an integrated design approach might not lead to higher cost, add-on solutions at the end of the process always increase cost.

Perception-Based Engineering
Perception-based engineering is a term coined by professors at Purdue who were working on a proposal that involved a number of existing, collaborative research projects between engineering and psychology, such as automatic speech recognition, control of machine noise, automotive interior design, air quality, touch interfaces, printer design, and highway design. Although at first glance these projects may not have much in common, all of them were focused on getting a better understanding of how people perceive, process, and make decisions in response to intended or unintended stimuli—usually generated by machines or engineered systems. For example, the text and images produced by a printer are intended outputs whereas noise is an unintended output.

In perception-based research, engineers and psychologists work together to develop models of human information processing and decision making. By connecting engineering stimulus-prediction models to perception models and decision-making models, design options can be explored, compared, and optimized to reduce negative impacts and accentuate positive impacts.

Figure 1 is a sample of this model-building process from a study of how combustion variability impacts the sound quality of diesel engines (Hastings, 2004). Chen and Chiu (2004) used a similar model to address the perception of color banding in printed images (stripes that appear in the image in the print direction). Their goal was to design a printer-control algorithm to compensate for imperfections in printer components, thus removing perceptible banding in the printed images.


___________________________

Engineers do not always
use appropriate metrics to
quantify human responses.
__________________________

An understanding of effective human information processing can be extremely helpful in the development of “perceptive” machines—for example, word-recognition machines focused on the automatic detection of nouns and names (Surprenant et al., 1999); studies of multimodal gestures and speech (Harper and Shriberg, 2004); and haptics research being used to develop tactile displays of speech for people with hearing and visual impairments (Israr et al., 2006).

The overall purpose of perception-based engineering research is to integrate the ways people perceive, and are affected by, machinery outputs into the design of engineered systems:

• to understand how machine outputs (intended and unintended) impact people (e.g., comfort, annoyance, task performance and productivity, perception of product quality, etc.) as a basis for developing people-sensitive metrics for product design
• to understand how people process information (tactile, acoustical, visual, etc.) and use this understanding to emulate efficient human information processing to enhance machine performance
• to understand human-machine interactions to de-velop safer, more efficient, more user-friendly interfaces

Perception-based engineering is a component of human-factors research, but the emphasis is on the integration of human-response models into engineering product design rather than on the more traditional interface design, ergonomics, and human performance (Proctor and Van Zant, 1994; Salvendy, 2006). However, the dividing lines are permeable. Human-factors researchers at Purdue are active participants in perception-based research, and performance and productivity, as well as perception and comfort, should all be taken into account in engineering designs.

Metrics for Quantifying Human Responses
Although incorporating human response into product design may seem like common sense, it is often omitted, particularly for products with which users have little direct interaction. Even in products where attention is paid to the user-interface design, human response to unintended stimuli (e.g., noise, heat, vibration, etc.), which directly affect the surrounding environment, is often ignored. When engineers do consider human responses, they do not always use good metrics to quantify them.

Statistics based on A-weighted sound-pressure levels are widely used, even when this weighting is inappropriate because of the very high level of sound. A-weighting should only be applied to relatively quiet sounds (40 dB or less). As a reference, 40 dB(A) is approximately the quietest level of background noise in a city during the day (Beranek, 2005). Humans are more insensitive to lower frequency sounds and very high frequency sounds than to sounds at frequencies in between; but this relative insensitivity is more pronounced for quieter sounds. Thus when A-weighting is used on high-level sounds, low-frequency components are attenuated more than they should be. In spite of this, A-weighting is very commonly used for sound-pressure levels above 40 dB.¹

Another inappropriate criterion is the sum of squares of error (i.e., pixel differences between ideal images and reproduced images) in the evaluation of image quality. If perfection were possible, this might work, but, because there is always some imperfection, using this type of metric does not necessarily lead to better quality images (e.g., a lower error score may produce a poorer image). A different type of metric, such as the image-fidelity assessor developed by Wu et al. (2001), which is based on an understanding of what people look at in images, is a better way to determine where in the image more detailed image processing would make a difference.

Modeling Human Responses
If the modeling of human responses is included in engineering design curricula at all, it is usually in a superficial way. I often get the impression from my engineering colleagues that they do not think human response can be modeled. Even though we all observe people responding in similar ways to similar inputs, when engineers are challenged to incorporate human-response models into the design process, they tend to focus on the variability of people’s responses rather than data that might explain common response characteristics. Because they do not understand the influence of context and individual experiences on the response, they conclude that people’s reactions are “random,” and there is no point in modeling them. (Human-factors engineers are an exception, of course, but they are not involved in the design of most machines and engineered systems.)

In contrast, engineers gladly accept the challenge of modeling highly complex engineering systems, even when they are not certain of the exact mechanisms responsible for the observed behavior. In this context, they are comfortable with the concepts of approximation and variability, and they may develop strategies for reducing the influence of noise based on estimates of parameters from experimental data, account for un-modeled dynamics and environmental conditions, develop experimental strategies to improve their understanding of the mechanisms at play, and incorporate statistical properties into their models. They rarely feel that these significant uncertainties prevent them from building models.

Psychophysical Techniques
An interesting aspect of the work of my colleagues in quantitative and cognitive psychology is psycho-physics and the methodologies used to control and understand biases. From them, I have learned that great care must be taken in designing and executing experiments to prevent uncontrolled biases or confounding variables from obscuring the behavior of interest or its causes. When engineers and psychologists collaborate in their research and learn more about the complexities of human-response testing, engineering students often come away with great respect for their counterparts in psychology.

Although engineering measurements seem straightforward by comparison, in some sense, all test engineers grapple with similar problems (e.g., (un)controllable environments, system memory, external inputs [noise], repeatability, etc.). There are also similarities in the model-building and data-analysis methods used in both fields. For example, an analysis technique used to identify the number of independent sound sources generated by a machine is essentially the same as the technique used to identify the independent sound characteristics people hear when listening to a given set of sounds. This should not be surprising, because both are focused on understanding the underlying structure in the data.

The Acoustics Community
The acoustics community spans many disciplines—physics, music, engineering, human and animal physiology, medicine, sociology, and psychology, and acoustics conferences, which usually attract people from many different fields, can facilitate interdisciplinary interactions. Nevertheless, there often is very little dialogue between groups working on similar problems from different perspectives, even within the acoustics community. And dialogue does not always lead to an integrative approach to solving problems.

For example, research in psychoacoustics on how people judge the loudness of a sound is producing a wealth of knowledge. Models have been developed that incorporate many of the complex frequency and temporal attributes of the human hearing system, and it is now possible to predict the loudness of most sounds very accurately, even sounds that vary in time (Glasberg and Moore, 2002; Zwicker and Fastl, 1998). Reasonably good models have also been developed of how people judge other attributes of sound (e.g., sharpness, roughness, fluctuation, tonality). However, the understanding of sound perception and the corresponding models have not propagated to other parts of the acoustics community.
_________________________________

Research in psychoacoustics
is producing a wealth of
knowledge about how
people judge loudness.
_________________________

The distance of the design engineer from the people who experience the product outputs, including noise, is one problem. Goals set locally may not address the needs of the recipient of the product sound, which, if unwanted, is perceived as noise. An engineering designer working on noise control in aircraft engines, for example, is probably using a standardized metric developed decades ago that does not incorporate advances in our understanding of how people perceive sound. Thus the designer probably does a good job of reducing the metric value to “improve” the sound. The question is whether he or she is using the correct metric.

In this example, the chain of engine-aircraft-airline-passenger or community resident isolates engine-noise engineers, who have a profound influence on the comfort of passengers and community residents. When noise-control engineers make choices between different designs, they don’t always understand the implications of those choices.

Quieter Isn’t Always Better
It has been known for a long time that an A-weighted sound-pressure level can be decreased without making the sound quieter. Sometimes when loudness is reduced, sound components that were previously masked become audible, thus making the sound more prominent, or even seemingly louder (Hellman and Zwicker, 1987).

If these are high-frequency, time-varying, and/or tonal (discernible pitch) sounds, the overall result might be a less desirable sound. For example, reducing fan noise in refrigerators to reduce A-weighted sound pressure may make compressor noise (which is often time-varying, tonal, high-frequency noise) more prominent. Thus the resulting sound may be more annoying than the original sound (May et al., 1996).
_________________________

An A-weighted sound-
pressure level can be
reduced without making
the sound quieter.
_________________________

The automotive industry uses sound-perception and acoustical-engineering models to produce better sounding products (Lyon, 2003). A group of automotive engineers has developed guidelines to help other engineers develop sound-perception and preference models (Otto et al., 2001). Recently, noise-control engineers in other industries have shown more interest in the concepts of sound quality and product-sound design, especially because some noise characteristics can not be measured by the acoustic metrics they have been using.

Engineers are beginning to realize that if a product sounds good, people may think the product is good—and those people may be right! Noise is often a symptom of problems, and changing the sound by changing the design often leads to improving the product in other ways, such as making it more reliable, more energy efficient, and so on.

The product-sound community is slowly adopting a more psychoacoustics-based approach to noise measurement and taking the time to develop an understanding of what people hear and pay attention to in making decisions about products. Some sound engineers are even adopting sound-perception models.

Community Noise
In contrast, the measurement approaches used to assess community noise have changed very little for a number of decades. Although there have been some discussions about the structure of models relating noise to other factors, such as sleep disturbance or annoyance, the sound metrics used as inputs to these models are typically based on statistics of A-weighted sound-pressure levels (e.g., sound-exposure levels [SELAs], and day-night levels [L_dn]). Some researchers have advocated using loudness metrics (Schomer et al., 2001), but not metrics based on the most recent models of loudness for time-varying sounds (Glasberg and Moore, 2002; Zwicker and Fastl, 1998).

Several concerns have been raised about using A-weighted metrics to assess community noise. For example, there have been questions about using 65 L_dn contours around airports for planning, a level which, in the transportation data collected and analyzed by Schultz (1978), relates (on average) to about 15 percent of the population finding airport noise highly annoying. Schomer (2002) points out the high level of variability in the data (90 percent of the data from surveys at 65 L_dn falls between 5 and 28 percent of people being “highly annoyed”) and recommends that normalization methods be used to account for context and sound characteristics; this is very much in the spirit of perception-based engineering. Other issues related to A-weighted metrics are the influence of isolated noise events (to which time-averaged levels are relatively insensitive) and low-frequency noise (which has the potential to cause vibration and rattle).

The value of L_dn that corresponds to a particular percentage of the community being annoyed by noise is different for different transportation sources (Miedema and Vos, 1998). Thus predicting noise impacts in areas where multiple noise sources are present can be challenging. Some argue that knowledge of the noise source influences the response (e.g., people prefer trains to road traffic and thus find train noise less annoying than road noise, even at the same noise level).

Others argue that the differences only occur because Ldn incorrectly accounts for human response to low-frequency noise (Fastl et al., 1996), and that supplemental metrics, such as SELA (time and number of events above a certain level), which are easier to understand, should be used for reporting airport noise. But here again the “level” is A-weighted.

Taken together, these concerns call into question whether community noise is being measured in the best way. Although modeling community responses to long-term exposure to environmental noise is somewhat different from investigating preferences in product sound, I believe lessons learned in perception-based engineering would be helpful in developing appropriate metrics for assessing community noise.

A Mapping Initiative
A large noise-mapping initiative is under way in Europe, but we are not sure that appropriate metrics are being used to predict noise impacts or that the mapping process is flexible enough to update as we gain a better understanding of community response. Other questions have been raised about whether these maps will be interpreted correctly and whether it is possible to capture and convey the main attributes of an acoustical landscape (soundscape) in a visual representation (a map). Answering these questions may require a perception-based engineering project!

Summary
Perception-based engineering is a people-focused approach to engineering design involving both stimulus-prediction (engineering) and stimulus-perception (psychology) modeling. To date, much of the focus has been on unimodal modeling, such as modeling force feedback and force perception in teleoperator applications (Choi and Tan, 2004) or sound quality in diesel engines (Hastings, 2004). Modeling human response to multiple modalities is more challenging, but necessary, to a more holistic approach to design (e.g., making trade-offs between thermal and acoustic controls to maximize occupant comfort in energy-efficient buildings).

Engineers design and build machines and systems that have profound effects, both intended and unintended, on the quality of people’s lives. Consider, for example, machines in intensive care units in hospitals that help save lives but may also disturb sleep, a key element in patient recovery. Engineers must take into consideration all of the impacts of their designs, and perception-based engineering is one discipline that can help them to do this.

References
Beranek, L.L. 2005. Criteria for Noise and Vibration in Communities, Buildings, and Vehicles. Pp. 887–909 in Noise and Vibration Control Engineering: Principles and Applications, 2nd ed., edited by I.L. V?r and L.L. Beranek. New York: John Wiley & Sons.
Chen, C-L., and G.T.-C. Chiu. 2004. Incorporating Human Visual Model and Spatial Sampling in Banding Artifact Reduction. Pp. 2642–2647 in Proceedings of the 2004 American Controls Conference, Vol. 3. New York: IEEE.
Choi, S., and H.Z. Tan. 2004. Perceived instability of virtual haptic texture. I. experimental studies. Presence: Tele-operators and Virtual Environments 13(4): 395–415.
Fastl, H., S. Kuwano, and S. Namba. 1996. Assessing the railway bonus in laboratory studies. Journal of the Acoustical Society of Japan (E)17: 139–147.
Glasberg, B.R., and B.C.J. Moore. 2002. A model of loudness applicable to time-varying sounds. Journal of the Audio Engineering Society 50: 331–342.
Harper, M.P., and E. Shriberg. 2004. Multimodal integration for sentence unit detection. Proceedings of the 6th International Conference on Multimodal Interfaces, State College, Pa., October 2004. New York: ACM Press.
Hastings, A. 2004. Sound Quality of Diesel Engines. Ph.D. Thesis, Purdue University, West Lafayette, Ind.
Hellman, R., and E. Zwicker. 1987. Why can a decrease in dB(A) produce an increase in loudness? Journal of the Acoustical Society of America 82(5): 1700–1705.
Israr, A., H.Z. Tan, and C.M. Reed. 2006. Frequency and amplitude discrimination along the kinesthetic-cutaneous continuum in the presence of masking stimuli. Journal of the Acoustical Society of America 120(5): 2789–2800.
Lyon, R.H. 2003. Product sound quality: from perception to design. Sound and Vibration 37(3): 18–23.
May, P., P. Davies, and J.S. Bolton. 1996. Correlations between Objective and Subjective Evaluations of Refrigerator Noise. Pp. 2257–2260 in Proceedings of Internoise 96, Vol. 5, edited by F.A. Hill and R. Lawrence. Liverpool, England, July 30–August 2. St. Albans, U.K.: Institute of Acoustics.
Miedema, H.M., and H. Vos. 1998. Exposure-response relationship for transportation noise. Journal of the Acoustical Society of America 104(6): 3432–3445.
Otto, N., S. Amman, C. Eaton, and S. Lake. 2001. Guidelines for jury evaluations of automotive sounds. Sound and Vibration 35(4): 24–47.
Proctor, R.W., and T. Van Zant. 1994. Human Factors in Simple and Complex Systems. Boston, Mass.: Allyn and Bacon.
Salvendy, G. 2006. Handbook of Human Factors and Ergonomics, 3rd ed. New York: John Wiley & Sons.
Schomer, P.D. 2002. On normalizing DNL to provide better correlation with response. Sound and Vibration 36(12): 14–25.
Schomer, P.D., Y. Suzuki, and F. Saito. 2001. Evaluation of loudness-level weightings for assessing the annoyance of environmental noise. Journal of the Acoustical Society of America 110(5): 2390–2397.
Schultz, T.J. 1978. Synthesis of social surveys on noise annoyance. Journal of the Acoustical Society of America 64(2): 377–405.
Surprenant, A.M., S.L. Hura, M.P. Harper, L.H. Jamieson, G. Long, S.M. Thede, A. Rout, T.-H. Hsueh, S.A. Hockema, M.T. Johnson, P. Srinivasan, C.M. White, and J.B. Laflen. 1999. Familiarity and pronounceability of nouns and names. Behavior Research Methods, Instruments, and Computers 31(4): 638–649.
Wu, W., Z. Pizlo, and J.P. Allebach. 2001. Color Image Fidelity Assessor. Pp. 148–152 in Proceedings of the Image Processing, Image Quality, Image Capture Systems Conference (PICS-01). Montreal, Quebec, Canada, April 22–25, 2001. Springfield, Va.: Society for Imaging Science and Technology.
Zwicker, E., and H. Fastl. 1998. Psychoacoustics: Facts and Models, 2nd ed. New York: Springer-Verlag.



FOOTNOTES

1. In the early days, there was A-, B-, and C-weighting. B-weighting is rarely used today. C-weighting—which is basically flat-frequency weighting—is still used for some sounds.