M. d’Ischia, A. Concilio, A. Sorrentino
CIRA S. c. p. A. - The Italian Aerospace Research
Centre
Via Maiorise, Loc. Silvagni, 81043 Capua (CE),
Italy
Presented at the 4th European conference on Noise Control EURONOISE 2001, 14 - 17 January 2001, Patras, Greece. © EURONOISE.
Keywords : artificial neural network (ANN), comfort, noise, psycho-acoustics, vibration, virtual passenger.
This paper deals with activities carried out in IDEA PACI, a Brite-EuRam funded project. In terms of general performance, air transport technology is very advanced. Due to the increased use within shorter and shorter distances with more and more frequency, not only environmental impact, but also the internal, particularly vibro-acoustic, comfort has to be considered in the aircraft design process.
The tests showed a complex link between noise and vibration levels. One of the most significant and surprising results of the application of pure noise abatement systems, was that people gave a preference for high instead of low noise if at high vibration levels. In that case, lowering interior noise levels decreased comfort. The passenger welfare inside the aircraft is therefore a complex function of both noise and vibrations (not to cite other surrounding characteristics); as a result, the design objective was moved, coherently, towards comfort improvement from noise suppression.
The target of this basic research project was to define a general Scalar Subjective Comfort Index for commercial aircraft, and to set-up a numerical tool able to output that number by means of information on the external solicitation field (both sound and vibration). This last step is accomplished by creating a "Virtual Passenger" on the basis of the development of an Artificial Neural Network, aimed at simulating the transfer function between the stimuli (physical variables) and the generic response of the passenger (subjective impression). Both acoustic and vibration data were taken into account.
Once the software was ready, a first application concerning
its use as a design tool was carried out. Data obtained from finite element
numerical simulations of a typical fuselage section were used to extrapolate
the structural configuration that would have given rise to the most comfortable
response. A parallel study, aimed at defining the boundaries of the structural
parameters such that a certain comfort threshold was guaranteed, completed
the investigation.
To assume that the passenger welfare inside the aircraft is a complex function of both noise and vibrations, apart of other surrounding characteristics, is almost trivial. Also, to imagine that the design objective is to improve the comfort and not to suppress noise, seems obvious. Because comfort is a subjective matter, psychological studies play a fundamental role [2]. The objective of IDEA PACI is to establish an aircraft passenger noise and vibration comfort index to relate subjective perception to mechanical and physical design properties. This is accomplished by means of an Artificial Neural Network (ANN), meant to simulate a Virtual Passenger, i.e. the generic transfer function between external stimuli and human impressions. Though aimed at interior aircraft noise, the results enable other transportation means to improve general features of comparable construction and developing processes, avoiding or substituting expensive experimental tests by the use of ANN.
The main steps of the research may be summarised as follows: first of all, a psycho-acoustic study is carried out to identify suitable descriptors and to develop a tool that allows translating the environmental solicitation in subjective impressions. Then, experimental investigations both at ground and in flight are performed to produce a wide data-base for statistical investigations. On the bases of the acquired results, the architecture of the ANN is defined: it is successively trained and assessed on the experimental data coming from the main test campaign, so setting the aforementioned “Virtual Passenger”. As the last step, ANN finally predicts comfort levels on both experimental and numerical (FEM) data: the physical parameters that mainly affect the passengers' welfare during the flight are identified through proper simulations.
Being the project target, the identification of the relationship among
the physical and the modelled environment and the human response, psycho-acoustic
studies, numerical applications, experimental results and ANN definition
are strictly interdependent.
After, a comprehensive questionnaire, oriented to the passenger comfort inside aircraft cabins, has been designed by the University of Oldenburg (D). Jet, propeller or helicopter aircraft are targeted. Both vibration and noise disturbance are taken into account. Standard methods of evaluating test-subjects answers (factor analysis, paired comparison analysis, etc.) yield a set of psycho-acoustic parameters being identified as the most important descriptors contributing to a comfort index.
In order to have a sensibility about the test environments, the psycho-acoustic team (University of Oldenburg, ITAP - D, and NLR - NL) performed the so-called pre-tests, from which a guideline is defined for the identification of a proper questionnaire, starting from classical formulations. The differences among the used mock-up’s, as well as the difference between the available test articles and the real aircraft could, in fact, affect the subjective impression, in a manner or another.
Moreover, in the first part of the project, a qualitative analysis has been performed to relate the psycho-acoustic descriptors to noise and vibration fields; in fact, psycho-acoustic parameters are easily correlated with acoustical and signal-related quantities like spectral parameters, signal envelope and other time-dependent features.
The characteristics of the exterior disturbance signals are then related
to the subjective impression, directly, in a qualitative way. The problem
is solved in two steps. Structural acoustics is concerned with the relationship
between the exterior noise source and the produced interior noise field;
in detail, its objective deals with the reduction of the sound (and vibration)
field in enclosures. Psychological acoustics is instead concerned with
the effects of the interior sound field on the human comfort feeling; its
objective deals with the minimisation of the noise annoyance. Both the
experiences are combined, in order to address better the aircraft design
“to noise”.
Acoustic and vibration measurements and the passengers’ response evaluation to the correspondent environment are carried out, both in-flight, on a helicopter and at ground, on a helicopter cabin and two fixed-wing mock-up's. The results of this experimental campaign support the definition of a Scalar Comfort Index as a function of aircraft vibration and interior noise fields.
Pre-tests have been performed in order to take confidence with the test facilities planned to be used in the main test campaign, in view of preparing the questionnaire to be used during the psycho-acoustic tests. The test facilities have been provided by Alenia (I), Agusta (I) and Dornier (D), complete with interior furnishings and all the other necessary arrangements, including the noise and vibration excitation sources. Flight tests data, recorded in previous campaign, have been collected, stored and selected for the experiments.
The pre-tests made possible to test “on the field” the proposed methodologies. The performance of the pre- and the main tests are essentially devoted at determining the related physical and psycho-acoustical data. Here, for experimental psycho-acoustic data, the one are intended, coming from the evaluation of the questionnaire as submitted to the group of people under the defined excitation, in-flight or in the mock-up. In this way, two sets of records result: a first, describing the physical environment, and a second, defining the related human response. The aim of the “Virtual Passenger”, is to find the relationship (or transfer function) between these two general parameters. A data-base is then built, on the basis of which the ANN is trained and validated.
From a preliminary evaluation of the data, available at the end of the
pre-test campaign, it came out that would have been difficult to have a
sufficiently large experimental basis, so that all the possible configurations
were covered and just an ANN were defined, for all the types of aircraft.
In other words, the produced data-sets contain a limited amount of physical
psycho-acoustical correlated quantities; furthermore, dealing with restricted
variations of cabin environments (apart the noise-vibration signals, in
fact, an essential remark stands on the fact that the mock-up’s number
is fixed to three).
Statistical considerations led the choice of the minimum number of people to be involved in the experiments. After the preliminary test campaign, the proper procedures at the different mock-up’s and real aircraft, were issued.
The ATR42 fuselage section N.13 is used as test article during the experimental campaign, following illustrated. It is made of 6 bays and is 3.5 metres long. It is elastically suspended at the extremities, by means of steel springs. Two caps close the cylinder. The fore one is movable, and through that side the internal access is permitted, through a small staircase. The fuselage section is completely furnished, in the same way as the corresponding flying aircraft.
Figure 1: inside view of the furnished ATR42 mock-up (courtesy of
Alenia)
The ATR42 mock-up is prepared with both external and internal instrumentation, in order to get the data concerning the outside and inside excitation, and the interior vibration-acoustic field. The tests have been performed at the plants of Alenia in Pomigliano D’Arco (I), where the mock-up is placed. Fig.1 shows an outside view of the cabin section. In Fig.2 a complete external and internal view of the mock-up is shown, together with all the used instrumentation. Noise and vibration field are acquired in correspondence of the 12 seats, installed inside the mock-up; in particular for each place, a microphone is placed at the passenger head height and a three-axial accelerometer is positioned at his feet.
Psycho-acoustic tests are carried out by simulating the real flight conditions for different aircraft types (propeller, jet). The internal vibration-acoustic field is not homogeneous at each seat as it was evidenced during a great amount of experiences [3]. The main tests performance was scheduled in 5 main parts:
Figure 2: Scheme of ATR42 mock-up with relative instrumentation
(courtesy of Alenia)
Figure 3: Agusta 109 (courtesy of Agusta)
The accelerations along the 3 axis, together with the noise level inside the cabin, were measured during the tests. During the flight tests the University of Oldenburg and ITAP used proper instrumentation for binaural measurement evaluation. The test performance is structured according to the same schedule of the tests at Alenia; it holds also for ground tests.
Oleo-dynamic shakers are used as vibration exciters, because at very low frequencies and high forces, they yield lower noise than the electro-dynamic ones. For the acoustic excitation of the mock-up, 3 couples of loudspeakers were used. In order to verify if the acoustic and vibration environment inside the mock-up cabin were representative of the real flight, some accelerometers in the cabin floor and microphones were installed at the seat. Before each test begins, the sound and vibration field was verified to be sure the passengers were exposed to the same conditions.
Figure 4: A109 Airframe Reference Lines (courtesy of Agusta)
The vibro-acoustic field inside the mock-up is obtained using the available signals recorded in flight. The loudspeakers simulate the acoustic field while the shakers the vibration environment. All the noise and vibration sources in a helicopter are located overhead the cabin. Therefore the shakers and loudspeakers are installed overhead, where the main gearbox and rotor are located. The number of the excitation directions depends on the results of the pre-tests.
Configurations simulating jet aircraft flight and propeller aircraft cruise flight were presented to the subjects. In order to exclude sequence effects from the test, the presentations were randomised for each group of subjects. Hence, the signals used for generating the different configurations were recorded prior to the test and combined into different sequences.
For the listening test the different sequences are combined into different runs with the effect that each group of subjects was presented a different sequence of configurations during one run.
All measured signals at the seats with and without subjects were used for calculation of physical and psychoacoustic parameters and generation of multispectra. Both types of signal representations are used for training the ANN by respective partners. Calculated objective parameters are: sound level [dB], A-weighted sound level [dB(A)], loudness [sone], sharpness [acum], roughness [asper], fluctuation strength [vacil] and prominence ratio [dB]. These were evaluated directly from vibroacoustic field measurements through standard procedures. Corresponding subjective parameters are instead calculated by the questionnaires, by considering statistical techniques.
Also in this case, the test performance is structured according to the
same schedule of the tests at Alenia and Agusta.
Given the complexity of the relationship between these two classes of data, the chosen mathematical technique for the comfort simulation is based on Artificial Neural Networks. These tools are built to mimic the behaviour of the human brain and its ability to learn, by sharing the information among their basic elements (the electronic neurones). In this way, the characteristics of non-linearity and complexity of the simulated transfer function can be duplicated by adopting suitable network architectures.
Regarding the reliability of this technique, there is a full mathematical proof of the fact that multilayer neural networks are a class of universal simulators [6]. Given the type of the problem to be solved, i. e. the identification of a system, the three partners involved in the third task (CIRA, ONERA - F - and University of Patras) use this kind of network architecture to reach the solution. CIRA employs a three-layer network (a multilayer perceptron) trained with a standard back-propagation algorithm, while ONERA and the University of Patras use Radial Basis Function Networks.
In both cases, the network training procedure is a pure supervised learning, so several couples of input/output data must be given during the learning phase in order to update the network parameters (i. e.: the connection weights) and to define the final network structure. The main problem of a direct application of the procedure lies in the large number of input data and in the limited, and heterogeneous, number of experimental data.
The complete multispectra, indeed, define the power distribution of the vibro-acoustic fields in a given frequency band and its fluctuation in a given time interval. This means that the number of data values is relevant if compared with the number of tested conditions. Given that a large size of the input vector implies a big size of the network architecture and a relevant number of connection weights, it is clear that the learning algorithm cannot evaluate their correct values if there are few known conditions.
Furthermore, the fact that the experimental data come from different aircraft types can create problems during the training phase if the data distribution in the input and output space is too complex. For this reason, an analysis of the structure of both the input and output space can give important indications regarding the main features of the problem and the procedure to adopt for the definition of the network.
The input space is the subset of the real space in which every point represents a given recorded flight condition; since a complete multispectrum is very hard to analyse in this way, a clearer representation can be obtained by taking into account the objective psycho-acoustic parameters. These comfort indicators are evaluated directly from the physical data by using standard numerical procedures, so they can be used to characterise the spectra completely (at least from the point of view of vibro-acoustic comfort).
For this reason, the input space will be defined as a sub-domain of the R5 space in which the co-ordinates of every point are given by the objective psycho-acoustic parameters. A 2D representation of it, involving loudness and sharpness only (obtained from Alenia and Dornier tests), is given in Fig. 5.
Figure 5: input patterns classification
In this diagram the different data from different types of aircraft are clearly clustered in different areas of the domain, these areas are partially superposed too
The same consideration, on the other side, does not hold for the propellers; in these cases, the differences of comfort characteristics of the vibro-acoustic fields are much more evident and, even if a clustering for the same aircraft type still exists, the different clusters are clearly separated.
The output space can be defined and represented in a similar way. The co-ordinates of every point are given by the subjective psycho-acoustic parameters (evaluated on the basis of the questionnaires). A 2D representation of it, involving loudness and sharpness only (obtained from Alenia and Dornier tests), is given in Fig.6.
Figure 6: output patterns classification
In this diagram too, it is quite evident that there is a clustering similar to the one observed in the previous case. Here, with the only exception of the Alenia propeller data (that are much more dispersed), the representative point for the same aircraft type are clustered together.
This is a clear indication that it would be a sensed choice to define different virtual passengers for different aircraft; the coherent structure of both the input and output spaces gives homogeneous domains in which the Neural Network can operate the correct mapping.
Of course, the problem of the input vector size remains still open,
and it must be faced in order to have a good reliability of the simulations.
This reduction is obtained with suitable, purely numerical pre-processing
of data and/or with the formulation of suitable assumptions on their structure,
starting from physical considerations.
These considerations give the guidelines for the development of the last task of the project. The achieved goals are related to the analysis of the physical parameters that mostly affect the Comfort Index (CI) and the definition of the relationship between the CI boundaries and the input domain. In this way, the Virtual Passenger is integrated in a numerical tool aimed to the design to comfort.
The first part of the task is aimed to the employment of a FE model of Alenia ATR42; the calculation provides the physical vibro-acoustic fields for each seat and for different values of structural characteristics. The power spectra are computed up to 300 Hz because of the reliability limits of the calculation technique. For higher frequencies to adopt a different approach is mandatory, like the Statistical Energy Analysis (SEA).
The results of this step are the basis for the sensitivity analysis; here, the “quasi” optimal power spectrum is defined by pointing out the best set of structural parameters. In this way, a set of design indicators can be given for the reaching of the best comfort level.
The final step is related to the constraint generation: it consists,
mainly, in the definition of an inversion procedure for the ANN. By doing
so, to define unknown power spectra that correspond to CI below a certain
threshold is possible. In this way, it will be possible to point out the
structural parameters set in such a way to obtain optimal CI for a given
aircraft configuration.
All the tests, that were carried out, allow to point out, in a detailed manner, the different parameters that contribute to define the comfort status of a civil aircraft passenger, at least from the point of view of vibro-acoustic solicitations. Of course, the fact that the aircraft tests were carried out in ground mock-ups is a limitation for their complete reliability. Furthermore, given that the personal comfort is a complex function of many other variables, by taking into account the vibro-acoustic field only, the problem is greatly restricted.
But, apart of these experimental details, it is important to note that this first approach is widely opened to many further developments. This is quite evident from the point of view of the experimental activity (e. g. the involvement of airliners too in flight experiments) as well as from the point of view of the involved parameters (e. g. by taking into account the light distribution).
Regarding the definition and the assessment of the Virtual Passenger, the employed Neural Networks show a good effectiveness in “catching” the complex mapping between the physical data and the subjective comfort level. This is an important tool that can be used by aircraft manufacturers to add a new discriminating variable to their products, but not for them only; the knowledge of the Comfort Index, indeed, can be very important from the point of view of the marketing and for the airliners too.
The last part of the project gives other important indications for the
manufacturers. The activity developed, indeed, allow knowing the effect
of some structural changes on the Comfort Index. Furthermore, it is also
shown the way to define the structural design variations that are required
to obtain good comfort levels.
Authors wish to thank Prof. V. Mellert, Prof. B. Schulte-Fortkamp and Dipl.-Psych. J. Quehl (University of Oldenburg – D) for the preparation of the psychological investigation tools, the statistical evaluation of psycho-acoustic descriptors and Dipl.-Phys. H. Remmers (ITAP – D) for the physical data processing.
Nice thanks are given to Prof. D. Tsahalis and all the team at the University of Patras (GK) for their activities about the ANN development and Dr. P. Bourret (ONERA – F) for his precious advises concerning the set-up of those tools and the efforts produced in the sensitivity analyses.
Regarding this same point, the authors wish to recognise the contribution given by NLR. Particularly the support of Mr. G.D.R. Zon, concerning the organisation of the project, Mr. R. Maas for the assessment of the constraint generation procedures and Mr. D. van Touw, that accompanied the first development of the activities.
The authors express their vivid appreciation to the contribution provided by DaimlerChrysler Aerospace Dornier, through the people of Dr. I. U. Borchers, for his logistic supervision throughout the project, Dr. H. Prante for his support concerning the experimental plan and tests and Dr. R. Drobietz that accompanied the activities till the natural end.
A special acknowledge to Mr. F. Cenedese and Mr. M. Maggiore of Agusta (I), for their valuable involvement that produced a remarkable test campaign, including flight tests. The video recording of the experiments in Cascina Costa di Samarate (I) will remain as one of the most significant witnesses of the effort produced by the partnership.
The colleagues at Alenia (I) namely Mr. A. Paonessa and Mrs. A. Pezzolla, are gratefully cited, for their devotion to the objectives of the project, the experimental campaign plan and organisation and the operative support for the FE analyses.
Last, but not least, special thanks are given to the colleagues at CIRA,
Mr. L. De Vivo and Mr. P. Vitiello for the effort provided in the numerical
analyses concerning the ATR mock-up.