Master-thesis project ideas
Loudspeaker modeling and control
Time-varying nonlinear modeling and estimation
The behavior of many dynamical systems changes with time, e.g. loudspeaker, tube amp, tube compressors. Explicitly model system as a time-varying system, e.g. for loudspeaker \(R_e(t)\) and \(K(t)\) are time-varying.
- Design experiments and build data set that shows parameter drift when analyzed with time-invariant models.
- Propose model for explicit time-variation.
- Simulate time variation and compare with measurements for plausibility of the model.
- Develop offline estimation routines for estimating time-variation using Kalman-filters (Sarkka 2013; Simon 2006) and their nonlinear extensions.
- Benchmark models against data set.
- Develop real-time estimation routines, e.g. in C.
Flux instead of current in loudspeaker state
The physical quantity behind energy in the magnetic field is the magnetic flux, not the current in the voice coil. Can this inside be used in modeling?
- build loudspeaker with additional coil to measure change of flux directly
- design experiments measuring flux variation
- update or reformulate model with \(z=[\Phi, x, v]\)
- eddy-currents
- Cunningham-force
- compare measured change of flux with model predictions
- update model with insights
Gathering a high-quality, extensive data set of nonlinear behavior in loudspeakers
Modern data-driven methods need high quality data-sets for training and benchmarking.
- design and conduct high-quality experiments measuring different drivers under different conditions/input signals/input levels extending Franz M. Heuchel and Agerkvist (2022)
- implement and test some simple baseline models
- build a website and documentation on how to use the data
- get researchers interested in the modeling challenge
Discrepancy modeling
Mechanistic models seem only to get us to \(\sim5\%\) error. Model that discrepancy with nonparametric method (neural network, Gaussian process, etc.): \[ \dot x = f(x:\theta) + g(x;\theta)u + NN(x, u;\theta) \]
- gather data-set building ontop of Franz M. Heuchel and Agerkvist (2022)
- linearizable network architecture
- develop fitting procedure, e.g. multiple shooting, linear\(\rightarrow\)nonlinear\(\rightarrow\)discrepancy
Neural network architectures for linearizable transducers
We desire accurate models that are feedback linearizable.
- develop neural architectures for transducers that are feedback linearizable (Delgado 1996; Hache et al. 2022; Ellacott, Mason, and Anderson 1997, 8:181) or/and stable (Dahdah and Forbes 2022).
Learning parameters from linearization error
Linearization is one of the final goals of loudspeaker modeling. Learning loudspeaker that minimize the linearization error thus should give us the best model for that purpose.
- design experimental setup for online-parameter estimation similar to Klippel
- develop fitting routines for online fitting based on linearization error
- compare linearization performance to classic approach
Sound field analysis, synthesis and control
3D Sound field control: minimizing acoustic intensity or similar over a 2D microphone array
In sound field control, we would like to “shield off” entire 3D domains, e.g.houses, heads. Using sound propagation models for such control is useful (Franz M. Heuchel et al. 2020), but requires explicit modeling of loudspeakers. Could one use measurements over a 2D acoustic array and extrapolate from there to virtual microphone positions instead?
- understand and implement a feasible sound field analysis method
- reconstruct the sound field at virtual microphone positions
- develop a (simplified) theoretical framework for this application relating dimensions of source arrays, microphone arrays and distances to quite zone size
- develop a simulation of sound field analysis and the control approach
- experimentally verify the approach
- (optional) develop a real-time analysis and control approach
Bayesian sound field control (in combination with sound field analysis)
Choosing regularization parameters is tricky for sound field control. A Bayesian formulation would give a rock solid method for regularizing sound field control problems using uncertainty in the data. This is related to methods for robust optimization.
- investigate and compare approaches for computing sound-field control-filters that are robust to the uncertainties in transfer-function estimates
- develop a probabilistic sound field model, e.g. based on plane waves
- reconstruct probabilistic estimates of the sound field at virtual microphone positions
- combine the sound field reconstruction and control approaches
“Transparent sound portal”
Two people standing in two rooms in two different buildings. In each room, a “sound portal” of the size of the wall (with a 2D loudspeaker and 2D microphone array) connects the two rooms, such that the two can acoustically communicate with each other with the feeling of having the two rooms transparently connected through the portal.
- develop a sound field analysis and reconstruction method for this use case using sound field separation techniques (Franz M. Heuchel et al. 2018)
- test the method in simulations
- test the method by implementing it
- investigate how transparent the connection is by investigate the influence on room room acoustics (are the rooms acoustically coupled?)