Blind estimation of room acoustic parameters and the performance improvement based on information fusion using Dempster-Shafer theory

Modeling and estimating states or parameters of a dynamic system are necessaryfor monitoring and controlling a system since direct measurement of some systems mightbe difficult. The accuracy of the estimation is thus crucial in such cases. Therefore, thisthesis investigates estimation techniques for the state or parameters of dynamic systems bycombining information from multiple sources or models. The Dempster-Shafer theory (DSTor theory of evidence), an information fusion framework, is employed for this purpose. It allowsus to assign the degree of belief for each source (data-level fusion) and model (decisionlevelfusion). The investigations were conducted on two applications: (1) estimation of roomacoustic parameters and (2) state-of-charge estimation of a lithium-iron-phosphate battery.Firstly, this thesis proposes two approaches for blind estimation of parameters ofa room-impulse response (RIR) model. Both methods are based on the concept of the modulationtransfer function (MTF) and use a stochastic RIR model, namely the extended RIRmodel. The RIR model is used to derive room-acoustic parameters related to the quality ofthe sound field and hearing perception, such as reverberation time, clarity index, and speechtransmission index (STI). Since obtaining the RIR when the sound field is occupied is difficult,estimation from only an observed signal, called blind estimation, becomes necessary.In the first method, convolutional neural networks are utilized to estimate the RIR model’s parameters and the signal-to-noise ratio on seven-octave bands, and the second approachanalyzes the characteristics of observed power envelopes from the inverse RIR model. Theresults of each individual method demonstrate significant estimation accuracy. This studyalso discusses the potential of fusing these two models to improve the performance of theestimation. Furthermore, the estimated speech transmission index was successfully used forcontrolling speech intelligibility in a semi-open space.Secondly, the thesis explores the information fusion framework using the Dempster-Shafer theory (DST). The objective is to estimate the state of a nonlinear system, i.e., theremaining energy of a battery during usage in an electric vehicle, called battery state-ofcharge(SoC). Although lithium-iron phosphate batteries are widely used, accurate SoC estimationremains challenging due to their nonlinear characteristics, fluctuations during transientstates, and uncertainty of the sensors. The DST, bounded-error state estimation, andforward-backward propagation are employed to enhance SoC accuracy. Different types ofcurrent and voltage sensors are used to observe the model’s variables, and the degree ofbelief, called mass function, for each sensor, is optimized using the differential evolutionalgorithm. Experimental comparisons with a Kalman filtering-based approach show the effectivenessof the proposed method in accurately estimating the SoC of a lithium battery.This thesis has presented blind estimation techniques based on the concept of theMTF using convolutional neural networks and the deterministic formulation. The proposedmethods can simultaneously and blindly estimate five room-acoustic parameters and thespeech transmission index successfully. Subsequently, the study has explored an informationfusion technique leveraging the Dempster-Shafer theory and interval analysis to combinedata from multiple sources. The findings of this research demonstrate the feasibility ofcombining information from different sources and models, thereby enhancing the state andparameter estimation of a dynamic system.