Frequency Domain Blind Source Separation Algorithms And Application In High Speed Train Noise Components Separation

Blind source separation (BSS) aims at recovering individual source signals from their mixed observations, the word "blind" means that neither the sources nor the mixing environment is known. There are many potential applications of BSS techniques such as speech enhancement, robust speech recognition, analyzing EEG or fMRI signals, feature extraction, geological exploration, image denoising, hyperspectral image processing, etc.In this dissertation, we mainly focus on frequency domain blind source separation (FDBSS) algorithms. Since signals propagate in specified velocity in real-word mixing environment, different sources are mixed with each other, as well as their delays, attenuations, and reverberations, i.e., they are mixed in a convolutive manner. FDBSS algorithms utilize short time Fourier transform (STFT) to convert time domain convolutive mixture to frequency domain instantaneous mixture, then signals can be separated by well-studied instantaneous BSS algorithms in each frequency bin, so the entire separation problem is simplified. However, FDBSS algorithms suffer from the well-known "permutation ambiguity":separated signals in each frequency bin must be rearranged to the same order before output the final results, so, a permutation algorithm is required for this purpose as a postprocessing step after frequency bin-wise separation. The main contributions of this dissertation are as follows:1. Kernal and spectral methods for solving the permutation problem. Clustering techniques are broadly used to solve the permutation problem in FDBSS algorithms, however, some challenges still exist, for example, elongated datasets should be handled, and constraint from the background knowledge must be considered. Inspired by various successful applications of kernel and spectral clustering methods in machine learning and data mining community, we try to solve the permutation problem by these methods. In this work, the weighted kernel k-means algorithm is modified according to the specific requirement of the permutation problem, and the spectral interpretation of the kernel approach is also investigated. In addition, several kernel construction approaches are proposed to improving the permutation performance.2. Independent vector analysis using subband and subspace nonlinearity. Independent vector analysis (IVA) is a recently proposed technique, which can solve the FDBSS problem. Compared with the traditional frequency bin-wise instantaneous separation plus permutation correction approach, the largest advantage of IVA is that the permutation problem is directly addressed by IVA rather than resorting to the use of an ad hoc permutation resolving algorithm after a separation of the sources in multiple frequency bands. In this work, two updates for IVA are presented. First, a novel subband construction method is introduced, IVA will be conducted in subbands from high frequency to low frequency, and the fact that the inter-frequency dependencies in subbands are stronger allows a more efficient approach to the permutation problem. Second, to improve robustness and decrease noise, the IVA nonlinearity is calculated only in the signal subspace, which is defined by the eigenvector associated with the largest eigenvalue of the signal correlation matrix. When the two updates are used together, both separation performance and algorithm robustness are dramatically improved.3. Performance evaluation methods for FDBSS algorithms. In order to develop better BSS algorithms, how to evaluate the algorithm performance becomes a problem worthy to be investigated. In this work, we mainly focus on the evaluation problem for FDBSS algorithms. First, the uniform energy flow network is calculated from the mixing and the demixing system, or estimated from the original and the estimated source signals in frequency domain. Then, signals are decomposed according to their energy flow, and several performance indices are derived from the decomposition. The proposed method is especially suitable for BSS performance evaluation in simulated environments.4. High speed train noise components separation. High speed train noise level is an important factor with respect to passenger comfort and life quality of residents along the railway, thus determining how to attenuate the noise level is an important research direction that train designers care about. Studies show that train noise is a kind of mixed signal which is made up of train body vibration, rolling noise, aerodynamic noise, device noise, etc., separating individual noise components from the overall observations will provide some guide to train noise reduction design. Since BSS algorithms have many successful applications in speech and audio separation tasks, which are similar to the circumstance in the train, we try to use these techniques to perform train noise components separation. Two kinds of noise components are separated in this work:first, train transmission noise and structural noise are separated by a system identification method; then, train interior noise components are separated by BSS algorithms.All the works in this dissertation and all compared algorithms are integrated in a unified platform for BSS research and application, this platform is developed in Java, and the source code is available for public.