Parallel architectures for signal processing
This thesis presents the development of parallel architectures and algorithms for signal processing techniques, particularly for application to ultrasonic surface texture measurement. The background and context of this project is the real need to perform high speed signal processing on ultrasonic echoes used to extract information on texture properties of surfaces. Earlier investigation provided a solution by the nonlinear Maximum Entropy Method (MEM) which needs to be implemented at high speed and high performance. A review of parallel architectures for signal processing and digital signal processors is given. The aim is to introduce ways in which signal processing algorithms can be implemented at high speed. Both hardware and software have been developed in the project, and the signal processing system and parallel implementations of the algorithms are presented in detail. The signal processing system employs a parallel architecture using transputers. A feature of the design is that a floating-point digital signal processor is incorporated into a transputer array so that the performance of the system can be significantly enhanced. The design, testing and construction of the hardware system are discussed in detail. An investigation of some parallel DSP algorithms, including matrix multiplication, the Discrete Fourier Transform (DFT) and the Fast Fourier Transform (FFT), and their implementations based on the transputer array are discussed in order to choose an appropriate FFT implementation for our application. Several implementations of the deconvolution algorithms, including the Wiener-Hopf filter, the Maximum Entropy Method (MEM) and Projection Onto Convex Sets (POCS) are developed, which can benefit from the use of concurrency. A development of the MEM implementation based on the transputer array is to use the DSP as a subsystem for FFT calculations; this dual-system environment provides a significant resourse to be used to process ultrasonic echoes to determine surface roughness. Finally, the performance of the Projection Onto Convex Sets (POCS) algorithm in the field of ultrasonic surface determination and comparison with the Wiener-Hopf filter and the MEM are presented using simulated and real data. It is concluded that the parallel architecture provides a valuable contribution to high speed implementations of signal processing techniques.