Abstract
This paper presents a low-power, low-voltage implementation of the Wavelet Transform (WT) using analog switched-current (SI) filters. To address the challenges of approximating complex wavelet bases into stable electronic circuits, an Adaptive Genetic Algorithm (AGA) is employed to optimize the filter's transfer function. The resulting design utilizes SI differentiators in a cascaded structure, specifically targeting the first derivative of a Gaussian wavelet. Simulations demonstrate that this approach achieves high approximation accuracy and multi-scale functionality, offering a viable alternative to power-intensive digital signal processing for real-time applications.
Introduction
The Wavelet Transform (WT) is a fundamental tool in signal processing, enabling simultaneous localization in both time and frequency domains. While digital implementations are standard, they often fall short in low-power, low-voltage environments—such as wearable medical sensors—due to the energy demands of Analog-to-Digital (A/D) converters and complex digital logic.
Analog implementations offer a potential solution but face significant design hurdles. Approximating a mathematical wavelet base into a realizable circuit transfer function is difficult. Traditional methods like Padé approximation can lead to unstable filters, while $L_2$ optimization often converges on local optima, limiting accuracy. This research proposes a novel combination of Adaptive Genetic Algorithms (AGA) for mathematical optimization and Switched-Current (SI) filters for hardware realization.
Methodology
Optimization via Adaptive Genetic Algorithm (AGA)
The core of the design process involves finding an 8th-order transfer function that closely matches the target wavelet base—in this case, the first derivative of a Gaussian wavelet. The researchers utilized an AGA, which dynamically adjusts crossover and mutation probabilities during the search process. This adaptability prevents the algorithm from getting stuck in local optima, a common failure point for standard genetic algorithms or $L_2$ methods. To ensure circuit stability, the AGA was constrained to produce parameters that result in stable filter poles.
Circuit Realization using SI Differentiators
Once the optimal transfer function was determined, it was mapped to a hardware architecture. Unlike many analog designs that rely on integrators, this implementation uses Switched-Current (SI) differentiators. SI circuits are advantageous for low-voltage operation and are highly compatible with standard CMOS processes. The filter is constructed using a cascade of four biquad sections, providing a robust and scalable architecture.
Results and Discussion
Approximation Performance
The AGA-derived transfer function demonstrated exceptional fidelity to the ideal Gaussian wavelet. When compared against traditional Padé[7/8] and $L_2$ approximation methods, the AGA approach provided a significantly closer fit in the time domain, which is critical for accurate wavelet analysis.
Multi-Scale Signal Processing
A key feature of the Wavelet Transform is its ability to analyze signals at different scales (dilations). In this analog implementation, scaling is achieved by simply varying the clock frequency of the SI filter. Simulations conducted using ASIZ software confirmed that the circuit could successfully generate impulse responses for scales $a=1, 2, 4,$ and $8$ by adjusting the clock frequency from 100kHz down to 12.5kHz.
The results indicate that the frequency response remains stable and accurate even when accounting for real-world transistor imperfections, such as finite conductance and parasitic capacitance.
Conclusion
The integration of Adaptive Genetic Algorithms with Switched-Current analog filters provides a powerful framework for implementing the Wavelet Transform in low-power environments. By prioritizing time-domain accuracy through AGA and utilizing the stability of SI differentiators, the researchers have developed a system capable of high-performance, multi-scale signal processing without the overhead of digital conversion. Future work will focus on physical chip fabrication and extending this methodology to other complex wavelet bases.
Summary updated