Optimised reservoir computing for temporal signal classification in LiDAR

Reservoir Computing has emerged as a potent machine learning tool, distinguished by its efficacy in handling complex sequential signals with minimal training complexity. This thesis delves into the realm of applying RC to train and classify distance information gleaned from FMCW LiDAR. Validation is carried out through simulation and experiments. Subsequently, the parameters of reservoir computing are optimized using genetic algorithm.

FMCW LiDAR, an optical sensing technology, is adept at gauging target distances by capturing changes in the time-domain waveform of the IF signal. Noteworthy is the spectral broadening issue arising from nonlinearities in the light source's wavelength-scanning process, leading to degraded ranging resolution. This research addressed this challenge by adapting RC-based temporal signal classification, completely avoiding computing-heavy Fourier analysis in existing systems. The approach has been validated via successful simulation experiments. These experiments showcased RC's proficiency in processing time-domain waveform signals, even when grappling with similar low-frequency waveforms.

To validate RC's real-world applicability, experimental signals were subjected to various scenarios, including sequential, jumbled, and single waveform distance signals of different lengths. Remarkably, RC demonstrated impeccable performance with 100% accuracy, attesting to its robustness across diverse signal configurations.

To quantify the quality of results, the study introduced RMSE, a metric measuring the discrepancy between test and standard results. The close alignment between test results and standard RMSE (ranging between 0.03 and 0.04) underscored RC's strong performance.

Despite RC's prowess, the importance of parameter tuning cannot be overstated. The study employed a genetic algorithm to optimize six crucial parameters in the RC model: spectral radius, lambda (regularization rate), res_units (number of nodes in RC), activation function, scaling, and connectivity. The choice of RMSE as the objective function for optimization aimed to fine-tune RC for optimal performance.

After 80 generations of iterative optimization, the observed decrease in RMSE signifies the genetic algorithm's success in identifying parameter combinations that enhance RC's performance. However, it's noteworthy that the overall degree of RMSE decrease is moderate. This suggests that RC inherently possesses a degree of robustness, capable of delivering acceptable results without exhaustive parameter optimization.

In conclusion, the synergy of RC's robust performance, combined with thoughtful parameter optimization through a genetic algorithm, positions it as a reliable and adaptable tool for processing and classifying complex sequential signals in applications such as FMCW LiDAR.