Intelligent Fiber Sensors via Digital Signal Processing and Machine Learning

点击翻译按钮获取中文摘要

Optical fiber based sensors are sensitive and robust devices that are resilient to harsh external environment fluctuations (e.g. temperature, electromagnetic interference, radiation etc.) and serve well as a distributed sensing medium in infrastructure. Often used for high end calibration and control of manufacturing processes, their typical deployment involves bulky spectrum analyzers (electrical or optical), and costly tunable laser sources. The proposed research will address these challenges by applying detection schemes and signal processing techniques that have revolutionized digital communications in both wireline and wireless technologies. Information will be encoded in both the amplitude and phase of the optical signal to increase its sensitivity to the changing ambient environment, while removing the need for additional hardware at the transceivers. Estimation and calibration strategies will be developed to compensate for optical device fabrication defects and variations. As the environmental conditions vary relatively slowly (kHz) compared to the bandwidth requirement of a typical wireless (MHz)/ wireline (GHz) channel, the correspondingly matured semiconductor integrated circuit (high-speed electronic) originally intended for digital communications provides dramatic computation power at an extremely low cost. The excess computing bandwidth can be used to compensate for device artefacts via mature digital signal processing (DSP) technique common in communication and imaging. In a communication channel, where both the information channel and transmitted signal are varying in time, sophisticated techniques have been developed to estimate the sent messages with high fidelity. In the case of sensing, by restricting the probing signal in the transmitter, the receiver can readily estimate or sense the fluctuations in the medium by monitoring the probe signal. Finally, the digitization of the optical sensing signal means that common techniques in image recognition and machine learning can be applied to extract a particular signature associated with certain environmental parameters being monitored. The training duration has been significantly reduced due to the availability of high-end and affordable computation devices such as graphical processing units. Features of interest can be readily classified with high degree of certainty, whether it is a senior citizen falling on a sensor-enabled carpeted floor, or the seismic stability of an underground mining chute. The proposed research strives to apply commercially available electronics digital DSP and machine learning to the cutting edge of optical fiber sensing. As sensors continue to proliferate in our daily life, trainees will be provided hands-on experiences to integrate various state-of-the-art hardware and software components to address applications from the healthcare to the mineral sectors.