| Precision metrology,represented by high-precision measurement technology,plays a decisive role in various aspects of intelligent manufacturing,where precision requirements would reach the micron level.The frequency modulated continuous wave(FMCW)laser ranging technology is one of the important high-precision ranging approaches and shows advantages including high sensitivity,strong anti-noise and antiinterference capabilities,low cost,and the ability to measure speed and distance simultaneously.It has broad development prospects in the non-contact contour scanning of non-cooperative targets,precision instrument manufacturing,military defense,life science,and other fields.At present,the related research on the high-precision linear FMCW laser ranging system mainly revolves around factors such as the frequency-swept nonlinearity,the frequency-swept bandwidth,and the phase noise of the frequency-swept laser,as well as the spectrum analysis in the process of the distance information inversion.However,due to constraints such as the ranging principle,the optical system design,and the disturbances in complex environments,there would be information occlusion and loss during this process inevitably,and the continuous improvement of the system performance leads to a significant increase in system complexity and cost,making it difficult to achieve.From the perspective of the limitations of previous methods in FMCW laser ranging system research,this dissertation focuses on the application and exploration of the machine learning theories and approaches represented by deep learning and reinforcement learning(RL)and conducts the following research:(1)The construction of the FMCW laser ranging systems on experimental platforms.Based on the research foundation of the system principle and the system characteristics,this dissertation designs and implements an experimental platform for the FMCW laser ranging system based on the laser power balance module,taking it as the basic platform for the following research of ranging systems.After that,this dissertation presents a twostage control-based frequency-swept nonlinearity correction method,including the predistorted iteration algorithm and the optoelectronic phase-locked loop.The pre-distorted modulation signal increases the effective operating bandwidth of the phase-locked loop,enhancing linearization stability and accuracy and integrating the advantages of both methods.Moreover,this dissertation implements a FMCW dual-channel laser ranging system based on the resampling principle,suppressing the impact of frequency-swept nonlinearity on measurement results from the perspective of data processing and enhancing the distance information restoration precision of the foundational system.The aforementioned systems lay the foundation for the subsequent introduction of data-driven methods into the FMCW laser ranging system and provide abundant data resources.(2)The front-end system control algorithm based on deep RL.This dissertation focuses on the inherent nonlinear characteristic of the frequency-swept laser that has a significant influence on the performance of the FMCW laser ranging system and proposes the narrow frequency-swept nonlinearity correction method based on the model-based deep RL(MDRL)with the Actor-Critic structure.This dissertation transforms the frequency-swept nonlinearity correction task into the form of Markov decision processes,constructs a frequency-swept nonlinearity measurement system of the semiconductor laser,and establishes the system model based on its experimental data as the RL environment model so as to improve the data efficiency during the training process of the control policy.The agent of the MDRL algorithm realizes the mining and learning of the frequency-swept nonlinear characteristic of the semiconductor laser,and completes the off-line optimization of the linearization policy in the interaction with the environment model.The modulation signal generated under the guidance of the policy is applied to the frequency-swept nonlinearity measurement system of the semiconductor laser,which increases the side mode suppression ratio of the beat signal from 5.54 d B in the case without a control algorithm to 13.31 d B,and the spatial resolution of the system is increased fivefold.(3)In response to the issues of the rapid increase in dimensions of the action and state spaces caused by the broadband frequency sweep,as well as the overestimation bias in the MDRL algorithm,the enhanced MDRL(EMDRL)algorithm is proposed by introducing the twin Critic neural networks and the network update strategies.Applying the obtained modulated signal to the frequency-swept nonlinearity measurement system of the semiconductor laser,the bandwidth of the beat frequency signal obtained by the system is reduced by one order of magnitude compared to the uncontrolled case,and the spatial resolution is improved from 0.0759 m to 0.0072 m.The remarkable linearization performance of the EMDRL algorithm demonstrates the potential of deep RL methods in complex control tasks,along with strong generalization ability and linearization stability.Since the training of the MDRL and EMDRL algorithms,as well as the generation and updating of the modulation signal for the semiconductor laser,are all accomplished on the computer side,the direct application of these RL methods in the FMCW laser ranging system would be feasible while controlling the system complexity and cost.(4)The distance information inversion is a crucial technology in the FMCW laser ranging technology,and its accuracy is affected by various factors,such as the beat signal quality and the spectral analysis accuracy.In contrast to previous approaches that compensate for different influencing factors separately,this dissertation combines the improvement of original signal quality with spectrum analysis optimization from the perspective of back-end data processing and proposes the distance information inversion method based on the deep attention mechanism(DAM).The deep residual network and the multi-head self-attention mechanism are used to achieve the inversion of the distance information in the original data.During the test,the mean square error of the detected distance can reach sub-millimetre scale.In summary,based on system platforms including the FMCW laser ranging system,the MDRL,EMDRL,and DAM methods are proposed with a focus on the data-driven concept.These methods realize the key technologies of the FMCW laser ranging system,involving the learning of nonlinear characteristics of the frequency-swept light source,the offline optimization of control policies,and the high-precision reconstruction of the distance information in ranging scenarios.This dissertation validates the feasibility and advantages of using deep learning and RL methods to address complex dynamic system control and data processing challenges in the field of optics.It is of both theoretical and practical value. |
