随着现代化工业设备的精密化和智能化，各类故障的发生不可避免，复杂系统的安全性和可靠性问题也因此成为了当前的研究热点。目前在故障诊断与容错控制领域，基于解析模型的传统方法已得到较为全面的探索和发展。然而在实际应用中，传统方法受限于精确建模困难等问题。为了克服这一困难，本文着重于探索基于数据驱动的故障诊断和容错控制方法，及其在机器人系统上的应用。其中，强化学习作为一种数据方法，旨在通过环境交互来学习最优的控制策略，其特点在于能够摆脱对系统模型的依赖，同时具有处理系统不确定性的能力。因此，本文的研究工作主要分为以下几点：首先，本文针对机械臂系统的执行器故障，提出了一种基于数据的主动容错控制算法，主要包含故障检测，故障辨识和容错控制三个模块。故障检测利用动力学模型网络来生成残差信号，作为判断故障是否发生的依据；故障辨识利用一维卷积网络对残差信号进行特征提取和分类识别；容错控制采用强化学习辅助控制器对原系统进行故障补偿和矫正。主动方法会利用故障信息指导强化学习策略的使用，相比于被动方法，长期运行的稳定性得到了提高。其次，本文提出了一种基于数据的主动故障诊断方法，在系统发生故障后，通过设计主动输入实现故障辨识和容错控制两个目标，它们的集成设计有助于降低计算复杂度并提高诊断期间的安全性。本文先将输入设计这一双目标优化问题转化为约束优化问题，再利用约束强化学习算法进行求解。在策略优化中，以故障辨识为优化目标，以跟踪控制为优化约束。同时，本文提出了基于系统输出轨迹的故障分离策略，可逐步剔除不匹配的故障模态来确定系统当前的故障模态。该方法在空间机械臂这类复杂系统上得到了仿真验证。最后，基于模型方法和数据方法各自的优势，本文探索了二者结合的混合控制方法，以及基于强化学习的混合容错控制方法。通过对系统注入激励信号，可获得包含系统特征的输入输出轨迹，这些数据能够用来刻画系统状态方程并用于控制律的设计，从而实现混合控制。故障发生时，可用强化学习辅助控制器来处理故障带来的系统不确定性，将其与混合控制方法相结合，可最终实现未知线性系统下的混合容错控制。
As modern industrial equipment becomes increasingly sophisticated and intelligent, the occurrence of various faults is inevitable. For this reason, the safety and reliability of complex systems have become a research hotspot in recent years. In the field of fault diagnosis and fault-tolerant control, traditional model-based methods have been developed comprehensively. However, in practical applications, traditional methods are limited by difficulties in accurate modeling. In order to overcome such problems, this thesis will focus on exploring data-driven fault diagnosis and fault-tolerant control, as well as their application in robotic systems. Among them, reinforcement learning, as a data-driven method, aims to learn the optimal control strategy through interactions with the environment. It can avoid dependence on analytical models and has the ability to handle system uncertainties. Therefore, this thesis mainly includes the following contributions:Firstly, a data-driven active fault-tolerant control method is proposed for actuator faults of robotic manipulators, which mainly includes three parts: fault detection, fault identification, and fault-tolerant control. Fault detection uses a dynamic network to generate residual signals as a basis for judging the occurrence of faults? Fault identification uses a one-dimensional convolutional network for feature extraction and classification based on residuals? Fault-tolerant control is based on reinforcement learning auxiliary controllers to correct faults for the original system. Active methods make use of faultinformation to guide reinforcement learning strategies, which helps to improve long-termstability compared to passive methods.Secondly, this thesis proposes a data-driven active fault diagnosis method, which designs active inputs and then injects them into the faulty system to achieve two goals: fault isolation and fault-tolerant control. Integrated design of diagnosis and control helps to reduce computational complexity and enhance safety during diagnosis. To this end, the bi-objective optimization problem of input design is transformed into a constrained one, which can be solved using some constrained reinforcement learning algorithms. In this optimization, fault isolation is regarded as the objective, and tracking control is taken as a constraint. Meanwhile, this thesis designs a fault isolation strategy according to system outputs, by which mismatched fault modes can be gradually excluded to determine the current fault mode. The proposed methods are verified by simulations on complex systems such as space manipulators.Finally, this thesis explores a hybrid control method combining the advantages of both model-based and data-based methods. Also, a hybrid fault-tolerant control method based on reinforcement learning is proposed. In the proposed method, input and output trajectories that contain system information can be obtained by injecting excitation signals into the system. These data can be used to characterize the system state equation and thus design control laws to achieve hybrid control. The effect of faults can be seen as system uncertainties, which can be handled with reinforcement learning-based auxiliary controllers. Combining it with hybrid control methods, hybrid fault-tolerant control forunknown linear systems can be realized.
