毕业生知识库

作为新型潜航器，仿生机器鱼以其低噪声、高机动、高效率和高隐蔽性的优势，在民用和军事领域具有广阔的应用前景。水下环境感知传感器易因复杂水下环境、传感器自身寿命以及难以避免的环境碰撞等因素影响而发生故障，如何快速发现传感器故障并识别故障类型对提高仿生机器鱼的生存和可持续作业能力尤为重要。因此，面向仿生机器鱼的传感器故障诊断方法研究具有重要的理论意义和应用价值。本文以多关节仿生机器鱼为研究对象，围绕仿生机器鱼系统优化与动力学建模、基于深度学习的传感器故障诊断和基于虚实迁移的传感器故障诊断方法展开研究，主要研究内容如下:

一，优化仿生机器鲨鱼基础运动平台，搭建面向传感器故障诊断的实验平台，并建立其三维动力学模型。首先，在现有基础运动平台的基础上，引入双深度传感器实现深度传感器故障模拟和深度真值的采集，同时设计故障指示系统，实时可视化故障数据采集状态和故障诊断结果；其次，基于此仿生机器鲨鱼实验平台，建立其三维动力学模型，为仿真研究和虚拟故障数据集的构建奠定基础；最后，仿真实验和实物实验验证了三维动力学模型的有效性。

二，针对故障诊断准确率低和速度慢的问题，提出了基于空域图像融合和轻量化AlexNet网络的故障诊断架构。首先，基于仿生机器鱼故障诊断实验平台，利用故障注入法，采集定深运动模态关键数据，构建传感器故障诊断真实数据集；其次，基于格拉米角场方法，将时序数据转换成空域图像，并深入分析空域图像特性；另外，研究空域图像融合方法，设计基于粒子群算法的融合系数优化策略，实现空域图像的优化融合。此外，为了降低资源的依赖和提高计算效率，设计轻量化AlexNet神经网络架构，在确保故障诊断准确率的同时，降低时间复杂度和空间复杂度；最后，基于真实数据集，验证了所提网络架构具有更高的准确率和更快的速度。

三，针对现实世界中数据稀缺和标注困难的问题，探索虚实迁移在仿生机器鱼故障诊断领域的应用。首先，基于仿生机器鱼三维动力学模型，采用故障注入法，采集定深运动模态关键数据，构建用于训练迁移网络的虚拟数据集；其次，设计时空双流神经网络，并基于虚拟数据训练模型，深入挖掘故障数据的时序与空域特征；进而，通过真实数据集进行模型适应性调整，使模型适用于真实数据集的数据分布；最后，利用大部分真实数据测试网络，评估虚拟到真实故障诊断模型迁移的性能。

最后，对本文工作进行了总结，并对提升仿生机器鱼安全性和可靠性的研究方向进行了展望。

As a novel submersible, bionic robotic fish have broad application prospects in both military and civilian fields due to their advantages of low noise, high maneuverability, high efficiency and high concealment. Underwater environment sensing sensors are prone to failures due to the complex underwater environment, sensors' life cycle, and unavoidable environmental collisions. It is particularly important to quickly detect sensor faults and identify fault types to improve the survival and sustainable operation of the bionic robotic fish. Consequently, the research on sensor fault diagnosis methods for bionic robotic fish has important theoretical significance and application value. This thesis takes 
multi-jointed  bionic robotic fish as a research object and focuses on the system optimization and dynamic modeling of the bionic robotic fish, sensor fault diagnosis based on deep learning and sensor fault diagnosis based on simulation to real transfer. The main research contents are as follows:

Firstly, the basic motion platform of the bionic robotic shark is optimized to build an experimental platform oriented to sensor fault diagnosis and establish three dimensional dynamic model. To begin with, based on the existing basic motion platform, dual depth sensors are introduced to implement the fault simulation of the depth sensor and the acquisition of the true depth value. Meanwhile, fault indication system is designed to visualise fault data collection status and fault diagnosis results in real-time. Then, based on the bionic robotic shark experimental platform, the three dimensional dynamic model is established, which lays the foundation for the simulation study and the construction of the simulated fault dataset. In the end, simulation experiments and physical experiments validate the effectiveness of the three dimensional dynamic model.

Secondly, addressing the low accuracy and slow speed of fault diagnosis, a fault diagnosis architecture based on spatial image fusion and lightweight AlexNet network is proposed. Primarily, utilizing the fault injection method on the bionic robotic fish fault diagnosis experimental platform, key data of depth control motion mode are collected to construct the real sensor fault diagnosis dataset. Next, based on the Gramian angular field method, time series signals are transformed into spatial images, and the features of spatial images are deeply analyzed. Furthermore, spatial image fusion methods are researched, and a fusion coefficient optimization strategy based on the particle swarm optimization algorithm is designed to achieve optimal fusion of spatial images. Additionally, to reduce resource dependency and improve computational efficiency, a lightweight AlexNet neural network architecture is proposed, which ensures diagnostic accuracy while reducing time complexity and space complexity. In the end, based on the real dataset, it is verified that the proposed network architecture has higher accuracy and faster speed.

Thirdly, addressing the challenges of sparse data and difficult annotations in the real world, we explore the application of simulation to real transfer in the field of fault diagnosis for bionic robotic fish. Initially, utilizing the fault injection method on the bionic robotic fish three dimensional dynamic model, key data of depth control motion mode are collected to construct the simulated sensor fault diagnosis dataset which is used to train simulation to real network. Subsequently, a temporal and spatial dual stream neural network is designed, and the model is trained based on simulated data to deeply mine the temporal and spatial features of fault data. Following this, the model is adaptively tuned by the real dataset to make the model applicable to the data distribution of the real dataset. In the end, the majority of real data is utilized to test the network and evaluate the performance of simulation to real fault diagnosis model transfer.

Finally, we conclude the total work and look forward to the research directions to improve the safety and reliability of the bionic robotic fish.