【作者】 张玉祥；

【导师】 王晨希；

【作者基本信息】 西安交通大学 ， 机械硕士（专业学位）， 2024， 硕士

【摘要】 液体火箭发动机在太空任务中的关键作用不可忽视,其性能高、工作可靠的特点使其成为航天领域不可或缺的动力源。由于其复杂的热流体动力学系统,液体火箭发动机在高温、高压、强氧化等极端环境下工作,导致其成为运载火箭中故障的敏感多发部位。发动机失效对航天任务的影响巨大,提高液体火箭发动机的安全性和可靠性至关重要,故障检测与诊断技术成为确保液体火箭发动机安全可靠运行的关键技术之一。液体火箭发动机工作环境的极端性质和试车成本的昂贵,传统实验方法难以构建完备的故障模式库。基于计算机仿真技术的模型驱动方法有着低成本、安全性高、泛化能力强的优点,为构建完备的故障模式库提供了有效思路。然而,液体火箭发动机故障模式复杂多变,仿真模型无法完全模拟实际运行工况,仿真与实际存在一定的差异;人工智能领域迁移学习方法能够利用相似域学习到的知识用于目标域的分类。因此,论文研究模型驱动液体火箭发动机小样本迁移故障诊断,对于提高液体火箭发动机故障样本稀缺情况下的诊断可靠性和准确性具有重要意义。针对实际试车过程中故障数据缺少、故障模式与故障特征映射关系缺乏、故障难以复现等问题,建立液体火箭发动机动力学模型并在该正常模型基础上注入故障进行仿真构建故障数据库,有效克服了液体火箭发动机部分故障类型的数据样本不平衡以及数据缺少的问题。该方法根据液体火箭发动机运行原理以及部件之间的输入输出关系进行了模块化划分,并针对每个模块进行了动力学建模;调试连接每一个子模型,建立完整的火箭发动机模型。根据故障机理采用故障因子以及故障模型的方式在正常模型中注入转子卡滞、堵塞、泄漏等故障,通过时间信号控制模块控制故障在稳态运行过程中注入,记录故障发生时故障与故障表征之间的映射关系,建立故障模式库。最后,通过对比实际试车中故障征兆验证模型准确性和可靠性。针对液体火箭发动机仿真模型与实际试车存在差异导致仿真数据难以直接用于实际诊断的问题,研究仿真数据与发动机实际试车数据的迁移学习方法,实现从大量仿真数据中学习故障特征用于实际诊断。对仿真数据和实际数据进行预处理,通过归一化和梯度化统一数据结构便于作为神经网络的输入;通过数据堆叠的方式将时序信号排列为二维图像的形式,采用二维卷积提取数据参数之间的关联特征;针对数据结构特性搭建了卷积神经网络以及长方形卷积核用于提取故障特征。采用大量仿真数据预训练源域神经网络,学习仿真数据中的故障特征;将预训练得到的源域神经网络中的参数迁移到目标域神经网络,分别冻结不同的神经网络层,采用少量实际数据对未冻结层进行微调以适应实际数据分布,得到最佳的分类诊断模型。通过实际数据组成的测试集对模型进行了验证,模型对于故障类别的分类准确率可达95%,并采用混淆矩阵和t-SNE图的形式可视化诊断结果验证模型的可靠性。针对液体火箭发动机中样本类别之间误诊代价不同的问题,采用代价敏感的方法,基于类别之间不同的重要程度,使得迁移学习网络在训练时专注学习代价高的训练样本类别,通过惩罚对高代价样本的重度惩罚有效降低误诊率并提高诊断正确率。该方法基于贝叶斯最优决策理论对卷积神经网络中的多类别交叉熵损失函数进行修改,将成本矩阵系数分别注入到多类别交叉熵损失函数内和多类别交叉熵损失函数外进行训练对比;依据专家知识对不同样本类别进行重要程度划分,构建不同比例的成本矩阵,使得卷积神经网络在迭代训练时放大重要样本误诊项损失,依此惩罚误诊样本使得网络专注学习该样本类别的训练。通过训练比较得到最佳的成本矩阵注入位置以及成本矩阵内样本的比值关系,成本矩阵注入后模型对于故障类别的分类准确率提升到了97%,同时极大程度降低了误诊率。开展液体火箭发动机冷却管路故障模拟实验。根据某型液体火箭发动机冷却管路供应系统设计原理简图,根据原理简图参考管路实验台搭建案例,定制管路元件搭建实验台,搭建的实验台能够实现泄漏、堵塞、阀门误动等故障的模拟。通过温度、压力、流量传感器采集指定位置参数;针对其结构建立相应的管路模型进行仿真,通过仿真和实验构造源域数据集和目标域数据集,验证迁移学习算法有效性。最后,采用某型液体火箭发动机真实试车数据对迁移学习算法的准确性和可靠性进行了验证。更多还原

【Abstract】 The crucial role of liquid rocket engines in space missions cannot be ignored,as their high performance and reliable operation make them an indispensable power source in the aerospace field.Due to the complex thermal fluid dynamics system,liquid rocket engines operate in extreme environments such as high temperature,high pressure,and strong oxidation,making them sensitive and prone to faults in launch vehicles.Engine failure has a significant impact on space missions,and improving the safety and reliability of liquid rocket engines is crucial.Fault detection and diagnosis technology has become one of the key technologies to ensure the safe and reliable operation of liquid rocket engines.The extreme nature of the working environment and expensive testing costs of liquid rocket engines make it difficult for traditional experimental methods to construct a complete fault mode library.The model-driven method based on computer simulation technology has the advantages of low cost,high security,and strong generalization ability,providing effective ideas for building a complete fault mode library.However,the fault modes of liquid rocket engines are complex and variable,and simulation models cannot fully simulate actual operating conditions,resulting in certain differences between simulation and reality.Transfer learning methods in the field of artificial intelligence can utilize knowledge learned from similar domains for target domain classification.Therefore,the paper studied the model-driven small sample migration fault diagnosis of liquid rocket engines,which is of great significance for improving the reliability and accuracy of diagnosis of liquid rocket engines in the case of sample scarcity.In response to challenges such as a lack of fault data,the absence of a mapping relationship between fault modes and characteristics,and the difficulty in reproducing faults during actual test runs,we establish a dynamic model for liquid rocket engines.Faults are intentionally injected into this normal model for simulation,facilitating the construction of a fault database.This approach effectively addresses issues of imbalanced fault diagnosis data samples and data shortage in liquid rocket engines.The method is modularized based on the operating principles of liquid rocket engines and the input-output relationships between components,with dynamic modeling performed for each module.Debug and connect each sub model to establish a complete rocket engine model.Faults such as rotor jamming,blockage,and leakage,corresponding to fault factors and models,are injected into the normal model based on the fault mechanism.The fault is controlled by a time signal control module during steady-state operation,and the mapping relationship between the fault and its representation is recorded to establish a fault mode library.Finally,the accuracy and reliability of the model are verified by comparing fault symptoms in actual test runs.To address the challenge of disparities between simulation models and actual test runs of liquid rocket engines,rendering simulation data challenging for direct practical diagnosis,we explore a transfer learning method bridging simulation data and actual test run engine data.The approach involves preprocessing both simulation and actual data,standardizing the data structure through normalization and gradient,facilitating uniform input for neural networks.Temporal signals are organized into two-dimensional images via data stacking,and twodimensional convolution is applied to extract correlation features among data parameters.Convolutional neural networks,along with rectangular convolutional kernels,are constructed to extract fault features leveraging the inherent characteristics of data structures.The process entails pretraining source domain neural networks with a substantial amount of simulation data to learn fault features.Subsequently,parameters from the pretrained source domain neural network are transferred to the target domain neural network.Different layers of the neural network are frozen,and the unfrozen layers are fine-tuned using a limited amount of actual data to align with the actual data distribution,resulting in the optimal classification and diagnostic model.Model accuracy is verified using a test set comprising actual data,and diagnostic results are visualized through a confusion matrix and t-SNE graph to validate model reliability.To address the issue of different misdiagnosis costs among sample categories in liquid rocket engines,a cost sensitive approach is adopted.Based on the different importance levels between categories,the transfer learning network focuses on learning high cost training sample categories during training.By imposing severe penalties on high cost samples,the misdiagnosis rate is effectively reduced and the diagnostic accuracy is improved.This method is based on Bayesian optimal decision theory to modify the multi class cross entropy loss function in convolutional neural networks.The cost matrix coefficients are injected into the multi class cross entropy loss function and outside the multi class cross entropy loss function for training comparison;Based on expert knowledge,different sample categories are classified into importance levels,and cost matrices with different proportions are constructed to amplify the loss of misdiagnosis items in important samples during iterative training of convolutional neural networks.By punishing misdiagnosis samples,the network focuses on learning the training of that sample category.Obtain the optimal cost matrix injection location and the ratio relationship of samples within the cost matrix through training comparison.Conduct simulation experiments on cooling pipeline faults in liquid rocket engines.Based on the schematic diagram of the cooling pipeline supply system designed for a specific type of liquid rocket engine,and referring to the construction of a pipeline experimental platform,customized pipeline components to establish the experimental setup.This constructed platform can simulate faults such as leakage,blockage,and valve jamming.Collect specified location parameters through temperature,pressure,and flow sensors.Establish corresponding pipeline models for simulation based on its structure and construct source domain datasets and target domain datasets through simulations and experiments.In addition,the accuracy and reliability of the transfer learning algorithm were verified using real test data from a certain type of liquid rocket engine.更多还原