Research On Fault Detection And Diagnosis Methods Of Liquid-propellant Rocket Engine Based On Convolutional Neural Network

Fault detection and diagnosis methods are the basis of the health monitoring for liquid-propellant rocket engines(LREs)and have significant theoretical and engineering research value.However,the development in this field has been plagued by problems such as insufficient fault data and difficulties in fault feature representation.Convolutional neural network(CNN)has powerful capability in feature learning.The development of fault detection and diagnosis methods based on deep learning represented by CNN is one of the most important trends in the future.In this context,taking a large LRE as the main research object,some investigations are in-depth conducted focused on condition monitoring,fault detection,and diagnosis methods based on CNN for automatic feature extraction and one-class support vector machine(OCSVM)for anomaly detection.Aiming at the problems of the existing traditional condition monitoring methods,such as high dependence on experience knowledge and feature engineering,difficulties in feature representation,and high R&D cost,an engine condition monitoring method based on CNN is proposed with Le Net-5 as a prototype.The effectiveness of the method is demonstrated by some supersonic combustion test data.The results show that the proposed method gives full play to the feature learning ability of CNN,and can directly extract data features from original data and complete accurate recognition of the engine working states.Compared with other methods,the recognition results of the methods based on deep architectures are significantly better than that of the methods based on shallow architectures and traditional machine learning.Feature engineering can assist the shallow models and traditional machine learning methods to improve their feature learning abilities,but there are still significant performance gaps when compared with the CNN-based method.Due to diverse fault modes and few fault samples,it is difficult to establish a complete fault mode database for LRE.To solve this problem,a fault detection method for LRE is proposed on the basis of convolutional auto-encoder(CAE)and OCSVM,in which the CAE is used for automatic feature extraction and the OCSVM is used for fault detection based on the data features from CAE.Its effectiveness is demonstrated by the data of both the start-up process and the steady-state process from LRE ground tests.The influence of model capacity and sample length on the performance of the method is discussed systematically.The results show that this method can achieve timely and accurate results for LRE fault detection.Compared with other methods,the proposed method has significant advantages in detecting inchoate and weak abnormalities in data and can greatly advance the fault alarm time,especially for incipient and soft faults.The model capacity of the CAE should match the size of the data set to avoid over-fitting or under-fitting.Increasing the sample length can help improve the detection performance of the method,but it will delay the first detection time.Based on the CAE-1SVM-based fault detection method,the effective information in reconstruction error of CAE is further mined,and the diagnostic index—the relative root mean square(RRMS)of reconstruction error is constructed.The unsupervised fault diagnosis method is proposed based on the CAE reconstruction error analysis.The method is verified by the steady-state process data from LRE ground tests.The results show that the proposed diagnostic index can accurately track the deviation of each parameter from normal working conditions.The method can accurately achieve component-level fault isolation for faults entering the functional fault stage but has limited diagnostic capabilities for incipient faults.Given the problems of the fault diagnosis method based on the CAE reconstruction error analysis,a modular fault diagnosis method based on parallel CAE-1SVM(P-CAE-1SCM)is further proposed according to the structural characteristics of the LRE.The space and time complexities of the CAE in each module are quantitatively analyzed.The steady-state process data of the LRE ground test are used to demonstrate the effectiveness of the P-CAE-1SCM based method.The results show that this method can effectively detect potential faults and isolate faulty components at the same time.The complexity of CAE in each module is significantly reduced,and the overall operating efficiency of the diagnostic method is effectively improved.Taking background knowledge and monitoring knowledge in the field of health monitoring as examples,a data-level fusion method suitable for background knowledge and a feature-level fusion method suitable for monitoring knowledge are proposed respectively for the fusion of domain knowledge and LRE fault detection and diagnosis methods.The influence of the two types of knowledge fusion on the performance of fault detection and diagnosis method is analyzed according to the detection and diagnosis results of steady-state process data of LRE ground tests.The results show that the datalevel fusion method of background knowledge can incorporate background knowledge that is difficult to be explicitly expressed in data or features into the test data,and significantly improve the sensitivity of the fault detection method to incipient faults.The feature-level fusion method of monitoring knowledge may cause new features to overwrite the useful information in original data and reduce the performance of the detection and diagnosis method.