Technical Memorandum 104484 Sensor Failure Detection and Recovery by Neural Networks

This paper describes a new method of sensor failure detection, isolation, and accommodation using a neural network approach. In a propulsion system such as the Space Shuttle Main Engine, the dynamics are usually very complicated and sometimes not well known. However, the number of variables measured is usually much higher than the order of the system. This built-in redundancy of the sensors can be utilized to detect and correct sensor failure problems. The goal of the proposed scheme is to train a neural network to identify the sensor whose measurement is not consistent with other sensor outputs. Another neural network is trained to recover the value of critical variables when their measurements fail. Techniques for training the network with a limited amount of data are developed. The proposed scheme is tested using the simulated data of the Space Shuttle Main Engine (SSME) inflight sensor group. INTRODUCTION In 1980, a ground test of the Space Shuttle Main Engine (SSME) experienced an erroneous measurement of the Main Combustion Chamber pressure (Pc) [1]. Pc is used for the closed loop thrust level control as well as closed loop mixture ratio calculations. The failed sensor reading led the testing to a severely abnormal operating condition. An internal fire and subsequent explosion occurred as a result of the sensor failure. The engine was virtually destroyed. Also, during the course of the Space Shuttle program there have been numerous incidents of sensor failures which caused component damage, unnecessary shutdowns and delays of the program. In order to improve the operational reliability it is necessary to validate the measured sensor data, isolate any failed sensor and recover the failed critical measurement. There has been an extended effort in applying analytical redundancy to the sensor failure detection and isolation in the jet engine failure diagnosis problem [2]. In general, this approach utilizes the engine model and the Kalman Filter to detect and isolate sensor failures. This technique is strongly dependent upon a reliable system model which may not always be attainable in a complex system. This paper proposes that neural networks be trained by experimental data and learn the relationships among the redundant sensors. These networks are then used to check the validity of the sensor readings and provide an estimated value for failed sensors. This paper will first describe some of the system dynamics of the Space Shuttle Main Engine. The selection and the training algorithms of the neural networks are then presented, followed by the simulation results of the proposed approach. Finally, a discussion of the research is presented. The SSME DYNAMICS The Space Shuttle Main Engine under study is by far the most complicated and power intense machine among propulsion engines. A simplified description of the system operation follows [3,4]. There are three main engines in a space shuttle orbiter. Each engine produces a sea level thrust of 375,000 lb and a vacuum thrust of 470,000 lb. A schematic diagram of the propellant flows is shown in Fig. 1. Pressurized fuel, provided by the fuel tank, flowing through the low pressure fuel pump and the high pressure fuel pump, is fed to the regenerative cooling and the prebumers. A pressurized oxidizer tank provides the oxidizer which flows through the low pressure oxidizer pump and the high pressure oxidizer pump where the output flow splits into the two prebumers and the main combustion chamber as shown in Fig. 1. The dynamics of the system operation include: (1) the performance of turbopumps; (2) the heat exchange of the cooling flows: (3) the combustion of the two prebumers and the main chamber; (4) the control valve actions; and (5) the energy properties of oxygen and hydrogen in different phases. Most of these dynamic *NASA/OAI Summer Faculty Fellow at NASA Lewis Research Center.