Model-based detection of soft faults using the smoothed residual for a control system

Soft faults are a very common failure mode in electro hydraulic servo control systems in aero engines, including component characteristic drift, performance degradation and so on. Traditional fault detection approaches for soft faults haveproblems of time delay, inaccuracy and false alarms. In order to solve these problems, a model-based detection approach for soft faults is proposed in this paper. Firstly, a robust, time-saving and reliable modeling method is proposed. This method utilizes a mechanism model and a data-driven correction model to characterize the main features and uncertain features of the practical system, respectively. The data-driven correction model is implemented by a backpropagation neural network, which adopts the input of the actual system and the output of the mechanism model as its training data, and adopts the output of the actual system as its training label. This neural correction model plays an important role in compensating for the deviation between the mechanism model and the practical system. Secondly, a robust fault observer, which is based on the residual between the measurement output of the practical system and the model output of the hybrid mechanism-neural network model, is designed to detect the soft fault. The squared residual is smoothed by a median exponential filter in order to eliminate the disturbances of impulsive interference or noises in measurement. Finally, the detection of soft faults is implemented by comparing the smoothed residual with a presetting threshold. Several simulations have been performed to verify the effectiveness of the proposed method, and the simulation results prove the advancement of the proposed method.