Online adaptive status prediction strategy for data-driven fault prognostics of complex systems

Accurate fault prediction or remaining useful life (RUL) estimation can obviously reduce cost of maintenance and decrease the probability of accidents so as to improve the performance of the system test and maintenance. At the same time, it is difficult to apply the model-based methods for fault prediction in most applications for complex systems. Due to continuously improving of automation, increasing of sampling frequency and development of computing technology and memory capacity, it gradually promotes data-driven technology into practical methods. Therefore, data-driven fault prognostics based on the sensor or historical test data has become the primary prediction means of complex systems, such as Artificial Neural Networks (ANN), Support Vector Regression (SVR) and other computational intelligence methods. In the other hands, most of traditional forecasting methods are always off-line that are not suitable for on-line prediction and real-time processing. Furthermore, for some on-line prediction methods such as Online Support Vector Regression (Online SVR), there is conflicts and trade-offs between prediction efficiency and accuracy. In real prognostics and health management (PHM) systems, such as the operating status monitoring and forecasting of complex systems like airplane and aircraft, it requires that the algorithms are flexible and adaptive in realization of balance between prediction efficiency and accuracy to meet different complicated requirements. To solve the problem above, an on-line adaptive data-driven fault prognosis and prediction strategy is presented in this paper. Considering the complex characteristics of trends and neighborhood of time series data, the multi-scale reconstruction strategy is applied to effectively reduce the size of on-line data and preserve rich history knowledge of samples. Therefore, the prediction efficiency could be improved and faster forecasting can be achieved with adaptive multi-scale sub-models. To evaluate the proposed prediction strategy, we have executed experiments with Tennessee Eastman (TE) process data. Experimental results with TE process fault data prove its effectiveness. The experiments and tests confirm the algorithms can be effectively applied to the on-line status monitoring and prediction with excellent performance in both efficiency and precision. New on-line fault status prediction strategy shows better prospect in real-time and on-line application for complex system. It can be applied in industrial fields for system maintenance and prognostics and health management.