Joint model for residual life estimation based on Long-Short Term Memory network

Abstract With the rapid development of industrial internet of things, multi-sensor technology has been widely used in system condition monitoring and residual useful life (RUL) estimation, which plays a crucial role in preventing catastrophic failures and reducing maintenance losses. However, current prognosis using multi-sensor data faces the following challenges: (i) Manual feature selection and extraction and exploration of degradation failure mechanism for a complex domain require a substantial amount of human labor and expertise from the practitioner, which is seldom available. (ii) Data fusion and prognosis are usually divided into two separate steps, resulting in the lack of intrinsic relationship between the two tasks. (iii) The end-to-end prediction methods directly based on deep learning behave like the black-box and provide no information for degradation progression. To overcome these drawbacks, a prediction framework comprising two deep learning models is proposed in this paper. Through the proposed supervised joint training scheme, the framework not only provides a continuous visualization progression of system degradation but also ensures that the generated fusion signal effectively performs in RUL prediction. As a framework application, a joint model of RUL estimation based on an ordinary long short-term memory network is designed. In the experiment section, a simulation study is firstly performed to show its RUL prediction performance. Then, a case study is conducted using two public experimental data sets (the aircraft turbofan engine data and the milling). Furthermore, the advantage of the proposed joint model is validated by carrying out a comparison with other methods based on the same experimental data sets.