Approach for fault prognosis using recurrent neural network

In general, fault prognosis research usually leads to the research of remaining useful life prediction and performance prediction (prediction of target feature), which can be regarded as a sequence learning problem. Considering the significant success achieved by the recurrent neural network in sequence learning problems such as precise timing, speech recognition, and so on, this paper proposes a novel approach for fault prognosis with the degradation sequence of equipment based on the recurrent neural network. Long short-term memory (LSTM) network is utilized due to its capability of learning long-term dependencies, which takes the concatenated feature and operation state indicator of the equipment as the input. Note that the indicator is a one-hot vector, and based on it, the remaining useful life can be estimated without any pre-defined threshold. The outputs of the LSTM networks are connected to a fully-connected layer to map the hidden state into the parameters of a Gaussian mixture model and a categorical distribution so that the predicted output sequence can be sampled from them. The performance of the proposed method is verified by the health monitoring data of aircraft turbofan engines. The result shows that the proposed approach is able to achieve significant performance whether in one-step prediction task, in long-term prediction task, or in remaining useful life prediction task.