Remaining Useful Life (RUL) prediction is one of the most common activities to ensure the reliability of a degradation system. In previous RUL prediction schemes based on RNN autoencoder, the multi-dimensional sensor data for each timestep made an equal contribution to the generation of the embedding vector during the encoding process. Besides, the single embedding vector carries the burden of decoding the entire multi-timestep information. To overcome the above shortcomings, two improvements are proposed: (1) For the embedding vectors to highlight critical timestep information, weights are assigned to each timestep information through an attention mechanism. (2) To reduce the decoding burden on a single embedding vector, a skip connection is introduced at each step of the decoding process to improve BiGRU decoding capabilities. The prognostic performance of the proposed method BiGRU-AS is evaluated on two publicly available datasets: the C-MAPSS dataset (simulation dataset) and the milling dataset (experimental dataset). Compared to the latest prediction methods, the experimental results show that the proposed method is competitive in RUL prediction for mechanical systems.
28 June 2022 in Measurement Science and Technology
Remaining useful life (RUL) prediction, as an essential task for prediction and health management, can improve the reliability of degradation systems. In this paper, multi-dimensional mechanical monitoring readings are treated as natural language sequences and inputted to a natural language processing model, Transformer, to extract semantic compression vectors. The health indicators (HIs) are constructed from the changes of compressed semantic vectors during the mechanical system operation, reflecting the mechanical health status at each monitoring point. Subsequently, combined with the similarity matching method, the RUL prediction is achieved by weighted summation. Transformer mainly consists of attention layers and residual connection layers, which contain no recurrence and convolution. The attention layer can highlight the critical time step information during the HI construction. Moreover, the output of different time steps can be obtained simultaneously in a parallel computation manner. The proposed scheme is conducted on two publicly available datasets for experimental and evaluation metrics comparison. The experimental results show that the proposed scheme can achieve competitive prediction results, with some of the metrics reaching the highest. An autoencoder only containing attention mechanism can effectively extract degradation feature representations for the adaptive construction of unsupervised HIs.
As an important part of urban big data, traffic flow data play a critical role in traffic management and emergency response. Traffic flow data contain multi-mode characteristics, which need to be deeply mined. To make full use of multi-mode characteristics, we use a 3-order tensor to represent the traffic flow data, considering “temporal-spatial-periodic” characteristics. To recover the missing data of traffic flow, we propose the Missing Data Completion Algorithm Based on Residual Value Tensor Decomposition (MDCA-RVTD), which combines linear regression, univariate spline, and CP decomposition. Then, we predict the future traffic flow data by using the proposed Traffic Flow Prediction Algorithm Based on Data Completion Strategy (TFPA-DCS). The experimental results show that recovering the missing data is helpful in improving the prediction accuracy. Additionally, the prediction accuracy of the proposed Algorithm is better than gray model and traditional tensor CP decomposition method.
Remaining useful life prediction can assess the time to failure of degradation systems. Currently, numerous neural network-based prediction methods have been proposed by researchers. However, most of the work contains an implicit prerequisite: the network training and testing data have the same operating conditions. To solve this problem, an adversarial discriminative domain adaption prediction method based on adversarial training is proposed to improve the accuracy of cross-domain prediction under different working conditions. First, an LSTM feature extraction network is constructed to mine the source domain data and the target domain data for deep feature representation. Subsequently, the parameters of the target domain feature extraction network are adjusted based on the idea of adversarial training to achieve domain invariant feature mining. The proposed scheme is experimented on a publicly available dataset and achieves state-of-the-art prediction performance compared to recent unsupervised domain adaptation prediction methods.
5 July 2021 in Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit
As the key running part of the railway freight transportation system, the wheel not only bears the load of the vehicle, but also ensures the running and steering of the car body on the rails. The frequent high-speed friction with the rail and brake is the main reason for early failure of wheelset tread. Therefore, real-time status monitoring and early fault diagnosis of wheel treads have become key technical issues that must be solved in the reform of the railway freight maintenance system. In this paper, an adaptive hybrid Simulated Annealing Cuckoo Search algorithm (SA-ACS) is proposed and applied to the Deep Belief Network (DBN). The SA-ACS-DBN algorithm is used to improve the training speed and convergence accuracy of the diagnosis model. Finally, it is found through the comparison experiment of wheel tread fault data that the data results prove the feasibility of the SA-ACS-DBN model in the application of wheelset fault diagnosis.