Modern engineered systems generally work under complex operational conditions. However, most of the existing artificial intelligence (AI)-based prognostic methods still lack an effective model that can utilize operational conditions data for remaining useful life (RUL) prediction. This paper develops a novel prognostic method based on Bi-Directional Long Short-Term Memory (BLSTM) networks. The method can integrate multiple sensors data with operational conditions data for RUL prediction of engineered systems. The proposed architecture based on BLSTM networks includes three main parts: 1) One BLSTM network is used to directly extract features hidden in the multiple raw sensors signals; 2) another BLSTM network is employed to learn higher features from operational conditions signals and the learned features from the sensors signals; 3) fully connected layers and a linear regression layer are stacked to generate the target output of the RUL prediction. Unlike other AI-based prognostic methods, the developed method can simultaneously model both sensors data and operational conditions data in a consolidated framework. The proposed approach is demonstrated through a case study on aircraft turbofan engines, and comparisons with other existing state-of-the-art methods are also presented.
16 March 2022 in IEEE Transactions on Transportation Electrification
State of health (SOH) is critical to the efficient and reliable use of lithium-ion batteries (LIBs), especially in electric vehicle (EV) applications. Recently, electrochemical impedance spectroscopy (EIS)-based technique has been proven to be effective for SOH estimation of LIB. However, existing EIS-based methods failed to consider the impact of ambient temperature and battery state of charge (SOC), leading to the limited flexibility of these methods under dynamic environments. In this work, a novel EIS-based method is proposed for battery SOH estimation considering variations of temperature and SOC. An equivalent circuit model (ECM) is first introduced, in which the solid electrolyte interface (SEI) resistance and charge transfer resistance are employed to map their relationship with SOH under variant temperature and SOC. Subsequently, a probabilistic model, taking charge transfer resistance, temperature, and SOC as input variables, is developed for LIB SOH estimation. Experimental study indicates that the estimation error of the proposed method is around 4% when simultaneously considering the temperature and SOC effects. Moreover, the estimation error can reach 1.29% under certain conditions (e.g., 80% SOC at 30 °C). Both results of estimation error are better than the existing EIS-based methods, which indicates that the proposed method is more flexible for SOH estimation with higher precision.
28 December 2021 in IEEE Transactions on Industrial Informatics
Cross-domain fault diagnosis methods based on transfer learning attempt to leverage knowledge from a domain with sufficient labeled samples to a different but related domain with few or even nonlabeled samples. These methods have been widely investigated in the past years. Notwithstanding the efficacy, most existing approaches assume that the label spaces of training and testing data are the same. However, this assumption is not practical in actual applications because new fault category usually happens in the testing stage. A cross-domain open-set transfer diagnosis method is presented in this article to manage the aforementioned problem. Domain adversarial model is employed to discriminate known from unknown target instances. Moreover, multiple auxiliary classifiers introduce a weighting module to evaluate the distinguishing domain knowledge to provide target instances with representative weights. The new adversarial domain adaptation network with diverse supplementary classifiers can effectively identify the unknown and known fault categories in the target domain and bridge the domain shift between the common fault category of the source and target domain. Experiments on two bearing datasets show the effectiveness and advantage of the proposed method.
8 November 2019 in IEEE Transactions on Reliability
This article proposes a novel deep learning based fusion prognostic method for remaining useful life (RUL) prediction of engineering systems. The proposed framework strategically combines the advantages of bidirectional long short-term memory (BLSTM) networks and particle filter (PF) method and meanwhile mitigates their limitations. In the proposed framework, BLSTM networks are applied for further extracting, selecting, and fusing discriminative features to form predicted measurements of the identified degradation indicator. Simultaneously, PF is utilized to estimate system state and identify unknown parameters of the degradation model for RUL prediction. Hence, the proposed fusion prognostic framework has two innovative features: first, the preprocessed features from raw multisensor data can be intelligently extracted, selected, and fused by the BLSTM networks without specific domain knowledge of feature engineering; second, the predicted measurements with uncertainties obtained from the BLSTM networks will be properly represented by the PF in a transparent manner. Moreover, the developed approach is experimentally validated with machining tool wear tests on a computer numerical control (CNC) milling machine. In addition, the popular techniques employed in this field are also investigated to compare with the proposed method.
A detailed analysis is performed on a dataset of failure and maintenance records from various onshore wind farms located in different geographical areas for the safety, risk, reliability, availability, and maintainability characterization of wind turbines. Specifically, characteristics related to failures, including the criticality of failure modes, failure frequencies, failure rates, and lifetime distributions of components, are analyzed to support the failure identification and failure prevention of wind turbines. Additionally, characteristics of maintenance, including typical maintenance measures of failures, policies for spare components, delayed maintenance, as well as related times such as reaction time, travelling time, and mean time to repair, are provided to support the maintenance management of wind farms. Based on the operational data analysis results, a reliability influence factor-based failure data correction approach is presented to transfer the onshore data to floating offshore turbines by modeling the differences in failure occurrences based on experts’ judgment. A comprehensive comparison with existing studies validates the performance of the proposed approach.
Health prognostics within the Internet-of-Things (IoT) paradigm face several challenges, including data privacy, client drift, and prediction accuracy. Federated learning (FL), as an emerging decentralized machine learning paradigm, has the potential to address these challenges by integrating multiple data silos in a distributed and privacy-preserved fashion. This paper develops a novel personalized federated transfer learning (PFTL) framework for customized health prognosis of multiple heterogeneous clients. The framework starts with a powerful initial global prognostic model that is pre-trained using a publicly available dataset in a central server. The pre-trained global model is then distributed to the local clients and fine-tuned separately on their respective private datasets. The fine-tuned local prognostic models are uploaded to the central server for dynamic weighted model aggregation. The aggregated model is then distributed to each client for implementing domain adversarial training to obtain a fine-grained local prognostic model. The proposed PFTL framework embeds a multi-scale attention module and a multi-head self-attention module parallelly into the deep learning-based prognostic model, which is shared between the central server and each local client. Through experimental verifications from lab testing-based and open-source fast-charging lithium-ion batteries datasets, we demonstrate that the proposed method can achieve accurate cycle-life prediction without compromising data privacy.