2 August 2024 in IEEE Transactions on Power Electronics
Accurately predicting the remaining useful life (RUL) of power insulated gate bipolar transistor (IGBT) modules is crucial for high-speed trains. Challenges under actual train operations, including significant uncertainty, multivariability, and insufficient full-lifecycle datasets of performance degradation parameters (PDPs), hinder the accurate RUL prediction. Thus, a hybrid RUL prediction approach is proposed. To address the significant uncertainty in PDP, variational mode decomposition (VMD) is utilized to segregate low-frequency trend information from the high-frequency uncertainty information, enabling distinct prediction approaches for diverse data. Besides, kernel density estimation-bidirectional long short-term memory network (KDE-BiLSTM) is proposed to precisely quantify and forecast the significant uncertainty information. Moreover, to tackle the multivariability of PDP, a Wiener-based transferable trend information modeling technique is introduced. Furthermore, to predict PDP trends with insufficient datasets, a model-based method employing regularized particle filter-RIME-least squares support vector machine (RPF-RIME-LSSVM) is proposed for trend information prediction. RIME-LSSVM solves the missing observation problem in the RPF prediction phase. Ultimately, the integration of the trend and uncertainty information yields the final RUL prediction. The proposed method was validated utilizing Infineon 6500 V/750 A IGBT modules. The maximum RUL prediction error was 9000 cycles, validating the method's effectiveness. The prediction error of at least 7000 cycles below the baseline models demonstrated the method's superiority.
25 August 2023 in IEEE Sensors Journal
The train axle temperature sensor data exhibit two distinct abnormal patterns: point anomalies and subsequence anomalies, which reflect the intricate operational state of the axle and train. Precise detection of multi-category anomalies within the axle temperature data is crucial for ensuring the safe and stable operation of trains. However, the coexistence of multiple anomaly types leads to diminished accuracy in existing anomaly detection methods. Hence, we propose an unsupervised unified detection model, leveraging the power of the prediction model, targeting both point and subsequence anomalies. Our model begins by utilizing long short-term memory (LSTM) networks to predict the axle temperature, thereby generating a residual sequence. To ameliorate the error accumulation effect of LSTM, a weighting mechanism is introduced during the prediction process. Subsequently, the kernel density estimation (KDE) method is adopted to perform piecewise non-parametric fitting on the distribution of the residual sequence, enabling the determination of non-parametric dynamic thresholds for point anomalies. Finally, using the detected point anomalies as boundary points, the residual sequence is partitioned into multiple subsequences of varying lengths. The improved local outlier factor (ILOF) method is applied to compute the anomaly score of each subsequence, enabling the recognition of subsequence anomalies. The evaluation experiments for anomaly detection were conducted on two axle temperature datasets. The F 1 -scores were all above 0.975, which verified the effectiveness of the model. Compared with the baseline models, the F 1 -score of the proposed model has increased by more than 5%, which verified the superiority of the proposed model.
26 May 2025 in IEEE Transactions on Instrumentation and Measurement
Accurate fault diagnosis of traction motor bearing is a necessary and indispensable part of the prognostics and health management of high-speed trains. Vibration signal analysis is usually utilized to diagnose bearing faults. However, challenges under actual train operations including strong multi-domain variability, high spectral complexity, high spectral uncertainty, and weak fault characteristics of vibration signals hinder accurate fault diagnosis. Thus, a novel fault diagnosis method based on spectrum transition entropy (SpTE) is proposed. Firstly, a symbolization method is constructed to represent the time-frequency distribution (TFD) of signals, completely considering spectral amplitude and phase information. Besides, to address the strong multi-domain variability, the transition probability matrix is introduced to describe the state transition process of the spectrum over time in the symbolic TFD (STFD). Moreover, to accurately quantify high spectral complexity and spectral uncertainty, the entropy value is calculated using the joint probability distribution of the symbolic state and its one-step transition state. Furthermore, a parameter optimization strategy is designed to enable the adaptive determination of SpTE parameters for enhancing sensitivity to weak faults. Finally, these extracted optimized features are identified using a classifier to achieve fault classification. The effectiveness of the proposed method is verified by three cases. The highest fault recognition rates are 99.67%, 99.30%, and 100%, respectively, which are better than baseline methods. The experimental results show that the proposed method has higher accuracy and provide a new approach for diagnosing traction motor rolling bearings.
5 November 2025 in IEEE Transactions on Power Electronics
The remaining useful life (RUL) prediction and predictive maintenance (PdM) fusion framework for high-power insulated gate bipolar transistor (IGBT) modules ensures timely maintenance decision-making, enhancing their operational reliability. However, existing frameworks suffer from low-robustness uncertainty management and insufficient consideration of both reliability information and random failure mode, causing RUL prediction and PdM decision deviations while limiting method applicability. Thus, an improved framework is proposed. The framework employs weighted multimodel (WMM)-enhanced particle filter (PF) to predict performance degradation parameters (PDPs). WMM provides more reliable observations for PF across device lifespans, robustly mitigating the variance accumulation and thereby reducing RUL prediction uncertainty. Based on PDP predictions, the RUL distribution and the reliability function characterizing fatigue failure are derived. By integrating the reliability function with random failure information, the reliable RUL considering both failure modes is acquired, extending the RUL prediction process to safety-critical scenarios. Furthermore, the framework employs deep reinforcement learning to obtain optimal PdM strategies. The reward function considers not only conventional maintenance cost and availability terms but also reliability predictions, enabling the PdM strategy adapt to scenarios with different reliability requirements. Finally, the effectiveness and superiority of the proposed framework were validated using full-lifecycle thermal resistance degradation data from 6500 V/750 A IGBT modules.
4 August 2025 in IEEE Transactions on Instrumentation and Measurement
Recently, Wiener-based remaining useful life (RUL) prediction methods are widely employed to ensure device reliability. However, for high-power insulated gate bipolar transistor (IGBT) modules whose performance degradation parameters (PDPs) contain substantial noise components, existing methods introduce considerable uncertainty in RUL predictions, significantly increasing maintenance costs. Thus, a Wiener-based RUL prediction method integrated with a comprehensive uncertainty reduction strategy is proposed. Under the existing Wiener-based RUL prediction framework, the drift coefficient is modeled by replacing the divergent random walk process in conventional methods with the convergent Ornstein–Uhlenbeck (OU) process, which describes the time-varying degradation rate of the device more precisely and diminishes the RUL uncertainty. Accordingly, a dedicated parameter estimation process for the Wiener degradation model incorporating the OU process is developed. During particle filter (PF)-based prediction of the PDP trajectory, a model augmentation technique is introduced. By employing the long short-term memory (LSTM) network to extend the PDP dataset to the full life cycle, the PF maintains continuous operation throughout the lifespan of the device. This approach prevents the accumulation of variance in PDP predictions and further reduces RUL uncertainty. Ultimately, the efficacy and superiority of the proposed method were validated using the simulation dataset and the degradation dataset from the Infineon 6500 V/750 A IGBT module.
1 January 2026 in IEEE Transactions on Instrumentation and Measurement