Predicting the long-term lifetime of power MOSFET devices plays a central role in the prevention of unprecedented failures for power MOSFETs used in safety-critical applications. The various traditional model-based approaches and statistical and filtering algorithms for prognostics have limitations in terms of handling the dynamic nature of failure precursor degradation data for these devices. In this paper, a prognostic model based on LSTM and GRU is developed that aims at estimating the long-term lifetime of discrete power MOSFETs using dominant failure precursor degradation data. An accelerated power cycling test has been designed and executed to collect failure precursor data. For this purpose, commercially available power MOSFETs passed through power cycling tests at different temperature swing conditions and potential failure precursor data were collected using an automated curve tracer after certain intervals. The on-state resistance degradation data identified as one of the dominant failure precursors and potential aging precursors has been analyzed using RNN, LSTM, and GRU-based algorithms. The LSTM and GRU models have been found to be superior compared to RNN, with MAPE of 0.9%, 0.78%, and 1.72% for MOSFET 1; 0.90%, 0.66%, and 0.6% for MOSFET 5; and 1.05%, 0.9%, and 0.78%, for MOSFET 9, respectively, predicted at 40,000 cycles. In addition, the robustness of these methods is examined using training data at 24,000 and 54,000 cycles of starting points and is able to predict the long-term lifetime accurately, as evaluated by MAPE, MSE, and RMSE metrics. In general, the prediction results showed that the prognostics algorithms developed were trained to provide effective, accurate, and useful lifetime predictions and were found to address the reliability concerns of power MOSFET devices for practical applications.

As the core component of power electronic systems, health monitoring of Metal Oxide Semiconductor Field Effect Transistor (MOSFET) is extremely crucial. In this article, a hybrid failure precursor prediction model based on machine learning techniques is proposed. It consists of an isolation forest method and a long short-term memory network (LSTM). The proposed model extracts information from different aspects of the input data to make predictions and can be sensitive to abnormal data behavior. By detecting the abnormality in the curve and predicting its future behavior, the model can give early warning of the power MOSFET failure and help avoid unexpected accidents. Besides, the model uncertainty is discussed. Two main factors that affect the model uncertainty of the proposed model are evaluated. To reduce the model uncertainty, a BNN is used to quantify the uncertainty of the proposed model with different parameters. The performance of the proposed model is verified based on the power MOSFET data collected from the accelerated life tests. The experiment results indicate satisfying performances of the proposed model because it can not only give early warning of MOSFET failures but also provide more stable prediction results with less model uncertainty compared to other existing models.

Intelligent power module (IPM) is a critical component of a power conversion system, and accurately predicting the lifetime of the key internal component, the insulated gate bipolar transistor (IGBT), is critical to the safe operation of the system. However, numerous challenges, such as monitoring difference of the internal state, variable degradation processes, and diverse field strengths, make it difficult to predict the IGBT lifetime accurately. Therefore, this paper proposes a remaining useful life (RUL) prediction method based on multi-source data fusion under limited information conditions. A prediction model combining Position Attention Module (PAM) and Long Short-Term Memory Network (LSTM) is designed to accommodate the large variability of the degradation process to enhance the ability to capture the spatio-temporal dependence of degradation features. In addition, to address the difficulty of internal state monitoring and the diversity of operating field strengths, this study constructs a Digital Twin (DT) model of the intelligent power module based on Finite Element Analysis (FEA) to comprehensively analyze the internal state of the device and provide accurate data support for RUL prediction. Meanwhile, to tackle the problems of insufficient health indicators, incomplete failure information, and end-of-life (EOL) fluctuation, a multi-indicator double-layer determination mechanism is proposed to improve the accuracy of the predicted EOL. Eventually, the prediction of the RUL of intelligent power module is achieved under limited observation information. The proposed method is experimentally verified on a 10 A/1200 V IPM from STMicroelectronics with accurate prediction results, and its superiority is further proved by comparing it with other methods.

In recent years, deep learning technologies have shown significant advancements and have been widely adopted in remaining useful life (RUL) prediction approaches. However, existing approaches primarily focus on learning the relationship between measured aging data of different time steps, neglecting the information contained within individual variables. Furthermore, most current approaches require a high computational cost to achieve accurate prediction. To address these limitations, this paper proposes a novel deep learning model called Lightweight Variable Dependency Aware Convolution Neural Network (LVDACNN) to analyse the measured aging data and make RUL prediction. The model consists of a data processing part and several LVDACNN blocks. The data processing part is designed to reduce the noise interference. Besides, an embedding scheme in the data processing part is proposed to enhance the variable relationship information; The designed LVDACNN block utilizes the time information in each variable and variable relationship to achieve accurate RUL prediction with low computation cost. Besides, to reduce the irrelevant information introduced by variable relationship learning, a channel attention mechanism is proposed. The LVDACNN model is evaluated using two public benchmark datasets for RUL prediction: The C-MAPSS dataset and the N-CMAPSS dataset. The experiment results demonstrate its superior performance compared to state-of-the-art methods.