Increasing the reliability of automotive power electronics
At present, information about the health state of power-electronic modules, such as DC/DC converters used in heavy-duty hybrid/battery electric vehicles, is typically inferred from statistical analysis of data from on-board sensors. Note that these traditional approaches don’t resource to actual physical phenomena behind the degradation mechanisms and, thus, their prognosticating capabilities are very limited. Predicting the Remaining Useful Life (RUL) becomes particularly challenging for components operating under specific or varying conditions or whose design differs from the original design for which the historical benchmark data have been acquired.
A promising way to achieve higher reliability of the automotive power electronics, explored in the iRel40’s industrial pilot IP-13, is through the application of advanced hybrid approaches combining data-driven AI/ML models with Physics-of-failure (PoF) methods. By utilizing the knowledge of specific mechanisms governing the degradation process, such approaches allow for enhancing the performance of the predictive models in the detection of incipient failures, especially for specific mission-load profiles.
The main focus of this industrial pilot has been on the key components of DC/DC converters, that is, on power transistors which subject to varying loads and temperatures with time suffer from bond-wire lift-off. As a starting point, in order to establish a database for developing and testing new predictive models, a comprehensive set of run-to-failure sensor data has been generated by power cycling SiC MOSFETs under diverse loadings, see Figure 1. Since all types of predictive models require typically at some stage training on complete run-to-failure trajectories, the application of the acquired data goes far beyond the objectives of the iRel40 project.
Figure 1: Schematic summary of key steps and results obtained in the frame of the iRel40’s industrial pilot IP-13.
To facilitate the acquisition and flow of data along the value chain, which also fulfills one of the objectives of the iRel40 project, a customized Python-based framework for processing and analyzing these sensor data has been developed. Importantly, this framework has been instrumental in enabling AI integration as a prognostic tool and in the future can be further adapted for processing various types of sensor data obtained via component cycling and having the form of time series.
In the next step, the training sensor data have been used to develop a universal method that allows for removing the effects of the operating point and extraction of accumulated degradation, see Figure 1. In particular, this accumulated degradation is derived from the drain-source ON-state voltage using operating conditions (input current and ambient temperature) and it corresponds to the sought-for data feature. This method/feature is of crucial importance for developing both data-driven and Physics-informed predictive models applicable to different devices working under varying experimental conditions. Furthermore, the proposed method can generally be adopted for extracting accumulated damage in various electronic components where the primary degradation mechanism is associated with crack propagation due to cyclic stress, induced by temperature swings, and resulting from a mismatch between coefficients of thermal expansion (CTE) for metal and semiconductor elements.
Employing the derived training data, machine/deep learning (ML/DL) architectures for predicting the Remaining Useful Life (RUL) of SiC MOSFETs have been implemented. In particular, three models based on deep neural networks suitable for handling tasks involving time series were tested and evaluated: a temporal convolutional neural network
(TCN), a long short-term memory neural network (LSTM) and a convolutional gated recurrent neural network (Conv-GRU). In Figure 2, a summary of RUL plots for each of the three models is shown for test data obtained for two example devices in the lowest accelerated round of the power-cycling experiment. It can be observed that the models can capture the general trend of decreasing RUL, with varying accuracy. Interestingly enough, the Conv-GRU model outperforms other models and is less computationally complex, being trained with fewer parameters.
Figure 2: Comparison of RUL predictions obtained with various (data-driven) ML/DL models for two example devices in the test data set.
Although such data-driven ML/DL models are usually very accurate and reliable in yielding device-specific predictions, they are not interpretable in terms of the relation between variations in input variables/data and the source of degradation mechanisms. Moreover, the performance of such models deteriorates when they are applied to data collected under operational conditions significantly deviating from those used when having acquired the training data. A possible solution allowing for circumventing this limitation is to supply the algorithm with additional information based on Physics-of-failure (PoF).
For this reason, a hybrid model assuming the bond-wire lift-off as the dominating failure mode and using a damage accumulation scheme based on Paris’ crack law has been developed and tested, see Figure 1. This model uses essentially only two inputs: the average junction temperature swing and the temperature-compensated drain-source ON-state resistance at the peak temperature of the current cycle. Using only these two input values, the
model is capable of predicting RUL with surprising accuracy for the range of constant current loads determining cycling conditions under which the training data series have been acquired. Importantly, this model is a crucial step towards building more elaborate prognostic schemes for RUL-determination of SiC power MOSFETs in actual working conditions, using physics-informed neural networks (PINNs).
To model the DC/DC converter and its electronic circuit (in particular, the H-bridge together with its four SiC MOSFET), the digital twin has been constructed using the ANSYS/TwinBuilder software. For this purpose, a detailed model of SiC MOSFET has been set up in high-fidelity 3D-simulation tools including the electro-thermal sources from the current flow through the transistor, see Figure 1. Next, this electro-thermal model has been validated with the available temperature measurements from a power-cycle rig. Importantly, such a high-fidelity model provides detailed information about the thermal response of the MOSFET junction temperature. As a result, utilizing data from different operational cycles (voltage and currents), the digital twin computes the temperature response of the transistors. The output information of the junction temperature of the transistors can then be further used in RUL calculations.
To conclude, the key achievement of the iRel40’s industrial pilot IP-13 is a demonstration of the ability to monitor health and prognosticate the RUL of wire bonds in power transistors. In particular, the obtained results provide a proof of concept for an approach allowing for estimation of the health state and deterioration of the power-electronic components from operational data. The novel aspect of the studied methods is that they allow for taking into account the effect of varying operational conditions on the reliability of the system. When implemented in a DC/DC converter, this approach should enable the converter to learn the operating behavior of a specific truck, which essentially means that this approach should make it possible to identify deviations from the normal operation occurring over time more accurately.
Wilhelm Söderkvist Vermelin (RISE), Mattias Eng (RISE), Maciej Misiorny (QRTECH), Jonas Bredberg (EDRMEdeso), Henrik Tryggeson (EDERMedeso), Poli Ninos (Scania)
 Wilhelm Söderkvist Vermelin et al., “Data-Driven Remaining Useful Life Estimation of Discrete Power Electronic Devices”, Proceedings of the 33rd European Safety and Reliability Conference (ESREL 2023), 2595, Eds. Mário P. Brito, Terje Aven, Piero Baraldi, Marko Čepin and Enrico Zio, Research Publishing, Singapore, pp. 2595-2602 (2023), available online.
 Mattias Eng et al., “A Simple Hybrid Model for Estimating Remaining Useful Life of SiC MOSFETs in Power Cycling Experiments”, Proceedings of the Asia Pacific Conference of the PHM Society 2023, vol. 4, no. 1 (2023), available online.
Power electronics, MOSFET, Wire bond, Prognostics and health management, Remaining useful life, Deep learning, Physics-of-failure, Power cycling.