Highly automated driving places very high safety requirements on the reliability of the power supply in the vehicle. To meet these high requirements, new vehicle electrical system designs and redundancy concepts are being investigated. Replacing conventional melting fuses with electronic fuses can provide more voltage and current signals that can be used for the analysis of the system state and to establish advanced power system diagnosis to increase safety and reliability.
In general, faults in the power supply can have many root causes and may lead to component failures or damage. On-line diagnosis and fault-handling systems must be able to reliably detect, identify, and then isolate the faults. Ideally, this must happen in real time so that countermeasures can be initiated quickly.
In this contribution, a data-based and real-time diagnostic concept is investigated with a focus on detection and identification of faults. Compared to previous contributions the focus here is on low-cost hardware that can be integrated into automotive power supply systems. Based on measurement data from electronic fuses, a centralized evaluation of the on-board power supply system within a power distribution unit, consisting of several electronic fuses, is analyzed. A laboratory setup was created to automatically inject faults (e.g. wiring faults) to capture training, test and validation data. The data is recorded using a microcontroller (STM32 Nucleo-64 G474RE) in combination with development boards for electronic fuses (Infineon BTS50005-1LUA). At the same time, measurements are also performed using a high-resolution USB oscilloscope for comparison. The presentation focuses on the differences in sensor quality and diagnosis accuracy between the high performance oscilloscope measurements and the low resolution on-board measurements that can be expected from an electronic fuse. The detection times of different on-board power supply system faults through evaluation with LSTM (Long Short-Term Memory) neural networks are discussed and analyzed. The networks are then implemented on a Jetson Orin Nano for run-time evaluation.
