tinyAI Forum on PdM & Anomaly Detection: DC Series Arc Fault Detection (Anomaly Detection) using…
DC Series Arc Fault Detection (Anomaly Detection) using 1D-CNN
Adithya THONSE, Embedded Software Engineer, Texas Instruments
In photovoltaic grids, charging stations, and home inverters, there exist AC as well as DC systems. Unlike in the case of AC systems, as there is no zero-crossing point in the current waveform for DC systems. When wires get worn off or contacts get broken, the resultant DC arc is more sustainable as a result, it stands to be a greater threat and a huge fire hazard. Series arc fault leads to a reduced fault current resulting in increased impedance and, hence, cannot be detected by traditional protection devices. As a result, extensive research has been conducted on different techniques for series arc detection.
The UL 1699B safety standard provides test procedures to qualify an arc fault detection device. It includes tests for both the effectiveness of detecting as well as immunity to false detection. Frequency-domain-based arc fault detection methods involving spectral analysis by Fourier transform methods are the most popular. Using Fast Fourier Transform makes signal processing computationally efficient, and has been implemented in microcontrollers for household purposes extensively in the current world. However, the spectral signature changes as the device undergoes aging and thereby ends up lowering the accuracy of the conventional algorithm. Since Time-domain methods are computationally most efficient, but conventional algorithms are prone to nuisance tripping, it’s time to use Deep Learning to tackle such problems.
The data used in this experiment was created with an arc generation setup under laboratory conditions for a variety of voltages (200V, 300V, 400V) and different current (5A,6A,7A) combinations. Raw data was sampled at 313kHz.
The solution presented here is a variable resolution deep learning classification network comprising 4 layers of CNN with a channel depth of 16 each followed by a Fully connected network. The computational complexity of the network is 557k MACs. Multiple CNNs were experimented with to arrive at the simplest and best network to tackle this problem. The dynamic resolution here implies that the time domain input sequence is comprised of 750 data points. This means that the sequence duration chosen for it to be identified as an arc or not varied between 30ms to 250ms whilst simultaneously varying the frequency in the opposite direction. That is, for a 250ms window – data was sampled at 3kHz to effectively give 750 data points, and a 50ms window was sampled at 15kHz to give a 750 data point input. This dynamic scaling ended up boosting the system arc detection accuracy to upwards of 97% compared to the 91% accuracy given by the conventional FFT algorithm. These experiments are tested on a TI 120MHz C2000TM MCU with 69kB on-chip RAM.
The added benefit of this system lies in the ability to retrain this system dynamically on the field which is very much a necessity due to the aging of the system. Given the conventional FFT-based approach, this would be a hard ask as the frequency characteristics change in an unpredictable direction which makes it harder to retrain. Although, the presented work is showcased for arc detection but can be extended for any application involving real-time 1D time series data.
source
