Introduction
Accurately measuring shock signals is critical for product design and validation across various industries—from consumer electronics to aerospace. Capturing these signals in high-g environments, however, is notoriously difficult. While high-end sensors provide the necessary accuracy, they are prohibitively expensive. Conversely, low-end sensors are affordable but suffer from structural and material limitations, resulting in noisy and inaccurate readings.
Traditional calibration methods, such as using a Hopkinson bar, are physically complex and often only calibrate basic parameters like peak value or pulse width. This ignores the overall shape of the shock signal, which is essential for calculating the Shock Response Spectrum (SRS)—a crucial metric for structural testing. This paper introduces a deep learning approach to calibrate low-end sensors so their outputs match the fidelity of high-end equipment, offering a cost-effective software-based alternative to expensive hardware.
Proposed Methodology
The authors treat sensor calibration as a time-series translation task, proposing a novel neural network architecture that decouples the prediction of the signal's overall shape from its peak value.
The architecture consists of two main branches:
- Encoder-Decoder for Shape: The raw, noisy signal is normalized and fed into an encoder-decoder network to reconstruct the overall clean shape of the signal.
- Peak Prediction Network (PPN): Because the absolute peak value is critical in shock testing, a separate PPN branch takes the global feature vector from the encoder and the raw peak value to predict the true, calibrated peak.
The network is trained using a combined loss function: $L_2$ loss for shape matching, $L_\infty$ loss for relative peak alignment, and an $L_1$-style loss for absolute peak value prediction.
Experimental Setup and Dataset
As no public dataset existed for this problem, the authors created the first dataset for industrial shock signals. Using a drop shock test platform, they conducted 660 tests, capturing paired signals from both low-end and high-end sensors mounted to a drop table. The dataset covers accelerations up to 8,000 g, with signals sampled at 200kHz and standardized to a 15 ms duration (3,000 data points).
Results and Discussion
The proposed network was compared against traditional approaches, including Low Pass Filters (LPF), Linear Regression (LR), and standard Auto-Encoders (AE).
The results demonstrate significant improvements:
- Peak Error: The raw low-end sensor had a relative peak error of 13.5%. The proposed network reduced this to 5.7%, outperforming the standard AE (6.9%) and vastly outperforming the LPF (48.8%).
- Shape Error: The raw signal shape error was reduced from 228.6 to 35.2.
- Qualitative Performance: As shown in the frequency domain, the SRS curve of the calibrated low-end signal closely aligns with the high-end sensor's curve, proving its viability for real-world industrial applications.
A key insight is that the global shape of the shock signal is highly predictive of the true peak value; removing global features from the PPN caused the peak error to nearly double.
Conclusion
This study shows that deep learning can effectively bypass the physical limitations of cheap piezoceramic materials by learning the complex, non-linear noise profiles inherent in high-g impacts. This data-driven approach significantly lowers the financial barrier to high-g shock measurements, paving the way for wider adoption of deep learning in hardware calibration.
Summary updated