A Deep Learning Approach to Recover High-g Shock Signals From the Faulty Accelerometer

Abstract

A deep learning based approach is proposed to accurately recover shock signals measured from a damaged high-g accelerometer without modifying the hardware. We first conducted shock tests and collected a large dataset of shock signals with different levels of acceleration by using an efficient experimental apparatus. The training data is composed of a pair of signals simultaneously obtained from a faulty accelerometer and a high-end accelerometer (served as the ground truth). A customized autoencoder neural network is designed and trained on this dataset, aiming to map the faulty signals to their reference counterparts. Experimental results show that, with the help of deep learning, shock signals can be accurately recovered from the faulty measurements. Compared with conventional approaches that require diagnosing and replacing faulty parts, the proposed data-driven method demonstrates a highly promising solution that allows recovering corrupted signals without introducing extra work to upgrade the hardware at almost zero cost. The dataset and code of this work are made publicly available on GitHub at https://github.com/hope-yao/Sensor_Calibration.

On Fault Diagnosis for High-G Accelerometers via Data-Driven Models

“…Shock environments are described by acceleration-time signals generally, which can be measured by accelerometers [6]. However, high-g accelerometers sometimes can get damaged during the shock tests due to the severe impact environment [7]. These damaged accelerometers cannot measure accurate shock signals, which would lead to measuring uncertainties of various unpredictable levels, and, hence, test failures.…”

“…The major damage types of accelerometer's inner components include cantilever fractures, wire bond shearing, solder joint loss, chip cracks, [10], [11] etc. Correspondingly, inner component damages will cause the waveform variation of the accelerometer's outputs, such as the peak truncation [7], noise pollution [8], and baseline drift [12]. Therefore, it would be of great value to be able to automatically diagnose the accelerometer's fault type through its readings.…”

Self-validating high-g accelerometers through data-driven methods

“…Under many circumstances, however, the acquired shock data is unique, and repeating the pyroshock test can be extremely difficult or costly [7]. A typical example is that when the to-be-tested electronic device and the accelerometer are both damaged at the same time during the high-g shock test [8]. Unfortunately, it is quite costly to repeat this destructive test, and, meanwhile, the faulty accelerometer cannot provide accurate measurements to help analyzing the damage causes of the defective electronic device.…”

“…These successes also inspired us to apply this powerful tool into the self-validation of high-g accelerometers. In our previous work, two deep neural networks (DNNs) were developed to recover shock signals of two forms of fault respectively [8,9]. In addition, we also proposed an ensemble learning-based model to identify the fault types of high-g accelerometers by integrating several data-driven-based sub-classifiers [27].…”

“…This hybrid method integrates the spline interpolation algorithm and Savitzky-Golay filtering and provides improvements in multiple metrics with lower calculation costs. Inspired by these works, a hybrid strategy is proposed in this work to self-validate high-g accelerometers by integrating the ensemble kNNs and the DNN-based recovery method in our previous works [8,9,27].…”