Micro-Electro-Mechanical Systems (MEMS) accelerometers play a critical role in safety-oriented applications, necessitating precise calibration methods that adhere to National Institute of Standards and Technology (NIST) guidelines of 2% accuracy. Traditional calibration techniques, relying on physical stimuli during the production phase, are often cost-prohibitive and time-consuming. This paper presents a novel calibration methodology utilizing electrical stimulation, specifically a DC stimulus, to excite the accelerometer. This approach simplifies the calibration electronics compared to AC stimulus calibration methods. By applying six distinct DC steps and measuring the resulting change in capacitance, we establish a correlation between the electrical response and the physical sensitivity of a selected batch of sensors. We employ machine learning algorithms to develop a predictive model that estimates the sensitivity of additional sensors within the batch based solely on the DC stimulus data. Experiments show that the proposed DC stimulus-only calibration method can achieve a root mean square (RMS) error of 1.15% for lot-to-lot process variations and 0.36% for within-lot process variations. This innovative calibration strategy not only enhances accuracy but also significantly reduces both testing costs and time.

Micro Electro-Mechanical Systems (MEMS) accelerometers are utilized in safety-critical applications, including airbags, aircraft, medical devices, as well as various consumer electronic applications. Despite providing highly accurate results, they require calibration during production and periodic recalibration due to potential degradation over time. The National Institutes of Standards and Technology (NIST) stipulates that the accuracy of motion sensors used in safety-critical applications must be maintained within a 1% error margin. In our research, we propose an electrical stimulus-based calibration method for sensors during production and in-field use. In-field electrical stimulation and calibration can facilitate prolonged sensor operation without the need to remove the sensor from its environment. Although electrical stimulation has been suggested as a replacement for physical stimulation to reduce testing costs for sensors, it has not yet demonstrated the ability to meet the 1% error requirement mandated by NIST safety standards. We propose an incremental sensor-based model that can correlate sensor sensitivity degradation and process variation to its electrical response for in-field monitoring. Simulations demonstrate that the model can forecast sensitivity changes within a 1% error margin. Additionally, we have developed a cost-effective rotating test platform for calibrating and measuring accelerometer sensitivity. This method utilizes wireless technology to transmit accelerometer data to a computer, eliminating the necessity for lengthy cables. The rotating platform can produce various accelerations at different distances along the radius by spinning at different RPMs, leveraging centripetal force to apply a fixed acceleration at a set RPM.

This paper introduces a sensor calibration and gas source localization method designed for the challenging task of detecting gas leaks in open environments using multiple low-cost sensors that can be randomly distributed in a coverage area. Low-cost sensors, such as Metal Oxide resistive sensors, suffer from very large sensor-to-sensor variations in their detection range, noise floor, and limited sensitivity. The proposed approach begins with a self-calibration process that leverages known source locations to fine-tune their performance parameters. The independent calibration of sensors results in more precise measurements. Furthermore, we propose a computationally efficient technique that achieves high-precision gas source localization. Our methodology is compared with existing localization techniques to demonstrate its effectiveness.

In this work, we present a method for detecting anomalous chemical sensors using contrastive learning-based framework. In many practical systems, an array of multiple chemical sensors are used. Some of the sensors may malfunction due to sensor drift and chemical poisoning. In standard contrastive learning, the aim is to learn representations that will have maximum agreement among data samples of the same concept while having a minimal agreement with data samples from other concepts. In this work, we adapt standard contrastive learning to learning useful representations for out-of-distribution sample detection. Furthermore, we compare the proposed framework with the cosine similarity measure and a novel similarity measure based on the ℓ1norm. Our experimental results show that our approach achieves higher AUC scores (93.6%) than baseline methods (90.1%).

Micro Electro-Mechanical Systems (MEMS) accelerometers are used in safety critical applications, such as airbags and airplanes. While providing very accurate results, they can degrade over time due to many wearout mechanisms. According to the National Institutes of Standards and Technology (NIST), the accuracy of motion sensors used in safety critical applications needs to be maintained within 1% error. In-field electrical stimulation and calibration can enable long-term sensor operation without removing the sensor from its environment. While electrical stimulation has been proposed to replace the physical stimulation to reduce testing cost for sensors, it has not yet been shown to achieve the 1% error requirement as required by the NIST standard of safety. In this paper, we propose an incremental sensor-based model that can relate the degradation in the sensitivity of the sensor to its electrical response for infield monitoring. In order to extract such a relation, we need to generate multiple sensitivity states for the sensor however, which is not possible using the normal mode of operation. We propose to temporary place the sensor in an enhanced state where the sensitivity can be changed also via electrical signalling, thereby generating an adequate number of measurements to solve for model coefficients. We show through simulations and hardware experiments that the model can predict the sensitivity changes within 1% error.