Publications

Smoldering fires are characterized by the production of early gas emissions that can include high levels of CO and Volatile Organic Compounds (VOCs) due to pyrolysis or thermal degradation. Nowadays, standalone CO sensors, smoke detectors, or a combination of these, are standard components for fire alarm systems. While gas sensor arrays together with pattern recognition techniques are a valuable alternative for early fire detection, in practice they have certain drawbacks-they can detect early gas emissions, but can show low immunity to nuisances, and sensor time drift can render calibration models obsolete. In this work, we explore the performance of a gas sensor array for detecting smoldering and plastic fires while ensuring the rejection of a set of nuisances. We conducted variety of fire and nuisance experiments in a validated standard fire room (240 m(3)). Using PLS-DA and SVM, we evaluate the performance of different multivariate calibration models for this dataset. We show that calibration models remain predictive after several months, but perfect performance is not achieved. For example, 4 months after calibration, a PLS-DA model provides 100% specificity and 85% sensitivity since the system has difficulties in detecting plastic fires, whose signatures are close to nuisance scenarios. Nevertheless, our results show that systems based on gas sensor arrays are able to provide faster fire alarm response than conventional smoke-based fire alarms. We also propose the use of small-scale fire experiments to increase the number of calibration conditions at a reduced cost. Our results show that this is an effective way to increase the performance of the model, even when evaluated on a standard fire room. Finally, the acquired datasets are made publicly available to the community (doi: 10.5281/zenodo.5643074).

JTD Keywords: Calibration, Chemical sensors, Co2, Early fire, Early fire detection, En-54, Fire alarm, Fire detection, Fire room, Fires, Gas detectors, Gas emissions, Gas sensors, Pattern recognition, Public dataset, Sensor arrays, Sensors array, Signatures, Smoke, Smoke detector, Smoke detectors, Standard fire, Standard fire room, Support vector machines, Temperature, Toxicity, Volatile organic compounds

Metal Oxide (MOX) gas sensors rely on chemical reactions that occur efficiently at high temperatures, resulting in too-demanding power requirements for certain applications. Operating the sensor under a Pulsed-Temperature Operation (PTO), by which the sensor heater is switched ON and OFF periodically, is a common practice to reduce the power consumption. However, the sensor performance is degraded as the OFF periods become larger. Other research works studied, generally, PTO schemes applying waveforms to the heater with time periods of seconds and duty cycles above 20%. Here, instead, we explore the behaviour of PTO sensors working under aggressive schemes, reaching power savings of 99% and beyond with respect to continuous heater stimulation. Using sensor sensitivity and the limit of detection, we evaluated four Ultra Low Power (ULP) sensors under different PTO schemes exposed to ammonia, ethylene, and acetaldehyde. Results show that it is possible to operate the sensors with total power consumption in the range of microwatts. Despite the aggressive power reduction, sensor sensitivity suffers only a moderate decline and the limit of detection may degrade up to a factor five. This is, however, gas-dependent and should be explored on a case-by-case basis since, for example, the same degradation has not been observed for ammonia. Finally, the run-in time, i.e., the time required to get a stable response immediately after switching on the sensor, increases when reducing the power consumption, from 10 minutes to values in the range of 10–20 hours for power consumptions smaller than 200 microwatts.

JTD Keywords: Robot sensing systems, Temperature sensors, Heating systems, Gas detectors, Power demand, Sensitivity, Electronic nose, gas sensors, low-power operation, machine olfaction, pulsed-temperature operation, temperature modulation

Low-cost gas sensors allow for large-scale spatial monitoring of air quality in the environment. However they require calibration before deployment. Methods such as multivariate regression techniques have been applied towards sensor calibration. In this work, we propose instead, the use of deep learning methods, particularly, recurrent neural networks for predicting the gas concentrations based on the outputs of these sensors. This paper presents a first study of using Gated Recurrent Unit (GRU) neural network models for gas concentration prediction. The GRU networks achieve on average, a 44.69% and a 25.17% RMSE improvement in concentration prediction on a gas dataset when compared with Support Vector Regression (SVR) and Multilayer Perceptron (MLP) models respectively. With the current advances in deep network hardware accelerators, these networks can be combined with the sensors for a compact embedded system suitable for edge applications.

JTD Keywords: Robot sensing systems, Predictive models, Logic gates, Gas detectors, Training, Temperature measurement, Support vector machines

The intermittency of the instantaneous concentration of a turbulent chemical plume is a fundamental cue for estimating the chemical source distance using chemical sensors. Such estimate is useful in applications such as environmental monitoring or localization of fugitive gas emissions by mobile robots or sensor networks. However, the inherent low-pass filtering of metal oxide (MOX) gas sensors-typically used in odor-guided robots and dense sensor networks due to their low cost, weight and size-hinders the quantification of concentration intermittency. In this paper, we design a digital differentiator to invert the low-pass dynamics of the sensor response, thus obtaining a much faster signal from which the concentration intermittency can be effectively computed. Using a fast photo-ionization detector as a reference instrument, we demonstrate that the filtered signal is a good approximation of the instantaneous concentration in a real turbulent plume. We then extract transient features from the filtered signal-the so-called “bouts”-to predict the chemical source distance, focusing on the optimization of the filter parameters and the noise threshold to make the predictions robust against changing wind conditions. This represents an advantage over previous bout-based models which require wind measurements-typically taken with expensive and bulky anemometers-to produce accurate predictions. The proposed methodology is demonstrated in a wind tunnel scenario where a MOX sensor is placed at various distances downwind of an emitting chemical source and the wind speed varies in the range 10-34 cm/s. The results demonstrate that models optimized with our methodology can provide accurate source distance predictions at different wind speeds.

JTD Keywords: Gas detectors, Chemical sensors, Signal processing, Machine learning, Time series analysis