<p>When equipped with a reliable calibration model, Low-Cost Sensor (LCS) can be relied upon as an effective option for gas concentration estimation, providing robust and high spatio-temporal resolution data to replace large-scale analytical instruments. In this paper, we present ProxySense, a rapid and efficient approach for gas concentration estimation. The ProxySense pipeline consists of gas sensing unit made up of array of metal oxide LCS, data pre-processing including an effective approach based on Variatioanl Autoencoders (VAE) for handling missing sensor data and Long Short Term Memory Reccurrent Neural Network (LSTM-RNN) prediction model. We investigate the capability of ProxySense in exploiting the deep characteristics that exist in multi-sensors’ responses to predict the concentration of a gas for which no specific sensor is included in a multi-sensor device. We evaluated ProxySense for benzene (C6H6) and carbon monoxide (CO) concentration predictions and compared the performances to multiple baselines by means of prediction error characterization. We further studied the relationship between model performance and training length and showed ProxySense to be highly accurate for gas concentration prediction even for small number of training period.</p>

<p>When equipped with a reliable calibration model, Low-Cost Sensor (LCS) can be relied upon as an effective option for gas concentration estimation, providing robust and high spatio-temporal resolution data to replace large-scale analytical instruments. In this paper, we present ProxySense, a rapid and efficient approach for gas concentration estimation. The ProxySense pipeline consists of gas sensing unit made up of array of metal oxide LCS, data pre-processing including an effective approach based on Variatioanl Autoencoders (VAE) for handling missing sensor data and Long Short Term Memory Reccurrent Neural Network (LSTM-RNN) prediction model. We investigate the capability of ProxySense in exploiting the deep characteristics that exist in multi-sensors’ responses to predict the concentration of a gas for which no specific sensor is included in a multi-sensor device. We evaluated ProxySense for benzene (C6H6) and carbon monoxide (CO) concentration predictions and compared the performances to multiple baselines by means of prediction error characterization. We further studied the relationship between model performance and training length and showed ProxySense to be highly accurate for gas concentration prediction even for small number of training period.</p>

No abstract