Till Örebro universitet

Gas Distribution Mapping (GDM) is essential in monitoring hazardous environments, where uneven sampling and spatial sparsity of data present significant challenges. Traditional methods for GDM often fall short in accuracy and expressiveness. Modern learning-based approaches employing Convolutional Neural Networks (CNNs) require regular-sized input data, limiting their adaptability to irregular and sparse datasets typically encountered in GDM. This study addresses these shortcomings by showcasing Graph Neural Networks (GNNs) for learningbased GDM on irregular and spatially sparse sensor data. Our Radius-Based, Bi-Directionally connected GNN (RABI-GNN) was trained on a synthetic gas distribution dataset on which it outperforms our previous CNN-based model while overcoming its constraints. We demonstrate the flexibility of RABI-GNN by applying it to real-world data obtained in an industrial steel factory, highlighting promising opportunities for more accurate GDM models.

Gas Distribution Mapping (GDM) is a valuable tool for monitoring the distribution of gases in a wide range of applications, including environmental monitoring, emergency response, and industrial safety. While GDM is actively researched in the scope of gas-sensitive mobile robots (Mobile Robot Olfaction), there is a potential for broader applications utilizing sensor networks. This study aims to address the lack of deep learning approaches in GDM and explore their potential for improved mapping of gas distributions. In this paper, we introduce Gas Distribution Decoder (GDD), a learning-based GDM method. GDD is a deep neural network for spatial interpolation between sparsely distributed sensor measurements that was trained on an extensive data set of realistic-shaped synthetic gas plumes based on actual airflow measurements. As access to ground truth representations of gas distributions remains a challenge in GDM research, we make our data sets, along with our models, publicly available. We test and compare GDD with state-of-the-art models on synthetic and real- world data. Our findings demonstrate that GDD significantly outperforms existing models, demonstrating a 35% improvement in accuracy on synthetic data when measured using the Root Mean Squared Error over the entire distribution map. Notably, GDD appears to have superior capabilities in reconstructing the edges and characteristic shapes of gas plumes compared to traditional models. These potentials offer new possibilities for more accurate and efficient environmental monitoring, and we hope to inspire other researchers to explore learning-based GDM.

The Sharp GP2Y1010AU0F is a widely used low-cost dust sensor, but despite its popularity, the manufacturer provides little information on the sensor. We installed 16 sensing nodes with Sharp dust sensors in a hot rolling mill of a steel factory. Our analysis shows a clear correlation between sensor drift and accumulated production of the steel factory. An eye should be kept on the long-term drift of the sensors to prevent early saturation. Two of 16 sensors experienced full saturation, each after around eight and ten months of operation.

To better understand the dynamics in hazardous environments, gas distribution mapping aims to map the gas concentration levels of a specified area precisely. Sampling is typically carried out in a spatially sparse manner, either with a mobile robot or a sensor network and concentration values between known data points have to be interpolated. In this paper, we investigate sequential deep learning models that are able to map the gas distribution based on a multiple time step input from a sensor network. We propose a novel hybrid convolutional LSTM - transpose convolutional structure that we train with synthetic gas distribution data. Our results show that learning the spatial and temporal correlation of gas plume patterns outperforms a non-sequential neural network model.

Gas distribution mapping is important to have an accurate understanding of gas concentration levels in hazardous environments. A major problem is that in-situ gas sensors are only able to measure concentrations at their specific location. The gas distribution in-between the sampling locations must therefore be modeled. In this research, we interpret the task of spatial interpolation between sparsely distributed sensors as a task of enhancing an image's resolution, namely super-resolution. Because autoencoders are proven to perform well for this super-resolution task, we trained a convolutional encoder-decoder neural network to map the gas distribution over a spatially sparse sensor network. Due to the difficulty to collect real-world gas distribution data and missing ground truth, we used synthetic data generated with a gas distribution simulator for training and evaluation of the model. Our results show that the neural network was able to learn the behavior of gas plumes and outperforms simpler interpolation techniques.

A fully designed low-cost sensing node for air pollution monitoring and calibration results for several low-cost gas sensors are presented. As the state of the art is lacking information on the importance of an active ventilation system, the effect of an active fan is compared to the passive ventilation of a lamellar structured casing. Measurements obtained in an urban outdoor environment show that readings of the low-cost dust sensor (Sharp GP2Y1010AU0F) are distorted by the active ventilation system. While this behavior requires further research, a correlation with temperature and humidity inside the node shown.

Wireless sensor networks provide occupational health experts with valuable information about the distribution of air pollutants in an environment. However, especially low-cost sensors may produce faulty measurements or fail completely. Consequently, not only spatial coverage but also redundancy should be a design criterion for the deployment of a sensor network. For a sensor network deployed in a steel factory, we analyze the correlations between sensors and build machine learning forecasting models, to investigate how well the sensor network can compensate for the outage of sensors. While our results show promising prediction quality of the models, they also indicate the presence of spatially very limited events. We, therefore, conclude that initial measurements with, e.g., mobile units, could help to identify important locations to design redundant sensor networks.

Wireless sensor networks provide occupational health experts with valuable information about the distribution of air pollutants in an environment. However, especially low-cost sensors may produce faulty measurements or fail completely. Consequently, not only spatial coverage but also redundancy should be a design criterion for the deployment of a sensor network. For a sensor network deployed in a steel factory, we analyze the correlations between sensors and build machine learning forecasting models, to investigate how well the sensor network can compensate for the outage of sensors. While our results show promising prediction quality of the models, they also indicate the presence of spatially very limited events. We, therefore, conclude that initial measurements with, e.g., mobile units, could help to identify important locations to design redundant sensor networks.

Air pollution within industrial scenarios is a major risk for workers, which is why detailed knowledge about the dispersion of dusts and gases is necessary. This paper introduces a system combining stationary low-cost and high-quality sensors, carried by ground robots and unmanned aerial vehicles. Based on these dense sampling capabilities, detailed distribution maps of dusts and gases will be created. This system enables various research opportunities, especially on the fields of distribution mapping and sensor planning. Standard approaches for distribution mapping can be enhanced with knowledge about the environment's characteristics, while the effectiveness of new approaches, utilizing neural networks, can be further investigated. The influence of different sensor network setups on the predictive quality of distribution algorithms will be researched and metrics for the quantification of a sensor network's quality will be investigated.