The rapid implementation of Internet of Things (IoT) technologies in industrial and air quality monitoring applications has resulted in large amounts of data being acquired. In addition, the integration of the artificial intelligence of things (AIoT) enables decision-making at edge nodes, involving the execution of data-driven models close to the data source. Data loss has become a limiting factor in AI-based applications. One way to address missing data is to leverage data from co-existing sensors, which are potentially correlated, to impute missing values. This article evaluates a set of missing value imputation (MVI) techniques in an edge node environment, where constraints on data availability and computational complexity must be met. Specifically, models such as multiple imputation by chained equations (MICE), k-nearest neighbors (KNN), and models using variational autoencoders (VAE) are evaluated. A comprehensive evaluation is presented covering different scenarios of missing data in terms of the percentage of missing values, bursts of missing values, and the size of the data windows stored on edge nodes. This evaluation uses real sensor data from an air quality monitoring network and an industrial sensor network. The results show that the VAE is able to obtain good imputation performances (R 2 greater than 0.90) while providing good uncertainty quantification (UQ) estimates with around 90%–95% of samples falling within the estimated confidence intervals. Moreover, a transfer learning-based VAE has been shown to adapt to the non-stationary nature of IoT sensing signals, and all methods have proven to be efficient in terms of execution time for the edge setting. • Missing value imputation for edge IoT applications. • Window-based models for the edge. • Variational autoencoders (VAE) for imputing missing values in IoT. • Uncertainty quantification (UQ) using heteroscedastic VAEs. • Transfer learning-based VAE for imputation in the edge.

The use of low-cost sensors (LCS) in Internet of Things (IoT) networks offers a promising way to improve air quality monitoring. However, there is a major concern regarding their long-term accuracy due to continuous data drift, which requires frequent data recalibration. To address this, we present window-based uncertainty drift detection and recalibration (W-UDDR), a unified system that automates the entire process. Our system uses a Bayesian approach with Gaussian processes (GP) to automatically and accurately detect when sensor output needs to be corrected. To achieve this, the uncertainty of the estimates is quantified using predictive confidence intervals alongside a window system that detects the need for recalibration in real time. We validate our approach using an air quality real-world sensor deployment, systematically assessing key performance metrics such as the frequency of recalibration and the required sample size. Our results show that W-UDDR successfully triggers automatic events when it detects drifts, achieving significant long-term accuracy improvements ranging from 40% to 95%. Hence, this tool provides an automated real-time mechanism to facilitate long-term sensor deployment maintenance.

Recently, the use of Internet of Things (IoT) technologies has been widely adopted, enabling large-scale monitoring and data acquisition. Together with the so-called artificial intelligence of things (AIoT), data-driven models are to be executed close to the data source. Nevertheless, one potential problem is the quality of the data used to feed these models. Data loss is a common issue in these systems and can hinder the use of monitoring data and the application of artificial intelligence (AI) models. This paper evaluates a set of missing value imputation techniques, emphasizing their application in an edge setting. Specifically, offline and sliding window variants are evaluated to meet data availability and computational complexity criteria. The benchmarking has been performed using the data of a real IoT air quality monitoring platform. The results demonstrate that the windowed variants can approach the offline performance using reduced sliding windows of 25 to 100 samples per sensor.

A digital twin (DT) is a virtual representation of a physical object or system created using data obtained through numerical simulation or sensor measurements from Internet of Things (IoT) technologies, among others. These measurements allow the virtual model to be specific to a particular object or system, enabling diagnosis, predictive, and prescriptive maintenance. In this paper, we present the development of a DT for an IoT monitoring network with applications in Industry 5.0. To achieve this, we employ reduced-order models (ROM) based on the proper orthogonality decomposition (POD) technique to reconstruct a field of interest, e.g., temperature in a cold room or on a pig livestock farming, from a reduced set of optimally placed sensors. The results demonstrate that optimal placement can achieve low reconstruction errors even when only a few sensors are deployed - for example, three or four.

Intelligent computing at the IoT edge allows tasks that are normally performed in the cloud to be performed at the IoT node, enabling low-latency applications and more efficient control and management of services. Among the tasks that can be performed on the IoT edge is improving data quality, such as outlier detection. This task is challenging because IoT nodes have fewer computational and storage resources than the cloud. In this paper, we propose an outlier detection framework adapted to IoT edge nodes that is lightweight, consumes few resources, adapts to changes in signal trends, and has the ability to provide real-time responses. The framework presented is based on the use of two windows. A sliding window for the current data, which captures the short-term changes in the signal, and a window that stores a summary of historical values, which captures whether the values are within the long-term range of the signal. We show how models using two windows reduce the number of false positives compared to models using only one window. A version for near real-time applications is also proposed, which improves detection by identifying trend changes in the signal at the cost of delaying the decision by a few samples.

The development of Internet of Things (IoT) technologies has led to the widespread adoption of monitoring networks for a wide variety of applications, such as smart cities, environmental monitoring, and precision agriculture. A major research focus in recent years has been the development of graph-based techniques to improve the quality of data from sensor networks, a key aspect of the use of sensed data in decision-making processes, digital twins , and other applications. Emphasis has been placed on the development of machine learning (ML) and signal processing techniques over graphs, taking advantage of the benefits provided by the use of structured data through a graph topology . Many technologies such as graph signal processing (GSP) or the successful graph neural networks (GNNs) have been used for data quality enhancement tasks. This survey focuses on graph-based models for data quality control in monitoring sensor networks. In addition, it introduces the technical details that are commonly used to provide powerful graph-based solutions for data quality tasks in sensor networks, such as missing value imputation, outlier detection , or virtual sensing. To conclude, different challenges and emerging trends have been identified, e.g., graph-based models for digital twins or model transferability and generalization.

The introduction of high- and low-cost Internet of Things (IoT) sensors in air quality monitoring networks, in addition to providing a cost-effective solution for monitoring pollutant levels, also brings with it the challenge of ensuring data reliability. These sensors can present anomalies in the data and identifying them is a challenging task. In this article, we propose a spatiotemporal correlation recurrent autoencoder anomaly detection (STC-RAAD) architecture that unlike the other existing architectures, in addition to the temporal correlation present in the sensor data, also involves information from the neighboring sensors in the monitoring network for better reconstruction. The performance of STC-RAAD is compared with the other methods on two different real data sets, for two categories of anomalies, outperforming the other existing approaches with an average increase of 38% for the detection rate and 36% for the precision.

Air pollution monitoring platforms play a very important role in preventing and mitigating the effects of pollution. Recent advances in the field of graph signal processing have made it possible to describe and analyze air pollution monitoring networks using graphs. One of the main applications is the reconstruction of the measured signal in a graph using a subset of sensors. Reconstructing the signal using information from neighboring sensors is a key technique for maintaining network data quality, with examples including filling in missing data with correlated neighboring nodes, creating virtual sensors, or correcting a drifting sensor with neighboring sensors that are more accurate. This article proposes a signal reconstruction framework for air pollution monitoring data where a graph signal reconstruction model is superimposed on a graph learned from the data. Different graph signal reconstruction methods are compared on actual air pollution data sets measuring O3, NO2, and PM10. The ability of the methods to reconstruct the signal of a pollutant is shown, as well as the computational cost of this reconstruction. The results indicate the superiority of methods based on kernel-based graph signal reconstruction, as well as the difficulties of the methods to scale in an air pollution monitoring network with a large number of low-cost sensors. However, we show that the scalability of the framework can be improved with simple methods, such as partitioning the network using a clustering algorithm.

Existing air pollution monitoring networks use reference stations as the main nodes. The addition of low-cost sensors calibrated in-situ with machine learning techniques allows the creation of heterogeneous air pollution monitoring networks. However, current monitoring networks or calibration techniques have limitations in estimating missing data, adding virtual sensors or recalibrating sensors. The use of graphs to represent structured data is an emerging area of research that allows the use of powerful techniques to process and analyze data for air pollution monitoring networks. In this article, we compare two techniques that rely on structured data, one based on statistical methods and the other on signal smoothness, with a baseline technique based on the distance between nodes and that does not rely on the measured signal data. To compare these techniques, the sensor signal is reconstructed with a supervised method based on linear regression and a semisupervised method based on Laplacian interpolation, which allows reconstruction even when data is missing. The results, on data sets measuring O3, NO2, and PM10, show that the signal smoothness-based technique behaves better than the other two, and used together with the Laplacian interpolation is near optimal with respect to the linear regression method. Moreover, in the case of heterogeneous networks, the results show a reconstruction accuracy similar to the in-situ calibrated sensors. Thus, the use of the network data increases the robustness of the network against possible sensor failures.

This article investigates the calibration of low-cost sensors for air pollution. The sensors were deployed on three Internet of Things (IoT) platforms in Spain, Austria, and Italy during the summers of 2017, 2018, and 2019. One of the biggest challenges in the operation of an IoT platform, which has a great impact on the quality of the reported pollution values, is the calibration of the sensors in an uncontrolled environment. This calibration is performed using arrays of sensors that measure cross sensitivities and therefore compensate for both interfering contaminants and environmental conditions. This article investigates how the fusion of data taken by sensor arrays can improve the calibration process. In particular, calibration with sensor arrays, multisensor data fusion calibration with weighted averages, and multisensor data fusion calibration with machine learning models are compared. Calibration is evaluated by combining data from various sensors with linear and nonlinear regression models.

This paper shows the result of the calibration process of an Internet of Things platform for the measurement of tropospheric ozone (O3). This platform, formed by 60 nodes, deployed in Italy, Spain, and Austria, consisted of 140 metal-oxide O3sensors, 25 electro-chemical O3sensors, 25 electro-chemical NO2sensors, and 60 temperature and relative humidity sensors. As ozone is a seasonal pollutant, which appears in summer in Europe, the biggest challenge is to calibrate the sensors in a short period of time. In this paper, we compare four calibration methods in the presence of a large dataset for model training and we also study the impact of a limited training dataset on the long-range predictions. We show that the difficulty in calibrating these sensor technologies in a real deployment is mainly due to the bias produced by the different environmental conditions found in the prediction with respect to those found in the data training phase.