By Mehdi Roopae
The Autonomous Internet-of-Things (A-IoT) is about device-to-device connection and interaction, so as to provide economic benefits in efficient time. A network of multiple drones applied in smart cities as a kind of A-IoT helps autonomous vehicles by mapping the aerial informatics beyond what the car's sensors can discover and then device-to-device communication to share all the required information for a safe and efficient transportation system. The connections between A-IoT devices can change very fast. Moreover, when a new A-IoT device with novel technology features and communication protocols is connected, the other devices in the network need to connect with the former and start interaction by exchanging data. Understanding connections among huge active devises in an A-IoT network and their impacts is the key to identifying perspectives for new plans and services. In addition, connections are more than lines between devices and each one includes attributes concerning the connected devices. Attributes may reveal the nature of connection, the data linked to the connection, the other direct or indirect related connections, when the connection was created, and more. Any modifications in the A-IoT network will consequently cause rapid changes in the connection attributes.
Activities of millions of A-IoT devices generate vast amounts of data that can be analyzed by assessing graphs, which model the mechanisms of spread and evolution of information. A graph model of information spreading in an A-IoT ecosystem can uncover the strategies employed by devices when escalating their connection circles. A graph model consists of information from various A-IoT sources, in which nodes are accumulating rich and time-varying information collected from autonomous devices and an edge is a metric shows degree of interaction between them. These graph models have the capability of tracking the behavior of the network quickly and adaptively, while its shape is changing by the spread of information due to the communication activity among devices. Therefore, graph models are essential to discovering, capturing and defining complex interdependencies and relationships between various types of A-IoT devices. Graph models are proposed to be used, in order to easily model and navigate grids of data with tremendously great performance. Recently, they have also been implemented to extract maximum information from the Internet of Things.
Graph models enable A-IoT networks to understand millions of interconnections between devices on a network and then share the data based on any type of relationship. This requires the development of graph models to be designed adequately “rich”, so as to detect the behavior of A-IoT devices, but also “simple” enough, so as to be rigorously analyzed. Most of the available models suffer from scalability, which is the main restriction of identity resolution models, and, which also limits the ability to scale up when huge amounts of information is captured from unstructured data. Therefore, there is a tremendous need to extract the structure and dynamics of identity resolution, using deep graphs to ensure that data from multiple A-IoT sources of heterogeneous and unstructured data can be integrated.
A traditional dynamical IoT graph captures the edge dynamics of hierarchical networks is to model the states vector of an A-IoT network and the weights of the links between any pair of A-IoT devices. To determine the dynamics of the change in the state-variable, a conventional approach considers the rate of change the weights as a function of all state variables.
Deep learning attempts to model multi-level perceptions in infinite device connections in A-IoT networks by using a deep IoT-graph to learn attributes and actions from large-scale unlabeled devices. An A-IoT network is a complex system with a huge number of state variables and needs a large amount of local and global information about the interaction between devices. Deep IoT-based graph by discovering almost all possible information is a powerful approach for modeling complex IoT networks. A deep IoT-graph model can extract almost all the local information around an autonomous device, and the link between this device and others is created based on the conditional probability of all the extracted features such as content of a message, series of message, information environment and so on. Therefore, the traditional adaptation law mentioned above could be improved by incorporating the weight of connection between the pair of autonomous devices, according to the conditional probability between all deep extracted features for that devices.
The introduced deep learning IoT-graph model, (i) extracts feature for each A-IoT device from pre-trained Deep Neural Network by transfer learning to provide better feature representation of the devices considering all interactions and dependencies and, (ii) applies two layers of Long-Short Term Memories to model device connections and predict degree of relevancy between those connections in an A-IoT ecosystem, which is more robust and accurate in compare with traditional intelligent systems.
In a deep-learning approach, the network must be large enough to have the capacity to understand all possible patterns existing in an IoT ecosystem. The scalability and flexibility of deep learning distinguish it as a powerful approach for a real-time system like a smart sensor in the continuously changing environment of A-IoT grids. The proposed deep learning IoT-graph model is capable to extract multiple scale features in a multi-level representation and it could be designed based on the scalability and accuracy required for the output of the system. A human expert may assist as the supervisor, who controls the degree of feature extraction and assigns LSTM modules. The training of deep learning, transfer learning, of IoT-graph model takes time for real time environment. However, with the huge progress in various aspects of cloud-fog computing platforms, it is feasible to handle all computational tasks required for the proposed deep learning automatic autonomous-connection update system of an A-IoT ecosystem.
Mehdi Roopaei, IEEE Senior Member, is a research scientist in the Open Cloud Institute at the University of Texas at San Antonio. His research interests include deep learning control of large scale dynamical systems and deep machine learning. Roopaei has a PhD in artificial intelligence from Shiraz University, Iran. He’s a members of AIAA and ISA.
To have the Bulletin delivered monthly to your inbox, join the IEEE Smart Grid Community.
To view archived articles, and issues, which deliver rich insight into the forces shaping the future of the smart grid, please visit the IEEE Smart Grid Resource Center.