Sensor networks are the backbone of the Internet of Things (IoT) and Industry 4.0. They provide the data that drives automated systems: both feedback for control loops and status information for all of the components (nodes) of the connected systems. Sensor data has to be reliable and available as closely as possible to real time, that is, with minimum latency. How this data is communicated is therefore critical. There are different topologies for establishing sensor communication networks. Understanding them is key to building a successful system.
The term “network topology” refers to the arrangement of communication devices in a network. These devices can belong to one of several classes, including servers, clients, peripherals, switches, hubs, repeaters, and routers. The structure of the interconnections can be viewed as how the devices are either physically connected (physical topology), or how they are logically connected (logical topology).
Basic Topology Structures
Of the basic topologies, the point-to-point interconnection strategy is the easiest to understand. As seen in Figure 1, two systems are connected via a single network medium. The advantage of this strategy is that it's easy to make the connection and permanently associate the two endpoints. With no additional hardware needed, only the two endpoints present potential points of failure and the speed of the connection is limited only by the bandwidth of the interconnect medium and the physical connections to the endpoints.
We can extend the basic point-to-point strategy with packet or circuit-switching technologies, which allow the connection to be established and torn down on demand. This assumes we add some sort of signaling protocol to tell the intermediary switches which endpoints need to be connected. The introduction of switches, however, adds complexity as well as additional potential points of failure.
A step up, is the bus topology, in which each of the network nodes is connected to a common broadcast medium referred to as the “bus”. Each node therefore has the potential to see the network traffic for every other node on the network segment — all nodes are part of a unified broadcast domain.
The problem with this topology is that as the number of network nodes increases, the potential that two nodes would try to transmit at the exact same time also increases. This is referred to as a collision — these network nodes are said to be in the same collision domain. If two stations collide, they would stop transmitting, wait a random amount of time and then try to retransmit the packet with the hope that the random delay would help avoid additional collisions. Because of this randomness there is no guarantee of how long it would take to transmit the message.
Moving up in complexity, we have the ring topology. Essentially the ring network is like a bus network where the two ends of the cable are connected to each other (Figure 2). Each network station has two connections: one to each of the adjacent network nodes. Collision avoidance in ring networks can be accomplished by time-division multiplexing (TDMA) or with a special three-byte frame called a token. Given that we know the number of nodes on the network, we can determine the maximum amount of time needed to transmit a message.
Since a station cannot transmit without the token, there is no collision domain. Therefore, the only practical limit in the size of the network is the maximum token rotation time that we can tolerate.
Even though token-passing rings can guarantee performance, each node has to retransmit the token to the next node in the ring. Therefore, ring topologies suffer from the potential that every node becomes a potential point of failure. If a node failed while it held the token, the token would be lost and would need to be regenerated, introducing considerable delay.
In order to avoid the potential for failure of the ring, two counter-rotating, rings are often used. That way, if one ring fails, the other could still function. Each node then has a chance to transmit on either ring when it gets the token on that particular segment. With two tokens circulating in the ring, the latency of the total network is greatly reduced.
If we use the point-to-point network topology with a centralized hub switch or router, to which each network node is connected, we end up with the star network topology (Figure 3). In the star topology, each network station is indirectly connected to every other node connected to the same hub. This approach, however, requires more cabling and the central hub becomes a single point of failure. In spite of all of the cabling required, with advancements in technology and the transition from coaxial cable to unshielded twisted pair (UTP), the star topology is now relatively easy to set up and maintain.
If we eliminate the central hub and connect each of the nodes to each of the other nodes using point-to-point links, we have a mesh topology. Mesh networks can be either fully connected — each node is connected to every other node (Figure 4), or partially connected — only some nodes have connections to more than one node (Figure 5). The multiple connections help increase the network's reliability by providing more than one path to a particular node.
Unfortunately, since the number of connections grows quadratically with the number of nodes, having large, fully-connected mesh networks is impractical. The partially connected mesh is a good compromise: we provide redundant links where needed, and only single links if redundancy is not required, as for non-critical network nodes.
Finally, if we take a collection of star networks and connect the hubs together using something like a bus topology, we have the tree (hierarchical) topology (Figure 6). The individual network nodes in the tree topology are referred to as leaves. If we lose a link to a particular leaf node, then only that node is isolated from the network. If we lose one of the hubs, then an entire section of the network could be isolated. With a star-wired bus topology, each cluster could represent a broadcast domain for every other cluster. Intelligence built into the hub/switch could then provide filtering and address learning to limit the span of broadcasts to the local cluster. This reduces the chit-chat between clusters to better support size scaling.
Logical vs. Physical Topology
At the beginning of this article, we defined network topology as having both a physical and a logical representation. For example, physically, the topology may look like a star, but logically look like a bus. An example of this duality is found in today's Ethernet. Each of the nodes is connected to a centralized switch as you would expect from a star topology. However, the switch implements a switch-wide broadcast domain, wherein each node can broadcast to every other node attached to that switch. Since each port on the switch represents its own, unique collision domain, we shouldn't see collisions unless the switch is malfunctioning.
Another example is a star-wired ring network. The so-called “ring-in-a-box” implements standard token-passing from station to station, but the physical connectivity is still wired as a star. Although the centralized hub remains a single point of failure, we can use UTP cabling and standard star wiring to avoid the multiple points of failure found in the traditional token ring network.
Logical connectivity in the network potentially has very little bearing on the physical topology. Network designers will often use the well-understood physical star wiring approach with diverse logical connectivity based on the mission requirements. Star topologies are easy to cable and current circuit technologies are mature enough to minimize the issues of the central point of failure of the hub.
Wireless connectivity can be approached using the same network topologies. For example, the common Wi-Fi access point is physically and logically a star network topology. If we connect the access points to a central switch, we have a star of stars topology. If we then connect a router, we transition from the star of stars to a packet-switched topology.
Examples include Bluetooth, which is usually point-to-point (although Bluetooth Mesh was recently announced with the Bluetooth 5 specification). The Thread Group (e.g., the Nest thermostat) uses a partial mesh topology, as do some of the modes of the Zigbee specification. The Advanced Metering Infrastructure (AMI) for home power meters also uses a partial-mesh topology.
Partial-mesh network topologies provide some real advantages for wireless sensor networks. Through the use of routing protocols, we can extend the range of wireless links as long as there is an intermediary node that is within range for each subsequent hop in the network.
This discussion is intended to bring some clarity to the different ways networks of sensors can be set up and interconnected. The patterns of these network interconnections are more formally called topologies. Each of these topologies has strengths and weaknesses, so it is important to understand the available choices and how they impact system requirements. The design engineer has to also consider that logical interconnections don't have to conform to the underlying physical topologies on which they are implemented. Creating a sensor network requires careful consideration of all of the possibilities.
This article was written by Mike Anderson, CTO and Chief Scientist for The PTR Group, Inc. (Ashburn, VA). For more information, Click Here .