Robbrecht van Amerongen about: The 'dumber' the intelligent edge the better

What is an intelligent edge?

An intelligent edge is the application of logic to data where it is generated. So in a factory, train, tunnel or stable, or on a wind or solar park, flood barrier or airport. The word 'intelligent' can be discussed at length, says Robbrecht van Amerongen, Head of Strategy at AMIS Conclusion. "We consider a solution to be 'intelligent' if it is able to return data and commands to the physical device, which then makes independent decisions based on this data. These may well be 'stupid' decisions, or rather, simple logic is applied. For example, at 100 pieces, the crate is full and a new crate must be placed under the machine."

Why an intelligent edge?

There are several reasons to choose edge computing. First, for real-time applications, latency is often an issue. The closer to the source the data is processed, the lower the latency. When it comes to large data, such as video images, network costs are an important argument. If all data has to be sent to the cloud continuously, you often need a fast connection and costs quickly add up. And for some solutions, you simply can’t afford to add extra dependencies; consider safety systems in factories or fire detection. You want these systems to function even when there is no network connection, cloud availability or central database. It is also not always possible to process data in a central cloud due to laws and regulations. And even if it is legally allowed, you still have to ask yourself whether you want to send your IP or privacy-sensitive data to the cloud.

What makes it so complex?

The complexity depends entirely on the application, says Robbrecht. "The longer the IT chain, the more complicated it is to have a complete overview of everything and the more places there are where something can go wrong." Sometimes all it takes is one relay not working properly to cause wrong decisions. Robbrecht concludes:  "The complexity is usually not in developing the algorithm, but in ensuring that everything continues to go well in all parts of the chain. In addition, you must be able to prove what happened. More and more processes require an audit trail, software bill of materials and version management from a compliance perspective. You must be able to mathematically demonstrate at any time which logic led to a particular (wrong) decision. The more logic - and certainly artificial intelligence - there is, the more difficult it becomes to subsequently deduce how a decision was made. You also need to consider that different versions of this logic may run side by side, for example, if not all updates are installed at the same time. You must also be able to guarantee that all components continue to communicate well with each other after such an update."

Where are the risks?

As with any application, with intelligent edge solutions, the output depends on the input. The problem with a lot of locally generated input is that the quality can be very variable. Examples include camera images that become blurred due to water vapour or sensors that produce many outliers in their measurements.

You also have to be careful with self-learning algorithms that form a black box, says Robbrecht. "As I emphasized earlier, you must still be able to explain how the algorithm arrives at a certain outcome. If you make everything smart, devices can start to work against each other. 'Dumb' devices have predictable behaviour via a clear if-then-else decision tree. If many more devices start working smartly and are therefore less predictable, this could also have the opposite effect. Consider smart energy devices in combination with an energy contract with flexible prices. There is a chance they will all switch on at the same time, exceeding your grid connection limit. If all companies in the area react the same, this could even lead to serious disruptions in the network."

Another point of attention are the unhappy flows. "It's generally not that difficult to fully automate a happy flow. But does the algorithm also make the right decisions in the event of a very rare unhappy flow? And if a wrong decision is made, what are the consequences? Especially when it comes to automatically intervening in physical processes, things can go very wrong with unhappy flows. If you don't think through these processes very well, you simply run great risks. After all, risk is chance times impact", Robbrecht warns.

Securing the edge computer can also be an issue. After all, it is often located in a cabinet somewhere in a factory hall, or even in a cabinet somewhere in the open field. "Physically, an edge computer like this is not easy to secure; you can just walk up to it, attach a device to it and take over a large part of the operational systems. The starting point should always be that every edge device must be treated as a potentially unfriendly guest in the network. A guest who must continuously prove that his behaviour is safe. Sometimes things go wrong unintentionally, like the edge computer on a ship that had a USB port. The employee on the ship used that port to charge his phone, resulting in a virus contamination in the primary process", says Robbrecht.

A final point of interest he brings to our attention is what happens to the data and algorithms when the device is sold. "If you sell your Tesla, what happens to the data that's still on the edge? You need to legally record these things. For example, think of a business unit that is being sold, but you don't want your IP to suddenly become part of that deal", Robbrecht warns. 

Practice

Conclusion has developed hundreds of intelligent edge applications for customers. From very 'dumb' applications such as counting certain things or actions, after which an action follows when a certain number is reached (scheduling maintenance after a certain number of operating hours) to very intelligent systems where cows receive feed individually tailored to them based on their weight, their past milk yield, temperature and other variables. Robbrecht: "We prefer to develop our solutions in such a way that the IT chain is clear, but this isn't always possible. In such cases, the systems provide support and we rely on the intelligence of the human operator to make the right choice."

The riskier the application, the more governance is involved. Robbrecht: "If an image recognition algorithm has to select the right potatoes for French fries and a potato slips through that is actually too small, then that's not much of an issue. But if you have to be able to trust the device's decisions one hundred percent, it's a different story." Almost anyone with a reasonably new car is familiar with steering corrections or braking actions that the car takes independently for your safety. Meanwhile, in practice, they actually endanger your safety. "Therefore, especially in situations where the risks of making the wrong decision are high, the 'dumber' the intelligent edge is, the better", Robbrecht concludes.