Edge vs Cloud AI

• Big Data. Data that is high in volume, frequency and variety and growing exponentially all the time. The data can be structured or unstructured and is difficult toWhere does data get stored and analysed, Edge vs Cloud AI process using traditional methods such as spreadsheets and SQL databases. AI requires big data to converge on the problem more accurately. Typically training requires substantial amounts of data.
• IoT (internet of things). Network of computing devices that are embedded with sensors, software and other technology that collect and transfer data over the internet.
• AI. Mimics the problem solving and decision-making capabilities of the human mind to achieve intuitive outcomes. Within artificial intelligence is a broad range of scientific techniques such as machine learning, natural language processing, computer vision and reinforcement learning.
• Cloud Computing. On-demand computing services which are accessed through the internet, including data storage and computing power. A great advantage is that businesses don’t need to own and manage physical servers and IT services.
• Edge Computing. Computation and data storage close to the location of the device (computer, IoT device or edge server), using instant data which is generated at the location.
• Edge AI. Combination of edge computing and AI, which allows for the running of AI algorithms on a local device using real-time data that is locally acquired. Data is not transmitted to be analysed in the cloud; the computation occurs in the device. Real time AI analytics on the edge is typically performed using pre-trained models and inference.
• Cloud AI. Using artificial intelligence on a cloud computing infrastructure

The main difference between cloud and edge is the volume of data that can be processed, the latency and therefore the depth and scope of the analytics able to be conducted.

Edge AI applications are designed to learn, monitor, and predict the behaviours of specific equipment or a limited data field. The rapid results are due to the localisation of data being analysed, and the reduced volume of that data. Due to the speed at which the results are required (for example an imminent accident or threat), there is little time for the analytics to take place. The limited edge computing power also typically restricts the complexity and finality of the AI models in use, possibly reducing their capabilities to mostly statistical analysis. Predictive analysis requires constant retraining and tuning of model parameters which cannot be easily achieved at the edge.

Cloud AI or AI that takes place on a hosted server differs in that it can easily analyse vast amounts of data, even billions of data points. AI models can digest a wide range of data sources, taking a holistic view of a process, plant or even equipment and determine outlying contributing factors, which may not be factored in when using a localised Edge AI approach. Some of these factors may be external such as the weather or even un-connected processes at an industrial facility.

Cloud AI can also perform the same tasks as edge AI where high bandwidth connections are available. Common examples of this can be seen in services such as Google for image and speech recognition. But as more research is done, and hardware level AI acceleration is added to edge devices, some of the heavy lifting is moving to the edge. However, for the time being, training of anything but the simplest models is still performed in the cloud.

One VROC client was able to quickly deploy an AI model after discovering a high moisture reading, which could have led to a gas leak at their oil and gas platform. The AI model indicated that the root cause was a Gas Well, which is a very separate process and not inter-related at all to their issue. This client followed the recommendation and shut in this particular well, which resulted in the high moisture reading rapidly dropping and a significant safety incident being avoided. This client was able to determine the root cause because live data from multiple separate processes was being analysed simultaneously.

In this example the models ran on live and historical data and was able to give insights for a critical situation in a matter of minutes, helping this client avoid an incident which was estimated to cost close to one million USD. The benefit of a cloud-based AI platform in this situation, was that more data could be analysed, using more computing power and despite the volume of data, this analysis was still performed rapidly.

The use of historical data is also a key component when conducting artificial intelligence. Historical data can help improve model accuracy, as the AI can learn from similar historical conditions and failure modes. When combined with live data the AI can model and predict future conditions. Historical data is typically not stored at the edge, as there is limited storage available and therefore limits the complexity of the analysis that can be performed at the edge.

One other major difference worth noting is that between predictive and condition-based monitoring. Edge AI can detect conditions and thresholds which are either nearly met, met, or exceeded, making it more of a condition-based monitoring tool. Whilst it does provide for the most part an early notification, it may only be minutes or hours in advance. Whereas AI that is cloud based detects minute deviations and can start predicting issues often weeks in advance.

The more lead time, the more options operators have to plan interventions. Another VROC client was alerted 7 days in advance of a failure with its Water Pump. This lead time ensured that no downtime was experienced, and maintenance was a planned activity. This prediction resulted in an estimated savings of between $150,000 to $250,000 USD.

It is possible for an industrial business to benefit by using artificial intelligence both on the edge and in the cloud. In situations where connectivity is an issue and critical equipment needs to run continuously the Edge is a great way to access AI insights. Following on from this to further assist businesses with advanced time to failure, holistic root cause analysis across processes, and optimization insights enterprise wide, artificial intelligence can either be performed in the cloud or in a client’s own infrastructure.

VROC’s no-code rapidly deployed solution, OPUS, is unique in that it can be hosted using our own private data centre, on a client’s on-premises data centre, or on a client’s public cloud or by using the VROC Cube.

Interested in a demo, get in touch with our team.