Machine Learning at Edge Enables IIoT to Scale
Published by : Industrial Automation
The concept of Edge computing will see more adoption in the future, says Nir Rostoker.
The Edge and the Cloud IoT are two pillars of the Internet of Things. IoT devices generate huge amount of data. This data must travel from the device to the cloud for processing, and the output is then, pushed back to the device. Thus, there is a need to push large amount of data to the cloud every time, which in turn is very expensive and in many cases not possible; often, this also leads to a delay in decision-making due to round-trip delay or intermittent connectivity because of intermittent connectivity.
Industrial IoT is a key pillar to successfully execute digital transformation for most companies. Business processes enabled with intelligence from sensors and devices is key. We need a way to store, compute and action upon the IoT data at or close to the data source to improve response time and save bandwidth. This is when IoT edge comes into picture.
Edge computing is a form of distributed computing, meaning not all data will be processed in the cloud; some data will be processed at the edge. But remember, at the same time, not all data is processed at the edge as well. Because if you process everything at edge, then it becomes an on-premise deployment and so visibility and analysis of data across the edges will not be possible. Hence we are talking about distributed processing across Cloud and Edge in a hybrid form.
It is not just Edge computing or Cloud computing, Cloud-to-Edge hybrid is the way forward for many Industrial IoT deployments:
1. It allows customers to run IoT enabled business processes at the Edge and in the Cloud, allowing for consistent execution of business processes across the cloud and multiple Edge nodes.
2. Customers can train predictive or Machine Learning models in the Cloud and deploy them on the Edge.
3. Customers can design interoperable rules, and actions across the Cloud and the Edge.
Industry 4.0 requires computation intensive use cases at the edge. Edge analytics, especially machine learning, is widely used in the manufacturing plants for predictive maintenance and predictive quality use cases so bringing machine learning to edge enables Industrial IoT to scale. So, in this article, let’s focus on leveraging machine learning execution at edge.
First, let’s look at the challenges of machine learning model execution in cloud.
Cost: It is costly to bring high volume of data to the cloud for real-time inference.
Latency: Bringing data from the edge to the cloud will lead to higher network latency.
Throughput: It is costly to support large number of inference calls per second for deep learning models.
Security: Security concerns of user data risks sending data to the cloud.
Machine learning training in Cloud and execution at Edge
Sensor data coming from the device/sensor is relayed to the IoT cloud. The machine learning model is built and trained over the cloud and this trained model is then deployed to the individual edge nodes. ML inferencing happens at the edge node, an output of which can be used to trigger a business action.
Let us understand this phenomenon with a customer example – a large confectionary company produces cereal bars; too many broken bars or too small products leaving production leads to high wastage of resources. This issue can be addressed by implementing the above-mentioned phenomenon.
A camera is installed in the production line (‘the edge’) to take and post the processing images of the cereal bars moving on a conveyor. The predictive models are trained in the cloud, using the physical information obtained from the defective and valid samples of the cereal bars. These trained predictive models are then deployed on the edge location. Now, in the real time, the predictive models take the data extracted from the images obtained from the camera, as input and interpret this data. The output of the model is a quality score that represents the quality level of the current production. This allows the company to take a real-time decision regarding approving/rejecting the cereal bar. So, in this case, predictive analytics at the edge improve the efficiency and quality of the production process in real time, while raw or aggregated sensor data sent to the cloud enables comparison of production quality across plants.
The concept of Edge computing will see more adoption in the future, as the phenomena of decentralised computing with data being processed locally becomes more prevalent, thus extending intelligent data processing to the edge orchestrated from the cloud enables Industrial IoT deployments to scale.
Advancement in hardware technologies and computing resources will provide the ability to run ML algorithms and artificial intelligence in small chips, leading to a proliferation of more edge use cases in the future.
The views expressed in this article are those of the author and may not reflect those of SAP.
Nir Rostoker is a Global VP, Head of Product Management, Internet of Things at SAP Labs, LLC (Palo Alto). He is responsible for a global team of product owners and product managers with the focus on product inbound and outbound activities.