The Internet of Things (IoT) will fundamentally change everyday life - and is already doing so today. This applies to the private lives of consumers as well as to companies and entire industries - be it in the smart city, where self-driving cars communicate with connected traffic lights and other parts of the infrastructure via sensors. Be it when data from a computer tomograph indicates when it is likely to fail. Or when movement data on soccer players predicts when a goal will be scored. The basis for all of this is analytics software that helps to interpret this IoT data. This is a considerable challenge. A large number of factors have to work together for the successful implementation of IoT projects. These include sensor technology and connectivity, of course, but also data management, analytics and algorithms and, last but not least, business aspects such as change management. And very importantly, no IoT project should be implemented without knowing the potential business case. After all, technology for technology's sake brings no business value.

IoT data is unknown or poorly analyzed at the time it is collected. It is transmitted at an extremely high frequency. Values are often missing, but there are also lots of irrelevant values (known as "noise"). There are also thousands of variables. Due to this data complexity, a decision has to be made: What data do I need? What information should be sent from the networked device to an analytical intelligence and the result of this analysis back to the corresponding device? How much data should be stored? These questions are crucial, as they determine the time required and the costs.

Identifying strong predictors - i.e. variables that allow the most reliable predictions to be made about certain issues - often turns out to be the proverbial search for a needle in a haystack. Sometimes there are only a few cases or outliers, such as when predicting machine failures or fraud attempts, which makes it difficult to recognize rules and patterns. The triad of classic analytics - accessing, storing and analyzing data (Access - Store - Analyze) - no longer works in the IoT. Instead, data often needs to be assessed very quickly (possibly even in real time) on the spot.

The analytics lifecycle consists of three stages: Data, modeling, application. In stage one, order is brought to the data chaos. The "hard work" consists of transforming the data, pre-processing it, filtering it intelligently, reducing the recognized dimensions and finally extracting features. In the second phase, a model is developed on the basis of this data in order to detect anomalies, such as a machine's drop in performance.

Analyze data where it is generated

Once the model is established, it is applied directly to streaming data. This means that it is analyzed using event stream processing directly during transmission, i.e. in the data stream and before it is saved - i.e. at the "edge" of the data path. This is where the term "edge analytics" comes from. The technology functions as an intelligent filter, which reduces the data transport. This enables the analytical system to avert potential problems in advance by shutting down a machine, issuing an alert or initiating another measure.

A variety of new analytical methods are used to detect signs of anomalies at an early stage. A so-called support vector data description, for example, can help to identify performance degradation in aircraft turbines. To do this, a "normal" average is calculated from the existing data and all newly generated data is evaluated for its deviation from this normality model. It is then possible that the deviation from the normal value increases as the machine approaches its end or a failure. Once a predefined threshold value is reached, an alarm is triggered and the turbine can be replaced in good time before a failure occurs. The analytical model is applied directly to the streaming data, eliminating the storage step. And only the data that could be valuable for later analysis is stored in the data warehouse.

Edge analytics in production

Many use cases for edge analytics can be found in production. For example, photos are taken of products to determine their quality. In the metalworking industry, for example, it is possible to determine at a high frequency whether the material has been heated correctly in the oven, draw conclusions about the quality based on this and decide whether it is a reject or whether the product should be fed back into the production process. A concrete use case could look like this: A chocolate manufacturer wants to check the white discoloration on the surface of its product, which occurs, for example, under the influence of cold. To do this, a camera takes a picture of each chocolate bar. Analytics can then be used to determine the degree of this "contamination" and goods above a certain threshold can be sorted out as broken for special sales.

New data requires new approaches. In order to extract relevant information from sensor and other IoT data, which in turn can be used to make intelligent business decisions, innovative analytical technologies are required as well as a new approach to the application of these technologies. Sometimes the third step has to come before the second, i.e. analysis before storage. In order to develop fresh solutions for new problems or even innovative business models from the information obtained, the technological prerequisites must be complemented by a change in thinking - both at management level and among employees on the production floor.

This year's SAS Analytics Experience in Milan also showed how IoT data can be put to good use with analytics. Practical examples ranged from predictive maintenance in production to predictions for efficient supply chain processes.

Dr. Nicole Tschauder, Analytics Expert at SAS DACH / ag