June 12, 2020
Description: Majority of AI approaches are based on the construct of training against historical data and then inferencing new data. While this is a sound and proven approach, a lot of IoT assets coming online don’t have historical data and we don’t necessarily have the time to wait.
Modern Machine Learning methods can be employed to understand the behavior of newly connected IoT assets as soon as they are connected. This allows organizations to begin “condition-based monitoring” for these assets while they collect enough historical data to begin creating predictive models. Condition based monitoring can be used to support use cases such as early detection of performance degradation, emerging safety issues etc. which are especially relevant in Industrial IoT.
Although we may be talking about billions of connected devices and hundreds of IoT platforms, the reality remains that only a small percentage of Industrial assets are connected. According to McKinsey, only 15% of industrial assets in production environment are connected. Now this number is increasing for sure but the data being collected from these newly connected assets is limited or non-existent.
The other reality to consider is the fact that most common AI models (especially deep learning) are strongly dependent on good training data. One can argue that the most important events that need to be modeled are failure events and for sectors like manufacturing where six sigma practices have been widely implemented, most organizations strive to have fewer than 3-4 defects per million parts. The rarity of these defects makes it challenging to train models.
These two realities could present a grim picture of the extent of operationalization of AI in Industrial environments. However, there is a middle ground. In this article, we will talk about Machine Learning approaches that do not require historical data and application scenarios. The focus use case for this article is Condition Based Monitoring (CBM) in an Industrial IoT setting.
CBM can provide timely intervention which can significantly increase the lifetime and operational efficiency of high valued assets. It can be the first step to get meaningful analytics on vast amount of data generated across supply chains and support AI based models to derive value from the data and add analytical maturity.
Machine Learning Approaches
There are several Machine Learning approaches that can be used in-stream against IoT data that does not need historical data. These methods can range from simple techniques such as Lag monitoring to more complex ones like Subspace Tracking.
Following are a few of these methods that are supported in SAS Event Stream Processing along with applicable scenarios.
Let’s see an example use case to detect anomalies in floodlights using Subspace Tracking method.
The data consists of energy consumption (kW) values captured every five minutes from six floodlight circuits in a smart campus over a span of three months. SST can detect malfunctioning circuits in the floodlights by tracking angle changes between principal components or by using principal component distances away from the mean.
Figure 1 above displays the angle change of the first principal component between consecutive
Figure 2 above can help determine which floodlight circuits are malfunctioning by comparing first
principal component over sliding windows. SST can give you a relatively clear indication of which part of
the system is getting out of control.
The real-time alerts can trigger maintenance activities to replace the faulty circuit to return it to normal operations. More details on the example are available on SAS Software’s GitHub under “Anomaly Detection in Floodlights for Smart Campus”.
Subspace tracking and other methods discussed above can be used in new IoT enabled asset without the need of historical data to gain insights and detect anomalies for Condition-Based Monitoring.
In a recent MAPI survey, 58% of research respondents reported that the most significant barrier to deployment of AI solutions pertained to a lack of data resources. However, organizations don’t necessarily have to wait to start on their Industrial IoT roadmap. There is immediate value to be had with the approaches described above while these challenges are overcome.
Original document: here