IoT analytics are a critical pillar of digital transformation. The speed and granularity of connected machine data give businesses unprecedented opportunity to transform processes, improve operational decision making and derive deeper insights, not to mention forming the basis of intelligent automation. With the proliferation of analytical techniques and algorithms, it can be helpful to organize different types of analytics into a few basic categories based on the insights required for their industrial IoT projects, including:
The preceding types of analytics are often thought of as an analytics maturity model, where greater value is unlocked by progressing to the next (and increasingly difficult) stage of the model. While there is certainly a kernel of truth to the broad presuppositions of this industry standard maturity model view, it can lead to a flawed approach to digital transformation-- it’s time to say goodbye to the analytics maturity model and instead adopt a new paradigm.
Naturally many people assume that most critical decisions required the most “advanced” analysis. In our experience, this instinct isn’t necessarily true always and should be examined. Each type of analytics serves different purposes and each can be tremendously valuable. Analytics is but a means to an end. “Predictive” for example is a capability, not a use case. Overexuberance to progress to that stage can be misguided. The business goals should determine the required insight, which given the constraints of the available connected and historical data should determine the type of analytics. Crawl-walk-run is still sage advice, but it can manifest in different adoption patterns that have little in common with the traditional analytics maturity curve.
What’s important when looking at the different types of analytics is what we need from the data, not the sophistication of the analytics technique.
The first thing to consider when determining what type of analytics to develop is the specific decision point(s) we are trying to improve- is it more appropriate for a human or a system to make that decision? The critical factors are time, context, and accountability/auditability. Humans need time to think and are limited in terms of the number of dimensions they can consider at any one time but they adept at interpreting information in context whereas algorithms are the opposite- they are lightning fast and can process a tremendous volume of data points with very subtle and/or complex patterns, but have no ability to grasp context outside of the data that’s fed to them.
Humans also understand responsibility. Business-critical decisions often require human insight and accountability. Where algorithms can help provide it may still be necessary to present all the contextual information to a human decision maker to best inform a judgment call.
High volume, repeatable decisions that trigger different workflows based on complex or subtle patterns are ideal for predictive and prescriptive algorithms.
With this paradigm in mind, let’s take a more in-depth look at each type of IoT analytics.
Descriptive analytics describes what happened historically or what’s happening now. Most business intelligence analytics fall into this category. What is important to measure and monitor in operations? Counts, KPI’s, benchmarks, trends and queries are all descriptive analytics that deliver historical data to the user/decision maker. Definitions for these digital objects are created to support basic operational and management decisions.
Descriptive analytics also includes statistical process control. Adding the dimension of time to a value opens up a wide array of additional insight that can be gleaned from a single measure. When have data rolling in every second and plot them, you’re able to derive rolling averages, deviations over time, and more, thereby allowing you establish a baseline of “normal” and monitor for rapid or unexpected changes that you want to be alerted to. Statistics can also summarize for relevant blocks of time such as a shift or production run.
Descriptive analysis is a foundational element that can deliver insight for decision makers in the near term while building historical data assets.
Diagnostic analytics is a deeper-dive into a specific area of interest. As opposed to a range of measures to track basic performance indicators and/or process variance, Diagnostic analytics seeks to uncover hidden patterns or root causes of events. It is used to examine not what happened but why it happened by determining what factors and events contributed to the outcome of interest. Diagnostic analytics uses the correlations between variables to provide deeper insight that descriptive analytics cannot, say, for example, to understand why the average downtime for a particular machine on the shop floor is higher than average this month. In some ways, diagnostic analytics is automating what analysts are often times trying to do manually when they engage in data discovery- querying, drill-down and “slice and dice” a data set looking for patterns.
This approach requires the analysis of a large swath of historical data. It can be time consuming and is well suited for investigations, problem solving and identification of opportunities to improve process parameters, policies or best practices.
Within manufacturing, diagnostic analytics can be used to deal with downtime, help with asset maintenance, ensure the efficiency of production cycles, and avoid inventory shortages and shipment delays. Diagnostic analytics is key to continuous improvement and lean manufacturing. While diagnostic analytics offers a lot of value on its own, it also serves as a foundation for predictive and prescriptive analytics.
Predictive analytics uses machine learning to analyze historical and current data to provide an assessment of future events. There are cases where you can start right away with predictive analytics, as many organizations already have a lot of data to work with.
Organizations have plenty of data from IoT sensors; more often than not, the bigger challenge is about how to unlock it for insights. The problem is linking real time operational data to historical data. While many organizations have sensorized machines and operations, perhaps even writing simple rules that can trigger alarms, they often struggle to associate that data with outcomes of interest that typically reside in legacy systems.
Generally speaking predictive analytics is good at further improving a process once you have already exhausted other means to improve it and have hit the wall of diminishing returns.
Predictions are probabilistic statements with an expiration date. They either come true, or they don’t. And the nature of accurate predictions is that they can be either self-fulfilling prophecies or self-defeating prophecies, meaning that the prediction itself will change the way we act and have an impact on the outcome. A perfect example is a failure prediction. If a predictive alert flashes across the screen and an operator reacts by slowing down a feed, thus successfully avoiding the failure, we have a self-defeating prophecy. For this reason, predictive models need care and tuning to keep them fresh. Other things in the environment change as well. Over time all models get stale and need to be refreshed.
It’s important to understand that advanced analytics is a continual process that never ends. Domain experts must interact with the data and the models to keep them fresh and aligned to business and operational goals.
To be fully effective, there needs to be an organizational change management perspective shift. Human-machine collaboration enables that computational ability and those algorithms to drive more and more insights, and to empower an employee to make better decisions. Ultimately, the creation and the management of that intelligence is a very human exercise. You need people to make sure those algorithms stay relevant, that they're learning from the right data, and that they are updated accordingly.
Factories and industrial organizations can benefit greatly from IoT sensor data integrated into existing workflows. Typically, downtime is an enormous cost for factories. When due to machine failure, for example, it can also be dangerous for workers.
Fortunately, IoT sensors attached to machinery can track everything from temperature to vibration and relay that information into a data analytics platform. Using predictive analytics, this data can be used to predict when particular equipment would need maintenance next, allowing plant managers to respond ahead of a device failure and circumvent unnecessary downtime. In more sophisticated setups, factory control systems can use information from data analytics platforms to shut down machinery ahead of a failure, helping prevent dangerous conditions. Predictive analytics can also be used to manage resources and forecast inventory, helping organizations improve operations and function more efficiently.
Prescriptive analytics is focused on determining the best solution or outcome among different choices based on a set of given parameters. Because prescriptive analytics can highlight implications for each decision option, it can be used to understand how to mitigate a future risk or how to evaluate a future opportunity. Once properly set up, prescriptive analytics can be used to automatically process new data continually, parsing millions of variables and constraints to improve the accuracy of predictions and provide better decision options.
Prescriptive analytics is a process-intensive task taking input from a variety of data sets from historical and transactional data to real-time data feeds and big data. It has been enabled by advancements in computing speed and the development of complex mathematical algorithms applied to data sets based in game theory, decision-analysis methods, and optimization. Prescriptive analytics are especially effective for industrial organizations for strategic planning and operational activities. By helping quantify the effect of future decisions to provide recommendations of one or more possible courses of action, they allow organizations to assess various possible outcomes based on their actions.
A factory could use prescriptive analytics, for example, to prevent the production of defective products by using relevant, contextual data to enable subsequent insights to drive action. Industrial organizations are already successfully using prescriptive analytics to optimize production, scheduling, and inventory in the supply chain. Prescriptive analytics are complex to enable; however, they have the ability to make a significant impact on an organization’s bottom line.
As we’ve seen here, each type of IoT analytics plays a valuable role in the manufacturing industry. Instead of having to follow a traditional maturity model where one type of analytics leads to the next one, you can delve into the appropriate type of analytics for your specific needs. Ultimately, determining what you need from the data you collect from IoT sensor data can help you determine what type of IoT analytics to apply. The insights you gain can help you with everything from business-critical decisions to intelligent automation.
IDC Analysts reveal the key to gaining value from your IoT data insights.