This document discusses using predictive analytics and artificial intelligence to solve industrial system problems. It describes Project Sherlock, a new predictive analytics solution from Rockwell Automation that uses machine learning to automatically build models for predicting issues using industrial device data without requiring data science expertise. Project Sherlock can perform descriptive, diagnostic, predictive and prescriptive analytics. A preview of the solution will be available in summer 2018 with general availability planned for 2019.
To help define analytics, we often use this chart.
The idea is that ultimately analytics are trying to help customers make a decision and take action from their data.
You can see from the chart, the red depicts what the system can do, and then the light gray is what the human or user must do to ultimately get to action.
DESCRIPTIVE
We start at the top, much of where we have invested and have solutions today is around descriptive analytics, historian, vantagepoint, performance… it’s all descriptive analytics when you are doing charts, reports, trends… etc. they are good examples of descriptive analytics, they tend to be “rear view mirror” systems that tell you what has happened.
DIAGNOSTIC
These types of analytics focus on “why” things happened, and then does more work for the user of the system… trying to get to the root cause of why issues or KPIs have arrived. FactoryTalk Analytics for Devices is a version of diagnostic analytics.
PREDICTIVE
Sometimes customers call these bottom two rows of analytics as “advanced analytics” but as you can tell from the chart, it is where the system starts to do even more work for the customer. In the case of predictive, this is where the system starts to tell the customer what will happen. Predictive maintenance is an area where RA has capabilities, and is an obvious example of predictive analytics; and in this case, the system uses advanced machine learning techniques to look at historical data (downtime tracking & process historical data) to then predict failures on assets by finding the “signature” of what happens before a failure (using the history) and then looking for that “signature” to happen again.
PRESCRIPTIVE
This is one of the most advanced analytical techniques, but where the system can go so far as to help determine what the customer should actually do (and even do it for them, if they so choose). This too is an area where RA has investments and R&D and several customer pilots working today.
FactoryTalk Analytics for Devices is a self-contained, ready-to-use appliance that works automatically on an industrial networks.
Through a variety of discovery strategies, the appliance finds Rockwell Automation intelligent assets (drives, controllers, switches, etc…) and self-configures data collection, analysis and display.
The system needs power and an Ethernet connection, and set-up is simple: just a few questions will get you started. What language do you speak, what time is it, and how will the appliance get an IP address. Once its on the network, the appliance does the rest. As it discovers devices on the network, it “digitizes” or “adopts” each device, collecting and historizing relevant data, analyzing the state of the device, and rendering the results in easy-to-understand ways.
At RA we will have a general purpose VersaView 5400 PC with industrial specs and no moving parts in April of 2017. With FTAD, we will take over this PC completely, treating it like an appliance… the customer NEVER interacts or installs items to/from Windows, the goal is we control the entire environment on this edge device, and the customer only ever uses a web browser or phone to use.
Shelby is available now on the Rockwell Automation eCommerce Portal. The site includes the current pricing for each region, but in general, expect the complete bundle to come in at under $5000 US. This includes the hardware, which is yours to keep, a one-year subscription to the software, and 8x5 tech support.
The subscription gets you continuous access to our latest and greatest device analytics, updated at least 3 times a year. Updates are simple and seamless – no firmware flashing required!
(Learn more about the Commerce Portal and how it works for Distributors on Noggin: https://noggin.gosavo.com/Post/Index/42695030?view=&srlid=55920385&srisprm=False&sritidx=0&srpgidx=0&srpgsz=25)
Project Sherlock is an upcoming system-level analytic appliance from the Shelby team, that begins moving us out of reactive, diagnostic-type analytics and into predictive, and eventually prescriptive analytics.
(recall or revisit the Intro to Analytics slides if necessary for positioning)
Like Shelby, we want smart technology to help our customers find the data that is already present, learn from it, and return value quickly and easily. But systems are more complex than individual smart devices (often systems are combinations of these devices) so we’re introducing our first Artificial Intelligence to help tackle this complexity…
We’ve had expert driven analytics for quite some time now – generally they’re getting more powerful, but not easier to use. With Sherlock we believe we can leverage new innovations in AI to begin to make these things easier. We’ll grant that at the beginning, it may not be as sophisticated as could be accomplished with an expert system – but it will be faster, easier and more affordable.
Sherlock’s experience will be targeted to a controls engineer – someone who knows their application, but doesn’t need to be a data scientist. We’ll start by identifying the output tag that matters for the thing you’re trying to product. This could be a measurement of quality, scrap, throughput, flow… its up to the user to choose the value that matters for the prediction they want to build.
Next, the user selects the other tags in the controller that might be related. A wide net can be cast – pick anything that you think might be involved. The Sherlock AI is smart enough to do its own “data cleansing” eliminating data points that aren’t correlated to the output you identified during operation.
Side note: If the data you want to include doesn’t exist as a controller tag, you’ll want to integrate that in somehow. Sherlock is also smart enough to identify if there is insufficient data to build a reliable prediction.
Once the inputs and outputs are identified, Sherlock needs to watch the operation running – this training period is “unsupervised”, meaning a human doesn’t need to be involved. Just start the operation and let Sherlock observe the data streaming through the controller as it runs. During this process the AI will find correlations, thin the data, and start building its model.
The actual predictive model building process is fully automatic. Sherlock will rapidly formulate theories of potential mathematical models of the observed operation, testing each theory until its confident it has selected the best possible physics-based model. The training time varies, depending on the complexity of the operation, but in a simple process unit, we’re talking about 20-25 minutes of unattended training. Sherlock will raise a bit in the controller to indicate that its ready to begin predicting.
Sherlock knows if it has enough data to accurately predict. If it doesn’t, it will tell you. Otherwise, if you’re ready to test it, you can monitor a bit in the control program to determine when Sherlock predicts a problem. How you respond is up to you. You can wire that bit up to a FactoryTalk Alarm and Event, add it to an HMI display, or monitor that value in an external maintenance system. If you have a Shelby, our goal is that FactoryTalk Analytics for Devices will automatically create an Action Card whenever a connected Sherlock spots a problem.
This chart shows the results of early Sherlock research and validation, when used in the real world. This was a power generation application, from one of our large customers. In the chart, a difference between a prediction and reality indicates either that the model isn’t working, or that there is a real problem with the process. Getting it wrong could be the difference between an accurate prediction and a false alarm. In this case study, Sherlock was monitoring normal operation – nothing went wrong.
In the best competing algorithm, shown with the green line, you can see the model starts accurately, but as the process goes through normal variation, it becomes less accurate, generating false alarms.
In Sherlock’s physics-based modeling, the algorithm is aware of normal physical process changes, and adapts, remaining accurate as it predicts the process. When Sherlock predicts a problem, you can be pretty confident that someone should look into it. For now, this is where we’ll stop: Sherlock learns a base-line, and lets you know when things appear to be heading in the wrong direction.
Over time, the product will improve to take advantage of more of the AI, giving you first recommendations about how to resolve a problem, and eventually making its conclusions available to you for closed-loop optimization, if you chose. That’s the future; at launch we’ll focus on problem detection and indication.