I attended a day presentation where they showed some interesting development in the field of Industrial Automation. They are really pushing for integrating factories into the internet via the internet to enable new data analytic opportunities. Here is a few quick notes I made of some of the key interesting things I noted on the day.
Factory based Machine Learning Appliance module
Usually when you think about Machine Learning, you would usually think about cloud based software crunching data and spitting out results. However what’s interesting is that Rockwell has opted for a AI module that fits right into a PLC.
Searching online, it’s called Project Sherlock; and the idea is that it is a device that sits on the PLC backplane, monitoring the datastream and detecting the various condition of the machinery.
Based on the data it then creates a predictive model of what it thinks is a working PLC system condition. It is then able to continuously monitor after training, alerting the user if there is an anomaly.
The main selling point according to Rockwell is that it is easier for non data scientist or field engineers to implement machine learning on the factory floor to detect failing machinery before the machine shuts down as well as the anomaly type. Example provided was monitoring a motor vibration to see if the motor has a loose or fracturing bearing.
Implementing the machine learning right at the PLC level, but not having it directly interact with the PLC is smart. If you remember the PetNet cloud outage incident, you realise that there is an importance from a risk-limiting perspective to not rely on the stability of remote connection. In the PetNet case, there was no local intelligence in the food schedule of the pet feeder, it was all online. Thus when there was an outage, the pet did not get fed.
Thus by implementing the AI module in the factory floor rather than purely the cloud, there is an inbuilt resilience to the system to allow the factory to function, even if disconnected from the internet. But at the same time, also allow for more advance analytics that are not connection stability sensitive (e.g. Long term mean time to failure prediction)
Basic Machine Learning Concepts To Understand
Classification Vs Regression
- Classification - Distinguishing between good or bad states (e.g. Pass or Failed quality of an apple)
- Regression - How well could a ML algo predict another future value, given a different input values. (e.g. Quality of apple, given image of an apple.)
Supervised vs Unsupervised vs Semi-Supervised
- Supervised - Used when a specific prediction is required (e.g. Percentage of apple that will passed quality inspection)
- Unsupervised - Used in detecting anomalies in a running process (Given known running motor, alert if it deviates)
- Semi-Supervised - (e.g. Detected if a motor is failing, and if possible… classify failure modes)
Machine Learning Types
Good for simple machine learning problems that is linear in nature.
- Cost Function - Quantity you want to minimise (e.g. error difference from desired motor speed)
- Gradient Descent - Algorithm that slowly moves towards the minimum cost function, via stepping down incrementally towards minimum cost
|* * | *--| * | * | * | *-| * | *---| * | * +--------------------
- Hyper Parameter Search: Grid search, is a different search algo that search both sides of the cost curve to find the answer faster.
|* * | * ---------------->| * | * |<---------------* | * ----------->| * | * |<-- * | * +--------------------
- There are plenty of information on neural networks on the web and its a vast field. So didn’t take much note on this.
It is a IIoT concept where given enough sensors is integrated into the factories and connected to the cloud, it would allow for predictive models to get advance enough that it is able to have an accurate enough conceptual state of the factory floor. Maybe you could describe it as the factory or the company being able to be virtually introspective of themselves. (e.g. How humans know where their limbs are.)
One example given is the tracking of a steel core though a steel mill. Being able to model a virtual representation of the product, would allow for triggering other factory processes based on the state of the product itself. (For example how much the steel core has cooled down since forging). (e.g. Having an IR camera monitor the heat outside of the product and thus know the spread of the heat inside the product)
This is as opposed to current factory process of simply waiting a fixed amount of time, which may be inaccurate due to variation in core material properties.
Overall process of machine learning process
Define the Problem space –> Get data –> explore the data for insights –> prep data —> create model –> gather recommended actions
Descriptive –> Diagnostic –> Predictive –> Prescriptive
- Descriptive : We have a motor
- Diagnostics : The motor is failing
- Predictive : The motor will fail in 2 weeks
- Prescriptive : Order motor so it will arrive before 2 weeks. Pay for express shipping if needed.
|Descriptive –> Diagnostic||Predictive –> Prescriptive|
|Basic Analytics||Advance Analytics|
Rockwell Automation Thin Manager
Rockwell Thin Manager is intended to be installed on the factory server rack and allows for the centralisation of various factory interfaces together in one thinclient framework. (E.g. IP cameras, HMD, remote desktops).
The server connects to various thin client terminals or android tablets, allowing for updating the server itself to upgrade all the thinclients, which would allow for easier security patching of the network and reduced attack surface for the factory.
The Thin Manager thin client remote operating system allow for sharing multiple virtual screens, so you can have different apps on different screens. E.g. monitoring a camera feed, while also checking the PLC’s HMD display of various valves state in the machinery.
Each terminal has two screens, one for the terminal specific interfaces, as well as a switchable screen for user specific displays that can be authenticated by fingerprint or RFID badge (e.g. Different screen for the floor manager vs the floor sparkie)
These two interfaces in the terminal screen are classified as:
- Terminal Context Tab (e.g. HMD)
- User Context Tab (e.g. User Apps)
Thin Manager is intended to be vendor agnostic, allowing for other companies to connect their device to Thin Manger via website, HMD, Remote Desktop Protocols or webcam.
Interesting Tech To Note
- CIP - Common Industrial Protocol
- Cisco Stealth Watch - Allows for network snapshot for security analysis
- Network Tagging - Allows for software defined networking for enterprise application. No need for VLAN setup.
Detecting Variable Speed Driver Faults
DC Drive Waveform (B) /| .......| |........(C) | \ ---- / \-------- \-/ (A)
- A : Load dip (Generating counter EMF?)
- B : Regeneration - Motor is pushing and overshooting
- C : If too high and sharp, this may trigger motor protection, shutting down the motor with an error condition.
- Overcurrent trip on a VSD drive, try checking the encoder wiring.
- IGOT Overtemp… Excessive load? High ambient temperature?