0:00 Introduction 9:00 What is Rockwell doing in the AI space? 14:00 Difference between automation vs autonomy 17:50 How do customers opt in for data sharing? 23:15 How data rich industrial environments are... 26:10 Industrial data privacy and protection... How is Rockwell leading? 33:23 Vision ML Application use case question 39:40 How to approach a machine learning project? Custom vs off the shelf 47:32 How do system integrators / partners work with Rockwell to resell solutions? 52:05 Another ML application, how to communicate application needs to Rockwell? And how are they making it easier? 58:10 What are Rockwell's latest offering for Edge Solutions? (Logix Edge) 1:04:00 The importance of native integration / high speed applications 1:06:11 How do manufacturers get started with ML / AI?
This was super insightful from both of you. Agree with most of what was discussed - with one notable exception. I don't think most folks "should start with open source". Just like you shouldn't roll your own SCADA, you shouldn't roll your own AI/ML platform if something viable already exists (many of which are built on OSS underpinnings). Start with configurable/extensible solutions that allow rapid proof of value at little or no licensing costs.
Fair point Rick - if I correctly recall the part of the convo you were referring to, I was proposing that open source frameworks can be used to quickly evaluate potential use cases prior to a structured implementation with a commercial-grade product - e.g. quick analysis of predictive performance and potential gains in a python lab notebook. Many commercial products provide this exploratory features for this reason.
Great conversation guys! Discussion of use cases requiring both proprietary data plus commonly generated process or vision data to solve are particularly useful. Interesting to hear Jordan say he thinks there's a community willing to share their data to create ML-optimized, open source process models. I'm not so convinced.
For machines/ assets ( condition monitoring , vibration analysis, predictive analysis, etc ) people will be willing ( already shared ) but that should fall under the domain of OEM / SPMs for helping them improving their product, supplying spares in time of machines distributed all across various Geo locations (Initially when Thingsworks came out PTC used to give an example of a coffee grinding /making machine OEM how he was collecting data from various locations they had supplied their machines to ....... ....for special process know how , no way , no innovator will share it. Sell it may be !
You are correct, in fact we should *assume* that customers are not willing. Willingness comes from a clear value proposition, and use cases that aren't core to their IP or the methodological advantage. An example of "core IP" is the analogy walker gave of a beverage company allowing an AI vendor to train on their recipe. However, many companies grant certain data in certain circumstances if it means that they have access to a public model with superior performance because other companies have also opted-in. End users allowing asset condition feedback to the OEM is a good example. We also have examples related to common product quality issues in vision systems.
Hey Walker, I know you discussed pose estimation very shortly. Is there any guidance you could give a young career professional trying to get into this space? My back ground is in kinesiology, and am building experience in industrial ergonomics. I find Iot 4.0 interesting but feel that my background isn’t a match to get into a tech demanding job that builds such technologies? Am I reaching to say that ergonomics and system integrations are slowly merging? I feel the pressure to keep up with technology and that if I don’t I will not provide the best possible service or solution for those I try to help. It seems everyone is trying to build some ml/ai solution but I know you said the best way to work with pose estimation is to work with what’s been built. Any thoughts would be much appreciated, I know I probably sound a bit confused.
34:12 *loaded into a truck* by a crane system is an optimization program running in may dockyards, its kind of pick and place where not only the final position is important but also exact weight and shape of goods , from where the goods needs to be picked up and what acceleration and optimal parabolic path movement should be conducted at best possible speed for each load so that there is zero over shoot, ZERO knocking (accidents) avoiding obstacles in the path by other goods placed in its path , or swing happening (over /undershoot), this application has no AI model running . Similar application is pick and placing missiles from one ship to another where in mistakes can result in blasts and mind you both the ships are oscillating by the sea /oceanic waves. This is not at all a vision problem it’s a pure drive control along with an excellent pick and place optimization and path creation application ( Mathematical solution). I may have misunderstood walkers’ description, he could clarify. As per me he selected a wrong solution for the given problem. This solution have been used for ages …… even before the birth of AI !!! I am unable to digest the 30K solution given by walker, that is not a right solution and I dispute this claim !! I am 100% sure Rockwell dives group will give you a better solution or even Siemens and ABB who have been doing is for ages!!
Autonomous truck loading/un-loading is still an unsolved problem - vision systems are already used here, though they are insufficient on their own. Unloading pallets with different sized, packaging configurations and centers of gravity is still very difficult - loading pallets when encountering variations in truck geometry and position is also difficult. Vision can provide valuable feedback to a control system to help with the problem - think of semantic segmentation between different pallets, size estimation, etc.
@@jordanreynolds2751 There are programs regarding arranging different( size shape and weights) pallets optimally for a given 3D space or in Container( case here being a truck ) its geometrical calculation done extensively. Loading unloading is another problem which is solved by optimal path definition, speed , acceleration, steady movement retardation following a parabolic path from pick up position to deposit ( based on first para and the parcel being picked up in some sequence) and its a function done by sending various speed set points and tension control values dynamically to dive systems based on mathematical calculations and this is linked to its weigh , Center of gravity (one cant have skewed packaging having strange CGs, bad for any transportation too). Vision based in really not required where in the positioning of the truck can easily be handled in simpler manner to ensure they are in synchronized position. There is a crane operator even though he has a limited role to play in deciding the path and speed of operations. But that is for safety and emergency purpose so not totally Autonomous.
@jordanreynolds2751 There are programs regarding arranging different( size shape and weights) pallets optimally for a given 3D space or in Container( case here being a truck ) its geometrical calculation done extensively Loading unloading is another problem which is solved by optimal path definition( Again Mathematically calculated) , speed , acceleration, steady movement retardation following a parabolic path from pick up position to deposit ( based on first para and the parcel being picked up in some sequence) and its a function done by sending various speed set points and tension control values dynamically to dive systems based on mathematical calculations and this is linked to its weigh , Center of gravity (one cant have skewed packaging having strange CGs, bad for any transportation too). Vision based in really not required where in the positioning of the truck can easily be handled in simpler manner to ensure they are in synchronized position. There is a crane operator even though he has a limited role to play in deciding the path and speed of operations. But that is for safety guidelines.
CAN any one tell me why 2nd level reply gets deleted in youtube comments ? I am unable to give counter reply to the comment made on my comments !! Is this a feature or a new bug in RU-vid ? How do we mark this very issue appearing this page directly to youtube for this very presentation ?
Not sure Ravi, it looks like most of your recent replies are there in the comment threads on this video. Sometimes RU-vid does not display it the best…
@4.0Solutions Its a strange way of replying ... to reply to your comment i need to reply to my original comment and then it gets placed below your comment based on the date and time !! I cant directly reply to your comment , IF I DO THAT it gets deleted !!
16:23 *you realize that what this represents is a fundamental paradigm shift in the defin definition of a control system* I strongly disagree to this statement , car driving is totally different from process control , in driving you have rule based but dynamics of control is not specified and in control plant process systems programs you have clear definition of how dynamics have to occur based on process knowhow !! It not rule based , so mapping the 2 is not a fair process !! In process control the programing is never going to be disrupted, its never operator based ( driver based ) operator skills are to play only along side he controls that are well predefined by process experts and fine tuning can be at best done by operator , he can never bring out a control philosophy for different verticals unlike the driver who can operate in different terrains !! Under process control random events are exceptions and in vehicles driving random events are the one that needs to be taken care of or controlled of always even if the driving path is consistent
Ravi, we could probably spend 90 minutes on this topic alone. Autonomous driving and process control are notably different application domains, but both are relevant targets for emerging technology patterns in areas like perception (vision, soft-sensing) and control. Driving and process control have sort of inverse complexity trade-offs: in driving, the complexity is more-so in the perception of the environment and the ability system to reject environmental disturbances, but for a fixed environment the variables and dynamics of the vehicle are simpler. In process control it’s the opposite - you have less uncertainty in the environment, but more uncertainty in the variables and dynamics of the process itself. There is a reason that neural-network methods have been used for system identification in process industries since the early 90s, and these methods are being adopted faster now that compute is more widely available. Process industries often have a significant number of input variables, alongside multiple competing control objects (outputs). The relationships between these variables can be nonlinear, time-dynamic, and confounding. Controlling a process requires that a multi-objective optimization tradeoff is solved in real time. The processes also change over time. This makes system identification and system adaption difficult, and many companies are leverage learning methods like MLPs and RL with promising results. Look at what Yokogawa, AspenTech, Imubit and others are doing in this space. The intention of the commentary was not that autonomous driving and process control are equivalent application domains, but that some of the nascent technologies that we are seeing in driving have relevance to process control domains, and toward enhancing levels of autonomous decision making in industrial environments.
@@jordanreynolds2751 May be we can debate in some other forum , IMHO Autonomous Cars will be simpler to run ( whenever) purely on models rather than entire Process control,....... interesting you brought out AspenTech but the models there in my opinion are different in class, were they using MP and RL ?? I don know if they have now switched over to this. Have you worked on them ? There are Models ( Mathematical) which have been developed by Voestalpine in steel industry ( Was bought over by Siemens to restrict competition ( my opinion) broken down , ported their models into Siemens system and sold off their remaining sections. Similar to such models SMS Germany in steel industry is also a big player. These models have specific purpose and not to dive the control system in a different manner but to optimize the control points to get better efficiencies, tighter control resulting in better quality.
@jordanreynolds2751 May be we can debate in some other forum , IMHO Autonomous Cars will be simpler to run ( whenever) purely on models rather than Industrial Process control programs being replaced by models, Process control programs are here to stay for a very long time (Aliens may change our technological knowledge in which case it may happen sooner 😜)thus the logical programswont be replaced just be a model running with all sensors and actuators connected to it....... interesting you brought out AspenTech but the models there in my opinion are different in class and purpose, were they using MLP and RL ?? I don know if they have now switched over to this AND FOR WHAT REASONS !!. Have you worked on them ? There are Models ( Mathematical) which have ALSO been developed by Voestalpine in steel industry ( Was bought over by Siemens to restrict competition ( my opinion) broken down , ported their models into Siemens system and sold off their remaining sections. Similar such models SMS Germany has built in steel industry (SMS is also a big player). These models have specific purpose and DO NOT dive the control system in a different manner but optimize the control points to get better efficiencies, tighter control resulting in better OR CONSISTENT quality.
Here's the way I look at it - to the extent that system identification on the process is relatively easy, or the process dynamics are already known, explicit models are just fine (if not preferred). Where methods like MLPs or RL are useful is in system identification for complex process domains where traditional mechanisms are limited - think of multi-input multi-output systems with heavy non-linearity, lots of disturbances, and where adaptation of the process is relatively continuous. Or, alternatively, think of industrial robotics with complex objective functions and motion profiles. You can also leverage 'hybrid' modeling techniques where you start with any known mathematical models, and train a neural network to fit the data that is constrained by the known model (this is the PINN technique) Another thing to think about regarding RL is that traditional MPC methodologies require you to solve a multi-objective optimization problem in real time, within the control frequency. This limits the speed at which MPC systems operate. RL based methods that produce a 'policy' (which is a direct state-action neural network mapping) can run much faster (in linear time rather than polynomial time). So there are significant speed advantages during inference. Here is one interesting report from Yokogawa on this concept: www.yokogawa.com/us/news/press-releases/2022/2022-03-22/
