Applied Machine Learning: Beyond the Hype. E360 Annual Conference Atlanta, Ga. April 11 and 12

Transcription

1 Applied Machine Learning: Beyond the Hype E360 Annual Conference Atlanta, Ga. April 11 and 12 John Wallace Director Innovation, Retail Solutions Emerson Ron Chapek Director of Product Management Emerson

2 Applied Machine Learning: Digital Transformation The profound and accelerating transformation of business activities, processes, competencies and models to fully leverage the changes and opportunities of digital, data-driven technologies including the application of machine learning/artificial intelligence. 2

3 Applied Machine Learning: What Is Different? Sensors, switches, actuators, communication protocols, cloud storage and other core components of IIoT are not new. What is NEW are lower storage costs, more intelligence/computing power, ubiquitous networking and affordable subscription access to powerful analytics platforms. 3

4 Applied Machine Learning: New Business Models These technologies and the resultant (massive) increase in available data they generate will enable entirely new value-creation opportunities, business models and revenue streams. 4

5 Applied Machine Learning: Reality Check Many companies are in digital shock and are struggling to make the digital culture shift: Not started 37% Playing catch-up 24% On the adoption curve 14% Ahead of the adoption curve 25% 5

6 Applied Machine Learning: Focus on Value Generation Think big but start small, focusing on a specific task with a compelling business (value) proposition. Predictive asset management Asset life cycle management Maintenance cost optimization 6

7 Applied Machine Learning: Question How do you plan to implement AI technology in your enterprise? 7

8 What Is Machine Learning? Machine learning refers to being able to provide a computer with the ability to learn without programming. Machine learning is NOT Big Data, IoT, data analytics, dashboards, augmented (or virtual) reality, etc. There s a lot of math, but no magic. It s been around for awhile (1959) but recent events (i.e., Cloud processing, high-speed computer processors [CPUs], cost-effective data storage, etc.) have accelerated development and enabled real-world applications

9 Some Real-World, Everyday Examples; Machine Learning Is All Around Us How do they know what I am searching for? How do they know what to translate? It s the machine! (and a cloud) How do they know what I am saying? How do they know where to invest? (and connectivity) 9

10 thyssenkrupp Utilizing Machine Learning (Predictive Maintenance) to Drive Optimization in Elevator Maintenance Development championed by thyssenkrupp Innovation Center located at Tech Square 10

11 How Does Machine Learning Work? The Problem What We Are Trying to Do Inputs Process or Function Results or actions Inputs Machine Learning Prediction Model Results or actions Given a set of inputs, can we predict with sufficient accuracy a result or action taken as a result of the inputs? Note that Inputs can be anything (i.e., human language, sensor data, stock market data, etc.) and Results can be either human (i.e., translate English to French) or machine (i.e., predict a failure will occur). It Starts With Inputs and Resulting Actions (Data). 11

12 Simplified Machine Learning Process; Creating a Model Is a Data-Intensive, Iterative Process A Input 1 Input 2 Input 3 Result I11 I21 I31 R1 I12 I22 I32 R2 I13 I23 I33 R3 I14 BI24 Input 1 I34 Input 2 R4 Input 3 Result I11 I21 I31 R1 I15 I25 I35 R5 I12 I22 I32 R2 I16 I26 I36 R6 I13 I23 I33 R3 I17 I27 I14 I37 I24 R7 I34 R4 I18 I28 I15 I38 I25 R8 I35 R5 I19 I29 I16 I39 I26 R9 I36 R6 I17 I27 I37 R7 I18 I28 I38 R8 I19 I29 I39 R9 A B Input 1 Input 2 Input 3 Result I11 I21 I31 R1 I12 I22 I32 R2 I13 I23 I33 R3 I14 I24 I34 R4 I15 I25 I35 R5 I16 I26 I36 R6 I17 I27 I37 R7 I18 I28 I38 R8 I19 I29 I39 R9 Analyze and understand what you are trying to predict. Create training data set(s) and validation data set(s) from inputs and results. Use training data to evaluate different models performance and accuracy. Select initial model based on training data performance. Use validation data to check performance. Utilize new model to predict results based on new inputs

13 Other Examples: Machine Learning RTU Management ( Overlapping RTU s ) Sales Floor Area RTU s RTU s operating independently generate demand peaks which impact utility bills. Supervisory Control App learns (and predicts) response of space to RTU state and coordinates control. Coordination of RTUs facilitates comfort and reduces demand peaks. 13

14 Applying Machine Learning to Refrigeration Systems Data Cloud-Based Machine Learning Algorithms Operational Insights Sensors and Other Data Dashboard Delivers Machine Learning Algorithm Predicts Refrigerant Leak Additional Data Can Predict System Health and Performance. 14

15 Platforms Are Tools, but Not a Solution; Creating a Solution Requires Domain Knowledge and a Keen Understanding of the Problem The Tool Lots of very good machine learning platforms available today IBM Watson Microsoft Azure Google Amazon AWS Domain Expert The Solution Solution requires the right platform and a keen understanding of the problem Data Scientists Domain Experts Coders Machine Learning Platforms Make the Math Easier, but Still Need Domain Experts as Well as Other Roles to Create a Solution That Drives Value. 15

16 Keys to a Successful Machine Learning Deployment Lots of confusion, activity and buzzwords Domain knowledge key to understanding the problem to be solved and creating a solution Lots of data critical to creating a good model Models are only as good as the data used to create them Analyze data inventory to understand what is available and ensure key data is being collected A platform is not enough (but can help with the math) Start small, but look for something with impact 16

17 Thank You! Questions? DISCLAIMER Although all statements and information contained herein are believed to be accurate and reliable, they are presented without guarantee or warranty of any kind, expressed or implied. Information provided herein does not relieve the user from the responsibility of carrying out its own tests and experiments, and the user assumes all risks and liability for use of the information and results obtained. Statements or suggestions concerning the use of materials and processes are made without representation or warranty that any such use is free of patent infringement and are not recommendations to infringe on any patents. The user should not assume that all toxicity data and safety measures are indicated herein or that other measures may not be required. 17