You are on page 1of 47

AI & MACHINE

LEARNING
FA L L E D I T I O N
Contents

3 — Introduction to artificial neural networks in


control applications

9 — Proficy CSense: Predict Product Quality

10 — AI vision for monitoring manufacturing and


industrial environments

18 — Machine Learning and Predictive Analytics for Engineers

31 — Seven tips for implementing machine learning in


controls environments

39 — Artificial intelligence applied to mill optimization

43 — How artificial intelligence will change work

2
Introduction to artificial neural
networks in control applications  Back to TOC

Practical applications of artificial neural networks (ANNs) for control systems,


especially for non-linear systems, include simulating time-optimal controllers
and for ANN-based controlled system (plant) models. Such models, combined
with classical proportional-integral-derivative (PID) controllers, can enable
adaptive and other, more sophisticated, control systems.

W hen reading the title, “Introduction to artificial neural networks in control ap-
plications,” you might ask if it’s worth reading more about artificial neural net-
works (ANN), successfully resolved decades ago, especially with thousands of ANN
articles on the internet. (Answer is: “Yes.” Please continue.)

When I became interested in ANN, I went through hundreds of ANN articles. Only a
handful helped me to understand how the ANN work and can be implemented. Even
so, I couldn’t find any article about how a control system using an ANN could be im-
plemented. I spent hundreds of hours experimenting with Python programming and
ANN and control systems simulations, until I became satisfied with my understanding
of how an ANN works in control system applications.

Today’s professionals don’t have to solve ANN problems from the scratch. They can
use elaborated libraries of classes, functions, etc., provided in the support materials
(libraries) of many programing platforms/languages. However, if you need to first un-
derstand ANN basics for control systems, then this article should help.

3
Introduction to artificial neural networks in control applications

1. Basics of neural networks  Back to TOC


Neural networks are trying to imitate abilities of the human and animal brains. The
most important of those abilities is adaptability. While modern computers can outper-
form human brain in many ways, they are still “static” devices, and that’s why comput-
ers can’t use all their potential. The use of artificial neural networks tries to introduce
brain functionalities to a computer by copying behavior of nervous systems. We can
imagine a neural network as a mathematical function that maps a given input set to a
desired output set. Neural networks consist of the following components:

• one input layer, x,

• one or more hidden layers,

• one output layer, ŷ,

• a set of weights and biases between each layer, W and b,

• activation function for each hidden layer, σ.

Figure 1 shows the architecture of a 2-layer neural network. (Note: The input layer is
typically excluded when counting the number of layers in a neural network.) Biases
are shown as an additional neuron (node) in the input and in the hidden layer, having
a fixed, non-zero value, for example 1, which is modified (multiplied) with the specific
weight coefficient values before they are added to the next layer nodes sums.

On the right-hand side of Figure 1, see one such (j-th) node of a hidden layer. At first,
4
Introduction to artificial neural networks in control applications

 Back to TOC

it sums up all the signals coming from the input Figure 1: Basic architecture of a neural
network. Courtesy: Peter Galan, retired control
(i-th) nodes, each affected – multiplied by its
software engineer
weight coefficient. Before the sum leaves the
node output, it goes through an activation func-
tion – limiter. The most popular limiter is implemented by the Sigmoid function, S(x), as
it can be relatively simply differentiated:

S(x) = 1 / (1 + e-x)

Similar nodes are used in the output layer. Just their input signals are coming from the
nodes of the hidden layer. 5
Introduction to artificial neural networks in control applications

 Back to TOC

The right values of the weight coefficients deter- Figure 2: Input – output mapping of a neural
network. Courtesy: Peter Galan, retired control
mine the strength of the predictions, that is, the software engineer
precision with which the input set will be trans-
formed into the output set. Figure 2 shows a typical input → output set mapping (on
the left side) and how the ANN can approximate such mapping by a linear function y
= W x + b (on the right side). A suitable mapping can be found even without applying
bias (offset), but generally adding the offset (bias) should yield better approximation.
Adding more hidden layers allows even a non-linear mapping.

A process of fine-tuning of the weight coefficients from the input and hidden layers
nodes is known as training the neural network. Each iteration of the training process
consists of the following steps:

• Calculation of the predicted output ŷ, a process known as feedforward

6
Introduction to artificial neural networks in control applications

• Updating the weight coefficients, a process known as backpropagation  Back to TOC

The output ŷ of a simple 2-layer neural network is calculated as:

ŷ = б(W2б(W1x + b1) + b2)

The predicted output, , will naturally differ from the desired output, , at least at the
beginning of the training process. How much it differs will tell us the Loss Function.
There are many available loss functions, but a simple, sum-of-squares error is a good
loss function.

Our goal in training is to find the best set of weights that minimizes the loss function.
Mathematically speaking, we need to find a loss function extreme (a minimum in our
case). Such a loss function doesn’t depend on just one variable (x). It can be a multidi-
mensional function having a complex shape with many local minimums. Our goal is to
find a global minimum of the loss function.

Let’s find a first derivation of the sum-of-squares error function, ∂Loss(y,ŷ)/∂W. Unfor-
tunately, our loss function does not depend directly on the weight coefficients, so we
need to apply the following chain rule for its derivation calculation:

7
Introduction to artificial neural networks in control applications

The first partial derivation yields: 2(y – ŷ )  Back to TOC

the second partial derivation yields: б’

and the last partial derivation is just: x

At the end we will get: ∂Loss(y,ŷ)/∂W = 2(y – ŷ) * б’ * x and this is what we will have to
implement as the backpropagation process. An advantage of using the Sigmoid func-
tion is a simple implementation of its derivative, б’ :

б’(u) = б(u)(1 – б(u))

The formula above describes how to proceed: Back propagate from the output of the
neural network to the hidden layer, which lays in front of the output layer, and which is
connected with the output layer via the W2 weights. The same process has to be ap-
plied again going from the hidden layer to the input layer, which is connected with the
hidden layer via the W1 weights. The first chain member in the loss function derivation
at the hidden layer, ∂Loss(h,ĥ)/∂ĥ, will have to be calculated differently, as we don’t ex-
plicitly know desired values of the hidden layer, h. We will have to calculate them from
the output values, as you will see in the following paragraph.

Peter Galan
Peter Galan is a (retired) control system designer with extensive experience in electronics,
control systems and software design. He worked for many companies like ZTS, GE, Husky,
Nortel, JDSU, Emerson (in Canada and the U.S.A.) and previously at the Technical Universi-
ty in Kosice (former Czechoslovakia). He holds a Ph.D. degree in Automated Control Sys-
tems and M.Eng degree in Applied Cybernetics from Czech Technical University in Prague. 8
Proficy CSense: Predict Product Quality

 Back to TOC


Proficy CSense: Predict Product Quality
In this demonstration we’ll see how a process engineer can create
a machine learning model to predict product quality that will help
reduce laboratory testing - using Proficy CSense. You will see
how to feed the optimal data into the CSense architect where the
model can be created, then publish to an external system such as
a SCADA or a DCS to improve quality and save cost.

9
AI vision for monitoring
manufacturing and industrial  Back to TOC

environments
Manufacturers can benefit from AI machine vision technologies by increasing
uptime, leverage preventive maintenance and more.

I n traditional industrial and manufacturing environments, monitoring worker safety,


enhancing operator efficiency and improving quality inspection were physical tasks.
Today, AI-enabled machine vision technologies replace many of these inefficient, la-
bor-intensive operations for greater reliability, safety, and efficiency. By deploying AI
smart cameras, further performance improvements are possible since the data used to
empower AI machine vision comes from the camera itself.

AI-enabled machine vision


In 2020, the AI machine vision market size for manufacturing and industrial environ-
ments was $4.1 billion, which, according to an IoT Analytics report, is forecast to grow
to $15.2 billion by 2025. That is a compound annual growth rate (CAGR) of 30% com-
pared to the 6.5% CAGR of traditional machine vision deployments. This high CAGR
is because the next generation of real-time edge AI machine vision is not limited to
quality assurance and product inspection applications.

Worker safety is a top priority in manufacturing and industrial settings, and AI-enabled
smart cameras help automate monitoring and inspection in these environments. It is
essential to ensure protection for employees, contractors, and other third-party op-
erators who work in potentially unsafe environments, such as dangerous mechanical
10
AI vision for monitoring manufacturing and industrial environments

equipment or hazardous materials. behavior and position (POSE) detection generates  Back to TOC
information that indicates whether machine operators are in danger, following stan-
dard operating procedures (SOP), or working in ways that enhance productivity and
efficiency. Automated optical inspection (AOI) also increases the speed and accuracy
of quality control.

AI for smart worker safety


Fatalities in industrial settings are not unheard of. When evaluating worker safety, fa-
cilities also must consider non-fatal work-related injuries. In addition to the emotional
trauma, there is often a financial consideration to take into account.

Industrial and manufacturing sites traditionally use human supervision and light cur-
tains for ensuring worker safety. However, human supervisors, who cannot be every-
where and see everything, are fallible and safety light curtains have their inherent
limitations.

Geo-fencing
In modern smart factories, people often work in potentially hazardous areas with dan-
gerous equipment, such as robotic arms. Safety light curtains protect personnel from
injury by creating a sensing screen that guards machine access points and perimeters.
However, they occupy lots of floor space, are difficult to deploy and lack flexibility. In
some instances, the safety light curtain’s limited response time may create additional
issues.

Conventional machine vision solutions use IP cameras and AI modules that are flexible
and easy to deploy but come with considerable latency and are, therefore, unsuitable
11
AI vision for monitoring manufacturing and industrial environments

 Back to TOC

for situations requiring an im- Figure 1: Safety light curtains take up floor space, are difficult to
mediate response. deploy, and lack flexibility; some are reaching limitations regarding
responsiveness. All-in-one artificial intelligence (AI) smart cameras,
such as Adlink Neon-2000 series, decrease latency by capturing
An all-in-one AI smart camera images and performing AI-related operations before sending results
can address this latency prob- and instructions to related equipment, such as the robotic arm. This
minimize delays, reduces space and bandwidth requirements, and the
lem. It captures images and camaras are easy to deploy and maintain. Courtesy: Adlink 12
AI vision for monitoring manufacturing and industrial environments

performs all AI-related operations before sending results and instructions to related  Back to TOC
equipment, such as the robotic arm [Figure 1]. Compared to light curtains and con-
ventional machine vision implementations, using an all-in-one smart camera minimizes
delays, reduces space and bandwidth requirements, and is easy to install and maintain.

Real-time machine vision AI offers additional benefits to augment worker safety by


alerting users if they enter an unsafe zone and logging that information for retraining
purposes. Logging data from past events can be helpful in the future. For example, if
a worker approaches a hazardous area, instead of the robotic arm shutting down com-
pletely, it could go into a functional safety process loop. These can help improve work-
er safety as well as increase the factory’s operating efficiency.

Smart refueling with AI vision


When a fuel truck arrives at a manufacturing facility, it introduces the potential for mul-
tiple safety issues smart AI vision can resolve. First of all, the truck may roll if the brake
is not applied correctly or fails. Training the AI machine vision system to monitor the
truck for movement enables it to raise an immediate alarm should the status change.

Facilities also must consider the operator’s location during the refueling process be-
cause there are different types of zoning breaches. It is critical to ensure that all work-
ers on-site understand that there are safety risks. For example, it is necessary to place
traffic cones at the four corners of the truck and ensure that the operative refueling the
truck is wearing the appropriate personal protective equipment (PPE) – AI smart vision
can perform all safety checks to confirm procedures are met correctly [Figure 2].

Immediate alerts from the AI machine vision system can warn operators of a safety
13
AI vision for monitoring manufacturing and industrial environments

breach and prevent  Back to TOC


injury. It also creates ac-
countability; if someone
enters an unsafe zone
without the appropriate
PPE, the logged imag-
es can flag errors and
educate employees to
prevent future mistakes.

POSE detection
teaches repetitive injury avoidance Figure 2: Although supervisors can be
present to reinforce safety procedures, it is
For the manufacturing industry, “cycle time” is a not always possible. If a person breaches a
critical performance index for production efficien- hazardous area, AI smart machine vision can
cy. It represents the amount of time a team spends sound an immediate alarm. Courtesy: Adlink

producing an item until the product is ready for


shipment. Monitoring employee behavior and position with AI smart camera technolo-
gy helps enforce SOP and improve worker efficiency.

POSE detection from live video plays a critical role, enabling the overlay of digital con-
tent and information on top of the analog world. POSE describes the body’s position
and movement with a set of skeletal landmark points such as a hand, elbow or shoulder.

AI machine vision enables factory operators and workers to focus on how physical po-
sitions affect their work. POSE data can help train operators to learn where they should
place their arms and hands to work more ergonomically and efficiently [Figure 3].
14
AI vision for monitoring manufacturing and industrial environments

Tracking whether an  Back to TOC


operator is present at
their workstation on
the production line also
automates and verifies
timesheets. Making sure
they are following SOP
helps ensure quality con-
trol and line balancing. Figure 3: POSE detection on an electronics
manufacturing line can help increase productivity,
as well as improve order, quantity, and line
AI smart AOI for contact lens balance. Courtesy: Adlink
inspections
Manual product quality inspection is time-consuming, inconsistent and can create
production line bottlenecks. Conventional AOI machine vision can detect easy-to-find
defects faster than professional quality control staff due its exceptional accuracy and
efficiency. However, when a fault is difficult to detect, such as a flaw on a contact lens,
these machine vision systems reach their limits for accuracy and consistency.

While most manufacturers randomly sample test products for flaws, this is not possible
on contact lens production lines as every lens needs inspecting. Quality control staff
can only view up to 4,000 lenses per shift, creating production bottlenecks. False dis-
covery rates and missed detections also are inevitable.

As contact lenses are transparent, implementing machine vision-based detection has


been a challenge for manufacturers. Conventional AOI relies on fixed geometric al-
gorithms to discover defects, but acquiring quality images from transparent objects
15
AI vision for monitoring manufacturing and industrial environments

 Back to TOC

is challenging, which results in unacceptable Figure 4: AI smart AOI can detect even
detection performance. minute defects in transparent contact lenses,
significantly improving inspection rates
compared to the previously used manual quality
Collecting data using AI smart cameras to train control processes. Courtesy: Adlink
the AI algorithms and iterate on inspection
performance gains provides a more favorable solution. The AI smart system can identi-
fy the most common defects, including burrs, bubbles, edges, particles, scratches, and
more [Figure 4], and maintain inspection logs for customer reference.

Each AI smart camera can inspect 50x more lenses than manual visual inspection, with
accuracy improvements from 30 to 95%.

16
AI vision for monitoring manufacturing and industrial environments

Machine vision with AI increases uptime, safety  Back to TOC


Manufacturers that leverage robust, real-time data from AI machine vision technolo-
gies can increase uptime, leverage preventive maintenance enhance productivity and
worker safety and many other workplace benefits.

The AI machine vision applications require AI algorithms for deep learning. The soft-
ware experts that develop AI algorithms need a smart, reliable platform for executing
AI model inferencing. AI smart cameras with pre-installed edge vision analytics (EVA)
software can address many issues common to conventional AI vision systems, improve
compatibility, speed up installation, and minimize maintenance issues.

To successfully deploy an AI vision project, it may take engineers as long as 12 weeks


to conduct a proof of concept. It takes considerable time to overcome the learning
curve of choosing optimized cameras and the AI inference engine, retraining AI mod-
els and optimizing video streams. EVA software simplifies these steps with its pipeline
structure and shortens the PoC time to up to 2 weeks, serving as an excellent starting
point to kick-start the AI vision project.

Chia-Wei Yang
Chia-Wei Yang, director, business and product center, IoT Solution and Technology
Business Unit, ADLINK

17
Machine Learning Leverage Domain Expertise to Drive Optimization
– Without Needing a Data Scientist

and Predictive T oday, staying competitive means progressing on a dig-


ital transformation journey, including machine learning

Analytics for and predictive analytics. Not only can industrial organi-
zations capitalize on the IoT opportunity, optimize oper-

Engineers
ations and generate greater profitability, but engaging in
the latest technologies also helps to attract and retain the
best talent.

Fortunately, the journey to success with machine learn-


ing and analytics doesn’t mean that process engineers
need to be data scientists. Proven processes and software
technologies make analytics do-able for every industrial
organization.

Engineers Are the Process Experts


Process engineers have exceptional domain expertise to
put together process models – or Process Digital Twins –
and be able to interpret the models. This is the foundation
for improving competitive advantage and success with
analytics.

To drive analytics and improve processes, process engi-


neers can align domain expertise to five capabilities:
Machine Learning and Predictive Analytics for Engineers

1. Analysis - au-  Back to TOC


tomatic root
cause identifi-
cation acceler-
ates continuous
improvement

2. Monitoring –
early warnings
reduce down-
Figure 1: Process engineers can align
time and waste domain expertise to five capabilities.

3. Prediction – proactive actions improve quality, stability, and reliability

4. Simulation – what-if simulations accelerate accurate decisions at a lower cost

5. Optimization – optimal process setpoints improve throughput at acceptable


quality by up to 10%

Advanced analytics techniques are available to industrial process engineers to fulfill on


these capabilities. To support the journey to machine learning and analytics, GE Digital
provides analytics technology training in the form of a self-serve detailed demo videos
and application advice.

Additionally, while today’s software features enhanced ease of use and no-code imple-
mentation extensible with Python, process engineers can still lean on product experts
19
Machine Learning and Predictive Analytics for Engineers

in combination with their own domain expertise to mine data and leverage analytics to  Back to TOC
improve operations.

Success with Predictive Analytics


As an example, a leading food manufacturer was able to drive down customer com-
plaints by more than 33% through analytics. The manufacturer had struggled with
weight control on a cube-shaped product.
Make the cubes too heavy, and the manufac-
turer was giving away product or producing
watery product if the excess weight was due
to too much water. When the cubes were too
light, the company was in regulatory jeopar-
dy as well as having trouble compacting the
product into a stable cube shape.

The team used Proficy CSense to get a


complete, correlated-by-lot and period pic-
ture of: ingredient specs, process variables
as run, and lab data – using the software to look for Figure 2: A leading food manufacturer
was able to drive down customer
controllable factors that correlated to excess give- complaints by more than 33% through
away and then comparing periods with better weight analytics, using Proficy CSense.
control to the factors that were true then. Now, when
the team sees how a raw material variance was successfully corrected for or a process
disturbance was overcome, that understanding is embedded into a new material spec,
recipe or SOP. The smart analysis with Proficy CSense yielded other benefits as well.

20
Machine Learning and Predictive Analytics for Engineers

Another example involves applying a smart predict project at a pulp and paper manu-  Back to TOC
facturer to predict Critical to Quality (CTQ) KPIs to improve productivity and eliminate
wastewater regulatory issues. As a final example, a partner in mining delivered an Ad-
vanced Process Control solution that increases throughput by 10% using smart optimi-
zation technology.


Predict Product Quality with Analytics
See how a process engineer can create a machine learning
model to predict product quality that will help reduce
laboratory testing.

Analytics and Control Loop Optimization


As a use case example, process optimization is key in manufacturing, and control loops
are the critical components. “Out of tune” loops can affect the quality of the product,
the material and energy consumption, and ultimately increase the risk of contamination
in regulated industries. AI and machine learning can be used to improve and optimize
21
Machine Learning and Predictive Analytics for Engineers

control loops to generate  Back to TOC


big savings and reduce
risks.

What is a control loop? A


simple form of a process
controller is the thermo-
stat which maintains the
temperature of a room
according to a given
setpoint. It operates as a closed loop control device, Figure 3: AI and machine learning can
be used to analyze and improve control
trying to minimize the difference between the room
loops to optimize performance KPIs,
temperature and the desired one. generate big savings and reduce risks.

The industrial version is the PID (Proportional-Integral-Derivative) control loop - an es-


sential part of every process applications. PID loops have been around for a very long
time. The first pneumatic instruments featuring a proportional controller were devel-
oped by Taylor Instrument Companies at the beginning of the 20th century.

Nowadays, loop controllers are available as standalone devices called single loop con-
trollers, but the most common version is a piece of code that resides in a PLC (Process
Logic Controller) or a DCS (Distributed Control System). It makes it easier to combine
them to create advanced control diagrams like cascade or feed-forward control, or
split range required for the complex control of food and beverage, chemical, oil and
gas operations, and more.

22
Machine Learning and Predictive Analytics for Engineers

Control Loops: Small Tweaks, Big Impact  Back to TOC


A lot of literature exists that describe the behavior of PID loops and how to tune them.
However, it still represents a challenge for many as all the processes are different.

The complexity of the process is obviously one of the criteria. Heat jacketed devices
such as kettles, dryers, reactors or pasteurization units can be hard to control. Using
steam, the heat
transfer is not uni-
form which might
result in an overshoot
during uptimes,
making the control
loops difficult to
tune. Note that this is
less prevalent using
water. Traditional
Figure 4: An advanced analytics system
cascaded loops will only solve part of the problem. An can solve problems by looking at
advanced analytics system such as GE Digital’s Proficy historical data – including from multiple
sources – to create a model of the
CSense can help by looking at historical data to cre-
actual profile and recommend new
ate a model of the actual profile and recommend new settings accordingly.
settings accordingly. The model will take into account
the change of parameters such as viscosity and steam pressure which affect the heat
transfer coefficient and the flow pattern.

Two apparently similar machines might require different settings as they are equipped
with sensors that will react to change in a slightly different way. This might be because
23
Machine Learning and Predictive Analytics for Engineers

they use different technologies – a glass vessel vs a steel vessel, which by nature have  Back to TOC
different inertia - or simply because their characteristics vary over time: aging valves,
deviating sensors, etc.

Loop tuning therefore doesn’t happen once. It must take place on a regular basis if
done manually and customized for each asset. Another option is real time monitoring
using AI and machine learning. Analytics make use of a suite of components to deter-
mine and understand the causes of process
deviation in industrial environments. Engi-
neers and data scientists can analyze, monitor,
predict, simulate, and optimize and control
set points in real time.

For process optimization, analytics solutions


need to provide multiple capabilities: process
modelling and troubleshooting as well as on-
line deployment and real time monitoring.

Data is prepared, visualized, and rules-based, Figure 5: Analytics solutions


data driven process models can be constructed. Using need to provide multiple
capabilities: process modelling
these models, root causes of process deviations are iden- and troubleshooting as well as
tified, so processes can be optimized. online deployment and real time
monitoring.

Predictive Analytics and Sensor Health


Applications for predictive analytics are endless, but a possible first step: engineers
can use analytics to monitor sensor health.
24
Machine Learning and Predictive Analytics for Engineers

Bad sensor data can mean lost product, downtime, compliance issues, and safety risks  Back to TOC
as well as a dirty data foundation for digital transformation and continuous improve-
ment programs. Industrial organizations need to have good data that can be leveraged
for operations, ad hoc analysis, and enterprise analytics.

Over time, sensors tend to deviate, impacting processes and operations. But, it’s time
consuming – and impossible for most organizations – to manually determine if and why
sensors are working or failing ahead of increasing risk.

Predictive Analytics and Sensors


Engineers can employ a predictive analytics app such as Proficy Sensor Health to
continuously monitor and analyze sensor data. Users can target anomalies and mini-
mize their potential impact. The analytics app provides an easy way to automate the
detection of bad sen-
sors, where data is
deviating from normal
conditions.

When an anomaly is de-


tected, the app can gen-
erate alarms to speed
repairs, replacements,
and recalibrations.

By using predictive ana- Figure 6: Follow these strategy steps to good

25
lytics to monitor sensor health, engineers can: sensor health, leveraging predictive analytics and
machine learning.
Machine Learning and Predictive Analytics for Engineers

• Reduce downtime: Sensors are often used to provide indications that equipment  Back to TOC
is running correctly. Incorrect readings can lead to equipment failure or damage.
Early detection of a sensor that is no longer giving accurate or consistent results
can provide advance warning that enables maintenance to replace or recalibrate
the sensor before the worst happens.

• Improve product quality and compliance: Sensors are often used for measuring
the results of a product or to ensure the ambient surroundings of a process are
within specification. If the sensors used to measure the product or environment
are not accurate or functioning correctly, it can lead to a product being out of
specification. Providing early warnings can reduce costs related product recalls or
scrapped product.

• Ensure sensor data quality: Ensuring data quality in downstream analytics is


part of IoT-fueled improvements. If the intent is more advanced use of analytics
for a process, the need for ensuring data quality is critical

How to Select the Right Predictive Analytics Solution


Every industrial organization knows that it’s imperative to move ahead at light speed
with predictive analytics. Consequently, new analytics startups and consulting compa-
nies are showing up every week.

How does an engineer select the right predictive analytics solution?

The following advice provides guidance for engineers.

26
Machine Learning and Predictive Analytics for Engineers

1. Make sure that “predictive analytics” isn’t buzz hiding risky, unproven software  Back to TOC
and newbie “experts.”

For example, GE Digital has offered and implemented analytics solutions for
more than 15 years, serving industrial organizations around the world across a
diverse set of industries.

Work with a partner that you can trust and know will support you for the long term.

2. Select industrial advanced analytics with a troubleshooting component that en-


ables engineers to rapidly trou-
bleshoot continuous, discrete,
or batch manufacturing process
performance by mining insight
from available sensor and produc-
tion data.

Seamless connectivity, rich visu-


alization, and predictive analytics
enable engineers to analyze op-
erating scenarios, quantifying the
impact that operational changes
will have on key performance
metrics and identifying causes for Figure 7: To support the full IoT value journey,
performance variation. look for capabilities from simple calculations to
predictive machine-learning models to real-time
optimization and advanced-control algorithms.
27
Machine Learning and Predictive Analytics for Engineers

Additionally, to support the full IoT value journey, look for capabilities from sim-  Back to TOC
ple calculations to predictive machine-learning models to real-time optimization
and advanced-control algorithms.

3. Be sure that the analytics package enables engineers to rapidly develop analytic
solutions – supporting improvements in production throughput, yield, quality, and
efficiency with significant margins.

A comprehensive analytic solution-development environment provides visual


analytic building blocks to build and test calculations, predictive analytics, and
real-time optimization and control solutions with connectivity to real-time and his-
torical data sources and drag-and-drop access to rich functional libraries.

Plug-and-play connectivity to historical and real-time data sources and automa-


tion systems make for faster configuration. Built-in support for data quality makes
real-time data cleaning and validation easy.

4. Confirm that the analytics package can speed deployment with templates for
greater efficiency.

Engineers should be able to save analytics solutions as reusable templates for


easy deployment to similar assets or process units. Additionally, while the analyt-
ics troubleshooting component should enable engineers to find answers faster
with analytics-guided data mining and process-performance troubleshooting, the
development/configuration capabilities should allow them to more easily capture

28
Machine Learning and Predictive Analytics for Engineers

expert knowledge and best practices into high-value analytic templates for rapid  Back to TOC
enterprise-wide deployment.

5. Focus on analytics solutions developed with engineers – not just data scientists –
in mind.

With an analytics package that is accessible to engineers, teams can create a Pro-
cess Digital Twin for smarter operations.

Visual drag-and-drop analytics accelerate time to value and reduce depen-


dence on data scientists and programmers. Online demos enable rapid mas-
tery of the software with easy-to-follow demonstrations and guided simula-
tions.

As previously mentioned, a rapid development environment is critical. The best


solutions provide rapid wizard-driven data mining for engineers for fast time-to-
insight, an easy visual drag-and-drop environment for subject matter experts and
engineers, and analytic solution templates without programming for simple calcu-
lations, data cleaning, maths, statistics, machine-learning models, real-time opti-
mization, and advanced process control.

From Small Projects to Multi-Plant Optimization


All automation and process engineers can and need to develop capabilities in analytics
and machine learning to remain competitive – both at an individual professional level as
well as to help their industrial organization – in our world of digital transformation.

29
Machine Learning and Predictive Analytics for Engineers

Over time, engineers can go from small projects to pilots to multi-plant optimization  Back to TOC
with deep application of analytics. Engineers’ deep domain expertise provides a foun-
dation for modelling processes and developing the analytics that are game changers in
very specific applications. The combination of applied analytics technology with those
Process Twin models uncovers hidden opportunities for improvement over and over
again.

If you’re ready to optimize with analytics, GE Digital’s Proficy CSense turns raw data
into real-time value with a Process Digital Twin. The software uses AI and machine
learning to enable process engineers to combine data across industrial data sources
and rapidly identify problems, discover root causes, and automate actions to continu-
ously improve quality, utilization, productivity, and delivery of production operations.

30
Seven tips for implementing
machine learning in controls  Back to TOC

environments
Machine learning (ML) applications are evolutionary by nature, so it is
important to understand how they work and keep looking for new ways to
apply them to deliver an automation advantage.

M achine learning (ML) technologies have proven their value in real-world manufac-
turing applications. However, many engineers and plant managers still think of
ML as a nebulous, mysterious force floating around in the cloud somewhere that only
data scientists can access. ML is not nebulous or inaccessible, but provide an automa-
tion advantage. While there are many approaches to ML projects, and more continue
to emerge, the best way to implement ML in an industrial context is incorporating it
directly into the controls environment. Consider these seven tips to get started.

1. Choose the right project for machine learning


Engineers and plant managers always should assess which types of applications are a
good fit for ML and how they can effectively use ML. Just because ML is new to many,
exciting and becoming more accessible does not mean ML is the panacea for every
unsolved engineering challenge. ML implementations within the machine controller
can offer huge innovation and competitive leads.

Applications for ML in machine control often fall into the category of application prob-
lems that are challenging to solve using traditional programming algorithms. If a pos-
sible application is identified, users should implement it as a prototype quickly and in
31
Seven tips for implementing machine learning in controls environments

 Back to TOC

an agile environment. The agility of the project Figure 1: Upgrading control platforms can
protect against future obsolescence, increase
team is a decisive factor. ML projects are, by
system flexibility, enhance technology or meet
nature, evolutionary and iterative processes. budgetary or cost of ownership requirements.
They do not fit into a tight, set framework. The Courtesy: Beckhoff Automation

general project flow starts with data collection,


data preparation, model training, testing, more data collection, model refinement, etc.,
until the model produces optimal and reliable predictions. 32
Seven tips for implementing machine learning in controls environments

2. Select an integrated software solution for ML  Back to TOC


Since the goal is to run ML in a real-time controls environment, the more integrated
a solution is, the better. This allows fast machine reaction times based on ML model
predictions. Some vendors offer inference engines for ML models that are incorporat-
ed into the standard automation software, with the execution of models such as neural
networks directly in the real-time environment.

An integrated approach supports nearly unlimited fields of machine application, such


as ML-based solutions for quality control and process monitoring/optimization applica-
tions. For example, fully automated and system-integrated quality control applications
based on existing machine data, such as motor currents, speeds and tracking errors,
enable the machine controller to give accurate product quality predictions on 100% of
goods produced. This works 24/7 and at a rapid cycle time.

Continuous process monitoring and process optimization are sequential steps: If a pro-
cess can be monitored with a trained model, the machine can send a message to the
machine operator to adjust the process to maintain quality. The next step is to learn
from the experienced machine operator and train the model to offer parameterization
suggestions or simply make the necessary parameter adjustments autonomously.

3. Understand the control hardware requirements


When most people think of ML, it brings up an association of needing very large com-
puter hardware and many GPU cores. For training ML models with massive data sets,
this significant computing power might be required. However, running the resulting
trained models, that is, the inference, requires less-powerful hardware. On optimized
solutions, the inference engine runs on the hardware side of powerful industrial PCs
33
Seven tips for implementing machine learning in controls environments

 Back to TOC

(IPCs), with access to the high computing capacity Figure 2: A large Chinese food company
of modern CPU architectures. This approach makes uses TwinCAT Machine Learning to achieve
the highest possible quality level in the
the execution of learned models more efficient by packaging of instant noodles. This image
using specialized CPU command set extensions in shows the standard ML workflow and
combination with optimized CPU cache memory an anomaly detected and displayed in
TwinCAT Scope View the system integrator
management. Also, the trend towards more and used to implement the solution. Courtesy:
more processor cores per CPU supports the accel- Beckhoff Automation
34
Seven tips for implementing machine learning in controls environments

erated execution of neural  Back to TOC


networks.

A close look at the trained


model is always necessary.
Just like “handwritten”
source code, it could be
large and inefficient, which
creates longer execution
times than lean and opti-
mized source code. The
ML models should always
be adapted and optimized
Figure 3: Beckhoff introduced TwinCAT ML
to the task. With the right hardware and solid at Hannover Messe 2019, and since then,
source code, it’s possible to execute neural net- the number of industrial applications using
trained neural networks has continued to grow
works in the microseconds range with a stan-
exponentially. Courtesy: Beckhoff Automation
dard PC-based machine controller.

4. Start planning early, but keep exploring options


An ML project is an evolutionary process. Engineers should start as early as possible in
the value chain at the machine builder. More refined model predictions take significant
data for training. More data often means more refined and accurate models. Therefore,
not every application will have the most optimal solution for a new machine design
when first put into operation at an end customer.

However, during the machine lifecycle, new relevant data can be collected and evalu-
35
Seven tips for implementing machine learning in controls environments

ated. This helps ML models constantly improve. To support this process, some vendors  Back to TOC
set up the inference engine so it can load new models on the fly without stopping the
machine, restarting the PLC or recompiling source code.

Many applications already have an existing machine with a controller that does not
include ML functionality. End users want the machine to optimize their production and
are now considering ML use. In these cases, an open-control machine-control concept
plays a decisive role. Choose a solution that readily connects to third-party controls
through various interfaces. In many cases, this can allow ML implementation on an
older machine by installing an IPC that has read access to the essential data from the
existing controller and hosts the inference for implementing, for example, inline prod-
uct quality predictions.

5. ML projects require collaboration among data and domain


engineering experts
Since machine knowledge and database expertise are critical in ML applications, the
project team should be made up of several different experts. The main player is the
domain expert, that is, a mechanical engineer or the expert for linear actuators or the
forming process. The domain expert wants to solve a specific ML challenge so they
have a goal in mind and know the interrelationships in their machines. Next is the data
scientist, who is responsible for data analysis. Working together, these two must define
essential machine variables, which is important for the defined goal. The data scientist
works closely with the domain expert to shed light on the meaning of certain data pat-
terns and behaviors.

A data scientist alone, without feedback from the domain expert, lacks the know-how
36
Seven tips for implementing machine learning in controls environments

to make these projects  Back to TOC


effective. Some machine
builders already have
data science depart-
ments, possibly only an
individual resource, and
take on this task them-
selves. Others need
more intensive support,
which is available from
some vendors or special-
ized automation inte-
grators. Factor the level
of support needed into
which solution to pick.
Figure 4: If data are continuously collected
from the machine and the classification result is
6. Protect IP while gathering ML also saved, the data scientist and the domain
expert can subsequently analyze in more detail
training data precisely those process sequences in which
A resulting trained ML model can have a huge an anomaly was detected. Courtesy: Beckhoff
Automation
impact for a company’s competitiveness. There
also has been engineering and collaboration
between experts of different domains and specialists and lots of collected data. There-
fore, when working on ML projects, always have the necessary sense of proportion re-
garding data and IP protection. After the ML models are trained, they will be deployed
to production or end user facilities. The trained model can, and in most applications
should, be protected from being copied and used in an unauthorized manner. These
37
Seven tips for implementing machine learning in controls environments

software protection mechanisms in some platforms can protect not only ML models,  Back to TOC
but also PLC source code and deployed code.

7. Understand how to train ML models to detect anomalies


It can be hard to train for anomalies when engineers don’t know what they are. Many
paths can lead to the destination. A simple, easy-to-describe example is to train a clas-
sification model with only one known class – the “no anomaly” class. Present the mod-
el in training with only data that contains no anomalies, and define this set of data as
“class A.” When the algorithm is used in the process, it recognizes the “class A” again.
However, it also recognizes when the data have some other unknown structure, and
then reports an unspecified anomaly.

It bears repeating: ML is an evolutionary process. If data are continuously collected


from the machine and the classification result is also saved, then the data scientist and
the domain expert can analyze in more detail those process sequences in which an
anomaly was detected.

If necessary, a model update can then be used for anomaly detection and to narrow
down the detected case in more detail. As users continue gaining experience and un-
derstand the tools, they will see how many opportunities and approaches there are to
automate machine performance and quality optimizations with machine learning.

Daymon Thompson
Daymon Thompson, automation product manager – North America, Beckhoff Auto-
mation.

38
Artificial intelligence applied
to mill optimization  Back to TOC

Variability controlled with artificial intelligence (AI) software. A digital


twin-based AI optimizer was deployed to run alongside advanced process
controllers (APCs). Results include a 1% increase in mill production
throughput with annual topline impact of $3-4 million for the mine. See 4
objectives for recurrent neural networks and self-adaptive tuning.

T he grinding mill is one of the largest pieces of equipment used in the mining and
minerals industries, and artificial intelligence has been applied to help advanced
process control increase throughput 1% and decrease variability for millions of dollars
in annual impact for the mine.

Grinding mills often pulverize hard ores and as a result, is subject to large forces that
affect its wear, life, and safety. The grinding circuits often consists of sag (semi-autoge-
nous grinding) mills, ball mills, pebble crushers, screens and hydro cyclones to produce
the desired particle size distribution in the powdered ore. Mills often process ores of
variable hardness, shape, and size. This inherent variability is important to account for
during the control of mills since there is always a possibility of a build-up of ore inside
the mill to an extent that it damages the equipment or causes an overload condition.
As a result, it is not uncommon to find plants running below peak conditions.

Changing raw materials required tighter controls; AI helps


A large gold mine operation had similar issues. An investment in advanced process
controllers were not delivering the intended benefit of maximizing production at the
grinding mills. The APC models were tuned for a certain type of ores and were strug- 39
Artificial intelligence applied to mill optimization

gling to maintain performance in the light of changed ore conditions. The grinding mill  Back to TOC
was processing ores of variable size, shape, hardness, etc., and had to be controlled to
mitigate the possibility of ore build-up inside the mill and subsequent overload or run-
away conditions. In fact, this dynamic variability of ore-feed along with other process
uncertainties soon created a bottleneck of automated control operations by the APC
and often had to be turned off intermittently for plant operators to take direct control.

The APC parameters needed constant adjustments from experts, which was not avail-
able at the mines. Since the mill’s APC system lacked the ability to automatically and
adaptively deal with the inherent variability and uncertainty, operators ran conservative
strategies that reduced the chance of runaway or overload conditions. However, this
also reduced the mill’s throughput, which meant reduced gold production.

Digital twin model driven by AI


AI-driven digital twin models offer dynamic process optimization that can predict pro-
cess uncertainties and recommend optimized control strategies to increase through-
put and reduce the chance of failures. The digital twin models use highly performant,
adaptive self-tuning AI models that can capture operating dynamics, predict unknown
quantities and find optimal control parameters in real time.

In this case, the digital twin-based optimizer was deployed to run alongside the APC.
As part of the digital twin model development, a detailed exploratory data analysis of
one year’s worth of historical process time-series data at 1-minute frequency intervals,
lab analysis of ore feed data, production and maintenance logs were undertaken to
glean model features and identify baseline operating data for model training and test-
ing.
40
Artificial intelligence applied to mill optimization

4 objectives for recur-  Back to TOC


rent neural networks,
self-adaptive tuning
The heart of the digital twin
model was based on recurrent
neural networks with self-adap-
tive tuning mechanism to take
into account of process changes
with an aim to meet the following
objectives:

1. AI-based approach with a


supervisory level of deci-
sion-making support for
APC

2. Coordinated constraint
management across the mill
circuit to get the best out
of mill controllers including
APC

3. Cognitive learning-based
approach, factoring in oper-
ator and metallurgical engi-
neer expertise and insights
41
Artificial intelligence applied to mill optimization

4. Continuous adaptation of AI-models to dynamically account for ore and process  Back to TOC
variations.

Typical advisories from the digital twin optimizer would include mill ore feed, power,
load, water additions and mill speeds.

The trained digital model deployed online with live data ingestion capabilities from
plant distributed control system (DCS), lab computer systems and asset management
systems, continuously predict, and run optimizers to provide adjustments of the APC
control parameters in real-time. This allowed the APC to adapt with changing condi-
tions and enable the grinding mill to perform at higher throughput levels.

The installation of AI software caused no disruptions to existing operations. This as-


pect is an important consideration when considering AI solutions at a plant. Process
disruptions or additional hardware and instrumentation can cause potential roadblocks
to successfully implement an AI program.

The outcome from this AI implementation showed an uptick of more than 1% increase
in mill production throughput with annual topline impact of $3 to $4 million for the
mine.

Dominic Gallello
Dominic Gallello is CEO of Symphony AzimaAI.

42
How artificial intelligence will
change work  Back to TOC

MIT researchers offer insight on the challenges and potential of artificial


intelligence (AI) and how they will change the nature of work.

M IT Professor Thomas Malone, Professor Daniela Rus, and Robert Laubacher col-
laborated on “Artificial Intelligence and the Future of Work,” a brief that pro-
vides a comprehensive overview of AI today and what lies at the AI frontier. Malone
and Rus offer insights in an overview on their research.

Q: You argue in your brief that despite major recent advances, artificial
intelligence is nowhere close to matching the breadth and depth of per-
ception, reasoning, communication and creativity of people. Could you
explain some of the limitations of AI?

Rus: Despite recent and significant strides in the AI field, and great promise for the
future, today’s AI systems are still quite limited in their ability to reason, make deci-
sions, interact reliably with people and the physical world. Some of today’s greatest
successes are due to a machine learning method called deep learning. These systems
are trained using vast amounts of data that needs to be manually labeled. Their perfor-
mance is dependent on the quantity and quality of data used to train them. The larger
the training set for the network, the better its performance, and, in turn, the better the
product that relies on the machine learning engine. But training large models has high
computation cost. Also, bad training data leads to bad performance: when the data
has bias, the system response propagates the same bias.

43
How artificial intelligence will change work

Another limitation of current AI  Back to TOC


systems is robustness. Current
state-of-the-art classifiers achieve
impressive performance on bench-
marks, but their predictions tend
to be brittle. Specifically, inputs
that were initially classified cor-
rectly can become misclassified
once a carefully constructed but
indiscernible perturbation is added to them. Thomas Malone (left) is director of the MIT
An important consequence of the lack of ro- Center for Collective Intelligence and the Patrick
J. McGovern Professor of Management. Daniela
bustness is the lack of trust. One of the wor- Rus is director of the Computer Science and
risome factors about the use of AI is the lack Artificial Intelligence Laboratory, the Andrew and
Erna Viterbi Professor of Electrical Engineering
of guarantee that an input will be processed
and Computer Science, and a member of the MIT
and classified correctly. The complex nature of Task Force on the Work of the Future. Courtesy:
training and using neural networks leads to sys- Massachusetts Institute of Technology

tems that are difficult for people to understand.


The systems are not able to provide explanations for how they reached decisions.

Q: What are the ways AI is complementing, or could complement, human


work?

Malone: Today’s AI programs have only specialized intelligence; they’re only capable
of doing certain specialized tasks. But humans have a kind of general intelligence that
lets them do a much broader range of things.

44
How artificial intelligence will change work

That means some of the best ways for AI systems to complement human work is to do spe-  Back to TOC
cialized tasks that computers can do better, faster, or more cheaply than people can. For
example, AI systems can be helpful by doing tasks such as interpreting medical X-rays, evalu-
ating the risk of fraud in a credit card charge, or generating unusual new product designs.

And humans can use their social skills, common sense, and other kinds of general
intelligence to do things computers can’t do well. For instance, people can provide
emotional support to patients diagnosed with cancer. They can decide when to believe
customer explanations for unusual credit card transactions, and they can reject new
product designs that customers would probably never want.

In other words, many of the most important uses of computers in the future won’t be
replacing people; they’ll be working with people in human-computer groups – “super-
minds” – that can do things better than either people or computers alone could do.

The possibilities here go far beyond what people usually think of when they hear a
phrase like “humans in the loop,” Instead of AI technologies just being tools to aug-
ment individual humans, we believe that many of their most important uses will occur
in the context of groups of humans – often connected by the internet. So we should
move from thinking about humans in the loop to computers in the group.

Q: What are some of your recommendations for education, business, and


government regarding policies to help smooth the transition of AI tech-
nology adoption?

Rus: In our report, we highlight four types of actions that can reduce the pain associat-
ed with job transitions: education and training, matching jobs to job seekers, creating 45
How artificial intelligence will change work

new jobs, and providing counseling and financial support to people as they transition  Back to TOC
from old to new jobs. Importantly, we will need partnership among a broad range of
institutions to get this work done.

Malone: We expect that – as with all previous labor-saving technologies – AI will


eventually lead to the creation of more new jobs than it eliminates. But we see many
opportunities for different parts of society to help smooth this transition, especially for
the individuals whose old jobs are disrupted and who cannot easily find new ones.

For example, we believe that businesses should focus on applying AI in ways that don’t
just replace people but that create new jobs by providing novel kinds of products and ser-
vices. We recommend that all schools include computer literacy and computational think-
ing in their curricula, and we believe that community colleges should offer more reskilling
and online micro-degree programs, often including apprenticeships at local employers.

We think that current worker organizations (such as labor unions and professional as-
sociations) or new ones (perhaps called “guilds”) should expand their roles to provide
benefits previously tied to formal employment (such as insurance and pensions, career
development, social connections, a sense of identity, and income security).

And we believe that governments should increase their investments in education and
reskilling programs to make the American workforce once again the best-educated in
the world. And they should reshape the legal and regulatory framework that governs
work to encourage creating more new jobs.

MIT Task Force on the Work of the Future


46
AI & Machine
Learning
Thank you for visiting the AI & Machine Learning eBook!
Content Archive

2021 Winter Edition If you have any questions or feedback about the contents
in this eBook, please contact CFE Media at
2021 Fall Edition
customerservice@cfemedia.com
2021 Summer Edition

2021 Spring Edition We would love to hear from you!

You might also like