Original subtitles

hi my name is David Boyle and I'm a

technical specialist with IBM Watson IOT

I work across all the industries with

the focus of helping clients better

manage the performance of their assets

in this video I'll be discussing how the

equipment maintenance assistance

solution is used to capture the

knowledge of an ageing workforce and how

its underlying AI is used to help field

technicians increase their productivity

please note that throughout this video

I'll be referring to equipment

maintenance assistant as EMA for short

now before I dive into this solution

itself I think it's important to

understand the motivations behind it at

a high level companies with capital

equipment have always been challenged

with lowering operating cost and

increasing production it used to be that

the preventative maintenance approach of

replacing parts every X number of run

hours was the only way to address these

challenges

eventually SCADA systems PLC's and IOT

analytics allowed operators to become

more proactive by monitoring the

condition of their assets in real time

or by predicting failures and extending

maintenance intervals these insights

have proven to be valuable in

understanding the performance of their

assets but they are not resolutions no

matter the approach a field technician

still needs to go out and service that

equipment that being said the key to

actually lowering operating costs and

actually increasing production there's a

combination of insights and effective

field technicians and so in order for

IBM to have developed EMA and provide

complete end-to-end maintenance

solutions we had to first understand the

challenges of the field technician from

that perspective we see that there's a

continued increase in the complexity of

machinery and the systems they make up

at the same time the availability of

equipment expertise is shrinking as a

result of an aging workforce so when you

couple these two challenges it causes

technicians to have to spend more time

in the field and revisit equipment more

than once which in other words means an

increase in mean time to repair and a

decrease in first-time fix rate these

metrics are associated with operating

cost and production levels and the way

we're addressing them directly is with

EMA

so what is EMA and how does one use it

simply put ei

EMA is an AI digital assistant that

field technicians can use at the point

of work to surface diagnosis and repair

recommendations from a corpus of tribal

knowledge EMA's ability to transfer this

knowledge from subject matter experts to

less experienced field technicians is

what allows them to make the best

repairs the first time and in a timely

manner thus addressing first-time

fixed-rate and mean time to prepare now

in order to get started with EMA the

knowledge base that bill refer to must

be prepared this is typically done by

subject matter expert and can include

both structured and unstructured

documentation such as OAM manuals repair

guides historical work orders online

forums and diagnosis models to name a

few from there the underlying watson

algorithms will automatically identify

and enrich the content found within the

uploaded documentation to help us search

be in accuracy once this process is

complete the AI will need to be trained

on relevancy that is what are the

questions being asked what are the

responses and how should they be

prioritized this second step in the

preparation phase is typically performed

by the same subject matter expert and

involves a simple iterative process of

asking questions and rating responses

until a high degree of confidence is

achieved once the SME is comfortable

with the results EMA is ready to be

deployed so now when work orders are

issued from an enterprise asset

management system such as Maximo or sa

ppm the less experienced field

technician can use that context to

prepare recommended parts tools and

safety equipment before heading out to

the site once at the site the same field

technician can leverage a diagnosis

model to help determine root cause or

use EMA's query capabilities to gather

step-by-step procedures you need to make

a quick and accurate fix then when they

go to close work orders the field

technician can also provide feedback to

help the underlying AI continuously

improve and then add that to the growing

knowledge

future technicians can take advantage of

it and the demonstration I'm about to

give will look at how an SME would

prepare EMA as well as how an automotive

technician would use EMA to determine

why a car won't start and how it should

be repaired now as I just stated the

desired outcome of using EMA is to

transfer knowledge from subject matter

experts to less experienced field

technicians I also mentioned that the

first step in doing so is to build out

the knowledge base from which they will

extract those repair recommendations

that being said we'll first take on the

role of a subject matter expert or an

SME and look at how this process is done

using EMA's document manager and the

diagnosis model manager in the document

manager the SME is able to create

various collections of documents as they

relate to specific assets so as you see

in the shared environment there's a

number of different assets with their

own distinctive collections but in our

case we want to create a collection

that's related to a car so to do that we

would hit the create button we would

give a name to the collection in this

case my car we would define the

collection type which in this case we'll

keep as documents and then we'll need to

associate this collection with one in

Watson discovery and this is what the

Watson discovery interface looks like

and as you can see there are collections

similar to what we saw in the document

manager so what you would do here is hit

the upload and essentially replicate

this collection in Watson discovery and

then you would come back here select it

from the drop-down and continue and this

is arguably the most important step

because without this the SME would not

be able to understand their

documentation and also they would not be

able to perform relevancy training on

that documentation now normally I would

hit create at this point but I've

actually gone ahead and created two

collections for the car so scroll down

take a look at the first one called my

car this collection contains various PDF

documents containing either OAM manuals

repair guides in any other

troubleshooting procedures if we come

back out of the document manager we can

look at the second collection titled my

car work orders which as you can imagine

contains historical work orders that the

field technicians can use as a reference

point to see what other technicians have

done on these repairs in the past now

both of these collections are very

useful in situations where the

technician has some sort of context

related to the repair that they have to

make and that may come from work orders

or from predictive maintenance systems

or even condition based maintenance

systems but there are a lot of times

where the field technician must also

perform a diagnosis to determine the

root cause before they understand the

types of questions that they need to ask

so another way in which the SME can

build out the knowledge base is by

creating a diagnosis model and as you

can see I've already created one for the

car so we'll go ahead and open that one

up here the SME has a canvas where they

can drop caused nodes and symptom nodes

and then define them and link them

together using probabilities so from the

technician perspective the technician is

able to go through a list of symptoms

click on the ones that they hear see

smell and so on and then run a diagnosis

to determine the most likely cause now

as far as preparing the knowledgebase

goes the SME has completed their job so

the next step in this phase as I had

shown earlier in that diagram was that

we need to now train the knowledgebase

and since we're already in the diagnosis

model manager we'll go ahead and look at

how we can train the diagnosis model

then after that

we'll go into Watson discovery to see

how we would do the same with the

documents we uploaded earlier so if the

SME were to open up the data tab they

would see three different methods for

training the diagnosis model the first

one is to map historical work orders to

instances where symptoms occurred and

causes were found another way in which

they could train this is through

feedback and right now we don't have any

feedback in the system but when we look

at this from the field technician

perspective we'll see how the field

technician uses EMA's interface to send

feedback through to the diagnosis model

now the third way of training a

diagnosis model is to create predefined

examples so this is arguably the

quickest way in which we can train the

diagnosis model to understand real-life

situations where symptoms were found and

causes occurred and so what happens is

as you start to use these various

methods of training the diagnosis model

the AI will start to dwarf the

probabilistic nature of the solution and

then override it with more realistic

examples of again where symptoms

occurred and causes were found so this

includes the training portion for the

diagnosis model manager so to get into

the process of training the documents

that we uploaded earlier we'll have to

go into Watson discovery now the first

thing I want to show you is how we

improve the search speed and efficiency

of Watson discovery by using what is

known as smart document understanding so

to do that I will use a sample

collection called SDU test to

demonstrate this as I've already done

some training to the my card collection

so once we open this up we will see

details around this collection we will

see some of the enrichments that were

added in the enrichments are is

basically data that has been added to

the existing content to help with the

search process so we see some

enrichments that have been added already

as far as entities go concepts sentiment

analysis and categories so this was done

automatically but what we can do through

smart document understanding is we can

configure this data to and to include

new and better enrichments so as you can

see through Watson discovery we get a

view of the document and what Watson has

come back as or come back with as

identified fields so we see initially

that everything is labeled as it is

what's known as a text entity and so the

text entity alone is not very efficient

in terms of search speed in accuracy so

one thing that the SME can do to improve

this is to use field labels to identify

different objects within their document

so if they drag over things such as the

title and drag over things such as these

subtitles they can begin to teach Watson

to identify these objects on their own

so as they upload new documents to a

collection these documents will be

annotated automatically and the subject

matter expert will no longer need to

provide these sort of annotations and

here in a second we'll talk about why

it's important to do these annotations

but I want to continue annotating for

right now because I want to show you the

real-life machine war machine learning

happening right here in front of us so

I'll go ahead and finish up this page

here

with some annotations and then as you

see here Watson has started to

incorporate machine learning and

understand these objects within within

the documents without me even having to

tell it to do so and to continue

training it a little further we'll go

ahead and highlight this footer here

submit this page and again we can see

that Watson has now picked up the footer

of this document and has correctly

identified that there are no subtitles

found within here now the reason it's

important to do this smart document

understanding is that we can then manage

these fields afterwards so for the

subject matter expert and their

documents they know that there is not no

relevant information or no information

of value to the technicians and things

like the answer the author the footer

the header may be the question and table

of contents for instance so now we're

filtering out what gets returned to the

field technician so if it's not a value

we don't want to give it to the

technician also what we can do is split

up the document based on one of those

fields so if I choose subtitle for

instance this original document will get

split into as many new documents as

there are subtitles with the subtitle

being the subject line of each new

document and what this does is this

provides granularity in the search

process so we have new documents that

Watson can then identify very quickly in

return answers in what I call answer

units to the field technician in a

shorter period of time and to extend the

enrichments even further we can add in

the entities such as the subtitle and

choose which enrichments we want to

include on each of those fields

now this process of the smart document

understanding is very useful in terms of

search speed and efficiency but it

doesn't necessarily address the concern

of accuracy and to do that we need to

look at how we perform relevancy

training on a collection so we'll first

open up the my car collection and

there's one thing that I want to point

out so I already did training on this

collection and what we saw earlier was

that I uploaded six documents but

because I've annotated these they have

now been broken into thirty four

distinct documents that Watson can then

take advantage of to search with with

better granularity but to do the

relevancy training we would click on the

performance tab here select view all and

perform relevancy training and then

choose the collection that we want to

train and I talked about this briefly

earlier the process of training a

collection on relevancy is to ask it

questions that a common field technician

would ask so for instance we can look at

one of these questions I asked earlier

where I asked what tools do I need to

remove a battery and what the SME can

then do is validate these results by

clicking rate results to then see the

smaller documents we talked about

earlier and rate them as relevant or not

relevant and what that does is it

teaches Watson what the question is how

it's supposed to be responded to and

then how to prioritize these documents

and we'll look at this again in the end

user interface to see what it looks like

from the field technician perspective

and we'll do that here shortly because

we are now finished with the process of

training the knowledgebase so to look at

this from the field technician

perspective we will jump to the sample

application now for this let's imagine

that an automotive technician just had a

car towed to the shop because the

customer could not get it to start on

their way to work that

this particular customer is of value to

the shop and has requested to pick it up

on their way back from the office that

same day that being said the technician

now has to repair the vehicle in a

timely manner analysts also get it done

correctly as to avoid having it back in

the shop later that week so the first

thing that the technician is going to

want to do is they're going to want to

determine the root cause of this failure

so here they can select the model that

we looked at earlier related to the car

and then what the technician can then do

is assess what are the things that I

hear smell see hopefully not taste and

so on and what they see is that the

lights are turning on but there is no

sound coming from the car they could

then submit these symptoms and the

diagnosis model will run in the

background to determine the most likely

cause and if you click on the most

likely cause the technician can then use

the support documentation we uploaded

earlier to find the best recommendation

for making this repair and because we're

working in a shared environment I have

to filter by the asset that I want to

focus on but what comes back to me is

one of those answer units I had

mentioned earlier which is a snippet of

the larger documents we uploaded in the

first step and I also get a confidence

level that's relatively low because of

the amount of training that's been

incorporated into this but another way

the technician can add feedback is

through the simple thumbs-up or

thumbs-down and that's one way to train

it before the technician moves on with

actually making the repair it is

suggested to also provide feedback in a

more detailed format and this is very

important because this is something that

we can incorporate into the

knowledgebase to let future technicians

use if they encounter similar problems

now we're gonna go look at this from the

perspective of a field technician who

now has the context they need from a

diagnose

is to then ask some questions along the

way so from the query a sample

application within EMA our technician

might find himself stuck on some parts

or have concerns about certain steps

within the process of repairing a car

that won't start so for instance we can

say or ask the question of how do I

safely disconnect the jumper cables hit

entered and EMA will use that that

context to then go out in search for the

best recommendation so if we again

filter by the collection of interest by

the asset of interest we can see various

recommendations based off of our

training based off of the SMEs training

and the technician can then use these

answer units to help troubleshoot the

car so what we're doing overall with the

EMA solution is we're streamlining the

process that the technician goes through

in searching for information that they

can use to make a repair and in doing so

we're reducing the mean time to repair

or rather reducing the sorry yeah

reducing the mean time to repair an

increase in the first time fix rate on

this alone so I want to come back to the

graph we looked at earlier because we

see that machine complexity has remained

the same and it likely always will if

not increased but what has changed was

the equipment expertise so we've closed

the gap between expertise and the

complexity of machinery and as you saw

in this demonstration the automotive

technician was given the required

knowledge to make the right fix the

first time in a shorter period of time

so this means that the potential for a

decrease in the operating cost and an

increase in earnings has just been made

available simply by focusing on

improving the productivity of fields

if you think this solution is something

of interest to you I recommend reaching

out to either myself or Hina purohit who

is our offering manager to explore the

solution further for client references

please see the links in the comments

below thank you for watching this video

I hope you enjoyed it

[Music]