Analyze Existing Sensor Data to Detect Abnormal Equipment Behavior with Amazon Lookout for Equipment
Analyze Existing Sensor Data to Detect Abnormal Equipment Behavior with Amazon Lookout for Equipment
Successfully implementing predictive maintenance requires using the specific data collected from all of your machine sensors, under your unique operating conditions, and then applying machine learning (ML) to enable highly accurate predictions. However, implementing an ML solution for your equipment can be difficult and time-consuming. In this tech talk, we will introduce you to Amazon Lookout for Equipment, which allows you to analyze the data from the sensors on your equipment to automatically train a machine learning model based on your equipment data – with no machine learning experience required. Lookout for Equipment uses your unique ML model to analyze incoming sensor data in real-time and accurately identify early warning signs that could lead to machine failures. This means you can detect equipment abnormalities with speed and precision, quickly diagnose issues, take action to reduce expensive downtime, and reduce false alerts.
Learning Objectives:
-Learn how Lookout for Equipment handles data from up to 300 sensors in one ML model, along with historical logs, to build a custom ML model and give you accurate alerts when your equipment behaves abnormally
-Learn how you can use data from Lookout for Equipment to set up automatic actions to be taken when anomalies are detected, such as filing a trouble ticket or sending an automatic alarm
-Learn how you can use data from Lookout for Equipment to set up automatic actions to be taken when anomalies are detected, such as filing a trouble ticket or sending an automatic alarm
To learn more about the services featured in this talk, please visit: https://aws.amazon.com/lookout-for-eq… Subscribe to AWS Online Tech Talks On AWS:
https://www.youtube.com/@AWSOnlineTec…
Follow Amazon Web Services:
Official Website: https://aws.amazon.com/what-is-aws
Twitch: https://twitch.tv/aws
Twitter: https://twitter.com/awsdevelopers
Facebook: https://facebook.com/amazonwebservices
Instagram: https://instagram.com/amazonwebservices
☁️ AWS Online Tech Talks cover a wide range of topics and expertise levels through technical deep dives, demos, customer examples, and live Q\u0026A with AWS experts. Builders can choose from bite-sized 15-minute sessions, insightful fireside chats, immersive virtual workshops, interactive office hours, or watch on-demand tech talks at your own pace. Join us to fuel your learning journey with AWS.
#AWS
Content
1.04 -> hello everyone and welcome today we're
2.879 -> going to be going through
3.84 -> amazon lookout for equipment detecting
6.319 -> abnormal equipment behavior by analyzing
8.72 -> industrial sensor data my name is brent
11.04 -> swidler i'm the product manager for
13.2 -> lookout for equipment so the objective
16.72 -> of
16.96 -> amazon lookout for equipment is to be
18.72 -> able to model
20.08 -> the behavior of industrial assets uh in
23.199 -> order to enable
24.32 -> predictive maintenance and i'm going to
26.08 -> be going through a lot more in
27.92 -> in depth on how the service functions
29.599 -> and and where it's applicable
33.36 -> just quickly i want to go through where
36.16 -> lookout for equipment
37.36 -> stands in the aws ml stack
40.719 -> so we have three layers to the aws ml
44.239 -> stack there's the infrastructure layer
46 -> which
46.559 -> includes tensorflow pytorch mxnet
50.399 -> the middle layer which is amazon sage
52.96 -> maker
54.16 -> which allows you to build your own
56.079 -> custom models
57.76 -> and then the top layer is managed
59.76 -> service managed ai services
62 -> there's a variety of different managed
63.359 -> ai services and over here on the right
65.68 -> we see the
66.72 -> industrial bucket of services including
69.6 -> panorama
70.799 -> amazon monotron look out for equipment
73.36 -> and look out for vision
77.119 -> so over the next hour or so i'm going to
79.52 -> be going through
81.36 -> lookout for equipment and the challenges
83.36 -> that it addresses in the
85.119 -> equipment monitoring and predictive
86.72 -> maintenance space
88.799 -> how the service works and some service
91.52 -> overview
92.72 -> i'm going to go through a demo of how
94.72 -> you can access and use
97.2 -> lookout for equipment some common use
99.6 -> cases and customer references
102 -> and then some getting started guy
104.399 -> getting started resources and where you
106 -> can access all of the information you
107.52 -> need to you know
108.32 -> in order to get started and then we're
110.079 -> going to end with
111.439 -> some q a
115.759 -> so in the maintenance space uh there's
118.399 -> been an
118.799 -> evolution of maintenance applications
120.399 -> over the past number of years
122.32 -> where the idea is to be able to move
125.6 -> from
126 -> reactive maintenance where you're
128.16 -> waiting until something breaks in order
129.599 -> to fix it
130.959 -> until you can move all the way over to
132.56 -> prescriptive maintenance and that means
134.16 -> you know first going through reactive
135.84 -> and then preventive
138 -> which is you know what you do with your
139.68 -> car changing its oil every six months or
141.76 -> however many miles
143.36 -> and so you maintain it on regular
144.72 -> intervals and most of our focus today is
147.12 -> going to be on this condition based
148.48 -> slash predictive maintenance
150.239 -> which most commonly uh uses you know
153.599 -> single variable alarms some statistical
155.92 -> analysis
157.04 -> physics-based modeling methods and then
159.36 -> also what we'll go through in more
160.879 -> detail is
161.599 -> ml based approaches so the objective is
165.36 -> to move from a reactive all the way to
167.12 -> prescriptive so condition based and
168.879 -> predictive can tell you
170 -> exactly how your assets performing now
172.239 -> and then prescriptive in the future is
174.56 -> giving you a lot more details into
176.16 -> exactly what you should be
178.56 -> doing to diagnose a specific issue
182.48 -> and in reality a lot of different uh
185.28 -> facilities and plants and operations
187.36 -> have a mixture of all of these
189.44 -> uh depending on the criticality of the
191.36 -> asset its operation
193.12 -> um it depends on what level of
195.36 -> investment warrants
196.48 -> uh moving towards um moving you know
199.68 -> further down this
200.8 -> evolutionary chain what we've seen
204.48 -> is a lot of companies are very heavily
206.48 -> moving towards uh
207.92 -> condition based and predictive
209.2 -> maintenance and closer to prescriptive
211.519 -> um
224.959 -> in in maintenance technologies will uh
227.599 -> save
228.08 -> an estimated 630 billion dollars across
230.72 -> a variety of industries
232.64 -> uh some of this is from you know failure
235.2 -> issues
236 -> uh i'm assuming a lot of it is from uh
238.799 -> avoiding
239.519 -> uh downtime and then there's uh other
243.04 -> you know optimizations as well as uh
246.319 -> quality assurance and all of those sorts
247.92 -> of things that add up in aggregate to
249.519 -> about 630 billion
252.48 -> by 2025. now because
255.84 -> this is such a big number um we've
258.959 -> you know there's been a lot of
260.32 -> investment rolling into equipment health
262.639 -> monitoring
264.08 -> which really involves putting a bunch of
266.32 -> sensors all over your equipment
268.96 -> investing in data storage so all of
270.639 -> those sensors take readings
272.08 -> all the time and deposit data into um
275.36 -> into what's most often a data historian
278.72 -> connectivity which is either you know
280.8 -> remote facilities connectivity
282.639 -> or connectivity to the cloud
286 -> analytics which is what we're going to
287.44 -> go into a lot more detail in today
289.44 -> and then also visualizations and the
291.68 -> ability to actually give
292.96 -> a maintainer uh some you know
296 -> visual aspects to actually help them
297.759 -> diagnose and take corrective action
301.039 -> so a lot of investments been rolling
302.56 -> into the space across this entire chain
304.56 -> of applications sensors data storage
306.72 -> connectivity
307.6 -> analytics and visualization but in
310.32 -> reality what that really means is that
312.24 -> companies are really drowning in data
313.84 -> trying to figure out how to get value
315.12 -> from it put a lot of sensors on a lot of
316.88 -> equipment
317.759 -> um you know they're all valuable and
320.88 -> they're all giving you meaningful
322 -> information but at a certain point it's
323.52 -> too much
324.8 -> and from that it's estimated that only
328.24 -> about five percent of the actual data
330 -> that's being generated from all of these
332.24 -> um all of these sensors is actually
334.639 -> being used
337.44 -> and the reason for this uh is because
340.16 -> there's a lot of information to be
341.68 -> parsed through
343.039 -> um for any given asset you can have uh
346.32 -> you know 20 30 50 up to 300 i use tags
350 -> and sensors interchangeably
352.32 -> you can have you know a wide variety of
354.56 -> different time series sensors coming off
356.24 -> of an
356.72 -> asset and so what you're left with is is
359.759 -> a you know a variety of data sets that
361.919 -> look like this
363.039 -> that you have to sort of sift through to
364.72 -> figure out what is actually a pattern of
366.479 -> behavior that you really care about
369.28 -> and how do you spot uh how do you spot
371.039 -> the problem in
372.319 -> you know when you're not only talking
373.68 -> about three you know 300
376.24 -> sensors on a specific asset but then
377.919 -> you're talking about also having
379.6 -> 50 assets in a plant and five plants and
381.759 -> all you know as it starts to grow
383.919 -> the complications of the complexity of
386.16 -> the problem starts to scale as well
392.4 -> so the current analytic methods um do
395.36 -> have some limitations in this but they
397.039 -> all
397.28 -> they also do have some positive aspects
399.6 -> of how they address some of these
400.8 -> problems
401.68 -> uh the most simplistic and easy to do is
404 -> a single variable analysis where you say
406.56 -> uh you know you can set a threshold
408.08 -> saying my temperature is too high or my
409.599 -> temperature is too low
411.44 -> and that can start to give you alerts
413.36 -> and start to make sense of some of this
414.88 -> data
416.08 -> however you do end up with a situation
417.84 -> where you have a lot of alarms
419.84 -> and often times you're you're reliant on
422.88 -> you know one specific value to give you
425.12 -> a lot of predictive capabilities in
426.8 -> what's actually happening in an asset
429.919 -> uh the next phase in analytics methods
432.24 -> is really physics-based modeling where
434.479 -> the engineer most often who's designed
436.639 -> this equipment can sit down and actually
438.4 -> type out
439.84 -> or simulate a physic physical equation
442.479 -> to denote how that equipment could
444.16 -> operate
446.319 -> and then as we move from single variable
449.12 -> to
449.599 -> physics based the third step here is
452.24 -> supervised machine learning applications
454.4 -> and i call out supervised machine
455.919 -> learning applications specifically
457.599 -> because
458.319 -> this is really um oftentimes you know
461.599 -> we'll you'll hear the promise of
463.44 -> predictive maintenance
465.039 -> um the idea with this type of
467.199 -> application is that you would expect
469.36 -> that you would say oh i have a bunch of
470.8 -> failures now i can use those as labels
473.44 -> and predict exactly what's happening you
475.599 -> know here's
476.639 -> 50 examples of this asset failing can i
479.44 -> now
479.84 -> use those as labels to predict when that
481.68 -> asset's going to fail
483.68 -> and while it sounds logical in in this
486.639 -> method
487.52 -> it really is not as lo it's not as easy
490.879 -> in practice
492 -> as this requires a lot of examples of a
494.879 -> specific issue
496.96 -> it requires a lot of historical data
499.68 -> sets and it requires a lot of
502.16 -> subject matter experts time to go
505.12 -> through and actually
506.56 -> make sense of the historical aspects of
509.039 -> what has happened in
510.08 -> you know the historical maintenance logs
512.32 -> and put them in a format and structure
514.399 -> that can actually be used
515.76 -> by ml techniques
518.959 -> and what we've been hearing um you know
521.12 -> across the customers that we work with
523.2 -> is that
523.919 -> um smes really feel like a lot of the ml
527.2 -> applications are
528.56 -> are overly consuming of their time and
530.8 -> effort
531.6 -> um and you know we have a quote here
533.519 -> from you know one that we've heard from
535.12 -> is that the climb is not worth the view
537.519 -> uh it just seems like it's way too much
539.36 -> upfront effort to get this going
542.48 -> so um there are some analytics methods
544.88 -> that can help sift through some of this
546.16 -> data
546.64 -> uh they do have some value to them of
549.12 -> course because they've been
550.16 -> you know the single variable and physics
551.92 -> based approaches have been being
553.2 -> used for a while but oftentimes there's
556.48 -> not a clear way to get like the very
558.56 -> direct insights at the speed that's
560.08 -> required
561.6 -> and the complication really is that
564.72 -> industrial assets are difficult there's
567.68 -> a lot of variability and there's a lot
569.92 -> of
570.88 -> complexity to this to predictive
573.279 -> maintenance applications
576.08 -> we specifically use unsupervised
579.68 -> techniques to try and overcome these
581.36 -> problems
582.32 -> but the overall you know the problem
584 -> space can be whittled down to three main
585.76 -> issues
587.44 -> and the reason why analytics on
589.12 -> industrial equipment is really
590.56 -> challenging
591.2 -> one is that every asset is unique so if
593.68 -> you have two
594.48 -> pumps right next to each other even if
596.16 -> you're lucky enough that they have the
597.519 -> exact same
599.44 -> types of tags sensors on them
602.72 -> it's often the case that there's rarely
604.88 -> two things that operate exactly the same
606.959 -> they might have different maintenance
608.079 -> histories different model numbers
609.92 -> different parts components you know
612.56 -> and then when you look at you know not
613.92 -> just pump pump a next to pump b
616.72 -> but also um you know a whole plant
619.68 -> versus
620.399 -> uh you know another plant maybe a wind
622.64 -> farm in alaska versus a wind farm in
625.68 -> central america and all of these things
627.6 -> start to have so much
628.72 -> variability in their the way that their
631.92 -> data is that
633.279 -> every asset is generally uh needs to be
636 -> modeled
636.959 -> independently and in a unique fashion
640.8 -> the other issue is that downtime events
642.88 -> are rare so
643.839 -> if you heard me say this with the
644.959 -> supervised applications um there's very
648 -> very rarely a case where you have 50
650.56 -> examples of the exact same failure
652.399 -> occurring
653.279 -> um in in you know some level of history
656.16 -> of the asset
658.839 -> um and so if downtime events are rare
661.839 -> it's really hard to structure one of
663.36 -> those supervised applications because
665.6 -> you only really have a small sample size
668.079 -> of the
668.64 -> the types of issues that you're looking
670.16 -> for
672.48 -> and lastly implementation is difficult
674.72 -> so uh
675.68 -> you know moving data from uh on-premise
679.68 -> uh you know the sensors to a historian
682.8 -> to the cloud
683.76 -> to you know running inferencing all of
685.76 -> this requires a non-trivial amount of
688 -> software engineering effort
690.8 -> so here we have every asset is unique
693.12 -> downtime events are rare and
694.56 -> implementation is difficult
696.48 -> so what we've seen is that we need a
698.16 -> solution that can do a number of things
700.32 -> we need a solution that can customize to
702.32 -> the number of sensors the unique data
704.64 -> and the operating conditions of each
705.839 -> asset
707.519 -> meaning that we build a custom model for
709.519 -> each individual asset regardless of
712 -> the changing inputs that go from asset a
715.04 -> to asset b
717.2 -> um we the solution that does not require
719.44 -> the luxury of having many labeled
720.959 -> examples of failure so i mentioned that
722.56 -> we use a lot of unsupervised ml
724.48 -> techniques here
725.6 -> um this allows us to get around the fact
728.079 -> that there's
728.839 -> um there's often times a lack
731.76 -> of labeled applications and labeled
734 -> examples of issues that have occurred in
735.6 -> the past
737.68 -> uh we need a solution that can be
739.04 -> adjusted over the lifespan of an asset
741.36 -> so
741.92 -> another wrinkle in this complication of
743.92 -> this of you know industrial equipment
745.76 -> monitoring is that
747.12 -> um as an asset changes over its lifespan
751.36 -> so does its data set in you know its
754.72 -> definition of behavior of normal
756.48 -> behavior
757.6 -> so for instance like i always go to
760.24 -> examples with my car
762.079 -> so a car with zero miles on it will have
764.48 -> a different pattern of behavior than a
766.399 -> car with a hundred thousand miles on it
768.88 -> both of those data patterns could be
771.04 -> considered normal but they are
772.56 -> different and this pattern of behavior
775.04 -> needs to be tracked over its life cycle
778.48 -> we also need a solution that can be
780 -> scaled across a large number of assets
782.24 -> and asset types
783.279 -> so we're not built you know here we're
786 -> not looking at
786.88 -> you know uh look out for equipment being
789.04 -> something that's like
789.92 -> specifically for this type of pump in
791.68 -> this type of application
793.36 -> we're we're building for the ability to
795.92 -> ingest data from a wide variety of
797.68 -> assets in a number of different
798.959 -> industries
800.959 -> and lastly we feel that this needs to be
803.76 -> an
804.32 -> automated solution so someone can use it
806.56 -> without any ml experience
808.32 -> so we can ask a user a set of questions
810.399 -> on their data set
811.519 -> and given their answers to those
813.36 -> specific questions
814.639 -> um we think we can build an automated ml
817.68 -> that
818 -> uh lookout for equipment does build an
819.839 -> automated ml
821.6 -> model that can then be used for
823.12 -> real-time inferencing
826.32 -> so with those things in mind we designed
829.279 -> look out for equipment
831.199 -> and lookout for equipment is an
832.639 -> automated machine learning service
834.32 -> that uses your sensor data to detect
837.199 -> abnormal equipment behavior
838.959 -> so in that first you know mess of a
841.04 -> bunch of time series that we saw
843.519 -> instead of you know now you can take all
845.6 -> of that data
846.8 -> bring it all in to look out for
848.16 -> equipment and we will use
850.32 -> our approach towards modeling the
852.16 -> behavior of that equipment
853.6 -> and we can tell you exactly what's
855.279 -> normal and when normal behavior is
857.519 -> happening in that asset
858.8 -> and alert to any abnormality in the
860.72 -> behavior of a specific asset
863.04 -> so over here on the right hand over here
864.959 -> on the left you see all this mess of
866.56 -> data going into look out for equipment
868.959 -> we model out the behavior and at the
871.12 -> output you'd see
872.079 -> you know everything is normal so
873.76 -> nothing's wrong you don't have to do
874.959 -> anything
875.76 -> to hey we're seeing an abnormality and
878.48 -> most of this is
879.44 -> since you know the rank ordered list of
882 -> exactly what sensor is telling us what's
883.92 -> happening here
886.959 -> in a little bit more detail here is the
889.519 -> generalized workflow
890.959 -> of how this should work so we start off
893.68 -> with historical data
895.199 -> in what's most often um either a
899.68 -> data lake or in a historian and then we
903.04 -> also have
904.16 -> maintenance records um and then those
906.8 -> maintenance records as well as the
908.079 -> historic historical data and i'll get
909.92 -> into what the formats are for this
911.519 -> when i go through the demo are then
914.16 -> ingested
915.04 -> directly into amazon s3
919.6 -> and look out for equipment set of apis
921.92 -> just read
923.04 -> and write directly from s3 so if you can
926.32 -> get your data into s3 you can use amazon
928.72 -> lookout for equipment
931.12 -> so all you have to do is ingest the
933.12 -> historical data
934.16 -> into look out for equipment select the
937.44 -> sensors that you want to use for a model
940.079 -> click train model and we automatically
942.16 -> train
942.72 -> uh the specific model to that asset i
945.6 -> give you the ability to evaluate the
947.199 -> performance with explainability
949.36 -> and then set up that model for real-time
951.6 -> inferencing in the cloud
954.56 -> and again i'm going to go through a demo
956.32 -> so i'm going to show you this entire
957.6 -> workflow
959.839 -> and how it works inside of the console
963.759 -> so just a little bit of some information
966 -> from under the hood
967.279 -> a typical ml workflow for
970.72 -> an application like this if we were
973.68 -> trying to do this
974.56 -> in a manual data science approach we
976.8 -> would have to do a bunch of data
978.16 -> cleaning steps
979.519 -> figuring out the right sensors aligning
981.839 -> different timestamps and puting values
984.32 -> then you'd actually get to the algorithm
986 -> selection and the hyperparameters
988.32 -> figuring out how to like create the
990.16 -> thresholds and score the
991.839 -> models and integrate the historical
994.16 -> failure labels
996 -> give out the feature rankings all of
998.24 -> this sort of stuff
999.92 -> is like step-by-step process and for
1002.48 -> anyone who's
1003.199 -> you know done a lot of data science work
1005.199 -> you could probably understand that every
1006.72 -> single one of these boxes is still high
1008.56 -> level
1009.279 -> and there's a number of steps underneath
1011.04 -> that also need to be thought through
1013.92 -> so we took it upon ourselves to automate
1016.48 -> a lot of this away
1017.519 -> so here now you can just take your input
1019.6 -> data all of those sensors
1022.079 -> dump it into lookout for equipment and
1024.24 -> you get a
1025.52 -> model at the end that can be set up for
1027.28 -> real-time inferencing
1028.799 -> and all of this uh all of the
1030.4 -> complications of all of these steps
1032.48 -> are just essentially uh cloud computer
1036.24 -> time
1036.64 -> under the hood without user needing to
1039.36 -> know any of them
1043.839 -> so with that i'm going to jump into a
1046.24 -> demo
1046.959 -> i'm going to go through a demo on an
1049.2 -> industrial pump
1050.96 -> the data set that i'm going to go
1052.24 -> through is publicly available and i'll
1054.24 -> show you where
1055.44 -> the resources are for you to access and
1058 -> use it
1059.12 -> so this specific pump is a pump with
1062.799 -> five different components to it we have
1065.2 -> an impeller
1066.08 -> a motor a housing a volute and a shaft
1068.16 -> and each one of these has
1069.919 -> six specific sensors on it they're just
1072.96 -> labeled sensor one two three four five
1074.96 -> all the way up until 30.
1079.039 -> this data set spans from january to
1081.36 -> november of 2019 and it has 30 sensors
1084.08 -> on it
1084.88 -> the sensors all take readings at a
1087.12 -> sample of once per minute
1088.88 -> and over that 11 month period there are
1091.679 -> 10 historical failure events that have
1093.679 -> occurred
1094.72 -> now i don't have a class for every
1096.96 -> single one of those
1098 -> failure events so we're not saying that
1099.6 -> this one's cavitation this one's
1101.039 -> misalignment
1102 -> or anything like that we're just saying
1104.16 -> here is a
1105.2 -> behavior of this equipment that was
1106.88 -> known to be abnormal inside of a given
1108.799 -> time range
1113.36 -> the data formatting looks like this
1118.16 -> let me fill those back in
1121.6 -> okay so we have our just our timestamp
1124.559 -> as the first column
1125.76 -> in the data set and every single other
1128.48 -> column
1128.96 -> is just a sensor in that data set so
1131.919 -> this is the
1133.6 -> file specifically for the um for the
1136.96 -> housing
1138.48 -> but we have our timestamp and every
1139.919 -> single sensor you don't have to format
1142 -> it
1142.559 -> with five cents or six sensors in one
1145.12 -> file you can have one you can have
1146.96 -> three thousand um this is just one way
1149.679 -> to frame the
1150.72 -> frame the problem so we're going to
1153.76 -> ingest this into s3 and then
1155.84 -> once this is ingested into s3 we'll
1158.799 -> we'll use
1159.52 -> uh look out for equipment and we'll
1160.96 -> bring it in to look out for equipment
1162.16 -> for modeling
1166 -> so i've now entered into uh the lookout
1169.039 -> for equipment
1170.48 -> uh console that can be accessed from
1173.6 -> just uh you know looking for lookout for
1175.679 -> equipment in
1176.64 -> the aws console and i've labeled the
1179.36 -> data set uh
1180.4 -> pump example webinar and i'm going to
1183.28 -> ingest
1184 -> a new data set into the service so here
1187.36 -> all i have to do is go find
1188.88 -> that where i've put that csv file
1192.08 -> inside of s3 and i'll do that right now
1195.36 -> so now inside of s3 i've located all of
1197.919 -> the folders that have all of those csv
1200 -> files so the one that i just showed you
1202.32 -> for the pump main housing this is that
1204.88 -> file
1205.36 -> that just has all of those sensors
1206.96 -> affiliated with them in that time series
1208.799 -> format
1210.48 -> and i'm just going to select uh
1214.159 -> the folder that contains all of those
1217.52 -> components inside of them
1221.44 -> the last step is just to create a new
1223.28 -> role
1225.12 -> and click ingest so what's happening
1228.88 -> under the hood here is that
1230.559 -> we are now taking data from the s3
1233.919 -> bucket where
1234.64 -> those csv files live and bringing them
1237.2 -> into
1238 -> look out for equipment that can then be
1239.679 -> used for modeling
1244.08 -> so now that my data has been
1246.64 -> successfully ingested into the service
1248.96 -> we'll continue through this workflow so
1251.12 -> you see the step one
1252.24 -> create data set so we've done that step
1253.84 -> two we've just ingested the data set
1256 -> and now step three i can create a model
1258.4 -> so i'm just gonna go click
1259.84 -> create new model we'll call this
1263.44 -> a pump webinar
1267.36 -> model
1271.039 -> and now you can see that all of the
1274.84 -> sensors
1276.96 -> um all of the sensors from this specific
1280.24 -> from that s3 folder are now ingested
1283.28 -> into the service over here
1285.039 -> so i have my volute my shaft the
1288.96 -> pump main housing the motor uh the
1291.84 -> impeller and all of this so
1293.52 -> if i just wanted to model out a sub
1295.6 -> component of this pump
1296.88 -> let's say the volume i can just click on
1299.36 -> all of the sensors for the volute
1301.76 -> or if i wanted to model out you know any
1304.08 -> other you know
1305.28 -> sub system of the pump i could just
1307.36 -> select those sensors
1308.88 -> in this case we're just going to select
1310.4 -> all of them because
1311.919 -> we want to mod we want to use all of
1313.6 -> these sensors for modeling purposes
1319.36 -> uh the next phase is to identify the
1321.76 -> historical maintenance events
1323.28 -> as labels again this is optional so if
1325.76 -> you don't have historical label events
1328.32 -> we can just go ahead without this step
1330.88 -> but for those of you that do and want to
1332.64 -> know how to structure this
1334.799 -> the historical label event file just
1336.96 -> looks like this
1338.24 -> so we have column a and column b column
1341.039 -> a
1341.52 -> is a start time of an abnormal behavior
1344.08 -> event and column b
1345.44 -> is the end time often times when a
1347.76 -> failure occurs
1349.28 -> the theory here is that inside of this
1352.24 -> window of time
1353.44 -> so this one day window of time there is
1356.159 -> abnormal behavior
1357.44 -> prior to a specific failure occurring
1361.76 -> let's say that i forget to change the
1363.28 -> oil in my engine
1365.679 -> over time what's going to end up
1367.12 -> happening is the
1369.039 -> temperature of my engine is going to be
1370.4 -> increasing as the amount of lubrication
1372.72 -> becomes
1373.679 -> less and less and my engine becomes less
1375.52 -> efficient so this pattern of behavior of
1377.44 -> my car engine
1379.6 -> will start to be abnormal until the car
1383.28 -> engine fails
1384.88 -> so inside of a window of time prior to
1387.52 -> any equipment failure
1389.12 -> there is some behavior pattern if it's
1391.2 -> detectable that should be
1393.2 -> um alert that should be told here so
1396.24 -> here we have our two time stamps
1398.4 -> our start time and an end time of when
1400.72 -> an abnormal behavior event occurred
1402.72 -> in the history of this asset so all we
1406.159 -> have to do for a lookout for equipment
1407.76 -> is just take that file and uh point us
1411.039 -> directly to that file location
1414.32 -> so in this case we just have that file
1416.48 -> that i just showed have the example
1417.919 -> labels
1419.28 -> it is just this labels one file i just
1422 -> have to
1422.4 -> select the folder i click choose
1425.679 -> and create that role
1430.799 -> so now we've ingested our industrial
1433.76 -> data into the service in
1435.279 -> in the form of all of these sensors on
1436.88 -> these components i've given the
1439.12 -> information to those
1440.64 -> ranges of values ranges of time at which
1443.52 -> the equipment
1444.4 -> operated abnormally and then i'm just
1446.72 -> going to
1448.72 -> select a training time range and a
1451.84 -> testing time range
1463.039 -> so inside of this training range from
1465.6 -> january 1st 2019
1467.36 -> to august 2019 we're going to use
1470.559 -> all of the sensor data from all 30 of
1472.88 -> those
1473.679 -> sensors to use this to train our model
1476.799 -> between these times as well as
1478.64 -> using the labels from this range over
1481.919 -> here
1482.32 -> now keep in mind we're doing this as an
1484 -> unsupervised method so the labels here
1486.08 -> are not telling us specific failure
1489.039 -> modes instead they're telling us that
1490.4 -> there's just
1491.279 -> known abnormalities behind the scenes in
1494.32 -> those
1494.72 -> in those ranges and then we're going to
1497.36 -> evaluate the model from august 2nd
1499.76 -> to october 25th
1503.36 -> and so what inside of this range we are
1505.679 -> just taking the
1506.88 -> model that was trained at the top range
1509.12 -> and seeing how it would have performed
1510.88 -> if it was
1512.24 -> running on real-time data evaluating
1514.96 -> this pump
1515.6 -> between august and october
1519.2 -> the last step here is i could down
1520.64 -> sample the data set so in this case
1522.32 -> instead of every one minute i want to
1524.32 -> down sample to five minutes so now i'm
1526.32 -> just taking readings from this every
1527.919 -> five minutes
1529.52 -> and then i can click create
1532.64 -> when i click create all of that
1534.88 -> automated pipeline
1536.48 -> that we went through in the powerpoint
1538.559 -> that shows all the
1539.679 -> cleaning steps the algorithm selection
1541.84 -> all of that is going to happen
1543.12 -> under the hood as we train our model
1547.12 -> so for the purpose of this demo i've
1549.039 -> already done this
1550.4 -> so we don't have to wait for this
1552.08 -> specific model to train
1553.84 -> i'm just going to go back to our models
1556.72 -> here
1558.159 -> and you can see the pump demo webinar
1560.4 -> model
1566.32 -> just center this for everybody if i
1568.48 -> wanted to create a new model i can just
1569.76 -> do it up here
1570.72 -> now if i go to pump demo webinar model
1574.08 -> i get all the information for this model
1576.08 -> it took 26 minutes to train
1578.88 -> it's ready for inferencing and
1582.24 -> from a performance perspective um of the
1585.44 -> labels that were given
1586.64 -> we saw abnormal behavior events
1589.919 -> in eight of those label ranges eight out
1592.24 -> of the eight label ranges
1593.679 -> and the first abnorma the first anomaly
1597.36 -> occurs on average 23 hours and one
1600.72 -> minute before the end of the label range
1602.4 -> now what that really means is that
1604.559 -> on average every time that we're
1606.32 -> detecting an anomaly
1607.919 -> we're getting we're giving a 23 hour and
1610.159 -> one minute for
1611.039 -> warning time for when the issue is going
1614.48 -> to
1615.039 -> when the asset's going to fail or
1616.799 -> however we define that
1618.08 -> specific label there are also
1621.44 -> five abnormal behavior events detected
1623.84 -> outside of the label range
1625.36 -> so these could be thought of as false
1626.88 -> positives we call them abnormal behavior
1629.2 -> events because
1630.64 -> oftentimes there could be something
1632 -> going wrong that might not necessarily
1633.679 -> be
1634 -> a specific failure
1637.2 -> this is the scoring this is how we score
1638.96 -> these models ideally we want to make
1641.12 -> sure that we're alerting to
1643.12 -> anomalies or issues that are occurring
1646.32 -> prior to major
1647.52 -> equipment failures
1650.96 -> um from uh from just from the data set
1653.84 -> uh we used uh
1654.96 -> january to august to train and then
1656.96 -> august to october to test and the sample
1659.36 -> rate was five minutes
1661.52 -> now visually it looks like this we have
1664.32 -> our
1664.64 -> gray area where we have the labels
1667.679 -> of those time ranges those time
1669.919 -> intervals so here's
1672.159 -> august 8th to august 9th there was a
1674.399 -> range of behavior that was known to be
1676.159 -> abnormal
1677.84 -> and you know so on and so forth through
1679.919 -> all these gray areas
1681.44 -> and the top portion that's all clickable
1684.32 -> are the
1685.12 -> uh anomalies that are detected from the
1687.52 -> actual service so if we play this like a
1689.52 -> movie over a timezone over a time range
1692.24 -> you can start to see the alerts that you
1693.919 -> would get and the alerts are
1695.279 -> binary output so it's zero everything's
1697.76 -> normal so that's all this
1699.36 -> blank space over here or one the
1701.84 -> equipment is operating abnormally
1704.08 -> and that's what's happening in these
1705.76 -> brownish red
1708.159 -> anomaly lines here now again this is
1711.279 -> only part of the story so
1712.799 -> alerting to a specific issue is uh
1716.159 -> nice but ideally we want to get a lot
1718.48 -> more granular into exactly what's
1720.48 -> happening in this asset at this time
1722.96 -> so we offer uh this diagnostic tool
1726.08 -> to be able to alert to which of the 15
1728.88 -> which of the top 15 sensors
1730.88 -> are telling us that this this equipment
1733.44 -> is currently acting abnormally
1735.52 -> so for this specific anomaly we can see
1737.84 -> that most of the sensors
1739.279 -> issues are happening in the volute so
1741.44 -> you see this volume
1742.799 -> sensor 21 22 18. this is telling us that
1746.24 -> the specific
1747.039 -> issues in this specific in this abnormal
1750 -> behavior and whatever
1751.12 -> failure is about to occur is most likely
1753.2 -> going to occur inside of the volume
1756.399 -> now as i go to other failure modes
1759.6 -> or other anomaly ranges you can start to
1761.76 -> see that now
1762.96 -> the ranking of these features is
1764.799 -> changing so now
1766.559 -> the most prominent features are from the
1770.399 -> pump main housing
1772 -> and the pump main housing is a different
1774.159 -> component of this asset
1775.84 -> and the this anomaly range is now
1779.039 -> something that's actually occurring
1780.24 -> inside of the pump main housing and not
1782.08 -> necessarily
1783.039 -> in the volume anymore and i can continue
1786.24 -> down
1786.72 -> um each one of these so this now as i
1789.279 -> click on all of these i'm seeing that
1790.799 -> most of these issues are now in the in
1792.64 -> the
1793.36 -> in the shaft of this mode of this pump
1795.679 -> and not in the volute or the pump
1797.6 -> main housing and the idea here is that
1801.039 -> if you have an asset with upwards of 300
1803.44 -> sensors on it you know you can ingest
1805.2 -> all of that data into look out for
1806.799 -> equipment at once we can model the
1809.039 -> behavior
1809.679 -> the interactions between all of those
1811.2 -> sensors together and
1812.799 -> output results that say hey we're seeing
1814.799 -> some abnormal behavior in this asset
1816.96 -> and it's we're thinking it's because of
1818.72 -> these 5 10 15
1821.12 -> specific items or location you know
1823.919 -> specific sub components that might be
1825.52 -> causing it
1827.12 -> and give you the exact location in areas
1829.6 -> where you should be looking
1833.36 -> so now that i've um i'm at this step
1836.799 -> what i've we've ingested all of the
1839.12 -> historical data
1840.32 -> into uh look out for equipment we've
1843.279 -> given ourselves labels
1844.799 -> in order to train the model that we want
1846.799 -> we've trained our model
1848.08 -> we've looked at these results we are
1849.679 -> very happy with these results because
1852.24 -> there's a lot of detections inside of
1853.84 -> the label ranges and we can make sense
1856.159 -> of what's actually happening in these
1857.36 -> sensors
1858.24 -> now i want to put this in for real-time
1860.399 -> inferencing
1863.519 -> so i can just go to schedule inference
1866.159 -> at the very top
1867.279 -> and i'm just going to create name my
1869.039 -> scheduler and we'll call this
1870.64 -> um up where i have this labeled so we'll
1872.96 -> call it pump uh
1876.88 -> i don't know why it keeps changing uh
1878.32 -> we'll call this pump webinar scheduler
1881.84 -> and i now have access to the specific s3
1884.32 -> location in which
1885.6 -> um i want my uh real-time data to arrive
1891.12 -> so i want my real-time data to show up
1893.36 -> as real-time data inputs so i'm just
1895.44 -> going to select this folder
1896.88 -> this is where um you know your historian
1899.76 -> data moves to s3 or your data lake data
1902.72 -> wherever you're having your industrial
1905.6 -> data
1906.08 -> being stored in real time we can push it
1908.24 -> to this specific folder
1909.84 -> and look out for equipment we'll grab it
1912.08 -> for inferencing purposes so i'll choose
1913.84 -> that
1915.919 -> so now what's going to happen is that
1917.84 -> i'm just going to set a frequency of
1919.679 -> anywhere between 5 minutes
1921.36 -> and up to 1 hour that means that what
1924.32 -> we're telling lookout for equipment now
1926.399 -> and
1926.72 -> what we're scheduling is every let's say
1929.36 -> if i select 10 minutes
1931.12 -> every 10 minutes look out for equipment
1933.36 -> will wake up
1934.64 -> go and search this s3 bucket
1937.679 -> for any new data that's arrived from
1940.72 -> um from you know from real time from
1943.039 -> wherever the data is coming from
1945.919 -> we will run that through the model and
1948.72 -> output the inference results
1950.96 -> back to the output location that is also
1954.48 -> user defined
1955.36 -> so here i've just defined a bucket
1957.84 -> location that's real time data output
1960.48 -> and so again what's happening here is in
1963.919 -> uh every 10 minutes lookout for
1966.48 -> equipment
1966.96 -> wakes up searches the s3 bucket for new
1970.08 -> data
1970.799 -> runs it through inferencing and outputs
1972.88 -> the result in a json file format
1976.24 -> into this specific bucket the results
1979.279 -> will either be
1980.159 -> binary will be binary so it'll either be
1983.039 -> zero
1983.519 -> everything's normal or one everything's
1986.08 -> you know this we're currently seeing an
1987.519 -> abnormality
1988.96 -> and given that we're seeing that
1990.399 -> abnormality
1992.159 -> we will also then rank order all of the
1994.559 -> features as to what's telling us that
1996.32 -> this is abnormal
1998.72 -> now this is where lookout for equipment
2000.32 -> stops the objective here is to be able
2003.039 -> to take this output from this s3
2005.12 -> location and be able to visualize it
2007.039 -> inside of a
2008.48 -> monitoring solution either custom built
2011.76 -> or one that's already being used to
2013.519 -> monitor the equipment that
2015.039 -> that's generating this data
2019.6 -> so we've now um in this simple workflow
2023.36 -> we've
2024.08 -> ingested our data into the service we've
2026.96 -> given ourselves
2028.159 -> the label files we've selected our
2030.399 -> training times
2032.08 -> we've built our model and we've
2033.919 -> evaluated our model and now we have our
2035.84 -> model running in real time
2039.12 -> giving real time output on the equipment
2041.679 -> performance for this specific pump
2055.04 -> so um now that we're done with this pump
2057.359 -> example um just a little bit more on the
2059.599 -> use case and application
2061.599 -> um so ideally uh
2064.639 -> lookout for equipment is ideal for
2066.399 -> process industries you see
2068.079 -> four buckets here energy power utilities
2070.879 -> renewables and process manufacturing
2073.52 -> they all have some critical elements in
2076.48 -> in
2077.04 -> in common having assets that run
2080.56 -> continuously and with limited
2083.28 -> variability in their operating
2084.879 -> conditions
2086 -> so a pump in this example um can operate
2090.32 -> at
2090.72 -> higher speeds or uh lower speeds
2094.639 -> but it can operate with uh different
2098 -> fluid flows it could operate with
2099.599 -> different densities but in reality the
2102 -> it's its application is really confined
2104.56 -> to a number of
2105.52 -> uh to you know confined to a number of
2108.96 -> normal behaviors that can really be
2110.48 -> defined using ml
2112.24 -> now on the other side you know on the on
2114.4 -> the other end of the spectrum
2115.76 -> assets like uh construction equipment
2118.48 -> such as
2118.96 -> cranes uh maybe like robots or anything
2122.88 -> like that
2123.44 -> have such uh high amounts of variability
2126.24 -> in their usage that it's a lot more
2127.839 -> difficult to actually model out the
2129.28 -> behavior of a specific asset
2132.079 -> i would also bucket things like cnc
2134.16 -> machines into
2136.32 -> that more complex area as those sorts of
2140 -> applications tend to be
2142.56 -> you know if a cnc machine milling part a
2146 -> versus part b
2147.28 -> is a lot more complex as it um
2150.4 -> as each one of those parts might have a
2152.079 -> different behavior model to it
2156.96 -> so while i can't go into a lot of
2159.28 -> details
2160 -> on all of the use cases for specific
2162.56 -> customers
2164.079 -> we can talk at a high level about uh
2167.68 -> the customer applications siemens energy
2171.68 -> building a bunch of a suite of digital
2174.079 -> services
2175.04 -> for the energy sector um and
2178.24 -> looking at using the automated workflows
2180.32 -> for um
2181.92 -> making it really easy to build and
2183.599 -> deploy machine learning models across a
2185.44 -> variety of different asset classes
2188.96 -> without having to require smes to have
2191.52 -> any data science knowledge
2194.96 -> really siemens energy brings a lot of
2197.839 -> expertise in both
2199.52 -> equipment manufacturing as well as um
2203.2 -> you know mod modeling and understanding
2205.68 -> their equipment behavior
2207.119 -> um leveraging that with the ml output is
2210.24 -> is a great uh partnership here
2216.48 -> uh one other key customer here is sepsa
2219.119 -> energy
2219.839 -> so sepsa is looking at using uh look out
2222.56 -> for equipment
2223.359 -> bringing machine learning and insights
2225.04 -> closer to the subject matter experts
2227.52 -> so um you know some of the things we've
2229.76 -> been working on with sepsa is making
2231.44 -> sure that we can bring a lot of the
2234 -> ml output as uh as close and
2237.52 -> usab and very usable for the reliability
2240.8 -> and maintenance engineers
2242.48 -> this will allow them to make a lot more
2244.4 -> informed decisions and
2245.52 -> have a lot more actionable relevant and
2248.24 -> actionable results from
2250.32 -> uh from the outputs that they're seeing
2252.079 -> from different models
2257.599 -> we also have a a great partner network
2260.72 -> uh you'll see a lot of familiar names
2262.32 -> here especially siemens energy
2264 -> ge digital hitachi ventara we have you
2267.04 -> know consulting services with uh
2269.119 -> tensor iot rov assistant embassy of
2271.28 -> things
2272.56 -> we'll also work closely with our
2273.839 -> technology partners uh so osi soft
2277.359 -> and seek
2281.68 -> and then lastly for
2286.16 -> for getting started with amazon lookout
2287.839 -> for equipment
2290 -> always you can go to the aws.amazon.com
2294.48 -> console and just search for lookout for
2297.28 -> equipment
2298.56 -> that should land you on the product
2300.8 -> details page and all the information is
2302.64 -> there
2303.52 -> we also have links listed out here where
2306.56 -> you can get access to the user guide
2308.16 -> both
2308.56 -> in a powerpoint not a powerpoint a pdf
2312.72 -> document or an html file
2316.079 -> we also have free tier usage so uh for
2318.96 -> the first month
2319.92 -> you get a certain amount of usage free
2322.72 -> that includes
2324.48 -> 50 gigabytes of free ingestion 250 hours
2328.32 -> of free model training
2330.079 -> and 168 hours of free inferencing
2333.52 -> so that's up to a week of uh free model
2336.16 -> inferencing
2338.8 -> and then as well um you know there's a
2341.68 -> lot of information available on github
2343.599 -> and we'll keep populating this with more
2345.359 -> and more
2345.92 -> uh enablement tools uh if you go to just
2348.88 -> github.com
2350.88 -> amazon samples slash lookout for
2353.2 -> equipment demo
2354.4 -> um everything that i just went through
2356.32 -> on this demo is there
2358.079 -> as well as the data sets and example
2361.44 -> notebooks on
2362.48 -> a variety of different applications that
2364.4 -> um that make it easy
2366 -> really to get started
2370.72 -> and with that um i just want to thank
2373.2 -> everyone for
2375.92 -> you know for your time and we're going
2378.4 -> to start a q a session now so
2381.2 -> feel free to ask any questions that that
2384.16 -> come up
2385.2 -> i hope this was helpful and and thank
2390.119 -> you
Source: https://www.youtube.com/watch?v=RWCpG1Psf9c