Events Add an event Speakers Talks Collections
 
SigOpt Summit 2021
November 16, 2021, Online
SigOpt Summit 2021
Request Q&A
SigOpt Summit 2021
From the conference
SigOpt Summit 2021
Request Q&A
Video
A Novel Framework for Predictive Maintenance Using Deep Learning and Reliability Engineering
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
138
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About the talk

​​The oil and gas plant’s equipment usually has a long-life cycle. During its O&M (Operation and Maintenance) phase, since the accidental occurrence of offshore plant equipment causes catastrophic damage, it is necessary to make more efforts for managing critical offshore equipment. Prognostics estimation from multi-sensor time series data is useful to enable condition- based maintenance and ensure high operational availability of equipment. We propose a novel deep learning based approach for Prognostics that is useful in scenarios where: (i) access to labelled failure data is scarce due to rarity of failures (ii) future operational conditions are unobserved and (iii) inherent noise is present in the sensor readings. All three scenarios mentioned are unavoidable sources of uncertainty and often resulting in unreliable predictions to be considered for maintenance interventions and planning. The solution utilizes real time data from an operational oil and gas production facility in the UK North and combines engineering failure mechanics, reliability engineering and machine learning. A journey that involved building and optimise 200+ recurrent neural networks using SigOpt, a probabilistic optimization engine with an end-to-end cloud solution resulting in increased asset uptime, lower operating costs, and environmental impacts.

About speakers

Alex Lowden
Data Science Manager & Team Lead at Accenture
Shayan Mortazavi
Data Science Manager at Accenture

Alex Lowden is a Data Science Manager & Team Lead within the AI practice of Accenture where he is predominantly focused on applying machine learning to unlock value within the Energy & Chemicals industries. Alex has a unique blend of analytical and operational experience, working as a Petroleum Engineer in the UK North Sea on fast-paced operational projects, as well as multiple years’ experience building analytical solutions across a variety of industries ranging from Aerospace to Agriculture. He also holds a BSc in Mathematics, and an MSc in Petroleum Engineering from Imperial College London where he received the Colin Wall Award for his work on modelling transient temperature profiles in Ultra-Deep-Water wells.

View the profile

Shayan Mortazavi is a Data Science Manager at Accenture Industrial Analytics Group, where he is currently working within resources industries with interests in predictive maintenance, digital plant engineering, and optimization for upstream industries and the water industry. Shayan has over 15 years of experience in material science, engineering consulting, data science, and digital change. In his recent work, Shayan is developing scalable machine learning based solutions for the industry by applying Bayesian inference, dynamic control and graph network decision making systems. Shayan received his MSc from the University of Newcastle in Marine Science and his BSc in Fluid Dynamic and Structural Design. He is a chartered member of IMECHE and a winner of TD Williamson award for building a novel machine learning solution to detect and locate leaks in transmission pipelines.

View the profile
Share

I don't nobody will come to. Our talk is great to be here is great. The opportunity to present our work and collaboration with my colleague, Alex, Gordon and myself. Cheyenne with a zombie from Accenture, industrial groups are going to present a worse than spring worm, on a string of production, kitaco machineries in industrial. There are a number of different. What today's talk is going to specifically focus on predictive maintenance to maintenance. The

time. Party in the Big Data are increasingly being deployed as part of the intelligent management system for a number of studies and reports that they just hired. I knew that can be gained by incorporating Savings in maintenance costs, like 25% acid downtime by 60% and decreasing the schedule to repairs. What about persons are all significant numbers for the isotope ratios. Let's have a petition of different Meats with the strategies in the industrial maintenance is due to increased reliability

acid reactive. These approaches are simply living room assistant to run to failure and then fix one broken in a fight. But falling down more, symptomatic development and maintenance which are scheduled maintenance activities based on The Whispers approaches. For example, Cincinnati maintenance. Americold frosted maintenance, which is in effect, based on condition-based monitoring system in order to improve the performance. The trial of these preventive proactive maintenance,

which is even approach enables an issues before they occur. These approaches allows optimization of Maintenance in multiple different dimension cost. In order to see whether a better view of, what has been currency applied in the industrial maintenance is strategies. Based on the risk based on the availability of the data and physicality of that Machinery of production action, plants, mostly value-based expensive, resource-intensive. So what are these critical Machinery is the are often likes of rotating. Machinery is high

speed of a compressor pump. AC operational regimes and exposed to high temperature because of their application to are, usually at Austin modular design and with many parts and multiple, for each part of the production or manufacturing line, if an unexpected failure happens is a catastrophe. so why these type of missionaries are good candidates for predictive maintenance often, because of their complexity are equipped with 4,000, ranging from Cheshire vibration to read Psalms 24 of the asset at

any given time, real time data into the greatest control system for Add to their historians and recently in the cloud infrastructures. Because of the quality of the sensors that are maintained quite regularly and provide good reliability. So in the next part of Alex is going to do some of the pictures of maintenance. Great. Thank you. Thank you shine. So before we get into the framework, so we've developed for a machine-learning predictive maintenance. You wanted to know about the types of problems that we typically see in predictive maintenance, where

machine, I can be a fight and this list is not exhaustive. This is based purely on our experience of what we see in industry has four. Main problems. Machine learning can be applied in the first is a problem of equipment States or equipment house. This is really looking at identifying the current state of the system or component within the system determining and how serious is this component in a healthy state or an integrating state? If you want to apply this

type of each other equipment state, so I can call, and then you may be interested to know if Elements of a call. I like the engine or transmission system LLC or in recording state. Another common feature, we see it is time to vent, which is taking the previous feature of projecting phone service is looking at it. When you identify that equipment is in an unhealthy state is projecting the forecast in the actual remaining time until equipment, critically files or no longer can performance to buy something. Starr, elementary.

Chicopee, serious is root cause action in this is regarding operators to the source of health issues. So if there is an issue where I, where is it raining at, what's the likely mode? Failure, link to the symptoms to you. Sing? Define element is optimization, which combination of the previous three with considerations of resorts and risk, opening time to intervene to fix equipment to repair equipment, excetera makeup, that the framework that will talk to you about today. So it's just to set the scene a little bit. So this particular project

recently installed facility located in the UK North Sea, an ocean between the UK and Norway and its facility is a semi-submersible platform that sits in the middle of the ocean. Now fluids hydrocarbons, or we'll chat explode from reservoir on a gas or a liquid form reservoirs in the subsurface to through Wells to the seabed and then to stop stuff is in the structure Rises on heavy machinery. Typically you have a separation of Separation engine oil at water and gas and the pressurization for transportation to a lot

to prevent mechanical failures benefits. First, being that you improve safety and integrity being able to predict and prevent these failures. You have prevented catastrophic events, which can have an impact in terms of sexy at exit. The second do not have time. You can prevent over juice on plant downtime by predicting and preventing is it best to lower operating costs and reduce carbon emission in the event of an unplanned event? Due to failure often the hydrocarbon stream, Eddie Bridget to the flowers busted with can result in increased carbon emissions by building the problem most

of detecting and preventing. So what are the characteristics of building such a system to detect the health of equipment on such a green acid worm, you know every five characteristics to consider when building such a system. Now, the first is that any system your building to to text each other states of Machinery Machinery, into a guide operated two locations, where where they may be issued. The first limited availability of his typically, only Greenfield a set only have 23 years of

data, given limited supply of air date of Labor Day events. On top of this, you already have limited coverage of the multiple ways that Machinery to fail. So, forgiving piece of Machinery is everything. Many ways in which you can reach. Australia, State or the point at which you can no longer function and it's unlikely police search, bring your assets. You have a sufficient coverage of Vesperia married to apply as a supervisor, 20 framework new building tomorrow machine state. Or she needs to be able to

leverage the large volume of sensor data. Generated by these heavy machinery Expo, impressive package, which gas across the machine anywhere from Ms. Your frame of ocean needs to be highly interactive decisions. An intervention in the no secret. God, stop with the 500,000 US dollars. Is there any decisions? Any insights provided by the tool? Need to come with a degree of interest to help your creative makes those decisions? And the finals in maderas really scalability. The only sporting

facilities, you typically have a lot of diaper Genies with separatist compresses pumps each with a different function. To develop needs to be able to scan machine. In order to provide coverage of the asset is together, who is considered really necessitate, the use of an unsupervised of semi-supervised approach remodeling, the state of equipment on subsequent your assets. It's worth mentioning that common systems for this kind of problem exists. Expert systems. Why are you have thresholds based on experience

is typically can be intensive to my schedule. So, what is this? This this framework that we're proposing for monitoring the house of machinery and pointing operators to areas of defects consist of the five elements. Check. The first element is a building a projection for normal, or abnormal behavior. Behavior is Ford's prediction real time and then identify deviations from normal, using what we're fighting here and nonparametric, Dynamic. I will talk to you about. What's up is this

Chipotle adding mechanics area code and affects relationships through tool developed for Matrix, 3 days. Since I've seen them righted specific location. Donuts, first element, that modeling. Normal behavior. How do you go about doing this welding machines, option machines? Like these compressors typically contain thousands of Senses, might be vibrations a temperatures and pressures anytime series off a nonlinear non-stationary with extended periods of unusable data to Bob's. You have two of these machines in parallel the

failures. You may also have knowledge. Take me to bust of a short less critical sensors within the dose at now, by looking at the failure mechanisms and show listing, number Santa still give you sufficient coverage of the spell mechanism of subsets of that. So sweet to census data for modeling normal behavior on speed, sensor, search known that, what what you have information from complications caused between instances of between now? And the way that using the situation is as a sequence to sequence map

into taking a selfie. Now, the individual sensors, like individual temperature sensor with your dick to certain failure modes, you might have a uni dimensional 2nd dimension. In the case, where you have wealth is relative, the movement of a group of Senses. You can learn a multi-dimensional secret, a secret. Easton backs. Question of, how do you train a maintain such a large number of LSD. If you are using LSD, I'm stupid. This is really one of the challenging and interesting parts of such a framework is how to efficiently maintain on the floor. This number of a last year

stats that we took in order to expedite this, this this process. The first is around Ridnour, in the architecture search by start shutting off pictures of a galaxy, ambulance else to send a text with John. For vibrations, temperatures in a standardized yoyo architecture, to some extent by having your first last and handle extended periods of unusable data capacity number of neurons. Within juice, the requirement by searching for Life, disrespect multiple Matrix with a nice feature available with

Elation loss, training, time to finding the best model trains in the shortest amount of time. On top of this performing distribute distributor training with PlayStation optimization and paralyzation on Primus judge valuation. Also significantly helps reduce requirement FaceTime. Your valuation said, if I could promise you can be in parallel, evaluating multiple suggestion, simple, Tennessee. Given this projection no more, my collection, when I'll talk to you about the subsequent steps in Spanish. It's time to Alex. So

when we're considering getting a system to be a boss, dealing with multiple, I don't measure with multiple different types of some sort of an ulcer. On the other hand has Alex mentioned systems. Experience and undergo multiple different changes in the state of operation changes in the interview. Alex talks about Jehovah, Detroit. Lstms to handle temper ality and the change in the mood of the nation and uncertain generally developed by scientists in NASA. I'm going on, I'm at sacred geometry.

Constructions are often too easily and resulting to the high rate of false alarms because these are set up on the side of temporal effects of the sick. So, how does it work? Basically, at any given time, this set of thresholds are applied not to the signal itself, but to and residual or error signal on to detect animalistic. When said the air space, when a new set of observations of our that I received a wide range of possible. Special bands are optimal one based on population of the

largest rate of change properties of the error distribution. The ones, an anomalous sequence is detected is sequence. A subject calculation of scoring mechanics and calculated based on their area of the sequence using the second moment. The super guys talked about a way that you can make sense of these abnormalities maintenance engineer. I need to press the domain knowledge into our worship in effect is a method that links. Sensors on the importance of the sensors component.

How does it work at author receiving a set of observations using this weight of importance Matrix of their components? Something recalled. Escort is chocolate. Do the dishes, start using the logic solve reliability. Engineering, schools are ruled up, an aggregated system level. And by the same means from subsystem Level 2 system House of the system. In size are generated based on the moods of failure for components and allows their end users to text on. Genders have the

ability to reject, or accept allow the user to provide their feet. so in order to put everything into perspective, that we have developed based on what you see on the screen is an overview of the acid is a process blocks, our room shoes to fill up the process and the boxes are there, active physical assets, that are part of the mother and the militant white paint is the inside of a generator waiting for users to The front is system, overview level and drill down into the system level where the subsystem system is

broken down into major assemblies. In this case of a compressor gasket. Gearbox Parts, part of the compressors are comprising. This trending in time in order to get dog musics on what is going on inside. Each assembly are forming disassembly and the internals of a compressor. And a case of bearings, we see how historically whites have trended from Elsie, Elsie Reese, the region and analyze. What are the drivers, the bottom sensors that are driving. These abnormalities, in this case, a bunch of bearing temperature

on the upper and lower bearings are indicating higher and higher. In order to allow the user to investigate further. What is the source of the deviation of what sitting on the anomaly score, receipt or the actual measurements are deviating from the prediction from the girls and how many schools are increasingly indicating? The ones is animated scores are automatically resulting in generation of inside. Stern and users are able to take action and either accept or reject and a link to their

actions into the maintenance management system. This would provide an end-to-end overview of the artist based on the framework Outlet, Alex and provide some of the lyrics. Play stinky, stinky shine message here. But since production ization through this too was production. I did in 2018, numerous Valley events have been detected by the toe, and most importantly events for identify in a week before Bay Area cities Revenge, that could have resulted in, in, in

catastrophic consequences to a number please for this area. We want to just finished with a couple of learning SAS three major names from the building of this crime was the first, I really being around Stockton application of machine learning in Maine, quiet. Wedding of the main knowledge to be successful. If we're really advocating. This is bottom. Approach, a combining with the mechanics of area. The second is around the adoption of Sasha to. Now within the project, team had a number of change management staff

members who helped with the adoption of the tools. So I can totally state. Is it new to the engineering works like quite a significant cultural changes not to be underestimated. Allison Tolman. Just finished on. Is it the use of these dead shall rise? Like cigarettes is really critical for developing maintaining Sasha to the retraining a number. Like $200 yams would not have been possible without Sasha, to do not know how to thank you very much for your attention on your time.

Cackle comments for the website

Buy this talk

Access to the talk “A Novel Framework for Predictive Maintenance Using Deep Learning and Reliability Engineering”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Full access

Get access to all videos “SigOpt Summit 2021”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “Education, Training and EdTech”?

You might be interested in videos from this event

November 1 - 4, 2021
Lisbon
17
3.2 M
ai, brand, chatbot, fintech, it, maas, marketing, sport, startup, technoligies

Similar talks

Ke Ding
Principal AI Engineer Intel at SATG MLP – Applied ML team
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Marat Latypov
Assistant Professor of Materials Science and Engineering at University of Arizona
+ 3 speakers
Michael McCourt
Head of Engineering at SigOpt
+ 3 speakers
Paul Leu
Associate Professor at University of Pittsburgh
+ 3 speakers
Vishwanath Hegadekatte
Global R&D Manager at Novelis
+ 3 speakers
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Rafael Gomez-Bombarelli
Professor at MIT
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video
Access to the talk “A Novel Framework for Predictive Maintenance Using Deep Learning and Reliability Engineering”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
946 conferences
37606 speakers
14373 hours of content
Alex Lowden
Shayan Mortazavi