Week 7 - Lecture 31 : Advanced Prognostic Modeling Techniques and Data-Driven Approaches

Hi friends. So we are into our seventh lecture of uh prognostics and health management a systems approach. uh I'm bringing this element that is a systems approach uh because for complex engineering systems uh we must integrate prognostics and that starts from right uh conceptual stage design stage and then um commissioning operation uh stage and then plant goes into shutdown after uh tens of years of operation and then also there are some gap areas we find where uh especially about aging of the components uh where we need to have prognostics and uh diagnostics. So the uh topic or the subject today is fault diagnosis uh and machine learning. Uh fault diagnosis probably all of you must have heard about it. Whenever you build a machine and if uh it fails and then you try to understand what are the faults. If uh if uh you have power supply in your house and if the power supply is not uh failed uh then you try to diagnose the problem uh whether whether it is overload problem whether it is a short circuit somewhere or uh you know uh some uh some other fault related to end like bulb has fused how you'll know that uh because uh electricity passes through so many it is junction failure switch failure um you know breaker failure many things so fault diagnosis is is integral part of our life. uh then uh it is something like uh in factories also when the machines don't start uh we say what is the problem and that is where the fall diagnostic starts or uh you hear an unusual noise from a machine and then you ask your operator to stop and then try to do an investigation because fault should not propagate into failure you know or should not manifest into failure because once failure is there replacement of the components are required and that takes time and all that so let us repair it before that itself um whether it is civil, mechanical, electrical, everywhere uh fault diagnosis is a integral part of any system operation or maintenance. So uh in the first lecture uh I'll introduce this subject in a little detail and then finally we'll go to the second third fourth and all that because our idea is to learn the fall diagnostic techniques um and uh then how it can be embedded uh in machine learning so that our uh systems they come with the fall diagnostics uh available even condition based mainten also is something which is related to fall diagnostics. So um this week uh in the context of fall diagnostics and machine learning um the first lecture as I mentioned it is a introduction uh and then systems approach to fall diagnosis or diagnostics I'll be using these words interchangeably diagnosis or diagnostics you know and then because this was a so many techniques are uh there so um uh part A and part B are there they are covered in part A part A and part B and there is One more uh uh technique uh which has got its own uh uh self-standing it's called root cause analysis. It's a very serious and very comprehensive analysis uh because it is done in safety critical system uh or you know uh mega system uh where uh if the component fail the uh stakes are very high in terms of uh uh risk as well as in terms of the uh shutdown time because once the plant shut down deliverables stop and uh that's how so the idea is once you do the root cause anal analysis and you have reached the root of the failures or roots of the failure. take the correct corrective action program uh corrective action program uh such that this failures ideally speaking if the root cause analysis was correct and you know uh we we were able to get the uh root faults uh and that we have corrected that means it should not recur at least uh in uh coming time maybe five year 10 year or so that means we say that root cause analysis done was very successful because all those faults are not recurring and even the feedback analysis will tell us key those faults are nowhere there and then machine diagnostic learning uh machine learning and diagnostics we'll see how these diagn diagnostics they are implemented into machine learning uh we will not be covering all the tools and techniques but whatever uh coverage is there in this lecture probably it will be sufficient for you so the background as I mentioned uh fault diagnosis uh is a uh is a very is a very specialized domain. In fact, maintenance and operation stops uh in any big company or a uh center uh they are uh they are trained to do fall diagnostics. In fact, if you remove from the training the fall diagnostics, that means you uh you don't have required qualification to operate or uh maintain the uh machines actually. Um and then um u what is the origin of this word uh fall diagnosis uh to be uh to be very uh straight on this probably uh human existence and fall diagnostics they came uh around the same time because your survival was dependent on how uh strong your protection is there how uh how uh strong uh your survival instinct is there and that will be determined by some physical entities like your house, your tools that you are using and if your fault diagnosis goes wrong uh you end up into uh threat condition uh and you more no more remain in the uh uh safety paradigm actually uh and then diagnostics and prognostics. In fact, diagnostics normally it is said that uh it is the either the failure when you start or you try to see how the failures can occur and you that you line up or embed them into machine or you embed them into our training program and that's how diagnostics uh bearing failure you know that inner catch failure outer outer catch failure uh then uh you know ball failure and many misalignment all those things. So there are wellestablished diagnostic uh tools and techniques that are available and as on today uh your diagnostics expert or condition monitoring or in service inspection expert uh can tell what is what has failed the these things the learnings have gone so strong that looking at the vibration signature they will tell you whether inner race has failed outer race has failed or balls uh one or two balls uh they are they have failed or cage has failed. So uh we have gone quite a deep into it even even a human understanding of the fall diagnostic itself. This I'm talking about bearing this is for true for even any other uh now uh diagnostics we we do in the current context but prognostics is for the future. So in diagnostics we see how it can happen and in prognostics we say when it will happen it is remaining useful life. So it is prognostics is a futuristic but building a prognostic also you require a lot of diagnostic knowledge what you will embed into the machine. So you should have some ways and I probably in this lecture you'll understand how they overlay diagnostics and pro prognostics. Then machine learning and diagnostics yes it provides a uh knowledge of the plant or component or system built into the machine and those machine they detect there is something wrong okay online. So it becomes a support for the staff who are involved in operation and maintenance of the plant and diagnostic tools and technique. There are techniques um some techniques just for the sake of uh completion we'll do it. Okay. So and the root cause analysis it is a field in itself. It's a very serious field because sometimes you have to model theoretical studies here. Sometimes you have to go for interview. Sometimes you have to go for recording, re-recording, simulation. Uh those and in root cause analysis I would say every diagnostic technique has got its role. So when I perform a root cause analysis that means I am using the toolbox of the diagnostic technique. Um so if I have to tell what what is the or origin. So the immediate available literature shows that diagnostic or diagnosis has got a uh its origin in healthcare or medical field. You know when a person approaches doctor he will ask some question he will do some test and those things are called together in simple way diagnosis was done. So what what what was leaking or what is the disease that is found out and after that the only prognostic procedure starts. So uh the the available record shows this. But then if you look at it uh in a very uh documented way or something if if you are hellbent upon seeing the then Ayurveda uh Greeks philosophy and Egyptian philosophy they also showed there are there there were techniques that were available but they were most mostly we try to see some uh some u examination which involves looking at your eyes, ears, mouth, nose inspection and then and pulse rate of course. uh and then we we uh all this thing ensemble of uh checking that was done uh we try to say okay this is the problem and then accordingly the medicine was given uh and of course if required because in uh ayurvea or in ancient Indian philosophy there was techniques where even surgeries were performed there are documented uh evidences that are available so it was uh like trying to do some sort of investigation to uh surgery. It was a full and follow. These were the procedures actually. Uh and if we trace to the human existence that is go much uh much uh behind and try to see uh you know the human existence so you know what all is required and what all you should understand things around you how they work how they don't work. uh if I have some tool which I'm using uh during my uh when I go for uh you know hunting animals uh so I should know how my tool works and how it can fail. So diagnostic things they were very primitive tool as it looks like but then there were survival mechanism. So um so uh protections uh whether you build your house u uh even if it is a mud house we should know that it should not fall it should not cause you. So uh that is the traces we have from the history even common sense will tell us that diagnostics was part of it and then natural evolution occurred in diagnosis informal way to in very formal way. Um again we are going back into into the uh past and trying to see rational based approach evolved and uh protection from harm and all they discussed and management thing came into uh existence. Okay. Prognosis is relatively little known but qualitative meaning of prognosis were used even earlier also in medical term also. uh we might not use a very quantified commitment or statement of the remaining life but doctors may say uh it is 3 years, 4 years, 5 years and that may vary as the diagnosis uh progresses actually uh so it has evolved from empirical science uh to experimental methods and finally clinical methods. So this is the evolution and now it is on the shop floor. Clinical is a term which is used uh in the medical domain and shop floors we use a term in industrial or engineering domain. Uh and the defense um so somewhere along this uh healthcare uh diagnostic engineering diagnosis came into existence probably starting with the defense and then uh engineering in general. Okay. So um if you see uh science where is it is it it is being applied more one is the life sciences and then other one is defense. So in defense it is the uh before it touches normal industry it entered there actually because uh in defense uh equipment should not fail in the field and that's how the birth of reliability also is in the world war when the equipment they were taken to the field and they were failing. So a formal MI217 or these standards they took their birth and reliability field came into the existence. Okay. So now if I look at the plant life cycle you know uh it starts with the uh planning design commissioning operation and maintenance refurbishment. Refurbishment means extended maintenance is done on the plant to extend its life. So 5 years 10 years uh after 30 years or say 20 years or 40 years the refurbishment program so that we can operate the same plant because there it is a very expensive provision to shut down the plant permanently. But if uh even after refurbishment some 10 years or 20 years of operation um the plant is shut down then decommissioning starts and finally the area which was covered by the plant it is converted into green field. Okay that is ideally should happen actually. So the advances in diagnostics you can see all through during planning stage when even we uh start the uh building the system uh we are bothered about whether they fail more less and which is their weakness and all that. So uh in very uh then only we go for design stage uh because you know once we decide okay our uh our pump motor heat exchanger turbine they should be of this size and for maintaining them we require a main hole uh you know to inspect the inspect the tubes and all that. So, so provision for diagnostics is visualized right in the planning stage itself. And then finally once that uh uh conceptual things are there then design stage things are come into picture. And in design one has to find out the maintainability of the system. And maintainability why because we should be able to uh not only do the fault diagnosis we should be able to repair replacement should be done such that uh maintainability uh can be effectively seen along with the uh along with the surveillance also. Okay. Then procurement stage. Yes, we will go for selecting those components where there are provision for diagnostics and now maybe uh time to follow. We'll be asking for prognostic provision also or condition based monitoring provision also. Okay, because condition based maintenance they will give line of degradation or track of degradation and and then prognostics will derive information on the remaining useful life. So state-of-the-art equipment down the line have to follow prognostics also. Commissioning stage. Yes. Once the plant is commissioned all the things which are perceived earlier they were they will be documented and whether the fault can be removed fault can be diagnosed if there is uh there is some problem in commissioning stage itself it will be fixed because after plant gets into operation it becomes difficult and during plant operation yes we get the firstand experience of and so uh only still 5% or 10% areas will be you'll Mind that maintainability is not there and that is where it calls for the uh updating uh uh you know uh plant uh doing maintenance and uh you know trying to do some back fitting something so that you know operable plant is operable and maintainable and refurbishing stage it's a very something very important and now we want to track the components which have operated for 40 years but they will operate for 10 years so we put additional sensor it is very important in refrigeration stage. Whether it is a machine or a passive component like pipeline or vessel we have to monitor them online because their margins have come down and we should not go to uh you know uh threatened stage or we should not go to a stage where uh where events uh can happen actually. Okay. So that's how it is. So river stage is very important. It gives feedback of last 30 to 40 years and then some tools for diagnostics and prognostics are installed. So this is uh we are able to review the life cycle of uh and diagnostics and prognostics. Now traditional approach to diagnostics in operating plant you know that there are two category of component I have been repeating uh time and again uh process system and safety system broad category there are in between many safety support system uh you know um maintainable system non-maintainable system there are so so many but let us put them into two categories why because process systems can be they are continuously operating so they can be in fact uh easily monitored even a small leakage you'll know because the system was operating. If a motor is about to fail or you know the we'll know their signature if bearing was about to fail we will know the signature that this bearing for a recirculation loop uh uh has some problem and corrective actions are taken. So process systems yes uh online monitoring and all but in safety system online monitoring will not help because safety system remain uh uh remain uh standby. So if something is not operating uh complete train of electronics if it is remaining standby uh then how you'll know. So for this also there are some special facilities are built like built-in test that means electronics is there there is let's assume there is there is no current flowing from up to down end for actuation. Uh so what you do so you start giving small impulses and see whether the circuit is all right. So this is called built-in test facilities which are for electronics. Similarly for shutdown thing also somewhere leakage or some some um if you if you want to monitor bearing then physically we do the testing that is called test testing and surveillance. So for a small period they will be started compared to their full full-fledged operation and we'll know that and do the diagnostics whether it is all right or not. If they are not all right then again a complete diagnostics follow why it happened because oil problem or the bearing got aged or inner inner race outer race as as I mentioned there. So those things have to be fixed for safety system also and it is done during testing. Okay. So there is called test and surveillance program that you have to follow and that's how the diagnostics is implemented for process system safety system and uh if electronics has a different case because mostly it is a passive component so uh we have that provision. So um diagnostics and prognostics uh before we come to the uh uh discussion where is their overlap how they different uh okay so just a small comparison it will show that uh uh the analysis if we say purpose diagnostics it is present health of the uh component but in prognostics it it is future prediction that is future health um you know and go on doing tracking till it fails. So then we call it as a remaining useful life prediction you know and then root cause analysis is a failure investigation. Okay that failure should not happen that component should not fail and diagnosis that if it is a incipient failure we should catch in between if it has failed then we should take so it relates to present plant operation and maintenance while here futuristic. So these three we see the overlap over here because as I mentioned prognostics also if you built and embed into a machine learning tool a complete diagnostic has has to be done uh which component will fail how it will fail what will be its signature. So that knowledge helps us to build prediction uh module also. Uh and let's a typical case to uh to continue our discussion we have a motor uh bearing and a pump setup everywhere. It is such a generic setup which is there. Um so for monitoring the health of the uh bearing we have vibration and temperature sensors are provided and for the bearing and this is a bullseye. Bullsai means there is a small hole with a with a uh with a glass which shows the oil level. Uh if it is a oilbearing and then it is called bullsai actually and oil quality and quantity also you can check over here um both if if it is certain level no need to worry if it is more also then also it is not good. If it is less then also we have to maintain a minimum level where it it can go up and down in a small band actually you know. So uh we have this uh this setup and if I have to talk about lot of work is going on in prognostics and health management of bearing because bearings are critical component. If they fail uh you know uh then uh the shutdown extended shutdown might happen or if they are part of suppose this is a part of injection system let us say uh water injection for fire uh injection system then it is a safety system. So then also if on demand it doesn't come then also it is a problem. So it is better to understand the bearing first it is a construction feature. So bearing is a very simple mechanism. It has got inner race that is inner ring. Uh then this outer ring the outer race in between balls are there uh you know spherical balls are there and they are being held in position by cage ring. It is called black portion is show showing the cage. So now the science of vibration has come to this level. It can looking at the signature whether it is a frequency domain or time domain. We can make out whether it is a cage has failed or one or two ball has failed or inner race has failed or outer race has failed. Looking at the signature without machine is operating we can make out. Maybe in some lecture uh case studies when we discuss we'll talk about it. um uh in details actually. So and then uh one more important thing is there sometimes motor current signature uh uh or voltage variation deep in voltage and all they also provide critical information about the health of the motor. So what we have seen here prognostics and health management or diagnostics of motor bearing and pump and of course this pipe piping and all that uh and uh this motor current signature analysis uh we have. So uh this field is matured in the sense that uh using vibration signature or motor current signature analysis we can find out what is wrong. These two white things are one is voltage and the other one is current meter actually uh which gives us online recording. But during uh uh the prognostics we have to have put put some sensors over here for vibration and all that because radial actual shaft shaft misalignment bearing I told you inner race outer race so many things you have to do uh this thing and for motor it is a motor current signature analysis and so many other things like sometime there is a um conductors are there the insulation failure and all those kind of phenomenas are encountered we can say that so diagnostics and prognostics healthcare it it was inception was there invisibly we can see in the near 100 years or so. Uh and uh so and then in engineering it uh it all started some borrowed somewhere and somewhere the engineering knowledge is set and prognostics and diagnostics were started. Diagnostics um uh that you know we do for present uh situation and prognostics we do for future. It could be online or offline both. Uh sometimes it could be between online and offline in the sense that we we carry some instrument measure the vibration so it is not online but we are uh but we put the sensor there and that way we do on so many machines. So okay so that is also could be prognostic but it will be sort of offline because those signals are not coming to the control room. Okay. And then diagnostics also in similar way. Uh and typical temperature for a mechanical uh system it is could be vibration, corrosion uh you know there are various mechanisms are there and for prognostics also uh having the sensors installed there uh we can uh we can check its uh life and how uh uh we can estimate the remaining useful life. So rul is the integral part part of prognostics. Now if I do a comp a comp comparison and try to understand. So the purpose of uh diagnos is to detect uh whether online or offline mode and implement a corrective mention uh program. Prognostics like health for prediction of future conditions in general you know but if you tell the objective it is um uh it is estimating the remaining useful life so that management action can be implemented and it doesn't remain prognostics it becomes prognostics and health management and uh objective here is to find fault and to isolate or remove the fault and take a corrective action and techniques that are there online offline both are there for diagnostics also and uh For like for example for electronics I mentioned it is a bit uh built-in test facility that is provided online or instrumented such that we will no fault a small short impulse of microconds are sent across that so that the uh relay or something whatever was actuator that will not come into action. Uh so uh so for actuation of the final relay that means process will get disturbed. It it require a signal of more than 30 microcond. So in one or two microcond that signal will be sent without disturbing the electronic card you are able to know their conditions actually and uh uh techniques here are remaining useful life estimation it is based on the sensor or precursor parameter monitoring okay and domain specific requirement diagnostics or prognostic anything it will not have it will not be successful till we have domain uh specific uh knowledge okay that means domain And experts should be there when the diagnostics are being modeled or prognostic should be modeled. Um having a machine learning uh machine learning expert alone is not sufficient. For a quality diagnostics you require a domain expert. Uh so for like maintenance you should have a maintenance expert. For operation you should have operation expert and for similar same thing is for true for prognostics also. Science and technology it is diagnostics have landed into engineering labs and online monitoring are uh adequate but resource consuming prognostics is a resource consuming because it requires huge computational AI tools and then maintaining those things and all and it requires a huge data efficient system. Uh sometimes it is cloud data we are operating on cloud and all so internet of things they come into picture. So these are the things that are required attention on trend and past till uh present uh what what is going wrong or how it is evolving and it is here the question is when when this component is going to fail. So that means that will give us definition of uh understanding of remaining uh useful life. Okay. So uh now everything we are trying to say plant risk and reliability we are reducing. Uh so uh what is the role of diagnostics and uh in risk and reliability? If a component fails then uh reliability uh comes down. If a safety system fail it manifest as a uh risk parameter. Okay. Both are part of uh uh risk reduction program also and both are part of even increasing the reliability also and in turn how it is achieved it it is achieved through increasing the availability and maintainability risk is realized through reduction of partic initiating the frequency of initiating event that means uh if we are reducing that means we are increasing the plant reliability by uh by removing un unwarranted failure. failure uh trips or failures. Uh there is a need to pay attention to common cause failures and human factor. Somewhere we'll be touching upon in diagnostics module itself. Uh how we can work about the otherwise there are empirical uh things and empirical data that are available and we are working with them. uh so what we can do something better and of course human factor prognostics is very important because it is again one of the common cause uh factor into diagnostics and prognostics and in reliability as a whole machine learning in diagnostics the artificial neural network which has made a mark in the in the it manifested as a deep learning approach and then support vector machine expert system diagnostics they are uh they are used extensively uh it the crisp rule based or fuzzy rule based and then statistical methods are there of them basian approach. Uh we will be discussing one example also on this. Um so with this uh uh we can say uh that we have seen the background of diagnostics and machine learning plant uh life cycle uh and diagnostics what it means uh system uh context to diagnostics you know uh why we should have a systems approach actually in diagnostics and uh major features of diagnostics and of course the what are the overlapping area in diagnostics and prognostics they are not in isolation of course they have deliverables uh in the present and one is another prognostic is for the future. So popular machine learning methods they form part of it because now they are implemented uh either as a machine learning or as a deep learning approach.