 In this lecture we are going to speak about fault diagnostics and prognostics as we already know that in this course we are focusing mostly on CBM that is condition based maintenance. We will see what are the definitions of fault diagnostics and what is prognostics and how do we estimate the remaining useful life of a machine because of the fact that you know once we maintain a machine we would like to see or predict by certain tools how long this machine is going to last before we totally remove it out of service. So the essential elements of CBM or the prognostics health management cycle are as follows. As you all know the most important features are the sensors around a machine and then sensors are going to provide us data which has to be analyzed and all our decision is based on this analysis. But in the modern days CBM the trend is towards that the sensor whether I can do with single sensor whether I should have a multiple sensor whether the sensors are fault tolerant etc. Because as you know everything in this CBM depends on the data collected by the sensor. In the to begin with if the sensor itself is faulty and it is providing us wrong data our analysis will be totally wrong. So the most important feature of what the sensors nowadays are used in CBM are what is known as smart sensors and topic like sensor fusion whether a single sensor will do or a multiple sensors have to be used and then of course towards the end of this lecture I will give in a short case study on sensor fusion and how we will see the remaining useful life of a cutting tool can be predicted by using sensor fusion. Now once we have the data coming from this sensor we can classify the faults and also then try to predict how the fault is going to evolve and then if necessary we will go for maintenance. So the prognostics or estimation of the remaining useful life depends on my data analysis. So before we move to the next sections I would like to give you a definition of fault diagnostics which is detection, isolating and identifying an impending or incipient failure condition the affected component subsystem system is still operational even though at a degraded mode. So the fault diagnosis encompasses detection, isolation and identification so fault diagnosis has three important components one is detection, other is isolation and other is identification and also and fault diagnosis is also used to detect isolate and identify a component or system that has seized to operate. So in one case we have the machine still running, machine running in another case we have machine sees to operate as I was telling in the last class you know CBM is not good for a machine which is not operating so we will use mostly CBM for a case when the machine is running. So to define fault detection it is known as an abnormal operating condition which is detected and reported. So this is the definition of fault or failure detection that is an abnormal operating condition which is detected and reported. We will see the mathematical techniques how such faults can be detected can be isolated can be identified and so on in the subsequent sections. Next is the fault failure isolation so determining which component subsystem or system is failing or has failed. Next is the fault failure identification that is estimating the nature and extent of the fault. Now question is how do we diagnose these faults? Diagnosis traditionally are done by two important methods one is the model based method and other is data driven or signal based method. In model based methods people use certain mathematical models like the from the physics of the problem in model based techniques they have a mathematical technique known as for example I will give you a simple example here. Suppose I have a system a simple mass spring dashboard system and I have got certain response because of a fault in it. So I can write the equation of motion of this system and because of some force I am going to get a response though I have written as a vector and shown a single degree of freedom system but you could generalize that a system having a lot of m and k's and c's and it can be written in such a matrix form and this force is what is creating the response in which we are measured response which is measured and once I have the response measured and if I have some estimate of this m c and k I can through this equation find out this force. If this force is abnormal force is more than what is necessary I can say a fault has occurred in the system and this could be done by what is known as the residue generation. In the model based system one such method is the residual generation technique. However in the present course on CBM our focus will be towards mostly data driven or which is also otherwise known as signal based because this is very easy to understand in the sense I have a machine I put a sensor and I get my signal and everything is the total FDI is done based on the signals and that is traditionally people do such kind of FDI or CBM in the industry and this is very easy to do compared to the model based technique wherein in the model based technique you require lot of computation, lot of physical understanding of the system. For complex large systems this becomes very difficult to model the entire complex system. So traditionally people have been using mostly the data driven or the signal based technique and in this course we will be focusing our attention mostly towards signal based techniques. Though while we are on this topic I will also tell you some of the model based techniques other than the residual generation techniques which are used in the industry or are practiced by maintenance engineers to do a fault detection or identification in a machinery. Before I go to the model based fault diagnosis I should mention to you that this data driven fault classification is actually done by many methods. One is stored fault written library and then we have a fault classification. For example something like a lookup table we compare in a database signals already stored with the measured signal. If the deviations are large we can pretty safely say a fault has occurred. Another is this feature vectors now we can in this signal what are the features to be studied. I will focus more on this when I talk about the time domain and frequency domain analysis but to do a fault classification there are nowadays many techniques one is the neural network based methods which is very widely used in the industry the fuzzy logic and the Bayesian system and then these are mostly reliability and probability based models and then defaults can be diagnosed. So in a data driven fault classification there are techniques wherein this signal are analyzed based on certain neural network based on fuzzy logic and a decision is made on the fault classification. Could be by fault classification I mean whether fault are of two different types one is severe one is normal or one is something which we can live with. So such decisions can be made by a neural network systems and I will give you an example as to how this kind of study can be done for different machines. We have a case study for a cutting tool monitoring situation which I am going to tell and this data diagnostic methods there are many ways by which this data can be diagnosed. One simple is say for example with time we have a certain signal parameter which we are monitoring say for example line now with time these are my upper and lower bounds. So what could be happening that my system parameter etc. So this is the upper bound and this is the lower bound. I can set the limits of this upper and lower bound within the plus minus 3 sigma values of the measured parameter and I can be only alarmed that a fault has occurred in such situation fault has been detected. So simple alarm bounds can be used and then in fact fault can be detected and traditionally or historically in the industry people were used such people were using such alarm bounds and then you know you can detect a fault but nowadays the availability of many algorithms and fast computers many new methods of fault diagnostic methods have evolved and we will discuss few of them here. Many statistical techniques are available I will give you one simple example of regression analysis for example with time I know that this is the signal parameter and this is my alarm level that means if the signal will increase beyond this level I will be alarmed about the condition of the machinery. For example what I could do is suppose I have measured with time that these are the condition of my signal. So statistically what I could do is I could do a simple regression analysis and suppose my measurement has been done only at these points looking at the past data and taking the present data I could do a straight line fit where you know y is equal to m x plus c is the equation to this curve here this is y and this is x but this is a single component I could have multiple coefficient but the most important parameter of this is had I known this condition of this machine and I know the equation to the straight line I can always find out the remaining useful life from the past data and this has been also known as what is known as the trend analysis in the industry that means looking at the past data measuring the present data I can come up with regression and then estimate when this machine will have a failure. So there are many software available with lot of database where this kind of data are stored people do simple regression analysis and can predict the failures and even in industry if this kind of measurements are done this is going to give us a lead time before I do my maintenance so that I know the remaining useful life. So I have this kind of a lead time before the failure occurs to take corrective measures I could perhaps now you know because of this trend I could increase my maintenance activity increase the maintenance activities such that I can reduce the failure and subsequently maybe this levels could come down and with time you can take corrective measures and that was just a simple example using the straight line regression analysis or linear fits but nowadays when multiple parameters are involved it becomes impossible to find out a straight line fit and many of the relationships between the input to the system. So such an analysis it becomes very difficult humanly to measure know what are the input parameters output parameters so people use what is known as a neural network or artificial neural network to find out a relationship between the input and output and come up with models as to what kind of output levels will be predicted with as a function of time and then we can estimate the remaining useful life by such a technique there are many methods of RUL estimation one is the fuzzy logic classification other is the neural network classification and next one is wavelet neural network in this course we will not be focusing into more in more on model based for diagnostics or using neural network and other tools but mostly since it is a very basic course on machinery for diagnostics and signal processing we will be mostly spending our time on the fault detection through the looking at the alarm levels looking at the past history looking at the lookout tables looking at the historical data and then try to correlate the fault condition with the present signature and that is what we are going to focus in this course. Now that we have some idea regarding the data based fault detection methods I will also tell you about the certain physical model based methods which are used in well CBM may be in large sophisticated systems for example an Airbus A380 there are many many systems it is a very very complex systems and what kind of fault detection strategy is to be used so this could be a combination of data driven and also a model based. Some very fundamental problem in engineering is be it cement mills be it gear boxes casings etc is the development of fatigue crack and its propagation so once a crack has developed on a machinery component say some crack on a machine component the next question is always people have in their mind is how long is this crack going to be there can my machine still run under this crack condition etc and the most important equation to this is this crack growth parameter with the number of cycles of operation C0 is a material property NAN is also a material property and this is known as the famous various equation from such an equation we can always estimate the crack length or dimension for crack this is the number of cycles the material property this is the stress intensity level and this is again a material property we can decide on how many number of cycles this machine can be loaded before a crack grows from an initial length to a final length. And then once N is estimated we can then decide on the RUL for a crack component this being a basic course I will not go into details of all this because this will require lecture on fatigue failure and fatigue fracture mechanics but I will just want to mention to you that such techniques are available for the crack growth development and then once we know the final crack length which our machine can be subjected to we can decide and we can estimate the number of loading cycles and thus know the time remaining useful life of the machine. Another very important model based techniques which people use is the FEA. So, FEA has been used in model based diagnostics so what we can do you know if you go back to that same equation I was talking about simple FEA model of the system can be developed and this force can be estimated and then the model based fault diagnostic procedure or strategy can be developed for the fault detection. Now once we know have an idea regarding the fault detection, fault diagnostics, fault solution identification and once we have understood our system the most important parameter is the RUL and what is actually known as the fault prognostics. So, prognosis is the ability to predict accurately and precisely the remaining useful life of a failing component or subsystem because once the fault has been detected and isolated the next question is this prognostics to estimate the remaining useful life and all these mathematical tools of you know trend analysis artificial neural network etc could be used to predict RUL. Now with this what I would like to do is I will show you an example as to how this model based or data driven based technique can be used for fault detection in a machine. So we are going to talk about toolware monitoring case study in fact and this was the project which we did in IIT about a decade ago and this was sponsored by the department of information technology here. This project was to estimate the effect of multiple sensor based fusion using neural network to determine the useful life of a single point cutting tool. In fact in this example I mean this work has actually been published by the in the journal of mechanical systems signal processing and the details of this could be found in that reference. But here let me tell you what this tool condition monitoring is about. We have a phase milling operation being done on a machine. We used two cases or work material one was a steel and another was the aluminum. There were no cutting fluid use it is a dry cutting condition there is a single cutting tool in this insert and the cutting speed approximately was at 140 meters per minute or 557 rpm of the spindle. Feed and depth of cut were maintained less and the approximate toolware was 75 in such a condition this instrument or this milling machine was instrumented with a cutting tool force dynamometer was instrumented with the acoustic emission sensor with a vibration sensor with a microphone etcetera and also the current driving current being supplied to the spindle motor was also monitored. So, I will show you this current which was being supplied to the spindle motor was measured using an hall effect sensor because this had a variable frequency drive and one this rpm was 557 rpm the corresponding to a frequency of 18.56 hertz and if you look at the current which is being measured by the hall effect sensor which is driving the spindle motor you will see this current to be a sinusoidal signal and then a single frequency is shown in this current. Next is when we try to mill aluminum with dry cutting conditions and try to measure the vibration the cutting force and also the current and current also gets modulated because the load on the spindle motor and then you can see the side bands of the current. I will not go into details, but these are the features which could be extracted from this signals around this machining process being current force and vibration and once we try to machine steel in the same conditions the cutting forces the vibration levels and the current modulations are high and which is reflected in such a plot here. Now with this kind of analysis we thought that why not we instrument this machine with many more sensors which are more than required like voltage acoustic emission vibration cutting force in x y z direction sound pressure level and then we have a data equation system where when we do a signal processing and feature extraction by certain software and then also try to in the offline stage measure the tool wear condition by a microscope. So, we have the actual conditions and then they have an offline condition of the tool cutting tool wear and then we through artificial neural network we could try to estimate the tool wear and then try to find out at what tool wear and at what is the remaining useful life of the tool when we say a tool fails about in a 500 microns etcetera. So, how much time it would take for the tool to fail and this analysis was done through all the sensors taken together in the artificial neural network because it would be humanly impossible or physically impossible to have a mathematical model based system where in all these parameters are taken in and compared to the tool wear rather we thought of using artificial neural network and which is very powerful in the sense it is a black box which relates pretty well to the output level condition and the input level to the system. These experiments were done also the same tool used at a very very higher cutting tool depth of cut to bring in failure quickly and such levels were done in an industry where the actual the force level the cutting speeds the depth of cut etcetera were very very high and then you can see in the same industry how this spindle is being sensor is being monitored through and vibration sensor here through a force sensor sitting on here below behind the just below the work piece and this three phases you will see the detailed experimental conditions for such an estimation cutting speed depth of cut and then and you see there are these three phases which were done in the laboratory and then the severe phases were done in the industry and these are the typical values of the forces which we are measured in the laboratory force in the x component y component because this was a milling operation every cut there is to be a force increase and decrease vibration of the spindle in x y z direction and then the sound level and the spindle current once we also measure the spindle voltage and current we could multiply them to get the spindle power and these are all functions of time once the signals are obtained we could do some sort of a feature extraction and this is entirely done in a data driven method where you can do a time domain analysis or a frequency domain analysis but right now we will focus our attention into the signals in the time driven only and typical signal we may do some sort of a feature extraction by having a typical signal we have you filter it to remove the high frequency noises and then we have the loops brought together segmented signal we will remove this and this processes and bring in these loops and then we can again filter it so feature can be extracted from the signal low filtered signal and then this condition we tried to monitor the machining features being affects in this condition for different conditions of the machining time and how the for different conditions these forces varies according force affects as a function of time for the all the different conditions and these are actually estimated by a neural network similarly for F y prediction in the laboratory and then prediction in an industry similarly for the vibration in the x direction we have again machining time because once we know that the levels are higher or not allowable then we can also always predict the remaining machining time by which this curtain tool has to be replaced and this is for the current again machining time and for different phases this is for the machining features of the voltage so all these features could be extracted and it is up to you or us how to decide and what kind of signal features to extract or use for our data driven fault detection and this is for the power and similarly for the sound pressure level now once we have all this data and the features extracted we now will decide on what to do with this kind of data as I was telling you in the artificial neural network based for distance system we have certain amounts of inputs and certain amount of output in this case the inputs were the RMS force in x y direction the process parameters like the depth of cut the machining time cutting speed etc spindle rotation and then the spindle power and then we used artificial neural network with these conditions and try to develop an model and just to predict the force features we took number of data sets and the inputs are fx fy vcs0 we could estimate the tool where in the sense certain amount of training were done test data were used and then validation were also done and we can estimate that what is the this is the training with the test data and we have a good sort of training data with the measured values so we have established an allowable or I would say a reliable neural network model for doing such prediction and just to tell you that the filtering was done through a 4th order filter and in for the raw signal and the feature space it was a 2nd order filter so raw signal filtered signal and then single low filtered signal. So once the filtered features were used we in phase 1 phase 2 phase 3 and these are the cutting forces and there is a current and we could have the training of the neural network by these inputs so once we have the laboratory and industry results I we could estimate the tool were microns in microns for the laboratory case and for the case of the industry okay and we will know at what levels that this tool is to be replaced only through a such a model again we did again the sensor fusion in the sensor not relying on just only on fx fy or current but taking all of them together okay and then we are trying to estimate the tool where for the different cases the different parameters have been used and then we could do a better estimation of the model with unfiltered features somehow sometimes the stator is more so while using artificial neural network for a data driven for detection techniques it is always necessary that we do some amount of signal tree processing before applying such models and once we do the filtering you see this variations are much smoother and then we have a better prediction. So in this ANM this systems developed different strategies were implemented and force based strategies and for all the five phases gave us a prediction error of we have sensor fusion we could have cutting force we could have the vibration and we have the current okay. Now right in the beginning I had told you that we are also using parameters like acoustic emission overall sound because for a machine while we are trying to assess its condition we were pretty liberal in the time in terms of the sensors and instruments to be used and then during the process of measurements we saw that some of the parameters which are being provided or measured did not correlate well with the observations and there were certain random fluctuations okay. Now to avoid real estimation errors we did couple of things we did signal feature extraction in one case there was filtered data unfiltered so once we have the filtered data it eliminated the high frequency noise okay and then we once we have the unfiltered data the prediction was very good and we had a better prediction and we also avoided using the acoustic emission and the overall sound in the measurement method because the predictions were not up to the mark and in fact I will go back to one of the slides before. For example I focus my attention here there are wide amount of variations in the sound level with machining time and there is not a strong correlation between the measured sound level of course in a voltage level given here to the machining time and such a feature was actually not used because as you know sound gets easily contaminated from the background noise because of high presence of background noise SPL contamination occurs. So the signal to be monitored or to be carefully chosen that is very important so in this experiment of ours actually we though we measured acoustic emission and overall sound and in the latest stages we did not use them for the estimation of the artificial neural network because of the fact that they were contaminated and then we are not giving in signals which are only representative of the machine or the process because as I was telling you an ANN model if certain levels of inputs are given and I obtained an output if certain inputs are very erratic having lot of random errors I would not say that ANN model will not be able to predict the output correctly but I would like to have a robust model which would eliminate certain random errors so in this case we decided not to use the SPL and the acoustic emission signal and then if you go back to the cases of when we used FX, FY, current and voltage the predictions are much better compared to just measuring just the forces. So sensor fusion is also very important in the sense in this example you will see when the force based strategy test where for all 5 phases are conducted the prediction was 8% within 8% but only when current based strategies were used the prediction error was 14% which is very high but when we used sensor fusion using force, current and voltage the prediction was within 6.5 so prediction error is reduced okay. So in summary to do fault detection isolation and identification two important strategies are used one is the model based fault detection method another is the data driven or the signal based detection technology and in the example we saw that sometimes a combination of them is used sometimes just an feature extraction followed by a data analysis and then using a mathematical tool like an neural network can be used to predict the remaining useful life of the machine component. The example which we discussed we talked about the cutting tool single point cutting tool and then we decided that this single point cutting tool can have a remaining useful life predicted on the neural network model and in the case of cutting tool and now we used to the remaining useful life as 350 micron for example I always get this question posed by friends in the industry which is the right time to remove a cutting tool from the machining operation because sometimes people do this from the workers experience the operator they know their experience that the workpiece parameters the workpiece surface finish is not up to standard so then they are saying that something is wrong with the cutting tool it is time to change the cutting tool out of experience out of the machining time people have developed the strategies but if such simple neural network based models could be developed wherein we just monitor the machining parameters or the tool vibration or the cutting tool force put it in a neural network chip in fact nowadays modules are available with programmable functions for neural network wherein all this data which has been measured by signals can be put into this chip and this chip is going to give you a periodic warning as to know when this cutting tool is to be replaced and so on. So, fault prognostics or diagnostics has come a long way than the historical data from just an overall alarm level and the today's world we talk about using model based fault detection techniques using artificial neural network using wavelets and other signal processing techniques for doing a more robust detection of faults in systems and of course estimating the remaining useful life. And then in this example which we saw so a sensor based fusion for filter data also gives you the best value unfiltered and filtered so sensor fusion is another area which is being used for a robust fault detection and identification in machineries thank you.