Industrial Analytics: Condition Monitoring of Wind Turbines and Preventative Maintenance Using JMP (2021-US-30MP-867)
Alan Rezazadeh, Analytics Research Scientist, Southern Alberta Institute of Technology
Condition-based monitoring (CBM) on large sensory data sets requires complex specific capabilities, detecting gradual change over extended time in industrial processes. Application of CBM using SCADA (Supervisory Control and Data Acquisition) is an emerging big data analytics practice to detect the gradual degradation of equipment and find the most effective time to repair equipment, which is also a component of preventative maintenance discipline.
JMP offers a rich set of statistical and machine learning tools that can be used for wind turbine condition-based monitoring. Wind turbines that convert turbulent and uncertain wind energy to highly regulated electricity is a multivariate process with many parameters, subcomponents and interactions that require advanced analytical capabilitiesto detect subtle changes within the operations. This paper uses JMP to build wind turbine bearing temperature predictive models. After comparing predicted against the actuals, residuals are used for condition monitoring. The residuals resulting from the predictive models are used within JMP process control methodologies to quantify wind turbine degradation for decision making and monitoring. It also includes applications of advanced statistical tools such as PCA, clustering, neural networks, and non-normal distributions for data mining and process control within a wind farm.
Auto-generated transcript...
Speaker | Transcript |
| Alan Rezazadeh | Hello everybody, thank you for joining this presentation. My name is Alan Rezazadeh, working with Southern Alberta Institute of Technology. My position is |
| a research analytics scientist. We are going to talk about industrial analytics and as an example, focusing more on wind turbines and condition monitoring of wind turbines. | |
| Acknowledgments. | |
| CFREF, Canada First Research Excellent Fund, thank you for allowing me to participate within this conference. | |
| EDP Renewable Energy Wind Farm from Spain, thank you for sharing operational data, which is freely available for research and education. And above all, JMP for building this fantastic statistical discovery software. | |
| Agenda. We are going to talk about the scope of industrial analytics, what's included within industry analytics | |
| and how we would consider industrial equipment as a process; preparing and understanding data; and defining subcomponents, how find the goal in defining subcomponents; and at the end will be monitoring degradation and performance, which is called condition-based monitoring. | |
| So a scope of industrial analytics, we have very much everything included. So we have large time series of data coming in by sensors, which is IoT. | |
| And we want to do process optimization, knowledge discovery, decision making. Basically, we want to reduce pollution, we want to increase the output good output, such as power from wind turbines or from solar panels, or we want to monitor the health or condition of industrial equipment. | |
| So the first idea is a complex machinery, such as a gas turbine jet engine is basically a process. There is input, there is output. The input, in this case, which is a gas turbine, is air and fuel; the output is electricity and exhaust, some kind of fumes and gases. So we would break this | |
| a compressor, a combustion chamber, and a power generator. It's the simplest case. So now we have a process and subcomponents. So from now, I may change the term process and equipment interchangeably with them together. | |
| So how far we go in breaking down the process? So we want to break it down into smaller simpler subcomponents that we can actually describe it by a mathematical physical formula, such as aerodynamics or thermodynamics and so on. | |
| So first principles, from which a thing is known, Aristotle. So we're breaking down into smaller components. | |
| And this is a breakdown cycle for jet engine or also for gas turbines. So as we can see, over time, there will be some changes, some modification, some degradation if we call it. | |
| Also it's not only because of the wear and tear. It could be due to operational changes, so the air temperature would change, the electrical demand from the grid will change. | |
| As a result, the operational mode of the equipment will change. So now we want to differentiate between operational modes and degradation, so that means we have to know what is operational mode first. | |
| Let's say for a | |
| wind turbine, we have rotor, | |
| low speed shaft, we have gearbox, generator and a transformer. | |
| It's a complex system; we break it down into smaller subcomponents. | |
| hydraulics, | |
| lubrication, cooling and so on. So the input to the process is wind, output is power. So let's see it, wind versus power, which is power curve in wind energy industry. So on the | |
| X axis, we have a wind energy, the wind speed, and on the vertical axis, we have the power. So it is a typical power curve that is used for analysis of wind turbines. As we can see, it's a complex system. So at four meters per second, | |
| the system starts working. Before, it is in idle mode and it's waiting for the wind to blow. At four meters per second, it's the minimum wind speed that it can operate, | |
| so slowly it will start producing electricity. Above 12, it's getting strong, so they have to reduce the amount of energy they harvest from the wind, and that is done by changing the blade, angle of the blades. | |
| Above 25 meters per second, operating the equipment will be unsafe, so it gets turned off. | |
| So we have a live demo now, | |
| and | |
| we are going to produce the power curve. | |
| So | |
| wind speed versus power. So as we can see with the contour, the high concentration is around idling mode and lower speed of wind. So this is the wind regime in the system. | |
| And as the wind speed speeds up, it's getting stronger, more electricity is produced. As it becomes too strong then it's turned off. We can see the histograms. If you want to know how to make it on | |
| JMP, it would be quite simple. When the speed as X factor and the power for Y, grouping them by | |
| turbines. | |
| So getting back to presentation. | |
| Now we are starting to analyze the data. First is the univariate analysis, so we would analyze the main parameters, which is, let's say, wind speed. | |
| They would research the existing literature and existing engineering knowledge that if there is any mathematical and statistical model describing the wind. | |
| And there is one, and wind distribution usually conforms to Weibull. So then we would compare the wind against the Weibull. We would see that it's very close. | |
| However, between different turbines there are some minor ??? differences, so there is a little opportunity, research optimization opportunity. | |
| So Turbine 9 is different than 6, so the wind from 6 and 9 are different, so it could be due to interference, interaction between these turbines, which is fine. It happens in the industry. | |
| And it's an optimization opportunity to reduce this wake and it's a big well-known topic within the industry. | |
| So after univariate analysis, we go to typical data science | |
| preprocessing. Missing values, are there zeros? Are the zeros really meaning zero or they are missing values? So we handle those; it's not a topic of this talk. Outliers, we'll talk about outliers a little bit. | |
| So | |
| in my projects, we remove any value that is not physically possible. So if the temperature...air temperature is recorded 100, you know we don't have 100 degrees, so we remove it. But we would not change the other data based on our own judgment, so we keep them very much. | |
| And date/time convergence, so we do some date/time conversions to fit into JMP format So these are very much a standard. So one method of univeriate, | |
| multivariate outlier analysis is Mahalanobis, and it is nicely supported at JMP and we are going to look at it. So basically if we have four parameters, we produce one output | |
| and we look at this one. So Mahalanobis and these are the outliers. We are not going to remove these outliers but, however, these are the centers of focus and attention that we would look into them. So for live demo, | |
| using | |
| Graph Builder, we can easily plot them against the Mahalanobis and checking different turbines. | |
| So, for example, we have one little spike in here. It's above the upper control limits, so it is an outlier. So we check into it to see what happened, and you can see the parameters, how they were changed. | |
| In order to produce this Mahalanobis distance, we can use the multivariate features of JMP, | |
| producing | |
| under outlier analysis, Mahalanobis, and then saving them into the data. | |
| So getting back to presentation. | |
| So after we have finished the pre processing, | |
| we would start the data discovery. We want to see what data is telling us, so I gave a little background of power curve, of how the flow was working. So let's formalize it a little bit. | |
| At the beginning is idling, so waiting for a stronger wind. So there is something wind. We start. Rotary speed increases, there is more wind. | |
| We connect to the grid, that means the turbine is connected to the generator. The generator has enough speed now and it gets connected to the grid and produces electricity. | |
| Wind speed increases. We are producing good amount of electricity, but it is not at maximum. As the wind gets a stronger, we get to the maximum level of | |
| electricity production. As the wind gets still stronger and it's too much, we have to reduce the energy we get...we harvest from the wind, or we may want to stop completely because it's too strong. | |
| As conditions settle down, wind is slow again, we go to idling. So it's a loop; it's a flow chart. Now let's do it in with the data. | |
| So here is a three-dimensional data parking mode, we are waiting, idling. We are waiting, parking, we are not operating. | |
| Rotor picks up speed producing minimum electricity. That's connecting producing subrated | |
| and rated. And if it gets too strong, then they would go to this parking again and turn the operation off. So what we have here is clustering | |
| and comparing AICs. So if we have one cluster or two or three, we would compare the cluser...the results of clustering and we would pick up the number of clusters that give the | |
| best...I can't say the best...the minimum error, while the number is not too high. So let's see it in action, how it would work. | |
| So we are doing clustering and clustering up to 10. I didn't want to take it too long within the presentation. It takes about 10 to 15 seconds. So we are finding the clusters within Turbine 1. | |
| So | |
| each cluster is basically the concentration of | |
| perational modes so, then we would know how this turbine was working. So we have the AIC versus the number of clusters. | |
| I would | |
| plot the graph | |
| AIC versus number of clusters. | |
| So it's not as smooth as it should because I did use a number of iterations and tools and made it fast. However, for the production and for the actual work, it would take longer time. | |
| So, as we can see on 6, 7, 8, it's a good number of number of clusters, while the error has dropped. The errorwill marginally drop more | |
| if we continue with more clusters, however, it would add to the complexity and if we increase it too much, it might become overfitting. | |
| So | |
| now | |
| we plot them. | |
| So this is for Turbine 1, and as we plot them, we can see, there are some transitions and that's when we are going from high wind to parking because of the wind is too strong. | |
| So getting back to presentation. | |
| Now, how many cluster are the right number? So if there is one cluster, everything is one group, okay. It's not our concern. | |
| If there are two clusters, we can see the algorithm is picking up this group here. We call that idling, just waiting. Or three clusters, it is parking, rotor picks up speed and operations, so algorithm picks up this three. | |
| If we look into six, seven, or eight they make most sense. I picked up six for it it's a small number for this presentation. An engineer of the operations may pick up more, eight or nine of them. | |
| So a live demo, and we already saw this one. | |
| hierarchical, K-mean and normal. For industrial projects, I usually use normal mixtures. | |
| For business projects, hierarchical would make quite sense, because it gives a good breakdown of the logic. K-mean is computationally intensive and on large data sets it would be too time consuming. | |
| So the assumption with normal mixture is the clusters are convex under normal distribution, so that means they only have one peak. And it's very much the case. | |
| In the...when we have one operational mode, it's usually operates within that operational mode within one center, so normal mixture is a safe | |
| algorithm to use for the clustering. | |
| Now, getting to condition monitoring. We did all the clustering, okay. What are we doing with them? | |
| Let's talk about what they are going to monitor here. | |
| As the gears in the gearbox work, there is a touch point in the gearbox. Due to wear and tear, this touch point may shift over time. That's what we want to find. However, | |
| if the ventral line is working hard and there is so much torque into it, the touch point might be different than when they torque is not so strong and | |
| the wind is not as strong. So that means again, operational mode. So we want to compare this touch points within one operational mode, not when the operational modes are too different from each other. So for this | |
| monitoring, we can use a simple linear regression. | |
| So doing a linear regression | |
| based on different operational modes. So for grid connecting, | |
| R squared is 99.6%. We can see it's a good approximation. It's a nice regression line. | |
| 99.993. It's a good R squared, so we'll use this one. | |
| For rated operations, | |
| we did not get a good R squared, so we will exclude it. So practically we are going to put together for this operational mode that are giving us a good R squared for the analysis we want to do. Now getting back to demo. | |
| This is basically the old linear regression. It happened in just a matter of a few seconds. As we can see, there are 521,000 records and it was just two-three seconds, so it's excellent performance. | |
| So these are the operational modes we are going to select. Let's have one more look again to just confirm we are picking the right values. | |
| Fitting Y by X, we would plot the density contours. So what we like to see here are the regression line to pass through the centers of the density, and this is passing. It's a nice one. | |
| It is going to miss it, and that's the idling we chose not to use. | |
| It's a nice regression line. | |
| For rated production, it's missing the density points, so we aren't going to use this one. | |
| It's very much confirmed the observation that we had had with R squared. | |
| So we did this. What's coming next? | |
| We draw the residuals. Residuals of turbines. | |
| And we would analyze the residuals when the residuals are increasing. So let's have a live demo. That would make sense. | |
| So we are plotting the residuals for all the five turbines, only for the selected operating modes. So idling and rated, we exclude them. What happens if we plot the other ones? For idling, it's all over, it doesn't mean anything, isn't telling us anything. | |
| Same thing with the rated production. It doesn't say anything, it doesn't say much. | |
| However, with this together, | |
| we can see the residuals increased on some specific dates. And actually if we look at the log of the repairs, this was one repair, that was another repair and we see a nice gradual increase here. That was one more repair. | |
| Turbine 11, we didn't have problem with gearbox. It never...the | |
| residuals never increased. | |
| And there were repairs on Turbine 6, so this may show some issues that happened, but what was it... | |
| at what level...let's say at what level of residual is a real problem. | |
| It's subjective. We have to quantify. We have to look at what level of residual an engineer needs to take some actions. And at this moment, it's an open question very much in industry, an open subject for research. However, it is a tool used for | |
| helping engineers. So getting back to | |
| our presentation. So the trends are for monitoring, and at what level an engineer would take action and actually interfere and open the gearbox. That's an engineering decision. However, from a data perspective, you're helping the engineers to make decisions. | |
| In summary, a complex process has multiple operating modes. When we want to monitor | |
| a device, we have to know which operating mode we are going to monitor. Each operating mode highlights a portion of the operational processes. | |
| Clustering methods, we can use for isolating the different operating modes. Condition monitoring, using operating modes for detecting degradation, as we did. Monitoring subcomponents | |
| in consideration to physical law, as we use for the ratio. And if the system is complex and we cannot describe it with one physical law, then we would use other methods, such as | |
| K-nearest neighbor or the neural network. And it's very much at the end of a presentation, and thank you for your attention. |
