Reducing Unplanned Repair Costs Using Predictive Maintenance with MATLAB - MATLAB
Video Player is loading.
Current Time 0:00
Duration 14:26
Loaded: 0%
Stream Type LIVE
Remaining Time 14:26
 
1x
  • Chapters
  • descriptions off, selected
  • en (Main), selected
    Video length is 14:26

    Reducing Unplanned Repair Costs Using Predictive Maintenance with MATLAB

    Maureen Barker, Urenco

    Urenco provides safe, cost effective and reliable uranium enrichment services and fuel cycle products for power generation within a framework of high environmental, social responsibility and corporate governance standards.

    Ensuring the safe medium-term storage of legacy waste in a controlled environment presents several unique challenges. These include reducing exposure to hazardous working environments and minimising costs associated with unplanned manual inspection and repairs. Predicting the expected failure time, or equivalently the remaining useful life (RUL) of a storage container requires a data-driven modelling and simulation approach, fully utilising inspection data which is often time consuming and costly to gather.

    Discover how Urenco applies statistical modelling, Monte Carlo simulation and predictive maintenance algorithms to model the RUL of containers in a secure storage facility.

    Published: 4 Nov 2024

    Good morning, everybody. And many thanks to our host at MathWorks for providing me with the opportunity to present how we are Reducing Unplanned Repair Costs Using Predictive Maintenance with MATLAB. So a short biography of my career to date. I completed a mechanical engineering degree at John Moores University in Liverpool, being one of two women in my first year and then reducing to one in the last few years.

    I then moved into a variety of different industries, including rail, developing maintenance depot equipment, water and sewage treatment primarily from a software perspective, doing PLC programming, and SCADA systems. And then followed by some mechanical works, looking at the inlet works equipment and scoping that.

    Moving into the building and construction industry, including involvement with the building services on the Spinnaker Tower, and finally into nuclear, where I am currently and have been for the last 20 years based at the Capenhurst site. I'm primarily involved with storage and decommissioning, including care and maintenance of the legacy building and equipment prior to decommissioning works.

    So a brief history of URENCO. The Capenhurst site was used as a nuclear facility essentially since the 1950s and primarily under BNFL and its various subsequent aliases. And then using diffusion technology to produce the product for nuclear power plants. URENCO then started its first centrifuge plant in building E21 in the late 1970s and had an output of 200 tons of SWU CO2 per annum, which is essentially just the measure of good uranium you can use in the power plant industry. And it took three years to commission.

    URENCO Capenhurst then became a wholly owned subsidiary of the URENCO Limited following restructuring of BNFL in 1993. And it changed its name to URENCO UK in 2008. UUK operates world leading centrifuge technology to enrich uranium for nuclear power stations around the world, providing electricity to millions. The site now operates three plants and has a total production output of 4,600 tons of SWU per annum.

    Each useful part of the uranium to be available for onward use in power stations also produces a waste product, which is stored in containers on the Capenhurst site. Our problem statement, URENCO needs to manage an aging inventory of these waste containers as previously indicated. And a scheduled maintenance for containers, which are close to the end of life, would enable the contents to be transferred into new compliant containers ahead of the current planned processing.

    However, this would expose the workers to potentially unnecessary hazards and high inspection and replacement costs. Gathering data is hazardous, time consuming, and expensive. Existing data from historical inspections has been collated and provides a start point to determine container condition. The question was asked as to whether the maintenance schedule could be improved upon by modeling and simulating the container condition as a function of time usage and environmental conditions within the store.

    This is a photo of our storage containers, showing where we have taken the historical data from, looking at wall thickness measurements and visual inspection data. So the four positions indicated are the top dome or the blank dish, the upper and lower body or shell, and the bottom dome and the skirt dish. The four positions indicate to the four checkpoints that were undertaken for wall thickness measurements prior to the containers being brought into the current store.

    This slide just gives a brief overview of the timeline of the life cycle of these containers. So they were developed in the 1950s, the first container being manufactured around about 1957 and with further versions manufactured right up until the close to the end of processing in the early 1980s. So you can see there's a wide range in ages in each of these containers. They were also manufactured from steel with a zinc coating.

    The containers were stored externally, which meant that they were subjected to weather conditions. They were then refurbished and re-coated with another zinc coating and then transferred to an indoor store where they have been for the last 25 years. And they are planned to be processed within the next two to three years.

    So onto the model. Prior to compiling the model, all the data from previous inspections had to be gathered and sorted into a usable form for the MATLAB to process. A number of spreadsheets have been collated over many years, providing detailed information on corrosion, so the pit depth. These have been gathered from both visual and mechanical inspections. The wall thickness, which all containers were measured prior to being placed into the store and after their refurbishment.

    Corrosion, so an additional methodology was used to undertake non-destructive testing of the amount of corrosion that we had on these containers. And this information was also collated within another spreadsheet.

    Then we also had the temperature and humidity of the store. It is a big warehouse store. So, therefore, it's just ambient temperatures. But it tends to be a little bit cooler than you would have in your normal ambient. All of this information is stored, as I say, in the Excel spreadsheets and were all brought into the model.

    So the first imports of the data carries out checks that the information is in a format that can actually be processed, so minor changes or minor inputs. A variety of inputs from various different people that have carried out the inspections over the years meant that the information is presented in slightly different ways, so capitalization. Some people will put all capitals. Others will put the first letter capital, et cetera.

    So the program checks all of this and indicates whether or not there are some errors or some areas of concern that need to be corrected as part of the cleanup stage. And then any missing data is then allocated a value based upon other local readings in a similar location so that all other data is then confirmed as OK prior to being put through the process.

    Our model equation. So the ultimate aim of the model is to see what wall thickness is remaining once the corrosion rates have been applied to it and the data has been put into equation shown and provide the final anticipated wall thickness with a critical wall thickness of approximately 0.25 millimeters at a time that we would need to then undertake the scheduled maintenance.

    All the equations for the corrosion elements are taken from the British standard ISO standards 9223 and 9224, which state that the equations are valid for up to 20 years, after which time, if left untreated, the corrosion would need to be measured by other equations. So these equations require the factors of temperature function, relative humidity, annual average sulfur dioxide, and chloride deposition rate in the air, the latter two being location specific. These are all important as all of the factors that cause corrosion can have the greatest effect on the corrosion rate of these containers.

    As described in the previous slide, the corrosion parameters are estimated values based upon historical data and up-to-date temperature and humidity values within the store. The diagrams show how the spread of the average thickness, based on the number of years of the container has been in use for with a linear fit graph of thickness over time. As indicated above, those with missing data are allocated a figure based upon the standard deviation of containers in the local area, which in turn ensures that simulated thickness aligned to a normal distribution.

    The graphs here show the results for the upper body positions, and similar can also be produced for all three other locations on the container. Overall, the results highlight that a small number of issues can be monitored in these particular graphs. But as you can see, the wall thicknesses vary considerably over both sets of graphs.

    We then undertake a Monte Carlo simulation due to the high number of uncertainties in some of the model parameters. This produces confidence that the model is robust across all the parameter space. This kind of simulation distributes possible outcome values using probability distributions variables. Variables can have different assumptions about different results that are recurring.

    The probability of distribution becomes a more realistic way of describing uncertainty in the variables of a risk analysis. When this type of simulation is performed, the decision maker within a project has a scope of possible results and consequently also assesses the probabilities of each result that may take place.

    Along with the main advantages of using Monte Carlo, the simulation technique can not only show the results of a problem, but also the probabilities behind each result, i.e. the probability distribution. Graphical results from the generation of data make it easy to create graphs of different results. We can sample from these results to produce a range of the possible failure times.

    So each container is modeled individually to determine the current wall thickness measurement and estimated time to failure based upon the critical wall thickness of 0.25 millimeters. We simulate the current thickness. And this captures the uncertainty in the existing measurements. The normal distribution parameters are cylinder dependent.

    We simulate the corrosion depths. And this converts the categorical ratings to numerical depth values and makes use of the depth measurements where available. The uniform distribution parameters are also cylinder dependent. The simulation is then repeated for each position on the container, top, bottom, upper, and lower shell. And the minimum failure over all four positions then provides an overall cylinder failure time.

    From the model, we then get an output. So the output includes the detailed prediction of all the cylinders and their predicted failure dates for both best and worst, as well as the median values, so that further examination can be undertaken, if necessary, to ascertain if more detailed inspection is required or if we need to implement the scheduled maintenance.

    As further data becomes available, this is also added into the model and can be used as-- this can also be used as a comparison of previous year's results against the current year's results. The graphs show figuratively and clearly what timescales to be aware of and can act as benchmarks for all our future operations.

    The clear and concise report format means that our regulators and our senior staff can easily see and understand the results. It has been invaluable in this respect by providing detailed information to our regulators, the Office of Nuclear Regulation and the Nuclear Decommissioning Authority, and give indication that none of the nuclear site license conditions are likely to be breached.

    For future updates of the model, we anticipate that we can include any future failure modes as they come apparent. And any input data is most likely to be stored as in the same format as the current data. We undertake an annual update to the base information based on ongoing annual inspections. And with the application of a user interface, we could then easily share the model with colleagues or regulators that do not actually need a full understanding of or need to be proficient in the programming of the model itself.

    [APPLAUSE]

    View more related videos