Enhancing Cathodic Protection Data Through
Advanced Analytics and Machine Learning

(Tony)

Thank you very much Rebecca, and welcome everyone to the webinar. We are very excited to have another opportunity to address the Materials Performance and AMPP communities, and in a small way contribute to the advancement of our sector and shared practices.

Today’s topic is ‘Enhancing CP Data With Advanced Analytics’.

(Tony)

Now onto introductions and a brief overview of the discussion planned for today.

My name is Tony da Costa and I’m the VP of Research and Development at Mobiltex.  I head up the team that comes up with creative ways to turn customer product research information into useful industry-leading products for our customers.  My background is in electrical engineering with a specialty in communications systems.  Over the last 3 decades, I’ve put that background to use in various digital radio and remote data acquisition product R&D efforts.  That includes many of Mobiltex’s current product offerings.

Now, I’m pleased to introduce my colleague, Mill Jawed, our Director of Client Services.

(Mill)

Thanks Tony for the intro.  I provide technical expertise internally within Mobiltex and externally to clients to strategize solve problems using RMUs. Lately, I have been involved with RMU deployment strategies with several clients. I have worked in the cathodic protection industry for over 13 years and I’m working towards my CP4. I have experience in rectifier manufacturing, compliance CP surveys, CP system installs and managing CP projects.

And that would be all about me, back to you Tony.

(Tony)

Thanks Mill.  Today we also have another colleague with us, Matt Barrett, our Senior Data Scientist.

(Matt)

Hi everyone, my name is Matt Barrett. My role at Mobiltex is to help make sense of the large amount of data that Remote Monitoring Units generate. I develop models and analytical tools to understand historical trends and try and predict future behaviour from Cathodic Protection data. My background is in experimental physics and after completing a PhD, I’ve moved into the realm of industrial IOT, using data analytics and data science techniques to make big datasets more understandable. You’ll hear more about some of these techniques later on in the presentation.

Back to you Tony.

(Tony)

Today we will present enhancing CP data through the use of advanced analytical methods.

To orient ourselves, we will start with a very brief review of cathodic protection practices, the assets, and common activities that largely influence how workflow is conducted in our sector. These assets present the opportunity for monitoring – with frequencies historically defined by regulatory practice. I’ll next provide a bit of a history lesson on remote monitoring applications within our sector.

With remote monitoring defined, I’ll examine blockers that currently limit the full potential of data, then Mill will layout the key elements of data collection and finally Matt will delve into the enhancement of data through the use of analytics.

(Tony)

Now for a quick intro into remote monitoring in CP.

(Tony)

Pipelines, and other large steel structures, are commonly protected against electrochemical corrosion, or rust, by coatings that are applied to the external surface. However, inevitable defects in the coating, called holidays in industry terminology, allow contact between the metal in the pipe and the surrounding environment, which leads to rusting of the pipe at the coating defect location. This rusting eventually causes perforation of the structure, which will always create an undesirable and potentially dangerous outcome.

To counter act external corrosion, impressed current rectifiers and sacrificial galvanic anodes are installed to polarize the pipeline and prevent the normal chemical reaction that causes rust at the defect locations. The assets in front of you are all common components of a cathodic protection system, specifically applied to a pipeline. The impressed current rectifiers and sacrificial anodes, acting in independent systems, are the assets that deliver energy to reverse the corrosive chemical reaction in a manner that is very similar to recharging a battery. Test stations are locations, typically spaced at increments of one mile or less, where a variety of readings can be taken easily. Bonds are locations where two or more pipelines are electrically bonded, ensuring that the CP on one pipe passes over to the other.

To validate the operation of rectifiers and sacrificial anodes and their effectiveness in preventing corrosion, measurements are made of critical parameters along the pipeline at test stations.  These measurements are compared against standard protection criteria values to gauge proper operation of the rectifier and galvanic anode systems. All of this is structured within the framework of regulation, designed to ensure that pipelines are operated in a safe and sustainable manner. Let’s look a bit deeper.

(Tony)

To ensure that this protection system is working, and providing sufficient protection to mitigate corrosion, operators are required to take readings of these common assets on defined periods. Critical assets, like rectifiers and critical bonds, play a more important role in the protection system and are therefore require readings every 2 months or so. On the other hand, test stations, casings, and other less critical assets only require readings once per year, even though you could argue that a P/S OFF potential on a test station, or depolarized coupon, is a better indication that adequate protection is being provided.

For many pipelines, these visits can be dangerous and expensive to facilitate, with travel to remote locations requiring crews of multiple people for safety reasons.

(Tony)

At this stage it would be exciting to tell you that these requirements are dramatically changing, and that the adoption of remote monitoring is a nascent practice in our industry. Unfortunately, it’s just not the case. Remote monitoring of CP assets may be one of the most mature and robust industrial IoT applications in the world. The need to gather frequent data from remote and hard to reach assets has long ago influenced operators to seek out help from technology. While reducing trips to site, these technologies have incrementally improved upon their visibility into CP system status and performance.

Let’s see how this evolution happened. Towards the later part of the 80s and early 90s, CP specialists started to explore how remote technologies could be used for gain insight into system performance. The 1st and 2nd generation devices utilized private radio networks. Data was stored on a local server, an old mainframe located near the pipeline or mine. These were power hungry devices, almost always requiring mains power, and were therefore not very small nor adaptable beyond fixed locations. You can see here a 1st generation Mobiltex product, the SMC4, designed for remote monitoring and interruption of CP rectifiers.

Around 2007 the technology had matured enough that 3rd generation offerings were brought to the market. These products were the first to leverage decentralized communications systems, such as cellular and satellite technologies, reducing the amount of infrastructure owned by the pipeline operator. Furthermore, web-based data management platforms were introduced to help ease access to, and management of, collected data. No on premise servers were required anymore. These devices were generally smaller, consumed less power, and some, like the RMU2 here in the photo, were battery operated. One additional upgrade in the 3rd gen was the introduction of Bluetooth technologies which advanced how RMUs could be configured more effectively.

I should note that the RMU2 remains a popular remote monitoring device for bonds and rectifiers to this day.

Fast forward to 2018 or so, and we see the introduction of 4th generation devices. Communicating on IoT-use case centric cellular & satellite communications, these devices are more robust, cover a wider range of applications and use-cases, and do so consuming less power than ever before. Ultra low power consumption enables two-way communication and control, enabling users to schedule GPS synchronized interruption of bonds using battery powered RMUs. The data is now stored in a commercial cloud platform, which enhances speed and accessibility while providing exceptional data security features.  Finally, G4 devices establish the backbone for scaling across a wider variety of CP assets and pave the way towards a future where more data is a good thing, in terms of being able to leverage in intelligent, data-driven analytics.

The 4th generation is largely where we sit today.

(Tony)

With these shifts in technology, different asset classes and applications have become more cost effective for remote monitoring deployments.

The critical nature of rectifiers, and the fact they often coexist with a reliable power supply, ensured they were the most commonly monitored asset during the first generations.  Bonds, being highly critical in nature as well, saw strong interest with the 3rd generation designs, with battery powered units becoming more feasible.

With 3rd and 4th generation technology becoming more reliable, robust, and cost-effective, assets like coupons, AC mitigation and pipe to soil test stations are a key focus across the industry.

(Tony)

So let’s zoom out to 10,000 ft to understand what a 4th generation remote monitoring system looks like. Application specific devices are installed unobtrusively alongside CP assets, helping automate data collecting and increase visibility. Two way communication allows commands to be sent to the remote devices, reducing manual effort to execute field interruptions. Data is sent to a cloud-based data management and analysis platform, which allows for manipulation, analysis of trends, and review of network alarms.

So with that – I hope you’ve got a good handle on the concept of remote monitoring as it relates to CP.

(Tony)

Next, we’ll look at how we can utilize all of this data that is now available from remote monitoring systems and other data sources that may be related to CP.

(Tony)

Let’s first start by discussing two factors that limit the value of the vast amounts of data that are being collected.  First, there is siloed data.  Unfortunately, many organizations lack a cohesive approach to data collection and processing between different departments.  How many times has a CP professional had to ask a different department for GIS data, asset data or integrity models?  When that data is received, is it in a consumable format?  Is it a snapshot, or a live feed?  In addition to internal data sets, third parties can provide useful data related to soil conditions, temperature, weather, seismic activity and many other parameters that can impact the safe operation of a pipeline.  Integrating all of this data into analytics models requires custom programming of API connectors and frequently reorganization of the data itself.  If done within the organization, commitment of resources is required at the highest level.

The second factor limiting the full potential of collected data is simply that it becomes dark data.  Dark data is collected, used for a simple purpose and then tossed into a physical folder, a file system, or some other storage area never to be seen again due to lack search capabilities, lack of time and resources to further analyze or even awareness of its potential when combined with other data sets.  In some cases, the collected is just stored–IBM estimates that 90% of all IoT sensor data collected is never actually used.  90%!

Better data governance through the use of relational or time series databases instead of Excel sheets for data storage goes a long way to keeping that data easily accessible beyond the initial use case.  Data transforms can be applied to easily format the data for other purposes, if it’s accessible.

Organizations spend significant amounts to collect such data, they owe themselves a duty to ensure that the effectiveness of the collected data is maximized.

(Tony)

How do we go about taking all of these individual data sets and create valuable actionable reports?  Here is where the magic of data analytics and machine learning comes into play.  Once connectors are established to each of the data sources, a function, whether statistical, simplified model or a trained machine learning model, combines the data to produce an output report with the actionable information.  These functions are typically created by data scientists with experience in statistical models.  A significant amount of work is put into perfecting the models so that the validity of the output is maximized.

Later on, Matt will present an example of a machine learning analysis technique that combines multiple data sources to provide a predictive indicator of ground bed operation into the future.

(Tony)

Multi-tenant systems, such as our CorView data platform, with common capabilities for many users can drastically improve the utilization of collected data across a large user base.   First, the age-old adage of why reinvent the wheel applies here.  With a multi-tenant system, the development of a particular analysis technique or report type is created once, perfected, and shared with all users of the system.  If each individual user were to try developing the same system independently, each would likely repeat the same learning experiences and errors in implementation.

Next, a common data platform allows for better standardization across industry.  If company A uses a different analysis technique than company B, are the results going to be comparable?  Perhaps, but likely not.  Analysis techniques on a multi-tenant system are created with input from multiple stakeholders, resulting in the best ideas from multiple users being used to create those techniques.

Cost is another factor in the inclination towards the use of multi-tenant systems.  Not only would individual organizations be reinventing the wheel, but the development costs to create independent data platforms would be astronomical.  To implement advanced analytics tools requires expertise in data science, user interface design, coding and testing.  In addition, the systems require an underlying computing infrastructure.  With multi-tenant systems, the development costs are born once by the service provider, but the service is provided to a multitude of customers.  The service provider has an inherent motivation to continue innovating by providing new leading-edge capabilities across the entire customer base.

Finally, machine learning algorithms require training and testing.  The more data that is input, the better the algorithms will perform.  Often, especially with smaller pipeline operators, there may be insufficient data available to adequately train algorithms.  By allowing anonymized data to be used from all users of the multi-tenant system in the training process, training accuracies can be vastly improved, allowing the model to then be applied against any customer’s data with dependable results regardless of that customer’s data set size.  Data confidentiality is still maintained, but the benefit of the larger data pool is seen by all.

Ultimately, a comprehensive and cohesive analytics platform, rather than a patchwork of softwares, is necessary to yield reliable visibility and quality actionable output.

Now I’ll turn it over to Mill for a discussion on how some of the needed data is collected in the field.

(Mill)

Thanks Tony, as you mentioned, we’ll look at a comprehensive cathodic protection system and then a portfolio of remote monitoring products which have become reliable data gathering sources.

(Mill)

Before diving deep into the illustration, as you can see it is showcasing various cathodic protection scenarios ranging from impressed current systems to sacrificial systems and their pertinent items. There is a cased road crossing, pipelines parallel to each other, and pipelines crossing each other. AC mitigation components like ribbon anode, SSD and AC/DC Coupon, and last but not least, a test station with good old permanent reference electrode with several pipe connections. These assets in front of you are all common important components of any cathodic protection system.

To validate the operation of rectifiers, impressed current and sacrificial groundbeds, data measurements are made of critical parameters along the pipeline at test stations.  These measurements are compared against standard protection criteria values to gauge proper operation of the rectifier and galvanic anode systems. All of this is structured within the requirements of regulation, designed to ensure that pipelines are operating in a safe and sustainable manner.

Let’s look a bit deeper. To ensure that a system is functioning, and supplying sufficient protection to mitigate corrosion, operators must take data reads of these assets on defined periods. Critical assets, like rectifiers and critical bonds, play a more important role in the protection system and therefore require readings every 2 months or less. On the other hand, test stations, casings, and other less critical assets may require readings once per year.

For many locations, site visits can be unsafe, expensive or both to collect useful data that requires higher data collection intervals for analytics purposes. At this stage I am thrilled to tell you that these requirements are dramatically being met, and that the adoption of remote monitoring is becoming an increasingly normal practice in our industry. The need to gather frequent data from assets has long ago influenced operators to seek out help from technology. These technologies have incrementally improved upon their visibility into CP system status and performance.

(Mill)

Let’s start off with the main components of an impressed current cathodic protection system, a rectifier with its groundbed. At a minimum, we need the rectifier’s total output voltage and total current. Better yet, we would like to capture individual negative or positive circuits if more than one was designed. On the left, you can see the RMU3 rectifier monitoring and control product that is comprised of a measurement block that is connected through a simple instrumentation cable to the integrated transceiver-antenna block.  The middle picture shows, if more than one rectifier is in the same location, an RMU3 MUX can be utilized to capture up to 16 data points such as separate outputs or multiple negative or positive circuits. Further diving into the data sets from rectifiers comprising of up to 44 output circuits, an RMU3 XL unit can be utilized as shown in the right picture.

(Mill)

Now that we have covered the impressed current cathodic protection system, let’s touch on data from the critical bonds such as interference bond panels, current bonds and other non-critical bonds. We need to track the authenticity of bonds by monitoring directional current flow and open and closed-circuit potentials. In the left picture, for better life cycle trending, we are using a battery-operated device with over 7 years of battery life called RMU1INT1 in either a test station or a junction box. The right picture shows an RMU1ER for corrosion rate and metal loss data used with an ER probe in various applications with one of them being encased casing with corrosion inhibitors.

(Mill)

Here we have a simple solution for the structure to soil potential readings. An RMU1 or a cost conscious solution is the RMU1 Lite which captures both ON and instant OFF readings. If multiple interruptions are affecting the structure the RMU1 Lite captures the most electropositive reading thereby capturing the true OFF reading.

(Mill)

Now that we have covered the impressed current cathodic protection system, let’s touch on data from the sacrifical cathodic protection systems. We need to track the health of sacrifical anode bed by monitoring its current output and open and closed-circuit potentials. There are instances where sacrificial systems are part of impressed current systems and also need interruption to obtain true OFF potential data, RMU1INT1 is utilized with synchronous interruption. For more clarity on the groundbed individual anodes, an RMU2 is utilized to record the current flow from each anode. The unit is battery operated and therefore can be installed right inside the positive or negative splitter panels.

(Mill)

Here we have a summary table showing the portfolio of product line that collects useful data of CP system in its entirety. This table is included as a download from the resource section of this webinar. This is enough from me, so I would like to pass it on to Matt.

(Matt) Thanks Mill. Now that we know about the different types of data we have access to, let me talk about some of the techniques we use to make sense of that data.

(Matt) I know there’s a lot of excitement around Machine Learning based techniques, and we’ll get to that later. First I’m going to go into some detail about how we can use direct analysis techniques to better understand large amounts of data. These techniques are more about data organization and simplification than using complex ML algorithms, so it makes sense to start here. I’ll break this subject down into these 2 sections.

(Matt) First of all, statistical predictions and exceptions. The idea here is to use the accumulated historic data from a site which has been capturing recordings for a long-enough time period. Tomorrow’s measurement should look statistically similar to the measurements from the previous year, except if something physically changed on site. We define the expected range using some simple stats, calculating the mean and standard deviation of the measurements. For example, starting from the mean and adding and subtracting three times the standard deviation should give a range wide enough to capture 99.7% of new measurements. If your new measurement is outside of this range there’s a possibility that something physically happened on site; a rectifier tap was changed, a cable was broken, or a holiday in the coating has started to grow. Here’s an example of rectifier DC current and DC voltage, where this principle has been applied. The green lines are the statistical expectation range and you can see that the voltage is dipping below the lower line. This approach can be very powerful for large deployments of RMUs, since the calculations can be performed automatically, and exceptions will be caught in real-time.

(Matt) Next simplifying complex datasets through modelling. The goal here is to describe your data with the fewest possible parameters, which you can later use to compare and group different datasets to better what understand factors are at play.

(Matt) One of our most successful modelling applications so far has been modelling rectifier resistance. We begin by using Ohm’s law to combine Rectifier DC voltage and DC current into system resistance. A typical rectifier’s resistance measurements over the course of a year will change in response to the temperature and moisture variations of the soil, looking like a sine wave.

(Matt) If you think back to your trigonometry classes, you will remember we can describe a sine wave with 3 variables. The amplitude, period and phase. Amplitude tells us about the difference between the maximum and minimum resistance over the year. The period describes the time between two maximums and for this model is set to exactly one year, so we can ignore it in this case. The phase will shift the peak left or right, and in this case we use a cosine instead of a sine so that phase actually tells us the calendar date with maximum resistance. Additionally, we can add a linear component, with two additional variables describing a change of resistance over time and the average resistance of the system. In this example on the right side we take weekly measurements in blue dots and fit the data to our model with a green line. The x-axis here is calendar date, and y-axis is resistance.

(Matt) Now that we have a model which simplifies our datasets, how can we use it to make our lives easier? For a single rectifier system, we can use these values to inform when site visits or annual surveys might be completed. We can extrapolate between annual survey results, being more informed about how the system’s resistance varies throughout the year. The model can be used in conjunction with the statistics discussed previously to create a more sensitive expectation range that varies with the seasons.

(Matt) The real power of this approach is accessed when you apply the model to a large number of units. Comparisons can be made between different rectifiers, different pipelines or even different parts of the world. We can use this approach to understand what factors influence the rectifier system, and learn from other units to anticipate problems which may arise.

(Matt) Here’s one example of such a comparison. We’re using the phase variable, the one that tells us the date of maximum resistance, comparing hundreds of rectifiers across the United States. The colour of the marker on the map tells us when the maximum occurs, with red being closer to January, and blue being closer to December. The histogram on the right summarizes this, with the most common maximum system resistance occurring early March.

(Matt) Now let’s get into a few machine learning examples. We’ll talk about how machine learning can help us to combine and compare different data sources, and how we can make predictions of future behaviour.

(Matt) One of the advantages of the machine learning approach is that an endless number of features can be combined into a single model. If you don’t know the influence of a particular feature, add it to the model and see what effect it has. Our team began by introducing various datasets to the model which have been shown to have some influence on cathodic protection. We can add different publicly available data-sources such as historic weather forecasts, soil measurements, or even the location of AC Power lines or telluric readings in the case of AC mitigation. One important dataset will be the set-up of the rectifier-groundbed-pipeline system. We can gather details about these different components to see what effect something like anode depth or pipe diameter might have on CP readings. Finally, we can take CP measurements recorded using different techniques or at different times of the year to get a more complete picture of what’s happening on-site.

(Matt) In this example we combined soil temperature classifications from the US Department of Agriculture with the amplitude parameter of the cosine model. The histogram in the middle is split into two different soil temperature classes – hyperthermic in red representing regions in the southern US, with a mean soil temperature above 22C, and frigid in blue for regions with mean soil temperature less than 8C, often regions close to the Canadian border. You can see that the distribution of hyperthermic (red) is much tighter and closer to the left than the frigid in blue. This type of “separable” feature is a great candidate for an input to a machine learning algorithm. You can read more about how we used various soil classifications to create a model which predicts the seasonal parameters in the paper published for the NACE Corrosion conference in 2021, and you can find a link to the paper in the resources page.

(Matt) In another project we used different rectifier-groundbed-pipeline features which were provided by a pipeline operator, including pipe construction year, groundbed depth and diameter, soil types and rectifier type. In total there were 40 features considered. We wanted to see how these features would influence the long-term trend of rectifier resistance.

(Matt) To simplify the problem we assigned our historic resistance readings one of three classifications, steady, increasing or decreasing. We then trained the algorithm to predict this classification with only the pipeline operator’s feature list, without ever showing the algorithm any rectifier output measurements.

(Matt) Of the 800 or so rectifiers which were considered in this study, 60% showed a steady resistance trend, 30% increasing and 10% decreasing. You can see this is well distributed across different geographic areas, which means that longitude or latitude alone would not be sufficient to predict the trend.

(Matt) Finally, we trained the algorithm by associating the 40 features and the trend classification of steady, increase or decrease. We used a decision tree-type algorithm called xgboost which tries to maximize the accuracy by trying different combinations of features at different levels of the decision tree. We found that we can predict the trend with an 80% accuracy. This project was presented at the AMPP 2022 conference, and you can find a link to the paper in the linked resources page.

(Tony):

Thanks Matt.

Mobiltex designs, manufactures, and assembles all of our products in our new offices located in Calgary, Alberta, Canada, in the same building that houses our world-class customer service, sales, finance and marketing teams. As I explained earlier, we started developing remote monitoring products for the cathodic protection industry 31 years ago. We’ve evolved alongside our customers and partners, to now offer best-in-class remote monitoring and field-survey products for the cathodic protection and pipeline integrity industries, spanning oil & gas, water, power generation, and civil infrastructure sectors. Our products are depended on by many of the largest pipeline operators and distribution utilities in North America and around the world.

(Tony):

Over the years, we have utilized capabilities from the world’s foremost technology companies, allowing us to deliver the leading-edge products in our portfolio.  Through our industry relationships, we maintain open dialogs with leaders in the industry that have helped shape those same products.

(Tony):

We are very proud to partner with Cathodic Protection specialists across the world, and the logos on the screen now represent just a selection of our partners in North America. Please reach out to one of our partners, to Mobiltex customer reps and our service team, or to Mill, Matt and me directly, in case you have any questions and inquiries into our technologies and services.

(Tony):

We will now open-up the floor for questions.