<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Alexandra Gunderson on Medium]]></title>
        <description><![CDATA[Stories by Alexandra Gunderson on Medium]]></description>
        <link>https://medium.com/@alexandra.gunderson?source=rss-bbc3114804ff------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*S3LRjN0HpCMqLvMbueutHg.jpeg</url>
            <title>Stories by Alexandra Gunderson on Medium</title>
            <link>https://medium.com/@alexandra.gunderson?source=rss-bbc3114804ff------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 01:27:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@alexandra.gunderson/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Industrial Data Crisis — Overcoming the Challenges.]]></title>
            <link>https://medium.com/unifai/the-industrial-data-crisis-overcoming-the-challenges-49f774537598?source=rss-bbc3114804ff------2</link>
            <guid isPermaLink="false">https://medium.com/p/49f774537598</guid>
            <category><![CDATA[data-preprocessing]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[industry]]></category>
            <category><![CDATA[data-labeling]]></category>
            <dc:creator><![CDATA[Alexandra Gunderson]]></dc:creator>
            <pubDate>Wed, 14 Jul 2021 11:27:32 GMT</pubDate>
            <atom:updated>2021-07-14T14:16:11.424Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wD0b-a4XhhJ9UYZn0gGD8Q.jpeg" /></figure><h3>The Industrial Data Crisis — Overcoming the Challenges.</h3><p>Advances in AI and machine learning have been tremendous in recent years and impact our lives in many different and significant ways. We’ve seen self-driving cars become a reality, we are able to communicate with our homes using Alexa, and we are even able to write code with AI thanks to tools like Github Copilot.</p><p>Naturally, we expect similar advancements within heavy industries. Predictive maintenance has been hailed as the Holy Grail for industries including oil &amp; gas and maritime, where equipment failures are costly and safety is a priority.</p><p>The reality is disappointing, and that is that we have fallen short.</p><p>We’re constantly told that the technology is there — look at AlphaGo, DeepMind or Libratus consistently outperforming humans at more complex tasks.</p><p>Yet, in asset heavy industries, where AI has the potential to transform and improve operations, we’ve barely made it out of anomaly detection pilot purgatory, largely because the data available simply doesn’t lend itself beyond that.</p><p>The predictive maintenance revolution will be fuelled by data, but that <a href="https://www.linkedin.com/pulse/datasets-beat-models-lukasz-mentel">data needs to be carefully curated, annotated, and augmented</a> before it can harness any sort of predictive power…and building those datasets is very challenging.</p><p>At Unifai, we’re working on addressing these challenges by building the tools and methods necessary to develop robust industrial datasets. We believe that great innovation comes from collaboration, and that equipment operators and data scientists should be empowered to leverage their own data quickly.</p><h3>Challenge #1 — Timing 👵👴</h3><p>Heavy industries are ripe for disruption, just like finance, health and transportation before them. In fact, it’s one of the major appeals to apply machine learning — allowing operators to improve their throughput, energy consumption, and reduce emissions.</p><p>However, age also provides a major challenge, because many industrial assets (oil platforms, ships, hydropower plants, …) were designed, built, and put into production in an entirely different century…long before the cloud!</p><p>To put their age into perspective, here are a couple of fun facts:</p><ul><li><strong>Hydropower plants </strong>— the average age of hydropower plants operating in Norway today is <a href="https://www.hydropower.org/country-profiles/norway#:~:text=The%20current%20average%20age%20of,as%20well%20as%20environmental%20improvements.">46 years old</a>.</li><li><strong>Oil &amp; Gas — </strong><a href="https://www.norskpetroleum.no/en/">75%</a> of the remaining reserves in Norwegian fields began production before the year 2010. Half of the remaining reserves are from fields put into production before 2000.</li><li><strong>Shipping </strong>— The average age of general cargo ships is <a href="https://www.statista.com/statistics/1102442/age-of-world-merchant-fleet-by-vessel-type/">26 years old</a>.</li></ul><p>It’s clear that we have to be pragmatic and understand that there are inherent difficulties when it comes to retrofitting and collecting data from such installations.</p><p>But, there is also a huge upside here — tons of data often spanning a long period of time, and seasoned veterans that are closely familiar with the equipment and industries to learn from.</p><p>When life gives you lemons…🍋</p><h3>Challenge #2 — (Too Few) Labellers 👻</h3><p>Labeled data doesn’t happen by chance. It happens with people. But, who is labelling the data?</p><p>Many organisations have been able to gain a huge data advantage by outsourcing their labelling needs to so-called <a href="https://mitsloan.mit.edu/ideas-made-to-matter/hidden-work-created-artificial-intelligence-programs">ghost workers</a>, who label mostly images and text for the specific content that will be used in model training.</p><p>But, identifying objects in images (like cats, cars, dogs, and trees) is a relatively simple task, and you can build up an enormous dataset due to the sheer number of people who can label this type of content.</p><p>On the other hand, what do you see when you look at this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eHDGQKXZcceHfHY7" /></figure><p>If I asked you to point out when there was a failure, would you be able to point it out?</p><p>You would probably first ask for other information, like the sensor metadata (is it a temperature, pressure, speed sensor?) or contextual insights (when was the equipment shut down? when was maintenance performed?) to be able to say anything meaningful. All of this information in parallel is what would allow a specialist to differentiate between blockage, valve failure, sensor reading errors, etc.</p><p>Hoping that a data scientist can label and curate their data set on their own even with all the aforementioned information is still excessively optimistic. Data scientists are experts in statistical modelling — they aren’t equipment specialists or operators. We need to put aside the right resources to label complex equipment data.</p><h3>Challenge #3 — Infrequent Events 💥</h3><p>Equipment is, fortunately, built to fail <strong>very rarely</strong>. Rare is relative, so let’s look at an example using a compressor.</p><p>Compressors are used to increase the pressure of the fluid that they are processing. They are used in the oil &amp; gas industry (among others) throughout the extraction, transportation, and refining processes.</p><p>They are massive and critically important to these processes, and if they fail, you better hope you have an equipment specialist available. Because, while it might only fail once a year, that failure could be any of 33 different failure mechanisms, 20 different failure modes, and any of 60 components could be responsible for them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7ISDKx0pvT_xb7L7" /></figure><p>How can you build an algorithm to predict with any certainty which of these modes and components should be addressed?</p><p>It’s hard to say exactly what is the <strong>right number </strong>of examples you’d need in order to build such a model, but if we assume we need at least 10.000 examples of each scenario, then we’ll need quite a lot of data:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3Orl_kTS7YZS--f2" /></figure><p>This definitely feels daunting — what are you supposed to do? Wait hundreds of thousands of years before you get started with machine learning?</p><p>This type of thinking has led to a <strong>big data paradox</strong>, where we believe that more data means that we’re more likely to discover new and unknown relationships. The truth is <strong>we’re much more likely to succeed if we inject the data with expert opinion.</strong></p><p>We’d also be saving the earth in the process — did you know that training large models has a huge carbon footprint due to their significant energy consumption? In fact, it was found that the most wasteful model can emit more than <a href="https://www.technologyreview.com/2019/06/06/239031/training-a-single-ai-model-can-emit-as-much-carbon-as-five-cars-in-their-lifetimes/">626,000 pounds of CO2 equivalent in its lifetime</a> (five times the lifetime emissions of the average American car!!). 🌍</p><p>We can balance the need to accumulate labeled data, as well as the time that it takes to label it by:</p><ul><li><strong>Pooling the data across the fleet. </strong>Most industrial operators have more than one piece of equipment within each category.<strong> </strong>If a typical industrial operator has 100 compressors with 10 years of historical data, that is 1000 years of operational data!</li><li><strong>Augmenting the data </strong>by leveraging decades of engineering experience and knowledge. By calculating operational and performance indicators specific to the equipment, you can enrich your dataset with meaningful insights.</li><li><strong>Leveraging hybrid models </strong>to identify specific patterns that may be indicative of specific failure modes. For example, looking at the rate of change on a temperature sensor could help identify overheating, while on a pressure sensor, it could help to identify blockage or leakage. These methods can be employed to expedite the time to label, as well as alternatives to machine learning models running in live environments.</li></ul><h3>Challenge #4 — Lack of Efficient Tooling 💻</h3><p>Let’s say you’ve solved the first 3 challenges — you have lots of data collected nicely and readily available and you have a rotating equipment specialist that knows what to look for…how are you going to overlay insight with the data?</p><p>Time series data annotation and augmentation has been one of the most underserved domains in the field of data science. Search for image, text, and video preparation and you have tons of awesome tools and services ready to be exploited.</p><p>But <strong><em>time series</em></strong>? And industrial time series at that? Almost nothing, which means that most of the data preparation is still happening in Excel or in bespoke tools developed by eager (read: frustrated) developers; and this knowledge is then often shared over email and phone calls.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DeZAcLpN9vk8Bnupag4Iqw.gif" /><figcaption>Actual sample of data for one single day on a pump. Excel is great for many things, but looking at years of sensor data is not one of them. 😬</figcaption></figure><p>With all of this data and knowledge scattered throughout the organisation, it is impossible to leverage lessons learned and foster sharing or reusability of analytics.</p><p>At Unifai, we’re trying to address this challenge and the aforementioned ones by building the technology necessary to prepare, annotate, curate, and visualise industrial datasets.</p><h3>Rome wasn‘t built in a day, but they were laying bricks every hour.</h3><p>Empires aren’t built overnight and by addressing these challenges slowly and surely, you can build a strong foundation that will reap rewards. The organisations that are able to do this will see an increase in productivity and revenue (sometimes by <a href="https://www.accenture.com/us-en/insights/artificial-intelligence/roi-artificial-intelligence">up to 30%</a> 🥳).</p><p>As a good friend recently told me — this is <strong>no regrets work</strong>, and there is no downside to getting started. 🚀</p><p><em>If you want to discover how Unifai can help you save up to 80% of data preparation time and rapidly transform your sensor data into high quality datasets, sign up to be a beta user </em><a href="https://www.unifai.dev/become-a-beta-user"><strong><em>here</em></strong></a><em> today.</em></p><p><em>Unifai fuels industrial AI by focusing on the foundation — the data. Our platform enables experts to accelerate the creation and curation of high-quality datasets from industrial equipment and processes.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=49f774537598" width="1" height="1" alt=""><hr><p><a href="https://medium.com/unifai/the-industrial-data-crisis-overcoming-the-challenges-49f774537598">The Industrial Data Crisis — Overcoming the Challenges.</a> was originally published in <a href="https://medium.com/unifai">Unifai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>