Christoph Netsch’s Post

The good, the bad, and the ugly of AI 🚳 Reflecting on how AI's state-of-the-art has evolved lately, it is remarkable to see, how computers can perform the immensely complex tasks of arranging tens of thousands of pixels almost from randomness to create (somewhat 😉 ) coherent video frames. On the other hand, the video sequences clearly show that my dog 🐕 has a more grounded understanding of the physics of a bike race. As a biker, I find this video (source unknown) quite funny. Many have commented to make their point along the lines of "AI is stupid". Personally, I don't see much value in people's tendencies to antropomorphize AI by labeling it with human attributes. I would argue that even the most powerful AI models we interact with today don't inherently "understand" the world around them, even though some do a decent job at emulating that understanding most of the time. That doesn't render them useless: - AI algorithms are effective tools to obtain insights about patterns, anomalies, and other phenomena in data, where analytical methods are either not possible or cost-effective. - Embedding them in software designed with domain expertise to process data according to its meaning can result in outputs, that maximize traceability and create real business value. This pragmatic view on the capabilities as well as the limitations of AI is something, I truly appreciate that we share among the Alpamayo team. I believe it is the type of mentality #industrialAI needs to take AI from keynotes, whitepapers, and PoCs to the shopfloor.

🚰 Data Leakage - Sounds like a beginner’s mistake, but it happens more often than you'd think. Performance metrics for data-driven algorithms for condition and process monitoring are typically benchmarked using test data. Data leakage occurs when information from test data influences training, leading to overly optimistic performance results. It makes models seem better than they are at predicting new data. You’d think this is an absolute beginners mistake, but data leakage can sneak in through the back door in #conditionmonitoring and #processmonitoring. And if this has happened to you before, you’ll pay special attention to it and will even come across studies presenting overly optimistic benchmarks due to data leakage. 📑 Yesterday, Omri Matania, Roee Cohen, Eric Bechhoefer, and Prof. Jacob Bortman published “Test-Training Leakage in Evaluation of Machine Learning Algorithms for Condition-Based Maintenance.” They highlight two types of data leakage you should pay attention to: 1️⃣ Segment Leakage: Example: Splitting a 10-second vibration signal into ten 1-second segments. Issue: Segments from the same record in both training and test sets share characteristics unrelated to the fault type. 2️⃣ Record Leakage: Example: Using different records of the same fault type in both sets. Issue: The model learns specific patterns that don't generalize to new, unseen fault shapes. I'd like to add another type myself, where I am unsure whether it technically classifies as data leakage, but definitely leads to a misalignment of expectations and performance benchmarks: 3️⃣ Entity Leakage: Example: Data from multiple bearings of the same type. An entity corresponds to an individual bearing. Issue: People sometimes expect their model to generalize from one entity to another. When that is a requirement, then a benchmark will fail to demonstrate this unless data is split between entities. Otherwise, the algorithm might learn irrelevant patterns associated with individual machines that do not generalize well to new, unseen machines.

When we take consultation calls at Alpamayo there’s a question I always ask. 🔦 I enjoy going deep into the nitty gritty details of specific failure modes, sensors, data sharing, competing solutions… if that helps you to better understand whether your machine needs an integrated analytics software and whether PREKIT is the right one for you. But there is one thing I always need to ask you beforehand, because discussing any of these details is meaningless until the purpose is entirely clear. It ends up defining many of the requirements. „What is the top-ranking strategic priority driving your machine analytics project?“ 📝 We recently tallied up the answers we got from talking to the heads of internal digitisation teams and executives of equipment manufacturers throughout the past year as you can see in the picture below. To be honest, I was a bit surprised how (within our sample) product differentiation clearly dominated all other narratives. What do you think about the results? Would you have expected different answers?

When it comes to AI in predictive maintenance, I’d chose a low performing glassbox over a high performance blackbox any day. That is because in maintenance, data-driven tools don’t replace engineering sense. They rather empower operators and maintainers to make informed decisions and plan more effectively. An anomaly detection for instance, generally outputs some sort of health score or attention index. But it becomes much more effective as a decision support system, once the human is provided with deep context to a potential failure detection, so he can reason about its source and the plan the most appropriate course of action. Thinking about how AI and the user interfaces around it can be made more useful by providing rich and interpretable context has become one of my favorite projects to work on over the past months. These are a few features beyond the classic attention index, that we recently made accessible in the PREKIT user interface: ⚙ likely system element of origin ⚠ likely modes of failure 💬 OEM recommendations for appropriate interventions 📈 healthy envelope of operation for a given process parametrization 📉 examples of similar detections in past and from other machines 👁️🗨️ how familiar a model is with the current operational context What information would you need from your analytics software, so it is useful in your #conditionmonitoring and #processmonitoring use case?

Vervierfachung der Standzeiten. 70% Ammortisation in einem Jahr. Mehr Sicherheit. Die aktuelle SMM - Schweizer Maschinenmarkt zeigt ein anschauliches Beispiel, wie der Werkzeugmaschinenhersteller Leistritz AG in Partnerschaft mit UTTec GmbH & Co. KG seinen Kunden dank integrierter Prozessüberwachung handfeste Mehrwerte verschafft. ❗ Besonders spannend: die Story hinter der Entwicklung zeigt, dass gerade im Werkzeugmaschinenbau, sehr viel Hersteller-Know-How in derartigen Lösungen steckt: Jede Maschine fertigt verschiedene Produktserien mit unterschiedlichen Geometrien und Werkstoffen. Dabei kommen verschiedene Werkzeuge und Schnittparameter zum Einsatz. Zum Teil werden in einer Werkzeugmaschine ganz unterschiedliche Prozesse gefahren. 📉 Ein Analyseverfahren, das nicht in der Lage ist, diese ganzen Zusammenhänge mitzuberücksichtigen, vergleicht zwangsweise Äpfel mit Birnen, wenn es den aktuellen Prozess beurteilt. 📖 Die volle Geschichte kannst du hier lesen: https://lnkd.in/dPqKjiij

For the average AI-powered condition monitoring tool, 11% of the development effort goes into the algorithm. That’s data from an OEM in the transportation industry, who has become quite proficient with condition-based and predictive maintenance. And yet, we tend to spend a lot of time getting excited about shiny new AI algorithms, as if each one is the silver bullet we've been waiting for that finally makes AI for manufacturing timeseries data "easy". And we spend too little time talking about the analytics platforms that support those AI models, ensure great data quality, and actually make the endeavour a lot easier.