Blog

Chat with AI or Artificial Intelligence. Young businessman chatting with a smart AI or artificial intelligence using an artificial intelligence chatbot developed by OpenAI.

<h3>TL;DR</h3>When most people think about the complexity of Artificial Intelligence in industry, they immediately picture deep neural networks, advanced architectures, or models with dozens of layers.But the real challenge often starts much earlier — with something as “simple” as a linear regression. Let's dig in.<h3>The Hidden Complexity of Simplicity </h3>Imagine a model with just two parameters, running in real time, every second, 24/7 on the plant floor.That’s:<ul><li>1 model</li><li>Running every 1 second</li><li>For 30 days → 2,592,000 inferences per month</li></ul>And this is before you even validate whether the prediction can be trusted.<h3>Validating Reality: Millions of Operations </h3>Every inference should be checked against at least five basic validations:✔️ Is the input data present? ✔️ Is it within the expected scale? ✔️ Is the feature within range? ✔️ Is there an outlier? ✔️ Is the model still calibrated?<blockquote>Even with this conservative set of 5 checks, that single regression means 13+ million operations per month — just to ensure reliability.</blockquote>Now scale this up:<ul><li>Medium-sized industrial company → ~30 active models = ~400 million operations/month</li><li><span style="font-family: inherit; font-size: calc(18px * var(--body-fontSize, 1)); font-style: var(--body-fontStyle, normal);" data-mce-style="font-family: inherit; font-size: calc(18px * var(--body-fontSize, 1)); font-style: var(--body-fontStyle, normal);">Large industrial company → ~150 models = ~1.9 billion operations/month</li></ul>And this is with linear regression, one of the simplest ML techniques available.<h3>Why Talk About Linear Regression? </h3>Because despite its simplicity, linear regression remains one of the most powerful tools in industry.We’ve seen high-impact use cases such as:<ul><li>Virtual sensors with just one or two input variables</li><li>Simple and robust anomaly detection</li><li>Estimations of hard-to-measure variables in real time (e.g., NOx, CO, O₂)</li></ul>It works — reliably and efficiently — especially in industries such as mining, chemicals, and pulp &amp; paper.But “simple” doesn’t mean “trivial.” <blockquote>In OT, even a basic regression requires robust infrastructure, automation, and observability to remain trustworthy at scale.</blockquote><h3>The Real Lesson </h3>Never underestimate the power of linear regression — nor the complexity of keeping it alive in production.The question every industrial organization must ask is:👉 Is your data architecture and MLOps stack ready for this? 👉 Can you guarantee performance, reliability, and security at this level of scale?<h3>Final Thoughts </h3>At Aignosi, this is the daily reality we face with our industrial clients. There’s no glamour in it — but making AI work in OT means ensuring that every inference, every second, is dependable.And that requires more than algorithms. It requires industrial-grade AI infrastructure.

Why Keeping a Simple Linear Regression Alive in OT Is Harder Than You Think

<blockquote>The story of AI/ML at Samarco didn’t begin with a massive data science team or an ambitious enterprise-wide program. </blockquote>It began in 2022, when a handful of automation engineers developed a simple virtual sensor to estimate mass flow on a conveyor belt.That model worked. It helped operators detect anomalies earlier, avoided blockages in crushers and screens, and kept the process flowing. What seemed like a small step revealed a huge opportunity: if one conveyor could benefit from AI/ML, why not 30?But scaling wasn’t just about copying code. It required a holistic approach around People, Processes, and Technology.<h3>People: Building Internal Capability &nbsp;</h3>One of the most remarkable aspects of this journey is that Samarco itself developed 13 of the 15 models currently in production.This was only possible because the company created a dedicated Process Optimization team inside OT. Multidisciplinary by design, the team combined automation, process engineering, and data science skills. With Aignosi’s support, Samarco conducted an analytics maturity diagnosis to identify gaps, set priorities, and build internal skills.This organizational move was crucial: instead of outsourcing model development, Samarco empowered its own people to own the models — designing, deploying, and maintaining them. AI became part of the operational DNA, not a black-box service.<h3>Processes: Governance and Repeatability &nbsp;</h3>The next step was to standardize. Without clear governance, models risk becoming isolated “one-off experiments.” Samarco adopted AIOps for OT, incorporating best practices from MLOps, DataOps, and DevOps:<ul><li>Version control for models and datasets</li><li>Automated retraining pipelines</li><li>Continuous monitoring for drift and anomalies</li><li>End-to-end observability of the models in production.</li><li>Standardized workflows and templates for rapid replication This turned model deployment into a repeatable, reliable process. What once took months could now be done in days.&nbsp;</li></ul>These processes turned AI from a one-time success into a repeatable system, cutting deployment lead-time from months to days.<h3>Technology: SIENTIA™ as the Enabler &nbsp;</h3>The backbone of this transformation is SIENTIA™, an on-premise AIOps platform designed for OT environments. Unlike generic IT tools, it was built to meet industrial realities:<ul><li>Native OPC UA connectors for integration with SCADA, PLCs, and PIMS</li><li>100% on-premise deployment, addressing cybersecurity and latency requirements</li><li>Real-time inferencing, ensuring AI insights are available when operations need them</li><li>No-code templates, allowing engineers and operators to configure models without Python expertise &nbsp;</li></ul>This meant that Samarco’s own team could focus on process knowledge, while SIENTIA™ took care of governance, automation, and observability.<h3>&nbsp;The Result </h3><blockquote>In less than 90 days, Samarco went from 1 model to 15 models in production — and 13 of those were developed internally by its own Process Optimization team.</blockquote>This is more than scale. It’s organizational autonomy, where AI is not just a project but a capability owned by the operation itself.And this is only the beginning. The roadmap now points to 50 models by the end of 2025, positioning Samarco as one of the leaders in applying Industrial AI to mining operations in Brazil.We will bring more stories about how Samarco is continuing to scale their AI/ML in the shop floor.

AI/ML Scalability at Samarco's OT

<blockquote>Conveyor belts are the arteries of a mining operation. </blockquote>Yet, unplanned stoppages and failures in belt scales are costly and frequent. Samarco solved this challenge not with deep learning or complex architectures, but with something deceptively simple: a linear regression model.This virtual balance sensor estimates conveyor mass flow in real time, delivering the same function as a physical belt scale — but with higher reliability and lower maintenance.<h3>Simplicity That Works &nbsp;</h3>While often overlooked, linear regression remains one of the most powerful tools in industrial AI/ML. In Samarco’s case, it proved accurate enough to feed APCs directly, with retraining needed only once every 4 months.Its success shows an important lesson: <blockquote>AI in OT doesn’t start with deep learning — it starts with models that work and keep working.</blockquote><h3>The Auto-ML Template: A Samarco Idea Scaled Through SIENTIA™ &nbsp;</h3>One of the key breakthroughs came not from the algorithm itself, but from how the model was operationalized.The idea of creating a no-code template for conveyor models originated within Samarco’s own team. By incorporating this into SIENTIA™, the approach became part of the platform’s DNA:<ul><li>Drag-and-drop simplicity, so models could be built without coding.</li><li>Standardized workflows, ensuring governance and observability by default.</li><li>Accessibility for all skill levels: from senior process engineers to plant operators, anyone could configure and replicate the model in minutes instead of months &nbsp;</li></ul><h3>&nbsp;Impact on Operations &nbsp;</h3><ul><li>Reduced unplanned stoppages from conveyor overloads.</li><li>Reliable mass flow estimation, even when physical scales fail.</li><li>Scalability across 30+ conveyors, multiplying the impact.</li><li>Cultural empowerment, as more people across the organization can now build and deploy ML models safely. &nbsp;</li></ul>

Soft sensors for conveyor belts at Samarco

<h3>TL;DR;</h3>Using the same template created for <a data-mce-href="https://www.aignosi.com.br/blog/soft-sensors-for-conveyor-belts-at-samarco" href="https://www.aignosi.com.br/blog/soft-sensors-for-conveyor-belts-at-samarco" target="_blank" rel="noopener">virtual belt scales</a>, Samarco built a silica prediction model for flotation. With SIENTIA™’s automated retraining, even a simple regression delivers reliable, real-time insights without costly custom development.<h3>The case</h3>In iron ore flotation, silica content is a critical quality parameter. Traditionally, plants rely on laboratory analyses every two hours, leaving Advanced Process Controllers (APCs) operating with delayed data. This slows decision-making, increases reagent waste, and risks product quality.Samarco addressed this challenge with a virtual quality sensor, designed to predict silica grade every 5 minutes. Surprisingly, the model wasn’t built on complex deep learning, but on a multivariate linear regression.<h3>&nbsp;Reusing What Already Worked &nbsp;</h3>Instead of reinventing the wheel, <a data-mce-href="https://www.aignosi.com.br/blog/soft-sensors-for-conveyor-belts-at-samarco" href="https://www.aignosi.com.br/blog/soft-sensors-for-conveyor-belts-at-samarco" target="_blank" rel="noopener">Samarco reused the same no-code template originally created for conveyor belt mass flow models</a>. This was possible thanks to the flexibility of SIENTIA™, which allowed the Process Optimization team to adapt an existing template to a new use case.But flotation is a nonlinear process, and a simple regression alone wouldn’t be robust enough over time. This is where another key feature of SIENTIA™ came into play.<h3>Automated Retraining: The Game-Changer &nbsp;</h3>SIENTIA™ provided automated retraining every 4 hours, which compensated for process variability and data drift. This capability transformed what could have been an unstable model into a reliable, production-ready sensor.The combination of a simple template + automated retraining meant that Samarco could:<ul><li>Avoid exorbitant costs of custom model development.</li><li>Leverage its own team’s process knowledge.</li><li>Deploy a working solution much faster than traditional R&amp;D cycles. &nbsp;</li></ul><h3>Impact on Operations </h3><ul><li>Silica predictions delivered every 5 minutes instead of waiting 2 hours.</li><li>Direct integration with the flotation APC, enabling real-time reagent dosing.</li><li>Significant reduction in reagent waste and higher product quality stability. </li></ul><blockquote>This case is a good example that even “simple” models, when combined with robust MLOps features, can solve complex nonlinear problems. &nbsp;</blockquote>

Soft Sensors for quality prediction at Samarco

Artificial Intelligence, Technology, Robot, Futuristic, Data Science, Data Analytics, A.I.

TL; DR;"MLOps started in the cloud. Now it must survive on the factory floor."Google’s canonical MLOps architecture set the foundation for ML systems engineering — defining CI/CD pipelines, model registries, automated retraining and performance monitoring. It works well for cloud-native apps, SaaS platforms, and enterprise IT.But on the industrial shop floor? It struggles to survive. Let's dive into this topic in this article.<h3>The gap between IT and OT</h3>Industrial manufacturing lives in a different universe:- 🔌 **Intermittent or no cloud connectivity** - ⏱️ **Sub-second latency for critical systems** - 🔐 **Sensitive data locked behind OT firewalls** - 🧱 **Legacy equipment using OPC UA, SCADA, and PLCs**These constraints make the cloud-first architecture unfit for real-time control loops, predictive quality models, and embedded diagnostics.<h3>A new architecture is emerging: hybrid edge-cloud</h3>A recent scientific publication by Raffin et al. (2022, FAU Germany) proposes a reference architecture tailored to manufacturing. Unlike the centralized Google model, this architecture is **distributed, modular and event-driven**.It’s not about abandoning the cloud — it’s about **empowering the edge**. Here’s how:🧠 Inference at the edgeModels are served locally via lightweight containers, with inference times in the order of milliseconds. This ensures safety, uptime, and operational continuity even when offline.🧩 Microservices and domain-driven designEach pipeline component — from data acquisition to decision services — is a microservice deployed on the edge. Interfaces are designed for modularity, encapsulation, and adaptability to factory heterogeneity.&nbsp;🔄 CI/CD/CT adapted to OTA full DevOps loop is preserved, but adapted:- Models are versioned and tested centrally - Containers are deployed locally using GitOps or IaC - Continuous Training (CT) is triggered by drift or scheduling - Monitoring is embedded both locally and in the cloud&nbsp;📊 Governance and feature storesData is cached and served via local feature stores. Model drift, data quality and inference logs are managed on-site, with batch synchronization to cloud systems for retraining and collaboration.<h3>Why it matters</h3>This shift isn’t just technical — it’s operational and strategic. The success of industrial AI depends on:- Running models where the data lives - Empowering process engineers, not just data scientists - Handling millions of edge nodes without losing controlThe hybrid architecture enables AI to operate autonomously, securely, and explainably — without sacrificing agility or compliance.<h3>How aignosi is leading this transition</h3>At aignosi, we’ve implemented these principles in our platform — combining:- 🔹 Industrial-grade data collectors (OPC UA, MQTT, Modbus) - 🔹 Scalable edge-based orchestration - 🔹 Embedded MLOps governance and observability - 🔹 No-code/low-code interfaces for domain expertsWith our solutions, manufacturers are running dozens of AI models per site, even in air-gapped or partially connected environments.We believe the future of AI in manufacturing is not cloud-only — it’s cloud-aware, edge-native, and operator-centric.---📚 Reference: Raffin, T. et al. (2022). A reference architecture for the operationalization of machine learning models in manufacturing. Procedia CIRP, 115, 130–135.*

From Cloud to Factory Floor: The Evolution of MLOps Architectures for Manufacturing

Flying through the futuristic tunnel. Abstract 3D animation. Data Network, Virtual Reality, Quantum. The concept of illuminated corridor, interior design, spaceship, Galaxy, science, lab, technology, science

TL;DR;Efficient operationalization of machine learning (ML) and deep learning (DL) algorithms in the industrial sector presents significant challenges due to the inherently probabilistic nature of these methods, the constant need for retraining, and their critical dependence on high frequency data sampling/inference. To address these challenges, this article summarizes a proposed a hybrid cloud/edge architecture, integrating centralized computing resources (cloud) with decentralized, local resources (edge) as shonw in the original paper [1].<h3>Motivation and Context</h3>The integrated application of ML in Cyber-Physical Systems (CPS) is essential for smart manufacturing, delivering substantial gains in flexibility, agility, and operational robustness. However, effective implementation requires a systemic approach that enables both vertical and horizontal integration of data and models throughout the entire production chain—also known as Cyber-Physical Production Systems (CPPS). In this context, Continuous Delivery (CD) practices become essential to support the centralized and ongoing deployment of trained models across distributed industrial environments.Key Challenges in Industrial Implementation• Technological Heterogeneity: A wide range of equipment, non-standard communication protocols, and various data formats.• Data Management: The need for high-quality, adequately volumed, and efficiently integrated structured and unstructured data.• Sustainable Model Operation: Continuous monitoring, support for incremental learning, and guaranteed low latency due to the critical and time-sensitive nature of industrial processes.• Non-Technical Challenges: Shortage of skilled labor, high stakeholder expectations, and concerns about cost-effectiveness of tech implementations.<h3>Proposed Architecture</h3>The proposed hybrid cloud/edge architecture is composed of five key domains:1. Edge SystemHandles time-sensitive tasks such as data acquisition, local pre-processing, real-time inference, and direct interaction with physical devices (sensors and PLCs).2. Cloud InstanceConducts more complex, less time-sensitive tasks such as system-wide monitoring, analytical dashboards, and advanced incremental learning techniques.3. Integrated Data ManagementProvides economical, long-term storage for large volumes of data, with local caches enabling near-instantaneous access for critical industrial processes.4. Development and ExperimentationCovers the full cycle of ML/DL model creation, experimentation, and optimization, supported by rigorous version control and experiment tracking.5. Machine Learning Operations (MLOps)Involves continuous and automated processes for integration (CI), delivery (CD), and training (CT), which are essential to dynamically maintain and update ML models.<figure id="figure-m88yu3uksm8nepkbg" data-full-width=""><img id="m88yu3uksm8nepkbg" src="https://cdn.durable.co/blocks/cFci9mLApW7gxH4CyFdDsS27qHWyFzzA9Mvl4puTVGpRyAQTRDvP79QAvY5C6vMm.jpg" alt="" title="" class="object-cover bg-gray-100 rounded-lg md:rounded-xl lg:rounded-2xl selected" style="aspect-ratio: 16/9; object-position: undefined undefined;" data-corners="medium" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="true" data-mce-src="https://cdn.durable.co/blocks/cFci9mLApW7gxH4CyFdDsS27qHWyFzzA9Mvl4puTVGpRyAQTRDvP79QAvY5C6vMm.jpg" data-mce-style="aspect-ratio: 16/9; object-position: undefined undefined;" data-mce-selected="1"></figure>Proposed hybrid architecture for MLOPs/AIOps purposes in manufacturing.<h3>Advantages of the Hybrid Architecture</h3>• Significant Latency Reduction: Local inference ensures fast, effective responses in critical operations.• Optimized Bandwidth Usage: Local pre-processing greatly reduces the volume of data transmitted to the cloud.• Greater Flexibility and Scalability: Complex model training can be performed in the cloud while simplified versions run at the edge.• Improved Security and Operational Reliability: Minimizes external risks and ensures continuity of operations even when cloud connectivity is lost.Implementation Challenges• High technical complexity in integrating cloud and edge layers.• Substantial initial investment in advanced technological infrastructure.• Complex security management in distributed environments.• Ongoing need for maintenance and frequent updates on both layers.<h3>Conclusion and Strategic Relevance</h3>The hybrid cloud/edge architecture is essential for ensuring scalability and economic efficiency in deploying ML solutions in complex industrial environments. By combining the analytical power of the cloud with the speed and operational efficiency of edge computing, this approach holds significant potential to optimize operational performance, improve product quality, and enhance the overall robustness of production processes.This article contributes to both academic and practical understanding of the strategic importance of smart integration between cloud and edge computing, highlighting its critical role as a competitive differentiator in Industry 4.0.Reference:<a data-mce-href="https://www.sciencedirect.com/science/article/pii/S2212827122014974" href="https://www.sciencedirect.com/science/article/pii/S2212827122014974" title="A reference architecture for the operationalization of machine learning models in manufacturing" target="_blank" rel="noopener">[1] A reference architecture for the operationalization of machine learning models in manufacturing</a>

Blog

Why Keeping a Simple Linear Regression Alive in OT Is Har...

AI/ML Scalability at Samarco's OT

Soft sensors for conveyor belts at Samarco

Soft Sensors for quality prediction at Samarco

From Cloud to Factory Floor: The Evolution of MLOps Archi...

Hybrid Cloud/Edge Architecture for Operationalizing Machi...