FrameworX and Machine Learning

Introduction

Isabela Taccolini (Marketing, Tatsoft)

Hello everyone, and thank you for joining us. My name is Isabela, and I work in marketing here at Tatsoft. We’re very happy to have you with us for this special Technical Tuesday session.

Today, we’ll be talking about AI on the plant floor, with a focus on machine learning, ML.NET, and how these technologies are implemented inside the Tatsoft FrameworX platform.

We’ll start with a short presentation, followed by a live demonstration, and we’ll close with Q&A. Feel free to ask questions in the chat at any time.

This session is being recorded, and we’ll share the recording along with all referenced resources, examples, and tutorials after the event.

I’m joined today by:

• Dave Hellyer, VP of Business Development
• Eduardo, Automation Engineer
• Scott Gray, Senior Solutions Consultant

Let’s get started.

Tatsoft and the FrameworX Platform

Dave Hellyer (VP of Business Development)

Tatsoft has more than 30 years of experience building industrial real-time software. The company was founded by the original creators of InduSoft, and that background strongly influences our focus on OEMs, embedded systems, and large-scale industrial deployments.

Today, Tatsoft operates globally, with teams in the United States, Brazil, Europe, and Asia. Our platform, FrameworX, is used across industries including manufacturing, water and wastewater, food and beverage, life sciences, transportation, oil and gas, and data centers.

FrameworX is a modern SCADA and IIoT platform built on a managed .NET architecture, with native support for Python 3. It is a single, unified platform—not a collection of add-ons. Alarming, historians, security, scripting, MQTT, Sparkplug B, edge connectivity, device drivers, and visualization all live in the same environment.

You configure everything once and deploy the same project across edge devices, servers, virtual machines, or containers without rewriting the application.

This unified foundation is what makes advanced capabilities like machine learning practical. Because FrameworX is built on .NET, the same runtime is used for scripting, calculations, alarms, business logic, and machine learning—without the need for external execution environments.

When machine learning runs inside the operational platform, results become first-class industrial data. They can be historized, alarmed, visualized, and acted upon just like any other signal.

With that context, I’ll hand it over to Scott Gray to discuss why machine learning is relevant in industrial applications.

Why Machine Learning Matters in Industry

Scott Gray (Senior Solutions Consultant)

Machine learning is most commonly used to optimize processes and improve outcomes—but its value goes far beyond that.

One major benefit is knowledge retention. Experienced operators and maintenance technicians often understand subtle machine behaviors that are difficult to document. Machine learning allows you to capture and preserve that expertise, reducing reliance on specific individuals and making knowledge transferable across lines, plants, or sites.

Machine learning can help:

• Increase production output
• Reduce scrap and rework
• Shorten cycle times
• Reduce unplanned downtime
• Improve product quality
• Reduce energy consumption
• Improve scheduling and root cause analysis

From a return-on-investment perspective, some of the most common use cases include:

• Predictive maintenance
• Process optimization
• Defect detection
• Production forecasting

Rather than reacting to failures, machine learning enables maintenance and operations teams to act proactively based on real conditions.

Common Machine Learning Models in Industrial Use

Some common machine learning approaches include:

• Binary classification: Predicting one of two outcomes (for example, pass/fail).
• Multi-class classification: Identifying one condition among many possible outcomes.
• Regression: Predicting a numeric value such as energy usage or cycle time.
• Time series forecasting: Predicting future values like hourly production output.
• Anomaly detection: Identifying unusual behavior that deviates from normal operation.
• Clustering: Grouping data into operational states such as normal, stressed, or inefficient.
• Decision trees and gradient-boosted trees: Understanding which variables drive failures or inefficiencies.

It’s important to note that machine learning should not replace traditional control systems. PLCs and controllers continue to handle real-time control, while machine learning supports analysis and decision-making.

Why ML.NET?

ML.NET is Microsoft’s machine learning framework for .NET applications. It is particularly well suited for industrial environments where:

• Data is structured (temperatures, pressures, vibration, flow)
• Performance and execution speed matter
• Tight integration with .NET is required

ML.NET models are compiled and execute very efficiently. Because FrameworX itself is built on .NET, ML.NET integrates naturally into the platform.

Models can be trained using Visual Studio Model Builder, which guides users through data selection, training, evaluation, and code generation. Once validated, the generated code can be directly embedded into FrameworX projects.

Python remains a powerful option for deep learning and offline analytics, especially when using libraries like TensorFlow or PyTorch. However, for real-time, operational use cases, ML.NET often provides better performance and easier integration.

The key is choosing the right tool for the job.

ML.NET Demonstration in FrameworX

Eduardo Bogo (Automation Engineer)

In this demonstration, we show how ML.NET models can run directly inside FrameworX—using the same project environment that already manages tags, historian data, alarms, and visualization.

We implemented three practical examples that Scott mentioned earlier:

1) Regression: Predicting a numeric value (Quality)

In the regression example, we start with training data that contains three columns:

• Pressure
• Temperature
• Quality (the numeric output we want to predict)

After training the model, we feed current process values (in our case, simulated real-time values) into the model. Every time pressure or temperature changes, FrameworX runs the ML.NET model and outputs a predicted quality value. That predicted value can be written to a tag and used like any other signal—displayed on screen, historized, alarmed, and analyzed.

2) Binary classification: Pass/Fail outcomes

In the binary classification example, we use multiple process variables as inputs and predict whether the product passes quality control:

• 1 = pass
• 0 = fail

We also included a visual indicator on the screen:

• Green = pass
• Red = fail

This makes it easy for operators to understand model outcomes instantly.

3) Anomaly detection: Normal vs. abnormal states

In the anomaly detection example, we evaluate multiple process variables and determine whether the current operating condition is:

• Normal
• Abnormal

In addition to a normal/abnormal status, the model also provides a score. A higher score indicates the operating condition is closer to an abnormal state.

As a general recommendation, anomaly detection is often a strong first step for industrial machine learning—especially for predictive maintenance.

Where to download the demo and technical resources

We created a dedicated technical resource page for this session. On that page, you can:

• Download the ML.NET example project
• Review a detailed walkthrough of the solution
• Access related tutorials and supporting documentation

The example is structured using the “four pillars” methodology:

1. Data Foundation
2. Industrial Operations
3. Business Operations
4. User Interaction (Displays)

The ML.NET workflow (why it’s simpler than it looks)

Even though machine learning can seem complex, ML.NET reduces the workflow to a set of clear steps:

1. Load training data (CSV, database, historian export, etc.)
2. Define input and output classes
   1. Inputs = process variables
   2. Outputs = predicted result (depends on the model type)
3. Build a pipeline (choose transformations + algorithm)
4. Train the model
5. Evaluate the model (validate with a subset of known data)
6. Save the trained model
7. Run predictions with live field data

A practical predictive maintenance example is:

• Collect healthy operating data from an asset (3–6 months when possible)
• Train an anomaly detection model to learn the “normal state”
• Monitor live sensor values
• Alarm or notify when behavior deviates from normal

The advantage is reducing downtime and optimizing maintenance. The tradeoff is that you need sensors, data history, and good analytics practices.

Visual Studio vs. FrameworX: Training and deployment

For development, we recommend:

• Train and evaluate the model in Visual Studio (using ML.NET and Model Builder)
• Once performance is acceptable, deploy into FrameworX

In Visual Studio (regression example), we:

• Choose the regression scenario
• Select the algorithm (for example, LightGBM)
• Train with one method call
• Evaluate with ML.NET metrics
• Save the trained model

Inside FrameworX, the runtime scripting typically focuses on the last part:

• Load the saved model
• Pass real-time tag values as inputs
• Run prediction
• Write the output back to a tag

Because FrameworX is .NET-based, the ML.NET code structure is very similar between Visual Studio and the FrameworX scripting environment.

MCP: Connecting FrameworX to Large Language Models

In addition to ML.NET, we introduced MCP (Model Context Protocol), which allows FrameworX to securely expose operational data to large language models such as ChatGPT or Claude.

With MCP, LLMs can:

• Query live tag values
• Retrieve alarm status and history
• Access historical data
• Execute custom methods defined in FrameworX

This enables natural-language interaction with industrial systems while keeping control and security within FrameworX.

ML and MCP can also be combined—for example, querying anomaly detection results generated by ML.NET models using conversational AI.

Key Takeaways

• Start small with AI and machine learning projects
• Anomaly detection is often the best first step
• Machine learning improves over time through retraining and feedback
• Data quality, consistency, and context are more important than algorithms
• Machine learning works best when it is part of an operational system, not a standalone experiment

FrameworX provides a unified, contextualized platform that reduces friction and accelerates adoption of advanced analytics.