Leveraging Internal Know-how with Custom Addons:

An MLOps Case Study

Chemical engineer with +20 years of experience in Petrochemicals and Oil&Gas specialized in processes, control and advanced analytics. Currently focused in Efficiency & Sustainability at Moeve.

Miguel Ángel López Andreu /in/miguel-angel-lopez-andreu/

About Us

Pablo Míguez Domínguez

Mechanical engineer, MBA, with experience in digitalization projects related to emerging technologies and implementation of new ways of working through continuous improvement programs at Moeve.

/in/pablo-miguez/

The typical pipeline for a Machine Learning use case

Operators, process engineers and data scientists add value at different stages of the pipeline.

Identify need

Understand problem

Preprocess data

Build, test and deploy model

Gather data

By this stage, 80% of the work is done. Remaining work is repetitive and time-consuming.

Previous experience is not reusable. New use case requires new project.

Soft Sensors Soft sensors are models that use a combination of signals to predict a measurement. Seeq’s built-in capabilities allow SMEs to contextualize plant information for useful soft-sensing applications.

Use case types for Soft Sensors:

• Back-up of measuring devices • Real time estimation for monitoring & control • Prediction models and What-If scenarios

• Exploratory Data Analysis: graphical tools • Data Cleansing: conditions and filters • Feature Engineering: formulas • Model evaluation: underfitting vs overfitting • Models maintenance: residuals monitoring

Model & Predict in Seeq

For many use cases, Seeq’s native regression models are simply the right choice:

• Computationally efficient • Direct explainability with linear or polynomial

coefficients • Feature importance with p-value • Training windows based on domain-knowledge

and conditions • Easy scale-up through Seeq Formula tools /

Asset groups

Automate low- value tasks

Reuse pre-trained models

Empower non-data scientists with robust ML models

The Soft Sensor AddOn When a linear model is not enough, combine Seeq’s time series functionalities and Machine Learning in a no-code solution available within the Workbench environment

• Handle regression and classification problems • Automate preprocessing steps (data cleaning, feature

selection, scaling, one-hot encoding, PCA and SMOTE) • Test different models with cross-validation • Tune hyperparameters • Push predictions to a Seeq Worksheet

A quick look

Soft Sensor for Excess Oxygen in a Furnace Proper oxygen content in industrial furnaces flue gas is critical to achieve safety and energy efficiency. A soft sensor of a calibrated O2 analyzer can help:

• Detect analyzer drift and recalibration requirements

• Identify anomalies in main combustion parameters

Encapsulate solution and upload to private server for seamless integration.

Manage access via Seeq user groups.

D ev

el o

p in

g a

cu

st o

m A

d d

O n Build the Project

Take advantage of Seeq’s infrastructure to prototype in DataLab (Spy library and data sources).

Develop core features: • Custom functions • Custom tools • Custom visualizations

DeploymentIterative Refinement

Develop front-end based on user feedback (ipwidgets, ipyvuetify).

Refine core features and fail fast with AddOn Mode.

The Opportunity

Identify internal know-how that is not being fully leveraged due to varying skillset across individuals or departments.

Tackle bottlenecks in advanced analytics, focusing on repetitive, time-consuming tasks.

Update Maintain DeployVerify value

Thank you rb.gy/ieu58z

Slide 1 Slide 2 Slide 3 Slide 4 Slide 5 Slide 6 Slide 7 Slide 8 Slide 9 Slide 10