Data Driven Performance & Devaition Analysis of Industrial Chemical Weighing Systems - ELiAR

Project description

This project addresses how operational data generated by an automated industrial dosing system can be transformed into reliable insight to understand and reduce deviations between target and actual chemical amounts dosed during a production process.

Across roughly 90 connected machines deployed in factories worldwide, every dosing job produces detailed records of intended versus actual amounts, alongside alarms, waiting times, and machine utilisation, all stored in a central database.

The challenge was to explore this large, only partially structured dataset, identify which factors are most associated with larger deviations, and translate those findings into a validated, interactive dashboard that both technical and non-technical stakeholders can use to monitor performance, alongside concrete, data-supported recommendations for where monitoring and maintenance effort should be focused.

Context

The project sits in the production and manufacturing domain, specifically industrial automation for the textile dyeing industry. The client is a manufacturer of automated liquid chemical weighing and distribution systems, used by textile dyeing factories to dose chemicals precisely during the dyeing process. Accurate dosing directly affects dyeing quality, chemical waste, and production cost, so even small deviations between the target and the actually dosed amount can have meaningful downstream consequences for a factory's output and expenses.

The company manages more than 90 of these mechatronic systems, installed across multiple countries, factories, and equipment generations, and each system continuously generates operational data, including weighing executions, alarms, queue and waiting times, and utilisation, all consolidated into a central relational database.

At the start of the project, this data was not yet used in a structured, data-driven way to monitor or compare system performance; insight existed mainly at the level of individual factory reports rather than across the full fleet of systems. The project domain therefore combines industrial process monitoring, mechatronics, and business intelligence, since the work required understanding the physical dosing process, the way that process is captured as data through an ETL pipeline, and how that data can be modelled and visualised to support operational decisions for a global industrial client.

Results

The project's main outcomes are a validated Power BI dashboard, Streamlit dashboard, and a set of data-supported recommendations, both delivered to the client within an 18-week timeframe. The dashboard is built on a star-schema data model connected via DirectQuery to the client's PostgreSQL database, and is organised into four pages: a geography analysis, a deviation analysis, an alarms and reliability view, and a queue and workload view. It includes KPI cards, comparative charts across factories and chemical groups, scatter plots comparing target versus actual amounts, and time-based workload visualisations, giving both technical analysts and non-technical managers a way to monitor deviation behaviour across the full fleet of around 90 systems rather than per factory in isolation.

Underneath the dashboard sits a defined KPI framework covering deviation stability, alarm intensity and reliability, and system utilisation, grounded directly in SQL-based exploration of the operational data rather than abstract metrics. This analysis also surfaced concrete findings, for example which chemical groups, machine types, or factories are associated with larger deviations, and how alarm events relate to process outcomes, which were translated into recommendations on where closer monitoring or maintenance attention is likely to have the most impact.

Validation took place iteratively rather than only at the end: the dashboard went through six design iterations, each reviewed with the client and academic coach, moving from an initial single-page concept to the final four-page version. Concrete client feedback, such as changing geographic filtering from city level to region and country level and adding cross-filtering between systems, companies, and regions, was directly incorporated, and the underlying KPI definitions and dashboard wireframes were separately reviewed with the coach for analytical soundness.

In terms of TRL positioning, the result sits at roughly TRL 6: a working, interactive prototype that has been demonstrated and validated using real operational data in the client's actual business context, with direct stakeholder feedback shaping its design, but not yet integrated into the client's live reporting or operational pipeline. Reaching a higher TRL would require structurally embedding the dashboard into the client's regular reporting process, automating data refresh in production, and validating adoption over a longer operational period, steps that were explicitly out of scope for this project and are instead captured as recommendations for future work.

About the project group

Made up of three members, Joaquin, Almario, Khulan. Joaquin (Group Leader) studies in Mechanical Engineering in Fontys Engineering. Almario studies ICT & Business at Fontys ICT, and Khulan is an exchange student from Italy studying Economics and Data Science. We have worked during a time period of one semester.

The work was distributed evenly throughout the three group members and in the end Joaquin and Khulan are most responsible for the PowerBI Dashboard and Almario is for the Streamlit dashboard.