For years, continuous improvement has been synonymous with Pareto charts, team meetings, root cause analyses with post-its, and a great deal of operational intuition. In this article, I use the DMAIC framework (Define, Measure, Analyze, Improve, Control), widely adopted in Lean Six Sigma methodologies, as a guiding structure to reflect the three major phases of evolution I experienced—and later imagined—in transforming industrial maintenance. This is no coincidence: the DMAIC cycle helps structure thinking, ensure each step adds value, and connect operational excellence with data-driven decision making.
However, the arrival of data science and artificial intelligence has radically transformed how we identify and resolve issues on the shop floor. What once relied almost entirely on human expertise and manual analysis can now be accelerated and scaled with sensors, algorithms, and real-time dashboards.
To illustrate this evolution, I share a case I experienced firsthand as a maintenance manager: how we tackled recurring failures in car body conveyors in an industrial plant, and how that same approach could be reimagined today with data science and artificial intelligence tools. More than a story, this article aims to offer a critical reflection on each phase. Because it’s not just about technology—it’s about understanding what worked, what didn’t, and why.
Phase 1 – From Paper to Excel (2004–2007)
Applying DMAIC before sensors and algorithms
When I took on the role of maintenance and technical services manager, one of the biggest challenges was the recurring breakdowns in the conveyors that moved car bodies along the production line. Although the systems were controlled by PLCs, they were outdated models with no connectivity or structured failure logging. Every time a stoppage occurred, technicians responded, but there was no traceability or historical data to support analysis.
Define
We set a clear goal: significantly reduce conveyor stoppages. The problem’s scope and impact on production were defined collaboratively with technicians and operators.
Measure
We built a manual tracking system in Excel, a sort of rudimentary CMMS. We documented every incident: conveyor section, failure type, and downtime. This dataset enabled us to generate the first Pareto charts and identify some early trends.
Analyze
We ran structured analysis workshops with maintenance and technical teams. Using Ishikawa diagrams, brainstorming, and operational know-how, we identified root causes such as:
- Excessive wear in braking systems
- Brushes misaligned from the rail causing cascading damage
- Detached plastics impacting other segments
- Difficulty pinpointing failure origins due to lack of segmentation in the electrified monorail
Improve
We designed ad hoc solutions for each root cause: redesigns, tolerance adjustments, physical barriers, etc. Improvements were implemented and their impact measured.
Control
We introduced visual controls, checklists, and periodic inspections. Crucially, we strengthened the team’s skills with training in industrial networks and cost management to contextualize every intervention. It was a shift in mindset: maintenance wasn’t just about fixing problems, but preventing them.
Constructive reflection: This phase was critical in building a culture of structured improvement. However, it heavily relied on individual initiative. Data was incomplete, manually entered, and often biased. Analysis was mostly descriptive and based on human interpretation. Anticipating failures or prioritizing actions with economic insight was not yet feasible. Still, it laid a strong foundation.
Phase 2 – Digitalization Begins (2008–2014)
Applying DMAIC in the age of connected automation
Define
Though no longer a specific project, a general goal emerged: improve the reliability and traceability of conveyor systems through advanced automation. The focus shifted from ad hoc fixes to integrated visibility of plant-wide incidents.
Measure
With the rollout of Siemens S7 PLCs and TIA Portal, we began capturing real-time data from key process variables. Manual input was replaced by structured, continuous data acquisition.
Analyze
Thanks to better data quality and traceability (by section, shift, operator), we began seeing failure patterns. Dashboards revealed trends: concentrations of faults at certain times, on specific lines, or under defined operating conditions.
Improve
PLCs were reprogrammed to trigger early alarms. Some incidents even auto-generated SAP PM work orders. Maintenance routines were refined and inspections tailored using the new sensor infrastructure. Communications modules and Ethernet links were added, often without a data strategy—but they were building future readiness.
Control
We established monitoring routines with SCADA systems and connected platforms. Integration between plant systems and SAP was expanded. Control now involved digitally skilled technicians—but still lacked full analytical maturity.
Constructive reflection: Despite the clear infrastructure leap, the DMAIC cycle often stopped short. Improvements were localized, and analysis/control lacked deep analytics or strategic use of data. It was digitalization, but not yet data-driven management.
Phase 3 – From Infrastructure to Machine Learning (a vision for today)
Applying DMAIC in the era of industrial AI (hypothetically)
Unlike the previous phases, this third one is not a lived experience but a hypothetical reflection. After that stage, I moved into other professional domains and didn’t witness how the digital infrastructure evolved. At the time, I wasn’t fully aware of the true potential of data. However, after gaining in-depth training in data science and artificial intelligence, I now clearly see what could (and should) be done.
Looking back with this new lens, I recognize many missed opportunities—patterns, correlations, and decisions that went unnoticed simply because the data, tools, or analytical mindset weren’t there.
Define
We would start with a clear ambition: anticipate failures, prioritize resources, and maximize operational efficiency using AI and data science. Unlike previous stages, the real challenge here would be organizational—defining new roles, skills, and a culture built on data.
Measure
With IoT sensors and automated data capture from modern PLCs, we could measure critical variables like vibration, energy consumption, temperature, or cycles. This information, stored in SQL, MongoDB, or cloud-based data lakes, would become the new raw material of continuous improvement.
Analyze
We would apply machine learning models to uncover nonlinear relationships and hidden operational segments. Libraries like scikit-learn, xgboost, lightgbm, and catboost could support classification and regression; DBSCAN, HDBSCAN, and k-means for unsupervised clustering. Dimensionality reduction with PCA, t-SNE, and UMAP would help visualize latent patterns, while time-series forecasting with prophet, ARIMA, Facebook Kats, or even LSTM would predict cyclical behaviors.
Improve
Improvements would no longer come from intuition alone. We would build Digital Twins with tools like AnyLogic, SimPy, or API-connected platforms to simulate real and synthetic data scenarios. Maintenance planning could be optimized using optuna, hyperopt, or scikit-optimize, applying evolutionary and Bayesian algorithms. Prioritization would combine Monte Carlo simulations and scoring models based on business impact.
Control
Control would be achieved through dashboards built with Power BI, Streamlit, Plotly Dash, or Grafana, offering real-time KPI tracking and root cause traceability. Early warning systems could be integrated into workflows using Apache Airflow or Prefect, triggering automated actions. But the real control would come from trained professionals—data scientists on the shop floor and machine learning engineers in operations—who ensure the models evolve and translate into value.
Constructive reflection (future): This approach will only work if companies invest in people. It’s not about importing models; it’s about developing internal capabilities to understand, train, and act on data insights. Without human transformation, AI in manufacturing will remain just another unfulfilled promise.
Conclusion
The DMAIC cycle is more relevant than ever. What has changed is the speed, power, and depth with which we can go through it, thanks to modern technology.
What used to be captured in a notebook can now be predicted with sensors, algorithms, and intelligent visualizations. But the transformation has not only been digital—it has also been cultural.
As I highlighted in another article —Bridges are always a good idea (though we don’t always see it clearly)— this evolution requires more than infrastructure: it demands human bridges. Because data doesn’t speak for itself; it needs interpreters committed to impact.
That’s why, beyond models and algorithms, we still need something essential:
people with operational insight who understand the value of data and know how to speak its language.
And you, what phase are you in?
- What cultural, technical, or structural barriers have you encountered when trying to move from a traditional to a data-driven approach?
- What kind of profiles (or combination of profiles) do you think an industrial plant really needs to embrace impactful AI?
- How can we assess whether an organization is truly ready to apply continuous improvement driven by artificial intelligence?