EN / ES — Español

Case study · Industrial monitoring

Smart membrane monitoring

A module that tracks the health of each membrane train at a desalination plant and estimates when cleaning is due — not too late, not too early.

My role
Solution design and development of its formulas
Industry
Desalination
Validation
6 months on site
Status
In production
Row of blue industrial vessels in a machine room
Illustrative photo, not the plant in this case. Photo by Frantisek Duris on Unsplash.

01 · Context

Membranes that foul with use

A desalination plant produces water by pushing seawater through reverse osmosis membranes grouped in trains. Over time the membranes foul and must be chemically cleaned, taking part of the operation offline. The module was built within an industrial monitoring SaaS platform.

02 · Problem

Cleaning too late or too early both cost

Clean too late and the system works harder and uses more energy. Clean too early and chemicals and operating hours are wasted. And plant signals, with their gaps and outliers, hide the trend.

03 · Approach and method

From noisy signals to a suggested date

  1. 1Build an index

    Operating signals are combined into a health index that is comparable over time.

  2. 2Baseline

    Each cleaning sets a new starting point for the cycle.

  3. 3Filter outliers

    Anomalous values and gaps are excluded from the fit.

  4. 4Project

    A regression estimates when the index will cross the threshold, with its uncertainty.

100806045cleaning thresholdfiltered outliersuggested datedays since last cleaning →
Fig. 1 — The method in one figure: observations, filtered outlier, regression and projection with interval to the threshold. Synthetic data.

04 · Interactive demo

Try the monitor

A conceptual reconstruction of the module: zoom the time window and move the threshold to see how the suggested date, alerts and scenario versus a fixed schedule change.

Membrane monitorSynthetic data
Indicator today69
7-day trend−2.9 pts
Days to threshold16 d
ConfidenceHigh
100806040-119 d-74 d-29 d+15 d+60 dcleaningcleaningthreshold 62today
ObservedForecast and 95% intervalFiltered outlierNo data
Conceptual interface with synthetic data · not the deployed productInvented indicator · fixed seed

05 · Results

What was observed on site

Real results · no figures for confidentiality
  • Decrease: Lower specific energy consumption
  • Decrease: Lower total energy consumption
  • Increase: Higher flow after cleaning

Observed on one train during a 6-month on-site validation. The module contributed alongside other operational improvements: results are not attributed to it alone.

06 · My contribution

I took part in every stage

Technical workManagement and handover
  1. 01Technical workUnderstand
    • Understanding the operational and business problem.
  2. 02Design
    • Proof-of-concept design.
    • Module specifications.
  3. 03Build
    • Data integration.
    • Code for the index and threshold-projection formulas.
  4. 04Management and handoverManage
    • Task backlog and prioritization.
    • Task specification for the team.
  5. 05Validate and hand over
    • Audit of the on-site validation.
    • Handover to operations teams.

07 · Technologies

Tools

Model and analysis

PythonpandasNumPyscikit-learn

Platform (team stack)

DjangoCeleryTime-series database