Dataset and reproducibility package for robustness screening of temperature-compensated ATR-MIR milk-protein soft sensors
Description
This dataset is described in the Data in Brief article “Dataset and reproducibility package for robustness screening of temperature-compensated ATR-MIR milk-protein soft sensors” and supports the related Journal of Process Control manuscript “Robustness-Qualified Screening of Temperature-Compensated ATR-MIR Soft Sensors for Low-Solids Milk-Protein Monitoring”. The package contains the experimental-source inputs, MATLAB code, simulation definitions, analysis outputs and figure-source files used to evaluate temperature-compensated attenuated total reflectance mid-infrared (ATR-MIR) soft-sensor signals for low-solids milk-protein monitoring. It is based on seven fixed spectral-window protein-prediction signals obtained from one 5.06% reconstituted skim-milk formulation over a measured temperature interval of 2–20 °C and referenced to a matched Kjeldahl protein value of 1.610%. Repository contents include source-input tables, temperature-associated prediction-bias values, compensation-function coefficients, three 600-second temperature profiles, controller and plant settings, nominal screening outputs, 50-seed stochastic results, transport-delay sweeps, coefficient-perturbation outputs, acceptance-band checks, plant and controller-tuning sensitivity results, mechanistic trace files and figure-source data. MATLAB scripts, dependency information, random-seed lists, execution order, checksums and a complete file inventory are provided to support computational reproduction. The files may be reused to compare alternative compensation laws, ranking criteria, robustness thresholds, soft-sensor conditioning methods and supervisory-control assumptions within the stated low-solids, instrument and temperature boundaries. The package is intended as a method-development and computational-reproducibility resource. It does not constitute population-scale, multi-batch, cross-instrument, pilot-scale or plant-level validation data.
Files
Steps to reproduce
This repository supports the article “Robustness-Qualified Screening of Temperature-Compensated ATR-MIR Soft Sensors for Low-Solids Milk-Protein Monitoring.” The repository contains the source data, MATLAB scripts, compensation coefficient files, simulation outputs, robustness outputs, targeted final-check outputs, figure-source data, supplementary tables, and run-order documentation required to reproduce the main article findings. The computational workflow uses a low-solids ATR-MIR milk-protein source dataset derived from a separate experimental data record: Stefanidis, K. (2026), Temperature sensitivity of ATR-MIR protein prediction in a single low-solids reconstituted bovine skim milk formulation, Mendeley Data, V1, doi: 10.17632/mf2pgbjrbd.1. The active source dataset comprises one low-solids reconstituted bovine skim-milk formulation, a matched 1.610% Kjeldahl protein benchmark, seven fixed ATR-MIR spectral-window protein-prediction signals, and a 2–20 °C passive-warming sequence at 2, 5, 8, 11, 14, 17, and 20 °C. For the present article, those low-solids source predictions were converted into temperature-associated prediction-bias functions and compensation-law candidates. The MATLAB workflow then evaluates corrected and uncorrected protein signals as process-facing soft-sensor outputs in a supervisory benchmark. The nominal benchmark uses three within-envelope temperature profiles, Drift600, Spike600-20, and Cycle600, and two screening objectives, ProductionAcceptable and BalancedRecovery. The robustness workflow applies 50-seed stochastic testing, transport-delay testing at 0, 5, 10, and 20 s, targeted acceptance-band checks, targeted ±10% coefficient perturbation, plant/tuning sensitivity, and mechanistic trace export. The principal result reproduced by this repository is the validation-priority classification reported in the article. The primary promoted candidate is Cycle600 / ProductionAcceptable / Full MIR + PIDPHI + poly5. This candidate is retained in the nominal, stochastic, short-delay, and acceptance-band layers, but is coefficient-sensitive and plant/tuning-conditional. The repository therefore supports a reproducible soft-sensor screening workflow and validation-priority interpretation; it does not contain completed instrument-level, plant, or production-line validation data.
Institutions
- University of Southern QueenslandQueensland, Toowoomba