Analysis code and processed data for "Weight-based estimation of juice extraction rate and cumulative juice mass using artificial intelligence"

Abstract

This repository provides the analysis code, processed data, and model-ready inputs supporting the study “Weight-based estimation of juice extraction rate and cumulative juice mass using artificial intelligence.”The study evaluates six regression models—Gradient Boosting, CNN, LSTM, and adaptations of PatchTST, iTransformer, and TimeMixer—for estimating juice extraction rates from remaining raw material mass, operating settings, and material identity. The dataset comprises 48 experimental batches: nine orange, 25 pear, nine apple, and five carrot batches. Repository contents• Processed observations: 258,856 records after removal of identified erroneous measurements and reconstruction of the analysis time axis.• Model-ready inputs: 31,738 original evaluation time points, including 21-step histories of remaining material mass, extraction speed settings, and feeding speed settings, together with four material indicators and 5-s and 10-s mass-change rates.• Fixed batch assignments for five-fold cross-validation and fold-specific augmentation inputs generated exclusively from training batches.• Python code for hyperparameter search, model training, validation-based model selection, held-out test evaluation, and result plotting.• Reference results, including scores for 1,200 hyperparameter trials, 60 selected fold-specific configurations, and overall, fold-level, and batch-level evaluation metrics.• Installation requirements, execution instructions, data documentation, environment records, and file-integrity checks. Experimental designOriginal-data training is compared with training supplemented by synthetic sequences assembled from segments of the same raw material. For each model, the same 20 hyperparameter configurations are evaluated across five folds and both training conditions. Within each fold, the configuration with the lowest validation extraction-rate RMSE is selected and evaluated on the corresponding held-out test batches. Reported overall performance is calculated by pooling test predictions across the five folds, with each of the 48 batches evaluated once.The reference extraction rate is defined as the least-squares slope of cumulative juice mass over the preceding 10 s. The task therefore concerns current extraction-rate estimation rather than future forecasting. Cumulative juice mass is reconstructed by trapezoidal integration of estimated rates, initialized with the measured cumulative mass at the first evaluation time. Scope and reuseThis release supports reproduction of the model comparison from the supplied processed data and fixed model-ready inputs. It does not include the unmodified sensor acquisition files, the complete raw-data preprocessing and input-generation pipeline, or trained model checkpoints. Training the models generates new checkpoints. The plotting code produces analysis figures but does not include every manuscript-specific layout script.Synthetic sequences are provided for training augmentation and should not be interpreted as measured physical responses to operating-speed transitions. Numerical results may vary with software versions, hardware, and numerical execution settings.Please consult README.md, README_KO.md, and docs/DATA_DICTIONARY.md for installation, execution, and data interpretation. Third-party source information and license notices are retained in the accompanying files.