The transition toward Industry 4.0 has established data-driven forecasting as a cornerstone for enhancing operational reliability and optimizing predictive maintenance strategies. A preliminary study in this field, exploiting Gaussian Process Regression (GPR), has demonstrated the feasibility of long-term predictions, with increasing accuracy when decreasing forecast's horizon. To further enhance forecasting performance, this work proposes a Deep Learning framework based on Long Short-Term Memory (LSTM) networks. Although LSTMs, as artificial neural networks, require substantial datasets to achieve high generalization, such data volumes are increasingly available within modern industrial infrastructures. Thanks to the employment a direct multi-step forecasting strategy, the proposed LSTM architecture effectively captures long-range dependencies while avoiding the error accumulation typical of recursive models. The results obtained in this context proved superior to GPR baselines, as the LSTM's ability to model complex non-linear temporal patterns provides greater stability and precision over extended horizons. The proposed model reaches a substantially lower RMSE, with reductions of up to an order of magnitude. The methodology was validated through two distinct industrial case studies: a batch fermentation process and an industrial furnace subject to fouling. This dual validation confirms the robustness and adaptability of the framework across different sectors. Finally, the validated methodology has been integrated into PREDator, an open-access Python-based application with a graphical user interface designed to streamline the deployment of advanced predictive tools in both industrial and academic environments.
La transizione verso l’Industria 4.0 ha reso la previsione basata sui dati un elemento fondamentale per migliorare l’affidabilità operativa e ottimizzare le strategie di manutenzione predittiva. Tuttavia, persiste una significativa dicotomia tra previsioni a breve termine ad alta accuratezza e la capacità di effettuare previsioni robuste a lungo termine, necessarie per il processo decisionale industriale. I modelli ricorsivi standard soffrono spesso di accumulo di errore e \textit{exposure bias}, limitando la loro efficacia su orizzonti estesi. Questo lavoro affronta tali limitazioni proponendo un framework di Deep Learning basato su reti neurali ricorrenti Long Short-Term Memory (LSTM-RNN), che utilizza una strategia di previsione diretta multi-step per catturare dipendenze di lungo raggio senza richiedere l’iniezione di \textit{ground truth}. La metodologia viene validata attraverso due distinti casi di studio industriali: un processo di fermentazione batch e un forno industriale soggetto a \textit{fouling}. Un’analisi comparativa con Gaussian Process Regression (GPR) e la regressione polinomiale mostra che, sebbene i modelli probabilistici siano efficaci nel cogliere tendenze locali, l’architettura LSTM proposta garantisce una stabilità e una precisione superiori negli scenari di lungo periodo. Infine, la metodologia è integrata in PREDator, un’applicazione Python dotata di interfaccia grafica progettata per semplificare l’implementazione di questi avanzati strumenti predittivi sia in contesti industriali che accademici.
PREDator: an open-access deep learning framework for long-term industrial time series forecasting
Mazzitelli, Antonio
2024/2025
Abstract
The transition toward Industry 4.0 has established data-driven forecasting as a cornerstone for enhancing operational reliability and optimizing predictive maintenance strategies. A preliminary study in this field, exploiting Gaussian Process Regression (GPR), has demonstrated the feasibility of long-term predictions, with increasing accuracy when decreasing forecast's horizon. To further enhance forecasting performance, this work proposes a Deep Learning framework based on Long Short-Term Memory (LSTM) networks. Although LSTMs, as artificial neural networks, require substantial datasets to achieve high generalization, such data volumes are increasingly available within modern industrial infrastructures. Thanks to the employment a direct multi-step forecasting strategy, the proposed LSTM architecture effectively captures long-range dependencies while avoiding the error accumulation typical of recursive models. The results obtained in this context proved superior to GPR baselines, as the LSTM's ability to model complex non-linear temporal patterns provides greater stability and precision over extended horizons. The proposed model reaches a substantially lower RMSE, with reductions of up to an order of magnitude. The methodology was validated through two distinct industrial case studies: a batch fermentation process and an industrial furnace subject to fouling. This dual validation confirms the robustness and adaptability of the framework across different sectors. Finally, the validated methodology has been integrated into PREDator, an open-access Python-based application with a graphical user interface designed to streamline the deployment of advanced predictive tools in both industrial and academic environments.| File | Dimensione | Formato | |
|---|---|---|---|
|
Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
2.17 MB
Formato
Adobe PDF
|
2.17 MB | Adobe PDF | Visualizza/Apri |
|
PREDator: An Open-Access Deep Learning Framework for Long-Term Industrial Time Series Forecasting.pdf
accessibile in internet per tutti
Descrizione: PREDator: An Open-Access Deep Learning Framework for Long-Term Industrial Time Series Forecasting
Dimensione
7.6 MB
Formato
Adobe PDF
|
7.6 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/253455