Plastic recycling plants are essential to achieving the circular economy goals proposed by the European Commission. However, these plants often face operational inefficiencies and equipment breakdowns due to conditions that are difficult to monitor and predict. In particular, the grinder machine, used to shred plastic waste into flakes, experiences significant wear, necessitating blade replacement on a weekly basis. This preventive maintenance strategy does not consider the actual wear state or the remaining cutting power of the blades, leading to inefficiencies. In recent years, research on predictive maintenance has surged, driven by the Industry 4.0 revolution. These approaches typically begin with data collection and proceed with the development of models capable of capturing the actual health status of industrial machines. Inspired by these advancements, this work aims to forecast the optimal maintenance timing for the grinder blades, enabling timely scheduling and efficient allocation of maintenance resources. The study began with real-world data acquisition from an operating PET recycling plant, integrating existing automation infrastructure with a custom-built acquisition framework. This setup captured key signals such as motor current, production rate, temperature, and vibrations from the two bearings of the grinder. A remote web connection enabled the extraction of this data from a central database for offline analysis and pattern discovery. To determine the optimal maintenance timing, an optimization algorithm was developed to maximize the estimated annual production for each potential maintenance time within a production cycle. The algorithm produced a binary target vector, where positive labels indicated the need for maintenance. This vector was then used to train machine learning models to predict maintenance needs in advance. A range of models was evaluated, starting with a baseline logistic regression model, followed by decision forest algorithms, and culminating in advanced deep learning architectures. While the logistic model performed poorly due to the non-linearity nature of the problem, more complex models achieved high predictive accuracy, with some reaching up to 98%. The thesis concludes with a discussion of current limitations and possible future improvements to enhance robustness and adaptability. To the best of the author’s knowledge, this is the first predictive maintenance framework applied to a mechanical recycling grinder. The methodology developed in this work contributes to improving the efficiency of plastic recycling operations and supports the strategic objectives of Industry 4.0 and circular economy, paving the way for the next generation of smart recycling plants.
Gli impianti di riciclo della plastica sono fondamentali per il raggiungimento degli obiettivi di economia circolare promossi dalla Commissione Europea. Tuttavia, questi impianti affrontano frequentemente inefficienze operative e guasti ai macchinari, causati da condizioni difficili da monitorare e prevedere. In particolare, la macchina trituratrice, utilizzata per ridurre i rifiuti plastici in scaglie, è soggetta a un’elevata usura, rendendo necessaria la sostituzione delle lame su base settimanale. Questa strategia di manutenzione preventiva non tiene conto dello stato reale di usura né della capacità di taglio residua delle lame, generando inefficienze. Negli ultimi anni, la ricerca sulla manutenzione predittiva ha raggiunto una forte crescita, trainata dalla rivoluzione dell’Industria 4.0. Questi approcci iniziano generalmente con la raccolta dei dati, per poi sviluppare modelli capaci di catturare lo stato di salute reale delle macchine industriali. Ispirato da questi sviluppi, questo lavoro si propone di prevedere in anticipo il momento ottimale per effettuare la manutenzione delle lame della macchina trituratrice, consentendo una pianificazione tempestiva e un’allocazione efficiente delle risorse. Lo studio è stato avviato con l’acquisizione di dati reali da un impianto di riciclo di bottiglie in PET, integrando l’infrastruttura di automazione esistente con un framework di acquisizione personalizzato. Il sistema ha registrato segnali chiave come la corrente assorbita dal motore, il tasso di produzione, la temperatura e le vibrazioni misurate sui due cuscinetti della trituratrice. Una connessione web remota ha consentito l’estrazione dei dati da un database centrale per effettuare analisi offline e individuare pattern rilevanti. Per determinare il momento ottimale della manutenzione, è stato sviluppato un algoritmo di ottimizzazione volto a massimizzare la produzione annuale stimata per ciascun potenziale tempo di intervento all’interno di un ciclo produttivo. L’algoritmo ha generato un vettore binario di target, in cui le etichette positive indicano la necessitá di manutenzione. Questo vettore è stato successivamente utilizzato per addestrare modelli di apprendimento automatico in grado di prevedere il bisogno di manutenzione con sufficiente anticipo. Sono stati valutati diversi modelli, a partire da una regressione logistica come modello di base, seguita da algoritmi basati su foreste decisionali, fino ad arrivare ad architetture avanzate di deep learning. Mentre il modello logistico ha mostrato prestazioni modeste a causa della natura non lineare del problema, i modelli piú complessi hanno raggiunto un’elevata accuratezza predittiva, fino al 98%. La tesi si conclude con una discussione sulle attuali limitazioni dell’approccio proposto e sulle possibili evoluzioni future per migliorarne la robustezza e l’adattabilitá. Per quanto a conoscenza dell’autore, questo rappresenta il primo framework di manutenzione predittiva applicato a una trituratrice per il riciclo della plastica. La metodologia sviluppata contribuisce a migliorare l’efficienza degli impianti di riciclo e supporta gli obiettivi strategici dell’Industria 4.0 e dell’economia circolare, aprendo la strada alla prossima generazione di impianti intelligenti.
A machine learning framework for predictive maintenance of a plastic recylicling grinder
Li Manni, Gianluca
2024/2025
Abstract
Plastic recycling plants are essential to achieving the circular economy goals proposed by the European Commission. However, these plants often face operational inefficiencies and equipment breakdowns due to conditions that are difficult to monitor and predict. In particular, the grinder machine, used to shred plastic waste into flakes, experiences significant wear, necessitating blade replacement on a weekly basis. This preventive maintenance strategy does not consider the actual wear state or the remaining cutting power of the blades, leading to inefficiencies. In recent years, research on predictive maintenance has surged, driven by the Industry 4.0 revolution. These approaches typically begin with data collection and proceed with the development of models capable of capturing the actual health status of industrial machines. Inspired by these advancements, this work aims to forecast the optimal maintenance timing for the grinder blades, enabling timely scheduling and efficient allocation of maintenance resources. The study began with real-world data acquisition from an operating PET recycling plant, integrating existing automation infrastructure with a custom-built acquisition framework. This setup captured key signals such as motor current, production rate, temperature, and vibrations from the two bearings of the grinder. A remote web connection enabled the extraction of this data from a central database for offline analysis and pattern discovery. To determine the optimal maintenance timing, an optimization algorithm was developed to maximize the estimated annual production for each potential maintenance time within a production cycle. The algorithm produced a binary target vector, where positive labels indicated the need for maintenance. This vector was then used to train machine learning models to predict maintenance needs in advance. A range of models was evaluated, starting with a baseline logistic regression model, followed by decision forest algorithms, and culminating in advanced deep learning architectures. While the logistic model performed poorly due to the non-linearity nature of the problem, more complex models achieved high predictive accuracy, with some reaching up to 98%. The thesis concludes with a discussion of current limitations and possible future improvements to enhance robustness and adaptability. To the best of the author’s knowledge, this is the first predictive maintenance framework applied to a mechanical recycling grinder. The methodology developed in this work contributes to improving the efficiency of plastic recycling operations and supports the strategic objectives of Industry 4.0 and circular economy, paving the way for the next generation of smart recycling plants.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis_10529079 (1).pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
10.29 MB
Formato
Adobe PDF
|
10.29 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/10589/239157