Detecção de defeitos em máquinas de fabricação de copos de papel utilizando sinais acústicos e de vibração com algoritmos de aprendizado de máquina e redes neurais

Abstract

Early detection of defects in industrial machinery is essential to ensure high operational availability, reduce maintenance costs, and prevent productivity losses, particularly within the context of Industry 4.0. This study proposes and evaluates an approach for fault diagnosis in paper cup manufacturing machines by integrating acoustic and vibration signals to detect slack in Chain 1 of the transmission system, a critical component for equipment synchronization. Data were collected during a real production process in a packaging company, ensuring practical representativeness of operational conditions. Nineteen experimental acquisitions were conducted based on a structured Design of Experiments (DOE), ensuring systematic variation of machine operating conditions. In total, approximately 50 minutes of recordings were obtained. The acquired signals were segmented into 5-second windows with 50% overlap, resulting in a dataset comprising 1,453 instances. The modeling stage was structured into two complementary approaches. The first was based on manual feature engineering, extracting statistical and spectral descriptors from audio and vibration signals and using them as input to traditional statistical models (Logistic Regression and Linear Discriminant Analysis) and classical machine learning algorithms, including Random Forest, Support Vector Machine, Multilayer Perceptron, and a soft-voting Ensemble model. The second approach employed deep learning through Convolutional Neural Networks (CNNs) applied to Mel spectrograms extracted exclusively from the audio signal, enabling automatic learning of relevant time–frequency representations for fault diagnosis. Model evaluation was performed using stratified internal validation and an external test set composed of completely unseen experimental runs, ensuring a rigorous estimation of generalization capability. To mitigate stochastic effects, experiments were repeated across multiple independent executions, and average performance metrics along with their confidence intervals were analyzed. Results indicate that multisensory integration of audio and vibration signals significantly improved performance and robustness compared to single-sensor approaches, substantially reducing the gap between validation and external testing. The feature-based approach achieved an average external test accuracy of 94% in its best multisensory configuration, while the CNN-based approach reached an average accuracy of 92% using audio alone, demonstrating competitive performance even under less intrusive instrumentation. Overall, the findings confirm that multisource sensory integration enhances diagnostic robustness, while deep learning approaches provide a promising alternative in scenarios with instrumentation constraints. The proposed method shows strong potential for predictive maintenance systems and quantitatively advances the application of Artificial Intelligence techniques in industrial fault diagnosis.