Integrating Deep Autoencoders and Bayesian Inference for Diagnostics and Health Management of Industrial Inkjet Systems

Abstract

Modern printing industry requires extreme reliability to achieve zero-defect production. This study contributes to the paradigm shift from reactive to predictive maintenance in industrial inkjet systems, reducing downtime, material waste, and operational costs. We present an integrated Diagnostics and Health Monitoring strategy leveraging piezoelectric self-sensing with deep autoencoders and Bayesian inference for real-time fault diagnostics. The system captures residual pressure waves in the ink chamber—exploiting the dual actuator-sensor function of piezoelectric crystals—as "acoustic signatures" to detect subtle nozzle deviations from clogging, air bubbles, or mechanical wear. Nozzle pressure signals feed into a multimodal autoencoder (AE) architecture, where each AE specializes in a distinct fault class (jetting, non-jetting, deviated, intermittent). AE outputs combine with Gaussian Mixture Models (GMM) and Bayesian inference to provide high-confidence classification, even with imbalanced industrial datasets. Tests on industrial printheads (Ricoh MH5420) demonstrate 99.4% accuracy in detecting critical failures and enabling preventive maintenance. However, while the system excels at detecting fluidic obstructions, challenges remain in classifying deviated and intermittent faults. The prognostic layer reuses the jetting autoencoder's reconstruction error (RE) as a continuous Health Indicator, correlating pressure-induced degradation with nozzle health. Controlled experiments varying ink chamber pressure reveal a parabolic RE-pressure relationship, with minimum RE at nominal operating range. This enables early degradation detection, as RE increases progressively before functional failure, supporting condition-based maintenance strategies.