Monitoring the condition of belt conveyor idlers is critical for ensuring safe and efficient operation of industrial conveying systems. However, existing methods often suffer from limited scalability and delayed fault detection, particularly under variable environmental conditions. In this work, we propose MIRA, a transformer-based framework for monitoring idler roller anomalies, which detects and localizes faults using acoustic and vibration signals collected from low-cost sensors. MIRA employs a masked transformer-based autoencoder trained in an unsupervised manner to reconstruct normal patterns and detect deviations via reconstruction error. MIRA can also infer the fault location, enabling spatially aware anomaly detection without the need for labeled data. We validated the system on a custom-built conveyor belt testbed equipped with sensor units, each measuring sound and two-axis vibration signals. We evaluated MIRA on four types of idler faults across 14 roller locations and 6 belt speeds. The results show that MIRA achieves an anomaly detection accuracy of 98.70% and a fault localization accuracy of 96.09%, demonstrating its robustness and practical applicability in complex operational settings.
Mobile interaction is fundamentally tied to the screen, requiring users to look at and physically touch their devices by tapping on-screen controls. We introduce AcouLetter, a system for recognizing fine-grained, mid-air alphabetic thumb gestures above the screen that enables one-handed and non-contact input. Our method fuses acoustic signals with inertial sensor data to discern the subtle characteristics of each letter gesture, capturing the dynamics of the hand’s motion. We employ a multi-stage architecture that integrates a convolutional network for acoustic features and a transformer-based network for inertial data. AcouLetter first detects the initiation of the gesture using inertial measurements, determines which hand holds the phone, and then classifies the specific letter among 27 classes, including the 26 letters of the alphabet and a null gesture. The system achieves high accuracy and demonstrates robust performance even under ambient noise while a user is standing or walking. Since letter-shaped gestures are meaningful and readily recognizable, they support an intuitive system of expressive shortcuts in which letters act as mnemonics for complex functions. A user study further confirmed the practicality of this approach for one-handed, contact-free interaction. Our results demonstrate that this expressive alphabetic gestural interface is a feasible new modality for commodity smartphones.