Industry 4.0 has made predictive maintenance a practical necessity, but the dominant cloud-centric monitoring pipeline is increasingly limited by bandwidth cost, latency, and data confidentiality. This has accelerated a shift toward TinyML, where anomaly detection runs directly on edge microcontrollers. However, deployment on Cortex-M–class devices is constrained by a hard memory ceiling (2 MB), creating a “hardware conflict” for standard reconstruction-based solutions such as autoencoders: the decoder imposes a prohibitive memory tax, and point-wise reconstruction objectives tend to model nuisance detail (e.g., background noise) rather than fault-relevant physics, making detectors brittle under domain shift. This thesis proposes a decoder-free alternative: a self-supervised classification (SSC) framework for multi-sensor anomaly detection that replaces reconstruction with a discriminative pretext task, Sensor Channel Identification.