As demand grows for efficient edge computing systems, innovative architectures are crucial for achieving low-power, high-density data processing in resource-constrained environments. Compressed sensing (CS) and Analog In-Memory Computing (AIMC) offer promising pathways to meet these needs by enabling localized, efficient feature extraction and inference. This paper introduces an energy-efficient on-chip system that integrates CS with AIMC based on Phase-Change Memory (PCM) devices to enable robust feature extraction and inference. The proposed architecture employs CS for dimensionality reduction at the sensor level, generating low-dimensional feature vectors directly fed into a single-layer artificial neural network (ANN) implemented on PCM crossbars. To address inherent hardware non-idealities, we utilize hardware-aware (HWA) training combined with an attention-based regularization mechanism, improving both inference stability and drift resilience over extended periods. Performance evaluation on a face recognition task demonstrates that attention-enhanced HWA training effectively mitigates overfitting and maintains model accuracy under PCM drift conditions, highlighting the system's suitability for edge computing applications requiring low power consumption and long-term reliability.