MQTT is widely used in IoT systems but remains vulnerable due to its lightweight design. This paper proposes an interpretable deep learning-based intrusion detection framework that processes raw PCAP data through flow-based analysis. It introduces MQTTFlowLyzer for extracting protocol-aware features and presents the BCCC-IoT-MQTT-IDS-2025 dataset, which includes diverse attack scenarios. The framework leverages TabNet, GANDALF, and NODE to enable accurate and interpretable detection of known and novel attacks. Results show strong performance across attack types, with attention-based explanations providing insights into behavioral patterns and supporting zero-day threat identification.