NA-BEATs wins DCASE 2026 with 70.24% score, beating rivals by 5%
Two microphones and self-supervised learning detect machine faults with record accuracy
In a new paper on arXiv, Fujimura and colleagues from Mitsubishi Electric and other institutions tackle a real-world problem: detecting machine anomalies in noisy factories. They propose NA-SSL (noise-aware self-supervised learning), a framework that leverages two-channel audio recordings. One microphone sits close to the target machine, capturing its operational sounds, while a second, farther-away microphone records ambient background noise. The model uses the far microphone as auxiliary information to extract cleaner self-supervised learning (SSL) representations from the close-microphone signal, effectively isolating machine-specific patterns from environmental interference.
Tested on the official DCASE 2026 Challenge Task 2 development set, the NA-SSL frontends improved performance across three base SSL models: BEATs, EAT, and Dasheng, regardless of fine-tuning. The standout system, NA-BEATs, won the challenge outright with an official score of 70.24%, a substantial margin over the runner-up's 65.46%. This noise-aware design proves that incorporating spatial noise context into SSL pretraining significantly boosts anomaly detection robustness, bringing industrial predictive maintenance closer to real-world deployment.
- NA-SSL uses two microphones: one near the machine, one far away to capture background noise
- NA-BEATs model won DCASE 2026 Challenge Task 2 with 70.24% official score, 4.78% higher than second place
- Framework boosts three SSL models (BEATs, EAT, Dasheng) even without discriminative fine-tuning
Why It Matters
Enables reliable machine monitoring in noisy factories, reducing false alarms and unplanned downtime.