The prognosis of thermo-acoustic/combustion instability is usually accomplished by applying a priori knowledge about features of unstable operation and measuring deviation from those features using point values. In the present work, we adopt a different methodology, whereby the presence and extent of the signature of unstable combustion are learnt as an anomaly from the distribution of pressure oscillations during stable operation across several protocols. The protocols involve a transition from stable to unstable operation in a swirl combustor. It is inferred that the stable combustion in the present case is stochastic noise with a normal distribution containing values comparable with rootmean-square values at unstable operation with a p value 0.05–0.07. We exploit this feature to detect anomalies from flame intensity images, which represents the heat release rate fluctuations by manipulating their features to be a part of multivariate Gaussian distribution. To formulate this distribution, we employ a convolutional-neural-networkbased variational auto-encoder (CNN-VAE) and express the associated reconstruction loss as an anomaly metric. The anomalies obtained through CNN-VAE and integrated intensity fluctuations are then evaluated for their sensitivity against the unsteady pressure data. The analysis reveals that the CNN-VAE metric performs better than the integrated intensity fluctuations for predominantly all p values. © 2023, AIAA International. All rights reserved.