This paper proposes a novel technique for online unsupervised detection and isolation of abrupt critical faults. It is assumed that the system starts from a healthy condition while a fault can occur at any time point. The data stream is first processed by the gating model designed to detect and learn the characteristics of an unknown state, without forgetting information about previous states. During the optimization process, the loss function of the gating model is used as an indication of the fault appearance, while regularization and replay strategies are utilized to prevent information loss. An autoencoder is associated with each known state, which allows performing state classification by comparing reconstruction loss. Validation was conducted on test scenarios designed from two popular benchmarks: the Case Western Reserve University dataset for bearing faults and the dataset from Tennessee Eastman Process for chemical process faults. The focus was on critical abrupt failures which should not be undetected. The achieved results demonstrated the capabilities of the proposed model to memorize each fault, regardless of the introduction order, while achieving near-perfect fault isolation.