Isolation of abnormal changes in process variables is an integral component of fault diagnosis, as it provides evidential information for determining the root cause of a detected abnormal event. This task is challenging when the approach to diagnosis does not incorporate knowledge of the process' nominal behavior, but is instead established solely on historical process data. Though isolation of abnormal changes in variables may be facilitated by including historical process data for faults that have been previously diagnosed, inconclusive results will remain for unfamiliar faults. This paper presents a method for isolating abnormal changes in process variables with an autoencoder (AE) - a type of neural network configured for latent projection - and without prior knowledge of nominal process behavior or faults. The AE is optimized with nominal process data as well as a sparsity constraint to produce a sparse network. Probing into the sparse AE allows one to gain insight into the correlations that exist among the process variables during normal process operation. Movements in the AE's reconstruction space are interrogated alongside the acquired knowledge to isolate the abnormal changes in process variables. The method is demonstrated with a simulation of a nonlinear triple tank process, and is shown to isolate abnormal changes in variables for both simple and complex faults. (C) 2021 Elsevier Ltd. All rights reserved.