Fig. 4: The evolution of embeddings of all fifteen abnormal vehicles in Dataset Dahu from the input layer to the output layer of the dynamical autoencoder model with t-distributed stochastic neighbor embedding visualization. | Nature Communications

Fig. 4: The evolution of embeddings of all fifteen abnormal vehicles in Dataset Dahu from the input layer to the output layer of the dynamical autoencoder model with t-distributed stochastic neighbor embedding visualization.

From: Realistic fault detection of li-ion battery via dynamical deep learning

Fig. 4: The evolution of embeddings of all fifteen abnormal vehicles in Dataset Dahu from the input layer to the output layer of the dynamical autoencoder model with t-distributed stochastic neighbor embedding visualization.

Each subplot visualizes the dimension-reduced data of (a) the data input, (b) the inferred latent variable, and (c) the reconstructed observation. All snippets are identified as either abnormal or normal ones, marked as purple or green points in the output layer, respectively. One EV (named EV1) is highlighted. Its snippets are marked as red and blue points according to their abnormality determined by the dynamical autoencoder model. The predicted and observed curves of the three snippets (marked as ▴, ⋆, and •) selected from EV1 include two dimensions: maximum battery temperature, and minimum cell voltage. d The predicted, observed, and corresponding error curves of three charging snippets marked as ▴, ⋆, and •.

Back to article page