Fig. 1: Detection framework based on the model-constrained deep learning network.

The light blue part of the diagram constitutes the MCNN; a Modeling and state resolution of individual and overall control models: this model identifies real-time states through a specially designed pack control model and state estimator. The output state prediction parameters constrain the regression process of the network. b Overall state sequence prediction: this network combines battery state and vehicle state data to predict the pack state through a Bi-directional Long Short-Term Memory (BiLSTM)-based network. c Encoder based on the model coupling relationship: this encoder takes the outputs of (a) and (b), along with vehicle state data, as network inputs. It compresses and fuses hidden variables based on the physical significance of the model as a whole and deviations and obtains the single predicted state through decompression. d Residual monitoring module: this module extracts features from the outputs of (c), combining a directional linear layer and statistical regression for fault detection and fault classification. (#) Detailed presentation: this represents the corresponding network parts, including corresponding data flow, processing algorithm, network composition, etc.