The reconstruction error-based anomaly detection refers to constructing a data reconstruction model to determine whether a sample is an anomaly based on the reconstruction error on the test set. We propose a network architecture consisting of three channels, including random dropping, causal, and dilated convolution modules. The random dropping module is employed to stochastically discard sample information, compelling the network to rely on local context for current information reconstruction. The casual convolution module captures temporal dependencies in the data, and the dilated convolution module expands the convolutional receptive field. Together, these modules enhance the model’s ability to understand and reconstruct complex temporal patterns in 1D real-time signals. The proposed architecture can learn more comprehensive features through 3 scales, providing greater robustness. Validation on lithium-ion cell discharge voltage data demonstrates that our approach outperforms the baseline in terms of performance.