In practical resource-constrained environments, efficiently extracting the potential high-frequency fault-critical information is an inherent problem. To overcome this problem, this work suggests leveraging a bi-residual neural network named Bi-ResNet to extract the inner spatial-temporal high-frequency features using embedded spatial-temporal convolution blocks and intra-link layers. It can be considered as embedding a high-frequency extractor into networks without adding any parameters, helping shallow networks achieve the performance of deep networks. In our experiments, five advanced CNN-based neural networks and two baselines across a real-life dataset are utilized for synchronous motor electrical fault diagnosis to demonstrate the effectiveness of Bi-ResNet including one analytical, comparative, and ablation experiments. The corresponding experiments show: 1) The Bi-ResNet can perform better on low-resolution noisy data. 2) The proposed intra-links can help high-frequency components extraction and location from raw data. 3) There is a trade-off between intra-link number and input data complexity.