Integration of CNN and Attention Mechanism in Fault Identification of Substation Centralized Control Systems
Traditional methods of fault detection do not offer sufficient sensitivity toward significant local characteristics, leaving unsatisfactory results in both detection accuracy and response time. Thus, this article proposes a Convolution Neural Network model with an implementation of a Squeeze-and-Excitation layer, where multi-source time-series sensor data passes through one-dimensional convolutional layer for extracting local features, followed by the global average and pooling of the feature map to get channel statistics, followed by the processing of the information in a dimensionality-reducing fully connected layer with activation through the ReLU function. The dimensionality-increase fully connected layer outputs sigmoid-normalized channel weights. These weights are then multiplied by the original feature map on a channel-by-channel basis; the calibrated features are then fed into a fully connected classification layer to complete fault identification. Results demonstrate a 99.4% recognition accuracy for voltage mutation faults, with an average response time of 47.26ms, validating the key role of this approach in improving the real-time and robustness of power grid fault diagnosis.