Transformer multimode anomaly detection method based on convolution and LSTM unit

Traditional methods for detecting transformer multimodal data often suffer from complex network architectures and nonlinear mapping relationships, leading to ambiguous decision-making processes and unreliable output results. To address this, a multimodal anomaly detection method for transformers based on convolutional networks and LSTM units is proposed. The method calculates the average distance between data points within a sliding window and their spatial center, then compares this distance with a predefined threshold to identify potential anomalous data points. A cross-modal sensing module, composed of a linear layer, a multimodal feature fusion layer, and a feedforward network, is designed. In addition, a dynamic threshold model for transformer operational states is introduced, building upon traditional convolutional neural networks. By incorporating an expansion rate parameter and integrating control mechanisms such as the forget gate, input gate, and output gate into the long short-term memory network, the proposed transformer multimodal anomaly detection model is constructed. The experimental results demonstrate that the detection outcomes of the proposed method align with actual anomaly occurrences, with predicted values closely matching real measurements, achieving an accuracy of ∼98.8%. This approach effectively enables anomaly detection in transformer multimodal data, significantly enhancing the stability and security of power systems.