In order to improve the accuracy of feature extraction and multi-dimensional semantic information expression of Timesnet model, this paper proposes a semantic information extraction method based on graph neural network and Wavelet-Timesnet. The timesnet model introduces wavelet transform to enhance the ability to extract local semantic information in complex environments, and the graph neural network is used to enhance the representation ability of multi-dimensional information. The self-attention mechanism is used in the multi-dimensional semantic extraction model to improve the weight allocation effect between nodes. Taking the rapier loom multi-sensor model as an example, the accuracy of one-dimensional semantic information extraction reaches 99.29%, and the accuracy of multi-dimensional semantic information model reaches 92.8%, which is significantly better than other mainstream feature extraction algorithms. It plays an important role in improving the perception and prediction ability of current complex industrial field equipment.