The gearbox of tunnel boring machines(TBMs) is subjected to continuous strong torque during operation, resulting in a high wear failure rate. Therefore, to accurately identify the wear state of the gearbox and avoid further faults, a convolutional neural network(CNN)based TBM gearbox wear state recognition method is proposed. First, based on the wear state identification mechanism of the gearbox, the concentration values of nine elements in the spectral analysis results are selected as indicators for identifying the wear state of the gearbox. Next, an improved CNN model is established. Two convolutional layers are stacked to replace the depth of one convolutional layer in the traditional model to extract data features, and two convolutional kernels with different sizes are used to extract features at different levels. Finally, the model is applied to a practical engineering project, and the collected TBM gearbox oil spectral data and wear state labels are put into the model for training. The trained model is used to identify the wear state of the test set data. By achieving 95% accuracy rate, the research results suggest that the proposed method can accurately identify the wear state of TBM gearboxes, providing decisionmaking support for maintenance in actual construction.