Arc additive manufacturing has the advantages of high deposition efficiency, low cost, and unrestricted deposition shape and size. However, the dimensional accuracy of the formed parts by arc additive manufacturing is still difficult to guarantee accurately. The size of the deposited layer is one of the criteria for evaluating the quality of component formation, and it is crucial for judging processing quality and defect compensation. Therefore, real-time monitoring of the variation of the deposited layer size during the arc additive manufacturing process is of great significance for optimizing process parameters and ensuring the formation quality of additive manufacturing components. Arc spectral information can reflect the arc state, closely related to the forming quality. Therefore, studying the relationship between arc spectrum and deposited layer size is very important. This study used titanium alloy (TC4) material as the substrate and welding wire, and the arc plasma spectral signals were studied to investigate the correlation between GTAW additive arc spectral characteristics and deposited layer size. Firstly, a spectral acquisition system was constructed to collect arc spectral signals at different positions above the molten pool, around the molten pool, and below the tungsten electrode. Secondly, based on the principle of high spectral line separation, the wavelengths of 404.20 nm TiⅠ spectral line, 416.36 nm TiⅡ spectral line, 420.20, 434.81, 480.50, and 487.98 nm ArⅡ spectral lines, and 696.54 and 794.82 nm ArⅠ spectral lines were selected. The peak intensity features of these spectral lines were extracted, and the ReliefF algorithm was used to explore the correlations between different spectral line intensity features and deposited layer size. The results showed that among all the spectral lines, the peak intensity features of the 404.03 nm TiⅠ element spectral line, 416.36 nm TiⅡ element spectral line, and 794.82 nm ArⅠ element spectral line above the molten pool had a strong correlation with the deposited layer size. At the same time, combining the ReliefF algorithm, the correlations between peak intensity features of different positions and deposited layer size were studied. The results showed that the spectral line with the strongest correlation between the molten pool above and the deposited layer size was the 696.54 nm ArⅠ spectral line, and the spectral line with the strongest correlation between the molten pool around and below the tungsten electrode and the deposited layer size was the 794.82 nm ArⅠ spectral line. Furthermore, to reduce random errors, the PCA algorithm was used to fuse the intensity features corresponding to the three spectral lines with the highest correlation with the deposited layer size, and a new fusion feature was obtained. Then, combining the K-nearest neighbors algorithm, a deposited layer size prediction model was established. The fused feature and the feature values of samples with the highest correlation with the deposited layer size at different positions were extracted, and the accuracy of predicting the samples category based on these four features was calculated. The results showed higher accuracy for predicting the deposited layer size based on the fused feature. Finally, based on this new feature, combined with the threshold segmentation method, dynamic monitoring of the variation of the deposited layer size was achieved.