In the weaving process, yarn tension signals are adversely affected by a considerable amount of uncertain noise sequences, compromising the closed-loop control accuracy of yarn tension. Particularly challenging is the effective preservation of these features when confronted with sudden changes in yarn tension characteristics. To address this issue, we propose an Adaptive Wavelet Threshold Denoising (WTD) optimization method for yarn tension signals based on Empirical Wavelet Transform (EWT). The application of EWT decomposes yarn tension signals into components of different frequencies and scales, with wavelet thresholding used for threshold processing of the decomposed signals. The effectiveness of the proposed method is validated through simulation experiments and on-site data analysis. Results indicate that, compared to the PSO-VMD method and the FastICA method, the SNR after processing with the proposed method is improved by 8.55% and 26.29%, respectively. Root Mean Square Error (RMSE) shows that the denoising result curve of this method fits the simulated data curve more closely, and the sudden changes in the signal characteristics are better preserved. Experimental data verification demonstrates the superior performance of the proposed method in denoising tension signals with three different characteristics, with the SNR being maximally improved by 5.32dB while fully preserving the sudden changes in the signal. The proposed method exhibits excellent denoising effects in experiments on yarn tension signals collected at different speeds on a circular winding machine, with a maximum SNR improvement of 5.29dB. It adapts well to the changes in signals that occur under different operating conditions. This method provides a feasible solution for improving the stability and production efficiency of yarn tension in knitting systems.