Abstract: Objective Quantitative characterization of complexity of welding electrical signals is an important method to monitor and evaluate welding quality, approximate entropy and sample entropy can only describe complexity of welding electrical signals in a single scale, but cannot represent complexity in multiscale. Methods In this study, refined composite multiscale entropy (RCMSE) was introduced to describe complexity of electrical signals in multiple scales. Feature selection and support vector machine pattern recognition model was combined, then a weld width classification method based on composite multiscale entropy and support vector machine was proposed. Firstly, current signal was normalized and the multiscale complexity features were extracted by RCMSE. Then, separability of entropy value of each scale to weld width was evaluated by feature selection method, that was sequential forward selection (SFS), the highest correlation entropy value was obtained to form feature vector. Finally, feature vector was input into weld width multi-classifier of support vector machine (SVM) for training and recognition. Results RCMSE was applied to complexity feature extraction of arc current signal, sequential forward feature selection was use to select RCMSE feature values, and a method of weld width classification was proposed based on SVM. Conclusion RCMSE can effectively extract multiscale complexity feature of electrical signals and obtain a high recognition rate of weld width classification, which provided a method to measure multiscale complexity of signal in welding process.