<p>In this paper, a signal classification algorithm based on quantile spectrum is introduced. Firstly, in order to control the random fluctuation in the estimation of quantile spectrum, a Welch-type quantile spectrum estimation method is proposed. By dividing the signal into overlapping segments, the quantile periodogram of each segment is estimated by trigonometric quantile regression. The Welch-type quantile spectrum is obtained by averaging the estimations of the same quantile levels among different segments. According to the result of simulation, Welch-type quantile spectrum outperforms quantile periodogram, and the classification performance based on it is also better. Considering different quantile levels of the quantile spectrum as different information sources, an ensemble random forest algorithm is proposed. At each quantile level, the corresponding random forest is trained. The voting weights are determined based on the peak values of different quantile levels in the quantile spectrum, and the classification results are integrated under the weighted voting strategy. Treating Welch-type quantile spectrum as the input and weighted ensemble random forest as classification model, we proposed a signal classification algorithm called WWE-RF. In the situation of different noise disturbance, the classification accuracy of proposed algorithm is higher than the relevant algorithms. We use the proposed algorithm to classify rolling bearing signal data, and achieve high accuracy in the case of multi-classification.</p>