In order to predict surface roughness of the workpiece accurately, an approach for surface roughness prediction in machining based on multi-sensor fusion considering energy consumption is proposed. The power signal and the vibration signal of the turning process were collected firstly. Surface roughness was measured. The time domain and frequency domain characteristics of the vibration signal were extracted by using the EEMD and wavelet packet analysis. And the vibration features with the time domain feature of power signal, energy features and cutting parameters constructed the joint multi-eigenvectors. Then the KPCA was used to fuse the joint multi-eigenvectors to generate the fusion feature. Finally, the fusion feature was taken as the input characteristic of the SVM-based surface roughness prediction model. In addition，GA was used to optimize the relevant kernel parameters of SVM model to improve the prediction accuracy. For the prediction results obtained by multi-sensor fusion, the mean relative error was 4.91%, the maximum error was 0.111 μm and the prediction time was 9.24 seconds. The experimental results showed the proposed method had the highest prediction accuracy compared with the joint feature and single sensor. A comparative study with multi-sensor joint feature prediction method and signal-sensor feature prediction method shows the information collected by multi-sensor is more comprehensive and accurate, which ensures prediction accuracy and the prediction accuracy can be further improved by fusing the features.