As a field of research arc fault detection in domestic appliances has existed for a long time and many detection algorithms have been published, patterned or implemented on commercial products. None of them, however, guarantees perfect discrimination and all are susceptible to false negatives or false positives (i.e. indicating the absence of arcing fault, when in reality it is present, or recognizing normal functioning as an arcing condition). This phenomenon can be explained by the fact that all methods have been based on some features of arc fault which can be shared with load and network conditions such as noisy loads, the plugging-in or unplugging of appliances, the change of functioning mode of an appliance on a network and so on. A solution for limiting this phenomenon is multi arc-fault feature recognition. This research presents a method for finding and combining arc fault features in order to obtain better performance than using a single arc fault feature. The choice of arc-fault features and the algorithm for combining them are based on machine learning techniques. The method proposed here can be used for different network conditions and loads. The effectiveness of this method has been verified by a number of experimental tests including not only the requirements of standard arc fault detection, but also the most difficult situations such as multiple loads masking and transient loads.