For a characteristic that the edge distribution of faults under different working conditions is the same, but the condition distribution of each kind of fault samples is changed in scale and position,the invention provides a working condition self-adaptive deep neural network model and a variable working condition fault diagnosis method. The working condition self-adaptive deep neural network model is composed of five parts, including a source domain feature extraction module, a fault classifier, a target domain feature extraction module, a position scale conversion module and a domain difference regularization module. After the fault samples in a source domain are processed by the source domain feature extraction module and the position scale conversion module, the condition distributionof the fault samples is similar to the condition distribution of similar fault samples in a target domain. The difference of sensing data distribution under variable working conditions is overcome; amethod capable of eliminating the influence of the working conditions and acquiring information which only reflects the fault or performance degradation of a rolling bearing is provided, so that thefault diagnosis of the rolling bearing is more accurate; and the method has extremely high popularization values.