Under variable operation conditions, the fault diagnosis of rolling-element bearings encounters the problems of ambiguous characteristic frequencies and inconsistent disturbance features. Conventional diagnosis methods have adopted extra hardware or signal preprocessing to overcome these problems, which usually resulted in difficulties in implementation and incorrect diagnosis. Here, we propose a novel approach based on the strategy of extracting features that are insensitive to changes in operation conditions. This may have the advantages of preventing interference from signal resampling and preprocessing. First, we derive the bispectral expressions of vibration signals for rolling-element bearings in use. Next, we use simulated and measured experimental data detected by numerous fluid and vibration sensors to identify the proposed model. Finally, experiments on the fault diagnosis of bearings are performed to validate our proposed approach. Results show that the proposed bispectral distribution method has the advantage of insensitivity to the operation conditions. In experiments involving three levels of fault severity, our proposed diagnosis model always correctly identified the fault type of rolling-element bearings under different operating conditions. With the advantages of directly extracting the insensitive bispectral features without the requirement of additional hardware and signal preprocessing, our proposed approach is simple and easy to implement, giving it good application prospects in engineering practice.