The accuracy of fault diagnosis for engineering vehicles will decrease if the fault model is established by using traditional methods when the working conditions change. For the variable working conditions, it is not practical to build one fault model for each working condition. Therefore, a fault diagnosis method for engineering vehicles based on a multi-source transfer neural network algorithm is proposed. The proposed method is based on the convolutional neural network model and adds the framework of multi-source transfer learning, the output of the fully connected layer of the neural network is used as the input of multi-source transfer learning to calculate the inter-domain distance under multi-conditions, which is then added to the total loss to participate in the calculation of the network parameters. To validate the feasibility of the proposed method, two datasets were used to conduct the study. The first type of dataset is the bearing vibration data under three types of working conditions, with 8800 samples for each type, and the second type of dataset is the bearing vibration data under four types of working conditions, with 1683 samples for each type. The experimental results of classification accuracy show that compared with CNN and SVM methods, in the first type of data analysis, the average accuracy of the proposed method is improved by 1.95% and 48.57%, respectively; in the second type of data analysis, the average accuracy of the proposed method is improved by 3.17% and 26.39%, respectively. The proposed method requires only training data under limited working conditions to establish a fault analysis model, which can test the fault types under other unknown working conditions. The proposed method provides a feasible and effective method for fault diagnosis of engineering vehicles in extreme environments with variable working conditions.