The invention provides an industrial process fault detection method based on data neighborhood feature preservation and aims at solving the main technical problem of comprehensively considering the distance, time and angle neighborhood features of process data in a process of establishing a fault detection model. The method comprises the following steps: defining a corresponding neighborhood set for each sampling data point, wherein the neighborhood set comprises data samples similar to the data point in distance, time and angle; then, constructing a feature value problem to solve a projection transformation vector, and establishing a corresponding fault detection model on the basis; and finally, implementing online fault detection by use of the model. Compared with the traditional method, the fault detection model established by the method minimizes the risk of information loss, and a more reliable and accurate result can be obtained.