The invention provides an industrial process fault detection method based on data neighborhood characteristics and non-neighborhood characteristics maintenance. The main technical problem to be solved through the invention is how to comprehensively consider the neighborhood characteristics and non-neighborhood characteristics of process data such as distance, time and angle in the process of establishing the fault detection model. The method comprises the following steps: firstly, identifying the data samples for each sampling data point that correspond to and are close to the sampling points in terms of distance, time, and angle to compose a neighborhood set; using those non-neighboring samples to compose a non-neighborhood set; then, using a generalized eigenvalue problem to solve the projection transformation vector; and based on this, establishing a corresponding fault detection model; and finally, using the model to conduct online fault detection. Compared with the conventional method, the method can establish a fault detection model to achieve more reliable and accurate testing results.