Due to the complex structures of modern industrial processes, there are ubiquitous substantial couplings among variables, and thus faults can spread widely among them, which brings challenges to process maintenance. To address this issue, industrial root cause diagnosis (IRCD) technology has been developed, which determines the root cause of the fault from abnormal variables to facilitate targeted fault processing. IRCD consists of two main steps: structure inference and root cause identification. Structure inference establishes the information transmission structure among variables, and the root cause identification locates the root cause according to the structure. However, most of the existing IRCD reviews focus on structure inference without investigating the root cause identification step. In addition, they have not established systematic connections between various IRCD models and different process characteristics. To this end, a survey of IRCD studies is conducted in two stages: structure inference and root cause identification. First, four types of classical structure inference models are summarized according to their inference criteria. Then, considering the high-dimensional, nonlinear, and nonstationary nature of the process as well as the utility of mechanism knowledge, the variants and application scenarios of the structure inference models are sorted out. Afterward, the root cause identification methods are classified, including pure data-driven as well as hybrid knowledge-and-data-driven paradigms, covering six typical methods. The advantages and drawbacks of these methods are also analyzed. Finally, the unsolved challenges in IRCD are discussed, and future directions are put forward to provide references for follow-up research.