In process industries, complexity of operations and continuous exposure to hazardous substances, equipment and environment necessitates performing fault detection and diagnosis (FDD), and risk assessment via accident/failure modelling. Safety reports contain important information regarding health of process systems, hence FDD and risk assessment by text mining of reports is vital to identify potential faults, reveal insights on their behaviour/interactions, and predict hazard escalation into incidents. Here a novel hybrid framework is proposed where accident theory and hazard information are combined with text mining of incident descriptions to perform FDD and accident modelling. Information about hazards and failure events is used to generate syntactic rules of accident causation which are used to extract text chunks from incident descriptions depicting constituent faults/process safety events. Then an ensemble of unsupervised/semi-supervised models is applied on the chunks to identify hazardous elements and develop chain of events and fault trees showcasing failure propagation. The framework is applied on incident investigation reports of an Indian steel plant and 56 chains of events and corresponding 13 fault trees are identified. A comparison of the results generated by the framework with actual inputs provided by the HSE team of the plant showed accuracy of 85.42 %.