Digitising a large multi-source electrical diagram (ED) dataset is challenging due to the need to anonymise Regions of Interest (ROIs) that contain sensitive data such as personnel identification. Moreover, due to the difficulty of diagram anonymisation, there are limited openly available examples for deep learning model training. Reliance on manual annotation limits scalability and increases cost, resulting in insufficient data for generalisation. Even with data augmentation methods, the diversity of real-world styles would limit the application to limited use cases. This paper explores a method that combines text mining and clustering with feature extraction to accurately segment multiple ROIs and applies to industry-wide engineering drawings, but with a focus on EDs. Optical Character Recognition (OCR) text positional and contextual data serves as input for a multi-level density-based spatial clustering algorithm with multi-stage filtering which is applied according to image complexity. Simple layouts with ROIs in one area are identified with Level 1 clustering and filtering. Complex multi-location clusters are split into classes before Level 2 clustering within a smaller area. The proposed method achieved F1 scores of 94% and 75% for simple and complex layouts respectively on a real-world ED dataset, evaluated at a minimum Intersection over Union (IOU) threshold of 90%. Detection of tabular-structured title block ROIs was evaluated across three distinct datasets: the ED dataset, a synthetic raster dataset, and complex real-world piping and instrumentation diagrams (P&IDs), respectively, achieving F1 scores of 98%, 92%, and 42%, at an IOU threshold of 50%. Data and code available at: https://github.com/mayoT0/anonED .