Maintenance Work Orders (MWOs) are unstructured, technical short texts manually recorded by operations and maintenance technicians when identifying the need for maintenance work on equipment or assets. These records hold valuable information about equipment conditions and failures and there is commercial interest in using state-of-the-art information extraction models to process this data. However, accessing real-world data to train models is challenging due to confidentiality concerns. MWO records often contain proprietary information, and cost and failure data that should not be shared. Consequently, only a handful of small-scale datasets are available for training machine learning models. This research addresses the lack of readily available datasets by generating authentic and grounded synthetic MWO sentences. The proposed pipeline uses an existing graph database of annotated MWOs to extract equipment-failure relationships, which are then used by Large Language Models (GPT-4o mini) to produce synthetic MWO records. The challenge is to control the creation to be technically correct by respecting common engineering knowledge about equipment, its parts and their failure types. These synthetic MWOs are further refined to resemble the style of maintenance records generated by technicians by incorporating industry-specific jargon, abbreviations, and grammar. Turing test experiments show that subject matter experts could distinguish real from synthetic data only 57% of the time, demonstrating the potential of this approach. Additionally, evaluations showed that the synthetic data is 86% as natural and 95% as correct as real MWOs. While occasional deviations from the prompt were observed in the LLM-generated outputs, analysis indicated these deviations were not significant enough to compromise the overall quality of the data. This research offers a novel approach to generating grounded and authentic synthetic MWO datasets, which augment training data availability for domain-specific machine learning tasks.