OELAssist - Enhancing the Open Engineering Lab (OEL) Student Experience

Project leader(s): Dhouha Kbaier Andrew Mason

Theme: Technologies for STEM learning
Project faculties: STEM
Status: Archived
Date: November 2023 to November 2025

This eSTEeM project investigated how Generative Artificial Intelligence (GenAI) can enhance remote STEM learning by providing students with intelligent, real-time support during online laboratory activities. The study centred on the development and pilot testing of OELAssist, an AI-powered system that detects signs of student difficulty and delivers adaptive, personalised feedback using a combination of anomaly detection and large language models (LLMs).

The project addressed a significant pedagogical challenge: while the Open STEM Labs (OSL) enable students to control real laboratory experiments remotely, they lack immediate, responsive tutor feedback. Students often face uncertainty when performing experimental procedures, leading to frustration and reduced engagement. OELAssist was developed to provide a scalable, automated solution that can identify behavioural anomalies in student interaction data and deliver targeted feedback through natural language generation.

The pilot study focused on the Pressure Vessel Experiment from module T272 Core Engineering B. The study collected data from 2 module presentations in 2024 and 2025. In 2024 over 400 students interacted with the online laboratory between April and July 2024, generating more than 800,000 data points that recorded their motor control actions, power settings, and timing behaviour. The 2025 data collection period consisted of over 300 students contributing a further 600, 000 data points to the full data set used in this work.

These interactions were analysed using machine learning models, including Isolation Forests for unsupervised anomaly detection and Random Forests for supervised classification.

Anomalies such as prolonged motor use or exceeding recommended pressure thresholds were detected and passed through a Retrieval-Augmented Generation (RAG) pipeline, which contextualised the anomalies and prompted an LLM to generate immediate, situation-specific feedback. Example outputs included suggestions such as,

“You need to reduce the power of the motor to avoid increasing the pressure beyond 5MN and to decrease the amount of time the motor is running, as there's a risk of damaging the equipment. In the 'Pressure gauge video' panel, the dial measures pressure in the range 0–10 MN m−2 and has a red line at 7 MN m−2, which indicates the maximum allowable pressure for the equipment. You should not exceed this value, and the pressure is indicated by the position of the pointer on the dial.” or

“You might be being overcautious and it is okay to increase the power of the motor, to decrease the amount of time the motor is running in the reverse direction and enable you to complete the activity within the allotted time. Please be careful that you do not go below 0 MN m-2 pressure.”

Preliminary results confirmed that the system can accurately detect anomalous behaviours and generate pedagogically meaningful feedback. These findings highlight the potential of AI-driven adaptive systems to improve student experience, support retention, and promote equitable outcomes by offering timely, data-informed guidance.

The implications extend beyond the pilot: the integration of anomaly detection and generative AI presents a scalable framework for adaptive online learning. The next phase will focus on embedding OELAssist into live module delivery, evaluating student perceptions of AI-generated feedback, and exploring cross-module deployment across other Open Engineering Lab (OEL) experiments.