Evaluating the innovation capability of university research teams is critical for guiding policy and resource allocation in higher education. Traditional input–output metrics fail to capture the dynamic, multidimensional nature of scientific innovation processes in modern colleges. In this study, we propose an IoT-empowered, attention-enhanced LSTM framework that integrates real-time sensor data from smart laboratories and campus innovation centers to continuously monitor key research activities. We first construct a capability indicator system combining inputs (equipment usage, laboratory environmental parameters, researcher activity) and outputs (experiment throughput, publication metrics), all captured via 5Genabled IoT devices. An attention mechanism dynamically weights each indicator at every time step, allowing the LSTM to focus on the most informative features as innovation unfolds. To validate our approach, we conduct experiments on three datasets collected from university innovation labs over six months: climate-controlled bioengineering chambers (SMLCampus), soil-monitoring for agrotech projects, and power usage in maker-spaces. Compared with baseline MLP, vanilla LSTM, and BiLSTM models, our method achieves superior prediction accuracy of research output trends (e.g., on SMLCampus: RMSE = 0.024, MAE = 0.019, R² = 0.9999) and consistently higher anomaly detection precision in identifying workflow bottlenecks.
