From the hums and whirs of factory equipment to skin temperature and heartbeats, sensors track changes in machines and people over time. Like a social network where each person can influence their friends, how one sensor acts and the data it records can impact other sensors at a later point. For example, past fan speed can be linked to future device temperature, potentially helping predict the risk of overheating.
But many artificial intelligence (AI) models for analysing time-series data—measurements recorded over time—are not designed to spot correlations between Different sEnsors at Different Times (DEDT). “If you only look at the same sensor over time or all sensors at a single moment, you miss that delayed cross-effect,” said Min Wu, Head of the Model and Intelligence Chapter at A*STAR’s Institute of Advanced Intelligence and Computing (A*STAR IAIC). “By considering DEDT, you can identify which sensors influence each other, how long the delay is and how a local change later shows up elsewhere.”
Wu and A*STAR IAIC Research Scientist Yucheng Wang sought to capture correlations among sensors using a fully-connected spatial-temporal graph neural network (FC-STGNN), a type of AI that learns from connections between segments of sensor data. They worked with collaborators from Nanyang Technological University, Singapore; National University of Singapore; Singapore University of Technology and Design; and University of Glasgow, UK.
Their approach links every sensor to all others across every time, with the strength of each link depending on the delay between sensor cross-effects. “The model can learn both who is related to whom and how that relationship changes over time,” Wang explained.
However, FC-STGNN still slices time-series data into patches of a single fixed width. Patches can be thought of as time windows or intervals for recording a sensor’s signal. With fixed-size patching, patterns that unfold on varying timescales, from seconds to hours, could fly under the radar.
To solve this, the researchers used a Gaussian model—a bell shape with equidistant curves from the centre that represent the standard deviation (SD)—where a larger SD and broader curve yield a wider patch. Named Gaussian Adaptive Patching (GAP), the approach enables the model to extract the SD from the data and find a suitable patch size per sensor. “A small patch width captures short spikes, like a sudden vibration. A large width captures slow trends, like gradual machine wear,” said Wang.
When tested on 16 real-world datasets covering both prediction and classification tasks, GAP-enhanced FC-STGNN outperformed state-of-the-art approaches and even other GNN-based models, while needing fewer parameters and less training time.
By modelling these cross-sensor links, Wu and the team hope their approach can eventually help improve the reliability of predictive maintenance and health monitoring processes, like assessing the remaining service life of machines or recognising human activities from wearable devices.
The A*STAR-affiliated researchers contributing to this research are from theA*STAR’s Institute of Advanced Intelligence and Computing (A*STAR IAIC).
