Advantages
- Making it easier to train subsequent machine learning/AI algorithms in tasks involving prediction or classification
- Providing a reliable and robust representation for the time-series data as noise and other outliers are removed during the weighted averaging operation.
- Providing a scalable solution where one can analyze time-series data with any length in duration including any number of variables.
- Improving memory allocation and computational speeds to due to compressing potentially very high-volume data into a rich yet compact representative version.
Summary
Analysing multivariate and heterogeneous time-series data is highly challenging, particularly when the data is collected from multiple sensors operating under varying conditions and time durations. In applications such as cold-chain food transportation, sensors capture variables like temperature, humidity, and location across different shipments, producing datasets that differ in length, contain inconsistencies, and are often noisy. These challenges are further intensified by issues such as sensor malfunctions, battery failures, and uneven sampling rates. As a result, extracting meaningful insights and applying machine learning techniques becomes difficult. Conventional approaches are not well-suited to handle such complex and irregular data, leading to inefficient analysis, potential loss of critical information, and lower accuracy in prediction and classification tasks.
Our technology offers a new and practical way to simplify and understand complex time-series data collected from multiple sensors. Instead of analysing each sensor separately, it combines all the data into a single representative signal that captures the overall behaviour. It uses techniques like correlation, Dynamic Time Warping (DTW), and weighted averaging to align and merge data even when the signals have different lengths or quality. This approach helps reduce noise and inconsistencies while keeping the important information intact. As a result, the data becomes easier to visualize, analyse, and use in machine learning models for tasks like prediction and classification. It also makes the process more scalable and efficient, allowing it to be applied in real-world areas such as food supply chains, healthcare, and environmental monitoring. Overall, this technology makes it much easier to turn complex sensor data into useful and reliable insights.
Cold-chain time-series Based Machine Learning Pipeline. The images show the real environment for data collection in this work.
Desired Partnerships
- License
- Sponsored Research
- Co-Development