Fujitsu Laboratories Ltd. and Kumamoto University have jointly developed a technology to easily create the train data necessary to apply AI to time-series data, such as those from accelerometers and gyroscopic sensors.
Time-series data obtained from sensors does not include anything other than every-changing numerical data. Therefore, in order to create training data for use in machine learning, it was necessary to manually attach finely detailed labels to the data in accordance to its changing values, indicating what was done and when at each point where the numerical values changed. For example, data from accelerometers when a person goes running includes intermingled data from when a person is running, when they are walking, and when they are standing still. So in order to create training data for AI, the data needs to be separated into segments, and labeled as "Running" "Walking" and "Stopped."
Conventionally, to create this sort of training data, the typical process was to record a video of the behavior while measuring the time-series data, identify the type of behavior seen with the changes in the numerical values at a second-by-second level, and manually attaching the labels. Because this process required a significant amount of work and time, the application of AI to time-series data saw limited progress, and there was a demand for a technology to automate the labeling process and reduce the workload.
Fujitsu Laboratories and Kumamoto University claim they have enabled the automatic creation of highly accurate training data with appropriate labels for each action, just by manually attaching a single label to each longer time period, even if they include multiple actions, indicating the major action in that time period according to human judgement. Because this significantly reduces the number of man-hours required, this technology is expected to accelerate the use of AI with time-series data.
The development could enable easier installation of services such as fall detection, operational functionality checks, and abnormality detection for machines, in smartphones and various other devices.
Fujitsu Laboratories and Kumamoto University aim to conduct field trials using time-series data from a variety of fields, with the goal of commercializing this technology as a preprocessing technology for time-series data as part of Fujitsu Human Centric AI Zinrai, Fujitsu Limited's AI technology, during fiscal 2019.
How it works
Looking at time-series data, the new technology can learn characteristics of times when the same activity is ongoing and characteristics of times when the activity changes, and can then automatically extract appropriate time periods from time-series data with actions based on same characteristics.
Users attach a single broad label for long segments of data (for example, one hour), such as "running" if the majority of the segment is spent running. After a deep neural network is trained to predict such labels and the resulting estimated labels can be used to calculate the segment of the time-series data that most contributed to that prediction. Also by adding up the time periods that have a high degree of contribution as label candidates, this system can create training data capable of accurate prediction.
Fujitsu Laboratories and Kumamoto University conducted a trial where they attached labels to time-series data from accelerometers while performing mock work processes in a factory such as polishing. They confirmed that this technology was able to correctly label 92% of time periods. They judged that this was equivalent to the highly accurate results obtained when using data that was manually labelled in detail as training data.