Japanese multinational IT equipment and services company Fujitsu has collaborated with Inria, the French national research institute for digital science and technology, to develop a new technology that can automatically create AI models that detect anomalies by extracting the necessary information from time-series data.
The new technology has been incorporated into GUDHI — an open source Topological Data Analysis (TDA) library developed by Inria — and is currently available for free for users globally. According to a Fujitsu press release, this will not only promote the use of AI in companies, research institutions, and other organizations, but also enable the creation of AI models for a variety of use cases. Fujitsu says the feedback from these organizations will be reflected in ongoing tech improvements.
The technology provides an easy solution for software engineers that create AI categorization and anomaly detection models for time-series data, and can reduce required human labour to just one percent that of previous methods. This may ultimately allow even engineers with no specialized training to create anomaly detection models, and eventually to accelerate the deployment of new AI models in a variety of business fields.
Time-series data, including for example sensor data from IoT devices or biological data, consists of a wide range of information with complicated interconnections. Often subjected to severe volatility, it’s challenging to discern when meaningful patterns or anomalies occur in time-series data.
With the continued progress of AI technology in recent years there has been greater deployment of AI in many business fields. Despite demand for greater levels of automation, the most common means of creating AI models still involve manual work by specialized AI engineers. Moreover, the process of building new AI models continues to rely on trial and error, which requires a good deal of human labour that may lead to delays in field deployment.
In the case of time-series data such as heart rates and brain waves, it is often necessary to extract a range of data features across a range of different time windows. With the new technology, it only takes 20 minutes to create an AI model for detecting anomalous states, such as drowsiness, in human pulse data. The quick-build Fujitsu-Inria model had only one-tenth the average error rate of an AI model created over four days by AI specialist engineers using standard methods.
In another use case, it took just 10 minutes to generate an AI model for detecting internal damage in bridges. In an evaluation using the equivalent of 30 years worth of vibration data collected from accelerometers attached to a mock bridge deck plate, it took AI engineers over five days to build a model with the same detection performance.
The Fujitsu-Inria technology is based on a deep learning algorithm which is able to map the features as points on a chart with the axes defined by the length of the time period and the features of the behaviour of the waveform over that period.
The mapped charts are then compared, and the space is divided into regions where ordinary data points co-occur, regions where they do not, and regions where there is no overlapping data. The number of regions and the method of segmentation are then optimized, and the regions are extracted in the order of their degree of similarity.
For time-series test data, in order to determine if the data is anomalous or not, the features extracted from the input data using the TDA technology are mapped onto a chart, and the number of points which fall into the regions delineated above is counted. Multiplying the number of points that fall within each region with the degree of similarity for that region, and adding together all the regions outputs a degree of deviation value which can be used to determine the deviation of the input data from the standard.
Fujitsu and Inria will be presenting this technology at the 23rd International Conference on Artificial Intelligence and Statistics (AISTATS 2020) in June.
Journalist: Yuan Yuan | Editor: Michael Sarazen