Fujitsu: Technology of deep training for time series analysis of data with high accuracy

In March, 2016 Fujitsu Laboratories Ltd. ^[1] capable with a high accuracy to analyze time series of data announced development of technology deep. In perspective applications for Internet of Things time series of data can vary considerably therefore detection of patterns of their change appears for the person very difficult task.

Machine learning is the central artificial intelligence technology. In recent years in this area all attention was riveted on technology of deep training as to a method of auto eject of the characteristic values necessary for interpretation and assessment of the phenomena. Huge volumes of time series of data gather from devices, especially during an era of Internet of Things. Applying deep training to these data and classifying them with a fine precision, it is possible to carry out the further analysis with perspective of creation of new products and solutions and opening of the new directions of business.

The technology of deep training which is perceived as break in development of artificial intelligence provides the highest accuracy of image understanding and the speech, however it is still applicable only to limited data types. In particular, still it was difficult to classify precisely in the automatic mode changeable time series of the data arriving from the devices connected to Internet of Things.

Fujitsu developed technology of deep training at a basis of the chaos theory and topology for automatic accurate classification of changeable time series of data. This technology allows to process precisely even complex temporary data with the big amplitude of changes.

The technology uses the following procedures for training and classification:

1. Graphical representation of time series of data with use of the chaos theory

The numerical data arriving from sensors are represented using multidimensional surfaces as the work of a complex combination of dynamic movements. Directly the research of mechanisms of dynamic movements represents a difficult task, however creation of the diagram of dependence of these changes from time, allows to reveal characteristic trajectories for each mechanism of movements. Use of such graphic approach allows to carry out differences between time series given using schemes.

2. The quantitative description of charts using topology

As it is difficult to apply directly machine learning to the schemes created on a step 1, the Fujitsu company used the topological analysis ^[2] to express characteristics of charts as numerical values. In this method instead of functions which are usually connected with graphics images the analysis of quantity of openings in the scheme and the main characteristics of a form is carried out, and then data will be transformed to vector representation of properties.

3. Training and classification using convolution neural networks

The Fujitsu company processed the concept of convolution neural networks which study at the vector representations received on a step 2 and provide a possibility of classification of changeable time series of data.

In reference tests which were carried out to repositories ^[3] on classification of time series of the data collected from gyroscopes in devices of wearable electronics, the new technology showed accuracy about 85% that it is nearly 25% better in comparison with already available technologies. In tests by determination of a mental status of the person using a time series of data on brain impulses this method reached accuracy about 77% that it is about 20% better, than at the existing methods.

The technology developed by Fujitsu expands data types to which it is possible to apply deep training. Moreover, as it allows to classify very precisely time series of data with significant changes, opportunities for new analysis types open. For example, using the devices connected to Internet of Things it will be possible to reveal precisely anomalies in behavior of the equipment, to predict accidents at the plants, it is also possible to use technology in the analysis of the major signs in medical diagnostics and in the course of treatment. Similar use of technology, as expected, will allow to achieve considerable achievements in the different areas connected with artificial intelligence.