NEC develops status recognition technology through AI-driven time-series data analysis

• NEC develops status recognition technology through AI-driven time-series data analysis, aiming for application in the operational monitoring of social infrastructure

Tokyo, December 12, 2018: NEC Corporation (NEC; TSE: 6701) today announced the development of artificial intelligence (AI)-based "time-series data model free analysis technology" that recognizes the status of a rapid, highly accurate system by analyzing time-series data collected from sensors and allowing for searching. Application of this technology will enable the detection of abnormalities, fault diagnosis, and the prediction of breakdowns for important social infrastructure, including plants, roads, bridges, railroads, and vehicles.

This technology quickly and accurately recognizes the status (normal, abnormal, signs of abnormality) of a system by using the data collected and accumulated from sensors installed in social infrastructure, such as plant facilities.

Specifically, a feature extraction engine is automatically created by using deep learning to understand the data that is collected and accumulated in advance. With this engine, sensor data that is collected as analog data is converted into binary data with a compact data capacity. By learning the characteristic behaviors contained in the data, it becomes possible to find a past status similar to the present status simply by searching for similar binary data. For example, if this technology is applied to the operational monitoring of a system, the detection of an abnormality, fault diagnosis, and the prediction of breakdown based on signs of abnormality become possible.

Most conventional operational monitoring is carried out using a method where data obtained from a target is modeled using a mathematical model, at that point, it is then judged if a system is operating according to the model. However, this method takes time and effort for the construction, verification, and assessment (tuning) of a model.

This technology represents a new method, where a temporal change in data and the relationship between data is extracted as features without modeling data obtained from a target (model free). The data is converted into binary data, and the status of the target is assessed by making a comparison to the past binary data. In operational monitoring using time-series data, this AI technique finds the characteristics of data in the same manner as an observer does empirically.

Further, this technology is available from an early stage where there is only a small amount of data. It is possible to improve accuracy through operation. This method does not require modeling, allowing for immediate introduction.

NEC aims to implement this technology in a thermal power plant in FY2019, and to expand its application to other social infrastructure through extensive demonstrations and verification.

NEC positions the safety business as a global growth engine in its 3-year mid-term management plan up to FY2020, called the "Mid-term Management Plan 2020." This technology accelerates and enhances the development of solutions and services for the achievement of "NEC Safer Cities" and contributes to safe, secure urban development.

Larger view Flow chart of operational monitoring using this technology

Flow chart of operational monitoring using this technology

• Learning phase: Create a feature extraction component using collected and accumulated data. Time-series data collected and accumulated from a variety of sensors installed in an operation target is divided as many partial time series segments for a given length of time. The characteristic behaviors contained in each divided segment are learned repeatedly through deep learning. A feature extraction engine that converts time-series data into compact binary data is created automatically.
• Accumulation phase: Automatically extract the features of accumulated data and create a database. With the feature extraction component, all segments that have been created by dividing collected and accumulated data are converted into binary format and stored in a database. This compresses the data and allows for high-speed searching.
• Monitoring phase: Ascertain the status of the operation target by searching the feature database. With the feature extraction engine, the monitored time-series data are converted into binary data. Using this as a search key, similar binary data are searched for in the feature database, and the status of the operation target is assessed. This allows for the detection of abnormalities that were difficult to detect so far, fault diagnosis, and the prediction of breakdowns.

Features of this technology

• Automatically create a feature extraction engine that allows the features of data to be extracted. For feature extraction (conversion) suitable for searching, a feature extraction engine using this technology focuses on the following two points: (1) Temporal change in data and (2) Relationship between sensors. An engine has two feature engines in it that extract each feature. By synthesizing these results, features are ultimately converted as binary data. The inner engines are learned simultaneously through deep learning. Efficient learning is implemented according to NEC's unique learning index.
• The diagnosis and prediction of an abnormality are possible, as well as the detection of an abnormality. In addition to the detection of an abnormality, fault diagnosis becomes possible based on past experience by making a comparison with the features of data for a variety of faults that occurred in the past. This technology enables a clue that is noticed intuitively and empirically by a professional to be found from the information on past faults with similar features. Even if a fault occurs, this allows the recovery period to be shortened by implementing appropriate recovery action.

Further, by making a comparison with the features of data that appear in common before the occurrence of a fault, it is detected that the status before the occurrence of a fault and the current status have become the same, allowing the detection and prediction in advance of the possibility that a fault will occur after a specific time. This allows preventive action to be taken in advance, such as part replacement and system switching.

This technology was released at an international meeting on data mining in August 2018: 24th SIGKDD Conference on Knowledge Discovery and Data Mining (Host: London). (Note)

• (Note) Dongjin Song, Xia Ning, Wei Cheng, Haifeng Chen, Dacheng Tao. "Deep r-th Root of Rank Supervised Joint Binary Embedding for Multivariate Time Series Retrieval" SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) Released at the following URL. https://www.kdd.org/kdd2018/