STTR Phase I: Machine Learning-Based Smart Data Compression Solutions for Structural Health Monitoring Sensors

点击翻译按钮获取中文摘要

The broader impact/commercial impact of this Small Business Technology Transfer (STTR) Phase I project is to enable efficient monitoring of civil infrastructures and rapid decision-making on their structural safety. The conditions of aging structures are monitored using structural health monitoring (SHM) sensors. These sensors produce very large datasets. In this project, a data compression solution will be developed to reduce the size of such datasets by 90%, without losing important information. As an example, one sensor can fill up a 128 Gigabyte hard disk in about 6 hours, but with the data compression solutions, it will take at least 60 hours to fill the hard disk. Data compression is thus a critical factor for both storage (disk space) and efficient transmission of sensor data. A microchip with a built-in data compression algorithm will be developed. The sensors with microchips will need to be visited less often for data retrieval and dramatically less bandwidth and power will be required for data transmission over existing wireless networks. This will enable monitoring of structures in remote areas. The data compression will be applicable to various market segments, however the initial target market will be the SHM of aging structures within the oil and gas industry.This Small Business Technology Transfer (STTR) Phase I project aims to develop sensor data compression schemes and encoder/decoder devices utilizing deep learning methods. The proposed system will consist of a data encoder and decoder, which will autonomously learn the characteristics of the sensor data, extract relevant features, and transmit these using low bit rates. Even users without prior experience in machine learning will be able to train the deep neural network with transform domain layers for different sensor types. The software version of the system will allow for data processing and transmission over the Internet when the sensor is connected to a computer, making it possible to handle stored data on-site. The embedded hardware version will be designed for "edge" usage, meaning it will be implemented next to the sensor itself. This approach will ensure computational efficiency, particularly for the feature extraction part of the network, which needs to be executed at the edge. The project's focus will be on detecting pipeline leakage using high-frequency acoustic emission data on the developed microchip system. By reducing the data transmission bitrate of SHM devices, this system will enable continuous transmission of SHM data to the cloud or data centers.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.