Jon Amdur, Vice President, Senior Technical Manager, Kleinfelder
Unmanned Aerial Systems (UAS), as well as other remote sensing techniques, are designed to obtain more accurate and cost-efficient data with greater safety. However, a reoccurring challenge has been the ability to embrace the tools to process and analyze the data with efficiencies commensurate with the data collection process. To fully reap the benefits of the data collection tools, the consulting Engineering Industry needs to make greater progress on data management and analytics. Especially, understanding the range of benefits from deploying Artificial Intelligence (AI), for data analytics, predictive modeling, to the potential pitfalls of the AI “learning” process.
Recent improvements made by Unmanned Aerial Vehicle (UAV) developers, including endurance, lifting capabilities, and ease of use, have resulted in UAV systems that are more available and ubiquitous. Concurrently, sensors including various types of cameras, non-visual sensors, and LiDAR, as well as global positioning systems (GPS), have made data more precise and capable of measuring a wider array of attributes in the field. Improvements continue to drive the utility of UAS by giving engineers and scientists larger and richer data sets in which to analyze and draw conclusions. Despite the continued improvement and access to programs capable of taking raw data and converting it into useable information, such as converting point cloud data into Digital Terrain Models (DTM) or Digital Surface Models (DSM), there are still gaps that need to be filled. For example, the Engineering Industry has been slow in adapting AI tools, including Machine Vision and Machine Learning software to accelerate and improve analysis of the vast amount of routinely collected data.
Let’s consider scenarios where the application of AI could accelerate and improve railroad track, pipeline and transmission line inspections. These inspections collect large amounts of data over thousands of miles, including hundreds of thousands of photographs, huge point cloud data files, and multitudes of sensor readings or telemetry (thermal, audio, and accelerometer data, among others).
Using the rich data analytical tools, we have an opportunity to fully realize the potential of remote sensing tools and AI. The sooner we begin to lead the training and integration of these systems, the sooner we can capitalize on the benefits of the innovative technologies
The terabytes of data that are generated even on the smallest linear project need to be reviewed for quality assurance and control before being used for engineering analysis. Although there are clear benefits to using remote sensing for data collection including cost, efficiency, and safety, having an inspector review thousands of photos or digital data points does not achieve all the benefits that can be realized using new technologies and could lead to critical data errors as well as inconsistencies associated with different inspector’s ability to synthesize information. A mixture of Machine Vision and Machine Learning would help mitigate the inconsistencies between individual reviewers. With the capability for inspectors to “teach” Machine Vision systems, increased consistency can be achieved and more data processed, resulting in identification of key factors that lead to a more in-depth review of data and improved inspection accuracy. Of course, nothing is 100%, but the use of AI in the analysis would lead to marked improvements in quality, consistency, cost efficiency, and inspection safety, especially of linear facilities that generate large data sets.
During the “learning process,” while training the AI programs, sensitivity can be set to avoid “false negative” results, causing the programs to err on the side of caution and highlighting issues for technicians to review and validate. However, the training inspector can transfer biases to the Machine Vision program, thus perpetuating the bias, and delaying the ability to rely on the programming. In time, the system will “learn” from the data, reducing the initial bias programmed into the system. The sensitivity can be adjusted so progressively fewer “False Positives” are reported. Regarding costs, the short-term development phase versus the cost of the long-term implementation needs to be considered. Over time, the investment in perfecting the training will provide greater benefits in prevention of costly system failures, speed and accuracy of analysis, increased inspector and public safety, and a cost reduction of the program through implementation. The cost-benefit is realized over time and will exceed the initial costs of implementation.
In addition to the data collection and synthesizing processes, the data (and telemetry collected)can be further analyzed using Predictive Analytics to identify key factors and alerts to system failures before they occur. The capability to sort through large data sets using complex multivariate statistical tools and sensor data analysis can determine key telemetry indicating when a system will fail. For example, an inexpensive multi-output sensor could eventually detect data that would indicate a system failure within weeks or days. With this information, scarce maintenance dollars and resources can be maximized, focusing on eliminating costly infrastructure shutdown or situations that could put lives at risk.
Using the rich data analytical tools, we have an opportunity to fully realize the potential of remote sensing tools and AI. The sooner we begin to lead the training and integration of these systems, the sooner we can capitalize on the benefits of the innovative technologies.