A Novel Framework for Real-Time Condition Monitoring Using Signal Streams:

1D Sliding Window Residual Network (ResNet)

The existing AI-based condition monitoring techniques of dynamic systems have been known to be fraught with a set of recurring issues. Some of these problems are as follows:

​

1. Real-time monitoring requirements are usually not closely satisfied.

2. The interferences of Fault-Tolerant controllers are not usually accounted for.

3. Data quality and quantity limit the performance of ML-based monitoring techniques.​

 

This project presents a novel real-time time-series classifier, the 1D Sliding Window Residual Network (ResNet), to address these issues.

​

The idea is to first adopt a sliding window strategy for time-series data types. Each time-step of a dynamic system's output signals is associated with the system's operating mode. For example, if a dynamic system, in this case, an in-orbit satellite is operating nominally, its outputs should reflect that and the same goes for other operating conditions such as different fault modes. Every window that results from running the sliding window through the signals is labeled with the dynamic system's operating mode. By creating a comprehensive dataset of these windows, a ResNet model could be trained that is capable of achieving near real-time condition monitoring and fault classification. In this way, real-time condition monitoring is achieved and this process is illustrated by Animation 1.

​

Issues #2 and #3 are addressed using a Markov-randomized satellite model that exhibits the effects of random mode transitions and the Fault-Tolerant controller. By curating a comprehensive dataset, the data dependency issues of AI-based methods are also alleviated.