Supporting Viewpoints to Review the Lack of Requirements in Space Systems with Machine Learning

Identifying the insufficient requirements, such as missing or lacking requirements, is important to prevent serious accidents for CPSs (Cyber-Physical Systems), such as launch vehicles and spacecrafts which often required to be self-adaptive. In JAXA (Japan Aerospace Exploration Agency), several review boards in place to verify the requirements toward space systems from various viewpoints which often derived from the reviewer's experience about anomalies. However, the impossibility of assigning a well-experienced reviewer to all review opportunities highlights the importance of sharing their viewpoints. In this paper, we aimed to extract and exploit associations between development documents and archived anomaly reports about space systems, thereby identifying the lacking requirement in the review. An association between these two documents can be treated as a viewpoint of the review, which contributes to preventing previously experienced anomalies. To cope with this problem, we propose a CNN (Convolutional Neural Network) model that predicts the correlation of documents. A collection of judgments about meaningfully related pairs of text was prepared to train the proposed model to contribute to detect the lack of requirements. Experimental results have shown that the performance of the proposed method is significantly better than that of baseline methods (i.e., 71.0% in F-measure). Also, further investigation has shown that not only the word similarity, but different attributes are necessary to solve our problem.