Representation and Retrieving Situation Method in Situational Knowledge Base of CBR-System Based on Neural Network and Contrastive Learning

The case-based reasoning (CBR) method has proved to be one of the most effective in intelligent decision support systems for the operation of complex objects. CBR assumes the presence of a situational knowledge base (SKB), which stores descriptions ever existed or potentially possible problematic situations and those solutions that are recommended in these situations. When operating complex technological facilities, in order to identify situations, it is necessary to take into account large amounts of data of various types that come from various sources. This makes it difficult to use metrics to assess the similarity of situations. The article proposes to use the approach used in large language and multimodal neural network models to compare and select from SKB situations. A neural network that converts input vectors into embeddings is trained on examples of similar and dissimilar situations in such a way as to best separate the embeddings of dissimilar situations and converge similar situations. This approach can be used to store situations in vector databases (VDB) with built–in search mechanisms. This paper presents the results of a study of the architecture of a neural network, where the input module performs the task of forming embeddings, and the output module performs the task of comparing them with by calculating the similarity score for the final selection. It allows the use of a neural network both independently and in conjunction with a vector database, when the task of preliminary localization of a subset of relevant situations is assigned to the VDB tools.