The digital transformation of industry drives a growing demand for intelligent decision support systems capable of efficiently extracting knowledge from regulatory and technical documentation. This paper examines Retrieval-Augmented Generation (RAG) technology applied to a specialized corpus of railway regulatory documents. A comparative analysis is conducted for several methods aimed at improving RAG accuracy, including query processing strategies (HyDE, Step-Back, Multi-Query, Decomposition, Pseudo-Relevance Feedback, Recursive Refinement), two-stage retrieval with a cross-encoder, and fine-tuning of embedding models on synthetic question-regulatory clause pairs. The experiments demonstrate that domain-specific fine-tuning of embedding models yields the most significant improvement in retrieval quality. Cross-encoder reranking also provides a positive effect, particularly for models with initially less accurate ranking. At the same time, advanced query processing strategies do not lead to a substantial improvement in retrieval quality compared to the baseline approach. The constructed document corpus, question sets, and source code are made publicly available.

Currently, ground-penetrating radar (GPR) is often used to solve problems related to searching for and recognizing underground objects. This allows for obtaining rich information about the studied area at various depths, but manual interpretation of radargrams is labor-intensive. This paper proposes the GeoRTr method for automatically detecting underground objects of interest. It is based on an original neural network architecture with a GeoAttention mechanism. The network's input is A-scans from a GPR mounted on a wheeled robot moving in automatic or semi-automatic mode. The proposed neural network model accounts for spatial relationships between A-scans even when the robot's trajectory deviates from a straight line. Using a mobile wheeled robot and a GPR with an antenna mounted at a low altitude above the surface being analyzed, a unique dataset was collected and labeled for recognizing several categories of subsurface objects. Experiments with the developed neural network model on the collected dataset demonstrated a classification F1-score of over 40%, surpassing comparable methods. A comparative analysis of various model architecture variations identified two configurations for different ground-penetrating radar object search conditions.

Autonomous navigation of a mobile robot in human-centered environments requires a map that contains not only a geometric model of the environment for path planning, but also information about environmental objects (doors, furniture, office equipment, etc.). Scene graphs provide such a map representation, where nodes correspond to rooms, locations, and objects, and edges encode spatial connectivity or relationships between objects. Most modern scene graph construction methods have high computational complexity, and the graphs they produce are redundant for the purposes of autonomous robot navigation. This paper proposes a method for constructing Compact Scene Graphs (CSG), which is based on the computationally efficient topological mapping method PRISM-TopoMap and the association of semantic objects with locations on the topological map. The resulting scene graph enables route planning to objects via topological map locations and achieves high-precision localization. The proposed method was experimentally evaluated in the Habitat simulation environment. The experimental results demonstrate that the proposed CSG consumes significantly less memory than traditional metric maps and scene graphs, while providing reliable localization on the topological map through association with semantic objects.

The aim of this study is to compare One-Class Support Vector Machine (One-Class SVM) and DBSCAN for detecting statistical anomalies in oil-well monitoring data and to assess their joint use. Six parameters from a multiphase water-cut meter were analyzed; Principal Component Analysis (PCA) reduced them to three principal components preserving 89.86% of the total variance. In a sample of 8639 observations, One-Class SVM identified 424 anomalous points and DBSCAN identified 188, with 85 observations detected by both methods. For flow-regime classification by gas-content level, SVM with an RBF kernel (C = 100.0, gamma = 1.0) achieved 98.26% accuracy with an F1 score of 0.97-0.99 on a stratified 70/30 split. Five-fold stratified cross-validation yielded an F1_macro of 0.9853±0.0026. Joint interpretation separates points detected by both methods from SVM-only and DBSCAN-only detections for subsequent expert review.

In aerial monitoring systems, a single detection error may develop into a persistent false track and affect data passed to downstream components. The study aims to reduce such errors while bounding the scale of intervention in advance and retaining a traceable account of each decision. An external method is proposed that briefly delays low-support track-initiation episodes. The number of delays is controlled by an accumulated budget, while the support estimate, current budget balance, reason for intervention, and final status are recorded for each episode. The method was evaluated on seven video sequences from a publicly available aerial-imagery dataset containing small objects, partial occlusions, and substantial scale changes. Three tracking algorithms and a supervised ensemble of decision trees were compared. For ByteTrack, the delay was applied to 2.62 percent of track-initiation episodes, removing 425 false observations and 176 false initiations from the operational output, while the change in the combined precision-recall measure remained within the predefined tolerance. For OC-SORT and SORT, the reduction in false observations was not confirmed under the adopted statistical criterion. An examination of 29,385 initial video fragments found no violations of the predefined limits. The additional processing time averaged 0.901 milliseconds per frame. The results show that a short delay can reduce the number of false observations while keeping the scale of intervention under control.

The article presents a methodology for quantitative assessment of the completeness, reliability and quality of primary data collected within the framework of monitoring the infrastructure of intelligent transportation systems on public roads of the Russian Federation. The relevance of the study is determined by the need for formalized quality control of source information used for the formation of consolidated ITS equipment inventories and subsequent analysis of the state of digital road infrastructure. The proposed methodology is a tool for verifying the reliability and completeness of primary data, on the basis of which such an analysis is performed. The scientific novelty of the study lies in the development of a system of interconnected criteria and integral indices intended for quantitative assessment of the completeness and reliability of ITS monitoring data, as well as for the formation of a consolidated information quality index. The methodology includes the calculation of partial data quality criteria, the formation of an integral index of data completeness, an integral index of data reliability and an integral information quality index. The proposed approach ensures reproducibility of calculations, comparability of monitoring results and the possibility of subsequent automation of ITS data quality assessment procedures.

The article examines the problem of the fragmented application of transport planning tools and Intelligent Transport Systems (ITS) in urban agglomerations. The relevance of the study is driven by a 22% increase in the motorization rate in the Russian Federation between 2013 and 2023, alongside persistently high accident rates – in 2024, 132,037 road traffic accidents were recorded, resulting in 14,403 fatalities. The isolated use of these tools leads to local optimization, where improvements in one section worsen the situation on adjacent sections. The aim of this work is to analyze the effectiveness of the joint development of transport planning documents and local ITS projects, and to identify integration points. The methodology is based on systemic and legal and regulatory analysis, with the empirical foundation comprising data from Rosstat, the Ministry of Transport of the Russian Federation, and the results of ITS implementation in 62 agglomerations. Four types of integration points are formalized (informational, target-based, evaluative, and regulatory). A system of key performance indicators is proposed, covering criteria for road safety, traffic management, comfort, and environmental impact. Regulatory barriers are identified, including the absence of a mandatory transport model for the initial maturity levels of ITS. An iterative scheme «model - project - data - model» is substantiated. The conclusion is drawn that institutional changes are necessary to achieve a synergistic effect.

The paper systematizes the results of the scientific and practical conference «PRO//Movement. Transport Management Systems» (August 18-20, 2026, Nizhny Novgorod), dedicated to the digitalization of the transportation process, artificial intelligence, and the robotization of railway transport. The conference was attended by the management of the Russian Railways holding, the scientific sector complex, and manufacturing enterprises: managers, scientists, and engineers. The target model for managing freight flows, the «Digital Railway Station» project, and technologies such as virtual coupling, machine vision, predictive diagnostics, and physical artificial intelligence were discussed. Special attention is paid to the role of the scientific activities of JSC «NIIAS» in shaping a unified data architecture, models, and executive systems, as well as to issues of functional safety, economic efficiency, and the replication of developments. The prospects for the development of artificial intelligence in the railway industry are discussed.