Instrumentation Technology and Economy Institute

Software vulnerability detection is a critical research direction in software security. Pre-trained models can automatically learn deep semantic and structural features of code, significantly enhancing the accuracy and effectiveness of vulnerability detection methods. However, existing vulnerability detection models often struggle to comprehensively capture the multi-dimensional features of code using a single modality. To address this issue, this study proposes a multi-modal representation learning-based vulnerability detection approach. By proposing an innovative fusion of three code representation modalities—text sequences, abstract syntax trees, and data flow graphs—it achieves effective complementarity among multi-modal code features, enabling the model to efficiently capture key vulnerability-related features within code lines. Experimental results demonstrate that the proposed vulnerability detection method outperforms existing models, and proves that fusing diverse code modalities can significantly boost the performance of pre-trained models.

This work presents a breakthrough in fabrication of high-aspect-ratio micro gas chromatography (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \text{GC}$</tex>) columns. Silicon-glass hybrid <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \text{GC}$</tex> columns were achieved for the first time through anodic bonding silicon microfluidic channels and glass microfluidic channels. The designed maximum aspect ratio of the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \text{GC}$</tex> columns reaches to 18.3. The glass microfluidic channels were fabricated using thermal reflow technology, proving its feasibility in process high-aspect-ratio glass-based microstructures. The resulting <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \text{GC}$</tex> columns exhibit excellent separation performance, successfully resolving n-alkanes mixture (C5-C9) within 6.5 minutes. The column efficiency of the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \text{GC}$</tex> columns with channel width of 50 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \mathrm{m}$</tex> is 1306 plates <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$/ \mathrm{m}$</tex>, confirming the effectiveness of the proposed silicon-glass hybrid structure for rapid separation.

Abstract High-precision diffused silicon piezoresistive pressure sensors are widely used in industrial automation, aerospace, and other fields. During their metal encapsulation process, corrugated diaphragms are often connected to the core using laser welding technology. However, it remains unclear how the deformation of the corrugated diaphragm, induced by thermal stresses during this process, impacts sensor accuracy. Therefore, this paper investigates the evolution patterns of corrugated diaphragm morphology induced by laser welding thermal effects and quantitatively evaluates the mechanism by which this evolution affects sensor accuracy. First, we established and validated a coupled thermal-deformation finite element simulation model against actual morphology measurements. Second, based on elastic mechanics theory, the relationship between corrugated diaphragm deformation and sensor output error was derived. Finally, the ‘welding thermal effect-corrugated diaphragm deformation-sensor accuracy’ mechanism was established, enabling analysis of the sensor’s full-scale error, sensitivity offset, and hysteresis error. Results indicate that welding thermal effects cause diaphragm deformation (average corrugation height difference &lt;10 μm before and after welding, variance increases from 3.32 to 19.36, and the diaphragm unevenness coefficient increases from 0.056 to 0.126, with warping exceeding 15 μm between the third peak and fourth trough). This deformation subsequently impacts sensor accuracy, increasing full-scale error from 0.1141% to 0.2226%, sensitivity offset from 0.0159% to 0.0222%, and hysteresis error from 0.0013% FS to 0.0019% FS. Among these, warping is the key factor causing accuracy degradation. The above results provide reference and basis for optimizing the design of high-precision pressure sensors.

Accurate train positioning is crucial for ensuring the safety and operational efficiency of intelligent railway systems. Traditional Global Navigation Satellite System (GNSS) and Inertial Navigation System (INS) integration often performs reliably only under ideal conditions. However, in GNSS-denied environments such as tunnels, signal blockage can cause significant positioning errors. To overcome these limitations, this paper proposes a seamless train positioning framework that dynamically switches between Visual Place Recognition (VPR) /INS and GNSS/INS integration. The framework operates adaptively: in tunnel or signal-denied scenarios, where GNSS is unavailable, VPR/INS integration maintains accurate localization; in open-sky environments, GNSS/INS integration ensures continuous positioning. To enable VPR-based localization, a reference image database of the railway line is pre-constructed, and the NetVLAD model is employed to extract robust global image descriptors. A location-guided image retrieval strategy is then introduced to accelerate global feature matching, providing a coarse estimation of the train’s position based on retrieval results. To ensure the accuracy of the matching process, we have implemented a double-check mechanism that accepts matching results only when validated, thereby enhancing overall railway positioning safety. Field experiments conducted on the Lhasa–Nyingchi Railway show that during GNSS signal blockage, the proposed switching system effectively reduces positioning errors, achieving an overall RMS error of approximately 1.03m.

This paper improved the back-propagation neural network (BPNN) algorithm for recommending ideological and political courses by a squeeze-and-excitation network (SEnet) and a multi-head attention mechanism. Simulation experiments were conducted to compare the improved algorithm with two other recommendation algorithms, followed by ablation experiments. Moreover, the effectiveness of the recommendation algorithm was tested in actual teaching of ideological and political courses. The results demonstrated that the improved BPNN algorithm outperformed others and the SEnet and the multi-head attention mechanism significantly enhanced the accuracy of recommendations. The algorithm effectively improved students’ performance in ideological and political courses and was satisfied by the majority of students.

In this paper, we propose a 5G communication architecture integrating TMS (Transport Management System) in response to the latency sensitivity and reliability requirements of 5G terminal communication in industrial logistics process. Through the user plane function (UPF) sinking, dual-transmitter selective reception redundant link design, and dynamic QoS enhancement mechanism, deterministic low latency (<20ms@99.99/%) and ultra-reliable communication of industrial terminal equipment have been achieved. Experimental results demonstrate that in the scenario of dense scheduling of AGVs (Automated Guided Vehicles), the standard deviation of the end-to-end delay of the network is reduced from 4.7 ms in the traditional scheme to 1.2 ms, and the synchronization accuracy error is controlled within <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2.5 \mu \mathrm{s}$</tex>. This architecture supports multi-protocol terminal access, which provides key technical support for the intelligent upgrade of Industry 4.0.

The rapid growth in the number of special equipment has led to an exponential increase in inspection case data; however, this data's effective utilization remains limited, and traditional manual retrieval and processing methods are insufficient to meet intelligent transformation demands. This paper proposes an information extraction framework for special equipment inspection cases based on large language models. A professional domain ontology comprising 15 core elements was established, alongside a comprehensive annotation scheme covering event classification, equipment details, inspection procedures, defect identification, and remediation plans. The Qwen2.5-7B base model from the Qwen Tongyi series was efficiently fine-tuned on 3,000 labeled samples using Low-Rank Adaptation (LoRA) technology, significantly enhancing its capability to comprehend and extract domain-specific terminology and information. Experimental results demonstrate that the fine-tuned model achieves notable improvements in professional vocabulary recognition accuracy, target field localization, and structured output standardization. This approach effectively extracts key information from inspection cases, providing a robust technical foundation for constructing a special equipment safety knowledge graph and advancing intelligent decision support systems.