Abstract To tackle the challenge of insufficient diagnostic accuracy for rolling bearings under cross conditions, this paper proposes a fault diagnosis method based on a ConvNeXt-FECAM architecture, integrated within an enhanced conditional adversarial domain adaptation framework. Specifically, each one-dimensional vibration signal is first reshaped into a two-dimensional (2D) matrix, and a 2D fast Fourier transform magnitude transform is then applied to construct a frequency-enhanced input representation for the ConvNeXt-FECAM feature extractor. In the feature extraction stage, FECAM is embedded into ConvNeXt blocks to assign attention weights to learned feature channels, allowing the network to emphasize feature responses that are more useful for fault-related representation. To further alleviate domain shifts arising from varying operating conditions, a conditional-marginal joint domain alignment strategy is introduced by combining conditional adversarial training with multi-kernel maximum mean discrepancy. Extensive experiments on three rolling bearing datasets demonstrate that the proposed method achieves higher diagnostic accuracy and better transfer performance than representative comparison methods under target operating conditions. Furthermore, ablation studies confirm the contribution of the main components in the proposed framework. The results indicate that the proposed approach has potential for cross-condition fault diagnosis and condition monitoring of rotating machinery.
Abstract Fully automatic operation (FAO) systems are a key development trend in urban rail transit. Their testing and verification encompass train control, train-to-ground (T2G) communication, interface interactions, and operational scenarios. Field testing, however, faces practical constraints, high costs, and limited coverage of hazardous scenarios, making simulation-based testing indispensable. This paper proposes a zero on-site testing platform for FAO systems that integrates hardware-in-the-loop and system-in-the-loop techniques. The platform reproduces train-control systems, T2G wireless communication systems, and train operating scenarios in a laboratory environment. First, two representative train-control systems—the Chinese train control system (CTCS)-3/2+automatic train operation (ATO) and communication-based train control—are analyzed, and the overall architecture of the zero on-site testing platform is presented. Second, the platform is used to test two representative T2G wireless communication systems: Long Term Evolution for Metro (LTE-M) and 5G for Railway (5G-R). Transmission delay, throughput, and packet loss rate are analyzed to evaluate the platform’s link-level testing capability. The 5G-R system achieves an average end-to-end delay below 36 ms, a maximum downlink throughput of 33.2 Mb s − 1 , and a packet loss ratio below 1%. Moreover, multiple on-board ATO test instances are connected to a real trackside temporary speed restriction server through the 5G-R zero on-site testing platform. Concurrent registration tests involving up to 60 trains are conducted to validate the platform’s capability for multi-train interface and protocol conformance verification. Finally, a virtual-coupling operational scenario is simulated to evaluate train speed control, inter-train distance keeping, and responses to dispatching commands. The experimental and simulation results demonstrate that the proposed platform supports T2G communication measurement, multi-train registration verification, and the simulation of key operational scenarios for FAO systems. The platform thus provides a practical means of reducing field-testing costs, improving test coverage, and facilitating the verification of intelligent train-control technologies.
Abstract Fault diagnosis in complex manufacturing systems presents a critical challenge that requires not just feature extraction but also symbolic knowledge and expert reasoning for reliable, explainable diagnostics. This paper introduces FD-LVLM (fault diagnosis with large vision language model), a framework for intelligent mechanical fault diagnosis. We first propose a method for creating a fault diagnosis semantic dataset, which embeds expert-level diagnostic knowledge into structured textual descriptions paired with visual images of sensor signals generated by the continuous wavelet transform. This multi-source dataset is then used to fine-tune a LVLM using our proposed asymmetric QLoRA, a novel parameter-efficient strategy that selectively adapts the model’s visual pathways while preserving its core linguistic reasoning. To use the model’s fine-tuned knowledge and enhance its deductive power, we introduce a reasoning chain, a structured prompting framework applied during inference that guides the LVLM through a step-by-step, expert-aligned diagnostic process. This complete framework allows the model produce accurate fault classifications, fine-grained explanations, and actionable maintenance suggestions. Crucially, we validate the model’s explainability by demonstrating that its diagnostic process is a context-aware, progressive sharpening of visual attention, where the model’s focus dynamically shifts in response to the evolving textual diagnosis. This process is shown to be directly linked to the visual evidence for multiple fault types. Validated on two distinct datasets, FD-LVLM demonstrates significantly enhanced performance over deep learning benchmarks. This paper presents a robust, end-to-end methodology for developing diagnostic systems that are not only highly accurate but also inherently explainable.
Abstract To improve the inverse reconstruction stage of acoustic tomography (AT)-based temperature field measurement, a novel neural network called CPP-Net is proposed. It is a temperature reconstruction model built on a multi-scale deep unfolding neural network inspired by the Chambolle–Pock algorithm. The model embeds trainable modules through a deep unfolding framework and introduces a dual path fusion block architecture to facilitate multi-scale feature extraction and feature fusion in the signal domain. This design improves the structural fidelity of the reconstructed temperature distribution. A series of representative temperature field models are constructed for simulation studies, and the reconstruction performance is further validated through experiments on an AT-based temperature measurement platform. Simulation results show that CPP-Net achieves improved reconstruction accuracy across various temperature distribution scenarios, with an average relative error of 0.13% and a root mean square error of 0.17%, outperforming various traditional algorithms and deep learning-based methods. In experimental evaluations, CPP-Net also yields the lowest temperature reconstruction error at thermocouple sampling points, indicating its potential for practical AT-based temperature measurement applications.
Abstract Defect detection in drainage pipelines plays an essential role in the sustainable maintenance of urban infrastructure. Although deep learning has gradually supplemented traditional manual inspections, which are often inefficient and subjective, existing detection approaches still struggle with several inherent difficulties. Specifically, these include weak boundaries against complex textures, significant scale variations with geometric diversity among defects, and low-contrast regions under non-uniform lighting, all of which hinder accurate and robust detection performance. To deal with these challenges, this study proposes edge-enhanced multi-scale dynamic fusion (EMDF)-DEtection TRansformer (DETR), a novel detection framework built upon the enhanced real-time DETR (RT-DETR) architecture. The strategy consists of three essential components: an edge feature infuse backbone that incorporates explicit Sobel operator features to strengthen boundary perception; a modulated multi-branch auxiliary feature pyramid network that adaptively fuses multi-level features through dynamic weighting; and a Pola multi-view transformer module, which employs polarity-aware linear attention and multi-cognitive visual projection to enrich feature representation across scales and domains. Extensive evaluations on a self-constructed dataset containing ten categories of pipeline defects show that EMDF-DETR achieves a mean average precision (mAP) of 85.25 % and an F 1 score of 84.86 % , exceeding the baseline model by 3.42 and 4.62 percentage points, respectively. The method also surpasses several current state-of-the-art detectors across multiple evaluation criteria. Experiments and visual interpretations confirm the contribution of each component. Additional validation on public datasets further demonstrates its generalization capability, with mAP improvements of 3.09 % on NEU-DET and 1.90 % on GC10-DET over the baseline, underscoring the method’s strong potential for practical inspection tasks.
