Non-destructive testing (NDT), particularly X-ray inspection, is vital for industrial quality assurance, yet existing deep-learning-based approaches often lack interactivity, interpretability, and the capacity for critical self-assessment, limiting their reliability and operator trust. To address these shortcomings, this paper proposes <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">InsightX Agent</small>, a novel LMM-based agentic framework designed to deliver reliable, interpretable, and interactive X-ray NDT analysis. Unlike typical sequential pipelines, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">InsightX Agent</small> positions a Large Multimodal Model (LMM) as a central orchestrator, coordinating between the Sparse Deformable Multi-Scale Detector (SDMSD) and the Evidence-Grounded Reflection (EGR) tool. The SDMSD generates dense defect region proposals from multi-scale feature maps and sparsifies them through Non-Maximum Suppression (NMS), optimizing detection of small, dense targets in X-ray images while maintaining computational efficiency. The EGR tool guides the LMM agent through a chain-of-thought-inspired review process, incorporating context assessment, individual defect analysis, false positive elimination, confidence recalibration and quality assurance to validate and refine the SDMSD's initial proposals. By strategically employing and intelligently using tools, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">InsightX Agent</small> moves beyond passive data processing to active reasoning, enhancing diagnostic reliability and providing interpretations that integrate diverse information sources. Experimental evaluations on the GDXray+ dataset demonstrate that <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">InsightX Agent</small> not only achieves a high object detection F1-score of 96.54% but also offers significantly improved interpretability and trustworthiness in its analyses, highlighting the transformative potential of LMM-based agentic frameworks for industrial inspection tasks.

Intelligent Fault Diagnosis (IFD) has achieved notable advancements through deep learning technologies, yet it continues to confront significant challenges in complex industrial environments. Current methodologies exhibit limited frequency perception capabilities, which impedes the comprehensive capture of critical frequency components within fault signals, thereby affecting diagnostic accuracy. Furthermore, existing IFD approaches lack interpretability, making it challenging to correlate selected features with physical fault phenomena effectively, thereby diminishing the practical applicability of diagnostic outcomes. To address these issues, this paper presents an adaptive feature learning framework that integrates kurtosis-based explainable attention and frequency domain encoding(K-FDE) to enhance both the interpretability and accuracy of fault feature selection. The proposed framework incorporates a kurtosis-based interpretable fault feature selection module and an Fast Fourier Transform (FFT) based frequency domain encoding module, which together dynamically capture fault-relevant features while improving the interpretability of feature selection. Additionally, this framework introduces a Discrete Wavelet Transform (DWT)-driven frequency decomposition module and a Convolutional Neural Network (CNN)-driven feature learning module, facilitating detailed frequency decomposition from coarse to fine and multi-resolution frequency feature learning. Empirical evaluations on high-speed aerospace bearings and motor bearings datasets validate that the proposed method demonstrates exceptional noise robustness and superior frequency perception capabilities.

Forecasting the power consumption of the spacecraft is critical for optimizing its lifespan and task allocation. However, the complex electromagnetic environment of outer space introduces unavoidable noise into the collected electrical signals. Moreover, the various subsystems of a multipower spacecraft are affected differently by internal and external noise, making it challenging for the existing methods to effectively capture the features of long-term power consumption sequences. We propose adaptive frequency-pruning-enhanced (AFPE)-iTransformer, a robust time-series forecasting model designed for spacecraft telemetry forecasting under noise and long-range dependency conditions. The model combines three key components: Legendre memory projection for historical compression, adaptive top-kfrequency pruning for per-channel denoising, and an improved inverted transformer for cross-subsystem attention. Evaluated on three years of Mars Express (MEX) data, our method consistently outperforms the state-of-the-art baselines in both within-year and cross-year forecasting. It also achieves competitive efficiency, with fast model load time and moderate parameter size. While focused on power forecasting, the model’s modular design supports broader applications in telemetry and industrial forecasting. Model code and configurations are open-sourced for reproducibility.

The subdomain adaptation (SA) based intelligent cross-domain fault diagnosis methods aim to reduce the conditional distribution shift caused by variable working conditions. However, existing SA methods may be limited by the quality of pseudolabels, since misclassified pseudolabels will lead to alignment between irrelevant subdomains, resulting in erroneous category-invariant knowledge being accumulated. To tackle this, we present a dynamic subdomain pseudolabel correction and adaptation (DSPC-A) framework. Specifically, we propose an end-to-end pseudolabel correction algorithm, which integrates an auxiliary network to learn clean and general target label distribution from noisy pseudolabels. So that, the auxiliary network can guide the SA model to perform precise subdomain alignment using learned label distribution. Moreover, to allow the synergy training of the additional auxiliary network and SA model, we introduce an iterative learning strategy to dynamically perform pseudolabel correction and subdomain alignment. The iterative training makes two models complement each other, thus achieving better SA ability and diagnosis performance. The DSPC-A framework has been thoroughly verified under three fault diagnostic scenarios: cross load, cross fault severity, and cross mechanical equipment. Case study results demonstrate the superiority of the DSPC-A, which improves the SA performance by solely implementing simple pseudolabel correction methods without other complex techniques.

The growing prevalence of the photovoltaic (PV) systems has intensified the focus on fault prediction and health management within both the academic and industrial realms. Electroluminescence (EL) imaging technology, recognized as an advanced detection method, has substantiated its efficiency and practicality in identifying diverse defects. In this study, we introduce a novel framework for anomaly detection in the PV panel systems, leveraging multiscale linear attention and scale distribution alignment learning (MLA-SDAL). Initially, we employ a feature extraction framework based on the multihead linear attention to facilitate the deep-level feature modeling. This network excels in the high-dimensional feature extraction while optimizing the model complexity, achieving a lightweight design tailored for efficient deployment. Subsequently, an unsupervised anomaly detection framework is devised based on scale learning. This framework employs feature dimension transformation and generates efficient supervised signals for distribution alignment learning. This surrogate task enables the framework to adeptly capture and characterize the feature distribution of healthy samples. By gauging the consistency between the input data and the learned model, we precisely quantify the anomaly level of each instance, effectively executing anomaly detection. This approach not only bolsters the accuracy of anomaly detection but also enhances the model’s adaptability to intricate data distributions. Through experimentation on a genuine EL data set, our proposed framework demonstrates pronounced advantages. Comparative to the alternative machine learning or deep learning-based methods, its performance is notable. This accomplishment is poised to furnish robust support for practical applications in the PV panel anomaly detection within the industry.

Machinery fault diagnosis plays an important role in machine Prognostic and Health Management (PHM). Leveraging the abundant data obtained from the Industrial Internet of Things (IIoT), the health states of machines can be effectively recognized, thereby ensuring the safety of the mechanical system. However, the lack of noise robustness and insufficient frequency domain perception make traditional methods to extract weak fault-related signals difficult under highly noisy conditions in practical industrial scenarios. Therefore, a method with abundant frequency domain learning ability is urgently needed. To this end, this paper proposes a PHM framework, a soft thresholding Fourier feature refiner network (Soft-FFRNet), for highly noisy bearing vibration signal diagnosis. Specifically, this framework includes a Fourier feature refiner which selectively extracts and refines the feature in the frequency domain from the perspectives of amplitude and phase. It achieves the extension from the time domain to the frequency domain. In addition, the proposed framework utilizes several residual blocks with soft thresholding to effectively improve the noise robustness. Their thresholds can adaptively change during the training process. The high-speed aeronautical (HSA) bearing and the motor bearing datasets with different noise levels are used to evaluate this framework. The results show that the proposed framework can effectively diagnose the faults under highly noisy conditions.

Industrial visual monitoring (IVM) is crucial in enhancing the reliability and efficiency of manufacturing processes. Recently, large vision-language models (LVLMs) have demonstrated remarkable semantic understanding and natural language interaction capabilities, which provide a novel solution to IVM. However, LVLMs pretrained on common domains lack specific knowledge for IVM scenarios, causing insufficient adaptation to industrial image patterns and specialized textual corpora. In this article, we deeply studied the adaptation of LVLMs to IVM and proposed DefectGLM. First, we proposed the first large-scale multimodal wafer dataset as a reliable data basis for model domain generalization. Second, this model employs low-rank adaptation–based contrast visual adaptation to align with industrial image patterns and utilizes vision-language instruction tuning for professional knowledge alignment. DefectGLM is the first large-model-based wafer image recognition model, and can accurately identify 36 types of wafer defects and provide appropriate text descriptions. DefectGLM provides a new solution for the development of industrial large models.

The wheel wear status of high-speed trains (HSTs) is an essential indicator of their safety and reliability. However, due to the time-varying operating state of HSTs, noisy and complex non-stationary signals are collected. This makes it difficult for data-driven algorithms to learn valuable discriminative features from data. Therefore, this inspired us to introduce signal analysis methods with clear physical meaning to improve the interpretability and performance of prediction models. This paper proposes a novel multi-layer wavelet integrated convolutional neural network (MWI-Net) for predicting HST wheel-wear. Specifically, discrete wavelet transform (DWT) extends the feature learning space of CNN from the time domain to the wavelet domain, thereby capturing the frequency features that are difficult to learn in the time domain. As a remarkable information space, the DWT can effectively alleviate the frequency aliasing problem, enabling MWI-Net to distinguish valuable frequency information from complex signals. In particular, the proposed dynamic frequency aggregation mechanism endows MWI-Net with excellent frequency analysis and feature selection capabilities. Experiments on the real operation dataset of CRH1A HSTs show that MWI-Net accurately predicts the wheel wear curves, which is more competitive than existing deep learning methods. Furthermore, we demonstrate the feature learning mechanism inside MWI-Net through visual analysis and illustrate how it optimizes and extracts valuable features layer by layer.

Machine intelligence fault prediction (MIFP) is crucial for ensuring complex systems’ safe and reliable operation. While deep learning has become the mainstream tool for MIFP due to its excellent learning abilities, its interpretability is limited, and it struggles to learn frequencies, making it challenging to understand the physical knowledge of signals at the frequency level. Therefore, this article proposes a physically interpretable wavelet-guided network (WaveGNet) with deep frequency separation for MIFP, inspired by the sound theoretical basis and physical meaning of discrete wavelet transform (DWT). WaveGNet expands the feature learning space of CNN into the frequency domain, allowing for a better understanding of the physical insights behind the frequency level. Specifically, WaveGNet involves a derivable and learnable frequency learning layer (FL-Layer) consisting of a wavelet-driven frequency decomposition module and a convolution-driven feature learning module. Multiple DWT-driven FL-Layers are used in WaveGNet to achieve deep frequency decomposition and multiresolution frequency feature learning in a coarse-to-fine manner. The effectiveness of WaveGNet was evaluated in real high-speed train wheel wear monitoring and high-speed aviation bearing fault diagnosis cases. Experimental results showed that WaveGNet outperforms cutting-edge deep learning algorithms and has excellent fault diagnosis and prediction abilities. Furthermore, an in-depth analysis of the learning mechanism of wavelet-driven CNN from the frequency domain perspective was conducted.

Fault diagnosis is an essential means to ensure the regular operation of mechanical systems. The existing data-driven algorithms are developed based on the assumption that the given label is entirely correct. However, mislabeling is common, which often occurs in industrial applications. These methods will overfit these mislabeled samples, resulting in inferior generalization. To this end, this article proposes a novel multistage true label distribution learning algorithm. Specifically, based on the training characteristics of data-driven algorithms on noisy datasets, a novel multistage adversarial loss function (MSA-Loss) is proposed. MSA-Loss can make the model construct the true label distribution from noisy datasets, prevent the model from overfitting the noisy samples, and finally keep the model with good generalization. The proposed method can be easily applied to any existing data-driven algorithm to improve its performance on noisy datasets. Our method is verified on high-speed aeronautical bearing and motor datasets, which prove that MSA-Loss has an excellent performance in noisy label scenarios. It can significantly improve the potential of existing diagnostic models in practical industrial applications.

State-of-the-art Deep Neural Networks (DNNs) are promoting medical image processing. However, DNNs are susceptible to adversarial attacks, which could significantly deteriorate model performances and pose a threat to clinical diagnoses. One common method for prevention is through adversarial training, which is highly dependent on harnessing adversarial examples during the training stage. However, adversarial medical examples generated by many existing works are too perceptible to be adversarial examples. To improve the imperceptibility, we proposed a Regional Saliency Map Attack that generates an adversarial example by only perturbing a small number of pixels. Extensive experiments have shown that, on average, our method caused the same degradation in model performance by quantitatively less perceptible perturbations. Visualisations have also verified that the improvement in imperceptibility in an image is both global and regional.

Accurate and real-time fault diagnosis (FD) and working conditions identification (WCI) are the key to ensuring the safe operation of mechanical systems. We observe that there is a close correlation between the fault condition and the working condition in the vibration signal. Most of the intelligent FD methods only learn some features from the vibration signals and then use them to identify fault categories. They ignore the impact of working conditions on the bearing system, and such a single-task learning method cannot learn the complementary information contained in multiple related tasks. Therefore, this article is devoted to mining richer and complementary globally shared features from vibration signals to complete the FD and WCI of rolling bearings at the same time. To this end, we propose a novel multitask attention convolutional neural network (MTA-CNN) that can automatically give feature-level attention to specific tasks. The MTA-CNN consists of a global feature shared network (GFS-network) for learning globally shared features and K task-specific networks with feature-level attention module (FLA-module). This architecture allows the FLA-module to automatically learn the features of specific tasks from globally shared features, thereby sharing information among different tasks. We evaluated our method on the wheelset bearing data set and motor bearing data set. The results show that our method has a better performance than the state-of-the-art deep learning methods and strongly prove that our multitask learning mechanism can improve the results of each task.

The critical issue for fault diagnosis of wheel-set bearings in high-speed trains is to extract fault features from vibration signals. To handle high complexity, strong coupling, and low signal-to-noise ratio of the vibration signals, this article proposes a novel multibranch and multiscale convolutional neural network that can automatically learn and fuse abundant and complementary fault information from the multiple signal components and time scales of the vibration signals. The proposed method combines the conventional filtering methods and the idea of the multiscale learning, which can extend the breadth and depth of the feature learning process. Consequently, the proposed network can perform better. The experimental results on the wheelset bearing dataset demonstrate that the proposed method has better antinoise ability and load domain adaptability and can diagnose 12 fault types more accurately when compared with the five state-of-the-art networks.

Recently, deep-learning-based fault diagnosis methods have been widely studied for rolling bearings. However, these neural networks are lack of interpretability for fault diagnosis tasks. That is, how to understand and learn discriminant fault features from complex monitoring signals remains a great challenge. Considering this challenge, this article explores the use of the attention mechanism in fault diagnosis networks and designs attention module by fully considering characteristics of rolling bearing faults to enhance fault-related features and to ignore irrelevant features. Powered by the proposed attention mechanism, a multiattention one-dimensional convolutional neural network (MA1DCNN) is further proposed to diagnose wheelset bearing faults. The MA1DCNN can adaptively recalibrate features of each layer and can enhance the feature learning of fault impulses. Experimental results on the wheelset bearing dataset show that the proposed multiattention mechanism can significantly improve the discriminant feature representation, thus the MA1DCNN outperforms eight state-of-the-arts networks.