Adherence to standardized poses is paramount for ensuring safety in special operations, where non-standard poses — often resulting from negligence or inexperience — can precipitate personal injury and property damage. Human Pose Estimation (HPE) presents an avenue for the automated evaluation of behavior pose standardization. However, monitoring special operations necessitates exceptionally high HPE accuracy under stringent thresholds (e.g., Object Keypoint Similarity with threshold > 0.75) for reliable pose reconstruction to ensure safety. Complex environments introduce numerous non-causal confounding factors into feature representations, reducing the performance of existing HPE models, which are primarily reliant on feature correlations. To address this, Hierarchical Causal Decoupling Pose Estimation (HCDPE) model is proposed. Specifically, HCDPE decouples image features into causal and non-causal features. Meanwhile, by integrating multi-hypothesis structure and graph attention, a hierarchical causal gating module is proposed. This enables HCDPE to intervene in causal features at multiple granularities, thereby finely indexing causal correlations between features and pose representation. To optimize the above causal decoupling, the counterfactual causal effect estimation module is designed. This method guides the generation of loss by minimizing the mutual information between causal and non-causal features, thereby achieving causal effect estimation. Together, these components form a closed-loop causal inference chain, enabling HCDPE to achieve fine-grained HPE. Extensive experiments on a specialized dataset for special operations, and two public benchmarks, demonstrate the superior performance of HCDPE in engineering applications. The code is publicly available at https://github.com/onepionts/HCDPE.

With the rapid development of neural networks, strip steel surface defect detection, as an important task in computer vision, has achieved remarkable progress. However, state-of-the-art methods still face a tradeoff between accuracy and efficiency. High-performing models are usually large and computationally expensive, whereas lightweight models often suffer from limited detection accuracy. To address this issue, we first propose a spatial channel enhancement (SCE) module, which consists of a reparameterizable depthwise large-kernel convolution and a reparameterizable pointwise (RepPw) convolution. The proposed SCE module enlarges the receptive field and strengthens long-range spatial and channel interactions while preserving computational efficiency. Based on the SCE module, we propose a novel lightweight saliency model for strip steel surface defects, namely, reparameterization-driven depthwise separable large-kernel network (RepDSLKNet). RepDSLKNet employs SCE modules to build an encoder and a decoder, and utilizes cascaded channel attention (CCA) modules for the feature fusion. The lightweight architecture can effectively extract and fuse the semantic information and detailed features of strip steel surface defects, thereby improving the accuracy and speed of detection with a small model size. With an input size of $224 \times 224$ , our RepDSLKNet has only 0.47 M parameters and 0.42 G FLOPs during inference. Compared to the current state-of-the-art methods, our approach achieves a 19-fold improvement in throughput and a twofold reduction in latency. Experiments on two public strip steel defect datasets demonstrate that RepDSLKNet delivers competitive performance against state-of-the-art methods.

Recently, few-shot strip steel surface defect segmentation has received more and more concerns. However, the existing few-shot segmentation methods usually adopt the frozen encoder, which is pre-trained on the classification task and can only provide class-related knowledge. Therefore, we propose a novel method, namely pre-trained variational auto-encoder based latent gaussian process regression (LGPR), to conduct few-shot strip steel surface defect segmentation. Firstly, different from previous methods, the frozen Variational Auto-Encoder (VAE) based encoder and decoder, which are pre-trained by using the pixel-level self-supervised task (i.e., image reconstruction), can provide rich image-related knowledge. This ensures the effective characterization of defect regions. Secondly, by deploying a gaussian process regression in the latent feature space generated by the VAE-based encoder, pixel-level correlation between support features and query features can be efficiently built. This operation is non-parametric and doesn't bring any training overhead. Besides, we deploy transformer-based projectors to dig long-range contextual cues of support and query features. Extensive experiments are performed on two public datasets, and the experimental results clearly show that our model consistently outperforms the state-of-the-art models with a large margin. Both the codes and results are publicly available at https://github.com/Hlao-hub/LGPR.

Surface defect detection in strip steel is a critical task in industrial quality control. However, existing methods struggle with capturing both local details and global context effectively. In this paper, we propose the Global-Local Fusion Network (GLNet) for strip steel surface defect detection, which combines the advantages of VMamba's global feature extraction and CNN's local feature modeling. GLNet employs an encoder-decoder structure, where the encoder consists of two parallel branches: one based on VMamba for capturing global features and the other using ResNet50 for extracting local features. In the decoder, a Global-Local Fusion (GLF) module integrates these features using the Cross Prototype Objective Enhancement (CPOE) and Selective Spatial and Channel Attention (SSCA) modules. The CPOE module facilitates the interaction and fusion between global and local features, while the SSCA module digs the multi-scale information from the global feature through dynamic attention to guide the feature aggregation. Extensive experiments on the ESDIs dataset, demonstrate that GLNet achieves state-of-the-art performance in defect detection, surpassing 13 existing methods in both quantitative and qualitative metrics.

Multiple object tracking (MOT) is one of the key technologies for intelligent industrial information systems. Confidence fluctuation and identity switch are common occurrences in MOT, which can significantly decrease the tracking performance. To solve the above issues, we propose the evidence reasoning-based robust multiple object tracking method (ER-MOT). First, you only learn one representation (YOLOR) and deep simple online and realtime tracking (SORT) are used to calculate the tracking confidence of each target in the video stream. Then, ER-based dynamic update algorithm is proposed to enhance the robustness of the tracking method, which can convert the tracking confidence of the target into a piece of evidence of the target identity and dynamically update the current evidence by fusing historical evidence. In addition, the forgetting strategy and the simulated annealing algorithm are used to enhance the ER-based dynamic update algorithm performance. The effectiveness of the proposed method is verified on the multiple MOT Challenge datasets. Experimental results demonstrate that the proposed ER-MOT can effectively reduce the occurrences of identity switch and enhance the tracking robustness under disturbance scenarios including target occlusion, unstable video quality, and dynamic target changes.

Correlations among process variables and inconsistencies in alarm decision making are quite common in multivariate alarm analysis, resulting in a large number of false alarms and missed alarms. The greatest challenges in multivariate alarm analysis are therefore analyzing overall correlations among all process variables and making integrated alarm decisions. In this work, a novel correlation analysis-based multivariate alarm method is developed to address these problems. First, a statistical characteristic-driven decision making trial and evaluation laboratory (DEMATEL) is proposed that can analyze the overall correlations among all process variables. Second, the sample space model (SSM) and evidence space model (ESM) can be used to convert process data into reference alarm evidence. Third, online samples are transformed into alarm evidence by matching them with the ESMs and holistically considering the data-level correlations and the evidence-level reliability and weight; the comprehensive alarm evidence is obtained by fusing this matched alarm evidence generated from the information of highly correlated or even colinear variables via maximum likelihood evidential reasoning (MAKER), and thus, more accurate and integrated alarm decisions are made. A real case study shows the superiority of the proposed method, which can therefore be generalized to other multivariate industrial processes.Note to Practitioners—Multivariate industrial processes generally have a large number of process variables, and with the rapid transfer of energy, material, and information, these process variables interact with each other or are even colinear. The focus of this study is to develop a multivariate alarm method for the correlation analysis of process variables and fusion of complementary, redundant and contradictory process information. The information fusion concept takes the place of the conventional alarm mechanism. From the perspective of the precise characterization of process information, process data are transformed into alarm evidence instead of alarm data. In addition, the proposed method can fully consider the overall correlations among all process variables and fuse each piece of process variable information to yield correct and integrated alarm decision results. It is noted that the information fusion concept is universal and can be extended to other real multivariate industrial processes.

An experienced helmsman can always distinguish ships from a pile of radar blips in scenarios such as nearshore waters and inland rivers with a single glance. To replicate this intelligence, a novel approach called MRNet based on deep convolutional networks is proposed. It employs a highly customized neural network to extract critical information from successive radar scans, ranging from low-level characteristics to high-level semantics. The feature fusion network of MRNet is also built with a Depthwise Separable Convolution-based network, which reduces parameter size and calculational usage while improving overfitting issues significantly. In the final prediction procedure, a method based on weighted-box fusion and a Scylla-IoU function is used to accelerate convergence. A marine radar image dataset, namely radar3000, was established to validate the proposed approach. In the corresponding experiments, the recall, identification accuracy, and precision of MRNet reached 0.9663, 0.9418, and 0.9267 respectively. On the other hand, the parameter size and calculational consumption were controlled to only 34.41M and 21.55G respectively. Compared with the commonly-used fractal algorithms and the YOLO series, the MRNet can be described as significantly superior in the application of recognizing ships from marine radar blips, especially in crowded scenarios, which is very similar to human eyes, and can be of great use to navigation and coastal surveillance.

Abstract H7N9 avian influenza is a novel virus with high morbidity and mortality that threatens human health and life. Therefore, it is necessary to diagnose H7N9 avian influenza in a timely and rapid manner to prevent further transmission of the virus and greatly reduce the infection and mortality rates. This paper proposes an H7N9 avian influenza diagnostic model that is based on a multilayer belief rule‐based (BRB) inference methodology by considering five typical characteristics of influenza: epidemiology, clinical manifestations, complications, characteristics of imaging tests and positive pathogen test results. Specifically, the severity of H7N9 avian influenza is gradually identified by a multilayer BRB model, and then the diagnostic model is optimized by a genetic algorithm (GA) to improve the diagnostic accuracy. Finally, the feasibility of the model is verified by fivefold cross‐validation with a real clinical dataset. The performance of the proposed diagnostic model is compared with those of the BP neural network (BPNN) model and support vector machine (SVM) model, and the results show that the multilayer BRB model can achieve rapid and satisfactory diagnostic results for H7N9 avian influenza. The experiment shows that the accuracy of the BRB model for H7N9 avian influenza hierarchical diagnosis provided in this paper is 0.903, which is higher than 0.818 of the BP neural network (BPNN) modules and 0.844 of the support vector machine (SVM) models. Especially when diagnosing the suspected and confirmed degree of H7N9 disease, it is more realized satisfactory diagnostic accuracy.

Guaranteeing the safety of nearby buildings is essential in tunnel construction. In practice, it is implemented by closely monitoring the daily and accumulated settlements, which are dependent outputs. To accurately predict such outputs, a new approach using two features is proposed. First, a new concept of data credibility is proposed to represent the different levels of consistency among the multiple dependent outputs. Second, the training dataset is developed using data gathered from multiple phases based on the fuzzy number. The new approach is named credibility-based fuzzy incremental learning approach using the belief rule base (BRB), CI-BRB. The key contributions of the proposed CI-BRB approach are: 1) data credibility is defined and calculated rather than blindly assuming all data are accurate by default and 2) the training dataset is more representative as it includes both the current phase and more prior phases. Subsequently, a numerical case with three dependent outputs is designed to provide a detailed illustration, and a practical case is studied in a more comprehensive fashion. The case study results show that the proposed approach can produce superior results for modeling that adopts a none strategy. Additionally, further investigations validate the effectiveness of the strategy over incremental learning.

This article investigates how to make use of imperfect data gathered from different sources for inference and decision making. Based on Bayesian inference and the principle of likelihood, a likelihood analysis method is proposed for acquisition of evidence from imperfect data to enable likelihood inference within the framework of the evidential reasoning (ER). The nature of this inference process is underpinned by the new necessary and sufficient conditions that when a piece of evidence is acquired from a data source it should be represented as a normalized likelihood distribution to capture the essential evidential meanings of data. While the explanation of sufficiency of the conditions is straightforward based on the principle of likelihood, their necessity needs to be established by following the principle of Bayesian inference. It is also revealed that the inference process enabled by the ER rule under the new conditions constitutes a likelihood inference process, which becomes equivalent to Bayesian inference when there is no ambiguity in data and a prior distribution can be obtained as a piece of independent evidence. Two examples in decision analysis under uncertainty and a case study about fault diagnosis for railway track maintenance management are examined to demonstrate the steps of implementation and potential applications of the likelihood inference process.

When using the belief rule base (BRB) methodology to deal with the nonlinear causal inference problems, combinatorial explosion often occurs due to overnumbered antecedent attributes, resulting in poor performance. Therefore, this paper proposes a novel nonlinear causal inference approach based on vector-based BRB. In the modeling process of BRB, the original attributes are ranked by contribution rate and transformed into attribute vectors. Meanwhile, combined with the k-means method, appropriate referential vectors are obtained. Thereby a vector-based BRB can be established. In the inference process of BRB, the idea of full activation of vector-based rules is presented. By calculating the spatial matching degree of the testing sample and the referential vectors, activation weights of the rules which are used in the evidential reasoning algorithm are acquired. Experimental results of a nonlinear function with four-dimensional input and the pipeline leakage detection data show the effectiveness and superiority of the proposed approach.