Near-infrared spectroscopy provides a non-destructive and rapid route for soil analysis. However, conventional chemometric models based on nonlinear supervised learning remain limited for complex samples and multi-characteristics prediction. This paper proposes a chemometric framework based on spectral-characteristics fusion and minimization of representation mapping or prediction error. First, spectral-characteristics fusion integrated spectral data with single or multiple soil characteristics. Subsequently, quantitative models were constructed using a supervised spectral-to-compositional representation mapping model and a dynamic series forecasting model, which minimize the spectral-to-compositional representation mapping error and the prediction error, respectively. The dynamic series forecasting model was constructed based on dynamic sequential data analogous to time series data derived from spectral-characteristics fused data, with sliding windows covering all soil-characteristic positions. Experiments on organic samples from the LUCAS 2009 topsoil data showed that fused data with great continuity facilitated model construction, and the proposed framework achieved higher predictive accuracy than the selected conventional supervised learning baselines. This paper provides a near-infrared spectral-characteristics fusion and error-minimized prediction strategy for compositional analysis of complex samples, with potential applicability beyond soil analysis.

Casting defects on the surface of automotive engine blocks directly affect engine performance and service life, making rapid and accurate detection of these tiny defects crucial for improving automotive manufacturing quality. To overcome the limitations of traditional detection methods, such as low efficiency, heavy reliance on manual inspection, and frequent missed or false detections, this study proposes an improved YOLOv8-based automatic detection method for tiny defects on engine block surfaces, named YOLO-CFAEW. The proposed model integrates several key innovations: the introduction of a convolutional attention module (CBAM) into the YOLOv8 backbone to enhance feature extraction in both spatial and channel dimensions, and the design of the Receptive Field Attention Convolution (RFAConv) module, which dynamically adjusts the receptive field to better capture subtle defect features. Additionally, an improved multi-head detection structure is employed to optimize detection heads for defects at various scales, significantly improving the accuracy for tiny defects. A novel Wasserstein distance loss function is utilized to optimize the bounding box regression process, enhancing localization stability and accuracy of small defects. To balance detection accuracy and inference speed, the conventional coupled detection head in YOLOv8 is replaced with a decoupled detection head integrated with a lightweight GiraffeDet network structure. Experimental results demonstrate that the proposed YOLO-CFAEW model achieves outstanding performance on the expanded engine block defect dataset, with mAP@0.5 improved by 17.3

Infrared spectroscopy technology offers advantages such as non-destructive analysis, high efficiency, and low cost, making it highly valuable for quantitative soil monitoring. To address challenges such as the complex soil background in mining areas and the weak spectral response of iron, this paper proposes a soil detection framework called PB-ELM based on infrared spectroscopy. First, baseline drift was removed using 0.8-order Fractional Order Differentiation (FOD), and spectral dependencies were characterized using the SelectKBest method based on Normalized Mutual Information (NMI). Deep spectral features were extracted by stacking multiple layers of autoencoder extreme learning machines. This paper proposes the dynamic local penalty-based batch Bayesian Optimization (PBBO) for network architecture determination of the system. By leveraging a multi-core environment for parallel optimization, we address the inefficiency of Sequential Bayesian Optimization. The local penalty space is dynamically configured during the batch optimization process to enhance sampling diversity. Information gain is integrated into k-medoids clustering to refine candidate parameters. Compared to methods such as BRVFL and LS-SVM, PB-ELM achieved the best prediction performance, with approximately 95 % of the samples falling within the consistency limits and only a slight systematic bias, yielding RMSE of 2.3742 % under FOD0.8 and NMI600. Finally, this study achieved large-scale inversion of iron content at a 10-meter pixel resolution in the disturbance zone. The complete spatial evolution chain from the waste dump and downstream valleys to the mining area was mapped, demonstrating its potential for application in ecological restoration of mining areas.

Molten steel temperature is a core indicator in Consteel arc furnace smelting. Its accurate prediction is crucial for enhancing control precision, optimizing processes, and reducing costs. However, the extreme conditions inside the furnace and the limited number of temperature measurements make it difficult to continuously and accurately predict molten steel temperature. This paper proposes a mechanism constrained LSTM prediction model tailored to electric arc furnace processes. By modifying the loss function to incorporate the temperature variation range derived from the mechanism model into network training, it mitigates the issue of insufficient temperature labels. Concurrently, The parameter ranges in the mechanistic model were optimized using an improved particle swarm optimization algorithm to obtain new parameter ranges. The mechanistic model with these new parameters was then employed to guide the network training. Experimental results demonstrate that the proposed optimization algorithm enhances prediction accuracy.

Deep reinforcement learning (DRL) has been widely applied in microgrid energy management, yet existing methods typically rely solely on the statistical fitting of numerical observations and forecast data. This limitation predisposes agents to fit the stationary statistical patterns of training data. Consequently, in out-of-distribution scenarios lacking historical experience, agents struggle to utilize short-term forecast information to make correct risk-avoidance decisions, thereby reducing their robustness in uncertain environments. To address this, this paper proposes L2-Grid, a semantics-enhanced reinforcement learning framework that integrates the reasoning capability of Large Language Model (LLM) into the microgrid dispatch loop. This method utilizes an LLM as an auxiliary reasoning module to convert multi-source numerical observations and short-term forecasts into textual descriptions containing operational contexts. This paper designs a dual-stream gated cross-modal attention architecture to fuse pre-trained semantic embeddings with numerical time-series features. A global learnable scalar controls the overall scale of the semantic branch, while the state-dependent attention context allows its effective contribution to vary across operating conditions. Experimental results based on real-world microgrid data indicate that, compared to traditional purely numerical benchmark methods, L2-Grid exhibits superior economic performance and stability in out-of-distribution scenarios involving extreme weather. Furthermore, by visualizing cross-modal attention weights, this paper reveals the agent’s information attention preferences under different operating conditions, which helps enhance the trustworthiness of the algorithm in practical power system applications.

With the advancement of industrial digitalization, data models have been successfully applied in this field. However, industrial processes are complex and subject to various disturbances, inevitably leading to detected data containing label noise. To address this issue, we propose a continuous label noise filter based on an adaptive threshold and integrated iterative filtering framework. First, to overcome the limitations of existing adaptive thresholds, we design a novel adaptive noise evaluation criterion. This criterion generates adaptive threshold for identifying noisy samples based on the regressor's predictions for nearest neighbor samples. Second, the filter's framework integrates an iterative strategy with two ensemble strategies. Each iteration employs multiple regressors to collectively identify noisy samples, while the aggregated results from multiple iterations are used to recover erroneously discarded samples, dynamically determining noisy instances within the dataset. To evaluate the performance of the proposed noise filter, comparative analysis was conducted during the experimental phase against several representative labeled noise filters. The superior performance of the proposed method was validated on both public datasets and real-world industrial datasets.

Risk assessment is a vital component throughout the aircraft manufacturing process, playing a crucial role in ensuring safety and controlling costs. Traditional assessment methods largely depend on the subjective judgment of decision makers, the rationality of which cannot be effectively verified. Moreover, the subjective evaluations are often coarse and may result in inaccurate risk ranking outcomes. To address these limitations, the Analytic Hierarchy Process (AHP) is employed in this paper for risk assessment in aircraft manufacturing. Considering the shortcomings of conventional AHP in handling inconsistent judgment matrices and determining ranking weights, this study proposes two improved algorithms based on optimization models. The two models introduce the K-factor method to effectively identify and correct abnormal elements in inconsistent judgment matrices, and employ the COP criterion to fine-tune the ranking weights, thereby assisting decision makers in achieving more rational decisions. Experimental results demonstrate that the improved AHP can substantially reduce the number of modified elements of inconsistent judgment matrices while retaining more judgment information of decision-makers in the ranking process, which contribute to more reasonable risk assessment results.

To achieve accurate simulation of the copper flotation roughing process, this study proposes a hybrid modeling approach that combines mechanistic analysis with data-driven techniques. First, based on bubble dynamics theory and flotation process principles, the core calculation methods for concentrate grade and recovery are derived, constructing a mechanistic model framework for a four-cell series roughing circuit. Addressing the challenge that the key parameter of the flotation rate constant cannot be directly measured online, the Levenberg-Marquardt optimization algorithm is employed to determine its value. Subsequently, a LightGBM machine learning algorithm is utilized to build a flotation rate constant fitting model, enabling high-precision mapping of the complex non-linear relationships between ten process parameters and the flotation rate constant. Based on this hybrid model, a fullscale simulation of the copper flotation roughing stage in a concentrator is completed. Validation results demonstrate the model's good estimation accuracy for the concentrate grade.

While transfer learning-based anomaly detection techniques are popular, they often focus on known normal samples, neglecting insights from known anomalies. To address this gap, we present a novel weakly-supervised anomaly detection model using adversarial transfer learning. This model aims to achieve high performance despite limited known and unknown samples. Our approach introduces a dual-domain balanced centralization strategy in the anomaly detection module, combining the support vector description algorithm with transfer learning to strengthen hypersphere constraints in a weakly-supervised setting. In the adversarial learning phase for domain adaptation, we introduce a sample-based gradient reversal strategy to effectively integrate normal samples and segregate abnormal ones, optimizing scarce abnormal data utilization. To enhance the use of unknown samples, we employ a weight calculation method based on the Cumulative Distribution Function within the weakly-supervised anomaly detection module, leveraging sample feature distribution patterns for improved anomaly detection precision. We conducted three sets of anomaly detection experiments to validate our methodology across diverse scenarios, encompassing two image recognition and one process industry dataset. The experimental outcomes underscore the superiority of our proposed algorithm over current state-of-the-art methods in anomaly detection, affirming its efficacy and practical utility.

Owing to the scarcity of anomaly samples and the high cost of labelling, time-series anomaly detection usually relies on normality assumptions. However, precisely delineating the boundaries of normal data is difficult, which results in misclassification. Existing anomaly-assumption-based methods are constrained by coarse-grained augmentation strategies and a lack of domain knowledge, which lead to pseudo abnormal samples with insufficient validity. Meanwhile, the lack of integration and utilisation of typical abnormal features results in limited generalisation capabilities. These challenges are addressed in this study by proposing the EASTKB framework, which integrates enhanced anomaly simulation with the triplet contrastive knowledge base. First, structured sequence refactoring is employed to synthesise realistic and diverse abnormal samples by combining nonlinear trends with Gaussian noise. Next, normal and abnormal samples are processed using parallel spatiotemporal feature extractors to obtain deep representation. Finally, a triplet contrastive knowledge base comprising general, normal, and abnormal knowledge bases is designed. Furthermore, additive similarity and diverse contrast learning are used to dynamically refine feature boundaries and prevent homogenisation. Experimental results reveal that EASTKB outperforms state-of-the-art methods on six real-world datasets.

Deep reinforcement learning (DRL) methods for microgrid economic dispatch often suffer from reduced decision accuracy due to environmental changes within control periods. To address this challenge, this paper proposes an information-enhanced DRL approach (IE-DRL) that incorporates predicted future load and photovoltaic power generation into the state space, enhancing the robustness of dispatch decisions. The proposed method is evaluated using real-world microgrid data from Shandong Province, China. Experimental results demonstrate that the extended versions of the four baseline DRL algorithms (IE-DDPG, IE-SAC, IE-TD3, and IE-PPO) achieved reductions in operating costs of 5.63%, 12.85%, 7.87%, and 6.52%, respectively, validating the effectiveness and generalizability of the proposed approach.

Establishing a gas pressure model of the purification system to accurately estimate the system output pressure is a powerful guarantee to ensure the stable and efficient operation of the acid production process. This paper presents a study on the establishment of gas pressure models for three facilities in the purification system at the Guixi Smelter in China, including a pulse jet fabric filter, an electrostatic precipitator, and a drying tower. Through analysis of the operating mechanisms of these facilities, pressure mechanism models are established for the gas flows. Considering the nonlinearity and time-varying nature of the model, an improved extended Kalman filter (EKF) algorithm is proposed to perform online identification of the unknown parameters within the model. Compared to the first-order EKF, the improved algorithm achieves significantly better performance without incurring additional computational overhead. Field experiments on pressure estimation at the acid production site validate the reasonableness of the established models, as well as the efficacy of the proposed identification algorithm.

Ore grade is an important indicator for evaluating the mining value of a mine. Reasonable mining based on ore grade distribution can aid in sustainable development and grade control within mining areas. Therefore, this paper aims to utilize Sentinel-2 multispectral data and an extreme learning machine (ELM) to rapidly explore the grade distribution in iron ore mines. First, direct orthogonal signal correction and mathematical transforms are used to correct and enhance multispectral information. Then, we propose a rough set ELM (RS-ELM) to learn multispectral data for grade distribution exploration. Specifically, we first discretize the hidden layer matrix (H) of the ELM and calculate the dependency degrees of the nodes on each other according to the rough set method. Then, the nodes with high dependency degree are deleted to ensure the lightness of the network structure and reduce the complexity of the ELM. Furthermore, a whitening transformation is performed on H to improve numerical stability. Comparison experiments prove the effectiveness of the RS-ELM. Finally, the mine grade distribution exploration model is established by using Sentinel-2 multispectral data and RS-ELM, and the grade distribution map of the mine is drawn. The distribution map clearly shows the trend of grade distribution in the mine. Moreover, based on the cut-off and industrial grades, we categorize the mine area into three zones: non-ore, secondary ore, and main ore areas. This categorization provides a foundation for developing the mine’s mining plan and grade control.

In industrial settings, outliers can significantly affect equipment stability and potentially jeopardize system's regular operation and performance. Especially when confronted with new equipment or operating conditions, it becomes imperative to establish a precise model with a limited dataset to swiftly and accurately detect and manage outliers. To tackle this issue, we introduce a transfer learning-based approach aimed at quickly adapting to novel environments and constructing efficient anomaly detection models. This strategy merges transfer learning with conventional anomaly detection techniques to create resilient models within a weakly supervised framework. In contrast to conventional methods disregarding unknown outliers, our approach incorporates a Gradient Partial Reversal technique, employing a domain adversarial mechanism to gently segregate outliers at a distinct level from the anomaly detection algorithms. This strategy yields training outcomes comparable to supervised models. To validate the efficacy of our model, we conducted experiments across three scenarios: digit image detection (utilizing the MNIST and USPS datasets), object recognition (employing the Office-Home dataset), and rolling bearing anomaly detection. Our results show that the proposed algorithm significantly outperforms existing state-of-the-art methods in terms of detection accuracy and robustness.

Accurate load forecasting is crucial for the stable operation and efficient dispatch of power grids. However, directly uploading raw load data to the cloud poses privacy risks, while the complexity of multifrequency, non-stationary signals and the challenges in multiscale feature extraction and hyperparameter optimization hinder traditional forecasting methods. To address these issues, this article proposes a cloud-edge collaborative deep learning framework for power load forecasting, termed edge-cloud collaborative strategy (EC-CVNT). On the edge side, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and variational mode decomposition (VMD) are integrated to extract multiscale features and denoise the raw data, thereby reducing computational burden on the cloud and enhancing data privacy. The processed features are then transmitted to the cloud, where the Transformer model is optimized using the Newton-Raphson-based optimizer (NRBO) to improve forecasting accuracy under complex conditions. Validation experiments conducted on a microgrid in Yantai, Shandong Province, China, demonstrate that the integration of CEEMDAN and VMD significantly increases predictive accuracy. Comparative experiments with various hybrid models further verify that EC-CVNT offers substantial advantages in enhancing power load forecasting performance.

Laser-induced breakdown spectroscopy (LIBS), as a high-throughput and rapid-response laser spectroscopic technique, offers distinct advantages for the quantitative detection of heavy metals in soil and serves as an efficient analytical tool for ecological remediation in mining areas. This article proposes a high-precision detection framework that integrates feature selection, 2-D state transition spectrogram construction, and small-sample data augmentation with hybrid deep modeling to enhance the application of LIBS in soil heavy metal measurement. First, 1-D LIBS signal is transformed into the 2-D Markov state spectrogram to capture the dynamic transition relationships between spectral bands. A lightweight spectral regression architecture, termed the Kolmogorov-Arnold multiscale network (KAMS-Net), is proposed based on a convolutional neural network (CNN)-long short-term memory (LSTM) structure. This model synergistically leverages the strengths of multiscale convolution, temporal sequence modeling, and function approximation. The multiscale modeling capability is enhanced through a parallel multibranch structure. A novel KANs layer is constructed at the network's periphery, replacing conventional fixed-weight structures with the learnable activation function set based on B-spline function. The edge activation mechanism enables better nonlinear modeling expressiveness and improves the interpretability of the network. Finally, a local conditional variational autoencoder (L-CVAE) is developed to generate an augmented dataset, augmented LIBS spectrogram (AugLiSpg) based on the small sample condition. Extensive experiments show that the proposed method outperforms conventional 1-D models and various 2-D baseline networks, highlighting promising potential of 2-D architectures for soil quantitative analysis based on LIBS.

Microgrids are a crucial component of modern energy internet, offering reliable distributed generation solutions. However, communication failures in data acquisition terminals, caused by long-distance transmission and electromagnetic interference, can lead to data loss and negatively impact economic dispatch accuracy. This study introduces a novel deep reinforcement learning algorithm, the Relative Entropy Regularized Asynchronous Update Soft Actor-Critic algorithm, to address these challenges. Firstly, the algorithm estimates missing data during communication failures using predictive models, ensuring the system continues to operate with accurate information. Secondly, to handle the challenges posed by prediction errors, a relative entropy regularization strategy is applied during training to maintain policy stability. Finally, upon communication restoration, the algorithm asynchronously updates the reward function using actual historical data to correct cumulative errors introduced by the predictive model. Results demonstrate that this method significantly enhances economic efficiency under photovoltaic and load communication failures.

To address the challenge of quickly developing reliable anomaly detectors with limited sensor data in industrial processes, we propose a novel single-classification anomaly detection algorithm that combines Contrastive Learning and Transfer Learning. Our algorithm leverages abundant normal data from similar devices to migrate anomaly detection rules to target devices. Using a supporting data vector description algorithm, we derive anomaly detection rules for both the source and target domain devices. Then, Transfer Learning transfers these rules from the source domain to the target domain. Furthermore, we introduce an innovative approach that forms sample pairs from two domain samples and applies Contrastive Learning to ensure a uniform distribution of sample features, even with limited samples. This approach allows for extracting discriminative feature representations and facilitates rule migration. Experimental evaluations using the IEEE PHM Challenge 2012 dataset and the Arc Furnace dataset demonstrate the superior performance of our approach compared to existing state-of-the-art algorithms.

The integration of mechanism knowledge and data-driven technology to establish a deep network for predicting Mach numbers in wind tunnel systems has engineering and scientific significance. An intelligent hybrid model is first established to predict the Mach number in a wind tunnel system with input backlash and noise. A recursive least squares (RLSs) regression modeling strategy is constructed to obtain a regression model. The RLS can describe the main dynamic response between input and output signals. The backlash may reduce the excitation of the input signal, which will be reflected in the output signal and result in output prediction error and parameter estimation deviation. Based on the residual neural network (ResNet) structure, the long short-term memory (LSTM) network with excellent time series prediction performance and the graph attention networks (GATs) with strong non-Euclidean feature extraction ability are, therefore, deeply fused and adopted to estimate the error of the identification model. A sliding window is designed to divide the data and enhance the correlation of adjacent time series error data. A residual block composed of the LSTM network and convolution layer is designed to extract local features. By transforming the data into the topology structure, the GAT is used to capture the non-Euclidean features of the data, and the attention mechanism is adopted to achieve effective data feature updating. Finally, the results of industrial real wind tunnel experiments confirm the effectiveness and practicability of the proposed hybrid modeling algorithm.

The structure identification of block-oriented nonlinear models is of great significance; however, the research in this area is still far from mature. This paper focuses on the structure identification of time delay Hammerstein models with polynomial nonlinearity. Identification involves estimating the order of the linear part, the order of the polynomial function, and the time delay. The proposed algorithm is derived by analyzing specific information matrices, which are based on sufficient excitation conditions. Additionally, this paper investigates the structure identification of the Hammerstein systems with arbitrary nonlinearities under the influence of measurement noise. An improved criterion is proposed to identify the structure of the linear part, which can also offer important information for determining the order of the polynomial function. Simulation experiments validate the effectiveness of the proposed algorithm. (c) 2025 Published by Elsevier Ltd.