With the application of unmanned forklifts in industries, industrial pallet recognition becomes an increasingly key role. However, real-time pallet recognition with high precision remains challenging due to adverse industrial environment and limited computational resources on embedded devices. To address it, an improved hierarchical lightweight model based on you only look once (YOLO) (named YOLO-HL) is proposed. The YOLO-HL introduces the discrete wavelet transform and learning feature for preserving texture details and fusing cross-stage features respectively. Both modules can reduce the convolution operation burden and extract effective features from poor-quality images. In addition, the YOLO-HL drops the large-object detection layer and adopts SIoU loss. It can further reduce the computational burden and improve the model accuracy. Experimental results show that the YOLO-HL achieves higher accuracy and needs less computational resource than the common YOLO models. It can also be deployed on embedded in-vehicle system effectively.

Deep stochastic configuration network (DSCN) has demonstrated great potential in supervised machine learning tasks (e.g., data regression and pattern classification) due to its advantages of efficient model training and adaptive structure determination. However, its application in industrial fault detection scenarios remains unexplored, where only one class of normal training data is available. To handle this issue, this article presents a self-supervised DSCN model, called contrast learning-assisted one-class DSCN (CL-ODSCN), for fault detection in complex industrial systems. It consists of four contributions: 1) a ODSCN fault detection framework is built based on the self-learning philosophy, where feature information enhancement (FIE) module is designed between two consecutive feature extraction layers to capture rich feature information for effectively monitoring the anomalous behaviors; 2) for effective fault detection, a modified objective function using CL is constructed for strengthening the model discriminative capability for abnormal samples; 3) to implement automatic node configuration, a supervisory mechanism for model training is presented; and 4) a probabilistic fault explanation strategy is developed by combining the Shapley value method and the Bayesian inference theory to locate fault-related input variables. Two industrial cases, including a benchmark chemical system and a cementing pump unit, are used to validate the proposed model. The results demonstrate that the CL-ODSCN model achieves better fault detection performance than the compared models.

Deep learning neural networks have been widely adopted for developing quality prediction models in industrial processes. Despite their strong capability of nonlinear intrinsic features, the existing models have some notable drawbacks, such as insufficient capturing of local spatio-temporal features, high computational complexity of model training, difficult determination of deep model structure, and lack of model interpretability. To address these issues, this paper presents an efficient automated deep spatio-temporal feature learning framework for dynamic industrial process soft sensing, named Deep Convolutional Partial Least Squares (DeCPLS). The proposed approach introduces the convolutional Partial Least squares (CPLS) model as a basic feature extraction unit and stacks multiple CPLS layers to construct an efficient deep dynamic feature learning model. A layerwise training mechanism is presented to facilitate the automated determination of model structures and hyperparameters, thereby reducing the computational complexity. Furthermore, a model prediction error explanation mechanism is introduced to analyze prediction outcomes effectively. Compared to classical deep neural networks, the proposed method demonstrates the advantage of efficiently capturing local spatio-temporal features while maintaining acceptable computational complexity. Finally, the superiority of the proposed method is validated through a simulated industrial case study and a real-world industrial application.

Plunger pumps serve as the core of the fluid delivery system, and their fault diagnosis plays a critical role in ensuring system reliability and operational safety. As a prominent deep learning technique, convolutional neural networks (CNNs) have demonstrated significant success in the field of machinery fault diagnosis. However, conventional CNNs typically employ an end-to-end structure that directly processes raw measured signals as model input, often overlooking in-depth exploration of the underlying physical information and thus resulting in limited diagnostic accuracy under complex working conditions. To overcome this challenge, this article presents a physical information-assisted attention convolution neural network (PACNN) for more robust and accurate plunger pump fault diagnosis. Unlike the traditional CNNs, which rely solely on original measured signals, PACNN enriches the network inputs by incorporating time-domain and frequency-domain features extracted from the measured signals. These features provide richer physical insights and are beneficial for enhancing the data representation. Furthermore, for handling the information fusion of the augmented network inputs, an improved region-focused multihead attention (MHA) mechanism is designed by accounting for the scale difference between sensor data and physical features. In addition, the PACNN model incorporates a learnable convolutional kernel structure based on B-spline functions, thereby boosting the model's capability to interpret complex fault patterns. Finally, an industrial plunger pump dataset is utilized to validate the method's performance, and the results show that the proposed PACNN achieves an average diagnostic accuracy of 99.50%, significantly outperforming traditional methods.

Deep learning, particularly convolutional neural networks (CNNs), has been widely applied in fault diagnosis field. However, most existing models are trained solely on given class labels and ignore intra-class distribution discrepancies, which degrade the diagnostic performance. To address this issue, this paper proposes an Improved Convolution Attention Neural Networks for Fault Diagnosis of Plunger Pumps via Adaptive Clustering-Guided Subclass Reconstruction. Different from the traditional CNNs relying on the original class labels, the proposed model applies an adaptive subclass reconstruction strategy based on an improved Density Peaks Clustering (DPC) algorithm to generate reliable subclass labels. The construction of subclasses is beneficial to elaborating on fault pattern classification. Based on the subclass labels, the proposed framework builds a fault classifier by integrating a CNN for extracting local feature with a multi-head attention mechanism for describing global dependencies. Experimental results on industrial datasets show that the proposed method achieves an average classification accuracy of 96.67 %, outperforming conventional CNN and attention-based models, thus demonstrating its robustness and effectiveness for intelligent fault diagnosis in complex industrial environments.

Deep support vector data description (Deep SVDD) has emerged as a popular chemical process fault detection technology. However, traditional deep SVDD method focuses on the static relationship mining, but omits the dynamic fault evolving characteristic (i.e., fault symptoms shift across variables and time). Moreover, as an unsupervised one-class modeling method, deep SVDD fails to utilize prior fault knowledge for boundary optimization, resulting in insensitive fault detection. To address these problems, this paper proposes an enhanced deep SVDD method, called Probe-Recall Deep SVDD (PR-DSVDD), for more effective chemical process fault detection. In the proposed method, two novel modules of representation probe and memory recall are designed for evolving characteristic representation and decision boundary refining, respectively. In the representation probe module, a dynamic spatiotemporal evolution encoder is developed to facilitate adaptive attention mechanism. This module captures the evolving weight distribution of critical channels across time steps, effectively modeling the symptom shifts and position-sensitive patterns in fault propagation. The memory recall module constructs a contrastive memory mechanism that leverages a few fault prototype features to actively calibrate the decision boundary. Further, a fault evolution explanation diagram based on relative SHAP values and parallel coordinate system is presented to visualize the fault propagation procedure. Experiments on the Tennessee Eastman process demonstrate that PR-DSVDD is superior to the compared fault detection methods by achieving an average detection rate of 82.88% with a low false positive rate of 1.28%, and the proposed fault evolution explanation plot captures physically consistent fault propagation paths, offering physically consistent explanations.

Cementing pumps are key equipment in oilfield well cementing operations. Diagnosing their running status is a crucial and challenging task. To handle this issue, this paper forwards a consistency-enhanced dual-weighted broad learning system (BLS) for effective cementing pump fault diagnosis. First, the vibration signals of cementing pump are analyzed to obtain the features of time and frequency domains. To make efficient use of the time series properties of signals, a phase space reconstruction technology is introduced to generate the augmented features. Second, from two distinct perspectives of variables and samples, a consistency-enhanced dual weighting strategy is designed to improve the basic BLS. The variable-dimension weights are constructed by the partial F test method to highlight the features contributing largely to the classification performance, whereas the sample-dimension weights are developed by the sample divergence analysis for emphasizing the key samples. To further suppress the noise influence and enhance the model's robustness, a consistency constraint is introduced to the BLS optimization objective. Third, to provide a reasonable explanation of the decision result, an error gradient-based input attribution method is proposed to indicate which features are closely related to the classification results. Finally, a real cementing pump condition monitoring experimental platform is constructed and an application on the experimental dataset is used to validate the method's performance. The application results demonstrate that the proposed method has a better fault classification accuracy than the basic BLS method and provides a reasonable explanation for the decision procedures.

Anomaly detection in Industrial Control Systems (ICS) is critical for ensuring operational safety. However, the limited interpretability of existing deep learning models significantly constrains their practical deployment. This paper introduces the Internal and External Dual-View Interpretation (IEDI), an interpretable anomaly detection approach form a dual internal-external perspective. Through comprehensive attribution analysis, IEDI jointly uncovers the model’s internal reasoning process and the influence of input features on anomaly decisions. Furthermore, the interpretability results are leveraged to iteratively refine the model, forming an explanation-feedback closed-loop. Experimental results on the SWaT, WADI, and KDDCUP99 benchmark datasets demonstrate that the proposed method enhances anomaly detection performance, while offering highly transparent and interpretable decision support for the secure and reliable operation of industrial systems.

When incipient fault data of chemical processes is insufficient, transfer learning enables fault degradation trend prediction and process safety improvement. Traditional echo state network (ESN)‐based transfer learning generally utilizes a single source domain, while traditional multi‐source domain selection methods do not consider information complementation between source domains. Therefore, a multi‐source transfer learning based on ESN (MTL‐ESN) is proposed to improve chemical process time series prediction performance. In order to consider information complementation between source domains, a similar‐different source domain selection strategy is designed. To comprehensively utilize the transfer knowledge of multi‐source domains, transfer ESN sub‐models are established for selected source domains, and integrated to an ensemble ESN prediction model. For faults, especially mixed fault operating conditions, the proposed MTL‐ESN can select source domains correctly and predict variable trend accurately. Simulations on a continuous stirred tank reactor demonstrate the effectiveness of the proposed time series prediction method.

Stochastic configuration broad learning system (SCBLS) has demonstrated notable advantages in industrial soft sensing due to its computational efficiency and supervised node configuration. However, conventional SCBLS overlooks the dynamic characteristics of industrial data, which limits the prediction accuracy of soft sensor models. To address this limitation, this study proposes a two‐layer temporal SCBLS (TT‐SCBLS) for industrial soft sensor modelling. The proposed framework consists of two key modules: an output‐related dual‐layer temporal feature extraction module and an enhancement node stochastic configuration module. The former is used to build the primary mapped features, where the first layer employs quality‐related slow feature analysis (QSFA) to extract intrinsic slow features, mitigating the impact of data noise, and the second layer utilizes a cycle reservoir with regular jumps (CRJ) network to capture temporal dependencies in the process data. The later introduces the stochastic configuration algorithm for incremental node expansion at the enhancement layer, ensuring efficient model adaptation. Additionally, a kernel Shapley additive explanation (SHAP) model is integrated to quantify the contributions of input variables, enhancing model transparency. The proposed method is validated on two benchmark industrial systems, with experimental results confirming its superior prediction performance compared to existing approaches.

Stochastic configuration network (SCN) has become a favorable soft sensor model due to its innovative random parameter construction approach under inequality constraint. However, it may suffer from the disadvantage of model overfitting with the increase of hidden nodes. To handle this issue, an improved Regularized SCN based on Spatio-Temporal nearest neighbors (ST-RSCN) is presented for better nonlinear soft sensor development. Different from regularized SCN only with L2 regularization, a dually-regularized SCN optimization framework is designed, where the L2 regularization term and manifold regularization (MR) term are applied to enforce the constraints from the perspectives of model parameters and data structure. Peculiarly, considering the data dynamic property, the traditional spatial nearest neighbor selection method is upgraded by integrating temporal searching strategy. Two spatio-temporal neighbor searching strategies are formed by designing different neighbor determination orders in spatial and temporal domains. The efficiency of the developed ST-RSCNs is finally demonstrated by two industrial cases, including a debutanizer column process and a continuous stirred tank reactor. The outcomes in both supervised and semi-supervised scenarios indicate that ST-RSCNs have better prediction performance compared with several models with respect to stability and generalization. Note to Practitioners-Data structure characteristic is of great importance for soft sensor model performance in practice. However, this is omitted in the existing SCN models. This paper designs a new enhanced regularized SCN for soft sensing. Specifically, a unified optimization objective is established involving both L2 regularization for constraining output weight magnitude and manifold regularization for mining the underlying geometric structure information. Further, two spatio-temporal nearest neighbor strategies are introduced to better fit the dynamic data structure. ST-RSCN related principles and proofs are described in detail. The simulation results of DCP and CSTR cases show the proposed ST-RSCNs significantly reduce the model prediction root mean squared errors (RMSEs) and are applicable for supervised and semi-supervised domains. Additionally, the key parameters' selection methods are discussed, which contributes to a sound understanding of this paper.

For the version 5 of you only look once (YOLOv5) based fire detection method., the confidence scores of weak fire regions are usually low., which can affect the fire detection rate. For the above problem., a novel weak fire detection method based on the YOLOv5 network and principal component analysis (YOLOv5-PCA) is proposed. Firstly., the YOLOv5 network is used to process real-time monitoring images and produce potential fire regions with confidence scores. If the confidence score of a potential region is medium., the principal component analysis is further applied to fuse the pixel color features of the potential region. A pixel monitoring statistic is built to detect weak fire pixels. In a potential region., if the proportion of fire pixels exceeds 1/3., it is considered that there is fire., and a fire alarm is activated. Simulation studies on 5049 images from 25 fire videos and 424 images from 2 non-fire videos are used to verify the effectiveness of the proposed method. Compared with the YOLOv5 based method., the proposed YOLOv5-PCA method can effectively increase fire detection rate by 49.06%., and correct many misidentified non-fire regions.

Fault detection has great significance for chemical process safety with the development of science and technology. The conventional echo state network-based fault detection method does not highlight key fault features, and cannot forecast future fault trend after the occurrence of faults. For the above problems, a chemical process fault detection and trend analysis strategy based on key feature enhanced echo state network (KESN) is proposed. First, dynamic features are extracted by a detecting echo state network. Then, a weighting strategy is designed to enhance key features and increase fault detection rates. After detecting a fault, independent component analysis is utilized to extract independent key features. Future fault trend is forecasted based on the forecasting multi-KESN. Simulation results on the Tennessee Eastman process demonstrate the effectiveness of the proposed method.

The soft sensor model based on deep neural networks (DNNs) can only extract strong nonlinear features from industrial data, and suffers from high training complexity and difficulties in determining the model structure. In order to deal with these issues, a composite deep partial least squares (CDPLSs) modeling method is proposed in this work to jointly extract both linear and nonlinear features hidden in process data. In this method, a hybrid latent variable learning framework is designed, where the linear partial least squares (PLSs) and nonlinear PLS are connected in a parallel way for capturing the linear and nonlinear latent features, respectively. Furthermore, for the efficient nonlinear feature mining in this framework, a multilayer random PLS model is constructed which applies a layerwise random Fourier feature (RFF) mapping to implement the deep nonlinear transformation. Finally, the superiority of this method is demonstrated through two industrial case studies.

Stochastic configuration networks (SCNs) are widely used in the field of soft sensor modeling due to their advantages of good generalization performance and automatic model structure determination. The classical SCN-based soft sensors are usually effective when industrial processes only involve a single operation mode. In practical applications, however, operation mode variations are often seen because of many factors, including market demands, raw material changes, ambient temperatures, etc. In the historical modes, abundant labeled samples are collected. In the new operation mode, the labeled samples are, however, very scarce and cannot sufficiently support the effective training of soft sensor models. How to make full use of the historical modes to assist the soft sensor modeling of the new mode is, therefore, a meaningful and challenging problem. To handle this problem, this article proposes a transfer learning soft sensor modeling method based on 2-D domain-adaption SCN (TD-DASCN). In this method, a domain adaption SCN modeling framework is designed for transfer learning soft sensor development by fusing the abundant labeled samples from historical modes (source domain) and a few labeled samples from new modes (target domain). The feature alignment procedure is performed by using geodesic flow kernel method to reduce data distribution difference between source and target domains. For the sake of avoiding the possible negative transfer phenomenon, the source domain loss function is constrained according to the degree of contribution of the source domain samples in the transfer. Last, the effectiveness of the proposed method is verified by two industrial cases. Compared with the basic DASCN soft sensor method, the proposed method can reduce the average prediction RMSE value by 30.0% and 9.1% in the two tested cases, respectively.

Tight carbonate reservoirs are widely distributed in the world and have great economic exploration potential. The Ordovician Yijianfang Formation grainstones in the Shunnan area, Tarim Basin are considered to be hydrocarbon reservoirs. However, there are few visible pores from thin sections and cores, resulting in poor understanding of the pore systems in these tight limestone reservoirs. Core samples collected from 8 wells were prepared to characterize pore networks by combining the optical microscope, scanning electron microscope, helium gas porosimeter, nuclear magnetic resonance and micro-CT analysis. Carbon and oxygen isotopes and fluid inclusion thermometry were adopted to determine the timing of pore-related dissolution. For Yijianfang grainstones, porosity mostly ranges from 0.2 to 3.9%; permeability mostly ranges from 0.01 to 1 mD. Petrographic observations revealed that both intragranular macropores and micropores within grains constitute effective reservoir space. Besides, micropores could account for 50–90% of the total porosity. The intragranular pores only appear in the upper part of the cycle, and the micropores developed in upper and middle of the cycle. There are no pores in the lower part of the cycle. During brief subaerial exposure shortly after deposition, meteoric dissolution created intragranular pores and vugs in grainstones in upper part of the sedimentary cycle. Meanwhile, influenced by the freshwater-involved mineralogical stabilization, micropores transformed from between micrites to between calcite microcrystals within grains in grainstones in the upper and middle part of the cycle. In the burial diagenetic environment, vugs and interparticle pores were destroyed by compaction and burial cements. For packstones in the lower part of the cycle, mineralogical stabilization occurred without freshwater during progressive burial, leading to tight fabric with no or minor micropores. In summary, whether tight limestones can be effective reservoirs depends on whether freshwater is involved in dissolution and mineralogical stabilization. In general, tight grainstone reservoirs of shallowing-upward cycle could provide abundant micropores with few macropores due to the freshwater-involved dissolution and stabilization during subaerial exposure.

High-quality reservoirs of the Feixianguan Formation of the northeast Sichuan Basin contain abundant natural gas and are dominated by dolostones. Their origin is controversial, which impedes the exploration of these high-quality carbonate reservoirs. Petrographic observation and geochemical analysis were performed on various carbonates to trace the source of dolomitizing fluids and clarify the formation mechanism of the Feixianguan dolostones. Three types of dolostones were distinguished, including fine dolomite, fine-medium dolomite, and medium dolomite according to the grain size of dolomites. Fine dolomite was characterized by crystal sizes of 5-60 mu m and high total rare earth elements (sigma REEs) of 1.87-2.45 ppm. The thin layers of fine dolomite mainly occur at the top of the depositional cycles and have the highest delta 18O values and the lowest SmSN/YbSN values of 0.65-0.87 of the three types, suggesting that they were formed by evaporation pumping dolomitization from the restricted evaporated seawater. Fine-medium dolomite was characterized by the inheritance of the original particle texture with crystal sizes of 60-120 mu m. Medium dolomite was characterized by crystal sizes of 100-200 mu m and the highest porosity among the three types of dolostones. The similar REE patterns to seawater and the same 87Sr/86Sr ratios as Feixianguan seawater for fine-medium and medium dolomites indicate that the dolo-mitizing fluids originated from coeval seawater with the Feixianguan limestone. Both the high concentrations of Na and low concentrations of Fe and Mn imply the shallow burial or near surface condition of dolomitization. According to higher delta 13C and delta 18O values than coeval limestones, and the descending tendency of delta 18O values and Na concentrations from the top to bottom of the sedimentary cycles, fine-medium and medium dolomites were formed by reflux dolomitization of evaporated Feixianguan seawater restricted by beach bridges. However, slightly brighter cathodoluminescence of medium dolomite than fine-medium dolomite and the decreasing of delta 18O values as the crystal size increasing imply that medium dolomites were slightly recrystallized from fine-medium dolomites. Moreover, a small number of coarse dolomites in pores are derived from exogenous non-seawater fluids in the burial environment.

Just-In-Time Learning (JITL) has proven to be an effective tool for developing virtual sensors for complex industrial processes. However, most of the existing JITL-based virtual sensing methods are supervised, whereas labeled samples within a local area are usually particularly scarce. As a result, the performance of these JITL-based virtual sensors are unable to reach the fullest potentials. To this end, a semisupervised JITL (SsJITL) framework is proposed based on local label propagation (LLP), which is referred to as “LLP-SsJITL. ” In the LLP-SsJITL, a robust semisupervised relevant neighborhood dataset of the query sample is first formed by accounting for both spatially and temporally relevant information of the query sample. Then, with the aid of the technique of label propagation on the constructed neighborhood of the query sample, a semisupervised objective function is developed, which transforms the estimation task into an analytical model-free optimization problem. By doing so, abundant unlabeled data are fully exploited such that the shortcomings of the supervised JITL resulting from the scarcity of labeled samples are remedied. Two real-life industrial processes are used to evaluate the performance of the LLP-SsJITL framework, showing significantly improved performance by using the LLP-SsJITL framework.

The formation mechanism of high-quality reservoirs is a key issue for ultra-deep (>6000 m) oil and gas exploration. The Lower Cambrian Xiaoerbrak dolomite with a burial depth of more than 7000 m in the Tazhong area of the Tarim Basin still keeps a maximum porosity of up to 20%, which is rare in the world. Based on the identification of meter-scale sedimentary cycles in cored intervals, the main controlling factors on the formation of high-quality reservoirs were discussed. The upper part of the sedimentary cycle is mainly composed of stromatolitic dolomite and oolitic dolomite. Comparatively, quartz-bearing dolomite and algal-bonded dolomite, as the product of low energy depositional environment are dominant in the lower part of each sedimentary cycle. Vugs formed by meteoric dissolution widely exist in the middle-upper part of the sedimentary cycle due to brief to moderate long-term subaerial exposure, whereas the lower part shows low porosity with a few intercrystalline pores. It indicates that subaerial exposure dissolution caused by periodic sea level decline is the main source of pores. The stylolite density and thickness of insoluble residue decrease rapidly from the lower to upper part of each sedimentary cycle. The content of aluminosilicate minerals in dolomite is positively related to the stylolite density, indicating that the aluminosilicate minerals play a promoting role in the pressure solution process. Pure dolomites with the least aluminosilicate minerals content in the upper of the sedimentary cycles have the strongest resistance to pressure solution, which is beneficial to ultra-deep high-quality reservoir preservation. Coupling of sedimentation and diagenesis to illustrate the origin of ultra-deep high-quality dolomite reservoirs is proposed.