Multilayer Module Research Articles

DMFR-YOLO: An Infrared Small Hotspot Detection Algorithm Based On Double Multi-scale Feature Fusion

Abstract Hotspot automatic detection is an effective strategy to realize intelligent maintenance of photovoltaic modules. However, it is challenging to detect small hotspots accurately using deep learning-based object detection methods due to the weak information and small area of the target. This study presents a double multi-scale feature reconstruction (DMFR)-YOLO to detect small hotspots in PV modules. In the proposed DMFR-YOLO, the backbone of YOLOv8n is improved by eliminating redundant deep layers and introducing a shallow detection head to enhance its perception and recognition ability for small objects. Moreover, a weighted multi-layer feature reconstruction module is introduced to fuse the features of different depths and a multi-receptive field reconstruction module is designed to fuse the information of different receptive fields. Finally, ablation and comparison experiments are conducted to evaluate the performance of the proposed method. The model with both MLFR and MRFR can achieve a recall rate of 88.6%, with a mean average precision (mAP@0.5) of 93.3%. Experimental results demonstrate that the proposed DMFR-YOLO is able to achieve small hotspot detection in IR images with high accuracy.

Perception and reconstruction of temperature field in forgings based on physical model and CNN model

Accurately perceiving and reconstructing the temperature field of hot-deformed large forgings is essential for intelligent forging and enhancing forging quality. In this study, the surface temperature of large forgings is measured using an infrared thermal imager. The relationship between the grayscale of infrared images and the surface temperature of large forgings is analyzed. A physical model, incorporating a convolutional neural network (CNN) model, is established to ascertain the surface temperature of large forgings. The CNN model utilizes standard convolution and multilayer perceptron modules for the backbone network and implements an adaptive learning rate. The correlation between three channel information and the surface temperature besides the grayscale is considered in the CNN model. Lastly, this study proposes a method for reconstructing the temperature field of large forgings based on the developed perception model, enabling accurate surface temperature measurement even when the large forgings are covered by oxide skin.

Spatial layout optimization model integrating layered attention mechanism in the development of smart tourism management

Tourism demand projection is paramount for both corporate operations and destination management, facilitating tourists in crafting bespoke, multifaceted itineraries and enriching their vacation experiences. This study proposes a multi-layer self attention mechanism recommendation algorithm based on dynamic spatial perception, with the aim of refining the analysis of tourists’ emotional inclinations and providing precise estimates of tourism demand. Initially, the model is constructed upon a foundation of multi-layer attention modules, enabling the semantic discovery of proximate entities to the focal scenic locale and employing attention layers to consolidate akin positions, epitomizing them through contiguous vectors. Subsequently, leveraging tourist preferences, the model forecasts the likelihood of analogous attractions as a cornerstone for the recommendation system. Furthermore, an attention mechanism is employed to refine the spatial layout, utilizing the forecasted passenger flow grid to infer tourism demand across multiple scenic locales in forthcoming periods. Ultimately, through scrutiny of data pertaining to renowned tourist destinations in Beijing, the model exhibits an average MAPE of 8.11%, markedly surpassing benchmarks set by alternative deep learning models, thereby underscoring its precision and efficacy. The spatial layout optimization methodology predicated on a multi-layer attention mechanism propounded herein confers substantive benefits to tourism demand prognostication and recommendation systems, promising to elevate the operational standards and customer contentment within the tourism sector.

Hyperspectral image classification with token fusion on GPU

Hyperspectral images capture material nuances with spectral data, vital for remote sensing. Transformer has become a mainstream approach for tackling the challenges posed by high-dimensional hyperspectral data with complex structures. However, a major challenge they face when processing hyperspectral images is the presence of a large number of redundant tokens, which leads to a significant increase in computational load, adding to the model’s computational burden and affecting inference speed. Therefore, we propose a token fusion algorithm tailored to the operational characteristics of the hyperspectral image and pure transformer network, aimed at enhancing the final accuracy and throughput of the model. The token fusion algorithm introduces a token merging step between the attention mechanism and the multi-layer perceptron module in each Transformer layer. Experiments on four hyperspectral image datasets demonstrate that our token fusion algorithm can significantly improve inference speed without any training, while only causing a slight decrease in the pure transformer network’s classification accuracy.

DTA-GTOmega: Enhancing Drug-Target Binding Affinity Prediction with Graph Transformers Using OmegaFold Protein Structures

Understanding drug-protein interactions is crucial for elucidating drug mechanisms and optimizing drug development. However, existing methods have limitations in representing the three-dimensional structure of targets and capturing the complex relationships between drugs and targets. This study proposes a new method, DTA-GTOmega, for predicting drug-target binding affinity. DTA-GTOmega utilizes OmegaFold to predict protein three-dimensional structure and construct target graphs, while processing drug SMILES sequences with RDKit to generate drug graphs. By employing multi-layer graph transformer modules and co-attention modules, this method effectively integrates atomic-level features of drugs and residue-level features of targets, accurately modeling the complex interactions between drugs and targets, thereby significantly improving the accuracy of binding affinity predictions. Our method outperforms existing techniques on benchmark datasets such as KIBA, Davis, and BindingDB_Kd under cold-start setting. Moreover, DTA-GTOmega demonstrates competitive performance in real-world DTI scenarios involving DrugBank data and drug-target interactions related to cardiovascular and nervous system-related diseases, highlighting its robust generalization capabilities. Additionally, the introduced DTI evaluation metrics further validate DTA-GTOmega’s potential in handling imbalanced data.

Pre-training transformer with dual-branch context content module for table detection in document images

BackgroundDocument images such as statistical reports and scientific journals are widely used in information technology. Accurate detection of table areas in document images is an essential prerequisite for tasks such as information extraction. However, because of the diversity in the shapes and sizes of tables, existing table detection methods adapted from general object detection algorithms, have not yet achieved satisfactory results. Incorrect detection results might lead to the loss of critical information. MethodsTherefore, we propose a novel end-to-end trainable deep network combined with a self-supervised pretraining transformer for feature extraction to minimize incorrect detections. To better deal with table areas of different shapes and sizes, we added a dual-branch context content attention module (DCCAM) to high-dimensional features to extract context content information, thereby enhancing the network's ability to learn shape features. For feature fusion at different scales, we replaced the original 3×3 convolution with a multilayer residual module, which contains enhanced gradient flow information to improve the feature representation and extraction capability. ResultsWe evaluated our method on public document datasets and compared it with previous methods, which achieved state-of-the-art results in terms of evaluation metrics such as recall and F1-score. https://github.com/YongZ-Lee/TD-DCCAM

Dynamic Reorganization Patterns of Brain Modules after Stroke Reflecting Motor Function.

Advancements in neuroimaging technologies have significantly deepened our understanding of the neural physiopathology associated with stroke. Nevertheless, the majority of studies ignored the characteristics of dynamic changes in brain networks. The relationship between dynamic changes in brain networks and the severity of motor dysfunction after stroke needs further investigation. From the perspective of multilayer network module reconstruction, we aimed to explore the dynamic reorganization of the brain and its relationship with motor function in subcortical stroke patients. We recruited 35 healthy individuals and 50 stroke patients with unilateral limb motor dysfunction (further divided into mild-moderate group and severe group). Using dynamic multilayer network modularity analysis, we investigated changes in the dynamic modular reconfiguration of brain networks. Additionally, we assessed longitudinal clinical scale changes in stroke patients. Correlation and regression analyses were employed to explore the relationship between characteristic dynamic indicators and impairment and recovery of motor function, respectively. We observed increased temporal flexibility in the Default Mode Network (DMN) and decreased recruitment of module reconfiguration in the Attention Network (AN), Sensorimotor Network (SMN), and DMN after stroke. We also observed reduced module loyalty following stroke. Additionally, correlation analysis showed that hyper-flexibility of the DMN was associated with better lower limb motor function performance in stroke patients with mild-to-moderate impairment. Regression analysis indicated that increased flexibility within the DMN and decreased recruitment coefficient within the AN may predict good lower limb function prognosis in patients with mild to moderate motor impairment. Our study revealed more frequent modular reconfiguration and hyperactive interaction of brain networks after stroke. Notably, dynamic modular remodeling was closely related to the impairment and recovery of motor function. Understanding the temporal module reconfiguration patterns in multilayer networks after stroke can provide valuable information for more targeted treatments.

Meta-collaboration-based semantic contrast for inductive knowledge representation learning

Inductive knowledge representation learning aims at effectively representing new entities in emerging knowledge graphs based on current ones, providing positive support for reasoning facts with newly emerging entities. It requires a scheme that can derive the instant knowledge to generate rational representations for new entities. Recently, some approaches have been developed for it by capturing logical rules between entities, mining structural information of KGs, or learning multiple structural patterns of entities. However, there are two main challenges that they have yet to overcome. The first is the structural bias-induced semantic ambiguity, where new entities are typically represented in a manner akin to those with analogous structures, during which the identical structures are emphasized while the differences are overlooked. We refer to this as the structural bias, which will lead to the generated representations of new entities being semantically ambiguous. The second is the potential sparsity of new entities. The presence of new entities with low frequencies brings about the insufficient representations for these new entities even training a model from scratch with substantial costs. To this end, we propose a Meta-Collaboration-based Semantic Contrast (MCSC) framework with the aim of addressing both issues above. Following the current meta-task formulation scheme, we first acquire the relation-specific knowledge to produce entity representations through modeling relations as feedback. Then, a multi-layer graph neural network module is developed to aggregate sufficient neighbor information to update entity representations. In the final, a collaborative semantic contrast module is designed on both support and query sets to respectively overcome the structural bias-induced semantic ambiguity and the potential sparsity of new entities challenges. Such a meta-collaboration-based semantic contrast procedure transcends current meta-learning paradigms in inductive knowledge representation learning that fail to learn semantically discriminative entity representations. Experimental results of the inductive link prediction task on 12 benchmark datasets demonstrate the superiority of MCSC over SOTA baselines.

Multi-feature fusion dehazing based on CycleGAN

Under the foggy environment, lane line images are obscured by haze, which leads to lower detection accuracy, higher false detection of lane lines. To address the above problems, a multi-layer feature fusion dehazing network based on CycleGAN architecture is proposed. Firstly, the foggy image is enhanced to remove the fog in the image, and then the lane line detection network is used for detection. For the dehazing network, a multi-layer feature fusion module is used in the generator to fuse the features of different coding layers of U-Net to enhance the network’s recovery of information such as details and edges, and a frequency domain channel attention mechanism is added at the key nodes of the network to enhance the network’s attention to different fog concentrations. At the same time, to improve the discriminant effect of the discriminator, the discriminator is extended to a global and local discriminator. The experimental results show that the dehaze effect on Reside and other test data sets is better than the comparison method. The peak signal-to-noise ratio is improved by 2.26 dB compared to the highest GCA-Net algorithm. According to the lane detection of fog images, it is found that the proposed network improves the accuracy of lane detection on foggy days.

First results of the polar regional climate model RACMO2.4

Abstract. The next version of the polar Regional Atmospheric Climate Model (referred to as RACMO2.4p1) is presented in this study. The principal update includes embedding of the package of physical parameterizations of the Integrated Forecast System (IFS) cycle 47r1. This constitutes changes in the precipitation, convection, turbulence, aerosol and surface schemes and includes a new cloud scheme with more prognostic variables and a dedicated lake model. Furthermore, the standalone IFS radiation physics module ecRad is incorporated into RACMO, and a multilayer snow module for non-glaciated regions is introduced. Other updates involve the introduction of a fractional land–ice mask, new and updated climatological data sets (such as aerosol concentrations and leaf area index), and the revision of several parameterizations specific to glaciated regions. As a proof of concept, we show first results for Greenland, Antarctica and a region encompassing the Arctic. By comparing the results with observations and the output from the previous model version (RACMO2.3p3), we show that the model performs well regarding the surface mass balance, surface energy balance, temperature, wind speed, cloud content and snow depth. The advection of snow hydrometeors strongly impacts the ice sheet's local surface mass balance, particularly in high-accumulation regions such as southeast Greenland and the Antarctic Peninsula. We critically assess the model output and identify some processes that would benefit from further model development.

DAT-Net: Deep Aggregation Transformer Network for automatic nuclear segmentation

Detection and segmentation of cell nuclei in hematoxylin and eosin-stained tissue images pose critical clinical challenges, including intricate backgrounds, variable nuclear appearances, overlapping nuclei, and unclear boundaries. These complexities render automated instance segmentation a difficult research area. Thus, this paper proposes a deep aggregation transformer network for automatic nucleus instance segmentation. The method employs a robust Vision Transformer-based backbone to extract informative features from images. To fuse multi-level information, we propose the group aggregation multi-level feature fusion module, which merges low-level features for shape representation and high-level features for pixel edge segmentation, enhancing overall model performance. The depthwise integrated multilayer perceptron module enhances feature learning by combining the global context-dependent representation of MLP and the local representation of convolution through channel separation. This fosters feature learning across different modes, thereby improving network representation. The bidirectional path interaction feature processing module is introduced for multi-scale information extraction and fusion. This module enhances the entire feature pyramid hierarchy by transferring deep semantic information to shallow layers with robust spatial information via bidirectional paths. The Large kernel Attention module at each layer extensively improves feature extraction through a large receptive field. Our approach outperforms state-of-the-art models, including StarDIST, CPP-Net, TSFD-Net and TransNuSeg, by 7.8%, 6.3%, 1.5% and 3.5% in average panoptic quality for different cell nuclei segmentation categories across tissues of the Pannuke dataset. Furthermore, DAT-Net outperforms all other state-of-the-art models in terms of the F1 score for the five classes of cell nuclei in the PanNuke dataset.

Thermal flow and thermoelectricity characteristics in a sandwich flat plate thermoelectric power generation device under diesel engine exhaust conditions

Thermoelectric power generation (TEG) technology can be used to recover exhaust waste heat, but its effective application is restricted by bulky size and low thermoelectric conversion efficiency. The compactness evaluation of thermoelectric devices characterized by the power density is lacking. In this paper, a three-dimensional thermal flow and thermoelectricity multi-physics field model of a sandwich flat plate TEG device with multi-layer thermoelectric modules (TEMs) was established. An engine thermoelectric recovery bench test was carried out to verify the model and present the thermal parameters and output performance. The thermal, flow and electrical field distributions of the TEG device, as well as voltage, power and conversion efficiency with engine loads, and power density, were presented. The results show that the hot side temperatures, voltages and powers for each thermoelectric module (TEM) are greatly affected by the position occupied by the TEMs and engine loads. At full load and the coolant flow of 8.5 L/min, the device achieves the peak values of the power of 80.9 W and thermoelectric efficiency of 2.1 %, respectively. The device shows good compactness with the power density of 19.8 kW/m3, attributed to the reasonable arrangement of the collectors, coolers, and TEMs. This work can provide some insights of the relationship between the structural form and size, thermal parameters, electrical parameters and conversion efficiency of the sandwich type TEG device.

DRML-Ensemble: drug repurposing method based on feature construction of multi-layer ensemble.

Computational drug repurposing methods have been continuously developed in recent years to alleviate the high costs associated with drug development. As drug targets or the products of disease-related genes, proteins play an important role in drug repurposing. Although the potential has been demonstrated, heterogeneous graphs with proteins as independent nodes have yet to be studied, where extracting high-quality protein features from heterogeneous graphs poses a significant challenge. A novel drug repurposing model based on the feature construction of multi-layer ensemble (DRML-Ensemble) is proposed in this study. The performance of DRML-Ensemble, as evaluated on publicly available datasets, achieves an AUPR value of 0.93 and an AUROC value of 0.92, surpassing those of existing state-of-the-art methods. Additionally, DRML-Ensemble demonstrates its notable ability for drug repurposing in Alzheimer's disease. DRML-Ensemble is primarily composed of multiple layers of heterogeneous graph feature construction (HGFC). Each HGFC can extract protein features by leveraging the relationships between drugs, diseases, and proteins. These protein features are then utilized in subsequent layers to build drug and disease features, facilitating drug repurposing. By stacking multiple layers, optimal protein features can be obtained from the heterogeneous graph, consequently improving the accuracy of drug repurposing. However, an excessive· stacking of layers usually affect the model's training process, for example, causing problems such as overfitting; a multi-layer ensemble prediction module is designed to further improve the model's performance.

SQI-DOANet: electroencephalogram-based deep neural network for estimating signal quality index and depth of anaesthesia.


Monitoring the depth of anaesthesia (DOA) during surgery is of critical importance. However, accurate and real-time estimation of DOA remains a challenging task. In this paper, we proposed a signal quality index (SQI) network (SQINet) for assessing the electroencephalogram (EEG) signal quality and a DOA network (DOANet) for analyzing EEG signals to precisely estimate DOA. The two networks are termed SQI-DOANet.
Approach:
The SQINet contained a shallow convolutional neural network to quickly determine the quality of the EEG signal. The DOANet comprised a feature extraction module for extracting features, a dual attention module for fusing multi-channel and multi-scale information, and a gated multilayer perceptron module for extracting temporal information. The performance of the SQI-DOANet model was validated by training and testing the model on the large VitalDB database, with the bispectral index (BIS) as the reference standard.
Main results:
The proposed DOANet yielded a Pearson correlation coefficient with the BIS score of 0.88 in the 5-fold cross-validation, with a mean absolute error (MAE) of 4.81. The mean Pearson correlation coefficient of SQI-DOANet with the BIS score in the 5-fold cross-validation was 0.82, with an MAE of 5.66.
Significance:
The SQI-DOANet model outperformed three compared methods. The proposed SQI-DOANet may be used as a new deep learning method for DOA estimation. The code of the SQI-DOANet will be made available publicly at https://github.com/YuRui8879/SQI-DOANet.&#xD.

ALdamage-seg: A Lightweight Model for Instance Segmentation of Aluminum Profiles

Aluminum profiles are widely used in various manufacturing sectors due to their flexibility and chemical properties. However, these profiles are susceptible to defects during manufacturing and transportation. Detecting these defects is crucial, but existing object detection models like Mask R-CNN and YOLOv8-seg are not optimized for this task. These models are large and computationally intensive, making them unsuitable for edge devices used in industrial inspections. To address this issue, this study proposes a novel lightweight instance segmentation model called AL-damage-seg, inspired by the YOLOv8n-seg architecture. This model utilizes MobileNetV3 as the backbone. In YOLOv8n-seg, the role of C2f is to enhance the nonlinear representation of the model to capture complex image features more efficiently. We upgraded and improved it to form multilayer feature extraction module (MFEM) and integrates a large separable kernel attention (LSKA) mechanism in the C2f module, resulting in C2f_LSKA, to further optimize the performance of the model. Additionally, depth-wise separable convolutions are employed in the feature fusion process. The ALdamage-seg’s weight on the Alibaba Tian-chi aluminum profile dataset constitutes 43.9% of that of YOLOv8n-seg, with its GFLOPs reduced to 53% relative to YOLOv8-seg, all the while achieving an average precision (mAP) of 99% relative to YOLOv8-seg. With its compact size and lower computational requirements, this model is well-suited for deployment on edge devices with limited processing capabilities.

Whisper-SV: Adapting Whisper for low-data-resource speaker verification

Trained on 680,000 h of massive speech data, Whisper is a multitasking, multilingual speech foundation model demonstrating superior performance in automatic speech recognition, translation, and language identification. However, its applicability in speaker verification (SV) tasks remains unexplored, particularly in low-data-resource scenarios where labeled speaker data in specific domains are limited. To fill this gap, we propose a lightweight adaptor framework to boost SV with Whisper, namely Whisper-SV. Given that Whisper is not specifically optimized for SV tasks, we introduce a representation selection module to quantify the speaker-specific characteristics contained in each layer of Whisper and select the top-k layers with prominent discriminative speaker features. To aggregate pivotal speaker-related features while diminishing non-speaker redundancies across the selected top-k distinct layers of Whisper, we design a multi-layer aggregation module in Whisper-SV to integrate multi-layer representations into a singular, compacted representation for SV. In the multi-layer aggregation module, we employ convolutional layers with shortcut connections among different layers to refine speaker characteristics derived from multi-layer representations from Whisper. In addition, an attention aggregation layer is used to reduce non-speaker interference and amplify speaker-specific cues for SV tasks. Finally, a simple classification module is used for speaker classification. Experiments on VoxCeleb1, FFSVC, and IMSV datasets demonstrate that Whisper-SV achieves EER/minDCF of 2.22%/0.307, 6.14%/0.488, and 7.50%/0.582, respectively, showing superior performance in low-data-resource SV scenarios.

Multi-Output Prediction Model for Basic Oxygen Furnace Steelmaking Based on the Fusion of Deep Convolution and Attention Mechanisms

The objective of basic oxygen furnace (BOF) steelmaking is to achieve molten steel with final carbon content, temperature, and phosphorus content meeting the requirements. Accurate prediction of the above properties is crucial for end-point control in BOF steelmaking. Traditional prediction models typically use multi-variable input and single-variable output approaches, neglecting the coupling relationships between different property indicators, making it difficult to predict multiple outputs simultaneously. Consequently, a multi-output prediction model based on the fusion of deep convolution and attention mechanism networks (FDCAN) is proposed. The model inputs include scalar data, such as the properties of raw materials and target molten steel, and time series data, such as lance height, oxygen supply intensity, and bottom air supply intensity during the blowing process. The FDCAN model utilizes a fully connected module to extract nonlinear features from scalar data and a deep convolution module to process time series data, capturing high-dimensional feature representations. The attention mechanism then assigns greater weight to significant features. Finally, multiple multi-layer perceptron modules predict the outputs—final carbon content, temperature, and phosphorus content. This structure allows FDCAN to learn complex relationships within the input data and between input and output variables. The effectiveness of the FDCAN model is validated using actual BOF steelmaking data, achieving hit rates of 95.14% for final carbon content within ±0.015 wt%, 84.72% for final temperature within ±15 °C, and 88.89% for final phosphorus content within ±0.005 wt%.

Medical image classification: Knowledge transfer via residual U-Net and vision transformer-based teacher-student model with knowledge distillation

With the widespread integration of deep learning techniques in the domain of medical image analysis, there is a prevailing consensus regarding their efficacy in handling high-dimensional and intricate medical image data. However, it is imperative to acknowledge that while complex deep models exhibit a remarkable capacity for processing high-dimensional and intricate data, they often necessitate a substantial allocation of computational resources and time. Furthermore, lightweight models, despite their computational efficiency, tend to underperform when compared to their more intricate counterparts in terms of performance. Hence, the prevailing aspiration is to transfer the cognitive prowess of complex models to their lightweight counterparts. Addressing the aforementioned concern, this study proposes a knowledge distillation approach that encompasses joint feature and soft label transfer. It entails the transference of knowledge from the teacher model’s intermediate features and predictive outcomes to the student model. The student model leverages this knowledge to emulate the behavior of the teacher model, thereby enhancing the precision of its own predictions. Building upon this foundation, we introduce a Res-Transformer teacher model based on the U-Net architecture and a student model known as ResU-Net, which is grounded in residual modules. The Res-Transformer teacher model employs multi-layer residual attention during the downsampling process to capture deep-level features of the image. Subsequently, we have incorporated a Multi-layer Perceptual Attention module (MPA) for each skip connection layer, facilitating the integration of hierarchical upsampled information to restore fine-grained details within the feature maps. The ResU-Net student model enhances network stability through the utilization of residual modules and optimizes skip connections to recover any lost image information during convolutional operations. Lastly, we conducted experimental assessments on multiple disease datasets. The results reveal that the ACC of the Res-Transformer model achieves an impressive 96.9%. Furthermore, through the knowledge distillation method, rich knowledge is effectively transferred to the ResU-Net model, resulting in a remarkable ACC improvement of 7.2%.

Sintering behaviors, microstructure and properties of CaO–B2O3–SiO2 glass-ceramics for LTCC applications

The aim of this work was to investigate the impacts of MgO on the crystalline behaviors, microstructure, and properties of CaO–B2O3–SiO2 (CBS) glass-ceramics. In the study, the CBS samples with various MgO contents were developed via the melting and quenching route. Subsequently, a CBS casting tape and a multi-layered LTCC microcircuit module were created by using the tape casting method. Results indicated that MgO could break the glass network and promote the formation of liquid phase during the sintering process. A suitable amount of MgO additive promoted densification during firing process and improved the properties of the CBS materials. The CBS sample with 1.5 wt% MgO fired at 875 °C exhibited outstanding properties, including a dielectric constant (εr) of 6.5 (at 10 GHz), a dielectric loss (tan δ) of 8 × 10−4, a flexural strength of 167 MPa, a coefficient of thermal expansion (CTE) of 6.8 ppm/°C, and a thermal conductivity (λ) of 2.2 W/mK. In addition, this LTCC material demonstrated good co-firing compatibility with silver electrodes and superior temperature stability, showing its great potential for microelectronic applications.

Anomaly detection of wind turbine based on norm-linear-ConvNeXt-TCN

The supervisory control and data acquisition (SCADA) system of wind turbines continuously collects a large amount of monitoring data during their operation. These data contain abundant information about the operating status of the turbine components. Utilizing this information makes it feasible to provide early warnings and predict the health status of the wind turbine. However, due to the strong coupling between the various components of the wind turbine, the data exhibits complex spatiotemporal relationships, multiple state parameters, strong non-linearity, and noise interference, which brings great difficulty to anomaly detection of the wind turbine. This paper proposes a new method for detecting abnormal operating conditions of wind turbines, based on a cleverly designed multi-layer linear residual module and the improved temporal convolutional network (TCN) with a new norm-linear-ConvNeXt architecture (NLC-TCN). Initially, the NLC-TCN deep learning reconstruction model is trained with historical data of normal behavior to extract the spatiotemporal features of state parameters under normal operational conditions. Subsequently, the condition score of the unit is determined by calculating the average normalized root mean square error between the reconstructed data and actual data. The streaming peaks-over-threshold real-time calculation of the anomaly warning threshold, based on extreme value theory, is then used for preliminary fault monitoring. Moreover, by shielding the fault alarm for low wind speeds and implementing a continuous delay perception mechanism, issues related to wind speed fluctuations and internal and external interference are addressed, enabling early warning for faulty units. Finally, the effectiveness and reliability of the proposed method are validated through comparative experiments using actual offshore wind farm SCADA data. The performance of the proposed method surpasses that of other compared methods. Additionally, the results of the proposed method were evaluated using the uniform manifold approximation and projection dimensionality reduction technique and kernel density estimation.