Traditional Deep Learning Methods Research Articles

End-to-end multiple object tracking in high-resolution optical sensors of drones with transformer models

Drone aerial imaging has become increasingly important across numerous fields as drone optical sensor technology continues to advance. One critical challenge in this domain is achieving both accurate and efficient multi-object tracking. Traditional deep learning methods often separate object identification from tracking, leading to increased complexity and potential performance degradation. Conventional approaches rely heavily on manual feature engineering and intricate algorithms, which can further limit efficiency. To overcome these limitations, we propose a novel Transformer-based end-to-end multi-object tracking framework. This innovative method leverages self-attention mechanisms to capture complex inter-object relationships, seamlessly integrating object detection and tracking into a unified process. By utilizing end-to-end training, our approach simplifies the tracking pipeline, leading to significant performance improvements. A key innovation in our system is the introduction of a trajectory detection label matching technique. This technique assigns labels based on a comprehensive assessment of object appearance, spatial characteristics, and Gaussian features, ensuring more precise and logical label assignments. Additionally, we incorporate cross-frame self-attention mechanisms to extract long-term object properties, providing robust information for stable and consistent tracking. We further enhance tracking performance through a newly developed self-characteristics module, which extracts semantic features from trajectory information across both current and previous frames. This module ensures that the long-term interaction modules maintain semantic consistency, allowing for more accurate and continuous tracking over time. The refined data and stored trajectories are then used as input for subsequent frame processing, creating a feedback loop that sustains tracking accuracy. Extensive experiments conducted on the VisDrone and UAVDT datasets demonstrate the superior performance of our approach in drone-based multi-object tracking.

Military Equipment Entity Extraction Based on Large Language Model

The technology of military equipment entity extraction, a crucial component in constructing military knowledge bases, holds significant research value and theoretical importance for guiding the development and improvement of equipment support forces. In the military domain, equipment entities exhibit a phenomenon of nesting, where one entity is contained within another, and abbreviations or codes are frequently used to represent these entities. To address this complexity, this paper proposes a method named CoTNER for extracting entities. Initially, a large-scale language model is used to perform data augmentation with chain-of-thought on the original dataset, providing additional semantic and contextual information. Subsequently, the augmented dataset is fine-tuned on a small-scale language model to adapt it to the task of military equipment entity extraction and to enhance its ability to learn complex rules specific to the domain of military equipment. Additionally, a high-quality data filtering strategy based on instruction-following difficulty scoring is proposed to address the catastrophic forgetting issue that may occur during the fine-tuning of large language models. The experimental results demonstrate that the proposed military equipment entity extraction method outperforms mainstream traditional deep learning methods, validating the effectiveness of CoTNER.

Joint Image Processing with Learning-Driven Data Representation and Model Behavior for Non-Intrusive Anemia Diagnosis in Pediatric Patients.

Anemia diagnosis is crucial for pediatric patients due to its impact on growth and development. Traditional methods, like blood tests, are effective but pose challenges, such as discomfort, infection risk, and frequent monitoring difficulties, underscoring the need for non-intrusive diagnostic methods. In light of this, this study proposes a novel method that combines image processing with learning-driven data representation and model behavior for non-intrusive anemia diagnosis in pediatric patients. The contributions of this study are threefold. First, it uses an image-processing pipeline to extract 181 features from 13 categories, with a feature-selection process identifying the most crucial data for learning. Second, a deep multilayered network based on long short-term memory (LSTM) is utilized to train a model for classifying images into anemic and non-anemic cases, where hyperparameters are optimized using Bayesian approaches. Third, the trained LSTM model is integrated as a layer into a learning model developed based on recurrent expansion rules, forming a part of a new deep network called a recurrent expansion network (RexNet). RexNet is designed to learn data representations akin to traditional deep-learning methods while also understanding the interaction between dependent and independent variables. The proposed approach is applied to three public datasets, namely conjunctival eye images, palmar images, and fingernail images of children aged up to 6 years. RexNet achieves an overall evaluation of 99.83 ± 0.02% across all classification metrics, demonstrating significant improvements in diagnostic results and generalization compared to LSTM networks and existing methods. This highlights RexNet's potential as a promising alternative to traditional blood-based methods for non-intrusive anemia diagnosis.

A coordination attention residual U-Net model for enhanced short and mid-term sea surface temperature prediction

Sea surface temperature (SST) is crucial for studying global oceans and evaluating ecosystems. Accurately predicting short and mid-term daily SST has been a significant challenge in oceanography. Traditional deep learning methods can handle temporal data and spatial features but often struggle with long-range spatiotemporal dependencies. To address this, we propose a coordination attention residual U-Net(CResU-Net) model designed to better capture the dynamic spatiotemporal correlations of high-resolution SST. The model integrates coordinate attention mechanisms, multiple residual modules, and depthwise separable convolutions to enhance prediction capabilities. The spatiotemporal variations of SST across different areas of the South China Sea are complex, making accurate predictions challenging. Experiments across various regions of the South China Sea show the model’s effectiveness and robust generalization in predicting high-resolution daily SST. For a 10-day forecast period, the model achieves approximately 0.3 °C in RMSE, outperforming several advanced models.

A novel graph convolutional network-based interpretable method for chiller energy consumption prediction considering the spatiotemporal coupling between variables

The prediction of chiller energy consumption is critical for lowering the building’s energy consumption. Deep neural networks have been widely applied in this field. However, traditional deep learning methods only consider operational data and do not incorporate empirical knowledge and working condition information that are beneficial for chiller energy consumption prediction. The graph convolutional network (GCN) can handle this problem well, as it is able to incorporate empirical knowledge and working condition information into the network. However, it is important to avoid the excessive influence of empirical knowledge and working condition information on the model’s accuracy. To adjust the weight of the influence of the association graph and training set on model accuracy in the GCN network, a coefficient θ was introduced. Herein, a novel GCN (θ-GCN)-based method for the energy consumption prediction of chillers is proposed. This proposed method first combines empirical knowledge and working condition information to obtain the adjacency matrix among operational data and then constructs an association graph based on this adjacency matrix. Secondly, due to the temporal correlation of the collected dataset, a gated recurrent unit (GRU) is used for feature extraction. Lastly, the features extracted through GRU are input into the constructed association graph, and the θ-GCN network is used for training and energy consumption prediction. The proposed method was experimentally verified in an actual chiller system and compared with existing methods; the outcomes indicated that the established method could obtain better prediction accuracy.

FDADNet: Detection of Surface Defects in Wood-Based Panels Based on Frequency Domain Transformation and Adaptive Dynamic Downsampling

The detection of surface defects on wood-based panels plays a crucial role in product quality control. However, due to the complex background and low contrast of defects in wood-based panel images, features extracted by traditional deep learning methods based on spatial domain processing often contain noise and blurred boundaries, which severely affects detection performance. To address these issues, we have proposed a wood-based panel surface defect detection method based on frequency domain transformation and adaptive dynamic downsampling (FDADNet). Specifically, we designed a Multi-axis Frequency Domain Weighted Information Representation Module (MFDW), which effectively decoupled the indistinguishable low-contrast defects from the background in the transform domain. Gaussian filtering was then employed to eliminate noise and blur between the defects and the background. Additionally, to tackle the issue of scale differences in defects that led to difficulties in accurate capture, we designed an Adaptive Dynamic Convolution (ADConv) module for downsampling. This method flexibly compressed and enhanced features, effectively improving the differentiation of the features of objects of varying scales in the transform space, and ultimately achieved effective defect detection. To compensate for the lack of data, we constructed a dataset of wood-based panel surface defects, WBP-DET. The experimental results showed that the proposed FDADNet effectively improved the detection performance of wood-based panel surface defects in complex scenarios, achieving a solid balance between efficiency and accuracy.

Analysis of pedestrian second crossing behavior based on physics-informed neural networks

Pedestrian two-stage crossings are common at large, busy signalized intersections with long crosswalks and high traffic volumes. This design aims to address pedestrian operation and safety by allowing navigation in two stages, negotiating each traffic direction separately. Understanding crosswalk behavior, especially during bidirectional interactions, is essential. This paper presents a two-stage pedestrian crossing model based on Physics-Informed Neural Networks (PINNs), incorporating fluid dynamics equations to determine characteristics such as speed, density, acceleration, and Reynolds number during crossings. The study shows that PINNs outperform traditional deep learning methods in calculating and predicting pedestrian fluid properties, achieving a mean squared error as low as 10–8. The model effectively captures dynamic pedestrian flow characteristics and provides insights into pedestrian behavior impacts. The results are significant for designing pedestrian facilities to ensure comfort and optimizing signal timing to enhance mobility and safety. Additionally, these findings can aid autonomous vehicles in better understanding pedestrian intentions in intelligent transportation systems.

HyPhAICC v1.0: a hybrid physics–AI approach for probability fields advection shown through an application to cloud cover nowcasting

Abstract. This work proposes a hybrid approach that combines physics and artificial intelligence (AI) for cloud cover nowcasting. It addresses the limitations of traditional deep-learning methods in producing realistic and physically consistent results that can generalise to unseen data. The proposed approach, named HyPhAICC, enforces a physical behaviour. In the first model, denoted as HyPhAICC-1, a multi-level advection dynamics is considered a hard constraint for a trained U-Net model. Our experiments show that the hybrid formulation outperforms not only traditional deep-learning methods but also the EUMETSAT Extrapolated Imagery model (EXIM) in terms of both qualitative and quantitative results. In particular, we illustrate that the hybrid model preserves more details and achieves higher scores based on similarity metrics in comparison to U-Net. Remarkably, these improvements are achieved while using only one-third of the data required by the other models. Another model, denoted as HyPhAICC-2, adds a source term to the advection equation, it impaired the visual rendering but displayed the best performance in terms of accuracy. These results suggest that the proposed hybrid physics–AI architecture provides a promising solution to overcome the limitations of classical AI methods and contributes to open up new possibilities for combining physical knowledge with deep-learning models.

State-of-health estimation for lithium-ion batteries based on Kullback–Leibler divergence and a retentive network

Accurate state-of-health (SOH) estimation is crucial for the lithium-ion battery industry, as it underpins the safety, durability, and reliability of lithium-ion batteries. Currently, most researchers use various methods of health indicator (HI) extraction for the SOH estimation of batteries. However, these methods may require certain expertise and prior knowledge to achieve accurate modeling, being affected by measurement noise and other factors. To solve the abovementioned problems, three Kullback–Leibler (KL) divergence features based on partial voltage sequences are proposed as new HIs that are independent of prior knowledge and strongly correlated with SOH. Moreover, a modified retentive network is proposed to enhance SOH estimation accuracy and better utilize HIs than traditional deep learning methods, which have high training costs and insufficient accuracy. To ensure consistent extraction of KL divergence features across various experimental conditions and time intervals, a B-spline algorithm is utilized for interpolation. The effectiveness of the proposed method is validated through analysis of Pearson correlation coefficients and experiments conducted in four dimensions. Additionally, the potential of using the proposed method to compress data on the cloud-side is explored.

A Self-Attention Legendre Graph Convolution Network for Rotating Machinery Fault Diagnosis.

Rotating machinery is widely used in modern industrial systems, and its health status can directly impact the operation of the entire system. Timely and accurate diagnosis of rotating machinery faults is crucial for ensuring production safety, reducing economic losses, and improving efficiency. Traditional deep learning methods can only extract features from the vertices of the input data, thereby overlooking the information contained in the relationships between vertices. This paper proposes a Legendre graph convolutional network (LGCN) integrated with a self-attention graph pooling method, which is applied to fault diagnosis of rotating machinery. The SA-LGCN model converts vibration signals from Euclidean space into graph signals in non-Euclidean space, employing a fast local spectral filter based on Legendre polynomials and a self-attention graph pooling method, significantly improving the model's stability and computational efficiency. By applying the proposed method to 10 different planetary gearbox fault tasks, we verify that it offers significant advantages in fault diagnosis accuracy and load adaptability under various working conditions.

Predicting Energy Consumption for Hybrid Energy Systems toward Sustainable Manufacturing: A Physics-Informed Approach Using Pi-MMoE

Hybrid energy supply systems are widely utilized in modern manufacturing processes, where accurately predicting energy consumption is essential not only for managing productivity but also for driving sustainable development. Effective energy management is a cornerstone of sustainable manufacturing, reducing waste and enhancing efficiency. However, conventional studies often focus solely on predicting single types of energy consumption and overlook the integration of physical laws and information, which are essential for a comprehensive understanding of energy dynamics. In this context, this paper introduces a multi-task physics-informed multi-gate mixture-of-experts (pi-MMoE) model that not only considers multiple forms of energy consumption but also incorporates physical principles through the integration of physical information and multi-task modeling. Specifically, a detailed analysis of manufacturing processes and energy patterns is first conducted to study various energy types and extract relevant physical laws. Next, using industry insights and thermodynamic principles, key equations for energy balance and conversion are derived to create a physics-based loss function for model training. Finally, the pi-MMoE model framework is constructed, featuring multi-expert networks and gating mechanisms to balance cross-task knowledge sharing and expert learning. In a case study of a textile factory, the pi-MMoE model reduced electricity and steam prediction errors by 14.28% and 27.27%, respectively, outperforming traditional deep learning methods. This demonstrates that the model can improve prediction performance, providing a novel approach to intelligent energy management and promoting sustainable development in manufacturing.

Advancing soil property prediction with encoder-decoder structures integrating traditional deep learning methods in Vis-NIR spectroscopy

The technology for estimating soil properties using visible and near-infrared spectroscopy has been maturing, with corresponding advances and breakthroughs in deep learning models. In this study, based on the large soil spectral library LUCAS, we explore the potential of encoder-decoder structures to improve convolutional neural network regression predictions. By introducing an encoder-decoder structure into the feature channels of a six-layer CNN model (TRNN model), we significantly enhanced the performance of shallow CNN models and successfully carried out regression predictions for seven soil properties. We employed IntegratedGradients, DeepLift, GradientShap, and DeepLiftShap methods to interpret the output of the TRNN model. Our TRNN model, built on raw spectra, demonstrated high accuracy in predicting multiple soil properties, outperforming residual architectures, LSTMs, various CNN architectures, and other traditional machine learning methods proposed in previous studies. We also investigated the impact of multi-task output structures (TRNN 1-M and TRNN M−M) and single-task output structures (TRNN 1-1) on model performance. For the TRNN model with an encoder-decoder structure, multi-task output structures resulted in a reduction in performance. The TRNN showed outstanding results in regression analysis of the seven soil properties selected in this study (cation exchange capacity, organic carbon content, calcium carbonate content, pH, clay content, silt content, and sand content), with R2 values exceeding 0.93 for all seven properties. Different soil characteristics correspond to different wavelengths, with multiple characteristic peaks commonly observed. This research convincingly demonstrates the enormous potential of combining large model architectures with traditional deep learning approaches for predicting soil properties, which could significantly advance precision agriculture.

A graph-learning based model for automatic diagnosis of Sjögren’s syndrome on digital pathological images: a multicentre cohort study

BackgroundSjögren’s Syndrome (SS) is a rare chronic autoimmune disorder primarily affecting adult females, characterized by chronic inflammation and salivary and lacrimal gland dysfunction. It is often associated with systemic lupus erythematosus, rheumatoid arthritis and kidney disease, which can lead to increased mortality. Early diagnosis is critical, but traditional methods for diagnosing SS, mainly through histopathological evaluation of salivary gland tissue, have limitations.MethodsThe study used 100 labial gland biopsy, creating whole-slide images (WSIs) for analysis. The proposed model, named Cell-tissue-graph-based pathological image analysis model (CTG-PAM) and based on graph theory, characterizes single-cell feature, cell-cell feature, and cell-tissue feature. Building upon these features, CTG-PAM achieves cellular-level classification, enabling lymphocyte recognition. Furthermore, it leverages connected component analysis techniques in the cell graph structure to perform SS diagnosis based on lymphocyte counts.FindingsCTG-PAM outperforms traditional deep learning methods in diagnosing SS. Its area under the receiver operating characteristic curve (AUC) is 1.0 for the internal validation dataset and 0.8035 for the external test dataset. This indicates high accuracy. The sensitivity of CTG-PAM for the external dataset is 98.21%, while the accuracy is 93.75%. In comparison, the sensitivity and accuracy for traditional deep learning methods (ResNet-50) are lower. The study also shows that CTG-PAM’s diagnostic accuracy is closer to skilled pathologists compared to beginners.InterpretationOur findings indicate that CTG-PAM is a reliable method for diagnosing SS. Additionally, CTG-PAM shows promise in enhancing the prognosis of SS patients and holds significant potential for the differential diagnosis of both non-neoplastic and neoplastic diseases. The AI model potentially extends its application to diagnosing immune cells in tumor microenvironments.

Fast flow field prediction of pollutant leakage diffusion based on deep learning.

Predicting pollutant leakage and diffusion processes is crucial for ensuring people's safety. While the deep learning method offers high simulation efficiency and superior generalization, there is currently a lack of research on predicting pollutant leakage and diffusion flow field using deep learning. Therefore, it is necessary to conduct further studies in this area. This paper introduces a two-level network method to model the flow characteristics of pollutant diffusion. The proposed method in this study demonstrates a significant enhancement in flow field prediction accuracy compared to traditional deep learning methods. Moreover, it improves computational efficiency by over 800 times compared to traditional computational fluid dynamics (CFD) methods. Unlike conventional CFD methods that require grid expansion to calculate all operation conditions, the deep learning method is not confined by grid limitations. While deep learning methods may not entirely replace CFD methods, they can serve as a valuable supplementary tool, expanding the versatility of CFD methods. The findings of this research establish a robust foundation for incorporating deep learning methods in addressing pollutant leakage and diffusion challenges.

Deep finite volume method for partial differential equations

In this paper, we introduce the Deep Finite Volume Method (DFVM), an innovative deep learning framework tailored for solving high-order (order ≥2) partial differential equations (PDEs). Our approach centers on a novel loss function crafted from local conservation laws derived from the original PDE, distinguishing DFVM from traditional deep learning methods. By formulating DFVM in the weak form of the PDE rather than the strong form, we enhance accuracy, particularly beneficial for PDEs with less smooth solutions compared to strong-form-based methods like Physics-Informed Neural Networks (PINNs). A key technique of DFVM lies in its transformation of all second-order or higher derivatives of neural networks into first-order derivatives which can be computed directly using Automatic Differentiation (AD). This adaptation significantly reduces computational overhead, particularly advantageous for solving high-dimensional PDEs. Numerical experiments demonstrate that DFVM achieves equal or superior solution accuracy compared to existing deep learning methods such as PINN, Deep Ritz Method (DRM), and Weak Adversarial Networks (WAN), while drastically reducing computational costs. Notably, for PDEs with nonsmooth solutions, DFVM yields approximate solutions with relative errors up to two orders of magnitude lower than those obtained by PINN. The implementation of DFVM is available on GitHub at https://github.com/Sysuzqs/DFVM.

Multi-task deep learning strategy for map-type classification

ABSTRACT The information contained in a map is always represented by text, symbols, and map-type. Among them, map-type is a critical element that denotes the category and theme of map content, which can support map content extraction, map retrieval, and other map data mining tasks. However, the representations of map-type are always so complex and diverse that relies on multiple descriptive labels. Traditional deep learning methods regarding map-type classification are developed by single label, which only supports single-task classification. This means these approaches might overlook the common features among multiple map-type. In this paper, we propose a framework of multi-task deep learning strategy for employing the state-of-the-art deep convolutional neural network models, including ResNet50, MobileNetV2, and Inception-v3, to conduct efficient multi-label map-type classification. Specifically, we develop the dedicated classification module and label selection layer, and integrate them into the backbone of the deep convolutional network model. The experiments revealed that our proposed multi-task classification strategy can achieve greater accuracy in map-type classification, with less processing time required compared to state-of-the-art deep learning regarding map-type classification. This proves that multi-task classification strategy could be competitive to recognize and discover the complex map-type information.

An intensity-based self-supervised domain adaptation method for intervertebral disc segmentation in magnetic resonance imaging

Background and objective:Accurate IVD segmentation is crucial for diagnosing and treating spinal conditions. Traditional deep learning methods depend on extensive, annotated datasets, which are hard to acquire. This research proposes an intensity-based self-supervised domain adaptation, using unlabeled multi-domain data to reduce reliance on large annotated datasets.Methods:The study introduces an innovative method using intensity-based self-supervised learning for IVD segmentation in MRI scans. This approach is particularly suited for IVD segmentations due to its ability to effectively capture the subtle intensity variations that are characteristic of spinal structures. The model, a dual-task system, simultaneously segments IVDs and predicts intensity transformations. This intensity-focused method has the advantages of being easy to train and computationally light, making it highly practical in diverse clinical settings. Trained on unlabeled data from multiple domains, the model learns domain-invariant features, adeptly handling intensity variations across different MRI devices and protocols.Results:Testing on three public datasets showed that this model outperforms baseline models trained on single-domain data. It handles domain shifts and achieves higher accuracy in IVD segmentation.Conclusions:This study demonstrates the potential of intensity-based self-supervised domain adaptation for IVD segmentation. It suggests new directions for research in enhancing generalizability across datasets with domain shifts, which can be applied to other medical imaging fields.

Deep learning with local spatiotemporal structure preserving for soft sensor development of complex industrial processes

Data-driven soft sensors have emerged as indispensable tools for predicting quality variables in complex industrial processes because of their cost-effectiveness and ease of maintenance. In particular, soft sensors based on deep learning have been utilized in extensive research and successful applications in recent years. However, traditional deep learning methods capture hierarchical data features by minimizing global fitting errors, neglecting the local structural characteristics implied in the original data. In this paper, we propose a new deep learning method for soft sensor development. Utilizing autoencoders as the foundational architecture of our network, a new semisupervised strategy is adopted for layerwise pretraining optimization. On the one hand, more representative data features are extracted by maintaining the local spatiotemporal structure of the data; on the other hand, layer-by-layer supervised learning is employed to identify the critical features that are aligned with the ultimate task, which aids in obtaining the optimal network parameters and improving the resulting prediction accuracy. Subsequently, a local spatiotemporal structure-preserving stacked semisupervised autoencoder (LSP-SuAE) is established. To evaluate the feasibility and effectiveness of the proposed approach, experiments are carried out in a real industrial process. A soft sensor based on the LSP-SuAE is developed to predict the rate of ethylbenzene conversion during the dehydrogenation of styrene. The experimental results demonstrate that, compared to five other common or similar data-driven modeling methods, LSP-SuAE exhibits higher prediction accuracy and better stability.

DACLnet: A Dual-Attention-Mechanism CNN-LSTM Network for the Accurate Prediction of Nonlinear InSAR Deformation

Nonlinear deformation is a dynamically changing pattern of multiple surface deformations caused by groundwater overexploitation, underground coal mining, landslides, urban construction, etc., which are often accompanied by severe damage to surface structures or lead to major geological disasters; therefore, the high-precision monitoring and prediction of nonlinear surface deformation is significant. Traditional deep learning methods encounter challenges such as long-term dependencies or difficulty capturing complex spatiotemporal patterns when predicting nonlinear deformations. In this study, we developed a dual-attention-mechanism CNN-LSTM network model (DACLnet) to monitor and accurately predict nonlinear surface deformations precisely. Using advanced time series InSAR results as input, the DACLnet integrates the spatial feature extraction capability of a convolutional neural network (CNN), the advantages of the time series learning of a long short-term memory (LSTM) network, and the enhanced focusing effect of the dual-attention mechanism on crucial information, significantly improving the prediction accuracy of nonlinear surface deformations. The groundwater overexploitation area of the Turpan Basin, China, is selected to test the nonlinear deformation prediction effect of the proposed DACLnet. The results demonstrate that the DACLnet accurately captures developmental trends in historical surface deformations and effectively predicts surface deformations for the next two months in the study area. Compared to traditional LSTM and CNN-LSTM methods, the root mean square error (RMSE) of the DACLnet improved by 85.09% and 68.57%, respectively. These research results can provide crucial technical support for the early warning and prevention of geological disasters and can serve as an effective alternative tool for short-term ground subsidence prediction in areas lacking hydrogeological and other related data.

Research on green supply chain finance risk identification based on two-stage deep learning

As a resonance product between financial services and the upgrading of the green industry, green supply chain finance has garnered extensive attention in the process of ecological civilization construction. Effectively promoting the green transformation of small and medium-sized enterprises and achieving the "dual carbon" goals necessitate the avoidance of corporate green risks. However, the complex interdependence and information asymmetry among green supply chain finance enterprises result in data characteristics such as multi-source small samples and high-dimensional imbalance. To address these issues, this paper proposes a risk assessment model based on two-stage deep learning. In the first stage, we employ Generative Adversarial Network (GAN) to generate minority class default samples, and utilize Stacked Auto-Encoder (SAE) to extract data features with closed-form parameter calculation capability. In the second stage, the obtained features are input into a Deep Neural Network (DNN), and parameter learning and model optimization are conducted through joint training. Finally, to model low-order feature interactions, we integrate the Support Vector Machine (SVM) algorithm. The paper is grounded in the green innovation production of enterprises, collecting financial data of 176 upstream and downstream enterprises and corresponding core enterprise green indicators from 2013 to 2022. Experimental results demonstrate that GAN oversampling technique not only enhances the model's AUC metric but also significantly improves the F1 score. Compared with traditional deep learning methods, the proposed two-stage deep integration model effectively reduces training loss and exhibits superiority in identifying green supply chain finance risks.