Accuracy Degradation Research Articles

LDF-BNN: A Real-Time and High-Accuracy Binary Neural Network Accelerator Based on the Improved BNext.

Significant progress has been made in industrial defect detection due to the powerful feature extraction capabilities of deep neural networks (DNNs). However, the high computational cost and memory requirement of DNNs pose a great challenge to the deployment of industrial edge-side devices. Although traditional binary neural networks (BNNs) have the advantages of small storage space requirements, high parallel computing capability, and low power consumption, the problem of significant accuracy degradation cannot be ignored. To tackle these challenges, this paper constructs a BNN with layered data fusion mechanism (LDF-BNN) based on BNext. By introducing the above mechanism, it strives to minimize the bandwidth pressure while reducing the loss of accuracy. Furthermore, we have designed an efficient hardware accelerator architecture based on this mechanism, enhancing the performance of high-accuracy BNN models with complex network structures. Additionally, the introduction of multi-storage parallelism alleviates the limitations imposed by the internal transfer rate, thus improving the overall computational efficiency. The experimental results show that our proposed LDF-BNN outperforms other methods in the comprehensive comparison, achieving a high accuracy of 72.23%, an image processing rate of 72.6 frames per second (FPS), and 1826 giga operations per second (GOPs) on the ImageNet dataset. Meanwhile, LDF-BNN can also be well applied to defect detection dataset Mixed WM-38, achieving a high accuracy of 98.70%.

Analysis of the degradation in transmission accuracy of RV reducers considering the wear clearance between cycloidal wheels and pinwheels

To accurately determine the transmission accuracy degradation pattern of Rotate Vector (RV) reducers due to wear, a predictive method is proposed that simultaneously considers the initial clearance and wear clearance between the cycloid wheel and the pinwheel. The initial clearance is derived from a comprehensive gear system analysis, considering assembly errors, manufacturing errors and modifications. Building on this, the wear depth of the cycloid wheel tooth profile is numerically simulated using Archard's wear theory, revealing the wear distribution pattern under various operating conditions. A Gaussian regression model is used to predict the maximum wear amount for different operational durations. Based on the predicted wear amount, a multi-body dynamic simulation model incorporating wear clearance was established to analyze the transmission error under varying clearance conditions. Results indicate that the wear depth along the tooth profile initially increases and then decreases. The depth and region of wear along the tooth profile gradually increase with load and output speed, with the impact of load on the wear region being greater than that of output speed. Under the initial clearance condition, the maximum transmission error is 22.91″. As the wear clearance increases, transmission accuracy declines sharply. When the cumulative wear clearance reaches 0.2 mm, the transmission error exceeds the allowable backlash limit. This study provides a theoretical reference for predicting accuracy degradation and guides the manufacturing design of RV reducers.

Sensor-Aware Data Imputation for Time-Series Machine Learning on Low-Power Wearable Devices

Wearable devices that have low-power sensors, processors, and communication capabilities are gaining wide adoption in several health applications. The machine learning algorithms on these devices assume that data from all sensors are available during runtime. However, data from one or more sensors may be unavailable due to energy or communication challenges. This loss of sensor data can result in accuracy degradation of the application. Prior approaches to handle missing data, such as generative models or training multiple classifiers for each combination of missing sensors are not suitable for low-energy wearable devices due to their high overhead at runtime. In contrast to prior approaches, we present an energy-efficient approach, referred to as Sensor-Aware iMputation (SAM), to accurately impute missing data at runtime and recover application accuracy. SAM first uses unsupervised clustering to obtain clusters of similar sensor data patterns. Next, it learns inter-relationship between clusters to obtain imputation patterns for each combination of clusters using a principled sensor-aware search algorithm. Using sensor data for clustering before choosing imputation patterns ensures that the imputation is aware of sensor data observations. Experiments on seven diverse wearable sensor based time-series datasets demonstrate that SAM is able to maintain accuracy within 5% of the baseline with no missing data when one sensor is missing. We also compare SAM against generative adversarial imputation networks (GAIN), transformers, and k-nearest neighbor methods. Results show that SAM outperforms all three approaches on average by more than 25% when two sensors are missing with negligible overhead compared to the baseline.

Cluster partition-fuzzy broad learning-based fast detection and localization framework for false data injection attack in smart distribution networks

The distributed renewable energy generations, as accessible and easily targets for attackers, introduce an extra false data injection attack (FDIA) threat in the smart distribution networks. Scattered attack points and complex attack features hinder the elimination of potential threats. In this context, an FDIA fast detection and pinpoint localization framework is proposed. This framework identifies abnormal signals and attacked nodes from the unique topology structure and status contiguity of smart distribution networks, namely, spatial-temporal correlations of power grids, by using a cluster partition-fuzzy broad learning system (CP-FBLS). Unlike most existing FDIA detection methods, which are dedicated to high accuracy but neglect the urgent need for rapid detection in smart distribution networks, the proposed CP-FBLS framework maintains the fast computational nature of a fuzzy broad learning system (FBLS), while avoiding the accuracy degradation caused by high-dimension of data in large-scale smart distribution networks. Moreover, the multi-layer structure of the proposed framework recognizes the location of FDIA, bridging the research gap of attack localization. To comprehensively evaluate the proposed strategy, datasets containing various FDIA types are constructed. Numerical simulations based on the above datasets in modified IEEE 34-bus and 123-bus distribution systems are implemented. The results of the case studies showed that the proposed method can achieve 98.43 % accuracy with 0.34 ms detection time, realizing rapid detection and localization of various FDIAs with satisfactory accuracy.

A robust iterative algorithm for SINS/USBL integrated navigation based on dual hydrophone differential model

Aiming at the complex underwater environment which leads to large measurement errors and outliers in ultrashort baseline, a robust iterative Kalman filtering algorithm based on the differential model of dual hydrophones is proposed in this paper. Firstly, to achieve accurate navigation under large initial misalignment angles, the algorithm builds an inertial navigation error equation based on the right invariant error definition of Lie groups. Then a dual hydrophone differential model is proposed based on SINS/USBL tightly coupled. It successfully improves the accuracy of integrated navigation and significantly mitigating the issue of SINS/USBL positioning accuracy degradation caused by USBL slant distance noise that grows with distance. Finally, a robust iterative Kalman filter based on the dual hydrophone differential model is proposed to counter the threat to the vehicle’s safe operation and the accuracy loss caused by outliers. The algorithm’s superiority and effectiveness are verified through simulation and sea trial.

Non-contact measurement of structural dynamic displacement based on improved edge detection and video interpolation

In structural health monitoring (SHM), smartphones are increasingly used for non-contact dynamic displacement measurement using computer vision. To overcome limitations in capturing high-frame-rate data, a non-contact measurement method of structural dynamic displacement based on an improved edge detection algorithm and video frame interpolation is proposed. Initially, a smartphone is employed to capture the structure’s vibration video and then generates intermediary frames between adjacent frames that match the structure’s motion state by using a video frame interpolation algorithm to obtain a video with a high frame rate. Subsequently, the improved edge detection algorithm is applied to measure the displacement within the interpolated video. In order to address the problem that traditional edge detection may have false edges leading to the degradation of measurement accuracy, a feature point tracking technique is proposed based on the traditional edge detection technique, which effectively excludes the interference of false edges in ROI and improves the accuracy of displacement monitoring. Experimental validation on a cantilever beam and steel frame bridge demonstrates the method’s viability and precision, highlighting its effectiveness in improving displacement measurement accuracy compared to traditional methods.

Full-field three-dimensional system calibration for composite surfaces reconstruction

The accurate calibration of composite surface measurement systems for specular and diffused surface is of great importance in a number of industrial applications. This paper addresses the challenges in calibrating composite surface reconstruction by proposing a comprehensive full-field calibration method. The principal innovation is the construction of a comprehensive reference phase and the pixel-by-pixel calibration of system parameters within the same coordinate system. The details are as follows: (1) construct a complete reference phase for the composite surfaces through external parameter constraints phase fusion and interpolation; (2) calibrate the mapping relationship coefficients, compensate the accuracy degradation of the partial reflection. Experimental results validate the accuracy of the full-field calibration data, demonstrating the effectiveness and feasibility.

A Variation-Aware Binary Neural Network Framework for Process Resilient In-Memory Computations

Binary neural networks (BNNs) that use 1-bit weights and activations have garnered interest as extreme quantization provides low power dissipation. By implementing BNNs as computation-in-memory (CIM), which computes multiplication and accumulations on memory arrays in an analog fashion, namely, analog CIM, we can further improve the energy efficiency to process neural networks. However, analog CIMs are susceptible to process variation, which refers to the variability in manufacturing that causes fluctuations in the electrical properties of transistors, resulting in significant degradation in BNN accuracy. Our Monte Carlo simulations demonstrate that in an SRAM-based analog CIM implementing the VGG-9 BNN model, the classification accuracy on the CIFAR-10 image dataset is degraded to below 50% under process variations in a 28 nm FD-SOI technology. To overcome this problem, we present a variation-aware BNN framework. The proposed framework is developed for SRAM-based BNN CIMs since SRAM is most widely used as on-chip memory; however, it is easily extensible to BNN CIMs based on other memories. Our extensive experimental results demonstrate that under process variation of 28 nm FD-SOI, with an SRAM array size of 128×128, our framework significantly enhances classification accuracies on both the MNIST hand-written digit dataset and the CIFAR-10 image dataset. Specifically, for the CONVNET BNN model on MNIST, accuracy improves from 60.24% to 92.33%, while for the VGG-9 BNN model on CIFAR-10, accuracy increases from 45.23% to 78.22%.

Bidirectional Corrective Model-Contrastive Federated Adversarial Training

When dealing with non-IID data, federated learning confronts issues such as client drift and sluggish convergence. Therefore, we propose a Bidirectional Corrective Model-Contrastive Federated Adversarial Training (BCMCFAT) framework. On the client side, we design a category information correction module to correct biases caused by imbalanced local data by incorporating the local client’s data distribution information. Through local adversarial training, more robust local models are obtained. Secondly, we propose a model-based adaptive correction algorithm in the server that leverages a self-attention mechanism to handle each client’s data distribution information and introduces learnable aggregation tokens. Through the self-attention mechanism, model contrast learning is conducted on each client to obtain aggregation weights of corrected client models, thus addressing the issues of accuracy degradation and slow convergence caused by client drift. Our algorithm achieves the best natural accuracy on the CIFAR-10, CIFAR-100, and SVHN datasets and demonstrates excellent adversarial defense performance against FGSM, BIM, and PGD attacks.

Research on the measurement mechanism of a six-axis force sensor based on a flexible hinge series–parallel hybrid

Abstract Addressing the common limitations of current six-axis force sensors, including the degradation of strain gauge accuracy due to environmental influences and the necessity for a complete replacement when damaged, this paper proposes an innovative six-axis force sensor that integrates a hybrid serial–parallel structure featuring flexible hinges. The sensor’s bracket is designed using a combination of serial–parallel flexible hinges, which confer load distribution capabilities, are lightweight, have high strength, and facilitate replacement. The static force model and the overall stiffness model of the sensor were developed based on the 3-universal–prismatic–universal parallel mechanism theory, providing a theoretical foundation for evaluating the sensor’s performance. To validate the accuracy of the stiffness model, finite element software is utilized for simulation-based verification. Subsequently, a calibration experiment is performed on the sensor, yielding results that conclusively show an error margin of less than 5% between experimental and theoretical values, satisfying the design criteria for sensor measurement.

DuCFF: A Dual-Channel Feature-Fusion Network for Workload Prediction in a Cloud Infrastructure

Cloud infrastructures are designed to provide highly scalable, pay-as-per-use services to meet the performance requirements of users. The workload prediction of the cloud plays a crucial role in proactive auto-scaling and the dynamic management of resources to move toward fine-grained load balancing and job scheduling due to its ability to estimate upcoming workloads. However, due to users’ diverse usage demands, the changing characteristics of workloads have become more and more complex, including not only short-term irregular fluctuation characteristics but also long-term dynamic variations. This prevents existing workload-prediction methods from fully capturing the above characteristics, leading to degradation of prediction accuracy. To deal with the above problems, this paper proposes a framework based on a dual-channel temporal convolutional network and transformer (referred to as DuCFF) to perform workload prediction. Firstly, DuCFF introduces data preprocessing technology to decouple different components implied by workload data and combine the original workload to form new model inputs. Then, in a parallel manner, DuCFF adopts the temporal convolution network (TCN) channel to capture local irregular fluctuations in workload time series and the transformer channel to capture long-term dynamic variations. Finally, the features extracted from the above two channels are further fused, and workload prediction is achieved. The performance of the proposed DuCFF’s was verified on various workload benchmark datasets (i.e., ClarkNet and Google) and compared to its nine competitors. Experimental results show that the proposed DuCFF can achieve average performance improvements of 65.2%, 70%, 64.37%, and 15%, respectively, in terms of Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE) and R-squared (R2) compared to the baseline model CNN-LSTM.

Some effects of limited wall-sensor availability on flow estimation with 3D-GANs

In this work we assess the impact of the limited availability of wall-embedded sensors on the full 3D estimation of the flow field in a turbulent channel with Reτ=200\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$Re_{\ au }=200$$\\end{document}. The estimation technique is based on a 3D generative adversarial network (3D-GAN). We recently demonstrated that 3D-GANs are capable of estimating fields with good accuracy by employing fully-resolved wall quantities (pressure and streamwise/spanwise wall shear stress on a grid with DNS resolution). However, the practical implementation in an experimental setting is challenging due to the large number of sensors required. In this work, we aim to estimate the flow fields with substantially fewer sensors. The impact of the reduction of the number of sensors on the quality of the flow reconstruction is assessed in terms of accuracy degradation and spectral length-scales involved. It is found that the accuracy degradation is mainly due to the spatial undersampling of scales, rather than the reduction of the number of sensors per se. We explore the performance of the estimator in case only one wall quantity is available. When a large number of sensors is available, pressure measurements provide more accurate flow field estimations. Conversely, the elongated patterns of the streamwise wall shear stress make this quantity the most suitable when only few sensors are available. As a further step towards a real application, the effect of sensor noise is also quantified. It is shown that configurations with fewer sensors are less sensitive to measurement noise.

Optimal filtering of an information process depended on noninformative parameters described by Marcovian chain when radio signals are receiving

In radio navigation and radio location tasks tracking systems on radio signals parameters often operate when object acceleration has stepwise changing. If we synthesize various tracking systems using optimal filtering theory it is necessary to represent the estimated process as a multidimensional Marcovian process. But such representation for stepwise changed acceleration is a rough approximation that lead to estimates accuracy degradation. One of possible approach to such disambiguation is to describe acceleration as Marcovian chain with finite number of states. This approach is developed in the article. Objective. To synthesize an optimal filtering algorithm of information process depended on noninformative parameters described by Marcovian chain using the observation grouping method when receiving radio signals and to implement simulation for an illustrative task.

SpQuant-SNN: ultra-low precision membrane potential with sparse activations unlock the potential of on-device spiking neural networks applications.

Spiking neural networks (SNNs) have received increasing attention due to their high biological plausibility and energy efficiency. The binary spike-based information propagation enables efficient sparse computation in event-based and static computer vision applications. However, the weight precision and especially the membrane potential precision remain as high-precision values (e.g., 32 bits) in state-of-the-art SNN algorithms. Each neuron in an SNN stores the membrane potential over time and typically updates its value in every time step. Such frequent read/write operations of high-precision membrane potential incur storage and memory access overhead in SNNs, which undermines the SNNs' compatibility with resource-constrained hardware. To resolve this inefficiency, prior works have explored the time step reduction and low-precision representation of membrane potential at a limited scale and reported significant accuracy drops. Furthermore, while recent advances in on-device AI present pruning and quantization optimization with different architectures and datasets, simultaneous pruning with quantization is highly under-explored in SNNs. In this work, we present SpQuant-SNN, a fully-quantized spiking neural network with ultra-low precision weights, membrane potential, and high spatial-channel sparsity, enabling the end-to-end low precision with significantly reduced operations on SNN. First, we propose an integer-only quantization scheme for the membrane potential with a stacked surrogate gradient function, a simple-yet-effective method that enables the smooth learning process of quantized SNN training. Second, we implement spatial-channel pruning with membrane potential prior, toward reducing the layer-wise computational complexity, and floating-point operations (FLOPs) in SNNs. Finally, to further improve the accuracy of low-precision and sparse SNN, we propose a self-adaptive learnable potential threshold for SNN training. Equipped with high biological adaptiveness, minimal computations, and memory utilization, SpQuant-SNN achieves state-of-the-art performance across multiple SNN models for both event-based and static image datasets, including both image classification and object detection tasks. The proposed SpQuant-SNN achieved up to 13× memory reduction and >4.7× FLOPs reduction with < 1.8% accuracy degradation for both classification and object detection tasks, compared to the SOTA baseline.

Differences among the total electron content derived by radio occultation, global ionospheric maps and satellite altimetry

In recent years, significant progress has been in ionospheric modeling research through data ingestion and data assimilation from a variety of sources, including ground-based global navigation satellite systems, space-based radio occultation and satellite altimetry (SA). Given the diverse observing geometries, vertical data coverages and intermission biases among different measurements, it is imperative to evaluate their absolute accuracies and estimate systematic biases to determine reasonable weights and error covariances when constructing ionospheric models. This study specifically investigates the disparities among the vertical total electron content (VTEC) derived from SA data of the Jason and Sentinel missions, the integrated VTEC from the Constellation Observing System for Meteorology, Ionosphere and Climate (COSMIC) and global ionospheric maps (GIMs). To mitigate the systematic bias resulting from differences in satellite altitudes, the vertical ranges of various VTECs are mapped to a standardized height. The results indicate that the intermission bias of SA-derived VTEC remains relatively stable, with Jason-1 serving as a benchmark for mapping other datasets. The mean bias between COSMIC and SA-derived VTEC is minimal, suggesting good agreement between these two space-based techniques. However, COSMIC and GIM VTEC exhibit remarkable seasonal discrepancies, influenced by the solar activity variations. Moreover, GIMs demonstrate noticeable hemispheric asymmetry and a degradation in accuracy ranging from 0.7 to 1.7 TECU in the ocean-dominant Southern Hemisphere. While space-based observations effectively illustrate phenomena such as the Weddell Sea anomaly and longitudinal ionospheric characteristics, GIMs tend to exhibit a more pronounced mid-latitude electron density enhancement structure.

An aircraft engine remaining useful life prediction method based on predictive vector angle minimization and feature fusion gate improved transformer model

Remaining useful life (RUL) prediction is crucial for achieving intelligent and predictive maintenance of aircraft engines. In practical applications, advance prediction values smaller than the true values can prevent serious deferred maintenance accidents. Using asymmetric loss functions directly leads to noticeable accuracy degradation, and existing methods often fail to satisfy accuracy and advance prediction requirements. To address this problem, this paper proposes a novel RUL prediction method based on the Prediction Vector Angle (PVA) minimization and Feature Fusion Gate (FFG) improved Transformer network. Specifically, the FFG is proposed to enhance Transformer prediction accuracy by dynamically fusing global and local features. The concept of PVA is first introduced based on the tilting properties of the RUL descent process. The target of the prediction model is cleverly transformed from error minimization to PVA minimization through the cosine similarity loss function. Various experiments on the CMAPSS dataset demonstrate the effectiveness of the proposed method in achieving high accuracy and advanced prediction. Compared to the state-of-the-art method, RMSE is reduced by at least 2.94 % and Score by 7.00 %. Finally, the PVA minimization mechanism significantly improves long short-term memory and convolutional neural network performance. The proposed method is noteworthy for its superiority and applicability.

Analysis and Verification on Buckling Mechanism of Spatial Tow Steering in Automated Fiber Placement

Automated fiber placement (AFP) enables the efficient and precise fabrication of complex-shaped aerospace composite structures with lightweight and high-performance properties. However, due to the excessive compression on the inner edge of the tow placed along the curved trajectory, the resulting defects represented by buckling and wrinkles in spatial tow steering can induce poor manufacturing accuracy and quality degradation of products. In this paper, a theoretical model of tow buckling based on the first-order shear deformation laminate theory, linear elastic adhesion interface and Hertz compaction contact theory is proposed to analyze the formation mechanism of the wrinkles and predict the formation of defects by solving the critical radius of the trajectory, and finite element analysis involving the cohesive zone modeling (CZM) is innovated to simulate the local buckling state of the steered tow in AFP. Additionally, numerical parametric studies and experimental results indicate that mechanical properties and geometric parameters of the prepreg, the curvature of the placement trajectory and critical process parameters have a significant impact on buckling formation, and optimization of process parameters can achieve effective suppression of placement defects. This research proposes a theoretical modeling method for tow buckling, and conducts in-depth research on defect formation and suppression methods based on finite element simulation and placement experiments.

SGFNet: A semantic graph-based multimodal network for financial invoice information extraction

To meet the demand for a large amount of invoice entry work in the financial industry and improve the low accuracy of traditional manual entry, we construct SGFNet, a financial invoice information extraction network that integrates semantic graph associations and multimodal modeling. First, we construct a graph of strong and weak semantic associations between data within each modality based on the correlation of text content. Subsequently, we model the multimodal data in a unified structure, extract the text modal information of invoices along with corresponding image and layout modal information, and guide the fusion and embedding of multimodal data through semantic associations in the graph to produce a richer feature representation. Furthermore, semantically linked multimodal information is fed into an aggregated multimodal self-attention mechanism to establish effective connection between modalities. Finally, with the combination of supervised contrastive learning and smoothed Kullback–Leibler divergence in terms of loss functions, the accuracy degradation problem incurred by sample imbalance and convergence instability is reduced. In our experiments, we achieved F1 scores of 93.71% for the English financial invoice dataset and 96.27% for the Chinese dataset, indicating that the proposed method successfully extracts feature information from different data modalities, thereby achieving satisfactory information extraction results.

Efficient sparse spiking auto-encoder for reconstruction, denoising and classification

Auto-encoders are capable of performing input reconstruction, denoising, and classification through an encoder-decoder structure. Spiking Auto-Encoders (SAEs) can utilize asynchronous sparse spikes to improve power efficiency and processing latency on neuromorphic hardware. In our work, we propose an efficient SAE trained using only Spike-Timing-Dependant Plasticity (STDP) learning. Our auto-encoder uses the Time-To-First-Spike (TTFS) encoding scheme and needs to update all synaptic weights only once per input, promoting both training and inference efficiency due to the extreme sparsity. We showcase robust reconstruction performance on the Modified National Institute of Standards and Technology (MNIST) and Fashion-MNIST datasets with significantly fewer spikes compared to state-of-the-art SAEs by 1–3 orders of magnitude. Moreover, we achieve robust noise reduction results on the MNIST dataset. When the same noisy inputs are used for classification, accuracy degradation is reduced by 30%–80% compared to prior works. It also exhibits classification accuracies comparable to previous STDP-based classifiers, while remaining competitive with other backpropagation-based spiking classifiers that require global learning through gradients and significantly more spikes for encoding and classification of MNIST/Fashion-MNIST inputs. The presented results demonstrate a promising pathway towards building efficient sparse spiking auto-encoders with local learning, making them highly suited for hardware integration.

Bridge the gap between simulated and real-world data in optical fiber mode decomposition for accuracy improvement: A deep learning-based co-learning framework with visual similarity-based matching

In the area of optical fiber mode decomposition (MD), data-driven deep-learning (DL) algorithms have achieved remarkable strides with unlimited theoretically synthetic data, but the differences between theoretical and real-world data in the DL context are rarely investigated. The two differences, namely the information bottleneck and the domain discrepancy, between the simulation and realistic domain remain significant obstacles that limit the practical deployment of DL-based MD approaches in real-world optical systems. The paper aims to mitigate their influence on the accuracy limitation and the accuracy degradation of DL-based MD. First, we present a DL-based co-learning framework for MD for the first time. With the co-learning process, the knowledge of the teacher model learned from rich input information sources (near-field, far-field, and phase-distribution optical images) is transferred to the student model with the only near-field input source, which breaks the information bottleneck by allowing the student model to learn more knowledge in a simulation context. Furthermore, to settle the domain discrepancy between ideal data and practical observation, we propose a visual similarity-based matching to assign real values of modal coefficients to the practical optical images whose ground truth values cannot be easily acquired, which enables the transfer learning from the synthetic data to the real-world data for accuracy enhancement. Simulated and experimental results show that the proposed co-learning framework with visual similarity-based matching has realized high precision and robustness in MD simulated and practical tasks. By bridging the gap between theoretical and practical data, our study offers novel perspectives on DL-based MD practice.