Convolutional Neural Network Training Research Articles

Feature space separation by conformity loss driven training of CNN

Convolutional neural networks (CNNs) have enabled tremendous achievements in image classification, as the model can automatically extract image features and assign a proper classification. Nevertheless, the classification is lacking robustness to — for humans’ invisible perturbations on the input. To improve the robustness of the CNN model, it is necessary to understand the decision-making procedure of CNN models. By inspecting the learned feature space, we found that the classification regions are not always clearly separated by the CNN model. The overlap of classification regions increases the possibility to less perturbation induced input changes on classification results. Therefore, the clear separation of feature spaces of the CNN model should support decision robustness. In this paper, we propose to use a novel loss function called “conformity loss” to strengthen disjoint feature spaces during learning at different layers of the CNN, in order to improve the intra-class compactness and inter-class differences in trained representations. The same function was used as an evaluation metric to measure the feature space separation during the testing process. In conclusion, the conformity loss driven trained model has shown better feature space separation at comparable output performance.

WiKA: A Vision Based Sign Language Recognition from Extracted Hand Joint Features using DeepLabCut

The use of hand gesture recognition for sign language translation to address the communication gap between the hearing majority and the deaf population has had significant breakthroughs over the years. While the contact-based approach uses wearable devices, a vision-based solution is preferred owing to the convenience it offers and since it obviates the need for complicated gears. This study presents the development of WiKA, an open-source software designed to track the joints of the hands and interpret them into their corresponding sign language counterparts. DeepLabCut, a markerless pose estimation software, was employed to develop the Hand-Joint Tracking Model through the training of a sequential Convolutional Neural Network, utilizing extracted Hand-Joint features to predict the sign language alphabets (A-Z) and numbers (1-9) based on the positioning of the joints. The developed Hand-Joint tracking model exhibited a 4.92% training error and a 5.74% test error with a p-cut-off of 60%. On the other hand, the developed sign language recognition achieved a 96.44% prediction accuracy with only 0.0356% misclass. This model can be further integrated into mobile phones for seamless conversations between the signing and non-signing populations.

Lit me up: A reference free adaptive low light image enhancement for in-the-wild conditions

Images captured with different devices in uneven conditions (e.g., invariable lighting, low lighting, weather changes, exposure time, etc.) often lead to low image visibility and poor color and contrast, affecting the performance of computer vision and pattern recognition applications. The pre-trained convolutional neural networks (CNNs) solely rely on the training data and lack adaptation due to the uncertainty in the lighting conditions. Moreover, capturing large-scale datasets to train CNNs also raises the computational complexity and overall cost. This work integrates the knowledge and data and proposes a two-stage Uneven-to-Enliven network (U2E-Net), which rapidly learns to see in uneven conditions. A multiple-layered Uneven network learns to distinguish the reflection and illumination in the input images, and an encoder–decoder-based Enliven-Net contextualizes the illumination information. A key component in such ill-posed problems is to obtain information from priors and pairs; however, we present the compelling idea of information trade-off followed by decomposition consistency, thereby progressively improving the visual quality with the subsequent enhancement operations. To this end, we proposed a two-faceted framework that can work without depending on the data type. A novel color and contrast preservation strategy (CPS) is proposed following the decomposition of input data. CPS is integrated within the network to extract contrast in the darkest background regions.

From explanation to intervention: Interactive knowledge extraction from Convolutional Neural Networks used in radiology.

Deep Learning models such as Convolutional Neural Networks (CNNs) are very effective at extracting complex image features from medical X-rays. However, the limited interpretability of CNNs has hampered their deployment in medical settings as they failed to gain trust among clinicians. In this work, we propose an interactive framework to allow clinicians to ask what-if questions and intervene in the decisions of a CNN, with the aim of increasing trust in the system. The framework translates a layer of a trained CNN into a measurable and compact set of symbolic rules. Expert interactions with visualizations of the rules promote the use of clinically-relevant CNN kernels and attach meaning to the rules. The definition and relevance of the kernels are supported by radiomics analyses and permutation evaluations, respectively. CNN kernels that do not have a clinically-meaningful interpretation are removed without affecting model performance. By allowing clinicians to evaluate the impact of adding or removing kernels from the rule set, our approach produces an interpretable refinement of the data-driven CNN in alignment with medical best practice.

Using convolutional neural networks for stereological characterization of 3D hetero-aggregates based on synthetic STEM data

The 3D nano/microstructure of materials can significantly influence their macroscopic properties. In order to enable a better understanding of such structure-property relationships, 3D microscopy techniques can be deployed, which are however often expensive in both time and costs. Often 2D imaging techniques are more accessible, yet they have the disadvantage that the 3D nano/microstructure of materials cannot be directly retrieved from such measurements. The motivation of this work is to overcome the issues of characterizing 3D structures from 2D measurements for hetero-aggregate materials. For this purpose, a method is presented that relies on machine learning combined with methods of spatial stochastic modeling for characterizing the 3D nano/microstructure of materials from 2D data. More precisely, a stochastic model is utilized for the generation of synthetic training data. This kind of training data has the advantage that time-consuming experiments for the synthesis of differently structured materials followed by their 3D imaging can be avoided. More precisely, a parametric stochastic 3D model is presented, from which a wide spectrum of virtual hetero-aggregates can be generated. Additionally, the virtual structures are passed to a physics-based simulation tool in order to generate virtual scanning transmission electron microscopy (STEM) images. The preset parameters of the 3D model together with the simulated STEM images serve as a database for the training of convolutional neural networks, which can be used to determine the parameters of the underlying 3D model and, consequently, to predict 3D structures of hetero-aggregates from 2D STEM images. Furthermore, an error analysis is performed with respect to structural descriptors, e.g. the hetero-coordination number. The proposed method is applied to image data of TiO2-WO3 hetero-aggregates, which are highly relevant in photocatalysis processes. However, the proposed method can be transferred to other types of aggregates and to different 2D microscopy techniques. Consequently, the method is relevant for industrial or laboratory setups in which product quality is to be quantified by means of inexpensive 2D image acquisition.

A New Approach to Structural Damage Identification Based on Power Spectral Density and Convolutional Neural Network

In the field of structural health monitoring, vibration-based damage identification remains a formidable challenge. Key to this challenge is the establishment of a reliable association between observed vibration characteristics and the actual state of structural damage (e.g. stiffness reduction). This association not only accurately indicates the presence of damage, but also the location and severity of the damage. To solve this complex pattern identification problem, a large number of approaches, including deep learning, have emerged in recent years. In this paper, we propose a new structural damage identification method that utilizes the vibration information of the structure and a convolutional neural network based on Alex NET improvement. The method consists of calculating the acceleration response power spectral density of damaged and undamaged structures under impact loading separately, and then making a difference between the two power spectral data, and subsequently introducing these power spectral difference data into the convolutional neural network for training. The use of power spectral density analysis as a preprocessing step converts the time-domain signals into frequency-domain signals, and this conversion allows the convolutional neural network to capture and learn from the specific frequency characteristics of the data, thus facilitating the learning process of the neural network model. In this paper, the effectiveness of the method is critically evaluated through numerical simulation and experimental validation, and 3% and 5% noise are added to the numerical study to test the robustness of the method. During the convolution neural network training process, the optimal training mean squared error (MSE) is [Formula: see text] in the case of no noise addition; the optimal training MSE is [Formula: see text] in the case of noise addition. Both the results of simulations and experiments confirm the high accuracy and good robustness of the method in localizing structural damage.

METHOD OF GENERATIVE-ADVERSARIAL NETWORKS SEARCHING ARCHITECTURES FOR BIOMEDICAL IMAGES SYNTHESIS

Context. The article examines the problem of automatic design of architectures of generative-adversarial networks. Generativeadversarial networks are used for image synthesis. This is especially true for the synthesis of biomedical images – cytological and histological, which are used to make a diagnosis in oncology. The synthesized images are used to train convolutional neural networks. Convolutional neural networks are currently among the most accurate classifiers of biomedical images.
 Objective. The aim of the work is to develop an automatic method for searching for architectures of generative-adversarial networks based on a genetic algorithm.
 Method. The developed method consists of the stage of searching for the architecture of the generator with a fixed discriminator and the stage of searching for the architecture of the discriminator with the best generator.
 At the first stage, a fixed discriminator architecture is defined and a generator is searched for. Accordingly, after the first step, the architecture of the best generator is obtained, i.e. the model with the lowest FID value.
 At the second stage, the best generator architecture was used and a search for the discriminator architecture was carried out. At each cycle of the optimization algorithm, a population of discriminators is created. After the second step, the architecture of the generative-adversarial network is obtained.
 Results. Cytological images of breast cancer on the Zenodo platform were used to conduct the experiments. As a result of the study, an automatic method for searching for architectures of generatively adversarial networks has been developed. On the basis of computer experiments, the architecture of a generative adversarial network for the synthesis of cytological images was obtained. The total time of the experiment was ~39.5 GPU hours. As a result, 16,000 images were synthesized (4000 for each class). To assess the quality of synthesized images, the FID metric was used.The results of the experiments showed that the developed architecture is the best. The network’s FID value is 3.39. This result is the best compared to well-known generative adversarial networks.
 Conclusions. The article develops a method for searching for architectures of generative-adversarial networks for the problems of synthesis of biomedical images. In addition, a software module for the synthesis of biomedical images has been developed, which can be used to train CNN.

Eye-LRCN: A Long-Term Recurrent Convolutional Network for Eye Blink Completeness Detection.

Computer vision syndrome causes vision problems and discomfort mainly due to dry eye. Several studies show that dry eye in computer users is caused by a reduction in the blink rate and an increase in the prevalence of incomplete blinks. In this context, this article introduces Eye-LRCN, a new eye blink detection method that also evaluates the completeness of the blink. The method is based on a long-term recurrent convolutional network (LRCN), which combines a convolutional neural network (CNN) for feature extraction with a bidirectional recurrent neural network that performs sequence learning and classifies the blinks. A Siamese architecture is used during CNN training to overcome the high-class imbalance present in blink detection and the limited amount of data available to train blink detection models. The method was evaluated on three different tasks: blink detection, blink completeness detection, and eye state detection. We report superior performance to the state-of-the-art methods in blink detection and blink completeness detection, and remarkable results in eye state detection.

448 Deformable Medial Modeling to Generate Novel Shape Features of the Placenta Using Automated versus Manual Segmentations

OBJECTIVES/GOALS: In this study, we implemented deformable medial modeling as a morphometric approach in first trimester placentas to characterize morphometric differences between fully automated and manual segmentations. METHODS/STUDY POPULATION: Twenty placentas from singleton pregnancies between 11-14 weeks’ gestation were manually and automatically segmented from 3D ultrasound volumes. Automated segmentations were produced by a trained convolutional neural network pipeline. Dice overlap scores and volumes were computed between manual and automated segmentations. Deformable medial modeling was applied to both manual and automated segmentations to produce the following metrics: maternal and chorionic surface areas (SA), thickness, circumference, and diameter along the generated medial surface. Placental non-planarity was also determined as the greatest medial surface height difference. A paired t-test and simple linear regression was performed between manual and automated segmentations for each shape metric. RESULTS/ANTICIPATED RESULTS: Mean placental volume measurements between manual and automated segmentations were similar, with a percent difference of 3.28% and a mean Dice overlap score of 0.85 ± 0.07. There were strong, statistically significant (p <0.01) linear correlations with chorionic and maternal SA, SA difference, thickness, circumference, medial surface diameter, and medial surface height difference. No significant differences were noted with chorionic SA, thickness, circumference, maximum medial surface diameter, or medial surface height difference. However, statistically significant differences (p <0.01-0.03) were noted in maternal SA, SA difference, and mean medial surface diameter. Despite these differences, mean percent difference for all morphometric parameters was less than or equal to 10%. DISCUSSION/SIGNIFICANCE: A deformable medial model evaluate unique global and regional shape placental features with highly correlated values between manual versus automated placental segmentations. However, clinical studies are needed to determine if minor differences would impact the clinical utility of these features as potential indicators of placental function.

CNN-based model updating for structures by direct use of dynamic structural response measurements

This study presents a convolutional neural network (CNN)-based model updating method for structures using dynamic structural responses. In the presented method, the dynamic structural response is used directly for model updating, thereby omitting the manual process of extracting modal parameters. In the method, the type of stiffness variable in the target structure for model updating is selected, and a number of candidate models are generated by multiple selected stiffness variables in the structures and possible values of variables. Dynamic structural responses are then extracted by time history structural analysis of the candidate models, and a CNN sets dynamic structural response as input, and information on stiffness variables as output is introduced to learn the relationship between the two data. CNNs trained using dynamic displacement response and dynamic acceleration response are presented as Model_XT and Model_XTa, respectively. The presented models can rapidly estimate stiffness of the target structure by entering dynamic responses into the models. These models are applied to model updating for three example structures, and the stiffness estimation performance for model updating is evaluated. Furthermore, the length of data used for CNN training, the stiffness estimation performance according to the CNN input layer settings, and the stiffness estimation applied with limited data to the updated model are analyzed.

Monitoring manufacturing systems using AI: A method based on a digital factory twin to train CNNs on synthetic data

Modern cyber–physical production systems provide advanced solutions to enhance factory throughput and efficiency. However, monitoring its behaviour and performance becomes challenging as the complexity of a manufacturing system increases. Artificial Intelligence (AI) provides techniques to manage not only decision-making tasks but also to support monitoring. The integration of AI into a factory can be facilitated by a reliable Digital Twin (DT) that enables knowledge-based and data-driven approaches. While computer vision and convolutional neural networks (CNNs) are crucial for monitoring production systems, the need for extensive training data hinders their adoption in real factories. The proposed methodology leverages the Digital Twin of a factory to generate labelled synthetic data for training CNN-based object detection models. Regarding their position and state, the focus is on monitoring entities in manufacturing systems, such as parts, components, fixtures, and tools. This approach reduces the need for large training datasets and enables training when the actual system is unavailable. The trained CNN model is evaluated in various scenarios, with a real case study involving an industrial pilot plant for repairing and recycling Printed Circuit Boards (PCBs).

3D photogrammetry and deep‐learning deliver accurate estimates of epibenthic biomass

Abstract Accurate biomass estimates are key to understanding a wide variety of ecological functions. In marine systems, epibenthic biomass estimates have traditionally relied on either destructive/extractive methods that are limited to horizontal soft‐sediment environments, or simplistic geometry‐based biomass conversions that are unsuitable for more complex morphologies. Consequently, there is a requirement for non‐destructive, higher‐accuracy methods that can be used in an array of environments, targeting more morphologically diverse taxa, and at ecological relevant scales. We used a combination of 3D photogrammetry, convolutional neural network (CNN) automated taxonomic identification, and taxa‐specific biovolume:biomass calibrations to test the viability of estimating biomass of three species of morphologically complex epibenthic taxa from in situ stereo 2D source imagery. Our trained CNN produced accurate and reliable annotations of our target taxa across a wide range of conditions. When incorporated into photogrammetric 3D models of underwater surveys, we were able to automatically isolate our three target taxa from their environment, producing biovolume measurements that had respective mean similarities of 99%, 102% and 120% of those obtained from human annotators. When combined with taxa‐specific biovolume:biomass calibration values, we produced biomass estimates of 88%, 125% and 133% mean similarity to that of the ‘true’ biomass of the respective taxa. Our methodology provides a highly reliable and efficient method for estimating epibenthic biomass of morphologically complex taxa using non‐destructive 2D imagery. This approach can be applied to a variety of environments and photo/video survey approaches (e.g. SCUBA, ROV, AUV) and is especially valuable in spatially extensive surveys where manual approaches are prohibitively time‐consuming.

GAN-Based Data Augmentation with Vehicle Color Changes to Train a Vehicle Detection CNN

Object detection is a challenging task that requires a lot of labeled data to train convolutional neural networks (CNNs) that can achieve human-level accuracy. However, such data are not easy to obtain, as they involve significant manual work and costs to annotate the objects in images. Researchers have used traditional data augmentation techniques to increase the amount of training data available to them. A recent trend in object detection is to use generative models to automatically create annotated data that can enrich a training set and improve the performance of the target model. This paper presents a method of training the proposed ColorGAN network, which is used to generate augmented data for the target domain of interest with the least compromise in quality. We demonstrate a method to train a GAN with images of vehicles in different colors. Then, we demonstrate that our ColorGAN can change the color of vehicles of any given vehicle dataset to a set of specified colors, which can serve as an augmented training dataset. Our experimental results show that the augmented dataset generated by the proposed method helps enhance the detection performance of a CNN for applications where the original training data are limited. Our experiments also show that the model can achieve a higher mAP of 76% when the model is trained with augmented images along with the original training dataset.

Detection of lung cancer through SERS analysis of serum

Early detection and postoperative assessment are crucial for improving overall survival among lung cancer patients. Here, we report a non-invasive technique that integrates Raman spectroscopy with machine learning for the detection of lung cancer. The study encompassed 88 postoperative lung cancer patients, 73 non-surgical lung cancer patients, and 68 healthy subjects. The primary aim was to explore variations in serum metabolism across these cohorts. Comparative analysis of average Raman spectra was conducted, while principal component analysis was employed for data visualization. Subsequently, the augmented dataset was used to train convolutional neural networks (CNN) and Resnet models, leading to the development of a diagnostic framework. The CNN model exhibited superior performance, as verified by the receiver operating characteristic curve. Notably, postoperative patients demonstrated an increased likelihood of recurrence, emphasizing the crucial need for continuous postoperative monitoring. In summary, the integration of Raman spectroscopy with CNN-based classification shows potential for early detection and postoperative assessment of lung cancer.

REPrune: Channel Pruning via Kernel Representative Selection

Channel pruning is widely accepted to accelerate modern convolutional neural networks (CNNs). The resulting pruned model benefits from its immediate deployment on general-purpose software and hardware resources. However, its large pruning granularity, specifically at the unit of a convolution filter, often leads to undesirable accuracy drops due to the inflexibility of deciding how and where to introduce sparsity to the CNNs. In this paper, we propose REPrune, a novel channel pruning technique that emulates kernel pruning, fully exploiting the finer but structured granularity. REPrune identifies similar kernels within each channel using agglomerative clustering. Then, it selects filters that maximize the incorporation of kernel representatives while optimizing the maximum cluster coverage problem. By integrating with a simultaneous training-pruning paradigm, REPrune promotes efficient, progressive pruning throughout training CNNs, avoiding the conventional train-prune-finetune sequence. Experimental results highlight that REPrune performs better in computer vision tasks than existing methods, effectively achieving a balance between acceleration ratio and performance retention.

Alzheimer's MRI Recognition Based on Convolutional Neural Networks

Alzheimer's disease (AD) is a progressively worsening neurodegenerative disorder, where early diagnosis is crucial for implementing effective disease management strategies. In multimodal diagnostic pathways, Magnetic Resonance Imaging (MRI) plays a pivotal role as a non-invasive imaging technique in disease identification and progression monitoring. Deep learning, particularly Convolutional Neural Networks (CNN), can extract key biomarkers from complex imaging data. By training CNN models to automatically interpret MRI scans, radiology experts can utilize these advanced analytical tools for more efficient and consistent pathological assessments, thereby enhancing clinical decision support systems. This study develops a deep learning-based method for the automatic classification of Alzheimer's brain MRI images, using convolutional neural networks to categorize MRI images into four stages: no dementia, very mild dementia, mild dementia, and moderate dementia. A CNN model is constructed, learning distinctive features of the images through multi-level feature extraction and performing feature map visualization. Early stopping is employed to prevent overfitting. The model is trained on a training set and evaluated on a test set, with performance metrics including confusion matrix, accuracy, precision, recall, F1 score, Kappa coefficient, Matthew’s coefficient, ROC curve with AUC value, and PQ curve with AP value. The results show that the proposed model effectively differentiates between various categories of MRI images, providing valuable tools for early diagnosis and condition monitoring.

A review of convolutional neural networks in computer vision

In computer vision, a series of exemplary advances have been made in several areas involving image classification, semantic segmentation, object detection, and image super-resolution reconstruction with the rapid development of deep convolutional neural network (CNN). The CNN has superior features for autonomous learning and expression, and feature extraction from original input data can be realized by means of training CNN models that match practical applications. Due to the rapid progress in deep learning technology, the structure of CNN is becoming more and more complex and diverse. Consequently, it gradually replaces the traditional machine learning methods. This paper presents an elementary understanding of CNN components and their functions, including input layers, convolution layers, pooling layers, activation functions, batch normalization, dropout, fully connected layers, and output layers. On this basis, this paper gives a comprehensive overview of the past and current research status of the applications of CNN models in computer vision fields, e.g., image classification, object detection, and video prediction. In addition, we summarize the challenges and solutions of the deep CNN, and future research directions are also discussed.

Abstract 7169: A real-world toxicity-atlas shows that adverse events of combination-therapies commonly result in additive interactions

Abstract Purpose: Combination-therapies are a promising approach for improving cancer treatment, but it is challenging to predict their resulting adverse events in an real-world setting. Experimental Design: We provide here a proof-of-concept study using 15 million patient records from the US Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS). Complex adverse-event frequencies of drugs or their combinations were visualized as heatmaps onto a 2D UMAP grid. Adverse event frequencies were shown as colors to assess the ratio between individual and combined drug effects. To capture these patterns, we trained a convolutional neural network (CNN) autoencoder using 7,300 single-drug heatmaps. Additionally, statistical synergy analyses were performed based on BLISS independence or Chi-square testing. Results: The trained CNN model was able to decode patterns, showing that adverse events occur in global rather than isolated and unique patterns. Patterns were not likely to be attributed to disease symptoms given their relatively limited contribution to drug associated adverse events. Pattern recognition was validated using trial data from ClinicalTrials.gov and drug combination data. We examined the adverse-event interactions of 140 drug combinations known to be avoided in the clinic and found that near all of them showed additive rather than synergistic interactions, also when assessed statistically. Conclusion: Our study provides a framework for analyzing adverse events and suggest that adverse drug interactions commonly result in additive effects with a high level of overlap of adverse event patterns. These real-world insights may advance the implementation of new combination therapies in clinical practice. Citation Format: Asli Küçükosmanoglu, Silvia I. Scoarta, Bart Westerman. A real-world toxicity-atlas shows that adverse events of combination-therapies commonly result in additive interactions [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7169.

Fire Detection in Urban Areas Using Multimodal Data and Federated Learning

Fire chemical sensing for indoor detection of fire plays an essential role because it can detect chemical volatiles before smoke particles, providing a faster and more reliable method for early fire detection. A thermal imaging camera and seven distinct fire-detecting sensors were used simultaneously to acquire the multimodal fire data that is the subject of this paper. The low-cost sensors typically have lower sensitivity and reliability, making it impossible for them to detect fire at greater distances. To go beyond the limitation of using solely sensors for identifying fire, the multimodal dataset is collected using a thermal camera that can detect temperature changes. The proposed pipeline uses image data from thermal cameras to train convolutional neural networks (CNNs) and their many versions. The training of sensors data (from fire sensors) uses bidirectional long-short memory (BiLSTM-Dense) and dense and long-short memory (LSTM-DenseDenseNet201), and the merging of both datasets demonstrates the performance of multimodal data. Researchers and system developers can use the dataset to create and hone cutting-edge artificial intelligence models and systems. Initial evaluation of the image dataset has shown densenet201 as the best approach with the highest validation parameters (0.99, 0.99, 0.99, and 0.08), i.e., Accuracy, Precision, Recall, and Loss, respectively. However, the sensors dataset has also shown the highest parameters with the BILSTM-Dense approach (0.95, 0.95, 0.95, 0.14). In a multimodal data approach, image and sensors deployed with a multimodal algorithm (densenet201 for image data and Bi LSTM- Dense for Sensors Data) has shown other parameters (1.0, 1.0, 1.0, 0.06). This work demonstrates that, in comparison to the conventional deep learning approach, the federated learning (FL) approach performs privacy-protected fire leakage classification without significantly sacrificing accuracy and other validation parameters.

A Fast Reliability Evaluation Strategy for Power Systems under High Proportional Renewable Energy—A Hybrid Data-Driven Method

With the increasing scale of the power system, the increasing penetration of renewable energy, and the increasing uncertainty factors, traditional reliability evaluation methods based on Monte Carlo simulation have greatly reduced computational efficiency in complex power systems and cannot meet the requirements of real-time and rapid evaluation. This article proposes a hybrid data-driven strategy to achieve a rapid assessment of power grid reliability on two levels: offline training and online evaluation. Firstly, this article derives explicit analytical expressions for reliability indicators and component parameters, avoiding the computational burden of repetitive Monte Carlo simulation. Next, a large number of samples are quickly generated by parsing expressions to train convolutional neural networks (CNNs), and the system reliability index is quickly calculated under changing operating conditions through CNNs. Finally, the effectiveness and feasibility of the proposed method are verified through an improved RTS-79 testing system. The calculation results show that the method proposed in this article can achieve an online solution of second-level reliability indicators while ensuring calculation accuracy.