Convolutional Neural Network Training Research Articles

A non-intrusive load recognition method combining adaptive PSO algorithm and CNN model

Real-time monitoring of electricity usage details through load monitoring techniques is a crucial aspect of smart power grid management and monitoring, allowing for the acquisition of information on the electricity usage of individual appliances for power users. Accurate detection of electricity load is essential for refined load management and monitoring of power supply quality, facilitating the improvement of power management at the user side and enhancing power operation efficiency. Non-intrusive load monitoring (NILM) techniques require only the analysis of total load data to achieve load monitoring of electricity usage details, and offer advantages such as low cost, easy implementation, high reliability, and user acceptance. However, with the increasing number of distributed new load devices on the user side and the diversification of device development, simple load recognition algorithms are insufficient to meet the identification needs of multiple devices and achieve high recognition accuracy. To address this issue, a non-intrusive load recognition (NILR) model that combines an adaptive particle swarm optimization algorithm (PSO) and convolutional neural network (CNN) has been proposed. In this model, pixelated images of different electrical V-I trajectories are used as inputs for the CNN, and the optimal network layer and convolutional kernel size are determined by the adaptive PSO optimization algorithm during the CNN training process. The proposed model has been validated on the public dataset PLAID, and experimental results demonstrate that it has achieved a overall recognition accuracy of 97.26% and F-1 score of 96.92%, significantly better than other comparison models. The proposed model effectively reduces the confusion between various devices, exhibiting good recognition and generalization capabilities.

A Deep Learning Application for Building Damage Assessment Using Ultra-High-Resolution Remote Sensing Imagery in Turkey Earthquake

Rapid building damage assessment following an earthquake is important for humanitarian relief and disaster emergency responses. In February 2023, two magnitude-7.8 earthquakes struck Turkey in quick succession, impacting over 30 major cities across nearly 300 km. A quick and comprehensive understanding of the distribution of building damage is essential for efficiently deploying rescue forces during critical rescue periods. This article presents the training of a two-stage convolutional neural network called BDANet that integrated image features captured before and after the disaster to evaluate the extent of building damage in Islahiye. Based on high-resolution remote sensing data from WorldView2, BDANet used pre-disaster imagery to extract building outlines; the image features before and after the disaster were then combined to conduct building damage assessment. We optimized these results to improve the accuracy of building edges and analyzed the damage to each building, and used population distribution information to estimate the population count and urgency of rescue at different disaster levels. The results indicate that the building area in the Islahiye region was 156.92 ha, with an affected area of 26.60 ha. Severely damaged buildings accounted for 15.67% of the total building area in the affected areas. WorldPop population distribution data indicated approximately 253, 297, and 1,246 people in the collapsed, severely damaged, and lightly damaged areas, respectively. Accuracy verification showed that the BDANet model exhibited good performance in handling high-resolution images and can be used to directly assess building damage and provide rapid information for rescue operations in future disasters using model weights.

Screening of Mild Cognitive Impairment Subtypes Through the Training of 1D Convolutional Neural Network with QEEG Features

AbstractBackgroundThe pre‐clinical stage of dementia, mild cognitive impairment (MCI) carries various pathological pathways and respective prognoses. The two foremost causes of MCI and dementia that have been identified through autopsy diagnoses are Amyloid plaques and Lewy body, commonly referred to as Alzheimer’s disease (AD) and Lewy body disease (LBD) respectively. Although correct identification of pathological pathways plays a crucial role in effective designing of treatment methods, MCI subtypes are underexploited in the current diagnostic paradigm due to multiple reasons, mainly cost‐effectiveness. Hence, we utilized quantitative electroencephalography (QEEG) that are cheap and convenient to record in the distinguishment of MCI subtypes.MethodA total of 151 MCI patients’ EEG data were aggregated into 3 groups, pure AD (n = 29), pure LBD (n = 88), and mixed (n = 34). The mixed type was further subdivided into main AD (n = 31) and main LBD (n = 3) groups, in accordance with AD/LBD tendency exhibited by the patients. The clinical labelling was brought by the experienced experts of Yonsei Severance hospital, South Korea.1D SE‐ResNet‐based classification model was established for quantitative investigation of AD and LBD propensity. The power spectrum density (PSD) in dB/Hz scale were computed from the EEG data and were used to train the model. Due to the small number of data, augmentation was applied to the training data. The final dataset was split into 8 to 1 to 1 ratio (Train n = 7092: 2274 pure AD + main AD; 3694 pure LBD + main LBD, Validation n = 14: 6 pure AD + main AD; 8 Pure LBD + main LBD, Test n = 15: 6 pure AD + main AD; 9 pure LBD + main LBD).ResultThe classification model showed validation results of 85.7% accuracy, 83.3% AD sensitivity and 87.5% LBD sensitivity. Test results were at 92.8% accuracy, 83.3% AD sensitivity and 88.9% LBD sensitivity.ConclusionQEEG‐based deep learning classifier developed in this study successfully distinguished the two main causes of cognitive degeneration in MCI patients. Our model can help identify subtype‐specific spectral trends, which could also make contributions in the establishment of effective treatment methods for MCI.

A Deep Learning Approach to Improve the Control of Dynamic Wireless Power Transfer Systems

In this paper, an innovative approach for the fast estimation of the mutual inductance between transmitting and receiving coils for Dynamic Wireless Power Transfer Systems (DWPTSs) is implemented. To this end, a Convolutional Neural Network (CNN) is used; an image representing the geometry of two coils that are partially misaligned is the input of the CNN, while the output is the corresponding inductance value. Finite Element Analyses are used for the computation of the inductance values needed for CNN training. This way, thanks to a fast and accurate inductance estimated by the CNN, it is possible to properly manage the power converter devoted to charge the battery, avoiding the wind up of its controller when it attempts to transfer power in poor coupling conditions.

Biometric Authentication based on Liveness Detection Using Face Landmarks and Deep Learning Model

This paper describes the approach to active liveness detection of the face using facial features and movements. The project aims to create a better method for detecting liveness in real-time on an application programming interface (API) server. The project is built using Python programming with the computer vision libraries OpenCV, dlib and MediaPipe and the deep learning library Tensorflow. There are five modules in active liveness detection progress related to different parts or movements on the face: headshakes, nodding, eye blinks, smiles, and mouths. The functionality of modules runs through face landmarking through dlib and MediaPipe and detection of face features through Tensorflow Convolutional Neural Network (CNN) trained in two different approaches: smile detection and eye-blink detection. The result of implementing face landmarking shows an accurate result through the pre-trained model of MediaPipe and the pre-trained parameter of the dlib 68 landmarking model. And more than 90% classification model accuracy in precision, recall, and f1-score for both trained CNNs in detecting smiles and eyes blinking through the Scikit-Learn classification report. In addition, the prototype API is also implemented using the Python RESTful API library, FastAPI, to test the detection functionality in the prototype Android application. The prototype result is outstanding, as the model excellently requests and retrieves from the API server. The possible research path gives the success of real-time detection on API servers for easy implementation of liveness detection on low-spec client devices.

Intuitive Perception - Speech Recognition using Machine Learning

Abstract: Machine Learning is widely used to detect the movement of lips. It has been observed that the data generated through visual stir of mouth and corresponding audio are largely identified. This fact has been exploited for lip reading and for perfecting speech recognition. We propose a system that uses a CNN( Convolutional Neural Network) which is trained and used to descry the movement of lips and prognosticate the words being spoken. This trained CNN will be suitable to descry the words that are spoken within the videotape and display it in a textbook format. The CNN may also calculate on fresh information handed by the environment, knowledge of the language, and any residual hail. We hope to learn whether the application of machine literacy, more specifically the DNN( Deep Neural Network), could also be an applicable seeker for working the problem of lip reading. The main end of our design is to directly honor the expressions being spoken through automated lip reading

Reliability analysis of the vehicle-bridge system with uncertain parameters by SPEM integrating deep leaning method

The dynamic response of the vehicle-bridge (VB) system was the key problem in the assessment of railway vehicle running safety on bridges, which performed random vibration analysis due to the necessary uncertain parameter arising from either vehicle or bridge structure. Traditional random analysis methods usually required complete time-history calculated data and modal information to obtain the time-dominate statistics for evaluating the reliability of the vehicle running safety, whose computing was difficult and costs too much. This paper proposed a hybrid approach, which integrated the stochastic pseudo excitation method (SPEM) to optimized deep learning method (OLDM), to study the random vibration of uncertain VB system. The dynamic formulas of the uncertain system were produced by the SPEM, depending on which a neural network was established to construct the numerical model consisting of two functional modules, i.e., convolutional neural network (CNN) training with input rail irregularities and Bi-directional long short-term memory (BiLSTM) layer for the VB system response time-history prediction, i.e., CNN-BiLSTM. Due to the characteristics of calculated responses by SPEM, the BiLSTM cell was simulated by introducing the randomness of excitation and uncertain characteristics of parameters of the system into a portion of the cell, enabling the numerical model to convey the uncertainties of the VB system and obtained its stochastic response. For the proposed method, a vehicle running through a simply-supported railway bridge was applied to verify its accuracy and efficiency. According to the computed dynamic responses of VB system by the SPEM, a deep neural network was optimized for the investigated VB dynamic system. The dynamic statistics and reliability of vehicle were conducted through the optimized deep learning model and compared with the results from the conventional VB model by the Monte Carlo Method (MCM). The influence of training epoch, learn rate drop factor and vehicle running speed were investigated to examine the robustness of the deep learning approach.

Convolutional neural network for intrusion detection using blockchain technology

Cyber-physical systems (CPS) are becoming increasingly ubiquitous, connecting the physical world with the cyber realm. This convergence has exposed CPS to a growing threat landscape, necessitating robust intrusion detection systems (IDS) to safeguard critical infrastructure. Deep learning (DL) has emerged as a powerful tool for IDS, and convolutional neural networks (CNNs) have demonstrated exceptional performance in this domain. However, traditional IDS models are susceptible to data tampering and manipulation, compromising their integrity and effectiveness. Blockchain technology, with its inherent immutability and tamper-proof nature, offers a promising solution to enhance the security and reliability of IDS models. In this study, we propose a CNN-based IDS model that leverages blockchain technology to secure network traffic data. Our hypothesis is that integrating blockchain with CNNs can significantly improve the security and robustness of IDS models against data tampering and manipulation. To test our hypothesis, we employ a greedy-based genetic algorithm to select the most relevant features from network traffic data, followed by training a CNN model using the selected features. Finally, we evaluate the trained CNN model on a real-world dataset, demonstrating its ability to accurately classify network traffic as normal or intrusive. The results of our evaluation reveal that the proposed CNN-based IDS model achieves a classification accuracy of 99.2%, surpassing traditional IDS models. Moreover, our model exhibits enhanced resilience against data tampering and manipulation, demonstrating the effectiveness of blockchain integration in safeguarding the integrity of IDS models. Our findings underscore the potential of blockchain-enhanced CNNs as a robust and secure solution for intrusion detection in CPS, ensuring the integrity and protection of critical infrastructure.

FACILE: A capsule network with fewer capsules and richer hierarchical information for malware image classification

The struggle between security researchers and malware perpetuates an endless arms race. Recent studies indicate that converting malware into grayscale images and using convolutional neural network or capsule network to classify and identify them is a promising approach. Nonetheless, training convolutional neural networks demands a significant amount of data and parameters, posing a challenge to obtaining enough training samples for malware detection tasks, particularly for new malware families. Additionally, a high number of parameters translates to low classification efficiency. Capsule networks offer a promising approach to achieve higher-level feature representations through dynamic routing between capsules, but research on capsule networks is still nascent, and there are two major challenges for malware classification tasks: the huge number of parameters, which makes model training difficult, and capsules shedding, which results in information loss during routing. To address these issues, we propose a capsule network called FACILE that uses fewer capsules and richer hierarchical information. In the initial feature extraction stage, we perform multi-level feature extraction and fusion using dynamic convolution. In the routing phase, we introduce balance coefficients to enhance the model's representational power and stabilize the training process. We conducted experiments on the VIRUS-MNIST, MalImg, and BIG2015 datasets and found that FACILE requires only 8.1% of the capsules and 1.8%-3.3% of the parameters compared to the original CapsNet, while reducing the error rate to 8.087%, 1.149%, and 2.797%, respectively. Furthermore, FACILE demonstrates competitive performance with a small number of parameters as the training samples decrease compared to convolutional neural network-based models such as ResNet, EfficientNet, and VGG.

CWGAN-GP with residual network model for lithium-ion battery thermal image data expansion with quantitative metrics

Lithium batteries find extensive applications in energy storage. Temperature is a crucial indicator for assessing the state of lithium-ion batteries, and numerous experiments require thermal images of lithium-ion batteries for research purposes. However, acquiring thermal imaging samples of lithium-ion battery faults is challenging due to factors such as high experimental costs and associated risks. To address this, our study proposes the utilization of a Conditional Wasserstein Generative Adversarial Network with Gradient Penalty and Residual Network (CWGAN-GP with Residual Network) to augment the dataset of thermal images depicting lithium-ion battery faults. We employ various evaluation metrics to quantitatively analyze and compare the generated thermal images of lithium-ion batteries. Subsequently, the expanded dataset, comprising four types of thermal images depicting lithium-ion battery faults, is input into a Mask Region-based Convolutional Neural Network for training. The results demonstrate that the proposed model surpasses both traditional Generative Adversarial Network and Wasserstein Generative Adversarial Network in terms of the quality of generated thermal images of lithium-ion batteries. Moreover, the augmentation of the dataset leads to an improvement in the fault diagnosis accuracy of the Mask Region-based Convolutional Neural Network.

A method of concrete damage detection and localization based on weakly supervised learning

AbstractAutomatic inspection of concrete surface defects based on visual elements is crucial for the timely detection of security risks in infrastructure. Moreover, accurate determination of the geographical location of the detected defects is critical for subsequent maintenance and reinforcement tasks. This study employed convolutional neural network (CNN) training methods for detection and localization. This approach employs bounding boxes to confine damaged pixels and utilizes projection loss to foster similarity learning between pixels. In addition, geotags are automatically indexed through the vectorization of invariant features in a scene, which maintains high model accuracy and reduces training costs. The proposed method can detect and classify typical concrete defects in complex scenarios and accurately locate them without the use of external sensors. In addition, the proposed model can achieve pixel‐level defect detection and geographic location determination through the cost of bounding box annotation and automatic indexing. The proposed model was evaluated using 10,691 images of four typical concrete defects in various complex environments. The results demonstrate that the proposed method achieves a pixel‐level detection accuracy of 48.75 and a location accuracy of 83.69, showing a better performance than other methods.

Deep learning predicts therapy-relevant genetics in acute myeloid leukemia from Pappenheim-stained bone marrow smears

AbstractThe detection of genetic aberrations is crucial for early therapy decisions in acute myeloid leukemia (AML) and recommended for all patients. Because genetic testing is expensive and time consuming, a need remains for cost-effective, fast, and broadly accessible tests to predict these aberrations in this aggressive malignancy. Here, we developed a novel fully automated end-to-end deep learning pipeline to predict genetic aberrations directly from single-cell images from scans of conventionally stained bone marrow smears already on the day of diagnosis. We used this pipeline to compile a multiterabyte data set of >2 000 000 single-cell images from diagnostic samples of 408 patients with AML. These images were then used to train convolutional neural networks for the prediction of various therapy-relevant genetic alterations. Moreover, we created a temporal test cohort data set of >444 000 single-cell images from further 71 patients with AML. We show that the models from our pipeline can significantly predict these subgroups with high areas under the curve of the receiver operating characteristic. Potential genotype-phenotype links were visualized with 2 different strategies. Our pipeline holds the potential to be used as a fast and inexpensive automated tool to screen patients with AML for therapy-relevant genetic aberrations directly from routine, conventionally stained bone marrow smears already on the day of diagnosis. It also creates a foundation to develop similar approaches for other bone marrow disorders in the future.

Automatic caries detection in bitewing radiographs: part I-deep learning.

The aim of this work was to assemble alarge annotated dataset of bitewing radiographs and to use convolutional neural networks to automate the detection of dental caries in bitewing radiographs with human-level performance. A dataset of 3989 bitewing radiographs was created, and 7257 carious lesions were annotated using minimal bounding boxes. The dataset was then divided into 3 parts for the training (70%), validation (15%), and testing (15%) of multiple object detection convolutional neural networks (CNN). The tested CNN architectures included YOLOv5, Faster R-CNN, RetinaNet, and EfficientDet. To further improve the detection performance, model ensembling was used, and nested predictions were removed during post-processing. The models were compared in terms of the [Formula: see text] score and average precision (AP) with various thresholds of the intersection over union (IoU). The twelve tested architectures had [Formula: see text] scores of 0.72-0.76. Their performance was improved by ensembling which increased the [Formula: see text] score to 0.79-0.80. The best-performing ensemble detected caries with the precision of 0.83, recall of 0.77, [Formula: see text], and AP of 0.86 at IoU=0.5. Small carious lesions were predicted with slightly lower accuracy (AP 0.82) than medium or large lesions (AP 0.88). The trained ensemble of object detection CNNs detected caries with satisfactory accuracy and performed at least as well as experienced dentists (see companion paper, Part II). The performance on small lesions was likely limited by inconsistencies in the training dataset. Caries can be automatically detected using convolutional neural networks. However, detecting incipient carious lesions remains challenging.

Intelligent damage diagnosis method for offshore platforms based on enhanced stabilization diagrams and convolutional neural network

Considering that it is difficult to evaluate the damage state of offshore platform structures under environmental excitation by stochastic subspace identification (SSI) stability diagrams alone, an intelligent damage diagnosis method based on enhanced stability diagrams and convolutional neural network (CNN) is proposed. The data-driven SSI algorithm and covariance-driven SSI algorithm are utilized to identify stability diagrams of monitoring data, and the stability diagrams of the two algorithms are superimposed together for image enhancement. Further, the enhanced stability diagrams are used as input samples for CNN training to distinguish the damage state of the structure. In the meanwhile, the whale optimization algorithm is employed to optimize the hyper parameters of CNN to ulteriorly improve the recognition performance. The final test accuracy of CNN is 97.20%, and is 13.09% higher than before hyper parameter optimization, which indicates that the damage diagnosis method based on enhanced stability diagrams and CNN is reasonable and effective, and is expected to be applied to real-time damage diagnosis of offshore platform structures.

The Use of Artificial Intelligence for the Classification of Craniofacial Deformities.

Positional cranial deformities are a common finding in toddlers, yet differentiation from craniosynostosis can be challenging. The aim of this study was to train convolutional neural networks (CNNs) to classify craniofacial deformities based on 2D images generated using photogrammetry as a radiation-free imaging technique. A total of 487 patients with photogrammetry scans were included in this retrospective cohort study: children with craniosynostosis (n = 227), positional deformities (n = 206), and healthy children (n = 54). Three two-dimensional images were extracted from each photogrammetry scan. The datasets were divided into training, validation, and test sets. During the training, fine-tuned ResNet-152s were utilized. The performance was quantified using tenfold cross-validation. For the detection of craniosynostosis, sensitivity was at 0.94 with a specificity of 0.85. Regarding the differentiation of the five existing classes (trigonocephaly, scaphocephaly, positional plagiocephaly left, positional plagiocephaly right, and healthy), sensitivity ranged from 0.45 (positional plagiocephaly left) to 0.95 (scaphocephaly) and specificity ranged from 0.87 (positional plagiocephaly right) to 0.97 (scaphocephaly). We present a CNN-based approach to classify craniofacial deformities on two-dimensional images with promising results. A larger dataset would be required to identify rarer forms of craniosynostosis as well. The chosen 2D approach enables future applications for digital cameras or smartphones.

Improving deep learning performance by augmenting training data

Satellite image recognition is a crucial application of computer vision that has the po-tential to be applied in various fields such as disaster management, agriculture, and urban planning. The objective of this study is to determine the optimal amount of input data required and select the most effective methods of augmentation necessary for training a convolutional neural network (CNN) for satellite image recognition. To achieve this, we perform a series of experiments to investigate the effect of input data quantity on several performance metrics, including model accuracy, convergence, and generalization. Additionally, we explore the impact of various data augmentation techniques, such as rotation, scaling, and flipping, on model performance. The study suggests several strategies for identifying the saturation point and mitigating the effects of overtraining, in-cluding early stopping and dropout regularization. The findings from this study can significantly contribute to the development of more ef-ficient satellite recognition models. Furthermore, they can help improve the performance of existing models, in addition to providing guidance for future research. The study emphasizes the importance of carefully selecting input data and augmentation methods to achieve optimal performance in CNNs, which is fundamental in advancing the field of computer vision. In addition to the above, the study investigates the potential of transfer learning by pre-training the model on a related dataset and fine-tuning it on the satellite imagery dataset. This approach can reduce the amount of required data and training time and increase model performance. Overall, this study provides valuable insights into the optimal amount of input data and augmentation techniques for training CNNs for satellite image recognition, and its findings can guide future research in this area.

An Intermediate-Centric Dataflow for Transposed Convolution Acceleration on FPGA

Transposed convolution has been prevailing in convolutional neural networks (CNNs), playing an important role in multiple scenarios such as image segmentation and back-propagation process of training CNNs. This mainly benefits from the ability to up-sample the input feature maps by interpolating new information from the input feature pixels. However, the backward-stencil computation constrains its performance and hindered its wide application in diverse platforms. Moreover, in contrast to the efforts on accelerating the convolution, there is a rare investigation on the acceleration of transposed convolution that is identically compute-intensive as the former. For acceleration of transposed convolution, we propose an intermediate-centric dataflow scheme, in which we decouple the generation of the intermediate patch from its further process, aim at efficiently performing the backward-stencil computation . The intermediate-centric dataflow breaks the transposed convolution into several phases/stages, achieving feeding the input feature maps and performing the backward-stencil computation in a pipelining manner. It also provides four-degree computation parallelism and efficient data reuse of input feature maps/weights. Furthermore, we also theoretically analyze the irregular data dependence leveraging the polyhedral model, which constrains the parallel computing of transposed convolution. Additionally, we devise an optimization problem to explore the design space and automatically generate the optimal design configurations for different transposed convolutional layers and hardware platforms. By selecting the representative transposed convolutional layers from DCGAN, FSRCNN, and FCN, we generate the corresponding accelerator arrays of intermediate-centric dataflow on the Xilinx Alveo U200 platform and reach the performance of 3.92 TOPS, 2.72 TOPS, and 4.76 TOPS, respectively.

How deep convolutional neural networks lose spatial information with training

A central question of machine learning is how deep nets manage to learn tasks in high dimensions. An appealing hypothesis is that they achieve this feat by building a representation of the data where information irrelevant to the task is lost. For image datasets, this view is supported by the observation that after (and not before) training, the neural representation becomes less and less sensitive to diffeomorphisms acting on images as the signal propagates through the network. This loss of sensitivity correlates with performance and surprisingly correlates with a gain of sensitivity to white noise acquired during training. Which are the mechanisms learned by convolutional neural networks (CNNs) responsible for the these phenomena? In particular, why is the sensitivity to noise heightened with training? Our approach consists of two steps. (1) Analyzing the layer-wise representations of trained CNNs, we disentangle the role of spatial pooling in contrast to channel pooling in decreasing their sensitivity to image diffeomorphisms while increasing their sensitivity to noise. (2) We introduce model scale-detection tasks, which qualitatively reproduce the phenomena reported in our empirical analysis. In these models we can assess quantitatively how spatial pooling affects these sensitivities. We find that the increased sensitivity to noise observed in deep ReLU networks is a mechanistic consequence of the perturbing noise piling up during spatial pooling, after being rectified by ReLU units. Using odd activation functions like tanh drastically reduces the CNNs’ sensitivity to noise.

Detection of AI-Created Images Using Pixel-Wise Feature Extraction and Convolutional Neural Networks.

Generative AI has gained enormous interest nowadays due to new applications like ChatGPT, DALL E, Stable Diffusion, and Deepfake. In particular, DALL E, Stable Diffusion, and others (Adobe Firefly, ImagineArt, etc.) can create images from a text prompt and are even able to create photorealistic images. Due to this fact, intense research has been performed to create new image forensics applications able to distinguish between real captured images and videos and artificial ones. Detecting forgeries made with Deepfake is one of the most researched issues. This paper is about another kind of forgery detection. The purpose of this research is to detect photorealistic AI-created images versus real photos coming from a physical camera. Id est, making a binary decision over an image, asking whether it is artificially or naturally created. Artificial images do not need to try to represent any real object, person, or place. For this purpose, techniques that perform a pixel-level feature extraction are used. The first one is Photo Response Non-Uniformity (PRNU). PRNU is a special noise due to imperfections on the camera sensor that is used for source camera identification. The underlying idea is that AI images will have a different PRNU pattern. The second one is error level analysis (ELA). This is another type of feature extraction traditionally used for detecting image editing. ELA is being used nowadays by photographers for the manual detection of AI-created images. Both kinds of features are used to train convolutional neural networks to differentiate between AI images and real photographs. Good results are obtained, achieving accuracy rates of over 95%. Both extraction methods are carefully assessed by computing precision/recall and F1-score measurements.

AI-enabled crack-length estimation from acoustic emission signal signatures

Abstract This article addresses the classification of fatigue crack length using artificial intelligence (AI) applied to acoustic emission (AE) signals. The AE signals were collected during fatigue of two specimen types. One specimen type had a 1-mm hole for crack initiation. The other specimen type had a 150-micron wide slit of various lengths. Fatigue testing was performed under stress-intensity-factor control to moderate crack advancement. The slit specimen produced AE signals only from crack advancement at the slit tips whereas the 1-mm hole specimens produced AE signals both from crack tip advancement and crack rubbing or clapping. The AE signals were captured with a piezoelectric wafer active sensor (PWAS) array connected to MISTRAS instrumentation and AEwin software. The collected AE signals were preprocessed using time-of-flight filtering and denoising. Choi Williams transform converted time-domain AE-signals into spectrograms. To apply machine learning, the spectrogram images were used as input data for the training, validation, and testing of a GoogLeNet convolutional neural network (CNN). The CNN was trained to sort the AE signals into crack-length classes. CNN performance enhancements, including synthetic data generation and class balancing were developed. A three-class example with crack lengths of (i) 10-12 mm; (ii) 12-14 mm; and (iii) 14-16 mm is provided. Our AI approach was able to classify the AE signals into these three classes with 91% accuracy thus proving that the AE signals contain sufficient information for crack estimation using an AI-enabled approach.