Lamb mode identification based on lightweight CNN

Convolutional neural networks (CNNs) have achieved significant accuracy improvement in many intelligent applications at the cost of intensive convolution operations and massive data movements. To efficiently deploy CNNs on low power embedded platforms in real time, the depthwise separable convolution has been proposed to replace the standard convolution, especially in lightweight CNNs, which remarkably reduces the computation complexity and model size. However, it is difficult for a general convolution engine to obtain the theoretical performance improvement as the decreased data dependency of depthwise convolution significantly reduces the data reuse opportunity. To address this issue, a flexible and highperformance accelerator based on FPGA is proposed to efficiently process the inference of both large-scale and lightweight CNNs. Firstly, by sharing the activation dataflow between the depthwise convolution and pooling layers, the control logic and data bus of the two layers are reused to maximize the data utilization and minimize the logic overhead. Furthermore, these two layers can be processed either directly after standard convolutions to eliminate the external memory accesses or independently to gain better flexibility. Thirdly, a performance model is proposed to automatically explore the optimal design options of the accelerator. The proposed hardware accelerator is evaluated on Intel Arria 10 SoC FPGA and demonstrates state-of-the-art performance on both large-scale CNNs, e.g., VGG, and lightweight ones, e.g., MobileNet.

Unmanned aerial vehicles (UAVs) are invaluable technologies concerning their remote control and monitoring capabilities. Convolutional neural networks (CNNs), known for their high pattern recognition capabilities, are appropriate for forest fire detection with UAVs. Deep convolutional neural networks show substantial performance on hardware with high processing capabilities. While these networks can be operated in unmanned aerial vehicles controlled from ground control stations equipped with GPU-supported hardware, the execution on a typical UAV’s limited computational resources necessitates the use of lightweight, small-sized networks. To overcome these impediments, this article presents a lightweight and attention-based approach for performing forest fire detection tasks using UAV vision data (images acquired by cameras mounted on UAVs). In this paper, we also present comprehensive research for different approaches such as transfer learning, deep CNNs, and lightweight CNNs. Among the experimented models, the attention-based EfficientNetB0 backboned model emerged as the most successful architecture for forest fire detection. With the test accuracy of , the F1-score of , the recall of , and the precision of have strongly reinforced the efficiency of the EfficientNetB0-based model in wildfire recognition. Moreover, the network has a less parameter size than the experimented networks. It proves the model’s suitability for wildfire detection with UAVs having limited hardware resources.

<p>In recent years, there have been numerous forest fires, and fire identification technology has become increasingly influential in both academic and industrial fields. At present, most automatic fire alarm systems are limited to identification by sensors such as temperature, smoke, and infrared optics. One of the existing solutions is the method of image feature extraction, which does not need to rely on specific sensors and can be easily embedded in different devices. However, this method has the disadvantage that it is difficult to extract features from image data. To attack this issue, this paper proposes a lightweight convolutional neural network for forest fire recognition. Firstly, three-channel color images of three scenes are constructed as the input of the convolutional neural network, and the initial data are pre-processed and enhanced. Secondly, a deep convolutional neural network with multiple layers of convolution and pooling layers is constructed. Finally, the Softmax function is used to classify the fire recognition scenes. The experimental results show that our approach outperforms these selected techniques in the effectiveness and accuracy.</p> <p> </p>

Exposing counterfeit perfume products is significant for protecting the legal profit of genuine perfume manufacturers and the health of perfume consumers. As a holistic solution to the problem of perfume identification (PI) using an electronic nose (EN), the methods based on convolutional neural network (CNN) simplifies the inconvenient selection of methods and parameter values, which has traditionally complicated existing combinatory methods. However, existing CNN methods that can be used for EN-based PI were designed on the premise that the CNN model can be trained with plenty of computational resources in divide-body ENs. Aiming at PI with an integrated handheld EN, a novel light-weight CNN method, namely LwCNN, is presented for being entirely conducted on a resource-constrained NVDIA Jetson nano module. LwCNN utilizes a sequenced stack of two feature flattening layers, two one-dimensional (1D) convolutional layers, a 1D max-pooling layer, a feature dropout layer, and a fully connected layer. Extensive real experiments were conducted on an integrated handheld EN to the performance of LwCNN with those of four existing benchmark methods. Experimental results show that LwCNN obtained an average identification accuracy of 98.35% with model training time of about 26 s.

This study presents a comprehensive deep learning framework for diagnosing acoustic faults in electric motors. The framework uses Mel spectrograms and a lightweight convolutional neural network (CNN). The method classifies three motor states, engine_good, engine_broken, and engine_heavyload, based on audio recordings from the IDMT-ISA-ELECTRIC-ENGINE dataset. To prevent data leakage and ensure a robust evaluation, the study employed file-level splitting, session separation, 5-fold cross-validation, and repeated trials. The raw audio signals were transformed into Mel spectrograms and processed through a CNN architecture that integrates convolutional, pooling, normalization, and dropout layers. Quantitative metrics, including THD, spectral entropy, and SNR, further characterize the acoustic distinctions between motor states. The proposed model achieved a test accuracy of 99.7%, outperforming or matching state-of-the-art approaches, such as ResNet-18, CRNN, and Transformer classifiers, as well as traditional MFCC-based baselines. Noise robustness and sensitivity analyses demonstrated stable performance under varying SNR conditions and preprocessing settings. Feature-importance maps revealed that low-frequency regions (0–40 Mel bins) were key discriminative components linked to physical fault mechanisms. Computational evaluation confirmed the model’s real-time feasibility on embedded hardware with low latency and a modest parameter count. Though primarily validated on one motor type, external-domain testing revealed strong adaptability. Future work may incorporate transfer learning or multimodal fusion. Overall, the proposed framework provides a highly accurate, interpretable, and efficient solution for real-time motor fault diagnosis and predictive maintenance in industrial environments.

The application of deep learning technology has increased rapidly in recent years. Technologies in deep learning increasingly emulate natural human abilities, such as knowledge learning, problem-solving, and decision-making. In general, deep learning can carry out self-training without repetitive programming by humans. Convolutional neural networks (CNNs) are deep learning algorithms commonly used in wide applications. CNN is often used for image classification, segmentation, object detection, video processing, natural language processing, and speech recognition. CNN has four layers: convolution layer, pooling layer, fully connected layer, and non-linear layer. The convolutional layer uses kernel filters to calculate the convolution of the input image by extracting the fundamental features. The pooling layer combines two successive convolutional layers. The third layer is the fully connected layer, commonly called the convolutional output layer. The activation function defines the output of a neural network, such as 'yes' or 'no'. The most common and popular CNN activation functions are Sigmoid, Tanh, ReLU, Leaky ReLU, Noisy ReLU, and Parametric Linear Units. The organization and function of the visual cortex greatly influence CNN architecture because it is designed to resemble the neuronal connections in the human brain. Some of the popular CNN architectures are LeNet, AlexNet and VGGNet.

The fast and accurate series arc fault (SAF) identification method and its hardware implementation are the key to the development of arc fault circuit interrupter (AFCI) or arc fault detection device (AFDD). The SAF experiments under household multi-branch circuit conditions were conducted. And a novel SAF identification model based on lightweight one-dimensional (1-D) convolutional neural network was proposed. First, the main-circuit current signal was used as the input of the model. The 1-D convolutional layers and 1-D maximum pooling layers of the model were used to extract the features of the current signal. The fully connected neural network (FCNN) was used to identify whether or not there is a SAF in the circuit and determine the branch-circuit where the fault is located. Second, the second to fourth standard convolutional layers of the model were improved by using depthwise separable convolution, and the batch normalization layers were added to the model, so as to realize the optimal design of the model. Finally, the model was deployed to an embedded device and its performance was tested. When the sampling frequency is higher than 5 kHz, the accuracy of fault identification and fault line selection of the model in the embedded device is higher than 98.05% and 99.11%, respectively. The average runtime of single identification is 5.26ms. It meets the technical requirements of household AFCI or AFDD.

Lamb waves at ultrasonic frequencies travel with little attenuation in thin elastic plates, and we demonstrate their use in pulse-echo behavior to monitor plate integrity. We envision using a single PZT wafer-type transducer to generate waves and to receive reflections from distant flaw or boundary locations. However, Lamb waves generally have multiple modes, each of them highly dispersive, and in consequence pulse dispersion can become pronounced and can make difficult or impossible the interpretation of pulse-echo responses. We show that selective generation of the S0 wave will overcome those difficulties; therefore, selection of transducer dimensions and pulse characteristics to achieve selective generation should be considered mandatory for most intended applications. We first review the work of others identifying a basic relationship between transducer dimension and excitation frequency for selective generation of the S0 wave. We then summarize our extensive experimental studies of wafer-type transducers with particular attention to S0 and A0 mode behavior, both in transmission and reception. We next report our two-dimensional finite element simulation of the same problem performed in FEMLAB, requiring transient simulation of the coupled electromechanical problem. We simulate the piezoelectric response of the wafer-type transducer coupled to the elastic plate, both as transmitter and receiver, as well as the development of Lamb waves within the source region and their subsequent propagation along the plate. Simulations illustrate the development and separation of the S0 and A0 modes and reproduce the expected group velocities and dispersion behavior. We show good agreement between our experiments and our simulations regarding S0 mode behavior, and we summarize the results to guide a designer in choosing transducer dimensions. In particular, good selectivity between the S0 and A0 mode generation can be obtained with appropriate choice of transducer size and center frequency. We show the results of experiments on an aluminum plate in which excitation of a single PZT wafer-type transducer at 6.5 V (peak-to-peak) produces reflected signals of ample strength (tens of mV) from distant boundaries and from partial thickness flaws.

Optimizing convolutional neural networks on multi-core vector accelerator

In deep convolutional neural networks (DCNNs), model size and computation complexity are two important factors governing throughput and energy efficiency when deployed to hardware for inference. Recent works on compact DCNNs as well as pruning methods are effective, yet with drawbacks. For instance, more than half the size of all MobileNet models lies in their last two layers, mainly because compact separable convolution (CONV) layers are not applicable to their last fully connected (FC) layers. Also, in pruning methods, the compression is gained at the expense of irregularity in the DCNN architecture, which necessitates additional indexing memory to address nonzero weights, thereby increasing memory footprint, decompression delays, and energy consumption. In this article, we propose cyclic sparsely connected (CSC) architectures, with memory/computation complexity of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mathcal {O}(N\log {}N)$ </tex-math></inline-formula> , where <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$N$ </tex-math></inline-formula> is the number of nodes/channels given a DCNN layer that, contrary to compact depthwise separable layers, can be used as an overlay for both FC and CONV layers of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mathcal {O}(N^{2})$ </tex-math></inline-formula> . Also, contrary to pruning methods, CSC architectures are structurally sparse and require no indexing due to their cyclic nature. We show that both standard convolution and depthwise convolution layers are special cases of the CSC layers, whose mathematical function, along with FC layers, can be unified into one single formulation and whose hardware implementation can be carried out under one arithmetic logic component. We examine the efficacy of the CSC architectures for compression of LeNet, AlexNet, and MobileNet DCNNs with precision ranging from 2 to 32 bits. More specifically, we surge upon the compact 8-bit quantized 0.5 MobileNet V1 and show that by compressing its last two layers with CSC architectures, the model is compressed by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 1.5\times $ </tex-math></inline-formula> with a size of only 873 kB and little accuracy loss. Finally, we design a configurable hardware that implements all types of DCNN layers including FC, CONV, depthwise, CSC-FC, and CSC-CONV indistinguishably within a unified pipeline. We implement the hardware on a tiny Xilinx field-programmable gate array (FPGA) for total on-chip processing of the compressed MobileNet that, compared to the related work, has the highest Inference/J while utilizing the smallest FPGA.

Fine-tuning of pretrained convolutional neural networks (CNNs) has been proven to be an effective strategy for remote sensing image scene classification, particularly when a limited number of labeled data sets are available for training purposes. However, such a fine-tuning process often needs that the input images are resized into a fixed size to generate input vectors of the size required by fully connected layers (FCLs) in the pretrained CNN model. Such a resizing process often discards key information in the scenes and thus deteriorates the classification performance. To address this issue, in this paper, we introduce a scale-free CNN (SF-CNN) for remote sensing scene classification. Specifically, the FCLs in the CNN model are first converted into convolutional layers, which not only allow the input images to be of arbitrary sizes but also retain the ability to extract discriminative features using a traditional sliding-window-based strategy. Then, a global average pooling (GAP) layer is added after the final convolutional layer so that input images of arbitrary size can be mapped to feature maps of uniform size. Finally, we utilize the resulting feature maps to create a new FCL that is fed to a softmax layer for final classification. Our experimental results conducted using several real data sets demonstrate the superiority of the proposed SF-CNN method over several well-known classification methods, including pretrained CNN-based ones.

The key to preventing the COVID-19 is to diagnose patients quickly and accurately. Studies have shown that using Convolutional Neural Networks (CNN) to analyze chest Computed Tomography (CT) images is helpful for timely COVID-19 diagnosis. However, personal privacy issues, public chest CT data sets are relatively few, which has limited CNN's application to COVID-19 diagnosis. Also, many CNNs have complex structures and massive parameters. Even if equipped with the dedicated Graphics Processing Unit (GPU) for acceleration, it still takes a long time, which is not conductive to widespread application. To solve above problems, this paper proposes a lightweight CNN classification model based on transfer learning. Use the lightweight CNN MobileNetV2 as the backbone of the model to solve the shortage of hardware resources and computing power. In order to alleviate the problem of model overfitting caused by insufficient data set, transfer learning is used to train the model. The study first exploits the weight parameters trained on the ImageNet database to initialize the MobileNetV2 network, and then retrain the model based on the CT image data set provided by Kaggle. Experimental results on a computer equipped only with the Central Processing Unit (CPU) show that it consumes only 1.06 s on average to diagnose a chest CT image. Compared to other lightweight models, the proposed model has a higher classification accuracy and reliability while having a lightweight architecture and few parameters, which can be easily applied to computers without GPU acceleration. Code:github.com/ZhouJie-520/paper-codes.

The convolutional layers of standard convolutional neural networks (CNNs) are equivariant to translation. However, the convolution and fully-connected layers are not equivariant or invariant to other affine geometric transformations. Recently, a new class of CNNs is proposed in which the conventional layers of CNNs are replaced with equivariant convolution, pooling, and batch-normalization layers. The final classification layer in equivariant neural networks is invariant to different affine geometric transformations such as rotation, reflection and translation, and the scalar value is obtained by either eliminating the spatial dimensions of filter responses using convolution and down-sampling throughout the network or average is taken over the filter responses. In this work, we propose to integrate the orthogonal moments which gives the high-order statistics of the function as an effective means for encoding global invariance with respect to rotation, reflection and translation in fully-connected layers. As a result, the intermediate layers of the network become equivariant while the classification layer becomes invariant. The most widely used Zernike, pseudo-Zernike and orthogonal Fourier-Mellin moments are considered for this purpose. The effectiveness of the proposed work is evaluated by integrating the invariant transition and fully-connected layer in the architecture of group-equivariant CNNs (G-CNNs) on rotated MNIST and CIFAR10 datasets.

This paper presents a technique for estimating thermal-induced stress in constrained metallic plates using the group velocity of Lamb waves, the accuracy of which is crucial for assessing the structural integrity and serviceability of metallic structures. However, without the ability to gauge the current stress levels, obtaining such measurements is technically challenging. To overcome this, we propose a thermal stress estimation technique that uses changes in the group velocities of the fundamental symmetric (S0) and antisymmetric (A0) Lamb wave modes caused by thermal and stress variations. First, this study introduces a theoretical-based zero-crossing algorithm to measure the group velocities of S0 and A0 Lamb wave modes. Next, leveraging the acoustoelastic coefficients corresponding to the S0 and A0 modes, which are determined before the plate’s installation, this study generates the lines depicting the changes in group velocity induced by temperature variations (CT) for both the S0 and A0 modes. These CT lines are derived from the lines illustrating changes in group velocity due to thermal stress variations (CTS), which are obtained after plate installation. Ultimately, the generated CT lines can be used to estimate thermal stress throughout the entirety of the plate’s operational life span by isolating the distinct stress variation effects from the CTS lines. The numerical validation results show favorable accuracy in thermal stress estimation in a constrained plate subjected to temperature variation using both S0 and A0 Lamb wave modes, with average errors of 0.63 % and 0.91 %, respectively.

In this paper, we investigated several lightweight convolutional neural network (CNN) classifiers for person re-identification problems. Person re-identification can be described as finding a person by performing a query with a specific person image on some defined gallery of people images. We construct the lightweight CNN classifier by utilizing the lightweight residual network architectures originally used for the CIFAR dataset. The goal is to design a lightweight CNN classifier with a maximum of 3 million number of parameters that produce good accuracy on person re-identification problems. The last layer of the residual network is detached and changed by two fully-connected layers, one for learning the discriminant features, and the other fully-connected layer for calculated the final classification score in the training process. Experiment results show that the ensemble of lightweight residual networks achieved a good performance on Market-1501 (rank-1 accuracy of 89.46% with mAP of 84.48% for single-query and rank-1 accuracy of 92.23% with mAP of 88.07% for multi-query). Although the lightweight residual network does not achieve state-of-the-art performance, our analysis shows that the lightweight residual network has higher information density comparing with other state-of-the-art models. More information density means that the classifier is more efficient in terms of performance with the number of parameters in the classifier. Implementation of these experiments are available at https://github.com/rezafuad/ person-reid-lightweight-residualnet.