Multi-scale Convolutional Neural Network Research Articles

Adaptive multi-scale attention convolution neural network for cross-domain fault diagnosis

This paper proposes a novel approach named adaptive multi-scale attention convolution neural network (AmaCNN) to accurately detect cross-domain faults with very few labelled data. In AmaCNN, multi-scale feature fusion CNN (MSFFCNN) with a multi-level attention scheme (MLAS) extracts multi-scale less-noise features from source and target domains. Considering the domain shift and semantic difference in the two domain features, a cross-domain adaption (CDA) scheme is applied. Significantly, the extracted domain features are measured with correlation alignment (CORAL) distance to minimize the domain shift first. Then, semantic alignment (SA) loss aligns and separates domain-invariant features point-by-point. Therefore, the proposed AmaCNN could learn rich multi-scale, less-noise, domain-invariant, and semantic-alignment features using limited training samples to detect cross- fault accurately. The experimental results on three real data sets confirmed its priority and reliability. Besides, the in-depth analysis has confirmed each component’s effectiveness and CDA’s good generality.

Modeling the 4D discharge of lithium-ion batteries with a multiscale time-dependent deep learning framework

The lithium-ion battery (LIB) field is moving towards the direction of investigating spatially resolved physical phenomena in the 3D porous microstructure of electrodes. These pore-scale simulations give new insights into the local dynamics of lithiation/de-lithiation and charge transport. Nevertheless, the computational time of these simulations limits the integration of these models in optimization workflows of cycling conditions or electrode manufacturing processes.Machine learning models present a way of assessing in real-time the performance of materials. While several successful techniques for replicating simulations with machine learning have been proposed, this case study presents a more demanding problem, due to the necessity of understanding the behavior of heterogeneous 3D local data, as it evolves in time: this poses both a scientific and a technical challenge.To this end, we propose an autoregressive multiscale convolutional neural network model to predict relevant quantities at the pore-scale in the solid phase: the lithium concentration (in the active material) and potential (in the active material and carbon binder). These are ultimately used to reconstruct the battery discharge curve. 3D images of the electrode microstructures are the input to the network, trained with a dataset of finite element method simulations to predict the discharge behavior of the cathode side in lithium ion batteries.We propose this machine learning model as a proof-of-concept of the applicability of multiscale networks for time-dependent physics problems. The trained model exhibits very high accuracy (with errors lower than 2%) in forecasting the discharge behavior of new unseen cathodes.

Eff-PCNet: An Efficient Pure CNN Network for Medical Image Classification

With the development of deep learning, convolutional neural networks (CNNs) and Transformer-based methods have become key techniques for medical image classification tasks. However, many current neural network models have problems such as high complexity, a large number of parameters, and large model sizes; such models obtain higher classification accuracy at the expense of lightweight networks. Moreover, such larger-scale models pose a great challenge for practical clinical applications. Meanwhile, Transformer and multi-layer perceptron (MLP) methods have some shortcomings in terms of local modeling capability and high model complexity, and need to be used on larger datasets to show good performance. This makes it difficult to utilize these networks in clinical medicine. Based on this, we propose a lightweight and efficient pure CNN network for medical image classification (Eff-PCNet). On the one hand, we propose a multi-branch multi-scale CNN (M2C) module, which divides the feature map into four parallel branches along the channel dimensions by a certain scale factor and carries out a deep convolution operation using different scale convolution kernels, and this multi-branch multi-scale operation effectively replaces the large kernel convolution. This multi-branch multi-scale operation effectively replaces the large kernel convolution. It reduces the computational cost of the module while fusing the feature information between different channels and thus obtains richer feature information. Finally, the four feature maps are then spliced along the channel dimensions to fuse the multi-scale and multi-dimensional feature information. On the other hand, we introduce the structural reparameterization technique and propose the structural reparameterized CNN (Rep-C) module. Specifically, it utilizes multiple linear operators to generate different feature maps during the training process and fuses all the participants into one through parameter fusion to achieve fast inference while providing a more effective solution for feature reuse. A number of experimental results show that our Eff-PCNet performs better than current methods based on CNN, Transformer, and MLP in the classification of three publicly available medical image datasets. Among them, we achieve 87.4% Acc on the HAM10000 dataset, 91.06% Acc on the SkinCancer dataset, and 97.03% Acc on the Chest-Xray dataset. Meanwhile, our approach achieves a better trade-off between the number of parameters; computation; and other performance metrics as well.

Swin–MRDB: Pan-Sharpening Model Based on the Swin Transformer and Multi-Scale CNN

Pan-sharpening aims to create high-resolution spectrum images by fusing low-resolution hyperspectral (HS) images with high-resolution panchromatic (PAN) images. Inspired by the Swin transformer used in image classification tasks, this research constructs a three-stream pan-sharpening network based on the Swin transformer and a multi-scale feature extraction module. Unlike the traditional convolutional neural network (CNN) pan-sharpening model, we use the Swin transformer to establish global connections with the image and combine it with a multi-scale feature extraction module to extract local features of different sizes. The model combines the advantages of the Swin transformer and CNN, enabling fused images to maintain good local detail and global linkage by mitigating distortion in hyperspectral images. In order to verify the effectiveness of the method, this paper evaluates fused images with subjective visual and quantitative indicators. Experimental results show that the method proposed in this paper can better preserve the spatial and spectral information of images compared to the classical and latest models.

A novel hybrid heuristic adopted ensemble of deep learning models for COVID-19 detection framework using CT and X-Ray images

ABSTRACT In 2019, Corona Virus Disease (COVID)-19 has created an important impact on people’s health and economy because of its rapid spreading. Therefore, the earlier detection and diagnosis for handling the further spread is regarded as a significant factor. Computed Tomography (CT) images of the lungs are utilised for the detection process of COVID-19 infection. Chest X-Ray (CXR) and CT serves as the major source for the identification of COVID-19. Some of the classifiers have shown promising outcomes as well as better performance over various medical imaging applications and computer vision processes. Due COVID pandemic, researchers have utilised standard techniques for accurately identifying the coronavirus infection in lung images. Here, a new COVID-19 detection approach has been implemented with the adoption of ensemble classifiers to recognise the occurrence of COVID-19 infection at the starting phase. The required CT and the CXR images are garnered from publicly presented online sources. Meanwhile, the acquired images undergo the process of pre-processing using filtering methods and histogram equalisation techniques in order to ignore the unnecessary artifacts are presented in the images. Then, the detection of COVID-19 is made possible through Ensemble Classifiers (EC) such as Multi-scale Convolutional Neural Networks (CNN), Residual Attention Networks (RAN), Visual Geometry Group (VGG), and Long Short Term Memory (LSTM), where the variables within the machine learning strategies are tuned by Modified Mexican Axolotl Shuffled Frog-Leaping Optimization (MMASFLO). The efficacy of the newly developed ensemble classifier-based COVID-19 detection approaches will be tested through various benchmark datasets, and is tested with several existing approaches and conventional COVID-19 detection approaches.

SO-IMCKD processed signal improving MSCNN model’s fault diagnosis accuracy for drilling pump fluid end

Drilling pump is the ‘heart’ of drilling construction. The key to accurate fault diagnosis is to extract useful fault features from noisy raw signals. In order to improve the accuracy of fault diagnosis of drilling pump fluid end, this paper proposes a fault diagnosis method based on multi-scale convolutional neural network (MSCNN) combined with the snake optimization optimized maximum correlation kurtosis deconvolution (SO-IMCKD). First, the SO algorithm is employed to optimize the filter length and the shift number of IMCKD to process the raw signal, enhancing the fault features from the raw signal. Second, the continuous wavelet transform is used to convert the enhanced signals into time-frequency images which are input into an established MSCNN to extract the fault feature more effectively. Finally, by changing the training batchsize of the MSCNN model, the identification effect of the model to the normal state, minor damage, and serious damage of the fluid end is analyzed. The identification of nine states of the fluid end is successfully carried out, with an average diagnostic accuracy of 99.93%. Moreover, the adaptability of the proposed method is verified with the Mechanical Failure Prevention Technology Association dataset. The method has high accuracy and good adaptability, which has desired prospect for drilling pump fault diagnosis and bearing fault diagnosis.

Denoising method for weak velocity test signals under laser-induced shock waves based on attention-guided multi-scale convolutional neural network

In the research of characteristics of laser-induced shock waves via the laser interference velocity detection method, noise reduction for weak echo signals still relies on the manual setting of filtering parameters, alongside the problem of the loss of useful signals caused by global filtering and excessive filtering. To solve these problems, a multi-scale network for extraction of sparse features, in combination with the ECA channel module, was established in this study to acquire the features of multi-dimensional and multi-fine-granularity noise signals in a self-adaptive manner. By integrating various channel information of the deep/shallow features through skip connection, the deep features were fully observed without ignoring the impacts of the shallow features. Finally, the attention mechanism was adopted to adjust the weight of the features at different locations in the feature space to acquire an accurate representation of stochastic noise, and noise reduction for weak signals was achieved by reconstructing the module and removing the learned noise features from the noisy data. According to the simulation and test results, the method proposed in this study helped extract the noise information from weak signals in an accurate manner; unlimited by the forms and sizes of the noise and the working conditions and independent from manual setting of parameters, it eliminated the blind-source noise from weak echo signals effectively; with this method, the initial position and the coupled peak value were extracted accurately from the weak signals. In terms of the effects of the method, the phenomena of “peak clipping” and “breakpoint” caused by excessive filtering were avoided; compared to global filtering, time-frequency domain filtering and other methods, the method proposed outperformed the conventional ones for a peak value error of only 0.19%; as for the denoising velocity, simultaneous input of multiple pieces of data as well as simultaneous output of denoising outcomes in a real time manner were also achieved under this method, with the averaged processing time for a single piece of data (0.051 s) far superior to the result under other methods.

A method to detect internal leakage of hydraulic cylinder by combining data augmentation and multiscale residual CNN

AbstractDeveloping a method to detect internal leakage in hydraulic cylinder, which is used for Electro‐Hydrostatic Actuators (EHA), is important to prevent serious malfunctions for aircrafts. At present, the internal leakage in an EHA cannot be accurately detected only using operational data. This paper proposed a convolutional neural networks (CNN) based method to detect internal leakage in hydraulic cylinder according to the relationship between operational state parameters of EHA and leakage in the hydraulic cylinder. A method was presented to align multi‐source signals with different forms by using the motor current as a benchmark. Because the number of monitoring signals are relatively small, a feedforward neural network (FFNN) based data augment method is proposed to increase parameters of input data set. A general method on how to detect internal leakage by combining signals alignment, data augmentation and multiscale residual CNN was proposed. The experimental results show that the proposed method can be used to accurately detect internal leakage in a hydraulic cylinder operating under non‐stationary load and velocity conditions, and the detection accuracy reached 99.8%.

Real-Time Defect Detection Scheme Based on Deep Learning for Laser Welding System

Laser welding, as an important material processing technology, has been widely used in various fields of industry. In most industrial welding production and processing, high precision is required for welding parameters and fixed work pieces. However, in the process of laser welding, serious heat transfer effect will bring unpredictable welding deviations, and even a small deviation will lead to serious welding defects, which will affect the quality of the welded products. Traditional non-destructive testing methods have been widely used, but they have been proved to have some limitations. Existing laser welding defect detection schemes are mainly focused on the detection of post-weld defects, which requires a large amount of data, and the real-time detection cannot be guaranteed. In this paper, we propose a data acquisition system for collecting changes in physical characteristics during laser welding with the aids of multiple sensors. Based on the data originating from sensors’ system, an efficient laser welding defect detection model has been designed and investigated based on the MSCNN (multi-scale convolutional neural network), BiLSTM (bidirectional long short-term memory), and AM (attention mechanism). The final proposed MSCNN-BiLSTM-AM fusion detection model can achieve 99.38% detection accuracy, which make the laser welding system more efficient and more suitable.

Prediction of the tensile strength of friction stir welded joints based on one-dimensional convolutional neural network

Friction stir welding (FSW) is a complex thermo-mechanical coupling process. Tensile strength is an important evaluation index of the mechanical properties of welded joints. How to realize the real-time prediction of tensile strength of the friction stir welded joints to reflect the dynamic change of welding state is a problem in the field. To solve this problem, this paper presents a multi-scale one-dimensional convolutional neural network (Multi-scale 1D CNN) prediction model using time series data of temperature and axial force as inputs to realize the online prediction of tensile strength of welded joints. Firstly, FSW experiments are carried out to obtain time series data of temperature and axial force. Tensile strength values of the welded joints is obtained by tensile tests. The time series data and tensile strength values are fused as a dataset. Then Multi-scale 1D CNN, traditional 1D CNN and Multi-channel 1D CNN prediction models are established and trained with the dataset, respectively. Finally, by comparing the prediction performance of the three models, Multi-scale 1D CNN is proved to be more suitable for analyzing time series data to feedback the dynamic change of tensile strength of the joints during welding.

Multi-Scale Orthogonal Model CNN–Transformer for Medical Image Segmentation

Because of the limitations of convolution kernel, the traditional image segmentation network is not sufficient to obtain the context information, but the image segmentation task is very dependent on the context information. Transformer’s linear input can just get enough context information. In this paper, we propose a transformer segmentation network hyperfusion transformer based on a pyramid structure. First, the model divides the single-scale coding form into several-different-scale coding forms, and then fuses the decoding results. Second, in order to ensure the specificity of the output characteristics of each branch, we orthogonalize the results of a variety of different scales. By orthogonalizing in pairs, we can ensure that the results obtained by different branches are not completely similar to a certain extent, and reduce the redundancy of branch information. On the two datasets, the method in this paper surpasses a variety of classical models under multiple evaluation indexes, confirming that it is an effective segmentation method.

Multi-task learning for the bearing based on a one-dimensional convolutional neural network with attention guidance mechanism and multi-scale feature extraction

Fault type and fault degree identification are the main aim in the bearing multi-task learning. However, a large number of on-site accidents have shown that the bearing working condition plays an important role in bearing service life and fault diagnosis. In current studies, the bearing working condition identification task is often used for auxiliary tasks and is easily ignored. Thus, this paper studies the bearing multi-task learning, which regards the working condition identification task as an equally important task. However, simply adding the working condition identification task to the frequently used multi-task model will lead to a reduction in the overall performance of the network. To solve the network performance degradation problem, a succinct and effective multi-task one-dimensional convolutional neural network with attention guidance mechanism and multi-scale feature extraction (MAM-1DCNN) is proposed. Firstly, the time-series signal is selected as the input of the MAM-1DCNN model. Secondly, the shared network of the MAM-1DCNN model applies a double-layer multi-scale convolutional neural network structure to extract more complete information. Finally, the MAM-1DCNN applies an improved attention guidance mechanism to enhance the feature application ability of different branch tasks. Through two general bearings datasets, this paper verifies the effectiveness and generalisation ability of the MAM-1DCNN model.

Pipeline leakage aperture identification method based on pseudolabel learning

Aiming at the problem of insufficient label data in the pipeline leak detection field, this paper proposes a pseudolabel (PL) adaptive learning method based on multiscale convolutional neural network (MSCNN) with the idea of transfer learning for pipeline leak aperture identification. First, the convolutional and pooling layers for transfer learning feature extraction are improved by using a dual-channel MSCNN. Second, the KL divergence function after dimensionality reduction is used to calculate the distribution distance between the source domain and the target domain to improve the robustness of distribution alignment in high-noise environments. In addition, considering the interference of PL noise, this paper develops a pseudolabel (PL) dynamic threshold to achieve the purpose of PL adaptive updating. Compared with the fixed threshold, the improved PL learning (PLL) can effectively improve the prediction accuracy of the model. The effectiveness of the method proposed in this paper is verified by predicting pipeline leakage conditions at different distances and under different pressures. The comparative analysis results show that the method in this paper is superior to other transfer learning methods in terms of prediction accuracy, stability, and convergence speed.

An Improved Multi-Scale Feature Fusion for Skin Lesion Segmentation

Accurate segmentation of skin lesions is still a challenging task for automatic diagnostic systems because of the significant shape variations and blurred boundaries of the lesions. This paper proposes a multi-scale convolutional neural network, REDAUNet, based on UNet3+ to enhance network performance for practical applications in skin segmentation. First, the network employs a new encoder module composed of four feature extraction layers through two cross-residual (CR) units. This configuration allows the module to extract deep semantic information while avoiding gradient vanishing problems. Subsequently, a lightweight and efficient channel attention (ECA) module is introduced during the encoder’s feature extraction stage. The attention module assigns suitable weights to channels through attention learning and effectively captures inter-channel interaction information. Finally, the densely connected atrous spatial pyramid pooling module (DenseASPP) module is inserted between the encoder and decoder paths. This module integrates dense connections and ASPP, as well as multi-scale information fusion, to recognize lesions of varying sizes. The experimental studies in this paper were constructed on two public skin lesion datasets, namely, ISIC-2018 and ISIC-2017. The experimental results show that our model is more accurate in segmenting lesions of different shapes and achieves state-of-the-art performance in segmentation. In comparison to UNet3+, the proposed REDAUNet model shows improvements of 2.01%, 4.33%, and 2.68% in Dice, Spec, and mIoU metrics, respectively. These results suggest that REDAUNet is well-suited for skin lesion segmentation and can be effectively employed in computer-aided systems.

Lightweight single-image super-resolution via multi-scale feature fusion CNN and multiple attention block

Lightweight single-image super-resolution via multi-scale feature fusion CNN and multiple attention block

Production quality prediction of multistage manufacturing systems using multi-task joint deep learning

A multistage manufacturing system with multiple manufacturing stages is the key and main production mode for enterprises to achieve lean production. Due to the variation propagation between stages and multiple related quality prediction tasks, it is difficult to accurately predict the quality of multistage manufacturing systems with multiple tasks. Traditional single stage and single task quality prediction methods permit the multi stages and multiple tasks separately, which ignore multi-stage effects or the quality-related relationship between multiple quality output indicators and reduce the efficiency of quality prediction. In this paper, a production quality prediction framework based on multi-task joint deep learning is proposed to simultaneously evaluate the multi-task quality of all stages in a multistage manufacturing system. To be specific, variation propagation cumulative impact between multiple manufacturing stages is innovatively expressed by designing a multi-scale convolutional neural network with control gates (MCNN-CG) to extract and propagate data features. Production quality with multi-tasks at all stages is jointly predicted by designing a multi-layer multi-gate mixture-of-experts multi-task (ML-MMoE) model with reducing multi-task predictive loss simultaneously. The soft parameter-sharing strategy and multi-gate attention strategy are separately designed to ensure information sharing while learning personalized features of tasks to improve quality prediction accuracy of each task. In addition, a loss function based on homoscedastic uncertainty and regularization is designed to automatically learn the weight between multi-stage and multi-task losses. Experiments on multistage assembly test data of an inertial navigation manufacturing system show that the proposed method performs better than traditional models. Compared to the single-stage model, the proposed multistage model has an average improvement of 8.5%, 20.0% and 23.3% in R2, MAE and RMSE respectively in the second stage. Compared with the traditional multi-stage model, the proposed model has an average improvement of 1.7%, 6.2% and 9.8% in R2, MAE and RMSE respectively in the second stage.

TSception: Capturing Temporal Dynamics and Spatial Asymmetry From EEG for Emotion Recognition

The high temporal resolution and the asymmetric spatial activations are essential attributes of electroencephalogram (EEG) underlying emotional processes in the brain. To learn the temporal dynamics and spatial asymmetry of EEG towards accurate and generalized emotion recognition, we propose TSception, a multi-scale convolutional neural network that can classify emotions from EEG. TSception consists of dynamic temporal, asymmetric spatial, and high-level fusion layers, which learn discriminative representations in the time and channel dimensions simultaneously. The dynamic temporal layer consists of multi-scale 1D convolutional kernels whose lengths are related to the sampling rate of EEG, which learns the dynamic temporal and frequency representations of EEG. The asymmetric spatial layer takes advantage of the asymmetric EEG patterns for emotion, learning the discriminative global and hemisphere representations. The learned spatial representations will be fused by a high-level fusion layer. Using more generalized cross-validation settings, the proposed method is evaluated on two publicly available datasets DEAP and MAHNOB-HCI. The performance of the proposed network is compared with prior reported methods such as SVM, KNN, FBFgMDM, FBTSC, Unsupervised learning, DeepConvNet, ShallowConvNet, and EEGNet. TSception achieves higher classification accuracies and F1 scores than other methods in most of the experiments. The codes are available at https://github.com/yi-ding-cs/TSception

Time–Frequency Multiscale Convolutional Neural Network for RF-Based Drone Detection and Identification

Time–Frequency Multiscale Convolutional Neural Network for RF-Based Drone Detection and Identification

Multi-scale convolutional recurrent neural network for psychiatric disorder identification in resting-state EEG.

Accurate classification based on affordable objective neuroimaging biomarkers are important steps toward designing individualized treatment. In this work, we investigated a deep learning classification model, multi-scale convolutional recurrent neural network (MCRNN), to explore psychiatric disorder-related biomarkers by leveraging the spatiotemporal information of resting-state EEG (rsEEG) using a multiple psychiatric disorder database containing 327 individuals diagnosed with schizophrenia, bipolar, major depressive disorders, and healthy controls. All subjects were mapped to a shared low-dimensional subspace for intuitively interpreting the inter-relationship and separation of psychiatric disorders. Psychiatric disorders were identified using rsEEG with high accuracy ranged from 78.6 to 91.3% in patient vs. controls two-class classification, and 68.2% in four-class classification. The control-to-schizophrenia trajectory interpretated by the model was consistent with the disease severity in clinical observation. The MsRNN demonstrated a capability in extracting discriminative rsEEG biomarkers for psychiatric disorder classification, indicating its potential to facilitate our understanding of psychiatric disorders and monitoring interventions.

Learning Similarity Metrics for Volumetric Simulations with Multiscale CNNs

Simulations that produce three-dimensional data are ubiquitous in science, ranging from fluid flows to plasma physics. We propose a similarity model based on entropy, which allows for the creation of physically meaningful ground truth distances for the similarity assessment of scalar and vectorial data, produced from transport and motion-based simulations. Utilizing two data acquisition methods derived from this model, we create collections of fields from numerical PDE solvers and existing simulation data repositories. Furthermore, a multiscale CNN architecture that computes a volumetric similarity metric (VolSiM) is proposed. To the best of our knowledge this is the first learning method inherently designed to address the challenges arising for the similarity assessment of high-dimensional simulation data. Additionally, the tradeoff between a large batch size and an accurate correlation computation for correlation-based loss functions is investigated, and the metric's invariance with respect to rotation and scale operations is analyzed. Finally, the robustness and generalization of VolSiM is evaluated on a large range of test data, as well as a particularly challenging turbulence case study, that is close to potential real-world applications.