Multi-scale Convolutional Layer Research Articles

ConvMedSegNet: A multi-receptive field depthwise convolutional neural network for medical image segmentation

In order to achieve highly precise medical image segmentation, this paper presents ConvMedSegNet, a novel convolutional neural network designed with a U-shaped architecture that seamlessly integrates two crucial modules: the multi-receptive field depthwise convolution module (MRDC) and the guided fusion module (GF). The MRDC module’s primary function is to capture texture information of varying sizes through multi-scale convolutional layers. This information is subsequently utilized to enhance the correlation of global feature data by expanding the network’s width. This strategy adeptly preserves the inherent inductive biases of convolution while concurrently amplifying the network’s ability to establish dependencies on global information. Conversely, the GF module assumes responsibility for implementing multi-scale feature fusion by connecting the encoder and decoder components. It facilitates the transfer of information between features that are separated over substantial distances through guided fusion, effectively minimizing the loss of critical data. In experiments conducted on public medical image datasets such as BUSI and ISIC2018, ConvMedSegNet outperforms several advanced competing methods, yielding superior results. Additionally, the code can be accessed at https://github.com/csust-yixin/ConvMedSegNet.

MRNA-CLA: An interpretable deep learning approach for predicting mRNA subcellular localization

Messenger RNA (mRNA) is vital for post-transcriptional gene regulation, acting as the direct template for protein synthesis. However, the methods available for predicting mRNA subcellular localization need to be improved and enhanced. Notably, few existing algorithms can annotate mRNA sequences with multiple localizations. In this work, we propose the mRNA-CLA, an innovative multi-label subcellular localization prediction framework for mRNA, leveraging a deep learning approach with a multi-head self-attention mechanism. The framework employs a multi-scale convolutional layer to extract sequence features across different regions and uses a self-attention mechanism explicitly designed for each sequence. Paired with Position Weight Matrices (PWMs) derived from the convolutional neural network layers, our model offers interpretability in the analysis. In particular, we perform a base-level analysis of mRNA sequences from diverse subcellular localizations to determine the nucleotide specificity corresponding to each site. Our evaluations demonstrate that the mRNA-CLA model substantially outperforms existing methods and tools.

Multi scale encoder-decoder network with Time Frequency Attention and S-TCN for single channel speech enhancement

The goal of speech enhancement is to restore clean speech in noisy environments. Acoustic scenarios with low signal-to-noise ratios (SNR) make it quite challenging to extract the target speech from its noise. In the current study, to enhance noisy speech, we propose a feature recalibration based multi-scale convolutional encoder-decoder architecture with squeeze temporal convolutional networks (S-TCN) bottleneck. Each multi-scale convolutional layer in encoder and decoder is followed by time-frequency attention module (TFA). The recalibration based multi-scale 2D convolution layers are used to extract local and contextual information. Additionally, the recalibration network is equipped with a gating mechanism to control the flow of information among the layers, enabling weighting of the scaled features for noise suppression and speech retention. The fully connected layer (FC) in the bottleneck part of encoder-decoder contains a few neurons, which capture the global information from the multi-scale 2D convolution layer and reduce parameters. A S-TCN, inspired by the popular temporal convolutional neural network (TCNN), is inserted between the encoder and the decoder to model long-term dependencies in speech. The TFA is a highly efficient network component, that operates through two simultaneous attentions, one focused on time frames, and the other on frequency channels. These attentions work together to explicitly exploit positional information to create a two-dimensional attention map to effectively capture the significant time-frequency distribution of speech. Utilizing the common voice dataset, our proposed model consistently enhances results compared to the current benchmarks, as demonstrated by two extensively utilized objective measures PESQ and STOI. The proposed model shows significant improvements, with average PESQ and STOI scores increasing by 45.7% and 23.8% respectively for seen background noises, and by 43.5% and 21.4% for unseen background noises, when compared to the quality of noisy speech. Tests validate that the proposed approach outperforms numerous cutting-edge algorithms.

Image Classification Method Based on Multi-Scale Convolutional Neural Network

Traditional convolutional neural networks (CNNs) typically use fixed scale convolutional kernels for feature extraction when processing image classification tasks, while ignoring the multi-scale information present in the image. To overcome this limitation, we propose an algorithm based on multi-scale CNNs, which capture features at different levels by introducing convolutional kernels of different scales into the convolutional layer. In this study, we first designed a multi-scale convolutional layer consisting of multiple convolutional kernels of different scales to extract multi-scale features of the image. To further enhance classification performance, we introduced a multi-scale feature fusion module that can effectively fuse features of different scales and classify them through a fully connected layer. Then we conducted extensive experiments on several commonly used image classification datasets. The experimental results show that this network can not only effectively identify and locate hyperspectral image targets in different scenarios, but also reduce missed detections and false positives during the detection process. The average accuracy of the improved model has been improved, and the recognition accuracy of some small markers affected by external factors such as occlusion and lighting has also been improved. In addition, by comparing the detection effect of a single image, the progressiveness and anti-leakage ability of the improved model are proved. The image classification method based on multi-scale CNNs has broad application prospects in image recognition and feature extraction, and can provide valuable reference and reference for research in related fields.

A new LSTNet-based temperature prediction model for permanent magnet

Permanent magnet synchronous motors (PMSMs) can effectively protect against demagnetization by accurate permanent magnet (PM) temperature prediction; nevertheless, due to the nonlinear properties and intricate internal structure of PMSMs, accurate PM temperature prediction methods still encounter difficulties. This paper proposes a new PM temperature prediction model (LSTNet-Improved) based on long- and short-term time series network (LSTNet) to increase the prediction accuracy of PM temperature. By adding a multi-scale convolutional (MSC) layer to the convolutional neural network (CNN) layer of LSTNet, the short-term detail-dependent information between PMSM variables can be obtained by the model, and adds a CNN skip (CNN-skip) layer in place of the gate recurrent unite skip (GRU-skip) layer to find the long- and short-term local repetitive patterns between the variables, which yields additional feature information. Furthermore, a nonlinear bias term (MNB) is added to the original multi-head attention layer in order to solve the complex nonlinear relationship between multivariate time series. These layers are applied after the CNN-skip layer and gated recurrent unite (GRU) layers, respectively. This allows the model to become more robust and automatically learn to the complex nonlinear relationship between the sequences, focus on the important feature information, and lessen the impact of redundant information on the PMs’ temperature prediction task. The experimental results show that the goodness of fit (R2), root mean square error (RMSE) and mean absolute error (MAE) of the proposed model are 99.93%, 14.55% and 9.45%, respectively, and compared with LSTNet, the R2 is improved by 0.09%, the RMSE and MAE are reduced by 7.03% and 6.51%. These results confirm that the model can correctly predict the PM temperature, efficiently extract both long- and short-term patterns among the variables of the PMSM, and effectively focus on the key feature information in the intricate nonlinear series.

DSDCLNet: Dual-stream encoder and dual-level contrastive learning network for supervised multivariate time series classification

In recent years, deep learning approaches have shown remarkable advancements in Multivariate Time Series Classification (MTSC) tasks. However, the existing approaches primarily focus on capturing the long-term correlations of time series or identifying local key sequence fragments, inevitably neglecting the synergistic properties between global and local components. Additionally, most representation learning methods for MTSC rely on self-supervised learning, which limits their ability to fully exploit label information. Hence, this paper proposes a novel approach termed Dual-Stream Encoder and Dual-Level Contrastive Learning Network (DSDCLNet), which integrates a Dual-Stream Encoder (DSE) and Dual-level Contrastive Learning (DCL). First, to extract multiscale local-global features from multivariate time series data, we employ a DSE architecture comprising an Attention-Gated Recurrent Unit (AGRU) and a Dual-layer Multiscale Convolutional Neural Network (DMSCNN). Specifically, DMSCNN consists of a series of multi-scale convolutional layers and a max pooling layer. Second, to maximize the utilization of label information, a new loss function is designed, which combines classification loss, instance-level contrastive loss, and temporal-level contrastive loss. Finally, experiments are conducted on the UEA datasets and the results demonstrate that DSDCLNet achieves the highest average accuracy of 0.77, outperforming traditional approaches, deep learning approaches, and self-supervised approaches on 30, 23, and 27 datasets, respectively.

SatCoBiLSTM: Self-attention based hybrid deep learning framework for crisis event detection in social media

Social media platforms have emerged as vital sources of sharing real-time information during crises, enabling users to share critical details about disasters and ensuring timely awareness. This study addresses the daunting challenge of effectively detecting crisis events from a vast, noisy stream of short text data. To address the above-mentioned issue, we present a self-attention-based hybrid deep learning framework, SatCoBiLSTM, designed to extract hierarchical textual features and crucial information from textual data. The proposed model integrates a multi-scale convolution and BiLSTM layer to extract local and contextual relations, alongside a self-attention layer to select and focus on the most relevant crisis features. Various experiments have been conducted on three benchmark real-world crisis datasets to examine the proposed model’s performance. SatCoBiLSTM achieved an impressive F1-score of 96%, 94%, and 95% on the three public datasets. It also shows a promising improvement in the F1-score, compared to the state-of-the-art (SOTA) and baseline methods, by 1%, 1%, and 6%, respectively, showing its effectiveness in crisis event detection. An ablation analysis has been carried out to investigate the validity of each integrated layer in our model. The SatCoBiLSTM model’s capability to identify crisis-related information highlights its potential to enhance real-time awareness during disasters. Eventually, the study advances crisis event detection and lays the foundation for future research to process and handle short-textual data in noisy environments.

Reunion helper: an edge matcher for sibling fragment identification of the Dunhuang manuscript

The Dunhuang ancient manuscripts are an excellent and precious cultural heritage of humanity. However, due to their age, the vast majority of these treasures are damaged and fragmented. Faced with a wide range of sources and numerous fragments, the process of restoration generally involves two core elements: sibling fragments identification and fragment assembly. Currently, fragment restoration still heavily relies on manual labor. During the long practice, a consensus has been reached on the importance of edge features for not only assembly but also for identification. However, accurate extraction of edge features and their use for efficient identification requires extensive knowledge and strong memory. This is a challenge for the human brain. So that in previous studies, fragment edge features have been used for assembly validation but rarely for identification. Therefore, an edge matcher is proposed, working like a bloodhound, capable of “sniffing out” specific “flavors” in edge features and performing efficient sibling fragment identification accordingly, providing guidance when experts perform entity assembly subsequently. Firstly, the fragmented images are standardized. Secondly, traditional methods are used to compress the representation of fragment edges and obtain paired local edge images. Finally, these images are fed into the edge matcher for classification discrimination, which is a CNN-based pairwise similarity metric model proposed in this paper, introducing residual blocks and depthwise separable convolutions, and adding multi-scale convolutional layers. With the edge matcher, a complex matching problem is successfully transformed into a simple classification problem. In the absence of a standard public dataset, a Dunhuang manuscript fragment edge dataset is constructed. Experiments are conducted on that dataset, and the accuracy, precision, recall, and F1 scores of the edge matcher all exceeded 97%. The effectiveness of the edge matcher is demonstrated by comparative experiments, and the rationality of the method design is verified by ablation experiments. The method combines traditional methods and deep learning methods to creatively use the edge geometric features of fragments for sibling fragment identification in a natural rather than coded way, making full use of the computer’s computational and memory capabilities. The edge matcher can significantly reduce the time and scope of searching, matching, and inferring fragments, and assist in the reconstruction of Dunhuang ancient manuscript fragments.

An effective method for fault diagnosis of rotating machinery under noisy environment

Rotating machinery is widely utilized as mechanical equipment in the industrial field. However, due to the complex working conditions, the existing fault diagnosis methods have failed to address good results in practical applications. To improve the fault diagnosis performance of rotating machinery in a noisy environment, a new multi-scale convolution neural network (MSCN) based on a self-calibrating attention module is proposed. First, this thesis constructs a multi-scale convolution layer with a wide convolution kernel to form an efficient sampling structure at the filter level, which can filter out incoherent noise from the signals and extract rich features. Second, a multi-scale self-calibrating attention module is implemented with two identical self-calibrating convolutional networks to continuously focus on significant embeddings and adaptively combine information from different spatial dimensions. Third, the multi-dimensional characteristics are integrated by the feature cascade layer, and then the fault modes are identified though the classifier layer under noise. Finally, based on the Case Western Reserve University datasets and Paderborn University bearing datasets, the experimental results show that our proposed MSCN can significantly enhance the fault identification ability to rotate machinery in a noisy environment.

CapsNetYY1: identifying YY1-mediated chromatin loops based on a capsule network architecture

BackgroundPrevious studies have identified that chromosome structure plays a very important role in gene control. The transcription factor Yin Yang 1 (YY1), a multifunctional DNA binding protein, could form a dimer to mediate chromatin loops and active enhancer-promoter interactions. The deletion of YY1 or point mutations at the YY1 binding sites significantly inhibit the enhancer-promoter interactions and affect gene expression. To date, only a few computational methods are available for identifying YY1-mediated chromatin loops.ResultsWe proposed a novel model named CapsNetYY1, which was based on capsule network architecture to identify whether a pair of YY1 motifs can form a chromatin loop. Firstly, we encode the DNA sequence using one-hot encoding method. Secondly, multi-scale convolution layer is used to extract local features of the sequence, and bidirectional gated recurrent unit is used to learn the features across time steps. Finally, capsule networks (convolution capsule layer and digital capsule layer) used to extract higher level features and recognize YY1-mediated chromatin loops. Compared with DeepYY1, the only prediction for YY1-mediated chromatin loops, our model CapsNetYY1 achieved the better performance on the independent datasets (AUC > 0.99).ConclusionThe results indicate that CapsNetYY1 is an excellent method for identifying YY1-mediated chromatin loops. We believe that the CapsNetYY1 method will be used for predictive classification of other DNA sequences.

A novel bearing fault diagnosis method under small samples using time-frequency multi-scale convolution layer and hybrid attention mechanism module

Deep neural networks for bearing fault diagnosis have become the focus of research in recent years with its excellent feature extraction capability. However, the problem of diagnosis under small samples still needs to be solved in industrial applications, because bearings rarely work in the fault state in practice, resulting in the scarcity of fault data. To solve this problem, this paper proposes a new diagnosis model, a time-frequency multi-scale attention network, which structure allows the original signal and its transformed spectrum to be used as the input in parallel. A multi-scale convolutional layer is also designed to extract information from the signal at different scales to enhance the feature extraction capability of the network. In addition, a hybrid attention mechanism is added to integrate the redundant features and realize the complementarity between features. The experimental results of seven bearing diagnosis cases from two bearings show that the proposed method can achieve high diagnostic accuracy under small samples, which proves the superiority of the proposed method. The time domain signal and frequency domain signal were respectively used as input to train the model. By comparing the accuracy with the time-frequency combined signal as input, the superiority of the time-frequency domain signal as input is proved.

SCNet: A deep learning network framework for analyzing near-infrared spectroscopy using short-cut

In the past few decades, data-driven deep learning analysis has been developed and successfully applied in many domains, such as spectral signal analysis. The convolutional layer was regarded as an efficient feature extractor, and the residual learning framework was used to train the deep neural network. In this paper, by replacing identity blocks in a residual network with proposed short-cut structures, we present a new residual learning framework coupled with multiscale convolutional layers, whose advantages are discussed from the perspective of gradient descent during backpropagation and information theory. This study proved that the proposed short-cut structure can pass the loss gradient from the last layer to minimize the vanishing gradient problem. The experiments also showed that the multiscale design of our network can extract features from spectral signals more efficiently than traditional convolutional layers. The proposed new residual framework was tested on four public datasets to prove the network’s efficiency. Compared to the traditional residual network, the proposed method can decrease RMSE by 36.9% on average and the R2 reached 0.9544 on average.

Multibranch and Multiscale Dynamic Convolutional Network for Small Sample Fault Diagnosis of Rotating Machinery

Deep learning methods have been widely used in the field of fault diagnosis of rotating machinery. As one of the deep learning methods, multi-scale convolutional neural network (MSCNN) can achieve accurate fault diagnosis by extracting feature information of vibration signals at different scales. However, in practical applications, a small number of fault samples limit the performance of MSCNN. To overcome this problem, a small sample rotating machinery fault diagnosis method based on multi-branch and multi-scale dynamic convolutional network (MBSDCN) is proposed. Firstly, a feature splitting strategy is proposed, which splits a single input into multi-branch inputs to ensure that each multi-scale convolutional layer is matched to the appropriate input. Secondly, a channel reconstruction (CR) attention mechanism is designed, and CR attention mechanism can form a direct mapping of high-dimensional space with the fully connected layer through the reconstruction processing of channels, thus making the connection between feature channels closer. Finally, a novel multi-scale feature extraction model is constructed, in which multi-scale convolutional layers are applied to extract multi-branch input features, and the CR attention mechanism is used to calibrate the contribution of multi-scale convolution layers. The performance of MBSDCN is evaluated with CWRU rolling bearing dataset and UOC gearbox dataset and is compared with some advanced deep learning methods. The comparison results indicate that the proposed MBSDCN can accurately diagnose faults under small sample conditions. In addition, the experimental results in noisy environments indicate that the proposed MBSDCN has satisfactory anti-noise performance under small sample conditions and noisy environments.

Federal learning edge network based sentiment analysis combating global COVID-19

As one of the important research topics in the field of natural language processing, sentiment analysis aims to analyze web data related to COVID-19, e.g., supporting China government agencies combating COVID-19. There are popular sentiment analysis models based on deep learning techniques, but their performance is limited by the size and distribution of the dataset. In this study, we propose a model based on a federal learning framework with Bert and multi-scale convolutional neural network (Fed_BERT_MSCNN), which contains a Bidirectional Encoder Representations from Transformer modules and a multi-scale convolution layer. The federal learning framework contains a central server and local deep learning machines that train local datasets. Parameter communications were processed through edge networks. The weighted average of each participant’s model parameters was communicated in the edge network for final utilization. The proposed federal network not only solves the problem of insufficient data, but also ensures the data privacy of the social platform during the training process and improve the communication efficiency. In the experiment, we used datasets of six social platforms, and used accuracy and F1-score as evaluation criteria to conduct comparative studies. The performance of the proposed Fed_BERT_MSCNN model was generally superior than the existing models in the literature.

A Sound and Vibration Fusion Method for Fault Diagnosis of Rolling Bearings under Speed-Varying Conditions.

The fault diagnosis of rolling bearings is critical for the reliability assurance of mechanical systems. The operating speeds of the rolling bearings in industrial applications are usually time-varying, and the monitoring data available are difficult to cover all the speeds. Though deep learning techniques have been well developed, the generalization capacity under different working speeds is still challenging. In this paper, a sound and vibration fusion method, named the fusion multiscale convolutional neural network (F-MSCNN), was developed with strong adaptation performance under speed-varying conditions. The F-MSCNN works directly on raw sound and vibration signals. A fusion layer and a multiscale convolutional layer were added at the beginning of the model. With comprehensive information, such as the input, multiscale features are learned for subsequent classification. An experiment on the rolling bearing test bed was carried out, and six datasets under various working speeds were constructed. The results show that the proposed F-MSCNN can achieve high accuracy with stable performance when the speeds of the testing set are the same as or different from the training set. A comparison with other methods on the same datasets also proves the superiority of F-MSCNN in speed generalization. The diagnosis accuracy improves by sound and vibration fusion and multiscale feature learning.

Lung Disease Classification using Dense Alex Net Framework with Contrast Normalisation and Five-Fold Geometric Transformation

lung disease is one of the leading causes of death worldwide. Most cases of lung diseases are found when the disease is in an advanced stage. Therefore, the development of systems and methods that begin to diagnose quickly and prematurely plays a vital role in today's world. Currently, in detecting differences in lung cancer, an accurate diagnosis of cancer types is needed. However, improving the accuracy and reducing training time of the diagnosis remains a challenge. In this study, we have developed an automated classification scheme for lung cancer presented in histopathological images using a dense Alex Net framework. The proposed methodology carries out several phases includes pre-processing, contrast normalization, data augmentation and classification. Initially, the pre-processing step is accompanied to diminish the noisy contents present in the image. Contrast normalization has been explored to maintain the same illumination factor among histopathological lung images next to pre-processing. Afterwards, data augmentation phase has been carried out to enhance the dataset further to avoid over-fitting problems. Finally, the Dense Alex Net is utilized for classification that comprises five convolutional layers, one multi-scale convolution layer, and three fully connected layers. In evaluation experiments, the proposed approach was trained using our original database to provide rich and meaningful features. The accuracy attained by the proposed methodology is93%, which is maximum compared with the existing algorithms

Automated detection of sleep-arousal using multi-scale convolution and self-attention mechanism

In clinical, manually scoring by technician is the major method for sleep arousal detection. This method is time-consuming and subjective. This study aimed to achieve an end-to-end sleep-arousal events detection by constructing a convolutional neural network based on multi-scale convolutional layers and self-attention mechanism, and using 1 min single-channel electroencephalogram (EEG) signals as its input. Compared with the performance of the baseline model, the results of the proposed method showed that the mean area under the precision-recall curve and area under the receiver operating characteristic were both improved by 7%. Furthermore, we also compared the effects of single modality and multi-modality on the performance of the proposed model. The results revealed the power of single-channel EEG signals in automatic sleep arousal detection. However, the simple combination of multi-modality signals may be counterproductive to the improvement of model performance. Finally, we also explored the scalability of the proposed model and transferred the model into the automated sleep staging task in the same dataset. The average accuracy of 73% also suggested the power of the proposed method in task transferring. This study provides a potential solution for the development of portable sleep monitoring and paves a way for the automatic sleep data analysis using the transfer learning method.

RFNet: Fast and efficient neural network for modulation classification of radio frequency signals

Automatic Modulation Classification (AMC) is a well-known problem in the Radio Frequency (RF) domain. Solving this problem requires determining the modulation of an RF signal. Once the modulation is determined, the signal could be demodulated making it possible to analyse the signal for various purposes. Deep Neural Networks (DNNs) have recently proven to be successful in solving this problem efficiently. However, since deep networks consist of several layers resulting in a high number of trainable parameters, the hardware implementations of these solutions are resource-demanding. In order to address this challenge, we propose an efficient deep neural network referred to as RFNet to tackle the AMC problem efficiently. This network introduces the novel Multiscale Convolutional (MSC) layer to extract robust features in different resolutions. In addition, the network takes advantage of several Separable Convolution Blocks (SCB). These blocks employ pointwise and depth-wise convolutions to reduce network complexity. We further introduce RFNet+ and RFNet++ as extensions of RFNet with fewer number of parameters. These variants include fewer floating-point operations and hence a lower hardware implementation cost. Experimental results using the challenging RadioML 2018 dataset show that RFNet-32++ achieves an average classification accuracy of 56.09% over all Signal-to-Noise Ratios (SNRs) and an accuracy of 92.21% in+20dB SNR using only 3.1K parameters. The small number of parameters makes the RFNet family a promising solution for future AMC systems.

Breaking the resolution barrier: A low-to-high network for large-scale high-resolution land-cover mapping using low-resolution labels

Large-scale high-resolution land-cover mapping is a way to comprehend the Earth’s surface and resolve the ecological and resource challenges facing humanity. High-resolution (≤1 m) remotely sensed images can now be captured more easily, with wider coverage, as sensors and satellites develop. Nevertheless, the synchronous renewal of land-cover maps is still challenging when using the common land-cover mapping methods, due to the requirement for high-resolution land-cover labels. Abundant low-resolution (∼30 m) land-cover products are available for use as alternative label sources, but the resolution gap between these products and the growing volume of high-resolution imagery is a barrier yet to be overcome. In this paper, to break through this obstacle, we propose a low-to-high network (L2HNet) to automatically generate high-resolution land-cover maps from high-resolution images by taking only low-resolution land-cover products as the training labels, thus getting rid of the requirement for finely labeled samples during the large-scale map updating process. Firstly, to obtain the mapping results with rich details, we propose a resolution-preserving (RP) backbone that contains parallel multi-scale convolutional layers for extracting the high-resolution features from the images. Furthermore, to settle the label noise issue caused by the mismatched resolution, a confident area selection (CAS) module and a low-to-high (L2H) loss function, with weak and unsupervised strategies, are designed for obtaining reliable supervision information from the coarse labels. The experimental results obtained for six administrative states located in the Chesapeake Bay watershed of the United States show that L2HNet outperforms several of the state-of-the-art methods and the mainstream land-cover mapping methods in creating 1-m resolution land-cover maps by taking 30-m resolution land-cover products as training labels. As a further application, L2HNet was also adopted to produce the first 1-m resolution land-cover map with level II classification hierarchy for the entire state of Maryland in the United States, which covers an area of about 33,872 km2. The land-cover map of Maryland is publicly available at http://hipag.whu.edu.cn/L2HNet.html.

Rotating machinery fault diagnosis based on transfer learning and an improved convolutional neural network

Deep-learning-based methods have been widely used for rotating machinery fault diagnosis. However, they exhibit poor performance due to the severe data distribution difference under variable working conditions. Therefore, we first develop an improved convolutional neural network consisting of a multi-scale convolutional layer (MSC), a channel attention layer (CA), and an inception network structure (INS). Compared with other models, our model has strong feature extraction ability, fewer parameters and less training costs. Subsequently, based on transfer learning (TL), we propose the MSC-CA-INS-TL method. To improve the model’s generalization ability, we propose an appropriate fine-tuning strategy to coordinate with the model and pay attention to the accuracy of both source and target domains during migration. Bearing datasets and gear experimental platforms are used to verify the proposed method, and high fault diagnosis accuracy and stability are achieved under variable working conditions and small samples.