Improved deep residual shrinkage network
By introducing parallel space and channel attention mechanisms into the deep residual shrinking network, combined with the soft threshold function, the problem of extracting fault features in the noise background is solved, and the accuracy and robustness of mechanical fault diagnosis is significantly improved.
Patent Information
- Application Number
- CN202411754354.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-02
- Publication Date
- 2025-03-21
AI Technical Summary
Extracting effective fault characteristics in a noise background is a key problem, and prior art is difficult to effectively diagnose mechanical faults in high noise environments.
An improved deep residual shrinking network (DRSN-SCW) is proposed, which combines the parallel spatial attention mechanism and channel attention mechanism to automatically select the characteristics with the greatest diagnostic value through soft threshold functions to suppress noise interference.
该网络在高噪声环境下显著提高了故障特征的提取能力和诊断准确率,减少了对手动特征工程的依赖,适用于复杂工业场景。
Smart Images
Figure CN119670817A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of artificial intelligence algorithms, and in particular relates to a deep residual shrinkage network in the field of fault diagnosis. Background Art
[0002] With the vigorous development of modern industry, the intelligence level of industrial equipment has been significantly improved, and the structural design of equipment has become increasingly complex, and its safety issues have attracted much attention. Therefore, how to timely discover and solve the safety hazards of equipment plays an immeasurable role in modern industrial production.
[0003] Fault diagnosis technology aims to monitor and analyze the status of equipment or systems in real time, in order to identify potential faults at an early stage and take preventive measures to ensure the safe and efficient operation of the system. In traditional fault diagnosis methods, they usually rely on expert experience, rules, or analysis based on physical models. However, these methods require professional knowledge, and the modeling process is cumbersome and has poor adaptability when dealing with complex equipment or systems.
[0004] In recent years, data-driven fault diagnosis methods have gradually emerged, especially the introduction of machine learning technology, which has provided a new approach for fault diagnosis. Compared with traditional methods, data-driven methods do not require in-depth knowledge of equipment structure, and only require a large amount of historical fault data to complete the extraction and classification of fault features through machine learning models. However, considering that machine learning methods require the use of a large number of statistical characteristic parameters, deep learning methods based on convolutional neural networks (CNN) have gradually attracted attention and application. Compared with traditional machine learning methods, deep learning methods can automatically extract fault features and realize end-to-end training of mechanical equipment health status data, thereby improving diagnostic accuracy and generalization performance.
[0005] It is worth noting that many current deep learning methods rely on a large number of labeled fault signals with balanced categories, the same distribution of training and testing, in order to achieve high diagnostic accuracy. However, data in actual industrial applications often contain a lot of noise, especially the vibration signals and current signals collected by sensors, which are seriously affected by the external environment, increasing the difficulty of fault diagnosis. The one-dimensional signal collected by the sensor is interfered by environmental noise, which makes it difficult for the fault features to be extracted by the network. Although the fault information can be retained by continuous wavelet transform (Yan R, Gao RX, Chen X. Wavelets for fault diagnosis of rotary machines: A review with applications [J]. Signal processing, 2014, 96: 1-15.) or STFT operation (Wang LH, Zhao XP, Wu JX, et al. Motor fault diagnosis based on short-time fourier transform and convolutional neural network [J]. Chinese Journal of Mechanical Engineering, 2017, 30 (6): 1357-1368.), the whole process still requires manual adjustment of the threshold to eliminate some environmental noise or adjust the time-frequency transformation parameters to obtain a clear time-frequency diagram. Therefore, the entire training process is not end-to-end, and the overly redundant pre-task design will not only increase the data processing time, but also increase the model training and reasoning time.
[0006] Therefore, how to extract effective fault features in a noisy background becomes a key issue in fault diagnosis, and the effectiveness of the network has been fully verified through extensive experiments. Meeting the noise suppression requirements in industrial applications has become a technical problem that needs to be urgently solved by technicians in this field.
[0007] In order to extract effective fault features in a noisy background, the ResNet model has achieved good results in computer vision fields such as image classification and target positioning. However, in the field of mechanical fault diagnosis, monitoring signals usually come from complex industrial environments such as large rotating machines such as turbines or heavy trucks, which contain a lot of strong noise. When the ResNet network deals with monitoring signals in a strong noise background, its feature learning ability will be interfered by the noise, resulting in the inability to effectively extract fault-related features.
[0008] Zhao M, Zhong S, Fu X, et al. Deep residual shrinkage networks for fault diagnosis[J]. IEEE Transactions on Industrial Informatics, 2019, 16(7): 4681-4690. By using the soft threshold strategy to eliminate the influence of noise features in the monitoring signal, a deep residual shrinkage network with channel attention mechanism (DRSN-CW) is proposed, so that the residual network can automatically learn the threshold for eliminating noise. Compared with other networks, the DRSN-CW model has better feature learning ability under strong industrial background noise. The DRSN-CW model weights the key feature components through the soft threshold shrinkage strategy and the attention mechanism in the channel direction, and thereby eliminates the background noise of the monitoring signal, thereby achieving better diagnostic results. However, due to the limitations of the soft threshold function, in the denoising process, relying solely on the channel attention mechanism will cause the fault-related components in the monitoring signal to be weakened, resulting in a decrease in diagnostic accuracy. Summary of the invention
[0009] In order to solve the above technical problems, the present invention proposes an improved deep residual shrinkage network with parallel spatial attention mechanism and channel attention mechanism by correcting and improving the deep residual shrinkage network (DRSN-CW).
[0010] The specific plan is as follows:
[0011] The present invention provides an improved deep residual shrinkage network, comprising the following steps:
[0012] In step 1 of the present invention, a spatial attention mechanism module is added to learn data-sensitive location information. Inspired by the scSE (spatial-channel wise Squeeze and Excitation) module, the DRSN-CW model is improved. Specifically: the scSE module includes an sSE (spatial-wise Squeeze and Excitation) module, which is mainly manifested as a spatial attention mechanism for learning data-sensitive location information. Its core idea is to first obtain a new feature map with the same width and height as the original feature map by compressing the channel of the feature map, and then use the activated new feature map as an importance matrix. Finally, the matrix is applied to the original input feature map for correction, thereby completing the correction of spatial or temporal information.
[0013] Its specific approach is to assume that the size of the input monitoring signal is 1×W×1, and the size of the feature map output after the convolution layer is C×W×1. In the channel direction, a weight of F is first used. sq The input feature map is compressed by a 1×1×1 convolution to obtain a new feature map q, where the size of q is 1×W×1. The feature map q is then activated using the Sigmoid function σ(.) to obtain σ(q). Finally, σ(q) is element-wise multiplied with the input feature map y in the spatial direction to obtain the spatially weighted feature map z.
[0014] The above process is shown in formula (1) and formula (2):
[0015] q=F sq *y#(1)
[0016]
[0017] Among them, σ(q) represents the feature importance coefficient for each position of input x in time series, Represents the operation of multiplying each element in the spatial direction. The weighted feature map z represents the position-sensitive information of the model learning feature, so that the model pays attention to the fault information from the perspective of time series, achieving the purpose of suppressing noise in the spatial direction.
[0018] In step 2 of the present invention, a channel attention mechanism module is added to obtain the feature correlation between different channels. Inspired by the scSE (spatial-channel wise Squeeze and Excitation) module, the DRSN-CW model is improved. Specifically: scSE includes a cSE (channel-wise Squeeze and Excitation) module, which is mainly manifested as a channel attention mechanism for obtaining the feature correlation between different channels. Its core idea is to first obtain the weight of each channel by compressing the spatial information of the feature, and then activate the weight of each channel to the original input feature map channel for correction.
[0019] Its specific approach is to input feature map y = [y1, y2, ..., y c ](where feature y i ∈R C×W×1 , C and W are the number of channels and width of the current feature map respectively) use the GAP operation to compress the spatial dimension to obtain a vector q, whose dimension is C×1×1, and the vector q of the i-th channel i The calculation of is shown in formula (3):
[0020]
[0021] Finally, the activation vector is used for each channel of the input feature y Correction is performed to obtain the vector And the output z is shown in formula (4) and formula (5):
[0022]
[0023] in, Indicates the corresponding multiplication of each element in the channel direction. It contains an evaluation of the importance of each channel. After multiplying it to the input feature y, the information of unimportant channels can be suppressed while the information of key channels remains almost unchanged, thereby enabling the model to enhance its ability to capture key channel features.
[0024] In step 3 of the present invention, the application of two modules is combined at the same time to better capture the features useful for the diagnostic task. The design of the sSE module and the cSE module helps the model to better extract the information of key timing points and important channels. The scSE module combined in parallel can enhance the feature representation ability of the model from both spatial and channel aspects. Therefore, inspired by the design of the scSE module, the present invention improves the RSBU module and proposes a residual building unit (Residual Shrinkage Building Unit with Spatial-Channel Wise, RSBU-SCW) with parallel spatial attention mechanism and channel attention mechanism. The residual building unit (RSBU-SCW) fuses the spatial information and channel information of parallel attention, and then passes through the soft threshold function to achieve a better background noise suppression effect, and the effectiveness of the network is fully verified through extensive experiments, meeting the noise suppression requirements in industrial applications.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] 1. The spatial and channel attention mechanisms are added to enable the model to more accurately identify and amplify fault features while suppressing background noise irrelevant to the task, making the model more robust in high-noise environments and suitable for complex industrial scenarios.
[0027] 2. The spatial attention mechanism can capture the feature information of one-dimensional signals at different spatial positions, while the channel attention mechanism focuses on the correlation between different feature channels. By learning the spatial and channel distribution of important features, it can more effectively extract key features related to diagnostic tasks.
[0028] 3. Through the soft threshold function combined with the parallel attention mechanism, DRSN-SCW can automatically select the features with the most diagnostic value and suppress the interference of irrelevant information. This feature reduces the model's reliance on manual feature engineering in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 is a flow chart of the present invention;
[0030] Figure 2 This is the sSE module architecture diagram;
[0031] Figure 3 This is the cSE module architecture diagram;
[0032] Figure 4 This is the structural diagram of the RSBU-SCW module improved by the present invention. DETAILED DESCRIPTION
[0033] To further illustrate the various embodiments, the present invention provides drawings. These drawings are part of the disclosure of the present invention, which are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, a person of ordinary skill in the art should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.
[0034] With reference to the accompanying drawings, the present invention provides an improved deep residual shrinkage network, comprising the following steps:
[0035] Step 1: Add a feature spatial attention mechanism module to learn data-sensitive location information. Inspired by the scSE (spatial-channel wise Squeeze and Excitation) module, the DRSN-CW model is improved. Specifically, the scSE module includes the sSE (spatial-wise Squeeze and Excitation) module, which is mainly manifested as a spatial attention mechanism, such as Figure 2 As shown, it is used to learn data-sensitive location information.
[0036] Its core idea is to first compress the channels of the feature map to obtain a new feature map with the same width and height as the original feature map, then use the activated new feature map as the importance matrix, and finally apply the matrix to the original input feature map for correction, thereby completing the correction of spatial or temporal information.
[0037] Its specific approach is to assume that the size of the input monitoring signal is 1×W×1, and the size of the feature map output after the convolution layer is C×W×1. In the channel direction, a weight of F is first used. sq The input feature map is compressed by a 1×1×1 convolution to obtain a new feature map q, where the size of q is 1×W×1. The feature map q is then activated using the Sigmoid function σ(.) to obtain σ(q). Finally, σ(q) is element-wise multiplied with the input feature map y in the spatial direction to obtain the spatially weighted feature map z.
[0038] The above process is shown in formula (1) and formula (2):
[0039] q=F sq *y#(1)
[0040]
[0041] Among them, σ(q) represents the feature importance coefficient for each position of input x in time series, It represents the operation of multiplying the corresponding elements in the spatial direction. The weighted feature map z represents the position-sensitive information of the model learning features, so that the model pays attention to the fault information from the perspective of time series, achieving the purpose of suppressing noise in the spatial direction.
[0042] Step 2: A channel attention mechanism module is added to obtain the feature correlation between different channels. Inspired by the scSE (spatial-channel wise Squeeze and Excitation) module, the DRSN-CW model is improved. Specifically, scSE includes the cSE (channel-wise Squeeze and Excitation) module, such as Figure 3 As shown in the figure, it mainly manifests itself as a channel attention mechanism, which is used to obtain the feature correlation between different channels.
[0043] Its core idea is to first obtain the weight of each channel by compressing the spatial information of the features, and then activate the weight of each channel to the original input feature map channel for correction.
[0044] Its specific approach is to input feature map y = [y1, y2, ..., y c ](where feature y i ∈R C×W×1 , C and W are the number of channels and width of the current feature map respectively) use the GAP operation to compress the spatial dimension to obtain a vector q, whose dimension is C×1×1, and the vector q of the i-th channel i The calculation of is shown in formula (3):
[0045]
[0046] Finally, the activation vector is used for each channel of the input feature y Correction is performed to obtain the vector And the output z is shown in formula (4) and formula (5):
[0047]
[0048] in Indicates the corresponding multiplication of each element in the channel direction. It contains an evaluation of the importance of each channel. After multiplying it to the input feature y, the information of unimportant channels can be suppressed while the information of key channels remains almost unchanged, thereby enabling the model to enhance its ability to capture key channel features.
[0049] Step 3: The application of the two modules is combined at the same time to better capture the features that are useful for the diagnostic task. The design of the sSE module and the cSE module helps the model to better extract the information of key timing points and important channels. The scSE module combined in parallel can enhance the feature representation ability of the model from both spatial and channel aspects. Therefore, inspired by the design of the scSE module, the present invention improves the RSBU module and proposes a residual building unit (Residual Shrinkage Building Unit with Spatial-ChannelWise, RSBU-SCW) with parallel spatial attention mechanism and channel attention mechanism. The improved residual building unit (RSBU-SCW) is as follows Figure 4 RSBU-SCW fuses the spatial information and channel information of parallel attention and then passes through the soft threshold function to achieve better background noise suppression effect.
[0050] The present invention aims to allow the network to learn the key fault components and suppress unimportant channel information in time sequence, and then fuse the spatial information and channel information of parallel attention, and then pass through the soft threshold function to achieve better background noise suppression effect. The DRSN-SCW network proposed in the present invention contains sixteen RSBU-SCW modules, where the input is a 1×3600×1 monitoring signal and the output is the number of fault categories C. The design of the entire network layer is shown in Table 1.
[0051] Table 1 DRSN-SCW network layer design table
[0052]
[0053] In order to verify the effectiveness of the model, supervised learning is performed on balanced and unbalanced data for different networks to explore the effectiveness of the DRSN-SCW model proposed in this invention, and verification is performed on the Spectra Quest mechanical fault dataset and the CWRU bearing dataset. The mechanical fault diagnosis task is equivalent to a multi-classification problem, and its purpose is to diagnose different monitoring signals. In addition, for the evaluation of the classification effect of a single category, as shown in Table 2, the task is generally evaluated by a confusion matrix, in which the samples can be divided into four types according to their true categories and prediction results: true positive examples (True Positive, TP), true negative examples (True Negative, TN), false positive examples (False Positive, FP) and false negative examples (False Negative, FN).
[0054] Table 2 Confusion matrix for single category classification
[0055] True category Predicted as positive Predicted as a counterexample Positive example TP FN Counterexample FP TN
[0056] For each type of prediction result, this experiment uses accuracy as the main evaluation indicator, which is defined as shown in formula (6):
[0057]
[0058] In addition, according to the confusion matrix, the precision rate P and recall rate R can be calculated as shown in formula (7) and formula (8) respectively:
[0059]
[0060] In order to comprehensively consider the P and R indicators to evaluate the model, the F1_score of each category is defined as shown in formula (9):
[0061]
[0062] In this experiment, in order to highlight the importance of categories with fewer samples in the multi-class classification imbalance diagnosis task, this paper uses macro_F1 as a statistical indicator. Assuming that there are C categories in the dataset, then P macro and R macro They are defined as shown in formula (10) and formula (11) respectively:
[0063]
[0064] According to formula (10), formula (11) can be used to obtain macro_F1 as formula (12)
[0065]
[0066] In this section, the experiments uniformly use one-dimensional signals of size 1×3600×1 as input, and use ResNet10, ConvNet, DRSN-CW, and DRSN-SCW models with equivalent network parameters as comparison networks. The parameter comparison of different networks is shown in Table 3. In addition, these different networks are trained on the Spectra Quest and CWRU datasets for supervised learning, and the experimental results are shown in Table 4. The results with the highest indicators in the experiment are shown in bold.
[0067] Table 3 Comparative study of parameters of different networks
[0068]
[0069]
[0070] Table 4 Performance comparison of different networks under supervised learning
[0071]
[0072] From Table 4 we can see that:
[0073] First, compared with ConvNet, as the category imbalance factor α in the dataset gradually increases, the other three networks with residual structures benefit from more effective gradient backpropagation and feature extraction capabilities and can obtain better results. When α=100, the accuracy of the model with residual structure on the Spectra Quest and CWRU datasets is 8.375% and 9.572% higher than the lowest ConvNet, respectively.
[0074] Secondly, compared with the DRSN-CW model, with the same number of network parameters, the network proposed in this paper can significantly improve the diagnostic performance in three data distributions. When α=100, the accuracy on the Spectra Quest and CWRU datasets is improved by 6.75% and 2.286%, respectively, and macro-F1 is improved by 0.074 and 0.029, respectively.
[0075] Furthermore, the reason why the accuracy of the Spectra Quest dataset can be greatly improved is that the dataset itself contains background noise. The DRSN-SCW model proposed in this paper based on the parallel architecture of the channel attention mechanism and the spatial attention mechanism can better combine the soft threshold function to suppress noise. The CWRU bearing dataset contains less noise, and the effect of the DRSN-SCW model on improving performance is relatively less obvious.
[0076] In general, compared with networks such as ResNet10, ConvNet and DRSN-SCW, the DRSN-SCW model proposed in this paper achieves the highest scores in terms of accuracy and macro-F1 indicators, which can effectively prove the effectiveness of the DRSN-SCW model in fault diagnosis application tasks.
[0077] Although the present invention has been specifically shown and described in conjunction with the preferred embodiments, it should be understood by those skilled in the art that various changes may be made to the present invention in form and details without departing from the spirit and scope of the present invention as defined by the appended claims, all of which are within the scope of protection of the present invention.
Claims
1. An improved deep residual shrinkage network, characterized in that: The following steps are involved: (1) adding a spatial attention mechanism module, wherein the spatial attention mechanism module includes an sSE module for learning data-sensitive position information; (2) adding a channel attention mechanism module, wherein the channel attention mechanism module includes a cSE module for obtaining feature correlations between different channels; (3) The two modules of step (1) and step (2) are combined to better capture the features that are useful for the diagnostic task and suppress the features that are irrelevant to the current task. The network learns the key fault components in time sequence and suppresses unimportant channel information, fuses the spatial information and channel information that are paid attention to in parallel, and then passes through the soft threshold function to achieve better background noise suppression effect.
2. The improved deep residual shrinkage network according to claim 1, characterized in that The sSE module first obtains a new feature map with the same width and height as the original feature map by compressing the channel of the feature map, then uses the activated new feature map as the importance matrix, and finally applies the matrix to the original input feature map for correction, thereby completing the correction of spatial or temporal information.
3. The improved deep residual shrinkage network according to claim 2, characterized in that The specific method of the sSE module is as follows: Assuming that the size of the input monitoring signal is 1×W×1, the size of the feature map output after the convolution layer is C×W×1, and the channel direction is firstly calculated by using a weight F sq The input feature map is compressed by a 1×1×1 convolution to obtain a new feature map q, where the size of q is 1×W×1. The feature map q is then activated using the Sigmoid function σ(.) to obtain σ(q). Finally, σ(q) is element-wise multiplied with the input feature map y in the spatial direction to obtain the spatially weighted feature map z. The above process is shown in formula (1) and formula (2): q=F sq *y#(1) Among them, σ(q) represents the feature importance coefficient for each position of input x in time series, It represents the operation of multiplying each element in the spatial direction. The weighted feature map z represents the position-sensitive information of the model learning features, so that the model pays attention to the fault information from a timing perspective and achieves the purpose of suppressing noise in the spatial direction.
4. The improved deep residual shrinkage network according to claim 1, characterized in that The cSE module first obtains the weight of each channel by compressing the spatial information of the feature, and then activates the weight of each channel to the original input feature map channel for correction.
5. The improved deep residual shrinkage network according to claim 4, characterized in that The specific method of the cSE module is: for the input feature map y = [y1, y2, ..., y c ]Use GAP operation to compress in the spatial dimension to obtain vector q, whose dimension is C×1×1, and the vector q of the i-th channel i The calculation of is shown in formula (3): Finally, the activation vector is used for each channel of the input feature y Correction is performed to obtain the vector And the output z is shown in formula (4) and formula (5): Among them, y i ∈R C×W×1 , C and W are the number of channels and width of the current feature map respectively, Indicates the corresponding multiplication of each element in the channel direction; activation vector It contains an evaluation of the importance of each channel. After multiplying it to the input feature y, the information of unimportant channels can be suppressed while the information of key channels remains almost unchanged, thereby enabling the model to enhance its ability to capture key channel features.
6. The improved deep residual shrinkage network according to claim 1, characterized in that Step (3) enhances the feature representation capability of the model from the spatial and channel aspects through the scSE module which is a parallel combination of the sSE module in step (1) and the cSE module in step (2), so as to better extract the information of key timing points and important channels; and constructs a residual construction unit with a parallel special spatial attention mechanism and a channel attention mechanism, wherein the residual construction unit fuses the spatial information and channel information of the parallel attention, and then passes through a soft threshold function, so as to achieve a better background noise suppression effect; The effectiveness of the network has been fully verified through extensive experiments, meeting the noise suppression requirements in industrial applications.