Vibration signal processing method and device, computer equipment and computer storage medium

Abstract

The invention relates to a vibration signal processing method and device, computer equipment and a computer storage medium. A to-be-processed vibration signal is acquired; the to-be-processed vibration signal comprises a noise signal and a target vibration signal; the method comprises the steps of performing multi-scale feature extraction on a to-be-processed vibration signal by using a pre-constructed denoising model, removing a noise signal to obtain a multi-scale feature, performing feature enhancement and feature fusion processing on the multi-scale feature to obtain feature fusion data, and performing feature mapping processing on the feature fusion data to obtain a target vibration signal. Through the pre-constructed denoising model, multi-scale feature extraction, feature enhancement, fusion and mapping processing flows are integrated, noise signals in the to-be-processed vibration signals can be accurately removed, and the target vibration signals are effectively recovered. For a mechanical equipment operation state monitoring scene, the accuracy of a monitoring signal can be remarkably improved, fault misjudgment or missed judgment caused by noise interference is avoided, and a reliable basis is provided for equipment maintenance.

Description

Technical Field

[0001] The present application relates to the technical field of signal processing, and in particular to a vibration signal processing method, device, computer equipment and computer storage medium. Background Art

[0002] In the modern industrial field, the safe and stable operation of mechanical equipment is of vital importance. As an important representation of the operating status of mechanical equipment, vibration signals can reflect the working conditions of various components inside the machine. However, when actually collecting vibration signals, they are inevitably affected by a variety of interference factors, thus mixing in noise.

[0003] The current denoising algorithms are mainly based on signal processing technologies, such as filtering algorithms (low-pass filtering, high-pass filtering, band-pass filtering, etc.), wavelet transform, etc. These algorithms often require some preset conditions, such as the cutoff frequency of the filter, the selection of wavelet basis functions, etc. However, vibration signals have complex characteristics, and these preset conditions are difficult to adapt to the dynamic characteristics of vibration signals, resulting in poor denoising effects. Summary of the invention

[0004] Based on this, it is necessary to provide a vibration signal processing method, device, computer equipment and computer storage medium that can solve the above-mentioned technical problems.

[0005] In a first aspect, the present application provides a vibration signal processing method, the method comprising:

[0006] Acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal;

[0007] The pre-built denoising model is used to extract multi-scale features of the vibration signal to be processed, the noise signal is removed to obtain multi-scale features, the multi-scale features are enhanced and fused to obtain feature fusion data, and the feature fusion data is feature mapped to obtain the target vibration signal.

[0008] In one embodiment, the denoising model includes a pyramid network, a residual attention network and an output network; using the pre-built denoising model to perform multi-scale feature extraction on the vibration signal to be processed to obtain multi-scale features, performing feature fusion processing on the multi-scale features to obtain feature fusion data, and performing feature mapping processing on the feature fusion data to obtain a target vibration signal, including:

[0009] Multi-scale feature extraction is performed on the vibration signal to be processed through a pyramid network to obtain multi-scale features;

[0010] The multi-scale features are enhanced and fused through the residual attention network to obtain feature fusion data;

[0011] The feature fusion data is processed by feature mapping through the output network to obtain the target vibration signal.

[0012] In one embodiment, the pyramid network includes a multi-scale extraction layer, a stacking layer and a first convolution layer connected in sequence, and the pyramid network is used to perform multi-scale feature extraction processing on the vibration signal to be processed to obtain multi-scale features, including:

[0013] Performing multi-scale convolution processing on the vibration signal to be processed through a multi-scale extraction layer to obtain a plurality of first feature maps;

[0014] Performing fusion processing on each first feature map through a stacking layer to obtain a fusion feature;

[0015] The first convolutional layer is used to perform channel adjustment and feature optimization on the fused features to obtain multi-scale features.

[0016] In one embodiment, the multi-scale extraction layer includes a plurality of second convolution layers and a pooling layer; the vibration signal to be processed is subjected to multi-scale convolution processing by the multi-scale extraction layer, and a plurality of first feature maps are obtained, including:

[0017] Performing dilated convolution processing on the vibration signal to be processed through each second convolution layer and multiple dilation rates to obtain multiple initial feature maps;

[0018] Downsample each initial feature map through a pooling layer to obtain multiple downsampled feature maps;

[0019] The down-sampled feature maps are subjected to feature map concatenation processing to obtain first feature maps.

[0020] In one embodiment, the residual attention network includes a multi-level first residual convolution layer, a multi-level attention layer and a multi-level second residual convolution layer, the multi-level first residual convolution layer and the multi-level second residual convolution layer correspond to each other, and the multi-level attention layer is connected across the input end of the corresponding first residual convolution layer and the output end of the second residual convolution layer; the multi-scale features are enhanced and fused by the residual attention network to obtain feature fusion data, including:

[0021] Input each first input feature into the first residual convolution layer of each level for feature extraction and residual connection processing to obtain multiple first residual features; wherein the first input feature of the first residual convolution layer of the first level is a multi-scale feature, and the first input feature of the first residual convolution layer of other levels is the first residual feature output by the first residual convolution layer of the previous level;

[0022] The first input feature of the first residual convolution layer at each level is input into the attention layer at each level for feature enhancement processing to obtain multiple enhanced features;

[0023] Each second input feature is input into each level of the second residual convolution layer for fusion processing to obtain multiple second residual features; wherein, the second input feature of the first-level second residual convolution layer is the first residual feature output by the last-level first residual convolution layer, and the second input features of the other levels of the second residual convolution layer are the second residual features output by the previous level of the second residual convolution layer and the enhanced features output by the corresponding attention layer, and the feature fusion data includes the second residual features output by the last-level second residual convolution layer and the enhanced features output by the last-level attention layer.

[0024] In one embodiment, the first input feature of the first residual convolution layer at each level is input into the attention layer at each level for feature enhancement processing to obtain multiple enhanced features;

[0025] For each level of attention layer, the first input feature of the corresponding first residual convolution layer is input into the feature extraction layer for multi-scale local feature extraction processing to obtain multi-scale local features;

[0026] The multi-scale local features are input into the feature adjustment layer for global feature fusion and channel weight calculation to obtain channel weights;

[0027] The multi-scale local features and channel weights are input into the feature scaling layer for feature scaling operation to obtain enhanced features.

[0028] In one embodiment, the first residual convolution layer includes a convolution layer and a jump connection layer, and each first input feature is input into each level of the first residual convolution layer for feature extraction and residual connection processing to obtain multiple first residual features, including:

[0029] The first input feature is input into the convolution layer for local feature extraction to obtain the intermediate feature;

[0030] The intermediate features are added to the multi-scale features through the jump connection layer to obtain the first residual feature.

[0031] In a second aspect, the present application also provides a vibration signal processing device. The device comprises:

[0032] A signal acquisition module is used to acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal;

[0033] The denoising module is used to perform multi-scale feature extraction on the vibration signal to be processed using a pre-built denoising model, remove the noise signal to obtain multi-scale features, perform feature enhancement and feature fusion processing on the multi-scale features to obtain feature fusion data, and perform feature mapping processing on the feature fusion data to obtain the target vibration signal.

[0034] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0035] Acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal;

[0036] The pre-built denoising model is used to extract multi-scale features of the vibration signal to be processed, the noise signal is removed to obtain multi-scale features, the multi-scale features are enhanced and fused to obtain feature fusion data, and the feature fusion data is feature mapped to obtain the target vibration signal.

[0037] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0038] Acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal;

[0039] The pre-built denoising model is used to extract multi-scale features of the vibration signal to be processed, the noise signal is removed to obtain multi-scale features, the multi-scale features are enhanced and fused to obtain feature fusion data, and the feature fusion data is feature mapped to obtain the target vibration signal.

[0040] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0041] Acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal;

[0042] The pre-built denoising model is used to extract multi-scale features of the vibration signal to be processed, the noise signal is removed to obtain multi-scale features, the multi-scale features are enhanced and fused to obtain feature fusion data, and the feature fusion data is feature mapped to obtain the target vibration signal.

[0043] The above-mentioned vibration signal processing method, device, computer equipment and computer storage medium obtain the vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal; a pre-constructed denoising model is used to perform multi-scale feature extraction on the vibration signal to be processed, the noise signal is removed to obtain multi-scale features, the multi-scale features are subjected to feature enhancement and feature fusion processing to obtain feature fusion data, and the feature fusion data is subjected to feature mapping processing to obtain the target vibration signal. The present application integrates the multi-scale feature extraction, feature enhancement, fusion and mapping processing procedures through a pre-constructed denoising model, and can accurately remove the noise signal in the vibration signal to be processed and effectively restore the target vibration signal. For the scenario of monitoring the operating status of mechanical equipment, it can significantly improve the accuracy of the monitoring signal, avoid misjudgment or missed faults due to noise interference, and provide a reliable basis for equipment maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A diagram showing an application environment of a vibration signal processing method in one embodiment;

[0045] Figure 2 A schematic diagram of a flow chart of a vibration signal processing method in one embodiment;

[0046] Figure 3 A schematic diagram of a denoising process in one embodiment;

[0047] Figure 4 A schematic diagram of a multi-scale feature extraction process in one embodiment;

[0048] Figure 5 Schematic diagram of a multi-scale convolution process in one embodiment;

[0049] Figure 6 is a schematic diagram of a denoising model in one embodiment;

[0050] Figure 7 A schematic diagram of a process of feature enhancement and feature fusion processing in one embodiment;

[0051] Figure 8 A schematic diagram of a flow chart of feature enhancement processing in one embodiment;

[0052] Fig. 9 A schematic diagram of a process of feature extraction and residual connection processing in one embodiment;

[0053] Fig.10 is a structural block diagram of a vibration signal processing device in one embodiment;

[0054] Fig.11 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0056] First of all, before specifically introducing the technical solution of the embodiments of the present application, the technical background based on the embodiments of the present application is introduced.

[0057] Vibration signals play a vital role in monitoring and diagnosing the status of mechanical equipment, because vibration signals can reflect the operating status and potential hidden dangers of the machine. However, in the actual process of signal collection, due to interference from electrical interference, mechanical friction, environmental vibration, sensor noise, etc., these interferences are irrelevant to the operating status of the mechanical equipment and will seriously affect the quality of the signal. At the same time, these interferences will also mask the early fault characteristics of the vibration signal, or cause false fault alarms, thereby affecting the detection and diagnosis of the fault. In order to improve the quality of vibration signals, denoising is an indispensable operation. By processing the noise of vibration signals, the interference of noise in the signal can be reduced and the accuracy of fault detection and diagnosis can be improved.

[0058] The research on vibration signal noise processing algorithm is a complex issue, facing many key issues and challenges. First of all, there are many noise sources in the process of vibration signal acquisition. These noise interferences will make the collected vibration signal more complicated, making its analysis and processing more difficult. At the same time, vibration signals have non-stationary and nonlinear characteristics, and the denoising effect of traditional signal processing algorithms is very limited. Existing vibration signal noise processing algorithms are mainly based on signal processing technology and neural network algorithms. The denoising algorithm based on signal processing mainly processes the signal through signal decomposition, such as singular value decomposition, modal decomposition, etc., and signal transformation technology such as wavelet transform, and then denoises the signal through threshold denoising algorithm, and finally reconstructs the denoised signal. The signal processing algorithm based on singular value decomposition can process non-stationary and nonlinear signals, and has good stability. It has been widely used in the field of signal processing, but its computational complexity is large and the time complexity is high. Wavelet transform is also commonly used in signal denoising, which can provide information on signals at different scales and better capture the local characteristics of signals. Combined with the threshold method, it can achieve effective denoising of signals. However, the effect of wavelet transform depends on the selection of wavelet basis functions and the setting of decomposition layers, resulting in poor adaptability. Mode decomposition technology represented by empirical mode decomposition (EMD) has also shown certain advantages in the field of signal denoising. EMD can adaptively decompose signals without setting any basis functions. It also has good processing capabilities for nonlinear and non-stationary signals. It can also perform multi-scale analysis on signals and characterize the characteristics of signals at different time scales. For noisy signals, EMD can effectively separate noise components and achieve signal denoising. However, EMD also has some limitations. It is easy to produce mode aliasing during signal decomposition. There will also be noise components in low-frequency components. When the noise signal is strong, the noise components will be mistakenly identified as effective components, affecting the denoising effect. Deep learning algorithms have developed rapidly in recent years. The powerful feature extraction capabilities of deep neural networks can effectively deal with signal denoising problems. Deep learning models have strong feature extraction capabilities and good generalization capabilities. They can adapt to different signal environments and noises. They have good processing capabilities for signals with complex features such as vibration signals. At the same time, they can also achieve end-to-end signal denoising. However, neural network algorithms also have certain limitations. For example, the use of traditional denoising autoencoder denoising algorithms will lead to an increase in the number of parameters as the network depth increases, which is prone to overfitting. Using residual networks for signal denoising can increase the network depth while avoiding overfitting. However, this type of denoising algorithm mostly uses a single convolution kernel when extracting features, and the receptive field size is limited. The denoising effect is limited for signals with multi-scale characteristics such as vibration signals.

[0059] Based on this, the present application provides a vibration signal processing method, device, computer equipment and computer storage medium, aiming to solve the above technical problems.

[0060] The vibration signal processing method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. By obtaining the vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal; using a pre-built denoising model to extract multi-scale features of the vibration signal to be processed, removing the noise signal to obtain multi-scale features, performing feature enhancement and feature fusion processing on the multi-scale features to obtain feature fusion data, and performing feature mapping processing on the feature fusion data to obtain the target vibration signal. Among them, the terminal 102 can be but is not limited to various personal computers, laptops, smart phones and tablets. The server 104 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0061] In one embodiment, Figure 2 As shown, a vibration signal processing method is provided, and the method is applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0062] S201, obtaining a vibration signal to be processed.

[0063] The vibration signals to be processed include noise signals and target vibration signals. The vibration signals to be processed refer to the vibration signals collected by sensors during the operation of mechanical equipment. The signals include the target vibration signals generated by the operation of the mechanical equipment itself and the noise signals generated by various interference sources such as electrical interference, mechanical friction, environmental vibration, and sensor noise.

[0064] In the embodiment of the present application, in the actual application scenario, for example, a vibration sensor is installed at a key part of the mechanical equipment, such as an acceleration sensor installed at a bearing seat, a housing, etc. The sensor will collect the vibration signal generated by the mechanical equipment in real time during operation, and the signal is the vibration signal to be processed. During the collection process, the sensor converts the physical quantity of the vibration into an electrical signal, and converts the analog signal into a digital signal through a data acquisition card, and transmits it to a terminal device (such as a computer) for subsequent processing.

[0065] Another implementation method is to use a wireless sensor network to collect vibration signals. Multiple wireless sensor nodes are arranged around the mechanical equipment. After these nodes collect vibration signals, they transmit the signals to the aggregation node through wireless communication protocols (such as ZigBee, Bluetooth or Wi-Fi, etc.), and the aggregation node then sends all the collected vibration signal data to the terminal device. This method is convenient for arranging sensors in complex environments and reduces wiring costs and complexity, but it is necessary to consider the stability of wireless communication and the reliability of data transmission.

[0066] S202, using a pre-built denoising model to perform multi-scale feature extraction on the vibration signal to be processed, removing the noise signal to obtain multi-scale features, performing feature enhancement and feature fusion processing on the multi-scale features to obtain feature fusion data, and performing feature mapping processing on the feature fusion data to obtain a target vibration signal.

[0067] Among them, the denoising model refers to an algorithm model based on the encoder structure design, which can be composed of ASPP module, multi-scale attention mechanism, residual structure network, etc., and is used to denoise the vibration signal. The model can automatically learn signal features and realize end-to-end denoising process, and has strong adaptability and feature extraction capabilities.

[0068] In the embodiment of the present application, the vibration signal to be processed first enters the multi-scale extraction layer in the ASPP module. The multi-scale extraction layer includes multiple second convolution layers, and each second convolution layer uses a hole convolution with different expansion rates to perform a convolution operation on the input signal. For example, three second convolution layers are set, and the expansion rates of the hole convolution are 3, 6, and 9 respectively. When the input signal enters the second convolution layer with an expansion rate of 3, the convolution kernel samples the signal at a relatively small interval, focusing on extracting the local detail features of the signal and generating a preliminary feature map of the corresponding scale; the second convolution layer with an expansion rate of 6 samples at a slightly larger interval to obtain feature information of a slightly larger scale; the second convolution layer with an expansion rate of 9 samples at a larger interval to focus on the features of a wider area. The preliminary feature map output by each second convolution layer enters the corresponding pooling layer for downsampling operation. The pooling layer can adopt the maximum pooling or average pooling method. For example, the maximum pooling layer selects the maximum value in each local area as the output, which can highlight the significant features in the feature map, reduce the amount of data while retaining the main feature information, and obtain downsampled feature maps of different resolutions. Finally, the downsampled feature maps are concatenated in the channel dimension through the feature map concatenation operation to form multiple feature maps, each of which represents the characteristics of the signal at different scales and resolutions. These feature maps are output from the multi-scale extraction layer and enter the stacking layer of the ASPP module.

[0069] Another implementation method is to use the idea of ​​grouped convolution to achieve multi-scale feature extraction. The input signal is divided into several groups, each group is processed by a convolution layer with different parameters (such as different convolution kernel size or expansion rate), and then the output feature maps of each group are spliced ​​in the channel dimension to obtain multi-scale features. This method can reduce the amount of calculation to a certain extent and improve processing efficiency, but it may require a more sophisticated grouping strategy design.

[0070] To remove the noise signal and obtain multi-scale features, the superposition layer superimposes the feature maps of different scales from the multi-scale extraction layer so that the information of different scales can be merged in the channel dimension. Then, the superimposed features are downsampled through the average pooling operation in the pooling structure to further reduce the amount of data and retain the main feature information. After that, the downsampled features are subjected to nonlinear transformation and feature enhancement through operations such as convolution, batch normalization, and ReLU activation function. For example, the convolution operation can further adjust the representation of the features, batch normalization accelerates the training process and improves the stability of the model, and the ReLU activation function introduces nonlinear factors to enhance the model's ability to express complex feature relationships. Finally, upsampling is performed through the deconvolution structure to restore the resolution of the feature map and obtain the fused features, which integrate the feature information of different scales, that is, the multi-scale features after removing the noise signal.

[0071] Another alternative is to use a feature fusion method based on wavelet transform instead of using traditional convolution, pooling and deconvolution operations after stacking layers. The feature maps of different scales are decomposed by wavelet respectively, and then the features are fused in the wavelet domain, and then the fused features are obtained by inverse wavelet transform. This method takes advantage of the advantages of wavelet transform in time-frequency domain analysis and may be more effective in denoising certain types of vibration signals, but the computational complexity may be relatively high.

[0072] For feature enhancement, multi-scale features enter the first-level attention layer in the multi-scale attention network. In the feature extraction layer of the first-level attention layer, one-dimensional convolutions with different convolution kernel sizes (3, 5, 7) are used to convolve the input multi-scale features. The one-dimensional convolution with a convolution kernel size of 3 can finely capture the more local details of the signal, the one-dimensional convolution with a convolution kernel size of 5 can obtain the features of a slightly larger local area, and the one-dimensional convolution with a convolution kernel size of 7 focuses on the local features of a larger range, thereby extracting richer local feature information from the multi-scale features in parallel and obtaining multi-scale local features.

[0073] The feature adjustment layer performs a global feature fusion operation on the multi-scale local features, such as compressing all local features in the spatial dimension through global average pooling to obtain a global feature description. Then, based on this global feature description, the fully connected layer is used to calculate the weight of each channel. The fully connected layer determines the importance of each channel feature in the entire feature representation by learning the relationship between the global feature description and the channel, and obtains the channel weight.

[0074] The feature scaling layer scales the multi-scale local features according to the input channel weights, multiplying the feature value of each channel by the corresponding channel weight. Important channel features (corresponding to larger weight values) are enhanced, and less important channel features (corresponding to smaller weight values) are suppressed, thereby achieving effective adjustment of multi-scale local features, highlighting useful signal features, suppressing noise information, and obtaining the first enhanced feature.

[0075] Another way to achieve feature enhancement is to use a method based on the self-attention mechanism. The correlation between each position and other positions in the multi-scale feature is calculated, and the features are weighted according to the correlation score, so that the model can automatically focus on the important parts of the feature and enhance the useful signal features.

[0076] The above-mentioned vibration signal processing method, device, computer equipment and computer storage medium obtain the vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal; a pre-constructed denoising model is used to perform multi-scale feature extraction on the vibration signal to be processed, the noise signal is removed to obtain multi-scale features, the multi-scale features are subjected to feature enhancement and feature fusion processing to obtain feature fusion data, and the feature fusion data is subjected to feature mapping processing to obtain the target vibration signal. The present application integrates the multi-scale feature extraction, feature enhancement, fusion and mapping processing procedures through a pre-constructed denoising model, and can accurately remove the noise signal in the vibration signal to be processed and effectively restore the target vibration signal. For the scenario of monitoring the operating status of mechanical equipment, it can significantly improve the accuracy of the monitoring signal, avoid misjudgment or missed faults due to noise interference, and provide a reliable basis for equipment maintenance.

[0077] In an exemplary embodiment, based on the above embodiment, see Figure 3 The denoising model of the embodiment of the present application includes a pyramid network, a residual attention network and an output network; using the pre-built denoising model to extract multi-scale features of the vibration signal to be processed to obtain multi-scale features, performing feature fusion processing on the multi-scale features to obtain feature fusion data, and performing feature mapping processing on the feature fusion data to obtain a target vibration signal, including the following steps:

[0078] S301, performing multi-scale feature extraction processing on the vibration signal to be processed through a pyramid network to obtain multi-scale features.

[0079] Among them, the pyramid network is used to extract multi-scale features of the vibration signal to be processed. It can capture the characteristic information of the vibration signal at different scales through a series of specific network layer structures and operations, such as dilated convolutions with different expansion rates and pooling operations, just like a pyramid, presenting the full picture of the signal features from different levels, providing a rich and multi-dimensional feature basis for subsequent processing.

[0080] In the embodiment of the present application, in the multi-scale extraction layer, the group convolution method can be used to replace the traditional hole convolution with different expansion rates to perform multi-scale feature extraction. The vibration signal to be processed is divided into several groups, each group is processed by a convolution layer with different parameters (such as different convolution kernel size, step size, etc.), and then the feature maps output by each group are spliced ​​in the channel dimension to obtain multi-scale features. This method can configure the convolution parameters more flexibly according to the characteristics of the signal to a certain extent, and it is possible to improve the efficiency and pertinence of feature extraction, but it requires a more sophisticated design of the grouping strategy and convolution parameter settings to ensure that feature information of different scales can be fully and accurately extracted.

[0081] For the operations of the superposition layer and the first convolution layer, we can try to use a processing method based on wavelet transform to replace traditional operations such as convolution, pooling, and batch normalization. First, the superimposed feature map is decomposed by wavelet, and the decomposed coefficients are adjusted and processed in the wavelet domain. For example, the characteristics of the signal in different frequency bands are highlighted through the multi-resolution analysis characteristics unique to wavelet transform, and then the processed coefficients are restored to the feature map through the inverse wavelet transform. Finally, a wavelet convolution layer with a function similar to the first convolution layer is used to adjust the number of channels of the restored feature map and output multi-scale features. This method takes advantage of the advantages of wavelet transform in time-frequency analysis and may have better denoising effects on some vibration signals with special frequency characteristics, but the computational complexity is relatively high and needs to be weighed according to the computing resources and performance requirements of the actual application scenario.

[0082] S302, performing feature enhancement and feature fusion processing on multi-scale features through a residual attention network to obtain feature fusion data.

[0083] Among them, the residual attention network combines the network modules of the residual structure and the attention mechanism. The residual structure can effectively solve the performance degradation problem that may occur when the network depth increases. By establishing residual connections, it makes it easier for the network to learn useful feature representations; the attention mechanism can dynamically adjust the importance weights of different parts of the features according to the signal characteristics, highlight the useful signal features, and suppress the noise information, thereby improving the feature quality and the model's attention to key features.

[0084] In the embodiment of the present application, in terms of feature enhancement, a method based on a self-attention mechanism can be used to replace the traditional multi-scale attention mechanism. The self-attention mechanism calculates the correlation between each position and other positions in the multi-scale feature, and weights the feature according to the correlation score, so that the model can automatically focus on the important parts of the feature and enhance the useful signal features. This method may have better effects when processing long sequence vibration signals, but the computational cost is relatively high, and it needs to be weighed according to the computing resources and performance requirements of the actual application scenario.

[0085] For feature fusion operations, we can try to use dense connections instead of traditional residual connections and simple fusion methods. The enhanced features and residual features of each level are spliced ​​in the channel dimension, and then fused through a convolutional layer, so that the subsequent levels of the network can directly use the feature information of the previous levels, further improving the feature fusion effect, but it will increase a certain amount of calculation and model parameters, and it is also necessary to weigh the computing resources and performance requirements of the actual application scenario.

[0086] S303, performing feature mapping processing on the feature fusion data through the output network to obtain a target vibration signal.

[0087] Among them, the output network is mainly responsible for the final feature mapping of the feature fusion data after being processed by the residual attention network, and converting it into an output that conforms to the target vibration signal form. It achieves accurate conversion from processed feature data to actual usable vibration signals by adjusting the dimension of the features and performing necessary convolution operations.

[0088] In the embodiment of the present application, feature mapping processing can be performed by combining transposed convolution (deconvolution) and convolution layer. The feature fusion data is first upsampled by transposed convolution to increase the resolution of the feature map, and then the feature is adjusted by the convolution layer to make the output signal closer to the characteristics of the target vibration signal. This method can better restore the detailed information of the signal in some cases, but the computational complexity is relatively high, and it needs to be weighed according to the computing resources and performance requirements of the actual application scenario.

[0089] In the embodiment of the present application, through the multi-scale feature extraction processing of the pyramid network, it is possible to make full use of the hole convolution and pooling operations with different expansion rates, etc., to fully explore the feature information of the vibration signal at different scales. This allows the model to capture the local detail features (such as small vibration changes) and overall trend features (such as the overall vibration trend of the equipment) of the signal at the same time, and improves the ability to understand the complex characteristics of the vibration signal. Through the feature enhancement operation of the residual attention network, the attention mechanism can automatically adjust the weights of different channels and different local scale features according to the signal characteristics, effectively highlight the useful signal features, suppress noise information, and improve the quality and distinguishability of the features. This enables the model to focus more on the feature parts that are important for denoising in subsequent processing, thereby improving the denoising effect. Through the feature mapping processing of the output network, the feature fusion data after the previous series of processing can be converted into a form that matches the target vibration signal in terms of dimension and feature representation, and finally obtains the target vibration signal. This process effectively converts the feature information learned by the denoising model into an actually usable vibration signal, ensuring the effectiveness and practicality of the denoising process.

[0090] In an exemplary embodiment, based on the above embodiment, see Figure 4 The pyramid network of the embodiment of the present application includes a multi-scale extraction layer, an overlay layer and a first convolution layer connected in sequence. The embodiment of the present application is related to performing multi-scale feature extraction processing on the vibration signal to be processed by the pyramid network to obtain multi-scale features, including the following steps:

[0091] S401, performing multi-scale convolution processing on the vibration signal to be processed through a multi-scale extraction layer to obtain multiple first feature maps.

[0092] In the embodiment of the present application, the multi-scale extraction layer includes multiple second convolutional layers and pooling layers, which collaboratively complete the multi-scale convolution processing operation.

[0093] The plurality of second convolutional layers respectively use dilated convolutions with different expansion rates to perform convolution operations on the vibration signal to be processed. For example, three second convolutional layers are set, and the expansion rates of the dilated convolutions are 3, 6, and 9, respectively.

[0094] When the vibration signal to be processed enters the second convolution layer with a dilation rate of 3, the convolution kernel slides on the signal at a relatively small sampling interval to perform convolution operations. Due to the small dilation rate, this convolution layer can capture the local detail features of the signal more finely, such as being able to keenly perceive the small amplitude changes and short-term frequency fluctuations in the vibration signal, and generate a preliminary feature map of the corresponding scale.

[0095] For the second convolution layer with a dilation rate of 6, its convolution kernel slides on the signal at a moderate sampling interval for convolution. This layer can obtain feature information of a slightly larger scale. Compared with the layer with a dilation rate of 3, it can capture the combination of some related features in the local area, such as a relatively macroscopic local feature pattern formed by several adjacent small changes, and also generate a preliminary feature map of the corresponding scale.

[0096] The second convolution layer with a dilation rate of 9 allows the convolution kernel to slide on the signal for convolution at a larger sampling interval. This layer focuses on the features of a wider area, and can capture more macroscopic feature information such as the vibration trend of a certain component of the entire device in a larger range and the overall vibration characteristics, and outputs a preliminary feature map of the corresponding scale.

[0097] The preliminary feature maps output by each second convolutional layer then enter their corresponding pooling layers for downsampling. The pooling layer can use maximum pooling or average pooling. Taking maximum pooling as an example, the maximum value in the feature map is selected as the output in each local area. The advantage of this is that it can highlight the significant features in the feature map, while greatly reducing the amount of data, and to a certain extent suppressing the impact of noise on the features, thereby obtaining downsampled feature maps of different resolutions. These downsampled feature maps retain the main feature information and reduce the resolution, making the amount of data for subsequent processing easier to manage and more conducive to further fusion of features.

[0098] Finally, the downsampled feature maps are concatenated in the channel dimension through the feature map concatenation operation to form multiple first feature maps. Each first feature map now represents the features of the signal at different scales (determined by the dilated convolution with different expansion rates) and different resolutions (determined by the downsampling of the pooling layer), and is passed to the overlay layer as the output of the multi-scale extraction layer for the next step of processing.

[0099] S402, fusing the first feature maps by stacking layers to obtain fused features.

[0100] In the embodiment of the present application, after the superposition layer receives the multiple first feature maps from the multi-scale extraction layer, it first superimposes these feature maps of different scales in the channel dimension.

[0101] Next, the superimposed features are downsampled, and then the downsampled features are nonlinearly transformed and enhanced through operations such as convolution, batch normalization, and ReLU activation function. Finally, upsampling is performed to restore the resolution of the feature map and obtain the fused features.

[0102] S403, performing channel adjustment and feature optimization processing on the fused features through the first convolutional layer to obtain multi-scale features.

[0103] In the embodiment of the present application, the first convolution layer performs a convolution operation on the fused features from the superposition layer according to the preset convolution kernel parameters. The convolution kernel slides on the fused feature map, recombines and transforms the features by weighted summation, and adjusts the number of channels of the feature map so that the output multi-scale features are more in line with the requirements of subsequent network modules in terms of dimension and feature representation. For example, if the subsequent network module (such as the residual attention network) requires the input feature map to have a specific number of channels and feature representation, the first convolution layer achieves this goal by adjusting the convolution kernel parameters.

[0104] During the convolution process, the parameter learning of the convolution kernel can automatically optimize the representation of features, highlight the features that are useful for denoising, and suppress unimportant information. Specifically, the weights in the convolution kernel will be continuously adjusted according to the training data, so that during the convolution operation, those features that are highly correlated with the denoising effect are strengthened, and those features that may be noise-related or have little effect on denoising are suppressed, thereby further improving the quality and discrimination of the features.

[0105] At the same time, in order to improve the stability and training efficiency of the model, batch normalization can be performed after the convolution operation of the first convolutional layer. Batch normalization normalizes the features after convolution so that the features have a more consistent distribution between different batches of data, reduces the training difficulties caused by differences in data distribution, and accelerates the network training process.

[0106] Finally, the features after the above processing are activated by the ReLU activation function to introduce nonlinear factors, enhance the model's ability to express complex feature relationships, and transform linearly inseparable problems into linearly separable ones, so that the network can learn more complex feature representations, and finally obtain multi-scale features and output them to subsequent network modules (such as residual attention network) for further processing.

[0107] The embodiment of the present application uses the channel adjustment and feature optimization processing of the first convolutional layer to make the output multi-scale features more in line with the requirements of subsequent network modules in terms of dimension and feature representation, ensure the smooth flow of data in the entire denoising model, and provide suitable input for subsequent denoising processing.

[0108] In an exemplary embodiment, based on the above embodiment, see Figure 5 The multi-scale extraction layer of the embodiment of the present application includes multiple second convolution layers and pooling layers; performing multi-scale convolution processing on the vibration signal to be processed through the multi-scale extraction layer to obtain multiple first feature maps includes the following steps:

[0109] S501, performing dilated convolution processing on the vibration signal to be processed through each second convolution layer and multiple expansion rates to obtain multiple initial feature maps.

[0110] In the embodiment of the present application, multiple second convolutional layers are set in the multi-scale extraction layer, and each second convolutional layer is configured with a different dilation rate for dilated convolution processing. For example, three second convolutional layers are set, and the dilation rates of their dilated convolutions are 3, 6, and 9 respectively.

[0111] S502, down-sampling each initial feature map through a pooling layer to obtain multiple down-sampled feature maps.

[0112] In the embodiment of the present application, after receiving the initial feature maps output by each second convolutional layer, the pooling layer performs downsampling processing on each initial feature map. The pooling layer can adopt a maximum pooling or an average pooling method.

[0113] Taking max pooling as an example, when an initial feature map is downsampled by max pooling, the feature map is divided into several local areas (for example, a two-dimensional feature map is divided into several small square areas). Then, the maximum value in the feature map is selected as the output in each local area. The advantage of this is that it can highlight the significant features in the feature map, while greatly reducing the amount of data and suppressing the influence of noise on the features to a certain extent, thereby obtaining a downsampled feature map. In this way, by performing max pooling downsampling on each initial feature map, multiple downsampled feature maps can be obtained. These downsampled feature maps retain the main feature information and reduce the resolution, making the amount of data for subsequent processing easier to manage and more conducive to further fusion of features.

[0114] If the average pooling method is used, the initial feature map will also be divided into several local areas, and then the average value of the feature values ​​in each local area will be calculated as the output. The role of average pooling is to smooth the feature map, reduce data redundancy, highlight the overall trend of the feature, and suppress noise to a certain extent. By performing average pooling downsampling processing on each initial feature map, multiple downsampled feature maps can also be obtained. These downsampled feature maps also have the above advantages and are convenient for subsequent feature splicing and other processing.

[0115] S503: Perform feature map splicing processing on each down-sampled feature map to obtain each first feature map.

[0116] The embodiment of the present application performs feature map splicing processing on multiple downsampled feature maps obtained after the downsampling processing of the pooling layer. Specifically, these downsampled feature maps are spliced ​​in a specific dimension (usually a channel dimension).

[0117] For example, suppose there are three downsampled feature maps, which are obtained from different second convolutional layers after being processed by the pooling layer. The three downsampled feature maps are arranged in sequence along the channel dimension, that is, the first downsampled feature map is placed in the first layer of the channel dimension, the second downsampled feature map is placed in the second layer, and the third downsampled feature map is placed in the third layer, thereby forming a new feature map, which is one of the first feature maps. In this way, all downsampled feature maps processed by the pooling layer are concatenated in this way, and multiple first feature maps can be obtained.

[0118] Each first feature map now represents the characteristics of the signal at different scales (determined by the hole convolution with different expansion rates) and different resolutions (determined by the downsampling of the pooling layer). As the output of the multi-scale extraction layer, they will be passed to the superposition layer for further processing.

[0119] The embodiment of the present application integrates downsampled feature maps of different scales and resolutions through feature map splicing processing to form multiple first feature maps, thereby providing comprehensive and diverse input for subsequent overlay layer processing.

[0120] In an exemplary embodiment, based on the above embodiment, see Figure 6 The residual attention network of the embodiment of the present application includes a multi-level first residual convolution layer, a multi-level attention layer and a multi-level second residual convolution layer. The multi-level first residual convolution layer and the multi-level second residual convolution layer correspond to each other from beginning to end, and the multi-level attention layer is connected across the input end of the corresponding first residual convolution layer and the output end of the second residual convolution layer; please refer to Figure 7 The embodiment of the present application relates to performing feature enhancement and feature fusion processing on multi-scale features through a residual attention network to obtain feature fusion data, including:

[0121] S601, input each first input feature into each level of first residual convolutional layer for feature extraction and residual connection processing, to obtain multiple first residual features.

[0122] Among them, the first input feature of the first residual convolution layer of the first stage is the multi-scale feature, and the first input feature of the first residual convolution layer of other stages is the first residual feature output by the first residual convolution layer of the previous stage.

[0123] In the embodiment of the present application, for the first residual convolution layer of the first stage, its first input feature is a multi-scale feature. This layer first uses a one-dimensional convolution to perform a convolution operation on the input multi-scale feature. The convolution kernel slides along the time dimension on the feature map, and the weights in the convolution kernel are multiplied and summed with the elements at the corresponding positions on the feature map to extract the local feature information of the signal.

[0124] After the convolution operation, batch normalization is performed to speed up the training process and improve the stability of the model. Batch normalization normalizes the extracted features so that the features have a more consistent distribution between different batches of data, reducing the training difficulties caused by differences in data distribution.

[0125] Next, the normalized features are activated through the ReLU activation function, which introduces nonlinear factors, enhances the model's ability to express complex feature relationships, and transforms linearly inseparable problems into linearly separable ones, so that the network can learn more complex feature representations and obtain intermediate features.

[0126] Finally, the intermediate features after the above processing are added to the multi-scale features of the original input through residual connections to form residual connections to obtain the first residual features of the first level. This residual connection method enables the network to learn residual mapping, which helps to avoid performance degradation caused by increasing network depth, while retaining more useful information in the original input features.

[0127] For the first residual convolution layer of other levels, its first input feature is the first residual feature output by the first residual convolution layer of the previous level. Similarly, according to the above steps, a one-dimensional convolution operation is first performed to extract local feature information, and then batch normalization and ReLU activation function processing are performed to obtain intermediate features, and then the intermediate features are added to the first residual features of the input through residual connection to obtain the first residual features of this level. Through such layer-by-layer progressive processing, each level further extracts and optimizes features on the basis of retaining the original useful information, and gradually digs deeper into the feature information of the signal.

[0128] S602, inputting the first input features of the first residual convolutional layer at each level into the attention layer at each level for feature enhancement processing to obtain multiple enhanced features.

[0129] In the embodiment of the present application, after receiving the first input feature of the corresponding first residual convolution layer, each attention layer performs feature enhancement processing according to its internal structure and operation process. Taking the first level attention layer as an example, the layer includes a feature extraction layer, a feature adjustment layer, and a feature scaling layer.

[0130] In the feature extraction layer, the first input feature of the first residual convolution layer is convolved using one-dimensional convolutions with different convolution kernel sizes (such as 3, 5, and 7). One-dimensional convolutions with a convolution kernel size of 3 can capture more local detail features of the signal in a fine manner, such as capturing some short-term, small frequency fluctuations or amplitude changes in vibration signals; one-dimensional convolutions with a convolution kernel size of 5 can obtain features of slightly larger local areas, and may detect a relatively macroscopic local feature pattern composed of several continuous small changes in the local area; one-dimensional convolutions with a convolution kernel size of 5 focus on local features in a larger range, such as capturing the vibration feature change trend of a certain component within a certain range. Through this parallel convolution operation with different convolution kernel sizes, richer local feature information can be extracted from the input features to obtain multi-scale local features.

[0131] The feature adjustment layer performs global feature fusion operations on multi-scale local features, such as compressing all local features in the spatial dimension through global average pooling to obtain a global feature description. Specifically, all feature values ​​of each channel are averaged to obtain a value representing the overall feature of the channel, thereby compressing the entire feature map into a vector in the spatial dimension. This vector contains the global feature information of each channel, realizing the global fusion of multi-scale local features.

[0132] Then, based on this global feature description, the fully connected layer is used to calculate the weight of each channel. The fully connected layer determines the importance of each channel feature in the entire feature representation by learning the relationship between the global feature description and the channel, and obtains the channel weight. For example, if the features of a channel show characteristics that are more relevant to the target vibration signal features in the global feature description, then the channel will be given a larger weight.

[0133] The feature scaling layer scales the multi-scale local features according to the input channel weights, and multiplies the feature value of each channel by the corresponding channel weight. In this way, important channel features (corresponding to larger weight values) will be enhanced, while less important channel features (corresponding to smaller weight values) will be suppressed, thereby achieving effective adjustment of multi-scale local features, highlighting useful signal features, suppressing noise information, and obtaining enhanced features at this level. Through such a series of operations at each level of attention layer, multiple enhanced features are finally obtained. Each enhanced feature not only retains the multi-scale characteristics of the input features, but also highlights the feature parts that are important for denoising.

[0134] S603: Input each second input feature into each level of second residual convolutional layer for fusion processing to obtain multiple second residual features.

[0135] Among them, the second input feature of the first-level second residual convolution layer is the first residual feature output by the last-level first residual convolution layer, the second input feature of the other-level second residual convolution layers is the second residual feature output by the previous-level second residual convolution layer and the enhanced feature output by the corresponding attention layer, and the feature fusion data includes the second residual feature output by the last-level second residual convolution layer and the enhanced feature output by the last-level attention layer.

[0136] In the embodiment of the present application, for the first-stage second residual convolution layer, its second input feature is the first residual feature output by the last-stage first residual convolution layer. This layer first uses a one-dimensional convolution to perform a convolution operation on the input first residual feature. The convolution kernel slides along the time dimension on the feature map, and the weights in the convolution kernel are multiplied and summed with the elements at the corresponding positions on the feature map to extract the local feature information of the signal. Similarly, the size and parameter settings of the convolution kernel will be adjusted according to the specific network design and the characteristics of the vibration signal to be processed to meet the feature extraction requirements of different scales and frequency components.

[0137] After the convolution operation, batch normalization is performed to normalize the extracted features so that the features have a more consistent distribution among different batches of data, reducing the training difficulties caused by differences in data streams.

[0138] Next, the normalized features are activated through the ReLU activation function, which introduces nonlinear factors, enhances the model's ability to express complex feature relationships, and transforms linearly inseparable problems into linearly separable ones, so that the network can learn more complex feature representations and obtain intermediate features.

[0139] Finally, the intermediate features after the above processing are added to the first residual features of the original input through residual connections to form residual connections to obtain the first-level second residual features. This residual connection method enables the network to learn residual mapping, which helps to avoid performance degradation caused by increasing network depth, while retaining more useful information in the original input features.

[0140] For the second residual convolution layer of other levels, its second input feature is the second residual feature output by the second residual convolution layer of the previous level and the enhanced feature output by the corresponding attention layer. First, the second residual feature output by the second residual convolution layer of the previous level and the enhanced feature output by the corresponding attention layer are merged, which can be a simple addition of corresponding elements or weighted addition according to certain weights, so as to merge the two into a comprehensive feature. Then, the comprehensive feature is convolved by one-dimensional convolution to extract the local feature information of the signal, and then batch normalization and ReLU activation function processing are performed to obtain the intermediate feature, and then the intermediate feature is added to the merged comprehensive feature through residual connection to form a residual connection to obtain the second residual feature of this level. Through such a layer-by-layer progressive processing, each level further extracts and optimizes the features on the basis of retaining the original useful information, gradually digs into the feature information of the signal, and continuously fuses the enhanced features from the attention layer, so that the multiple second residual features finally obtained integrate multiple feature information, laying the foundation for constructing feature fusion data.

[0141] In an exemplary embodiment, based on the above embodiment, see Figure 8 The embodiment of the present application involves inputting the first input feature of the first residual convolution layer of each level into the attention layer of each level for feature enhancement processing, and the process of obtaining multiple enhanced features includes the following steps:

[0142] S701: For each level of attention layer, the first input feature of the corresponding first residual convolution layer is input into the feature extraction layer for multi-scale local feature extraction processing to obtain multi-scale local features.

[0143] In an embodiment of the present application, the feature extraction layer in each level of attention layer uses one-dimensional convolution with different convolution kernel sizes (such as 3, 5, and 7) to perform a convolution operation on the first input feature of the corresponding first residual convolution layer.

[0144] S702, input the multi-scale local features into the feature adjustment layer to perform global feature fusion and channel weight calculation processing to obtain channel weights.

[0145] In the embodiment of the present application, the feature adjustment layer first performs a global feature fusion operation on the input multi-scale local features. A common method is to compress all local features in the spatial dimension through global average pooling to obtain a global feature description. Specifically, all feature values ​​of each channel are averaged to obtain a value representing the overall feature of the channel, thereby compressing the entire feature map into a vector in the spatial dimension. This vector contains the global feature information of each channel, realizing the global fusion of multi-scale local features.

[0146] Then, based on this global feature description, the fully connected layer is used to calculate the weight of each channel. The fully connected layer determines the importance of each channel feature in the entire feature representation by learning the relationship between the global feature description and the channel, and obtains the channel weight. Specifically, the neurons in the fully connected layer are connected to the global feature description of each channel. Through the back propagation algorithm during the training process, the connection weights between neurons are continuously adjusted to adapt to the different input features, thereby determining the weight value that should be assigned to each channel. For example, if the features of a channel show characteristics that are more relevant to the target vibration signal features in the global feature description, then the channel will be assigned a larger weight.

[0147] S703, input the multi-scale local features and channel weights into the feature scaling layer to perform feature scaling operation to obtain enhanced features.

[0148] In the embodiment of the present application, the feature scaling layer performs a scaling operation on the multi-scale local features according to the input channel weights. Specifically, for the feature map of each channel, all its element values ​​are multiplied by the corresponding channel weights.

[0149] In this way, important channel features (corresponding to larger weight values) will be enhanced, while less important channel features (corresponding to smaller weight values) will be suppressed, thereby achieving effective adjustment of multi-scale local features, highlighting useful signal features, suppressing noise information, and finally obtaining enhanced features after enhancement processing. At this time, the enhanced features, while retaining the multi-scale local features, focus more on the feature parts that are important for denoising, providing more valuable input for subsequent fusion processing with other features.

[0150] In an exemplary embodiment, based on the above embodiment, see Fig. 9 The first residual convolution layer of the embodiment of the present application includes a convolution layer and a jump connection layer, and each first input feature is input into each level of the first residual convolution layer for feature extraction and residual connection processing to obtain multiple first residual features, including the following steps:

[0151] S801, input the first input feature into the convolution layer for local feature extraction processing to obtain an intermediate feature.

[0152] In the embodiment of the present application, when the first input feature is input to the convolution layer, the convolution layer performs a convolution operation on it using a one-dimensional convolution. The convolution kernel of the one-dimensional convolution slides along the time dimension on the feature map and performs a weighted sum operation with the elements at the corresponding position on the feature map.

[0153] For example, for a feature map corresponding to a vibration signal, the convolution kernel will start from the starting position of the feature map and slide in sequence according to the set step size. At each sliding position, the weight in the convolution kernel is multiplied by the element at the corresponding position on the feature map, and then these products are added to obtain a new value, which is the result of the convolution operation at that position. By continuously sliding the convolution kernel through the entire feature map, the feature map after the convolution operation, that is, the intermediate feature, can be obtained.

[0154] S802, adding the intermediate feature to the multi-scale feature through a skip connection layer to obtain a first residual feature.

[0155] In the embodiment of the present application, the jump connection layer realizes the addition operation of the intermediate features processed by the convolution layer and the multi-scale features of the original input. Specifically, in the network structure, the intermediate features and the multi-scale features are element-wise added at corresponding positions.

[0156] For example, assuming that the intermediate features and the multi-scale features are both in the form of two-dimensional matrices, then for each element at a corresponding position in the matrix, the element at that position of the intermediate feature is added to the element at that position of the multi-scale feature to obtain a new element. The new elements after the addition of all corresponding positions form a new matrix, and this new matrix is ​​the first residual feature.

[0157] This residual connection method enables the network to learn residual mapping, that is, the difference between output and input. During network training, when the network depth increases, traditional convolutional networks are prone to problems such as gradient disappearance or explosion, resulting in performance degradation. Through this residual connection, the network can more easily adjust parameters and learn useful feature representations because it directly focuses on the difference between output and input, rather than relying solely on the output of the convolutional layer.

[0158] At the same time, this connection method retains a lot of useful information in the original input features. Since the intermediate features are added to the multi-scale features of the original input, a lot of information in the original input features is retained, avoiding the loss of important information due to multiple convolutions and nonlinear transformations during the deep feature extraction process, laying a solid foundation for subsequent deep feature extraction and fusion.

[0159] In an exemplary embodiment, based on the above embodiment, the method of the embodiment of the present application further includes the following steps:

[0160] Step 1: Acquire a vibration signal to be processed including a noise signal and a target vibration signal;

[0161] Step 2: Input the vibration signal to be processed into the denoising model. The pyramid network in the denoising model inputs the vibration signal to be processed into each second convolution layer in its multi-scale extraction layer and multiple expansion rates to perform hole convolution processing on the vibration signal to be processed to obtain multiple initial feature maps; downsample each initial feature map through the pooling layer in the multi-scale extraction layer to obtain multiple downsampled feature maps; perform feature map splicing processing on each downsampled feature map to obtain each first feature map; fuse each first feature map through the superposition layer to obtain a fused feature; perform channel adjustment and feature optimization processing on the fused feature through the first convolution layer to obtain a multi-scale feature;

[0162] Step 3: Input each first input feature into the residual attention network in the denoising model, and perform local feature extraction processing on the convolution layer of the first residual convolution layer of each level of the residual attention network to obtain the intermediate feature; add the intermediate feature to the multi-scale feature through the jump connection layer of the first residual convolution layer to obtain the first residual feature;

[0163] Step 4: For each level of attention layer in the residual attention network, the first input feature of the corresponding first residual convolution layer is input into the feature extraction layer for multi-scale local feature extraction processing to obtain multi-scale local features; the multi-scale local features are input into the feature adjustment layer for global feature fusion and channel weight calculation processing to obtain channel weights; the multi-scale local features and channel weights are input into the feature scaling layer for feature scaling operation to obtain enhanced features;

[0164] Step 5: Input each second input feature into the second residual convolution layer at each level for fusion processing to obtain feature fusion data including the second residual features output by the last second residual convolution layer and the enhanced features output by the last attention layer; perform feature mapping processing on the feature fusion data through the output network to obtain the target vibration signal.

[0165] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0166] Based on the same inventive concept, the embodiment of the present application also provides a vibration signal processing device for implementing the vibration signal processing method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more vibration signal processing device embodiments provided below can refer to the limitations on the vibration signal processing method above, and will not be repeated here.

[0167] In one embodiment, Fig.10 As shown, a vibration signal processing device 900 is provided, comprising:

[0168] The signal acquisition module 901 is used to acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal;

[0169] The denoising module 902 is used to perform multi-scale feature extraction on the vibration signal to be processed using a pre-built denoising model, remove the noise signal to obtain multi-scale features, perform feature enhancement and feature fusion processing on the multi-scale features to obtain feature fusion data, and perform feature mapping processing on the feature fusion data to obtain the target vibration signal.

[0170] In one embodiment, the denoising model includes a pyramid network, a residual attention network and an output network; the denoising module includes:

[0171] An extraction unit is used to perform multi-scale feature extraction processing on the vibration signal to be processed through a pyramid network to obtain multi-scale features;

[0172] A fusion unit is used to perform feature enhancement and feature fusion processing on multi-scale features through a residual attention network to obtain feature fusion data;

[0173] The mapping unit is used to perform feature mapping processing on the feature fusion data through the output network to obtain the target vibration signal.

[0174] In one of the embodiments, the pyramid network comprises a multi-scale extraction layer, an overlay layer and a first convolution layer connected in sequence, and the extraction unit comprises:

[0175] A convolution subunit, configured to perform multi-scale convolution processing on the vibration signal to be processed through a multi-scale extraction layer to obtain a plurality of first feature maps;

[0176] A fusion subunit, used for fusing the first feature maps through a stacking layer to obtain a fusion feature;

[0177] The optimization subunit is used to perform channel adjustment and feature optimization processing on the fusion features through the first convolutional layer to obtain multi-scale features.

[0178] In one of the embodiments, the multi-scale extraction layer includes multiple second convolution layers and pooling layers; the convolution subunit is specifically used to: perform hole convolution processing on the vibration signal to be processed through each second convolution layer and multiple expansion rates to obtain multiple initial feature maps; downsample each initial feature map through the pooling layer to obtain multiple downsampled feature maps; perform feature map splicing processing on each downsampled feature map to obtain each first feature map.

[0179] In one embodiment, the residual attention network includes a multi-level first residual convolution layer, a multi-level attention layer and a multi-level second residual convolution layer, the multi-level first residual convolution layer and the multi-level second residual convolution layer correspond to each other, and the multi-level attention layer is connected across the input end of the corresponding first residual convolution layer and the output end of the second residual convolution layer; the fusion unit includes:

[0180] The residual processing subunit is used to input each first input feature into the first residual convolution layer of each level for feature extraction and residual connection processing to obtain multiple first residual features; wherein the first input feature of the first residual convolution layer of the first level is a multi-scale feature, and the first input feature of the first residual convolution layer of other levels is the first residual feature output by the first residual convolution layer of the previous level;

[0181] A feature enhancement subunit is used to input the first input feature of the first residual convolution layer at each level into the attention layer at each level for feature enhancement processing to obtain multiple enhanced features;

[0182] A fusion subunit is used to input each second input feature into each level of the second residual convolution layer for fusion processing to obtain multiple second residual features; wherein the second input feature of the first-level second residual convolution layer is the first residual feature output by the last-level first residual convolution layer, and the second input features of the other levels of the second residual convolution layer are the second residual features output by the previous level of the second residual convolution layer and the enhanced features output by the corresponding attention layer, and the feature fusion data includes the second residual features output by the last-level second residual convolution layer and the enhanced features output by the last-level attention layer.

[0183] In one of the embodiments, the feature enhancement subunit is further specifically used for: for each level of attention layer, inputting the first input feature of the corresponding first residual convolution layer into the feature extraction layer for multi-scale local feature extraction processing to obtain multi-scale local features; inputting the multi-scale local features into the feature adjustment layer for global feature fusion and channel weight calculation processing to obtain channel weights; inputting the multi-scale local features and channel weights into the feature scaling layer for feature scaling operations to obtain enhanced features.

[0184] In one of the embodiments, the above-mentioned first residual convolution layer includes a convolution layer and a jump connection layer, and the residual processing subunit is specifically used to: input the first input feature into the convolution layer for local feature extraction processing to obtain an intermediate feature; add the intermediate feature and the multi-scale feature through the jump connection layer to obtain a first residual feature.

[0185] Each module in the vibration signal processing device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0186] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig.11 As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a vibration signal processing method is realized. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0187] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0188] According to some embodiments of the present application, a computer device including instructions is also provided, such as a memory and a processor including instructions. The instructions can be executed by the processor of the computer device to complete the above method.

[0189] According to some embodiments of the present application, a computer program product is also provided, and when the computer program is executed by a processor, the above method can be implemented. The computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, part or all of the above method can be implemented in whole or in part according to the process or function described in the embodiment of the present application.

[0190] According to some embodiments of the present application, there is also provided a non-transitory computer-readable storage medium including instructions, such as a memory including instructions, and the above instructions can be executed by a processor of an electronic device to complete the above method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0191] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0192] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0193] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A vibration signal processing method, characterized in that: The method comprises: Acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal; A pre-built denoising model is used to perform multi-scale feature extraction on the vibration signal to be processed, the noise signal is removed to obtain multi-scale features, feature enhancement and feature fusion processing are performed on the multi-scale features to obtain feature fusion data, and feature mapping processing is performed on the feature fusion data to obtain the target vibration signal.

2. The method according to claim 1, characterized in that The denoising model includes a pyramid network, a residual attention network and an output network; the denoising model constructed in advance is used to extract multi-scale features from the vibration signal to be processed to obtain multi-scale features, the multi-scale features are subjected to feature fusion processing to obtain feature fusion data, and the feature fusion data is subjected to feature mapping processing to obtain the target vibration signal, including: Performing multi-scale feature extraction processing on the vibration signal to be processed by the pyramid network to obtain the multi-scale feature; Performing feature enhancement and feature fusion processing on the multi-scale features through the residual attention network to obtain the feature fusion data; The feature fusion data is subjected to feature mapping processing through the output network to obtain the target vibration signal.

3. The method according to claim 2, characterized in that The pyramid network includes a multi-scale extraction layer, a stacking layer and a first convolution layer connected in sequence, and the multi-scale feature extraction process is performed on the vibration signal to be processed by the pyramid network to obtain the multi-scale feature, including: Performing multi-scale convolution processing on the vibration signal to be processed through the multi-scale extraction layer to obtain a plurality of first feature maps; Performing a fusion process on each of the first feature maps through the stacking layer to obtain a fusion feature; The fusion feature is subjected to channel adjustment and feature optimization processing through the first convolution layer to obtain the multi-scale feature.

4. The method according to claim 3, characterized in that The multi-scale extraction layer includes a plurality of second convolution layers and a pooling layer; the multi-scale convolution processing is performed on the vibration signal to be processed by the multi-scale extraction layer to obtain a plurality of first feature maps including: Performing hole convolution processing on the vibration signal to be processed through each of the second convolution layers and multiple expansion rates to obtain multiple initial feature maps; Downsampling each of the initial feature maps through the pooling layer to obtain multiple downsampled feature maps; Perform feature map splicing processing on each of the downsampled feature maps to obtain each of the first feature maps.

5. The method according to claim 2, characterized in that: The residual attention network includes a multi-level first residual convolution layer, a multi-level attention layer and a multi-level second residual convolution layer, wherein the multi-level first residual convolution layer and the multi-level second residual convolution layer correspond to each other end to end, and the multi-level attention layer is connected across the input end of the corresponding first residual convolution layer and the output end of the second residual convolution layer; The performing feature enhancement and feature fusion processing on the multi-scale features through the residual attention network to obtain the feature fusion data includes: Input each first input feature into the first residual convolution layer at each level for feature extraction and residual connection processing to obtain multiple first residual features; wherein the first input feature of the first residual convolution layer at the first level is the multi-scale feature, and the first input feature of the first residual convolution layer at other levels is the first residual feature output by the first residual convolution layer at the previous level; Inputting the first input features of the first residual convolutional layers at each level into the attention layers at each level for feature enhancement processing to obtain multiple enhanced features; Each second input feature is input into each level of the second residual convolution layer for fusion processing to obtain multiple second residual features; wherein, the second input feature of the first-level second residual convolution layer is the first residual feature output by the last-level first residual convolution layer, and the second input features of the other levels of the second residual convolution layer are the second residual features output by the previous level of the second residual convolution layer and the enhanced features output by the corresponding attention layer, and the feature fusion data includes the second residual features output by the last-level second residual convolution layer and the enhanced features output by the last-level attention layer.

6. The method according to claim 5, characterized in that The first input features of the first residual convolutional layers at each level are input into the attention layers at each level for feature enhancement processing to obtain multiple enhanced features, including: For each level of the attention layer, the first input feature of the corresponding first residual convolution layer is input into the feature extraction layer for multi-scale local feature extraction processing to obtain multi-scale local features; Inputting the multi-scale local features into a feature adjustment layer to perform global feature fusion and channel weight calculation processing to obtain channel weights; The multi-scale local features and the channel weights are input into a feature scaling layer to perform a feature scaling operation to obtain the enhanced features.

7. The method according to claim 5, characterized in that The first residual convolution layer includes a convolution layer and a jump connection layer, and each first input feature is input into each level of the first residual convolution layer for feature extraction and residual connection processing to obtain multiple first residual features, including: Inputting the first input feature into the convolution layer to perform local feature extraction processing to obtain an intermediate feature; The intermediate feature and the multi-scale feature are added together through the jump connection layer to obtain the first residual feature.

8. A vibration signal processing device, characterized in that: The device comprises: A signal acquisition module, used to acquire a vibration signal to be processed; the vibration signal to be processed includes a noise signal and a target vibration signal; The denoising module is used to perform multi-scale feature extraction on the vibration signal to be processed by using a pre-built denoising model, remove the noise signal to obtain multi-scale features, perform feature enhancement and feature fusion processing on the multi-scale features to obtain feature fusion data, and perform feature mapping processing on the feature fusion data to obtain the target vibration signal.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images