Signal noise reduction method, system and device adaptive to time-frequency loss and storage medium

Abstract

The invention relates to the technical field of rotating machine fault signal noise reduction, and provides a signal noise reduction method, system and device adaptive to time-frequency loss and a storage medium, and the method comprises the steps: obtaining a rotating machine vibration simulation signal, and constructing an original data set; filling Gaussian white noise with different noise levels based on the original data set to obtain a noise pollution data set; stacking a plurality of convolutional layers through residual connection and jump connection based on the noise pollution data set, and constructing a ResTransUnet model; and performing noise reduction processing on the noise pollution data set based on a ResTransUnet model, and outputting a noise reduction signal. In this way, deep abstract features of noise signals are extracted by stacking a plurality of convolutional layers, local features and global features in vibration signals are effectively extracted and fused, a ResTransUnet model can better balance local and global information in a one-dimensional signal noise reduction task, the generalization ability of the noise reduction effect is improved, and the noise reduction efficiency is improved. And the noise reduction performance of the model under different noise levels is improved, so that the fault signal denoising effect is enhanced.

Description

Technical Field

[0001] The present disclosure belongs to the technical field of noise reduction for rotating machinery fault signals, and particularly relates to a signal noise reduction method, system, device, and storage medium with adaptive time-frequency loss. Background Art

[0002] In recent years, with the booming development of intelligent sensor technology and industrial Internet of Things, the industrial field has accumulated a large amount of high-dimensional, complex, and diverse data. In the actual industrial monitoring process, the collected signals are often interfered by various noises, posing challenges to fault diagnosis.

[0003] The sources of these noises mainly include complex industrial environments, problems with the accuracy and stability of sensors, assembly errors of mechanical structures, and changes in working conditions. When a device fails or is assembled abnormally, inevitable abnormal noises often occur. The abnormal noises generated by the fault are mixed with the background noise, resulting in distortion of the signal waveform, masking the characteristics of the fault signal, and increasing the difficulty of fault feature extraction.

[0004] Existing denoising methods include signal processing methods and data-driven methods. Traditional signal processing methods using signal processing usually require relying on expert knowledge to select appropriate filtering parameters, such as the mother wavelet function or threshold setting of wavelet transform. This subjective selection makes the noise reduction effect unstable and the generalization ability poor. Although the data-driven method has achieved good diagnostic accuracy in noise scenarios, this method mainly focuses on enhancing the anti-noise interference ability of the model rather than the noise reduction ability. In summary, it can be concluded that the method for denoising fault signals needs to be improved. Summary of the Invention

[0005] To solve the above problems, the present disclosure provides a signal noise reduction method, system, device, and storage medium with adaptive time-frequency loss, which combines the efficient local feature extraction ability of convolutional networks and the global feature modeling ability of Transformers, enabling the model to better balance local and global information in one-dimensional signal noise reduction tasks, making the noise reduction effect more stable, and thus improving the noise reduction ability.

[0006] In the first aspect of the present disclosure, a signal noise reduction method with adaptive time-frequency loss is provided. The method includes: obtaining a rotating machinery vibration simulation signal and constructing an original data set; filling Gaussian white noise with different noise levels based on the original data set to obtain a noise-polluted data set; stacking multiple convolutional layers through residual connections and skip connections based on the noise-polluted data set, and constructing a ResTransUnet model; performing noise reduction processing on the noise-polluted data set based on the ResTransUnet model and outputting a denoised signal.

[0007] With such settings, under residual connections and skip connections, by stacking multiple convolutional layers, deep abstract features of the noise signal are extracted, effectively extracting and fusing local and global features in the vibration signal, enabling the ResTransUnet model to better balance local and global information in the one-dimensional signal denoising task, improving the generalization ability of the denoising effect, and enhancing the denoising performance of the model under different noise levels, thereby enhancing the denoising effect on the fault signal.

[0008] In some embodiments, building the ResTransUnet model by stacking multiple convolutional layers based on the noise-polluted dataset through residual connections and skip connections includes: the ResTransUnet model includes an upsampling part, a downsampling part, and an intermediate layer; the upsampling and the downsampling are symmetrically designed, and the noise-polluted data is stacked through skip connections to form multiple convolutional layers; the intermediate layer is used to make a residual connection between the upsampling and the downsampling.

[0009] With such settings, in the downsampling, max-pooling operations are used between each module to gradually downsample the temporal features of the signal, compress the spatial dimension, and retain important features; the intermediate layer is connected through an improved residual structure, enabling the model to better capture multi-level features in the signal and enhancing the denoising effect.

[0010] In some embodiments, the downsampling includes a double-convolution module and a Transconv module. The double-convolution module is used to extract local features in the noise-polluted dataset, and the Transconv module is used to upsample the low-resolution feature map in the noise-polluted dataset to a high-resolution feature map.

[0011] With such settings, the combined use of the double-convolution module and the Transconv module enables the model to effectively extract and fuse local and global features in the vibration signal, thereby improving the denoising performance of the model under different noise levels.

[0012] In some embodiments, the double-convolution module includes at least two sequentially stacked convolutional, batch normalization, and Relu activation layers, and the formula is as follows:

[0013]

[0014] σ relu = max(0, x)

[0015] where μ and σ 2 respectively represent the mean and variance of this batch, γ and β are learnable parameters of the model, and ∈ is a very small value.

[0016] With such settings, it is convenient to extract deep abstract features of the noise signal.

[0017] In some embodiments, the downsampling further includes a Transformer module, which is used to dynamically assign different weights to each part of the noise signal in the noise-polluted data set, and capture the global features in the noise-polluted data set.

[0018] With such a setting, the perception ability of the model for long-distance features is enhanced, which helps to eliminate the noise of remote associations, thereby enhancing the ability of the model to extract global context information.

[0019] In some embodiments, denoising the noise-polluted data set based on the ResTransUnet model, and the output denoised signal includes: defining a loss function to initialize the ResTransUnet model; adjusting the learning rate during the model training process based on the initialized ResTransUnet model using a cosine annealing warm-up strategy; denoising the noise-polluted data set based on the adjusted ResTransUnet model and then outputting a denoised signal.

[0020] With such a setting, by combining the advantages of the time domain and the frequency domain through the loss function, the model can better adapt to different noise types and signal forms, and improve the generalization ability in different scenarios; through the cosine annealing warm-up strategy, the model can better learn coarse-grained global features in the early stage, and fine-tune local features by reducing the learning rate in the later stage, which helps the model balance the overall denoising effect and detail retention.

[0021] In some embodiments, the method further includes: training the denoised signal based on the ResTransUnet model to generate a benchmark deep learning model; performing fault diagnosis on the actually collected vibration signal based on the benchmark deep learning model.

[0022] With such a setting, it is convenient to diagnose the high-quality signal denoised by the ResTransUnet model, making the diagnosis of the denoised vibration signal more accurate and enhancing the diagnostic effect of the model.

[0023] The second aspect of the present disclosure provides a signal denoising system with an adaptive time-frequency loss, and the system includes a module for executing the signal denoising method with an adaptive time-frequency loss as described in the first aspect.

[0024] The technical effects brought by any possible implementation manner in the second aspect can be referred to the technical effects brought by the first aspect above, and will not be elaborated here.

[0025] The third aspect of the present disclosure provides an electronic device, and the electronic device includes: a memory for storing a computer program; a processor for implementing the signal denoising method with an adaptive time-frequency loss as described in the first aspect when executing the computer program.

[0026] For the technical effects brought about by any possible implementation manner of the third aspect, reference may be made to the technical effects brought about by the first aspect above, and details are not described herein again.

[0027] The fourth aspect of the present disclosure provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the signal noise reduction method with adaptive time-frequency loss as described in the first aspect.

[0028] For the technical effects brought about by any possible implementation manner of the fourth aspect, reference may be made to the technical effects brought about by the first aspect above, and details are not described herein again.

[0029] Compared with the prior art, the present disclosure has the following advantages:

[0030] ①Combining the efficient local feature extraction ability of the convolutional network and the global feature modeling ability of the Transformer enables the model to better balance local and global information in the one-dimensional signal noise reduction task, thereby improving the noise reduction effect.

[0031] ②The middle layer of the model is connected through an improved residual structure, which can better capture multi-level features in the signal and avoid overfitting caused by too many network layers.

[0032] ③During the training process, the cosine warm-up strategy is used to adaptively adjust the learning rate.

[0033] ④In order to make the model pay more attention to the frequency domain features of the noise reduction signal, a time-frequency domain hybrid loss function is constructed to optimize the training process of the model and retain the frequency information of the noise reduction signal.

[0034] Other features and advantages of the present disclosure will be described in the subsequent specification, and part of them will be obvious from the specification, or will be understood by implementing the present disclosure. The objectives and other advantages of the present disclosure can be achieved and obtained through the structures pointed out in the specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present disclosure. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a flowchart of the signal noise reduction method with adaptive time-frequency loss provided by the embodiment of the present disclosure;

[0037] Figure 2 The flowchart of fault diagnosis based on adaptive time-frequency loss under noise interference provided by the embodiments of the present disclosure;

[0038] Figure 3 The schematic diagram of the ResTransUnet model provided by the embodiments of the present disclosure;

[0039] Figure 4 The graph of the change of the loss value and the learning rate during the training process of the model provided by the embodiments of the present disclosure;

[0040] Figure 5 The comparison diagram before and after noise reduction of mild wear faults under -5dBSNR noise provided by the embodiments of the present disclosure;

[0041] Figure 6 The comparison diagram before and after noise reduction of mild wear faults under -5dBSNR noise provided by the embodiments of the present disclosure;

[0042] Figure 7 The probability density distribution curve graph of healthy signals under different noise interferences provided by the embodiments of the present disclosure;

[0043] Figure 8 The diagnostic result graph after noise reduction of different models provided by the embodiments of the present disclosure;

[0044] Figure 9 The confusion matrix graph after diagnosing the signals after noise reduction of different models provided by the embodiments of the present disclosure;

[0045] Figure 10 The block diagram of the signal noise reduction system with adaptive time-frequency loss provided by the embodiments of the present disclosure;

[0046] Figure 11 The block diagram of the structure of an electronic device provided by the embodiments of the present disclosure. Detailed implementation manners

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[0048] Some concepts that may be involved in the embodiments of the present disclosure are briefly introduced below.

[0049] Rotating machinery is the core power source of industrial production, and the stability and reliability of its performance are crucial for the efficient operation of the entire industrial system. Among these machines, rolling bearings and gearboxes are indispensable core components, which are mainly responsible for supporting the rotational motion of the machinery. A large number of studies have shown that the failures of rolling bearings and gearboxes are the main reasons for the shutdown of rotating machinery. Therefore, accurate fault diagnosis of rolling bearings and gearboxes is of great significance for ensuring the continuous operation of mechanical equipment and promoting the stable and safe production of modern industry.

[0050] In the actual industrial monitoring process, the collected signals are often disturbed by various noises, posing challenges to fault diagnosis. The sources of these noises mainly include the complex industrial environment, problems with the accuracy and stability of sensors, assembly errors of mechanical structures, and changes in working conditions. In addition, when the equipment fails or is assembled abnormally, inevitable abnormal noises often occur. The combined effect of these factors makes the fault signals mixed with the background noise, resulting in the distortion of the signal waveform, masking the characteristics of the fault signals, and increasing the difficulty of extracting fault features. To effectively process and remove noises, extract the key features in the fault signals, and conduct fault diagnosis, currently, signal processing methods and data-driven methods are commonly used.

[0051] Signal processing methods are based on prior knowledge of the characteristics of signals and noises, and use mathematical and statistical techniques to improve the quality of signals. Common ones include Fourier transform, wavelet transform, Kalman filter, median filter, etc. Data-driven methods focus on learning the distribution patterns of signals and noises from a large amount of data, reducing the dependence on expert experience, and usually involve machine learning or deep learning techniques. Data-driven methods rely on the powerful nonlinear modeling ability of deep learning models, can handle complex nonlinear relationships and non-stationary signals, are more suitable for denoising large-scale data sets, and have strong adaptability to new signals and noises in different environments.

[0052] However, traditional signal processing methods usually need to rely on expert knowledge to select appropriate filtering parameters, such as the mother wavelet function or threshold setting of wavelet transform. This subjective selection makes the noise reduction effect unstable and the generalization ability poor. Different noise levels or scenarios may require different parameter screening, and the process is very complex and cumbersome. In addition, traditional signal processing methods perform poorly in dealing with complex noises or non-stationary signals. Although data-driven methods have achieved good diagnostic accuracy in noise scenarios, this method mainly focuses on enhancing the anti-noise interference ability of the model rather than the noise reduction ability. This means that although the model can make accurate judgments in a noisy environment, it is difficult to generate a clear signal after noise reduction for further analysis, thus limiting the interpretability and application value of the method.

[0053] Therefore, this application proposes a signal denoising method with adaptive time-frequency loss, seeFigure 1 and Figure 2 , Figure 1 is the flowchart of the signal denoising method with adaptive time-frequency loss provided by the embodiments of the present disclosure, Figure 2 is the flowchart of fault diagnosis based on adaptive time-frequency loss under noise interference provided by the embodiments of the present disclosure. This method combines the efficient local feature extraction ability of the convolutional network and the global feature modeling ability of the Transformer. By constructing a ResTransUnet model for one-dimensional time-domain signal denoising, it can better balance local and global information, thereby improving the denoising effect. The intermediate layer of the model is connected through an improved residual structure to better capture multi-level features in the signal, while avoiding overfitting caused by too many network layers. During the training process, a cosine warm-up strategy is used to adaptively adjust the learning rate. In order to make the model pay more attention to the frequency-domain features of the denoised signal, a time-frequency domain hybrid loss function is constructed to optimize the training process of the model and retain the frequency information of the denoised signal. The denoising method includes the following steps:

[0054] S100. Obtain the vibration simulation signal of the rotating machinery and construct the original data set.

[0055] In this embodiment, the gear vibration signals of the rotating machinery simulation are collected through a fault test bench in a quiet environment and under different faults, and the original data set is constructed.

[0056] S200. Fill Gaussian white noise with different noise levels based on the original data set to obtain a noise-polluted data set.

[0057] In this embodiment, Gaussian white noise is added to the original data set to simulate the noise environment and form a noise-polluted data set. Specifically, the signal-to-noise ratio (SNR) index is used to measure the degree of environmental noise. The smaller the SNR, the smaller the proportion of the useful signal in the total signal and the more serious the noise. The formula is as follows:

[0058] SNR dB = 10 log(P signal / P noise ) 1)

[0059] where P signal signal represents the power of the original signal, and P noise noise represents the power of the Gaussian noise.

[0060] S300. Stack multiple convolutional layers based on the noise-polluted data set through residual connections and skip connections, and construct a ResTransUnet model.

[0061] In this embodiment, refer to Figure 2 and Figure 3 , Figure 3Schematic diagram of the ResTransUnet model provided by an embodiment of the present disclosure. The ResTransUnet model includes an upsampling part, a downsampling part, and an intermediate layer. The upsampling and downsampling are symmetrically designed, and multiple convolutional layers are formed by stacking noise-polluted data without data through skip connections. The symmetric design can use max-pooling operations between each module in the downsampling to gradually downsample the temporal features of the signal, compress the spatial dimension, and retain important features. The intermediate layer is used to perform residual connection between the upsampling and downsampling.

[0062] Exemplarily, the downsampling includes a double convolutional module and a Transconv module. The double convolutional module is used to extract local features in the noise-polluted dataset, and the Transconv module is used to upsample the low-resolution feature map in the noise-polluted dataset to a high-resolution feature map.

[0063] For example, the double convolutional module includes at least two sequentially stacked convolutional layers, batch normalization layers, and Relu activation layers to extract deep abstract features of the noise signal. The mathematical formula is as follows:

[0064]

[0065] σ relu =max(0,x) 4)

[0066] where μ and σ 2 represent the mean and variance of the batch respectively, γ and β are parameters that can be learned by the model, and ∈ is a very small value.

[0067] In some embodiments, the downsampling further includes a Transformer module, which is used to dynamically assign different weights to each part of the noise signal in the noise-polluted dataset to capture the global features in the noise-polluted dataset.

[0068] In this embodiment, a Transformer module is introduced after the convolutional module to enhance the model's ability to extract global context information. The multi-head self-attention mechanism in the Transformer module can dynamically assign different weights to each part of the signal, enhance the model's perception ability of distant features, and help eliminate noise with long-range associations.

[0069] Exemplarily, assume that the feature map output after feature extraction by the double convolutional module is where B is the batch size, N is the sequence length, and D is the dimension of the input. First, Y is mapped to queries, keys, and values through a linear transformation.

[0070] Q = YW Q , K = YW K , V = YW V 5)

[0071] Among them, is a linear transformation matrix, and N h is the number of attention heads, and d h is the dimension of each attention head.

[0072] Furthermore, calculate the scaled dot-product attention for each head:

[0073]

[0074] where is the scaling factor, which is used to prevent the gradient of the softmax output from vanishing due to the excessive dot-product value. W O is a trainable mapping matrix.

[0075] In some embodiments, the intermediate layer is composed of 2 stacked residual modules. By directly passing the input to the subsequent layers through residual connections, it can help the gradient to be more easily backpropagated in the network, avoiding gradient vanishing or explosion.

[0076] In this embodiment, for the complex signal after noise interference, the residual module can simultaneously learn low-order and high-order features at different levels. Among them, the low-order features retain the basic structural information of the original signal, and the high-order features retain the characteristic patterns of the noise or the details of the signal. The fusion of such multi-level features enables the model to better capture the feature information in the signal from different scales, retain the signal detail texture, thereby improving the noise reduction effect. For example:

[0077] H(x) = F(x) + x 8)

[0078] where x and F(x) are the low-order and high-order features respectively.

[0079] Exemplarily, the skip connections in the ResTransUnet model combine the high-resolution features in the encoder with the low-resolution features in the decoder.

[0080] In this embodiment, the encoder is the downsampling part, and the decoder is the upsampling part. The encoder extracts the high-level features in the vibration signal through downsampling operations, such as the overall trend and main frequency components of the signal. The decoder restores these high-level features to the resolution of the original signal through upsampling operations, and at the same time combines the high-resolution features transmitted by the skip connections to retain the detail information in the signal. Therefore, the high-resolution features in the encoder are combined with the low-resolution features in the decoder to ensure that both noise can be removed and important details in the signal can be retained during the noise reduction process, which helps to restore more detail information. For example:

[0081]

[0082] where F l and They are the output feature maps of the downsampling and upsampling modules in the l-th layer respectively.

[0083] Finally, the feature map is converted to the target number of channels through a 1×1 convolution to output the denoised signal.

[0084] S400. Denoise the noise-polluted data set based on the ResTransUnet model and output the denoised signal.

[0085] In this embodiment, a loss function is defined to initialize the ResTransUnet model; the cosine annealing warm-up strategy is used to adjust the learning rate during the training process of the model based on the initialized ResTransUnet model; the noise-polluted data set is denoised based on the adjusted ResTransUnet model and the denoised signal is output.

[0086] Exemplarily, when constructing and training the ResTransUnet model, a hybrid loss function is used to train the model. The hybrid loss function combines the advantages of the time domain and the frequency domain. By simultaneously optimizing the errors in the time and frequency domains, the model can better adapt to different noise types and signal morphologies, thereby improving the generalization ability in different scenarios; it not only ensures the restoration degree of the signal in the time domain but also retains the frequency structure of the signal, which can improve the denoising ability of the model, especially enhancing the perceptual quality and spectral fidelity.

[0087] For example,

[0088]

[0089] where represents the time domain loss, represents the frequency domain loss represents the hybrid loss, α and β are control weights, and fft is the Fourier transform. α(t) and β(t) are adaptive weights respectively, and ∈ is a small constant used to avoid the denominator being zero.

[0090] Exemplarily, when constructing and training the ResTransUnet model, the cosine annealing warm-up strategy is used to adaptively adjust the learning rate, so that the model can better learn the coarse-grained global features in the early stage, and fine-tune the local features by reducing the learning rate in the later stage, which helps the model balance the overall denoising effect and detail retention.

[0091] The warm-up strategy enables the model to learn with a relatively small step size in the initial stage, avoiding overly large gradient updates and preventing gradient explosion. The method of gradually increasing the learning rate allows the model to converge more easily in the initial stage of training. In the later stage of training, gradually decreasing the learning rate enables the model to perform more refined parameter adjustment near the optimal solution and avoid falling into local optima. Thus, it can be seen that cosine annealing can enable the model to explore the parameter space at different learning rate levels by periodically decreasing the learning rate, enabling the model to adapt to noise changes during training, especially on complex or non-uniform noisy data, and enhancing the generalization ability of the model.

[0092] For example, assume the initial learning rate is η0, the maximum warm-up learning rate is η max , the number of epochs for warm-up is T warmup , and the current epoch is t. Then the mathematical expression of the cosine annealing strategy is as follows

[0093]

[0094] where η0 is the initial learning rate, η max is the maximum warm-up learning rate, T warmup is the number of epochs for warm-up, η min is the minimum learning rate, T total is the total number of epochs, and t is the current epoch number.

[0095] Therefore, the embodiments of the present disclosure construct a ResTransUnet model, which can perform noise reduction processing on one-dimensional vibration signals. Considering that the convolutional network is limited to local perception, a transformer layer is introduced in the convolutional module to enable the model to better focus on long-term dependent patterns or trends in the signal, rather than just local noise features. In addition, the middle layer of the model is connected through an improved residual structure, enabling the model to better capture multi-level features in the signal and improve the noise reduction effect.

[0096] S500. Train the denoised signal based on the ResTransUnet model and generate a benchmark deep learning model.

[0097] In this embodiment, the benchmark deep learning model refers to a general and universal deep learning classification model. The deep learning classification model is a WDCNN model. The WDCNN model is a deep learning classification model based on a one-dimensional convolutional neural network (1D CNN), mainly used to process one-dimensional signal data, such as time series signals, speech signals, electrocardiogram signals, etc. The structure of the WDCNN model usually includes multiple convolutional layers, pooling layers, fully connected layers and other modules. In practical applications, the parameters and structure of the model can be appropriately adjusted and optimized according to specific data sets and task requirements.

[0098] S600 performs fault diagnosis on the actually collected vibration signals based on a benchmark deep learning model.

[0099] In this embodiment, after classifying and learning the denoised signals, the WDCNN model is used to perform fault diagnosis on the actually collected vibration signals, which can effectively perform fault diagnosis on the actually collected vibration signals, help to detect and handle equipment faults in a timely manner, and improve the reliability and operation efficiency of the equipment.

[0100] In some other embodiments, in order to prove the superiority and robustness of the noise reduction effect of the method of the present invention. This embodiment will compare the proposed model with the following several noise reduction methods: Unet1DRes, CDAE, and the original U-Net. Experiments are carried out respectively under the noise interference conditions of 0 dB and -20 dB, and each model runs 5 times and takes the average value to avoid random errors. As shown in Table 1 below:

[0101] Noise level Model RMSE↓ PSNR↑ FSCS↑ R2↑ 0dB Proposed method 0.00742 43.57922 0.99996 0.99916 U-Net 0.01920 35.08427 0.99976 0.99517 CDAE 0.09787 21.43072 0.99013 0.80703 Resunet 0.01623 36.18061 0.99982 0.99473 -5dB Proposed method 0.01295 38.38228 0.99989 0.99768 U-Net 0.02031 34.55440 0.99976 0.99479 CDAE 0.09890 21.15923 0.99060 0.81070 Resunet 0.04697 27.01726 0.99844 0.95804 -10dB Proposed method 0.02018 34.50782 0.99974 0.99453 U-Net 0.03027 31.03041 0.99944 0.98741 CDAE 0.10614 20.76100 0.98869 0.77048 Resunet 0.06843 23.52146 0.99664 0.90383 -20dB Proposed method 0.04535 27.77225 0.99852 0.96556 U-Net 0.05207 26.54773 0.99745 0.95256 CDAE 0.12123 19.11862 0.98831 0.66663 Resunet 0.19729 14.19527 0.96508 0.34790

[0102] Table 1 shows the experimental results of all types of signals after noise reduction by different models. Among them, ↓ indicates that the smaller the index, the higher the quality of the denoised signal, and ↑ indicates that the larger the index, the higher the quality of the denoised signal. It can be seen that the proposed model shows the best results in each index type. At the 0 dB noise level, the RMSE of the proposed method is only 0.00742, far lower than that of U-Net (0.01920) and Resunet (0.01623). Under the strong noise interference of -20 dB, the performance of all models has decreased significantly, but the proposed method still shows strong denoising ability. Its RMSE is 0.04535. Although it has increased compared with the 0 dB noise level, it is still 0.07588, 0.00672, and 0.15194 lower than CDAE, Unet, and Resunet, respectively, and still shows obvious advantages. PSNR and R 2 are 27.77225 and 0.96556 respectively, which are better than other models.

[0103] Exemplarily, refer to Figure 4 , Figure 4 is the change diagram of the loss value and learning rate during the training process of the model provided by the embodiment of the present disclosure, specifically the training result of the proposed ResTransUnet model for denoising fault signals under the 0 dB noise interference condition. As can be seen from Figure 4 , in the early stage of training, the learning rate gradually increases from 0.0001 to 0.001. In this stage, the training loss and test loss values decrease rapidly. In the later stage, the learning rate is slowly decreased to make the model training tend to be stable, which helps to avoid the overfitting risk caused by too large a learning rate.

[0104] Exemplarily, refer to Figure 5 andFigure 6 , Figure 5 This is a comparison chart before and after noise reduction for a mild wear fault under -5 dB SNR noise provided by an embodiment of the present disclosure. Figure 6 This is a comparison chart before and after noise reduction for a mild wear fault under -5 dB SNR noise provided by an embodiment of the present disclosure. Where fm is the meshing frequency of the gear. It can be seen that after adding noise with an SNR of -5 dB, the amplitude of the original signal increases from the original ±20 m / s 2 to ±40 m / s. 2 . It can be seen from the spectrogram that other spurious frequencies appear beside the meshing frequency and its harmonics, but the amplitudes of the meshing frequency and its harmonics are still relatively high. After adding noise with an SNR of -20 dB, the amplitude of the acceleration signal increases from ±40 m / s 2 to ±200 m / s. 2 The spectrum of the signal becomes extremely cluttered. This is because the noise introduces a large number of high-frequency components, causing the peak of the characteristic frequency of the original signal to be submerged. This not only interferes with the main frequency components in the original signal but also generates many useless frequency components, making it difficult to identify the true fault characteristics in vibration analysis and fault diagnosis, especially for data-driven intelligent diagnosis models. After denoising with the proposed ResTransUnet model, the main frequency components in the signal spectrum are relatively clearly restored. The meshing frequency of the original signal and its harmonics (nfm), etc., appear again, the noise components are significantly suppressed, and the spikes in the spectrum become significant again. It can be seen that the denoising method can effectively remove the noise in the spectrum while retaining the key frequency characteristics of the original signal.

[0105] Exemplarily, refer to Figure 7 , Figure 7 This is a probability density distribution curve graph of healthy signals under different noise interferences provided by an embodiment of the present disclosure. Among them, (a) is the signal denoising method with adaptive time-frequency loss proposed in an embodiment of the present disclosure, (b) is Unet, (c) is Resunet, and (d) is CDAE. It can be seen that the probability density distribution curve of the signal after denoising by the proposed ResTransUnet model is closest to the original signal and shows good robustness under the condition of noise interference from 0 dB to -20 dB, and it can still recover the original signal characteristics even under -20 dB noise interference. Compared with the original Unet model, the noise distribution after denoising by the proposed model is closer. In addition, ResUnet can effectively remove noise when the SNR is higher than -10 dB, but when the noise interference is -20 dB, the model cannot effectively remove the noise interference. The CDAE model has the worst denoising effect. When the noise interference is -20 dB, the probability distribution curve of the denoised signal shows a significant difference from the original signal.

[0106] Exemplarily, refer to Figure 8 ,Figure 8 The diagnostic result diagrams after denoising of different models provided by the embodiments of the present disclosure are respectively experimented under the condition that the noise interference ranges from 0 dB to -20 dB. Among them, (a) has 100 samples for each category, and (b) has 200 samples for each category. It can be seen that the proposed ResTransUnet model can achieve the highest diagnostic accuracy under different noise level interferences, which is better than other models. Figure 7 (b) shows that the proposed model reaches 96.463% and 94.814% respectively under 5 dB and 0 dB noises, and the standard deviations are relatively low (0.995 and 1.023), indicating that its diagnostic ability and stability performance are excellent in a high-noise environment. As the noise interference intensity increases, its accuracy rate decreases, but it still reaches 89.666% at -20 dB.

[0107] Exemplarily, refer to Figure 9 , Figure 9 The confusion matrix diagram after diagnosing the signals denoised by different models provided by the embodiments of the present disclosure. Among them, (a) is the signal denoising method with adaptive time-frequency loss proposed by the embodiments of the present disclosure, (b) is U-net, (c) is Resunet, and (d) is CDAE. It can be seen that all fault categories can be effectively identified after denoising by the proposed ResTransUnet model, and the recognition accuracy rate is above 80%, but there is confusion in the signals of the healthy category (label 4). After denoising using the original U-Net model, the model is confused in identifying mild crack and mild wear faults. When using the signal denoised by the Resunet model, there is confusion between mild spalling and mild wear. For the signal denoised by the CDAE model, there is confusion in most categories, especially the recognition rate is 0 in the discrimination of mild spalling and mild wear faults.

[0108] Based on the above method, the embodiments of the present disclosure also provide a signal denoising system 2000 with adaptive time-frequency loss corresponding to the above method, as Figure 10 shown, Figure 10The block diagram of the signal denoising system with adaptive time-frequency loss provided by an embodiment of the present disclosure. The signal denoising system 2000 includes a sampling module 210, an analog module 220, a RestTransTnet module 230, a denoising module 240, a training module 250, and a fault diagnosis module 260. The sampling module 210 is used to obtain the vibration simulation signal of the rotating machinery and construct the original data set; the analog module 220 is used to fill Gaussian white noise with different noise levels based on the original data set to obtain a noise-polluted data set; the RestTransTnet module 230 is used to stack multiple convolutional layers based on the noise-polluted data set through residual connections and skip connections, and construct a ResTransUnet model; the denoising module 240 is used to perform denoising processing on the noise-polluted data set based on the ResTransUnet model and output a denoised signal. The training module 250 is used to train the denoised signal based on the ResTransUnet model and generate a benchmark deep learning model. The fault diagnosis module 260 is used to perform fault diagnosis on the actually collected vibration signal based on the benchmark deep learning model.

[0109] Based on the same inventive concept as the above-disclosed content, correspondingly, the present disclosure also provides an electronic device 3000. As Figure 11 shown, Figure 11 The block diagram of an electronic device provided by an embodiment of the present disclosure. The electronic device 3000 of the embodiment of the present disclosure includes at least one processor 310 and at least one memory 320 that are electrically connected. The memory 320 is electrically connected to the processor 310. Among them, the memory 320 stores instructions executable by at least one processor 310, and the instructions are executed by at least one processor 310 so that at least one processor 310 can execute the method as described above.

[0110] It should be noted that the electrical connections between the above-mentioned various units do not necessarily mean direct connections between the circuits. Indirect connection methods can be applied to the embodiments of the present disclosure as long as the purpose of the present disclosure is achieved.

[0111] Based on the same inventive concept, the present disclosure also provides a computer storage medium. A computer program is stored on the computer storage medium, and when the computer program is executed by a processor, it is the method as described above. The storage medium may include various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc.

[0112] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. An adaptive time-frequency loss signal denoising method, characterized in that, The method includes: Obtaining the vibration simulation signal of the rotating machinery and constructing the original data set; Filling Gaussian white noise with different noise levels based on the original data set to obtain a noise-polluted data set; Stacking multiple convolutional layers based on the noise-polluted data set through residual connections and skip connections, and constructing a ResTransUnet model; Performing noise reduction processing on the noise-polluted data set based on the ResTransUnet model and outputting a noise-reduced signal.

2. The method according to claim 1, wherein The stacking of multiple convolutional layers based on the noise-polluted data set through residual connections and skip connections, and constructing a ResTransUnet model includes: The ResTransUnet model includes an upsampling part, a downsampling part, and an intermediate layer; The upsampling and the downsampling are symmetrically designed, and the noise-polluted data without data is stacked through skip connections to form multiple convolutional layers; The intermediate layer is used to perform residual connection on the upsampling and the downsampling.

3. The method according to claim 2, wherein The downsampling includes a double convolutional module and a Transconv module. The double convolutional module is used to extract local features in the noise-polluted data set, and the Transconv module is used to upsample the feature map with low resolution in the noise-polluted data set to a feature map with high resolution.

4. The method according to claim 3, characterized in that The double convolutional module includes at least two sequentially stacked convolutional layers, batch normalization layers, and Relu activation layers. The formula is as follows: σ relu = max(0, x) where μ and σ 2 represent the mean and variance of the batch respectively, γ and β are parameters that can be learned by the model, and ∈ is a very small value.

5. The method according to claim 3, characterized in that, The downsampling also includes a Transformer module, which is used to dynamically assign different weights to each part of the noise signal in the noise-polluted data set and capture the global features in the noise-polluted data set.

6. The method according to claim 1, characterized in that, The performing noise reduction processing on the noise-polluted data set based on the ResTransUnet model and outputting a noise-reduced signal includes: Defining a loss function to initialize the ResTransUnet model; Adjusting the learning rate during the model training process based on the initialized ResTransUnet model by using the cosine annealing warm-up strategy; Performing noise reduction processing on the noise-polluted data set based on the adjusted ResTransUnet model and outputting a noise-reduced signal.

7. The method according to claim 1, characterized in that, The method further includes: Training the noise-reduced signal based on the ResTransUnet model and generating a benchmark deep learning model; Performing fault diagnosis on the actually collected vibration signal based on the benchmark deep learning model.

8. Signal denoising system with adaptive time-frequency loss, characterized in that, The system includes a module for executing the signal noise reduction method of the adaptive time-frequency loss as described in any one of claims 1-7.

9. An electronic device, characterized in that, The electronic device includes: A memory for storing a computer program; A processor for implementing the signal noise reduction method of the adaptive time-frequency loss as described in any one of claims 1 to 7 when executing the computer program.

10. A computer storage medium, characterized in that, A computer program is stored on the computer storage medium, and when the computer program is executed by the processor, the signal noise reduction method of the adaptive time-frequency loss as described in any one of claims 1 to 7 is implemented.

Images