A vibration signal denoising method and system based on a self-supervised convolutional neural network

Abstract

This invention relates to the field of mechanical equipment condition monitoring technology, and discloses a vibration signal denoising method and system based on a self-supervised convolutional neural network. The method involves preprocessing the original vibration signal during the operation of the mechanical equipment to train a constructed one-dimensional convolutional-deconvolutional neural network model. Instead of relying on manually labeled clean samples, it utilizes signal statistical characteristics such as kurtosis, envelope spectrum sparsity, and root mean square amplitude to construct a self-supervised loss function. This allows the model to automatically learn to distinguish between impact components and noise, guiding the one-dimensional convolutional-deconvolutional neural network model to automatically separate impact components from noise. The convolutional-deconvolutional structure enables temporal feature extraction and signal reconstruction, thereby achieving signal denoising while maintaining consistency with the original amplitude and effectively enhancing impact characteristics.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring technology, and discloses a vibration signal noise reduction method and system based on a self-supervised convolutional neural network. Background Technology

[0002] Mechanical equipment commonly generates vibration signals during operation, which contain a wealth of information about the mechanical structure's condition and faults. Especially in rotating machinery (such as rolling bearings, gearboxes, and rotor systems), vibration signals are widely used as a key source of information for fault diagnosis. When structural defects (such as spalling, cracks, or foreign object impacts) occur in bearings or rotor systems, periodic impact responses are typically generated in the vibration signals. These impact signals reflect the characteristics of localized defects, with their energy primarily concentrated in the high-frequency range, making them one of the most direct physical indicators of mechanical fault characteristics.

[0003] However, in actual industrial environments, these impact signals are often overwhelmed by strong noise due to environmental background, testing systems, and local modes. Affected by these interferences, the signal-to-noise ratio of the vibration signal decreases significantly, the impact waveform is weakened or blurred, and the peak values ​​of fault characteristic frequencies in the envelope spectrum are difficult to identify, leading to a reduction in the accuracy of traditional diagnostic methods. To extract impact features, researchers have proposed various noise reduction and feature enhancement techniques, mainly including:

[0004] 1) Time-domain filtering and wavelet denoising: Noise is suppressed by bandpass filtering or multi-scale wavelet decomposition, but it depends on manual parameter settings and has poor robustness to non-stationary signals; 2) Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD): They can separate some noise components, but they have problems such as mode aliasing and endpoint effects; 3) Sparse representation and independent component analysis (ICA): can enhance sparse impact features, but is computationally complex and sensitive to parameters; 4) Neural networks: They can achieve non-linear feature separation, but they must rely on known clean signals as training labels, which are often unavailable in practical engineering.

[0005] Therefore, current research methods still have significant shortcomings when facing the problem of extracting impact signals under high noise: they lack a label-free self-learning mechanism; they are prone to impact amplitude attenuation or waveform distortion; and they are difficult to balance noise reduction performance with the preservation of physical characteristics. Summary of the Invention

[0006] The purpose of this invention is to provide a vibration signal denoising method and system based on a self-supervised convolutional neural network. By using a convolution-deconvolution structure, temporal feature extraction and signal reconstruction are achieved. Under label-free conditions, adaptive denoising and feature enhancement can be achieved for impact mechanical fault signals, which maintains the consistency of the original amplitude and effectively enhances the impact features.

[0007] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows: A vibration signal denoising method based on a self-supervised convolutional neural network includes: The raw vibration signal during the operation of the mechanical equipment is collected, and the raw vibration signal is preprocessed to obtain the preprocessed vibration signal; the preprocessing includes one or more combinations of DC removal, normalization, or bandpass filtering. A one-dimensional convolutional-deconvolutional neural network model is constructed, comprising convolutional layers, nonlinear activation function layers, deconvolutional layers, deDC removal layers, and amplitude constraint layers. The convolutional layers and nonlinear activation function layers are used to extract temporal features from the input signal. The deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. The deDC removal layer is used to remove DC from the waveform structure signal output by the deconvolutional layer. The amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure and to apply amplitude constraints to the waveform structure based on the root mean square to obtain a constrained and denoised output signal. A joint loss function based on constraints of kurtosis of the denoised output signal, sparsity of the envelope spectrum, and root mean square amplitude is established. The preprocessed vibration signal is used as input and the corresponding constrained denoised output signal is used as output. The one-dimensional convolution-deconvolution neural network model is trained to obtain a one-dimensional convolution-deconvolution neural network model that makes the joint loss function converge. The preprocessed vibration signal to be analyzed is input into the trained one-dimensional convolutional-deconvolutional neural network model pair to obtain the corresponding denoised output signal.

[0008] Furthermore, the method for extracting temporal features from the input signal using the convolutional layer and the nonlinear activation function layer includes: The input signal is locally weighted using the sliding convolution kernel of the convolutional layer to extract temporal features; the mathematical expression is: ,in The output feature map of the convolutional layer is at the 1st... The amplitude of each sampling point The original vibration signal at the 1st The amplitude of each sampling point The kernel length is 1. The convolution kernel weights, This is the internal index number of the convolution kernel. For the first convolution window One input signal sample value, For the bias of the convolutional layer, It is a non-linear activation function; Using the ReLU activation function extract The instantaneous excitation characteristics form a time-domain characteristic signal containing local impact features. .

[0009] Furthermore, the deconvolution layer reconstructs the input signal after temporal feature extraction into a waveform structure that matches the input signal. ,in This is the waveform structure signal output by the deconvolution layer. For deconvolution weights, For the deconvolution layer bias, This represents the deconvolution operation. It is a non-linear activation function.

[0010] Furthermore, the deDC layer performs deDC processing on the waveform structure signal output by the deconvolution layer to obtain a deDC signal symmetrically distributed around the zero point. ,in This is the waveform structure signal output by the deconvolution layer. This represents the number of sampling points in the signal sequence.

[0011] Furthermore, the root mean square of all data values ​​in the waveform structure is calculated using the amplitude constraint layer, and amplitude constraints are applied to the waveform structure based on the root mean square to obtain the constrained noise-reduced output signal. ,in The root mean square of the original vibration signal. , The original vibration signal at the 1st The amplitude of each sampling point The number of sampling points in the signal sequence. To remove the root mean square of the DC signal, , To prevent constants with zero denominators, the range of values ​​is from 1e-6 to 1e-8.

[0012] Furthermore, the method for training the one-dimensional convolutional-deconvolutional neural network model to obtain a one-dimensional convolutional-deconvolutional neural network model that converges the joint loss function includes: Establish joint loss function ,in: , , These are the preset weighting coefficients. , The kurtosis of the constrained, noise-reduced output signal. , , , The sparsity ratio of the envelope spectrum of the constrained denoised output signal. , , , The envelope spectrum of the noise-reduced output signal after constraint; The neural network is trained using a self-supervised approach, and the network parameters of the one-dimensional convolutional-deconvolutional neural network model are iteratively updated using the Adam optimizer until the joint loss function converges.

[0013] To achieve the above-mentioned technical effects, the present invention also provides a vibration signal denoising system based on a self-supervised convolutional neural network, used to implement the vibration signal denoising method based on a self-supervised convolutional neural network, comprising: The data preprocessing module is used to collect the raw vibration signals during the operation of mechanical equipment, and to preprocess the raw vibration signals to obtain preprocessed vibration signals; the preprocessing includes one or more combinations of DC removal, normalization, or bandpass filtering operations; The network model construction module is used to construct a one-dimensional convolutional-deconvolutional neural network model containing convolutional layers, nonlinear activation function layers, deconvolutional layers, deDC removal layers, and amplitude constraint layers. The convolutional layers and nonlinear activation function layers are used to extract temporal features from the input signal. The deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. The deDC removal layer is used to remove DC from the waveform structure signal output by the deconvolutional layer. The amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure and apply amplitude constraints to the waveform structure based on the root mean square to obtain a constrained, denoised output signal. The model training module is used to establish a joint loss function based on the constraint of kurtosis of the denoised output signal, the constraint of sparsity of the envelope spectrum, and the constraint of root mean square amplitude. The preprocessed vibration signal is used as input and the corresponding constrained denoised output signal is used as output to train the one-dimensional convolution-deconvolution neural network model, so as to obtain a one-dimensional convolution-deconvolution neural network model that makes the joint loss function converge. The noise reduction output module is used to input the preprocessed vibration signal to be analyzed into the trained one-dimensional convolutional-deconvolutional neural network model pair to obtain the corresponding noise reduction output signal.

[0014] Furthermore, in the network model construction module, the constructed convolutional layer locally weights the input signal using a sliding convolution kernel to extract temporal features; the mathematical expression is: ,in The output feature map of the convolutional layer is at the 1st... The amplitude of each sampling point The original vibration signal at the 1st The amplitude of each sampling point The kernel length is 1. The convolution kernel weights, This is the internal index number of the convolution kernel. For the first convolution window One input signal sample value, For the bias of the convolutional layer, It is a non-linear activation function; The constructed nonlinear activation function layer is activated by the ReLU function. extract The instantaneous excitation characteristics form a time-domain characteristic signal containing local impact features. ; The constructed deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. ,in This is the waveform structure signal output by the deconvolution layer. For deconvolution weights, For the deconvolution layer bias, This represents the deconvolution operation. It is a non-linear activation function; The constructed deDC layer is used to perform deDC processing on the waveform structure signal output by the deconvolution layer, resulting in a deDC signal symmetrically distributed around the zero point. ,in This is the waveform structure signal output by the deconvolution layer. The number of sampling points in the signal sequence; The constructed amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure, and to apply amplitude constraints to the waveform structure based on the root mean square to obtain the constrained noise-reduced output signal. ,in The root mean square of the original vibration signal. , The original vibration signal at the 1st The amplitude of each sampling point The number of sampling points in the signal sequence. To remove the root mean square of the DC signal, , To prevent constants with zero denominators, the range of values ​​is from 1e-6 to 1e-8.

[0015] Furthermore, the model training module includes: Loss function building unit, used to establish the joint loss function ,in: , , These are the preset weighting coefficients. , The kurtosis of the constrained, noise-reduced output signal. , , , The sparsity ratio of the envelope spectrum of the constrained denoised output signal. , , , The envelope spectrum of the noise-reduced output signal after constraint; The network parameter update unit is used to train the neural network in a self-supervised manner, and iteratively update the network parameters of the one-dimensional convolutional-deconvolutional neural network model through the Adam optimizer until the joint loss function converges.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention trains a one-dimensional convolutional-deconvolutional neural network model by preprocessing the original vibration signal during the operation of mechanical equipment. It does not rely on manually labeled clean samples, but instead utilizes signal statistical characteristics such as kurtosis, envelope spectrum sparsity, and root mean square amplitude to construct a self-supervised loss function. This enables the model to automatically learn to distinguish between impact components and noise, guiding the one-dimensional convolutional-deconvolutional neural network model to automatically separate impact components and noise. Through the convolutional-deconvolutional structure, temporal feature extraction and signal reconstruction are achieved, thereby achieving signal noise reduction while maintaining consistency with the original amplitude and effectively enhancing impact characteristics. Attached Figure Description

[0017] Figure 1 The flowchart is shown in Example 1 or 2, illustrating the vibration signal noise reduction method based on a self-supervised convolutional neural network. Figure 2 This is a block diagram of the vibration signal noise reduction system based on a self-supervised convolutional neural network in Example 1; Figure 3 This is a schematic diagram of the dual-rotor structure in Example 2; Figure 4 This is a comparison of the time-domain waveforms of the vibration signal before and after noise reduction in Example 2; Figure 5 This is a comparison of the envelope spectra of the signals before and after noise reduction in Example 2; The module consists of: 1. Data preprocessing module; 2. Network model construction module; 3. Model training module; 301. Loss function construction unit; 302. Network parameter update unit; and 4. Noise reduction output module. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0019] Example 1 See Figure 1 and Figure 2 A vibration signal denoising method based on a self-supervised convolutional neural network includes: The raw vibration signal during the operation of the mechanical equipment is collected, and the raw vibration signal is preprocessed to obtain the preprocessed vibration signal; the preprocessing includes one or more combinations of DC removal, normalization, or bandpass filtering. A one-dimensional convolutional-deconvolutional neural network model is constructed, comprising convolutional layers, nonlinear activation function layers, deconvolutional layers, deDC removal layers, and amplitude constraint layers. The convolutional layers and nonlinear activation function layers are used to extract temporal features from the input signal. The deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. The deDC removal layer is used to remove DC from the waveform structure signal output by the deconvolutional layer. The amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure and to apply amplitude constraints to the waveform structure based on the root mean square to obtain a constrained and denoised output signal. A joint loss function based on constraints of kurtosis of the denoised output signal, sparsity of the envelope spectrum, and root mean square amplitude is established. The preprocessed vibration signal is used as input and the corresponding constrained denoised output signal is used as output. The one-dimensional convolution-deconvolution neural network model is trained to obtain a one-dimensional convolution-deconvolution neural network model that makes the joint loss function converge. The preprocessed vibration signal to be analyzed is input into the trained one-dimensional convolutional-deconvolutional neural network model pair to obtain the corresponding denoised output signal.

[0020] In this embodiment, the original vibration signal during the operation of the mechanical equipment is preprocessed to train the constructed one-dimensional convolutional-deconvolutional neural network model. Instead of relying on manually labeled clean samples, a self-supervised loss function is constructed using signal statistical characteristics such as kurtosis, envelope spectrum sparsity, and root mean square amplitude. This allows the model to automatically learn to distinguish between impact components and noise, guiding the one-dimensional convolutional-deconvolutional neural network model to automatically separate impact components and noise. Through the convolutional-deconvolutional structure, temporal feature extraction and signal reconstruction are achieved, thereby maintaining consistency with the original amplitude and effectively enhancing impact characteristics while achieving signal noise reduction.

[0021] Based on the same inventive concept, this embodiment also provides a vibration signal denoising system based on a self-supervised convolutional neural network, used to implement the vibration signal denoising method based on a self-supervised convolutional neural network, including: The data preprocessing module 1 is used to collect the raw vibration signals during the operation of mechanical equipment, and to preprocess the raw vibration signals to obtain preprocessed vibration signals; the preprocessing includes one or more combinations of DC removal, normalization, or bandpass filtering operations; Network model construction module 2 is used to construct a one-dimensional convolutional-deconvolutional neural network model containing convolutional layers, nonlinear activation function layers, deconvolutional layers, deDC removal layers, and amplitude constraint layers. The convolutional layers and nonlinear activation function layers are used to extract time-domain features from the input signal. The deconvolutional layer is used to reconstruct the input signal after time-domain feature extraction into a waveform structure that matches the input signal. The deDC removal layer is used to remove DC from the waveform structure signal output by the deconvolutional layer. The amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure and to apply amplitude constraints to the waveform structure based on the root mean square to obtain a constrained and denoised output signal. Model training module 3 is used to establish a joint loss function based on the constraint of kurtosis of the denoised output signal, the constraint of sparsity of the envelope spectrum and the constraint of root mean square amplitude. The preprocessed vibration signal is used as input and the corresponding constrained denoised output signal is used as output to train the one-dimensional convolution-deconvolution neural network model, so as to obtain a one-dimensional convolution-deconvolution neural network model that makes the joint loss function converge. The noise reduction output module 4 is used to input the preprocessed vibration signal to be analyzed into the trained one-dimensional convolutional-deconvolutional neural network model pair to obtain the corresponding noise reduction output signal.

[0022] The model training module 3 includes: Loss function construction unit 301 is used to build the joint loss function. ,in: , , These are the preset weighting coefficients. , The kurtosis of the constrained, noise-reduced output signal. , , , The sparsity ratio of the envelope spectrum of the constrained denoised output signal. , , , The envelope spectrum of the noise-reduced output signal after constraint; The network parameter update unit 302 is used to train the neural network in a self-supervised manner, and iteratively update the network parameters of the one-dimensional convolutional-deconvolutional neural network model through the Adam optimizer until the joint loss function converges.

[0023] Example 2 See Figure 1 , Figures 3 to 5 This embodiment utilizes a dual-rotor test specimen to conduct bearing fault simulation tests. The experimental results are used to provide a detailed explanation of the vibration signal noise reduction method based on a self-supervised convolutional neural network, as described in this invention. The specific process of this method is as follows: Step 1: Collect raw vibration signals during the operation of the mechanical equipment; In this embodiment, the test specimen is a dual-rotor structure, such as... Figure 3 There are a total of 5 bearings shown. Bearing #1, bearing #2 (with an artificially damaged inner ring), and bearing #5 are connected to the low-pressure rotor; bearing #3 is connected to the high-pressure rotor; and bearing #4 is an intermediate bearing. Bearing #2 has its inner ring artificially damaged, and the fault characteristic frequency is 9.42 times the low-pressure rotor's rotational frequency. Accelerometers are placed in the bearing housings to collect vibration signals, with a sampling frequency of... 640,000 points of steady-state signal were extracted as the original vibration signal.

[0024] Step 2: Preprocess the original vibration signal to obtain the preprocessed vibration signal; the preprocessing includes one or more combinations of DC removal, normalization, or bandpass filtering operations; In this embodiment, the original vibration signal is normalized and DC-free processed, and then... Each segment is divided into 156 segments, which are used as training samples.

[0025] Step 3: Construct a one-dimensional convolutional-deconvolutional neural network model comprising convolutional layers, nonlinear activation function layers, deconvolutional layers, deDC removal layers, and amplitude constraint layers. The convolutional layers and nonlinear activation function layers are used to extract temporal features from the input signal. The deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. The deDC removal layer is used to remove DC from the waveform structure signal output by the deconvolutional layer. The amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure and to apply amplitude constraints to the waveform structure based on the root mean square to obtain a constrained, denoised output signal. In this embodiment, the neural network structure design adopts a single-channel one-dimensional convolution-deconvolution structure, wherein: The constructed convolutional layer extracts temporal features by locally weighting the input signal using a sliding convolution kernel; the mathematical expression is: ,in The output feature map of the convolutional layer is at the 1st... The amplitude of each sampling point The original vibration signal at the 1st The amplitude of each sampling point The kernel length is 1. The convolution kernel weights, This is the internal index number of the convolution kernel. For the first convolution window One input signal sample value, For the bias of the convolutional layer, It is a non-linear activation function; The constructed nonlinear activation function layer is activated by the ReLU function. extract The instantaneous excitation characteristics form a time-domain characteristic signal containing local impact features. ; The constructed deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. ,in This is the waveform structure signal output by the deconvolution layer. For deconvolution weights, For the deconvolution layer bias, This represents the deconvolution operation. It is a non-linear activation function; The constructed deDC layer is used to perform deDC processing on the waveform structure signal output by the deconvolution layer, resulting in a deDC signal symmetrically distributed around the zero point. ,in This is the waveform structure signal output by the deconvolution layer. The number of sampling points in the signal sequence; The constructed amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure, and to apply amplitude constraints to the waveform structure based on the root mean square to obtain the constrained noise-reduced output signal. ,in The root mean square of the original vibration signal. , The original vibration signal at the 1st The amplitude of each sampling point The number of sampling points in the signal sequence. To remove the root mean square of the DC signal, , To prevent constants with zero denominators, this embodiment takes... =1e-7.

[0026] In this embodiment, the kernel size of the convolutional coding layer The number of channels is 16; the non-linear activation layer uses the ReLU function; the kernel size of the convolutional decoding layer is... Channel number 1; then connect to DC layer and amplitude constraint layer.

[0027] Step 4: Establish a joint loss function based on the constraints of kurtosis constraint, envelope spectrum sparsity constraint, and root mean square magnitude constraint of the denoised output signal. In this embodiment, a joint loss function is established. ,in: , , These are the preset weighting coefficients. , The kurtosis of the constrained, noise-reduced output signal. , , , The sparsity ratio of the envelope spectrum of the constrained denoised output signal. , , , The envelope spectrum of the noise-reduced output signal after constraint; In this example, the loss weight is set to... , , .

[0028] Step 5: Using the preprocessed vibration signal as input and the corresponding constrained denoising output signal as output, train the one-dimensional convolution-deconvolution neural network model to obtain a one-dimensional convolution-deconvolution neural network model that makes the joint loss function converge. In this embodiment, the Adam optimization algorithm is used, with a learning rate of... The system was trained for 20 rounds on 500 simulated samples. The signal output was obtained after self-supervised training.

[0029] Step 4: Input the preprocessed vibration signal to be analyzed into the trained one-dimensional convolutional-deconvolutional neural network model pair to obtain the corresponding noise-reduced output signal; In this embodiment, Figure 4 This is a comparison diagram of the time-domain waveforms before and after vibration signal noise reduction in this embodiment. Figure 4 The gray dashed curve represents the original noisy signal, and the black solid line represents the denoised signal. It can be seen that before denoising, noise dominates the signal, and the impact characteristics are not obvious. After denoising, the periodic impact components in the signal are significantly recovered, background noise is significantly suppressed, the amplitude of the impact waveform is stable, and the period is consistent with the original fault characteristic frequency period, indicating that this method can enhance the impact characteristics of the fault.

[0030] Figure 5The image shows a comparison of the envelope spectra of the signals before and after denoising. The gray dashed line corresponds to the envelope spectrum of the original noisy signal, and the black solid line represents the envelope spectrum of the denoised signal. It can be seen that after denoising, the spectral peaks at the inner loop fault characteristic frequency and its 1 / 2 and 3 / 2 harmonics are significantly enhanced, and the main peak / noise ratio is significantly improved. This indicates that the method of this invention enhances the identifiability of the fault impulse frequency in the frequency domain.

[0031] To further verify the denoising and feature enhancement effects of this invention, key indicators such as root mean square (RMS), kurtosis, and signal-to-noise ratio (SNR) were calculated for the signal before and after denoising. The results are as follows: Table 1 Comparison of Key Signal Indicators Before and After Noise Reduction

[0032] As can be seen from Table 1, the effective value of the signal after noise reduction remains almost unchanged (0.547→0.547), indicating that the energy preservation module is effective and the output amplitude is not compressed or amplified; the kurtosis value increases significantly from 2.89 to 19.02, indicating that the impact and peak of the output signal are greatly enhanced and the impact events are more prominent; the signal-to-noise ratio increases by about 4.72dB, indicating that the noise energy is effectively suppressed and the proportion of useful features in the signal is significantly increased.

[0033] In summary, this embodiment achieves self-supervised denoising under unlabeled conditions by constructing a self-supervised loss function and utilizing the signal's own statistical characteristics (kurtosis, energy consistency, and envelope spectrum sparsity) as training targets, thus enabling network adaptive learning without manually labeled samples. The designed convolution-deconvolution structure can separate periodic impulses from random noise in the time domain. Combined with an effective value amplitude preservation module, the overall energy of the output signal remains consistent with the input. The impulse waveform of the denoised signal is clear, and the amplitude is not compressed, improving the problem of "impulse weakening" in traditional filtering or empirical mode decomposition methods, effectively suppressing noise and preserving impulse characteristics. Furthermore, by constraining kurtosis and sparsity, the kurtosis of the output signal is improved, thereby enhancing the impulse characteristics and improving the accuracy of impulse detection and feature recognition. Figure 4 , Figure 5 Table 1 verifies the noise reduction and impact enhancement effects of the present invention from the perspectives of time-frequency domain and key signal indicators.

[0034] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A vibration signal denoising method based on a self-supervised convolutional neural network, characterized in that, include: The raw vibration signals during the operation of mechanical equipment are collected, and the raw vibration signals are preprocessed to obtain preprocessed vibration signals. The preprocessing includes one or more combinations of DC removal, normalization, or bandpass filtering operations; A one-dimensional convolutional-deconvolutional neural network model is constructed, comprising convolutional layers, nonlinear activation function layers, deconvolutional layers, deDC removal layers, and amplitude constraint layers. The convolutional layers and nonlinear activation function layers are used to extract temporal features from the input signal. The deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. The deDC removal layer is used to remove DC from the waveform structure signal output by the deconvolutional layer. The amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure and to apply amplitude constraints to the waveform structure based on the root mean square to obtain a constrained and denoised output signal. A joint loss function based on constraints of kurtosis of the denoised output signal, sparsity of the envelope spectrum, and root mean square amplitude is established. The preprocessed vibration signal is used as input and the corresponding constrained denoised output signal is used as output. The one-dimensional convolution-deconvolution neural network model is trained to obtain a one-dimensional convolution-deconvolution neural network model that makes the joint loss function converge. The preprocessed vibration signal to be analyzed is input into the trained one-dimensional convolutional-deconvolutional neural network model pair to obtain the corresponding denoised output signal.

2. The vibration signal denoising method based on a self-supervised convolutional neural network according to claim 1, characterized in that, The method for extracting temporal features from the input signal using the convolutional layer and nonlinear activation function layer includes: The input signal is locally weighted using the sliding convolution kernel of the convolutional layer to extract temporal features; the mathematical expression is as follows: ,in The output feature map of the convolutional layer is at the 1st... The amplitude of each sampling point The original vibration signal at the 1st The amplitude of each sampling point The kernel length is 1. The convolution kernel weights, This is the internal index number of the convolution kernel. For the first convolution window One input signal sample value, For the bias of the convolutional layer, It is a non-linear activation function; Using the ReLU activation function extract The instantaneous excitation characteristics form a time-domain characteristic signal containing local impact features. .

3. The vibration signal denoising method based on a self-supervised convolutional neural network according to claim 1, characterized in that, The deconvolutional layer reconstructs the input signal after temporal feature extraction into a waveform structure that matches the input signal. ,in This is the waveform structure signal output by the deconvolution layer. For deconvolution weights, For the deconvolution layer bias, This represents the deconvolution operation. It is a non-linear activation function.

4. The vibration signal denoising method based on a self-supervised convolutional neural network according to claim 1, characterized in that, The deDC layer performs deDC processing on the waveform structure signal output from the deconvolution layer, resulting in a deDC signal symmetrically distributed around the zero point. ,in This is the waveform structure signal output by the deconvolution layer. This represents the number of sampling points in the signal sequence.

5. The vibration signal denoising method based on a self-supervised convolutional neural network according to claim 1, characterized in that, The root mean square (RMS) of all data values ​​in the waveform structure is calculated using the amplitude constraint layer, and amplitude constraints are applied to the waveform structure based on the RMS to obtain the constrained, noise-reduced output signal. ,in The root mean square of the original vibration signal. , The original vibration signal at the 1st The amplitude of each sampling point The number of sampling points in the signal sequence. To remove the root mean square of the DC signal, , To prevent constants with zero denominators, the range of values ​​is from 1e-6 to 1e-8.

6. The vibration signal denoising method based on a self-supervised convolutional neural network according to claim 5, characterized in that, The method for training the one-dimensional convolutional-deconvolutional neural network model to obtain a one-dimensional convolutional-deconvolutional neural network model that makes the joint loss function converge includes: Establish joint loss function ,in: , , These are the preset weighting coefficients. , The kurtosis of the constrained, noise-reduced output signal. , , , The sparsity ratio of the envelope spectrum of the constrained denoised output signal. , , , The envelope spectrum of the noise-reduced output signal after constraint; The neural network is trained using a self-supervised approach, and the network parameters of the one-dimensional convolutional-deconvolutional neural network model are iteratively updated using the Adam optimizer until the joint loss function converges.

7. A vibration signal denoising system based on a self-supervised convolutional neural network, used to implement the vibration signal denoising method based on a self-supervised convolutional neural network as described in claim 1, characterized in that, include: The data preprocessing module is used to collect the raw vibration signals during the operation of mechanical equipment, preprocess the raw vibration signals, and obtain the preprocessed vibration signals. The preprocessing includes one or more combinations of DC removal, normalization, or bandpass filtering operations; The network model construction module is used to construct a one-dimensional convolutional-deconvolutional neural network model containing convolutional layers, nonlinear activation function layers, deconvolutional layers, deDC removal layers, and amplitude constraint layers. The convolutional layers and nonlinear activation function layers are used to extract temporal features from the input signal. The deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. The deDC removal layer is used to remove DC from the waveform structure signal output by the deconvolutional layer. The amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure and apply amplitude constraints to the waveform structure based on the root mean square to obtain a constrained, denoised output signal. The model training module is used to establish a joint loss function based on the constraint of kurtosis of the denoised output signal, the constraint of sparsity of the envelope spectrum, and the constraint of root mean square amplitude. The preprocessed vibration signal is used as input and the corresponding constrained denoised output signal is used as output to train the one-dimensional convolution-deconvolution neural network model, so as to obtain a one-dimensional convolution-deconvolution neural network model that makes the joint loss function converge. The noise reduction output module is used to input the preprocessed vibration signal to be analyzed into the trained one-dimensional convolutional-deconvolutional neural network model pair to obtain the corresponding noise reduction output signal.

8. The vibration signal noise reduction system based on a self-supervised convolutional neural network according to claim 7, characterized in that, In the network model construction module, the constructed convolutional layers locally weight the input signal using sliding convolution kernels to extract temporal features; the mathematical expression is: ,in The output feature map of the convolutional layer is at the 1st... The amplitude of each sampling point The original vibration signal at the 1st The amplitude of each sampling point The kernel length is 1. The convolution kernel weights, This is the internal index number of the convolution kernel. For the first convolution window One input signal sample value, For the bias of the convolutional layer, It is a non-linear activation function; The constructed nonlinear activation function layer is activated by the ReLU function. extract The instantaneous excitation characteristics form a time-domain characteristic signal containing local impact features. ; The constructed deconvolutional layer is used to reconstruct the input signal after temporal feature extraction into a waveform structure that matches the input signal. ,in This is the waveform structure signal output by the deconvolution layer. For deconvolution weights, For the deconvolution layer bias, This represents the deconvolution operation. It is a non-linear activation function; The constructed deDC layer is used to perform deDC processing on the waveform structure signal output by the deconvolution layer, resulting in a deDC signal symmetrically distributed around the zero point. ,in This is the waveform structure signal output by the deconvolution layer. The number of sampling points in the signal sequence; The constructed amplitude constraint layer is used to calculate the root mean square of all data values ​​in the waveform structure, and to apply amplitude constraints to the waveform structure based on the root mean square to obtain the constrained noise-reduced output signal. ,in The root mean square of the original vibration signal. , The original vibration signal at the 1st The amplitude of each sampling point The number of sampling points in the signal sequence. To remove the root mean square of the DC signal, , To prevent constants with zero denominators, the range of values ​​is from 1e-6 to 1e-8.

9. The vibration signal noise reduction system based on a self-supervised convolutional neural network according to claim 8, characterized in that, The model training module includes: Establish joint loss function ,in: , , These are the preset weighting coefficients. , The kurtosis of the constrained, noise-reduced output signal. , , , The sparsity ratio of the envelope spectrum of the constrained denoised output signal. , , , The envelope spectrum of the noise-reduced output signal after constraint; The network parameter update unit is used to train the neural network in a self-supervised manner, and iteratively update the network parameters of the one-dimensional convolutional-deconvolutional neural network model through the Adam optimizer until the joint loss function converges.

Images