Mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution

Abstract

The application provides a mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution, and steps are as follows: collecting a mechanical fault vibration signal according to a preset sampling frequency; constructing a simulation data set by using a gear fault simulation signal; filling the simulation data set with different degrees of Gaussian white noise to obtain a noise-containing data sample set; constructing a multi-scale dynamic weighting module on an encoding part of a deep convolutional auto-encoding network, constructing a multi-dimensional fusion module on the deep convolutional auto-encoding network, and constructing a residual learning residual module through a jump connection line to construct an MDW-MDJRCAE model; and training the MDW-MDJRCAE model by using the noise-containing data sample set to obtain a trained gear denoising network model. The application fully utilizes one-dimensional convolutional neural network to construct noise multi-scale dynamic weighting features and two-dimensional convolutional neural network strong feature learning capability, and better denoising effect can be obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent noise reduction of rotating machinery fault signals, and in particular to a mechanical signal noise reduction method based on multi-scale dynamic weighted multi-dimensional residual convolution. BACKGROUND

[0002] As the cornerstone of modern industrial systems, unexpected shutdown of rotating machinery will cause huge economic losses, and even casualties, so monitoring the health status of rotating machinery is crucial to ensure its safe and stable operation. Vibration signals carry important information about the health status of machinery, however, rotating machinery is often in harsh environments with high temperature and high impact, and the measured vibration signals are severely disturbed by noise, and the key features are contaminated by noise. In order to effectively recover the original clean signal under noise interference and extract relevant and sensitive features, it is necessary to perform noise reduction processing on the vibration signal.

[0003] In the era of big data, deep learning models are widely used in the field of mechanical vibration signal noise reduction due to their strong representation and learning ability. In the signal denoising process, deep convolutional auto-encoding networks are often used to filter out noise components and directly reconstruct the original clean signal, and good denoising results are achieved. However, with the increase in network depth, the deep convolutional auto-encoding network is prone to gradient vanishing problems. And directly reconstructing the original high-dimensional signal leads to excessive number of auto-encoding network weights, making training more difficult, and in many cases, the original clean signal cannot be obtained.

[0004] For vibration signals, mechanical rotation and reciprocating frequency can make the signal complex and exhibit multi-scale characteristics, and features extracted by a single scale convolution kernel can easily cause information loss. At the same time, multiple scale convolution kernels have the same weight when learning signal features, i.e., different scale features have different contributions to the denoising result. In particular, when signal features extracted by multiple scale convolution kernels are fused, 1D CNN (Convolutional Neural Network) does not have the advantage of directly fusing multi-scale adjacent features as 2D CNN. The conversion method based on time domain, frequency domain and time-frequency domain converts one-dimensional signal to two-dimensional mapping, and combines 2D CNN to complete the vibration signal denoising process, which achieves good denoising results, but the denoising results are too dependent on the pre-processing of the signal, and the manual selection of features cannot be separated from expert prior knowledge.

[0005] In the Chinese patent, there is an application number CN 202010976852.X, and the name is a rolling bearing fault diagnosis method based on CEEMD and kurtosis weighted average threshold denoising, which belongs to the technical field of bearing fault diagnosis. The above-mentioned invention first collects the bearing vibration acceleration signal, decomposes the collected signal by CEEMD to obtain several IMF components; the wavelet threshold denoising is carried out on each IMF component, and the denoised component signal is used for reconstruction to obtain the filtered reconstruction signal, and the bearing fault characteristic frequency is identified by frequency spectrum analysis. The weighted average threshold denoising is a denoising method of linear combination of soft and hard thresholds according to the weight, which can effectively combine the advantages of soft and hard threshold denoising. The above-mentioned invention determines the hard and soft threshold weights in the weighted average threshold denoising method by calculating the size of each component kurtosis value, so as to obtain better denoising effect.

[0006] In addition, there is an application number CN202010272255.9 in the Chinese patent, and the name is a gear signal denoising method based on VMD and maximum overlap discrete wavelet packet transform. The combination of VMD and MODWPT is used to denoise the gear fault signal. The experimental results of the measured data show that the method of the above-mentioned invention can effectively filter out the noise interference and iteration error components in the signal, and realize the accurate extraction of the signal characteristics.

[0007] The above two patents use data processing methods based on vibration signal analysis theory to complete mechanical signal denoising. In the signal denoising process, the expert prior knowledge is excessively relied on, and the intelligent denoising algorithm based on deep learning theory is not involved. SUMMARY

[0008] In view of the technical problems that the existing rotating machinery signal denoising method excessively relies on the early signal processing and the denoising effect is poor in adaptability, the present application proposes a mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution. The one-dimensional and two-dimensional convolution neural networks are unified into an overall network framework to complete the reconstruction of the noise signal, the residual network is combined to subtract the noise, the clean signal is obtained, the signal denoising is realized, the influence of the environmental noise on the signal characteristics is effectively reduced, and the fault characteristics contained in the vibration signal are more prominent.

[0009] In order to achieve the above-mentioned purpose, the technical scheme of the present application is as follows: a mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution, the steps are as follows:

[0010] Step 1. Complete the collection of the mechanical fault vibration signal according to the pre-set sampling frequency;

[0011] Step 2. Construct a simulation data set by using a gear fault simulation signal;

[0012] Step 3. Fill the simulation data set with different degrees of Gaussian white noise to obtain a noisy data sample set;

[0013] Step 4. Construct a multi-scale dynamic weighting module on the encoding part of the deep convolutional auto-encoding network, construct a multi-dimensional fusion module on the deep convolutional auto-encoding network, and construct a residual learning module through a jump connection line to build a MDW-MDJRCAE model;

[0014] Step 5. Train the MDW-MDJRCAE model using the noisy data sample set to obtain a trained gear denoising network model;

[0015] Step 6. Input the actual mechanical fault vibration signal obtained in step 1 into the trained gear denoising network model to obtain a denoised signal.

[0016] Preferably, the method for constructing a simulation data set in step 2 is: obtaining a simulation signal of gear failure according to a gear failure signal simulation expression, randomly cutting the simulation signal according to a certain length to expand the simulation data sample, and standardizing the expanded simulation data sample to obtain a simulation data set.

[0017] The expression of the multi-component amplitude modulation and frequency modulation gear failure simulation signal is:

[0018] z(t) = (1 + a(t)) y(t)

[0019] a(t) = sin(2πf r t)

[0020] y(t) = cos(2πf m t + 2cos(2πf r t)) + 0.8cos(12πf m t + 2cos(2πf r t)) + 0.6cos(30πf m t + 2cos(2πf r t))

[0021] Where f m is the meshing frequency, and f m = zf r , z is the number of gear teeth, f r is the rotation frequency; z(t) is the gear failure signal, 1 + a(t) is the amplitude modulation signal, and y(t) is the carrier signal.

[0022] Preferably, the implementation method of step 3 is: filling the simulation data set with different degrees of Gaussian white noise to form noise-containing signals with signal-to-noise ratios (SNRs) of 8 dB, 10 dB, 12 dB, 14 dB, 16 dB, 18 dB, 20 dB, 22 dB, 24 dB, and 26 dB, respectively, to obtain a noise-containing data sample set; and the signal-to-noise ratio

[0023]

[0024] where s(i) represents the ith value of the original clean signal, denotes the ith value of the signal obtained by denoising, and N represents the length of the signal.

[0025] Preferably, the MDW-MDJRCAE model comprises a 1D convolution layer, a 1D activation layer, a dynamic weighting layer, a Concatenate layer, a 1D-2D conversion layer, a 2D convolution layer, a 2D activation layer, a 2D pooling layer, a 2D-1D conversion layer, and a 1D upsampling layer.

[0026] The multi-scale dynamic weighting module comprises four groups of parallel single-scale dynamic weighting modules and a Concatenate layer, the input signal passes through the four groups of parallel single-scale dynamic weighting modules to obtain four groups of single-scale dynamic weighting feature vectors, and the Concatenate layer is connected in parallel with the four groups of single-scale dynamic weighting feature vectors to obtain a 1D multi-scale dynamic weighting feature vector.

[0027] The multi-dimensional fusion module comprises a 1D-2D conversion layer, a 2D CNN module, a 2D-1D conversion layer, and a 1D CNN module connected in sequence, the 1D multi-scale dynamic weighting feature vector is stacked to generate a two-dimensional matrix through the 1D-2D conversion layer, the two-dimensional matrix is taken as the input of the 2D CNN module to obtain a noise feature two-dimensional representation; the noise feature two-dimensional representation is subjected to feature dimension reduction through the 2D-1D conversion layer to obtain a noise feature one-dimensional representation, which is taken as the input of the 1D CNN module to reconstruct a 1D noise feature; and the 1D noise feature passes through the 1D convolution layer to obtain a reconstructed noise signal.

[0028] Preferably, each single-scale dynamic weighting module comprises a 1D convolution layer I, a 1D activation layer I, and a dynamic weighting layer I connected in sequence, the 1D convolution layer I performs local convolution calculation on the input signal to extract high-dimensional signal features, the 1D activation layer I adopts a ReLU linear rectifier function to filter effective features, and the dynamic weighting layer generates dynamic weights during network training, the dynamic weights are multiplied with each single-scale feature vector output by the 1D activation layer I to perform dynamic weighting, and a single-scale dynamic weighting feature vector is obtained.

[0029] Preferably, the convolution kernel sizes of the 1D convolution layers I of the four groups of parallel single-scale dynamic weighting modules are 3x1, 9x1, 16x1 and 64x1 respectively.

[0030] The 2D CNN module is sequentially connected by a 2D convolution layer I, a 2D activation layer I, a 2D pooling layer I, a 2D convolution layer II, a 2D activation layer II, a 2D pooling layer II, a 2D convolution layer III, a 2D activation layer III and a 2D pooling layer III; the 2D convolution layers II and III directly fuse the feature vectors; the 2D activation layers I-III filter the effective features by using a ReLU linear rectification function; and the 2D pooling layers I-III extract the maximum values of adjacent local features by using maximum pooling.

[0031] The 1D CNN module is sequentially connected by a 1D upsampling layer I, a 1D convolution layer II, a 1D activation layer II, a 1D upsampling layer II, a 1D convolution layer III, a 1D activation layer III, a 1D upsampling layer III, a 1D convolution layer IV and a 1D activation layer IV; the 1D upsampling layers I-III fill and map the input features to a corresponding region of the output, perform a simple copy operation on the numerical values, and expand the length of the input data to twice the original length, so that the network completes the dimensionality increasing operation on the signal features in the decoding process; and the 1D noise features pass through the 1D convolution layer V to obtain a reconstructed noise signal.

[0032] Preferably, the residual module includes a residual module I and a residual module II; the residual module I is constructed by subtracting the output of the 1D convolution layer V of the deep convolution auto-encoding network from the input signal through a jump connection line; and the residual module II is constructed by adding the output of the Concatenate layer of the encoding part of the deep convolution auto-encoding network and the output of the 1D activation layer IV of the decoding part through a jump connection line.

[0033] Preferably, a mean square error function is selected as a loss function, the MDW-MDJRCAE model is trained, a gradient descent method is used for weight updating, each layer gradient is calculated by using an error back propagation method according to an error chain propagation rule, a learning rate is combined to update the parameters of the network in a timely manner during iteration, and a loss function and a signal-to-noise ratio mean value formula are combined to save the trained model; when the number of training times meets a model training termination condition, the training is stopped and the trained rolling bearing denoising network model is saved.

[0034] Preferably, the method for training the MDW-MDJRCAE model is as follows:

[0035] a.Forward calculation: for the input signal y=x+v, wherein x is a one-dimensional original clean vibration signal, v is a noise signal, and y is a noise-polluted vibration signal; input the input signal y into a deep convolutional auto-encoding network, and pass the multi-scale dynamic weighting module of the encoding part of the deep convolutional auto-encoding network through four parallel 1D convolutional layers to respectively perform multi-scale feature extraction on the vibration signal and perform dynamic weighting as follows: F ij = X ij W i ;

[0036] wherein X ij represents the result of the i-th convolution kernel and the j-th activation operation, W i represents the weight component corresponding to the i-th convolution kernel, F ij is the feature value of the i-th row and the j-th column of the weighted feature vector F; the Concatenate layer connects the dynamic weighted feature vectors obtained by the four dynamic weighting layers to form a 1D multi-scale dynamic weighted feature vector;

[0037] A multi-dimensional fusion module is constructed on the deep convolutional auto-encoding network, that is, a 1D-2D conversion layer is used to stack the 1D multi-scale dynamic weighted feature vector to generate a two-dimensional matrix as follows:

[0038]

[0039] wherein represents a 1D-2D stacking operation, F ij , G ij , M ij , N ij respectively correspond to the feature values of the i-th row and the j-th column of the dynamic weighted feature vectors extracted by 3x1, 9x1, 16x1 and 64x1 convolution kernels, n represents the length of each row of feature vectors, and K represents the constructed two-dimensional matrix;

[0040] The two-dimensional matrix K is taken as the input of the 2D CNN module to directly fuse the 1D multi-scale dynamic weighted feature vector, and a noise feature two-dimensional representation is obtained; a 2D-1D conversion layer is used to perform feature dimension reduction on the noise feature two-dimensional representation:

[0041]

[0042] wherein represents a 2D-1D operation, P is a noise feature one-dimensional representation, and P ab is the feature value of the a-th row and the b-th column of the one-dimensional representation P; the noise feature one-dimensional representation P is taken as the input of the 1D CNN module, and a series of 1D up-sampling layers, 1D convolution and activation operations are combined to reconstruct the 1D noise feature;

[0043] Residual module I connects the input signal with the output v' of the deep convolutional auto-encoding network through a skip connection line, i.e., subtracts the input signal y from the reconstructed noise signal v'; residual module II connects the output signal of the multi-scale dynamic weighting module and the output of the multi-dimensional fusion module through a skip connection 2 line, i.e., subtracts the output of the Concatenate layer from the output signal y of the 1D activation layer IV 11 The addition obtains the reconstructed 1D noise feature y1, and the 1D noise feature y1 passes through the 1D convolutional layer V to obtain the reconstructed noise signal v'; according to the noise superposition principle, the denoised clean signal y' = x + v - v' is obtained.

[0044] b. Weight update: the gradient descent method is adopted, the gradient of each layer is calculated according to the chain propagation rule of error and the error back propagation method; in the 1D-2D connection part, the gradient △K passed by the two-dimensional convolutional layer is:

[0045]

[0046] wherein, △F ij , △G ij , △M ij , △N ij correspond to the gradients of the feature values F ij , G ij , M ij , N ij .

[0047] In the 2D-1D connection part, the gradient △P passed by the one-dimensional convolutional layer is:

[0048]

[0049] wherein, △P ab is the gradient of the feature value P ab , a = 1-m, b = 1-n, m is the number of rows of the matrix, and n is the number of columns of the matrix.

[0050] The weight update of the parameter θ corresponding to the two-dimensional matrix K is: θ p+1 = θ p -η▽K;

[0051] The weight update of the parameter β corresponding to the one-dimensional matrix P is: β p+1 = β p -η▽P;

[0052] wherein, p represents the number of parameter updates, η represents the learning rate, θ p represents the parameter value of the two-dimensional matrix K at the pth iteration, and β p represents the parameter value of the one-dimensional matrix P at the pth iteration.

[0053] Preferably, the mean square error function is

[0054] wherein y' k represents the reconstructed clean signal, x k represents the original clean signal, and N is the data length;

[0055] Mean of SNR

[0056] SNR(d) is the signal-to-noise ratio SNR of the dth sample, and D is the number of samples;

[0057] Through time domain statistical analysis, the denoising effects of the trained gear denoising network model are comprehensively evaluated by using the denoising indexes of the signal-to-noise ratio estimation SNR1, the mean square error MSE and the cross-correlation index Xcorr, and the effectiveness of the MDW-MDJRCAE model for denoising of the vibration signal is verified in combination with the envelope spectrum of the original signal and the denoised signal.

[0058] Compared with the prior art, the beneficial effects of the present application are as follows: for the vibration signals collected in different states, the noise feature signal is obtained by using the deep convolution auto-encoding network, the original signal is subtracted from the noise feature signal by combining the residual network, and the clean signal is obtained; the mean square error function is taken as the loss function, the filter parameter of the multi-scale dynamic weighted multi-dimensional joint residual convolution auto-encoding network model is optimized by using the error back propagation algorithm, and the final signal denoising model is obtained in combination with the loss function and the MSNR formula. The present application fully utilizes the one-dimensional convolution neural network to construct the noise multi-scale dynamic weighted feature and the powerful feature learning ability of the two-dimensional convolution neural network, unifies the one-dimensional and two-dimensional convolution neural networks into an overall network framework, completes the reconstruction of the noise signal, and develops the multi-scale dynamic weighted multi-dimensional joint residual convolution auto-encoding network model for the signal denoising of the rotating machinery. The present application can obtain better denoising effect and good data adaptability, and can be used for the signal denoising of various rotating machinery faults. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0060] Figure 1 is the flowchart of the present application.

[0061] Figure 2 is the schematic diagram of the multi-scale dynamic weighting of the present application.

[0062] Figure 3 is the 1D-2D conversion diagram of the present application.

[0063] Figure 4 Figure 2D-1D conversion diagram of the present application.

[0064] Figure 5 Figure of noise learning of the present application.

[0065] Figure 6 Figure of gear fault signal before and after denoising of the present application, wherein (a) is the original vibration signal, and (b) is the denoised signal.

[0066] Figure 7 Envelope spectrum diagram of gear fault signal before and after denoising of the present application, wherein (a) is the envelope spectrum before and after denoising, and (b) is the envelope spectrum of the denoised signal. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0068] As shown in Figure 1 , a mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional joint residual convolution has the following steps:

[0069] Step 1. Data acquisition.

[0070] According to the pre-set sampling frequency, the acquisition of the mechanical fault vibration signal is completed.

[0071] Step 2. Constructing a simulation data set using a gear fault simulation signal.

[0072] According to the gear fault signal simulation expression, the gear fault simulation signal is obtained, the data is randomly cut according to a certain length, a simulation data set is constructed, and the standardization processing of the simulation data set is completed, wherein x is the simulation signal, mean(x) is the mean of the simulation signal x, sd(x) is the variance of the simulation signal x, is the standardized simulation signal.

[0073] The expression of the multi-component amplitude modulation and frequency modulation gear fault simulation signal is:

[0074] z(t)=(1+a(t))y(t)

[0075] a(t)=sin(2πfr t)

[0076] y(t) = cos(2pft + 2 cos(2pft) m t + 2 cos(2pft) r t + 2 cos(2pft) m t + 2 cos(2pft) r t + 2 cos(2pft) m t + 2 cos(2pft) r t)

[0077] where f m is the meshing frequency, and f m = zf r , z is the number of gear teeth, f r is the rotation frequency; z(t) is the gear fault signal, 1 + a(t) is the amplitude modulation signal, and y(t) is the carrier signal.

[0078] Step 3. Fill the simulation data set with different degrees of Gaussian white noise to obtain a noisy data sample set.

[0079] According to the SNR formula, fill the standardized simulation data set with different degrees of Gaussian white noise to construct a noisy data sample set. According to the set proportion, divide it into a training sample set and a test sample set, and use the training sample set to train the network denoising model.

[0080] According to the SNR formula, that is:

[0081]

[0082] In the formula, SNR is the signal-to-noise ratio, s(i) represents the i th value of the original clean signal, represents the i th value of the signal obtained by denoising, and N represents the length of the signal.

[0083] Add different degrees of Gaussian white noise to the simulation data set to form noisy signals with signal-to-noise ratios SNR of 8dB, 10dB, 12dB, 14dB, 16dB, 18dB, 20dB, 22dB, 24dB, and 26dB. In this way, the signal-to-noise ratio SNR is selected to enable the network to learn different degrees of noise components during the training process.

[0084] Step 4. Construct a multi-scale dynamic weighting module on the encoding part of the deep convolutional auto-encoding network, construct a multi-dimensional fusion module on the decoding part of the deep convolutional auto-encoding network, and construct a residual learning residual module through a jump connection line to build an MDW-MDJRCAE model.

[0085] The MDW-MDJRCAE model is composed of a 1D convolution layer, a 1D activation layer, a dynamic weighting layer, a concatenate layer, a 1D-2D conversion layer, a 2D convolution layer, a 2D activation layer, a 2D pooling layer, a 2D-1D conversion layer and a 1D upsampling layer.

[0086] First, a deep convolutional auto-encoding network is constructed to encode and decode signal features. A multi-scale dynamic weighting module is constructed in the encoding part of the deep convolutional auto-encoding network, that is, composed of four groups of parallel single-scale dynamic weighting modules and a concatenate layer, and each group of single-scale dynamic weighting modules is sequentially connected by a 1D convolution layer I, a 1D activation layer I and a dynamic weighting layer I. The 1D convolution layer performs local convolution calculation on the input signal to extract high-dimensional signal features, the 1D activation layer uses the ReLU linear rectifier function to filter effective features, and the dynamic weighting layer generates a self-defined parameter, that is, a dynamic weight, during network training. The dynamic weight is multiplied with each single-scale feature vector output by the 1D activation layer I to perform dynamic weighting and obtain a single-scale dynamic weighting feature vector. Appropriate convolution kernel sizes are set for each group of single-scale dynamic weighting modules. Without loss of generality, the sizes of the four groups of parallel 1D convolution layers are selected as 3x1, 9x1, 16x1 and 64x1, respectively, to extract features of different scales from the input signal. The signal passes through the four groups of parallel single-scale dynamic weighting modules to obtain four groups of single-scale dynamic weighting feature vectors, and then connects the concatenate layer to parallel the four groups of single-scale dynamic weighting feature vectors to obtain a 1D multi-scale dynamic weighting feature vector.

[0087] A multi-dimensional fusion module is constructed on a deep convolutional auto-encoding network, including a 1D-2D conversion layer, a 2D CNN module, a 2D-1D conversion layer and a 1D CNN module. The encoding part of the deep convolutional auto-encoding includes a multi-scale dynamic weighting module, a 1D-2D conversion layer and a 2D CNN module. The decoding part of the deep convolutional auto-encoding includes a 1D CNN module and a 1D convolution layer V. The 2D-1D conversion layer is connected between the encoding part and the decoding part. A 1D multi-scale dynamic weighting feature vector is stacked by the 1D-2D conversion layer to generate a two-dimensional matrix, which is used as the input of the 2D CNN (two-dimensional convolutional neural network) module to obtain a noise feature two-dimensional representation. The 2D CNN module is connected by a 2D convolution layer I, a 2D activation layer I, a 2D pooling layer I, a 2D convolution layer II, a 2D activation layer II, a 2D pooling layer II, a 2D convolution layer III, a 2D activation layer III and a 2D pooling layer III in sequence. The 2D convolution layer directly fuses the 1D multi-scale dynamic weighting feature vector, the 2D activation layer uses the ReLU linear rectification function to filter effective features, and the 2D-CNN pooling layer selects the maximum value of adjacent local features. The noise feature two-dimensional representation is reduced in dimension by the 2D-1D conversion layer to obtain a noise feature one-dimensional representation, which is used as the input of the 1D CNN module to reconstruct the 1D noise feature. The 1D CNN module is connected by a 1D upsampling layer I, a 1D convolution layer II, a 1D activation layer II, a 1D upsampling layer II, a 1D convolution layer III, a 1D activation layer III, a 1D upsampling layer III, a 1D convolution layer IV and a 1D activation layer IV in sequence. The 1D upsampling layer maps the input feature to a corresponding area of the output, performs a simple copy operation on the numerical value, and expands the input data length to twice the original length, so that the network completes the dimensionality reduction operation on the signal feature in the decoding process. The 1D noise feature is processed by the 1D convolution layer V to obtain a reconstructed noise signal.

[0088] The input signal and the output of the deep convolutional auto-encoding network (the output of the 1D convolution layer V) are connected by the skip connection I line, so that the input signal is subtracted from the output of the deep convolutional auto-encoding network, thereby constructing a residual module I (the addition in the residual is changed to subtraction). In the network training process, the input signal is subtracted from the noise characteristics learned by the deep convolutional auto-encoding network to obtain a clean signal. The output of the Concatenate layer of the encoding part and the output of the 1D activation layer IV of the decoding part are added by the skip connection II line to construct a residual module II, which solves the gradient vanishing problem in deep network training and improves the network training efficiency.

[0089] The original one-dimensional vibration signal is input into the MDW-MDJRCAE model, the noise feature signal is reconstructed through the deep convolution auto-encoding network, the multi-scale dynamic weighting module completes the multi-scale feature extraction and dynamic weighting of the noise signal, the multi-dimensional fusion module completes the direct fusion of the multi-scale dynamic weighted features of the noise and the reconstruction of the 1D noise features. The original signal is subtracted from the noise signal learned by the deep convolution auto-encoding network through the jump connection 1 line, to obtain a clean signal, complete signal denoising, and initialize the connection weights and bias parameters of each layer of the network, and set the initial learning rate to 0.001.

[0090] Step 5. The MDW-MDJRCAE model is trained using the constructed noise-containing data sample set to obtain a trained gear denoising network model.

[0091] The mean square error function is selected as the loss function, the MDW-MDJRCAE model is trained, the weight update adopts the gradient descent method, the error back propagation method is used to calculate the gradient of each layer according to the error chain propagation rule, and the learning rate is combined to update the parameters of the network in the iteration process. The loss function and the mean signal-to-noise ratio (MSNR) formula are combined to save the trained model. When the number of training times meets the model training termination condition, the training is stopped and the trained rolling bearing denoising network model is saved.

[0092] The mean square error function is

[0093] Where y' represents the reconstructed clean signal, x represents the original clean signal, and N is the data length. k k

[0094] The MSNR formula is

[0095] SNR(d) is the SNR of the dth sample. D is the number of samples.

[0096] The steps of training the MDW-MDJRCAE model are:

[0097] a. Forward calculation: for the input signal y=x+v, where x=(a1, a2, …, a N ) is the one-dimensional original clean vibration signal, i.e. the normalized simulation signal in step 2, v is the noise signal, and y is the vibration signal contaminated by noise.

[0098] The input signal y is input into the deep convolution auto-encoding network for encoding and decoding.

[0099] As Figure 2 ​​As shown, a multi-scale dynamic weighting module is constructed in the coding part of the deep convolutional auto-encoding network, that is, four parallel 1D convolutional layers are used to extract multi-scale features of the vibration signal (note: 3x1, 9x1, 16x1, and 64x1 convolutional kernels are selected), an activation layer and a dynamic weighting layer are connected after each 1D convolutional layer, and the dynamic weighting layer performs dynamic weighting, and the dynamic weighting formula is:

[0100] F ij = X ij W i

[0101] wherein X ij represents the result of the i-th convolutional kernel and the j-th activation operation, W i represents the weight component corresponding to the i-th convolutional kernel, and F ij is the feature value of the i-th row and j-th column of the weighted feature vector F.

[0102] A Concatenate layer connects the dynamic weighting feature vectors obtained by the four dynamic weighting layers to form a 1D multi-scale dynamic weighting feature vector.

[0103] A multi-dimensional fusion module is constructed on the deep convolutional auto-encoding network, which includes a 1D-2D conversion layer, a 2D CNN module, a 2D-1D conversion layer, and a 1D CNN module. That is, a 1D multi-scale dynamic weighting feature vector is stacked to generate a two-dimensional matrix by using a 1D-2D conversion layer, and the formula is:

[0104]

[0105] wherein represents a 1D-2D stacking operation, F ij , G ij , M ij , N ij correspond to the i-th row and j-th column feature values of the dynamic weighting feature vectors extracted by 3x1, 9x1, 16x1, and 64x1 convolutional kernels, respectively, n represents the length of each row of feature vectors, and K represents the constructed two-dimensional matrix. The two-dimensional matrix K is used as the input of the 2D CNN module to directly fuse the 1D multi-scale dynamic weighting feature vectors and obtain a noise feature two-dimensional representation.

[0106] A 2D-1D conversion layer is used to perform feature dimension reduction on the noise feature two-dimensional representation, and the formula is:

[0107]

[0108] wherein represents a 2D-1D operation, and P is a noise feature one-dimensional representation, and P abGiven the eigenvalue of the a-th row and b-th column of the corresponding matrix P, the one-dimensional representation of the noise feature P is used as the input of the 1D CNN module. By combining a series of 1D upsampling layers, 1D convolutions, activations, and other operations, the 1D noise feature is reconstructed.

[0109] Finally, using residual networks, such as Figure 5 As shown, residual module I and residual module II are constructed on the deep convolutional autoencoder network. Residual module I is constructed using jump connection line 1 (note: subtraction). Jump connection line 1 connects the input signal to the output v′ of the deep convolutional autoencoder network (note: the output v′ of the 1D convolutional layer V), that is, subtracting the input signal y from the reconstructed noise signal v′. Residual module II is constructed inside residual module I. Residual module II consists of a multi-dimensional fusion module and jump connection line 2. Jump connection line 2 connects the output signal of the multi-scale dynamic weighting module to the output of the multi-dimensional fusion module, that is, the output of the concatenate layer and the output y of the 1D activation layer IV. 11 The sums are used to obtain the reconstructed 1D noise feature y1. The 1D noise feature y1 is then passed through a 1D convolutional layer 5 to obtain the reconstructed noise signal v′ (Note: i.e., the noise reconstructed by residual module I). According to the principle of noise superposition, the denoised clean signal y′=x+vv′ is obtained.

[0110] b. Weight Update. During the network training of the MDW-MDJRCAE model, weight updates are performed using gradient descent. Based on the chain rule of error propagation, the gradient of each layer is calculated using backpropagation. In the 1D-2D connection part, the gradient ∇K passed from the 2D convolutional layer is:

[0111]

[0112] Among them, ▽F ij 、▽G ij 、▽M ij 、▽N ij Corresponding to F respectively ij G ij M ij N ij The gradient.

[0113] In the 2D-1D connection part, the gradient ▽P passed from the one-dimensional convolutional layer is:

[0114]

[0115] Among them, ▽P ab Corresponding to P ab The gradient.

[0116] The update of the parameter θ weights corresponding to the two-dimensional matrix K can be expressed as:

[0117] θp+1 = theta p - eta * deltar

[0118] The parameter beta weight update corresponding to the one-dimensional matrix P can be expressed as:

[0119] beta p+1 = beta p - eta * deltar

[0120] Wherein, p represents the parameter update times, eta represents the learning rate, theta p represents the parameter value of the matrix K at the pth iteration, beta p The matrix P represents the parameter value at the pth iteration.

[0121] The application establishes a training method of model parameters of a multi-scale dynamic weighted multi-dimensional joint residual convolutional auto-encoding network, deduces the mechanism of model training error propagation from a two-dimensional convolutional neural network to a one-dimensional convolutional neural network, and from a one-dimensional convolutional neural network to a two-dimensional convolutional neural network, realizes seamless connection of the two-dimensional convolutional neural network and the one-dimensional convolutional neural network, and updates the parameters of each layer of the entire model in the gradient back propagation process in combination with the learning rate.

[0122] Step 6. The actual mechanical fault vibration signal obtained in step 1 is input into the trained gear denoising network model to obtain the denoised signal.

[0123] The actual gear fault signal of the test bench is input into the trained model for signal denoising, time domain statistical analysis, signal-to-noise ratio estimation SNR1, mean square error MSE, cross-correlation index Xcorr and other denoising indexes are used to comprehensively evaluate the network denoising effect, and the original signal and denoised signal envelope spectrum are combined to verify the effectiveness of the MDW-MDJRCAE network model for vibration signal denoising.

[0124] Wherein, the formulas of SNR1, MSE and Xcorr are as follows:

[0125]

[0126]

[0127]

[0128] Wherein, s'(i) is the i th value of the original vibration signal, s''(i) is the i th value of the denoised vibration signal. s' is the original vibration signal, s'' is the denoised vibration signal, mu s′ , mu s″ are the mean values of the original vibration signal s' and the denoised vibration signal s'', sigma s′ , sigma s″The variance of the original vibration signal s' and the denoised vibration signal s", respectively.

[0129] The denoising effect is evaluated by SNR1, MSE and Xcorr. For the denoising model, SNR1 generally evaluates the ratio of useful components and noise components in the signal. MSE reflects the difference between vibration signals s'(i) and s"(i), and Xcorr measures the similarity of s' and s" at any two different times. The larger the SNR1, the smaller the MSE, the higher the Xcorr, the better the denoising effect.

[0130] Specific examples:

[0131] A mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional joint residual convolution, the fault signal denoising process is as shown in Figure 1 .

[0132] The experimental data in this experiment is the gear fault data set collected by the QPZZ-II rotating machinery vibration analysis and fault diagnosis test platform. The experimental platform mainly consists of a driving motor, a shaft assembly, a gear box, a brake, a base and an electrical cabinet. The gear damage fault is made by using the wire cut electrical discharge process, and the single tooth pitting fault is selected in the experiment, with a fault size of 0.4mm in depth, 1mm in width and 6 points in fault. The motor speed is 1470r / min, and the gear ratio of the large gear and the small gear is 75 and 55 respectively. According to the speed ratio, the large gear speed is 1078r / min. After calculation, the theoretical rotation frequency of the small gear is f r1 = 1470 / 60 = 24.5Hz, and the theoretical rotation frequency of the large gear is f r2 = 1078 / 60 = 17.9Hz. The following steps are used for processing:

[0133] Step 1. Data acquisition;

[0134] According to the pre-set sampling frequency, i.e. the sampling frequency is 12800Hz, the sampling time is 10s, and the data length is 128000, the vibration signal acquisition of the single tooth pitting fault of the gear is completed. According to the sampling theorem, the sampling frequency is set, and the data within 10s contains rich fault information.

[0135] Step 2. Simulation data set construction

[0136] According to the simulation expression of the gear fault signal, the gear fault simulation signal is obtained, and the data enhancement method is used, i.e. the gear fault simulation signal is randomly cut with a sliding window of 1024 length, and the simulation data sample expansion is carried out, so as to construct 660 simulation data samples, and the standardization processing of simulation data samples is completed.

[0137] Step 3. Construction of noisy data sample set.

[0138] According to the SNR formula, the normalized simulation data samples are added to different degrees of Gaussian white noise to form noise-containing signals with signal-to-noise ratios (SNRs) of 8 dB, 10 dB, 12 dB, 14 dB, 16 dB, 18 dB, 20 dB, 22 dB, 24 dB, and 26 dB, respectively, and a total of 6600 noise-containing sample sets. According to the set proportion of 7:3, the training sample set and the test sample set are divided to obtain 4620 noise-containing training sample sets.

[0139] Step 4. Constructing the MDW-MDJRCAE model.

[0140] The MDW-MDJRCAE model is composed of a 1D convolution layer, a 1D activation layer, a dynamic weighting layer, a concatenate layer, a 1D-2D conversion layer, a 2D convolution layer, a 2D activation layer, a 2D pooling layer, a 2D-1D conversion layer, and a 1D upsampling layer. The structure parameters of the MDW-MDJRCAE model are as follows: the network structure has a total of 34 layers, the noise signal is first subjected to 4 groups of single-scale dynamic weighting modules in parallel, each group of single-scale dynamic weighting module is sequentially connected by a 1D convolution layer I, a 1D activation layer I, and a dynamic weighting layer I to obtain 4 groups of single-scale dynamic weighting feature vectors. Further, the 4 groups of single-scale dynamic weighting feature vectors are connected in parallel through a concatenate layer to obtain a 1D multi-scale dynamic weighting feature vector. Then, a 1D-2D conversion layer, a 2D convolution layer I, a 2D activation layer I, a 2D pooling layer I, a 2D convolution layer II, a 2D activation layer II, a 2D pooling layer II, a 2D convolution layer III, a 2D activation layer III, a 2D pooling layer III, a 2D-1D conversion layer, a 1D upsampling layer I, a 1D convolution layer II, a 1D activation layer II, a 1D upsampling layer II, a 1D convolution layer III, a 1D activation layer III, a 1D upsampling layer III, a 1D convolution layer IV, and a 1D activation layer IV are sequentially connected, and finally a 1D convolution layer V is added to complete the reconstruction of the noise signal feature. The loss function is the mean square error function. The noise feature signal is reconstructed by the deep convolution auto-encoding network, the original signal is subtracted from the noise feature signal by using the residual network to obtain the clean signal, and the specific process is as follows:

[0141] For the MDW-MDJRCAE model, y = x + v, where x = (a1, a2, …, a N ) is a one-dimensional original clean vibration signal, i.e., the normalized simulation signal in step 2, v is the noise signal, and y is the noise-polluted vibration signal. The input signal y is input into the deep convolution auto-encoding network for encoding and decoding.

[0142] As Figure 2As shown in the deep convolutional auto-encoding network coding part, a multi-scale dynamic weighting module is constructed, that is, four parallel 1D convolutional layers 1 are used to extract multi-scale features of the vibration signal (note: 3×1, 9×1, 16×1, 64×1 convolution kernels are selected), and an activation layer and a dynamic weighting layer are connected after each 1D convolutional layer. The dynamic weighting layer performs dynamic weighting, and the dynamic weighting formula is:

[0143] F ij =X ij W i

[0144] Wherein, X ij represents the result of the i-th convolution kernel and the j-th activation operation, W i represents the weight component corresponding to the i-th convolution kernel, and F ij is the feature value of the i-th row and j-th column of the weighted feature vector F.

[0145] A concatenate layer is used to connect the dynamic weighting feature vectors obtained by the four dynamic weighting layers to form a 1D multi-scale dynamic weighting feature vector.

[0146] A multi-dimensional fusion module is constructed on the deep convolutional auto-encoding network, that is, a 1D-2D conversion layer is used to stack the 1D multi-scale dynamic weighting feature vector to generate a two-dimensional matrix, as shown in Figure 3 The formula is:

[0147]

[0148] Wherein, represents 1D-2D stacking operation, F ij , G ij , M ij , N ij correspond to the i-th row and j-th column feature values of the dynamic weighting feature vectors extracted by the 3×1, 9×1, 16×1, 64×1 convolution kernels, n represents the length of each row of feature vector, and K represents the constructed two-dimensional matrix. The two-dimensional matrix K is used as the input of the 2D CNN module, and the 1D multi-scale dynamic weighting feature vector is directly fused to obtain a noise feature two-dimensional representation.

[0149] As shown in Figure 4 , a 2D-1D layer is used to reduce the dimension of the noise feature two-dimensional representation, and the formula is:

[0150]

[0151] Wherein, represents 2D-1D operation, P is a noise feature one-dimensional representation, and P abIt is the eigenvalue of the a-th row and b-th column of the corresponding matrix P. The one-dimensional representation of the noise feature P is used as the input of the 1D CNN module. Combined with a series of 1D upsampling layers, 1D convolution, activation and other operations, the 1D noise feature is reconstructed.

[0152] Finally, using residual networks, such as Figure 5 As shown, residual modules ① and ② are constructed on the deep convolutional autoencoder network. Residual module I is constructed using jump connection 1 (note: subtraction). Jump connection 1 connects the input signal to the output v′ of the deep convolutional autoencoder network (note: the output v′ of the 1D convolutional layer V), i.e., subtracting the input signal y from the reconstructed noise signal v′. Residual module II is constructed inside residual module I. Residual module ② consists of a multi-dimensional fusion module and jump connection 2. Jump connection 2 connects the output signal of the multi-scale dynamic weighting module to the output of the multi-dimensional fusion module, i.e., the output of the concatenate layer and the output y of the 1D activation layer IV. 11 The sums are used to obtain the reconstructed 1D noise feature y1. The 1D noise feature y1 is then passed through a 1D convolutional layer V to obtain the reconstructed noise signal v′ (i.e., the noise reconstructed by the residual module I). According to the principle of noise superposition, y′=x+vv′ is used to obtain the clean signal y′ after noise reduction.

[0153] Step 5: Initialize the connection weights and bias parameters of each layer of the network, set the initial learning rate to 0.001, train the network model using the backpropagation algorithm, combine the loss function and MSNR formula, save the trained model, and stop training and save the trained gear denoising network model when the number of training iterations meets the model training termination condition.

[0154] During network training, weight updates are performed using gradient descent. Based on the chain rule of error propagation, backpropagation is used to calculate the gradient of each layer. In the 1D-2D connection, the error ∂K propagated from the 2D convolutional layer is:

[0155]

[0156] Among them ▽F ij 、▽G ij 、▽M ij 、▽N ij Corresponding to F respectively ij G ij M ij N ij The gradient.

[0157] In the 2D-1D connection part, the error ∽P passed from the one-dimensional convolutional layer is:

[0158]

[0159] Among them, ▽Pab corresponding P ab gradient

[0160] The parameter theta weight update corresponding to the two-dimensional matrix K can be expressed as:

[0161] theta p+1 = theta p - eta * gradient K

[0162] The parameter beta weight update corresponding to the one-dimensional matrix P can be expressed as:

[0163] beta p+1 = beta p - eta * gradient P

[0164] Wherein, p represents the parameter update times, eta represents the learning rate, theta p represents the parameter value of the matrix K at the pth iteration, beta p represents the parameter value of the matrix P at the pth iteration.

[0165] Step 6. Signal denoising.

[0166] The actual single-tooth pitting fault signal collected in step 1 is input into the trained model for signal denoising, and the denoising effect of the network is comprehensively evaluated through time domain statistical analysis, SNR1, MSE, Xcorr and the like, as shown in Figure 6 It can be directly observed from the time domain waveform diagram that the model effectively eliminates noise while well preserving the impact part, making the fault characteristics more obvious, and in Figure 6 (b), the quantitative index analysis can obtain SNR1=1.930, MSE=0.641, Xcorr=0.859. Further, envelope spectrum analysis is performed on the original signal and the denoised signal, as shown in Figure 7 , the blue line is the envelope spectrum of the original signal, and the red line is the envelope spectrum of the denoised signal, and it can be obtained that the noise component is well eliminated. From Figure 7 (b) the first 200Hz part of the denoised signal envelope spectrum, the pinion rotation frequency f r1 , the gear rotation frequency f r2 and its multiple frequencies can be clearly observed.

[0167] The present application uses a deep convolutional auto-encoding network to reconstruct a noise feature signal from vibration signals collected in different states, and combines a residual network to obtain a clean signal by subtracting the noise feature signal from the original signal; takes the mean square error function as the loss function, and uses the error back propagation algorithm to optimize the filter parameters of the multi-scale dynamic weighted multi-dimensional joint residual convolutional auto-encoding network model, and combines the loss function and the MSNR formula to obtain the final signal denoising model, so that the model has better denoising effect and good data adaptability, and can be used for various rotating machinery fault signal denoising.

[0168] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution, characterized in that, The steps are as follows: Step 1. Complete the collection of mechanical fault vibration signals according to the preset sampling frequency; Step 2. Construct a simulation data set using a gear fault simulation signal; Step 3. Fill the simulation data set with different degrees of Gaussian white noise to obtain a noisy data sample set; Step 4. Construct a multi-scale dynamic weighting module on the encoding part of the deep convolutional auto-encoding network, construct a multi-dimensional fusion module on the deep convolutional auto-encoding network, and construct a residual learning MDW-MDJRCAE model through a residual module constructed by a jump connection line; The MDW-MDJRCAE model comprises a 1D convolutional layer, a 1D activation layer, a dynamic weighting layer, a concatenate layer, a 1D-2D conversion layer, a 2D convolutional layer, a 2D activation layer, a 2D pooling layer, a 2D-1D conversion layer and a 1D upsampling layer; The multi-scale dynamic weighting module comprises four groups of parallel single-scale dynamic weighting modules and a concatenate layer, and the input signal is obtained through the four groups of parallel single-scale dynamic weighting modules to obtain four groups of single-scale dynamic weighting feature vectors, and the concatenate layer is connected with the four groups of single-scale dynamic weighting feature vectors to obtain a 1D multi-scale dynamic weighting feature vector; The multi-dimensional fusion module comprises a 1D-2D conversion layer, a 2D CNN module, a 2D-1D conversion layer and a 1D CNN module connected in sequence, the 1D multi-scale dynamic weighting feature vector is stacked to generate a two-dimensional matrix through the 1D-2D conversion layer, the two-dimensional matrix is taken as the input of the 2D CNN module to obtain a two-dimensional representation of noise features; the noise feature two-dimensional representation is reduced in dimension through the 2D-1D conversion layer to obtain a noise feature one-dimensional representation, which is taken as the input of the 1D CNN module to reconstruct the 1D noise feature; The 1D noise feature is subjected to a 1D convolutional layer to obtain a reconstructed noise signal; Step 5. Train the MDW-MDJRCAE model using the noisy data sample set to obtain a trained gear denoising network model; Step 6. Input the actual mechanical fault vibration signal obtained in step 1 into the trained gear denoising network model to obtain a denoised signal.

2. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution of claim 1, wherein, The method for constructing the simulation data set in step 2 is: obtaining a simulation signal of gear fault according to a gear fault signal simulation expression, randomly cutting the simulation signal according to a certain length to expand the simulation data sample, and standardizing the expanded simulation data sample to obtain a simulation data set; The expression of the multi-component amplitude modulation and frequency modulation gear fault simulation signal is: z(t) = (1 + a(t)) y(t) a(t) = sin(2πf r t) y(t) = cos(2πf m t + 2cos(2πf r t)) + 0.8cos(12πf m t + 2cos(2πf r t)) + 0.6 cos(30πf m t + 2 cos(2πf r t)) where f m is the meshing frequency, and f m = zf r , z is the number of gear teeth, f r is the rotational frequency; z(t) is the gear fault signal, 1+a(t) is the amplitude modulation signal, and y(t) is the carrier signal.

3. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution of claim 2, wherein, The implementation method of step 3 is: filling the simulation data set with different degrees of Gaussian white noise to form noisy signals with signal-to-noise ratios (SNRs) of 8dB, 10dB, 12dB, 14dB, 16dB, 18dB, 20dB, 22dB, 24dB and 26dB, and obtaining a noisy data sample set; the signal-to-noise ratio where s(i) represents the i-th value of the original clean signal, where s(i) represents the i-th value of the original clean signal, 4. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution of claim 3, wherein, Each single scale dynamic weighting module comprises a 1D convolution layer I, a 1D activation layer I and a dynamic weighting layer I connected in sequence, the 1D convolution layer I performs local convolution calculation on the input signal to extract high-dimensional signal features, the 1D activation layer I adopts a ReLU linear rectifier function to filter effective features, and the dynamic weighting layer generates dynamic weights during network training, the dynamic weights are multiplied with each single scale feature vector output by the 1D activation layer I, dynamic weighting is performed, and a single scale dynamic weighting feature vector is obtained.

5. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution of claim 4, wherein, The convolution kernel sizes of the 1D convolution layers I of the four parallel single scale dynamic weighting modules are 3x1, 9x1, 16x1 and 64x1 respectively; The 2D CNN module comprises a 2D convolution layer I, a 2D activation layer I, a 2D pooling layer I, a 2D convolution layer II, a 2D activation layer II, a 2D pooling layer II, a 2D convolution layer III, a 2D activation layer III and a 2D pooling layer III connected in sequence, the 2D convolution layers II-III perform direct feature fusion on the feature vectors, the 2D activation layers I-III adopt a ReLU linear rectifier function to filter effective features, and the 2D pooling layers 1-III select the maximum value of adjacent local features by maximum pooling; The 1D CNN module comprises a 1D up-sampling layer I, a 1D convolution layer II, a 1D activation layer II, a 1D up-sampling layer II, a 1D convolution layer III, a 1D activation layer III, a 1D up-sampling layer III, a 1D convolution layer IV and a 1D activation layer IV connected in sequence; the 1D up-sampling layers I-III fill and map the input features to a corresponding region of the output, perform simple copy operations on the numerical values, and expand the input data length to twice the original length, so that the network completes the dimensionality raising operation on the signal features in the decoding process; the 1D noise features pass through the 1D convolution layer V to obtain a reconstructed noise signal.

6. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution of claim 5, wherein, The residual module comprises a residual module I and a residual module II, the input signal and the output of the 1D convolution layer V of the deep convolution auto-encoding network are connected through a jump connection line, the input signal is subtracted from the output of the deep convolution auto-encoding network to construct the residual module I; the output of the Concatenate layer of the encoding part of the deep convolution auto-encoding network and the output of the 1D activation layer IV of the decoding part are added through a jump connection line to construct the residual module II.

7. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution according to any one of claims 4-6, characterized in that, The mean square error function is selected as the loss function, the MDW-MDJRCAE model is trained, the weight update adopts the gradient descent method, the error back propagation method is used to calculate the gradient of each layer according to the error chain propagation rule, and the learning rate is combined to update the parameters of the network in time during the iteration process; the loss function and the signal-to-noise ratio mean formula are combined to save the trained model, when the training times meet the model training termination condition, the training is stopped and the trained rolling bearing denoising network model is saved.

8. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution of claim 7, wherein, The method for training the MDW-MDJRCAE model is: a.Forward calculation: for the input signal y=x+v, wherein x is a one-dimensional original clean vibration signal, v is a noise signal, and y is a noise-polluted vibration signal; input the input signal y into the deep convolutional auto-encoding network, and pass the vibration signal through four parallel 1D convolutional layers in the multi-scale dynamic weighting module of the encoding part of the deep convolutional auto-encoding network to extract multi-scale features and perform dynamic weighting as: F ij = X ij W i ; wherein X ij represents the result of the i-th convolution kernel, the j-th activation operation, W i represents the weight component corresponding to the i-th convolution kernel, F ij is the feature value of the i-th row, the j-th column of the weighted feature vector F; the Concatenate layer connects the dynamic weighted feature vectors obtained by the four dynamic weighting layers to form a 1D multi-scale dynamic weighted feature vector; A multi-dimensional fusion module is constructed on the deep convolution auto-encoding network, that is, a 1D-2D conversion layer is used to stack the 1D multi-scale dynamic weighting feature vectors to generate a two-dimensional matrix as follows: wherein, represents a 1D-2D superposition operation, F ij , G ij , M ij , N ij respectively correspond to the feature values of the i-th row and j-th column of the dynamic weighted feature vector extracted by the 3x1, 9x1, 16x1, and 64x1 convolution kernels, n represents the length of each row of the feature vector, and K represents the constructed two-dimensional matrix. The two-dimensional matrix K is taken as an input of the 2D CNN module to directly fuse the 1D multi-scale dynamic weighted feature vectors, and obtain a noise feature two-dimensional representation; the noise feature two-dimensional representation is subjected to feature dimension reduction by using a 2D-1D conversion layer: wherein, represents a 2D-1D operation, P is a one-dimensional representation of the noise feature, P ab is the value of the feature in the a-th row and b-th column of the one-dimensional representation P; taking the one-dimensional representation P of the noise feature as the input of the 1D CNN module, combining a series of 1D up-sampling layers, 1D convolution, and activation operations, and reconstructing the 1D noise feature; The residual module I connects the input signal with the output v' of the deep convolutional auto-encoding network through a skip connection line, that is, subtracts the input signal y from the reconstructed noise signal v'; the residual module II connects the output signal of the multi-scale dynamic weighting module and the output of the multi-dimensional fusion module through a skip connection 2 line, that is, subtracts the output of the Concatenate layer from the output signal y of the 1D activation layer IV 11 The addition obtains the reconstructed 1D noise feature y1, and the 1D noise feature y1 obtains the reconstructed noise signal v' through the 1D convolution layer V; according to the noise superposition principle, the noise-reduced clean signal y' = x + v-v' is obtained. b. Weight update: the gradient descent method is adopted, the gradient of each layer is calculated according to the chain propagation rule of error and the error back propagation method; in the 1D-2D connection part, the gradient △K passed by the two-dimensional convolution layer is: wherein corresponding to the feature values F ij , G ij , M ij , N ij , respectively; In the 2D-1D connection part, the gradient passed from the one-dimensional convolutional layer is: wherein is the gradient of the eigenvalue P ab , a = 1 - m, b = 1 - n, m is the number of rows of the matrix, n is the number of columns of the matrix; The weight of the parameter θ corresponding to the two-dimensional matrix K is updated as: The parameter β weight corresponding to the one-dimensional matrix P is updated as: where p represents the number of parameter updates, η represents the learning rate, θ p represents the parameter value of the two-dimensional matrix K at the pth iteration, β p represents the parameter value of the one-dimensional matrix P at the pth iteration.

9. The mechanical signal denoising method based on multi-scale dynamic weighted multi-dimensional residual convolution of claim 8, wherein, The mean square difference error function is where y' = y - x k represents the reconstructed clean signal, x k represents the original clean signal, N is the data length; Signal-to-noise ratio mean SNR(d) is the signal-to-noise ratio SNR of the dth sample, and D is the number of samples; Through time domain statistical analysis, the denoising indexes of the signal-to-noise ratio estimation SNR1, the mean square error MSE and the cross-correlation index Xcorr are used to comprehensively evaluate the denoising effect of the trained gear denoising network model, and the effectiveness of the MDW-MDJRCAE model in denoising the vibration signal is verified by combining the original signal and the envelope spectrum of the denoised signal.

Images