A rolling bearing fault diagnosis method based on parallel convolutional neural networks

Abstract

This disclosure reveals a method for fault diagnosis of rolling bearings based on parallel convolutional neural networks, including: S100: acquiring the original vibration signal X of the rolling bearing under test. io (t); S200: For the original vibration signal X io (t) is preprocessed to obtain the preprocessed vibration signal X. i (t); S300: For the preprocessed vibration signal X i (t) Perform the first overlap sampling to obtain the first sample signal x. m (t); S400: Perform a second overlap sampling on the preprocessed vibration signal to obtain the second sample signal y. m (t); S500: Based on the first sample signal x m (t) and the second sample signal y m (t) Obtain the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test respectively; S600: Construct a fusion dataset based on the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test and input it into the trained parallel convolutional neural network to realize the fault identification of the rolling bearing under test.

Description

Technical Field

[0001] This disclosure pertains to the field of mechanical fault diagnosis and signal processing, specifically relating to a method for diagnosing rolling bearing faults based on a parallel convolutional neural network. Background Technology

[0002] With the deepening of Industry 4.0, electromechanical equipment is developing towards larger scale, automation, intelligence, and precision. Malfunctions and shutdowns of electromechanical equipment during operation can significantly impact production and cause economic losses. Therefore, ensuring the safe and stable operation of electromechanical equipment is of paramount importance, requiring the establishment of accurate and reliable health monitoring systems. Rolling bearings, as one of the key components of electromechanical equipment, are widely used in shipbuilding, automotive, wind power generation, aerospace, and other fields. Rolling bearings are easily damaged components in rotating machinery, frequently experiencing plastic deformation, fatigue fracture, pitting, and spalling under complex and harsh working conditions and alternating loads. Rolling bearing fault diagnosis is a crucial link in ensuring the normal operation of electromechanical equipment; therefore, in production practice, there is an urgent need for a high-precision, high-stability rolling bearing fault diagnosis method.

[0003] Currently, fault diagnosis methods for rolling bearings mainly fall into two categories: traditional methods and deep learning methods. Traditional methods primarily rely on time-frequency domain analysis of signals, such as fast spectral kurtosis plotting, minimum entropy deconvolution, variational mode decomposition, and wavelet decomposition. These methods process the bearing vibration signals detected by sensors, obtaining and plotting the frequency domain image of the bearing vibration signal. Based on the geometric parameters and operating conditions of the rolling bearing, the fault frequency is calculated, and the corresponding fundamental frequency and its harmonics are identified in the frequency domain image of the vibration signal. This is then compared with the calculated theoretical fault frequency to determine whether a fault exists at a certain location in the rolling bearing. These traditional fault diagnosis methods are mostly based on a combination of signal feature extraction and classifiers. Such algorithms are complex and require a high level of mathematical and physical knowledge, and users need extensive expert experience to judge the algorithm's processing results. These algorithms generally have applicable operating conditions, and the diagnostic results can vary significantly depending on the operating conditions and algorithm parameter settings, resulting in poor generalization of traditional fault diagnosis algorithms. Because rolling bearings are subject to many unpredictable and unquantifiable factors during the operation of electromechanical equipment, traditional fault diagnosis methods have certain limitations.

[0004] Deep learning methods use deep learning network models to automatically and intelligently learn the parameters of the input network and extract features. The network model then classifies the input and determines the type of bearing fault. Commonly used one-dimensional convolutional neural networks for fault diagnosis include WDCNN, and two-dimensional convolutional neural networks include ResNet50. These networks can intelligently complete the entire bearing fault diagnosis process by autonomously learning fault features from signals. While the deep learning-based rolling bearing fault diagnosis network described above can automatically extract features and classify faults, directly training the neural network to learn from raw bearing vibration signals can lead to difficulties in extracting effective features, increasing the training difficulty and potentially preventing convergence. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this disclosure is to provide a rolling bearing fault diagnosis method based on a parallel convolutional neural network. This method effectively avoids the interference of noise in the vibration signal by inputting one-dimensional and two-dimensional features extracted from the original vibration signal into the parallel convolutional neural network, thereby solving the problem of inaccurate fault diagnosis of vibration signal under interference.

[0006] To achieve the above objectives, this disclosure provides the following technical solutions:

[0007] A method for diagnosing rolling bearing faults based on parallel convolutional neural networks includes the following steps:

[0008] S100: Acquire the raw vibration signal X of the rolling bearing under test. io (t);

[0009] S200: For the original vibration signal X io (t) is preprocessed to obtain the preprocessed vibration signal X. i (t);

[0010] S300: For the pre-processed vibration signal X i (t) Perform the first overlap sampling to obtain the first sample signal x. m (t);

[0011] S400: For the pre-processed vibration signal X i (t) Perform a second overlap sampling to obtain the second sample signal y. m (t);

[0012] S500: Based on the first sample signal x m (t) and the second sample signal y m (t) Obtain the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test respectively;

[0013] S600: A fused dataset is constructed based on the one-dimensional tensor set and two-dimensional tensor set of the rolling bearing under test, and then input into a trained parallel convolutional neural network to identify the faults of the rolling bearing under test.

[0014] Preferably, in step S200, the original vibration signal X is... io (t) Preprocessing includes the following steps:

[0015] S201: Remove the original vibration signal X io The DC component in (t);

[0016] S202: The original vibration signal X after removing the DC component. io (t) is normalized.

[0017] Preferably, in step S201, the original vibration signal X is removed using the following formula. io DC component in (t):

[0018]

[0019] Among them, X i (t) represents the original vibration signal after removing the DC component, X io (t) represents the original vibration signal, X io (t) k Represents the original vibration signal X io The amplitude corresponding to the kth sampling point in (t), len(X) io (t) represents the original vibration signal X. io The number of sampling points in (t), where io represents the original vibration signal X. io (t) is the data sequence number, where t represents the original vibration signal X. io The time series of (t).

[0020] Preferably, in step S202, the original vibration signal X after removing the DC component is calculated using the following formula. i (t) is normalized:

[0021]

[0022] and

[0023]

[0024]

[0025] Among them, MIN(X) i (t) represents the minimization process, σ represents the scaling factor, and X i (t) max X represents iThe maximum value of X in (t), i (t) min X represents i The minimum value in (t).

[0026] Preferably, in step S300, the preprocessed vibration signal is subjected to a first overlapping sampling by using a sliding window.

[0027] Preferably, the position of the sliding window is calculated using the following formula:

[0028]

[0029] Where, N total This represents the pre-processed vibration signal X. i The number of sampling points N(t) total [(m-1)(L+step), mL+(m-1)step] represents all sampling points between the (m-1)(L+step)th sampling point and the mL+(m-1)stepth sampling point, where L represents the length of the sliding window, step represents the step size of the sliding window, and x m (t) represents the first sample signal, and m represents the first sample signal x. m The sequence number of (t).

[0030] Preferably, the preprocessed vibration signal X i The number of sampling points for (t) is calculated using the following formula:

[0031]

[0032] Where k represents the length coefficient, which is defaulted to k=2, and n represents the bearing speed. Z represents the sampling interval and Z represents a positive integer.

[0033] Preferably, in step S500, the step based on the first sample signal x m (t) Obtaining the one-dimensional tensor set of the rolling bearing to be tested includes the following steps:

[0034] S501: For the first sample signal x m Perform a Hilbert transform on (t) and calculate the first sample signal x. m The envelope signal A of (t) m (t);

[0035] S502: For the first sample signal x m (t) Perform integrated empirical mode decomposition to obtain the decomposed modal components u i (t), and based on the correlation coefficient α, the modal components u i (t) is reconstructed to obtain the reconstructed signal U. m(t);

[0036] S503: For the first sample signal x m (t) Perform variational mode decomposition to obtain the decomposed modal components v i (t), and based on the correlation coefficient β, the modal components v i (t) is reconstructed to obtain the reconstructed signal V. m (t);

[0037] S504: Envelope signal A m (t), reconstructed signal U i (t) and reconstructed signal V i (t) are respectively converted into a one-dimensional signal tensor A. t m (t), U t m (t) and V t m (t), and stack the signal tensors of the same category in the specified dimension dim=0 to obtain one-dimensional tensor sets T1, T2 and T3 respectively.

[0038] Preferably, in step S501, the first sample signal x is calculated using the following formula. m (t) Perform Hilbert transform:

[0039]

[0040] Among them, e m (t) represents the first sample signal x m (t) is the signal obtained after Hilbert transform, where t represents time, τ represents the integration variable, and m represents the first sample signal x. m The sequence number of (t).

[0041] Preferably, in step S501, the first sample signal x is calculated in the following manner. m The envelope signal A of (t) m (t):

[0042] Using the first sample signal x m (t) is the real part, and the signal e after Hilbert transform is used. m (t) represents the imaginary part, constructing the m-th sample signal x. m The analytic signal α of (t) m (t):

[0043] α m (t)=x m (t)+je m (t)

[0044] Take the analytic signal α mThe amplitude of (t) is used as the envelope signal A. m (t), that is

[0045]

[0046] Where, x m (t) represents the first sample signal, e m (t) represents the first sample signal x m (t) The signal after Hilbert transform.

[0047] Preferably, in step S502, the first sample signal x is calculated using the following formula. m (t) Perform integrated empirical mode decomposition:

[0048]

[0049] Among them, u i (t) represents the modal components after integrated empirical mode decomposition, i represents the i-th modal component obtained by decomposition, n represents the decomposition number, and r represents the modal component obtained by decomposition. n (t) represents the residual.

[0050] Preferably, in step S502, the correlation coefficient α is calculated using the following formula:

[0051]

[0052] in, Represents the modal component u i The average value of (t), Represents the first sample signal x m The mean of (t), where std represents the standard deviation operation;

[0053] The reconstructed signal U m (t) Modal components u with correlation coefficient α > 0.4 i (t) are obtained by adding them together, that is

[0054]

[0055] Where t represents t modal components with correlation coefficients α > 0.4.

[0056] Preferably, in step S503, a constrained variational model is constructed for the first sample signal x. m (t) Perform variational mode decomposition.

[0057] Preferably, the constrained variational model is expressed as:

[0058]

[0059] Among them, {vi {ω} represents the i-th intrinsic mode component after variational mode decomposition, I represents the number of decomposed signals, and {ω} i} represents the center frequency corresponding to the intrinsic modal component, δ(t) represents the Dirichlet function, * represents the convolution operation, and x m (t) represents the first sample signal, and π represents pi. The symbol represents the derivative, and j represents the imaginary unit.

[0060] Preferably, in step S503, the correlation coefficient β is calculated using the following formula:

[0061]

[0062] in, Represents the modal component v i The average value of (t), Represents the first sample signal x m The mean of (t), where std represents the standard deviation operation;

[0063] The reconstructed signal V m (t) is derived from the intrinsic modal components v with a correlation coefficient β > 0.4. i (t) are obtained by adding them together, that is

[0064]

[0065] Where t represents t intrinsic modal components with correlation coefficients α > 0.4.

[0066] Preferably, in step S400, the normalized vibration signal is subjected to a second overlapping sampling by using a sliding window.

[0067] Preferably, in step S500, the step based on the second sample signal y m (t) Obtaining the two-dimensional tensor set of the rolling bearing under test includes the following steps:

[0068] S5001: For the second sample signal y m (t) Perform continuous wavelet transform to obtain the corresponding wavelet coefficients, and plot the wavelet time-frequency diagram based on the wavelet coefficients;

[0069] S5002: Draw a Markov transition field diagram;

[0070] S5003: Draw the Gram corner field diagram;

[0071] S5004: Convert the wavelet time-frequency diagram, Markov transition field diagram, and Gram angle field diagram into two-dimensional tensors respectively to obtain the two-dimensional tensor sets T4, T5, and T6 of the rolling bearing under test.

[0072] Preferably, step S5001 includes the following steps:

[0073] S50011: Using the Python wavelet analysis toolkit PyWavelets to analyze the second signal sample y m (t) Perform continuous wavelet transform to obtain wavelet coefficients, specifically:

[0074]

[0075] Where WT(a,τ) represents wavelet coefficients, a represents the scaling factor, τ represents the time shift factor, * represents the conjugate function, and Ψ represents the conjugate function. a,τ (t) represents the wavelet basis function, and t represents the time series;

[0076] S50012: Draw a wavelet time-frequency diagram based on the time series t of the wavelet coefficients, the actual frequency series f, and the modulus abs(WT(a,τ)) of the wavelet coefficients.

[0077] Preferably, in step S5001, the pcolor function in the matplotlib plotting library is used to plot the wavelet time-frequency diagram based on the wavelet time series t, the actual frequency series f, and the modulus abs(WT(a,τ)) of the wavelet coefficients.

[0078] Preferably, the actual frequency sequence f of the wavelet is calculated using the following formula:

[0079] scalef = F s wcf

[0080] Where scale represents the scale sequence of wavelet f, F s Represents the second sample signal y m The sampling frequency of (t) is given by wcf, which represents the center frequency of wavelet f.

[0081] Preferably, step S5002 includes the following steps:

[0082] S50021: Based on the time sequence of a first-order Markov process, the second signal sample y is... m (t) is visualized as a complex network, yielding the Markov matrix;

[0083] S50022: Draw Markov transition field diagrams based on Markov matrices.

[0084] Preferably, step S5003 includes the following steps:

[0085] S50031: The second sample signal y in the Cartesian coordinate system... m (t) is normalized to the interval [-1,1] and then converted to polar coordinates.

[0086] S50032: Based on the second sample signal y m (t) The normalized polar coordinate representation is used to generate the Gram angular field matrix G through trigonometric functions;

[0087] S50033: Draw Gram angle field diagrams based on Gram angle field matrix G.

[0088] Preferably, in step S50032, the Gram angular field matrix G is represented as:

[0089]

[0090] in, Represents the second sample signal y m (t) Normalized time series points x n The polar coordinates.

[0091] Preferably, the parallel convolutional neural network includes a one-dimensional convolutional neural network and a two-dimensional convolutional neural network. The one-dimensional convolutional neural network includes a fully connected layer, a convolutional layer, a batch normalization layer, an activation function, and a pooling layer. The two-dimensional convolutional neural network adopts the InceptionV1 network.

[0092] Preferably, in step S600, the training of the parallel convolutional neural network includes the following steps:

[0093] S601: Collect the health signal of the healthy bearing and the fault signal of the faulty bearing respectively, then classify the fault signals according to different fault locations and fault degrees, and combine the health signal and the classified fault signal to obtain the original signal set;

[0094] S602: Obtain the one-dimensional tensor set and the two-dimensional tensor set of the original signal set;

[0095] S603: Combine the one-dimensional tensor set and two-dimensional tensor set of the original signal set with the labels of the fused dataset to construct the fused dataset, and divide the fused dataset into training set, validation set and test set;

[0096] S604: A large minimum loss function is preset. The parallel convolutional neural network is trained using the training set. At the same time, the parallel convolutional neural network is validated using the validation set during training. Whenever the loss function value on the validation set is less than the preset minimum loss function value, and the loss function on the training set is still decreasing, the minimum loss function value is set to equal the current loss function value on the validation set and saved. The model parameters of the network at this time are overwritten and saved. When training is completed, the trained parallel neural network is obtained.

[0097] S605: Test the trained parallel neural network using the test set. If the test passes, the trained parallel neural network is obtained; otherwise, the hyperparameters need to be adjusted and the parallel neural network needs to be retrained.

[0098] Preferably, in step S605, the classification performance of the model is evaluated using a confusion matrix and a t-distribution-random neighbor embedding to determine whether the parallel neural network test has passed.

[0099] Preferably, in step S604, before training the parallel neural network using the training set, it is necessary to convert the one-dimensional signal tensor A in the training set into a single array. t m (t), U t m (t) and V t m (t) assign different weights ρ k The merging is then performed using the following formula to obtain the one-dimensional fusion tensor x. f (t):

[0100] x f (t)=ρ1A t m (t)+ρ2U t m (t)+ρ3V t m (t)

[0101] Among them, A t m (t) is the tensor of the envelope signal; U t m (t) is the tensor of the integrated empirical mode decomposition reconstructed signal; V t m (t) is the tensor of the signal reconstructed by variational mode decomposition.

[0102] Preferably, the weight ρ k Calculated using the following formula:

[0103]

[0104] Where, ρ k kur represents the weights corresponding to the k-th type tensor. k Let be the kurtosis corresponding to the k-th type tensor; Let be the kurtosis of the three tensors.

[0105] This disclosure also provides a rolling bearing fault diagnosis device based on a parallel convolutional neural network, comprising:

[0106] The acquisition module is used to acquire the raw vibration signal X of the rolling bearing under test. io (t);

[0107] The preprocessing module is used to process the original vibration signal X. io (t) is preprocessed to obtain the preprocessed vibration signal X. i (t);

[0108] The first sampling module is used to process the preprocessed vibration signal X. i (t) Perform the first overlap sampling to obtain the first sample signal x. m (t);

[0109] The second sampling module is used to process the preprocessed vibration signal X. i (t) Perform a second overlap sampling to obtain the second sample signal y. m (t);

[0110] The tensor set acquisition module is used to acquire tensor sets based on the first sample signal x. m (t) and the second sample signal y m (t) Obtain the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test respectively;

[0111] The diagnostic module is used to construct a fused dataset based on the one-dimensional tensor set and two-dimensional tensor set of the rolling bearing under test, and input it into a trained parallel convolutional neural network to identify the faults of the rolling bearing under test.

[0112] This disclosure also provides a computer storage medium, including:

[0113] Memory is used to store multiple computer instructions;

[0114] A processor is used to execute computer instructions to implement the methods described in any of the preceding methods.

[0115] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0116] 1. This disclosure does not directly use the original vibration signal as input. Instead, it uses envelope analysis to extract one-dimensional fault features of the vibration signal and continuous wavelet transform to extract two-dimensional fault features of the vibration signal. The extracted one-dimensional and two-dimensional features are input into a parallel convolutional neural network, which effectively avoids the interference of noise in the vibration signal and extracts features rich in fault information from the vibration signal, thus better solving the problem of inaccurate fault diagnosis of vibration signal under interference.

[0117] 2. The parallel convolutional neural network constructed in this disclosure uses envelope spectrum and wavelet time-frequency diagram to build a diagnostic model, which has the advantages of accurate diagnosis and good fault classification effect. The accuracy of the network in identifying faults can reach more than 99%. Attached Figure Description

[0118] Figure 1 This is a flowchart of a rolling bearing fault diagnosis method based on a parallel convolutional neural network provided in one embodiment of the present invention;

[0119] Figure 2 This is a schematic diagram of overlapping sampling in a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0120] Figure 3 This is a schematic diagram of the structure of a one-dimensional convolutional neural network convolutional module in a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0121] Figure 4 This is a schematic diagram of the Inception module of a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0122] Figure 5 This is a schematic diagram of the SE-Inception module of a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention;

[0123] Figure 6 This is a schematic diagram of the parallel convolutional neural network structure in a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0124] Figure 7 This is a schematic diagram of a bearing test bench for a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0125] Figure 8 This is a wavelet time-frequency diagram of vibration signals corresponding to different fault categories in a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0126] Figure 9 This is a Markov transition field diagram corresponding to vibration signals of different fault categories in a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0127] Figure 10 This is a Gram angle field diagram corresponding to vibration signals of different fault categories in a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0128] Figure 11 This is a confusion matrix diagram of a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention;

[0129] Figure 12 This is a t-SNE visualization result of a rolling bearing fault diagnosis method based on a parallel convolutional neural network according to an embodiment of the present invention.

[0130] Figure 13 This is a diagram comparing the diagnostic accuracy of a parallel convolutional neural network with that of an existing convolutional neural network on the training set.

[0131] Figure 14 This is a schematic diagram comparing the diagnostic accuracy of a parallel convolutional neural network with that of an existing convolutional neural network on a validation set. Detailed Implementation

[0132] The following will refer to the appendix. Figures 1 to 14 Specific embodiments of this disclosure are described in detail. While specific embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0133] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out this disclosure; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of this disclosure. The scope of protection of this disclosure is determined by the appended claims.

[0134] To facilitate understanding of the embodiments of this disclosure, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of this disclosure.

[0135] In one embodiment, such as Figure 1 As shown, this disclosure provides a rolling bearing fault diagnosis method based on a parallel convolutional neural network, including the following steps:

[0136] S100: Acquire the raw vibration signal X of the rolling bearing under test. io (t);

[0137] S200: For the original vibration signal X io (t) is preprocessed to obtain the preprocessed vibration signal X. i (t);

[0138] S300: For the pre-processed vibration signal X i (t) Perform the first overlap sampling to obtain the first sample signal x.m (t);

[0139] S400: For the pre-processed vibration signal X i (t) Perform a second overlap sampling to obtain the second sample signal y. m (t);

[0140] S500: Based on the first sample signal x m (t) and the second sample signal y m (t) Obtain the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test respectively;

[0141] S600: A fused dataset is constructed based on the one-dimensional tensor set and two-dimensional tensor set of the rolling bearing under test, and then input into a trained parallel convolutional neural network to identify the faults of the rolling bearing under test.

[0142] The above embodiments constitute the complete technical solution of this disclosure. This embodiment employs envelope analysis and reconstructs the modal components from integrated empirical mode decomposition and variational mode decomposition, discarding modal components with low correlation coefficients to the original vibration signal, enabling the neural network to effectively extract one-dimensional fault features from the original vibration signal. Furthermore, it uses continuous wavelet transform to extract two-dimensional fault features from the vibration signal and Markov transition fields and Gram angle fields to extract two-dimensional spatial features of the vibration signal. This effectively avoids interference caused by noise in the original vibration signal, thereby improving the accuracy of rolling bearing fault identification.

[0143] In another embodiment, in step S200, the original vibration signal X is... io (t) Preprocessing includes the following steps:

[0144] S201: Remove the original vibration signal X io The DC component in (t);

[0145] In this step, the original vibration signal X is removed. io The DC component in (t) is determined by the following formula:

[0146]

[0147] Among them, X i (t) represents the preprocessed original vibration signal, X io (t) represents the original vibration signal, X io (t) k Represents the original vibration signal X io The amplitude corresponding to the kth sampling point in (t), len(X) io (t) represents the original vibration signal X. ioThe number of sampling points in (t), where io represents the original vibration signal X. io (t) is the data sequence number, where t represents the original vibration signal X. io The time series of (t).

[0148] S202: The original vibration signal X after removing the DC component. io (t) is normalized.

[0149] In this step, the original vibration signal X after removing the DC component is calculated using the following formula. io (t) is normalized:

[0150]

[0151] in,

[0152]

[0153]

[0154] Among them, MIN(X) i (t) represents the minimization process, σ represents the scaling factor, and X i (t) max X represents i The maximum value of X in (t), i (t) min X represents i The minimum value in (t).

[0155] By normalizing the original vibration signal Xio(t) after removing the DC component, it is possible to avoid the vibration signal X i The singular values ​​in (t) affect the neural network during training, thereby accelerating the convergence speed of the neural network.

[0156] In another embodiment, in step S300, the normalized vibration signal is first overlapped by using a sliding window.

[0157] In this embodiment, to increase the number of training samples, the normalized original vibration signal needs to be augmented through overlap sampling. Since one-dimensional fault signals have temporal and periodic characteristics, training samples can be collected from the normalized original vibration signal according to... Figure 2 The overlapping sampling is performed as follows: a sliding window of length L is used to sample the normalized original vibration signal. The sliding window moves by a step size of step each time, so that each signal segment obtained through sampling partially overlaps with the next signal segment, thus obtaining the m-th overlapping sample segment x. m (t) is represented as:

[0158]

[0159] Where, N total This represents the pre-processed vibration signal X. i The number of sampling points N(t) total [(m-1)(L+step), mL+(m-1)step] represents all sampling points between the (m-1)(L+step)th sampling point and the mL+(m-1)stepth sampling point, where L represents the length of the sliding window, step represents the step size of the sliding window, and x m (t) represents the first sample signal, and m represents the first sample signal x. m (The sequence number of t0.)

[0160] It should be noted that, to ensure that each sample during overlap sampling contains fault characteristic information of at least one revolution of the bearing, the preprocessed vibration signal x i The number of sampling points N in (t) should be no less than the number of sampling points within two revolutions of the bearing. To improve the computational efficiency of the subsequent fast Fourier transform, the number of sampling points should be a power of 2.

[0161] In summary, the vibration signal X after preprocessing during overlap sampling i The number of sampling points N of (t) should satisfy:

[0162]

[0163] Where k represents the length coefficient, with the default value being k=2, and the actual value can be adjusted as needed; n represents the bearing speed; and ΔT represents the sampling interval.

[0164] In another embodiment, in step S500, the step based on the first sample signal x m (t) Obtaining the one-dimensional tensor set of the rolling bearing to be tested includes the following steps:

[0165] S501: For the first sample signal x m Perform a Hilbert transform on (t) and calculate the first sample signal x. m The envelope signal A of (t) m (t);

[0166] In this step, the first sample signal x is expressed by the following formula. m (t) Perform Hilbert transform:

[0167]

[0168] Among them, e m (t) represents the first sample signal x m(t) is the signal obtained after Hilbert transform, where t represents time, τ represents the integration variable, and m represents the first sample signal x. m The index of (t);

[0169] Let the first sample signal x obtained by overlap sampling m (t) represents the imaginary part, and the signal e obtained after the Hilbert transform is... m (t) represents the imaginary part, constructing the m-th sample signal x. m The analytic signal α of (t) m (t):

[0170] α m (t)=x m (t)+je m (t)

[0171] Furthermore, take the analytic signal α m The amplitude of (t) is used as the envelope signal A. m (t), Where, x m (t) represents the first sample signal, e m (t) represents the first sample signal x m (t) The signal after Hilbert transform.

[0172] S502: For the first sample signal x m (t) Perform integrated empirical mode decomposition to obtain the decomposed modal components u i (t), and based on the correlation coefficient α, the modal components u i (t) is reconstructed to obtain the reconstructed signal U. m (t);

[0173] In this step, the first sample signal x is expressed by the following formula. m (t) Perform integrated empirical mode decomposition:

[0174]

[0175] Among them, u i (t) represents the modal components after integrated empirical mode decomposition, i represents the i-th modal component obtained by decomposition, n represents the decomposition number, and r represents the modal component obtained by decomposition. n (t) represents the residual.

[0176] The correlation coefficient α is calculated using the following formula:

[0177]

[0178] in, Represents the modal component u i The average value of (t), Represents the first sample signal x m The mean of (t), where std represents the standard deviation operation;

[0179] The reconstructed signal U m (t) Modal components u with correlation coefficient α > 0.4 i (t) are obtained by adding them together, that is

[0180]

[0181] Where t represents t modal components with correlation coefficients α > 0.4.

[0182] It should be noted that the correlation coefficient α reflects the degree of correlation between the modal components and the original vibration signal. The smaller the correlation coefficient, the weaker the correlation between the two, and these modal components can be considered as noise in the original vibration signal. Generally, a correlation coefficient < 0.4 is considered a weak correlation. Therefore, in this embodiment, modal components with a correlation coefficient > 0.4 are selected for reconstruction, discarding components in the modal components that are unrelated to the original vibration signal, thus avoiding interference caused by noise in the original vibration signal.

[0183] S503: For the first sample signal x m (t) Perform variational mode decomposition to obtain the decomposed modal components v i (t), and based on the correlation coefficient β, the modal components v i (t) is reconstructed to obtain the reconstructed signal V. m (t);

[0184] In this step, a constrained variational model is constructed for the first sample signal x. m (t) Perform variational mode decomposition, and the constrained variational model is expressed as:

[0185]

[0186] Among them, {v i {ω} represents the i-th intrinsic mode component after variational mode decomposition, I represents the number of decomposed signals, and {ω} i} represents the center frequency corresponding to the intrinsic modal component, δ(t) represents the Dirichlet function, * represents the convolution operation, and x m (t) represents the first sample signal, and π represents pi. The symbol represents the derivative, and j represents the imaginary unit.

[0187] The correlation coefficient β is calculated using the following formula:

[0188]

[0189] in, Represents the modal component v i The average value of (t), Represents the first sample signal x m The mean of (t), where std represents the standard deviation operation;

[0190] The reconstructed signal V m (t) is derived from the intrinsic modal components v with a correlation coefficient β > 0.4. i (t) are obtained by adding them together, that is:

[0191]

[0192] Where t represents t intrinsic modal components with correlation coefficients α > 0.4.

[0193] S504: Use the torch.tensor() function in the PyTorch library to extract the envelope signal A. m (t), reconstructed signal U i (t) and reconstructed signal V i (t) are respectively converted into a one-dimensional signal tensor A. t m (t), U t m (t) and V t m (t) (that is, change the data type from ndarray to tensor), and stack the signal tensors of the same category on the specified dimension dim=0 using torch.stack() in the pytorch library to obtain one-dimensional tensor sets T1, T2 and T3 respectively.

[0194] In another embodiment, in step S400, the normalized vibration signal is subjected to a second overlapping sampling by using a sliding window.

[0195] In this embodiment, the second overlapping sampling method is the same as the first overlapping sampling method, and will not be described again here.

[0196] In another embodiment, in step S500, the step based on the sample signal y m (t) Obtaining the two-dimensional tensor set of the rolling bearing under test includes the following steps:

[0197] S5001: For the second sample signal y m (t) Perform continuous wavelet transform to obtain the corresponding wavelet coefficients, and plot the wavelet time-frequency diagram based on the wavelet coefficients;

[0198] This step involves plotting the wavelet time-frequency graph, including the following steps:

[0199] S50011: Selecting the cmor3-3 wavelet as the basis function, and using the Python wavelet analysis toolkit PyWavelets to analyze the second sample signal y. m (t) is subjected to continuous wavelet transform to obtain wavelet coefficients.

[0200] Specifically, the second sample signal y is expressed by the following formula. m (t) Perform continuous wavelet transform:

[0201]

[0202] Where WT(a,τ) represents wavelet coefficients, a represents the scaling factor, τ represents the time shift factor, * represents the conjugate function, and Ψ represents the conjugate function. a,τ (t) represents the wavelet basis function, and t represents time.

[0203] It should be noted that the cmor3-3 wavelet exhibits an exponentially decaying oscillation, similar to the decaying component of the transient impact generated when a rolling bearing fails. Therefore, it is often used for fault feature extraction in rolling bearings, and its expression is as follows:

[0204]

[0205] Where α is the envelope factor, f c Let t be the center frequency of the wavelet, and t represent time.

[0206] S50012: Obtain the second sample signal y m After obtaining the wavelet coefficients (t), the pcolor function in the matplotlib plotting library is used to plot the wavelet time-frequency graph based on the time series t of the wavelet coefficients, the actual frequency series f, and the modulus abs(WT(a,τ)) of the wavelet coefficients. Furthermore, the obtained wavelet time-frequency graph can be converted into a three-channel true-color image with 3-channel RGB and a pixel value of 224×224, with a total of 7 categories, as detailed below. Figure 8 As shown.

[0207] Specifically, the actual frequency sequence f of the wavelet mentioned above is calculated using the following formula:

[0208] scalef = F s wcf

[0209] Where scale represents the scale sequence of wavelet f, F s Represents the sample signal y m The sampling frequency of (t) is given by wcf, which represents the center frequency of wavelet f.

[0210] S5002: Draw a Markov transition field diagram;

[0211] This step, drawing the Markov transition field diagram, specifically includes the following steps:

[0212] S50021: Based on the time sequence of a first-order Markov process, the second signal sample y is... m (t) is visualized as a complex network, yielding a Markov matrix, an example of which is shown below:

[0213]

[0214] Where, ω i,j This represents the frequency at which quantization box i transitions to quantization box j, where x is the timing data and q is the frequency. i ,q j This indicates a partition box.

[0215] S50022: Based on the Markov matrix and using the MarkovTransitionField function from the pyts library, plot the Markov transition field diagram, specifically as follows: Figure 9 As shown.

[0216] S5003: Draw the Gram corner field diagram;

[0217] This step of drawing the Gram angle field diagram specifically includes the following steps:

[0218] S50031: The second sample signal y in the Cartesian coordinate system... m (t) is normalized to the interval [-1,1] and then converted to polar coordinates.

[0219] S50032: Based on the second sample signal y m (t) The normalized polar coordinate representation is used to generate the Gram angular field matrix G through trigonometric functions;

[0220] In this step, the type of the Gram angle field matrix G is a Gram angle summation field, specifically represented as:

[0221]

[0222] in, Represents the second sample signal y m (t) Normalized time series points x n The polar coordinates.

[0223] S50033: Based on the Gram angle field matrix G, and using the GramianAngularField function from the pyts library, draw the Gram angle field plot, specifically as follows: Figure 10 As shown.

[0224] S5004: Use the transforms function in the Torchvision library to convert the wavelet time-frequency map, Markov transition field map and Gram angle field map into two-dimensional tensors respectively. The numerical range of a single channel of the image is [0,255]. Divide the tensor by 255 to obtain a normalized two-dimensional tensor. Two-dimensional image tensors of the same category are stacked in the 0th dimension to obtain the two-dimensional tensor set T4, T5 and T6 of the rolling bearing under test.

[0225] In another embodiment, step S600, the training process of the parallel convolutional neural network includes the following steps:

[0226] S601: Collect the health signal of the healthy bearing and the fault signal of the faulty bearing respectively, then classify the fault signals according to different fault locations and fault degrees, and combine the health signal and the classified fault signal to obtain the original signal set;

[0227] In this step, the rolling bearing vibration signal used in this embodiment comes from the signal collected by the rolling bearing condition monitoring test bench at the University of Paderborn. The bearing test bench is as follows: Figure 7 As shown in the diagram, the test bench consists of, from left to right, a motor, a torque measuring shaft, a rolling bearing test module, a flywheel, and a load motor. Experimental data is collected by installing rolling bearings with different damage types on the bearing test bench. All test bearings used on this test bench are model 6203 rolling bearings. Vibration sensors collect rolling bearing vibration signals at a sampling frequency of 64 kHz. Each signal segment is collected for 4 seconds, resulting in approximately 256,000 sampling points for the bearing vibration signals. This embodiment uniformly uses bearing vibration signals under the N15_M07_F10 operating condition, and its specific operating parameters are shown in Table 1.

[0228] Table 1 Operating Parameters

[0229]

[0230] In addition, this embodiment selected three different types of outer ring fault vibration signals and three different types of inner ring fault vibration signals, and combined them with normal bearing vibration signals to form original signal sets of seven different state categories, as shown in Table 2:

[0231] Table 2 shows the selected bearing vibration signals.

[0232]

[0233] S602: Obtain the one-dimensional tensor set and the two-dimensional tensor set of the original signal set;

[0234] After obtaining the original signal set, the following steps are required: First, preprocessing is needed to obtain a preprocessed signal. Second, oversampling is performed on the preprocessed signal to achieve data augmentation. The calculated number of signal sampling points for one revolution of the 6203 bearing is 2560. To ensure the oversampled signal segments contain complete fault information and reduce FFT computation time, 8192 signal sampling points were selected for oversampling, with an oversampling window movement step of 1000. Each category of original vibration signal was oversampled 200 times, resulting in 1400 sample signals. Third, a Hilbert transform was performed on the sample signals to obtain analytic signals, and the envelope signal of the analytic signal was calculated. Fourth, after performing an FFT on the envelope signal, amplitude normalization was performed, and the amplitude of the single-sided amplitude spectrum of the envelope spectrum was taken to construct a one-dimensional tensor set. After processing, a one-dimensional tensor with 1400 sampling points and 4096 sampling points was finally obtained. Fifth, oversampling of the original bearing vibration signal was performed to achieve data augmentation. The calculated number of signal sampling points for one revolution of the 6203 bearing is 2560. To ensure that the overlapping sampled signal segments contain complete fault information, the number of overlapping sampled signal points was selected to be 5120, the overlapping sampling window movement step size was 1000, and the original vibration signal of each category was overlapped 200 times, resulting in a total of 1400 sample signals. Sixth, the CMOR3-3 wavelet was selected as the basis function, and continuous wavelet transform was performed on the overlapping sampled signal segments. The resulting wavelet time-frequency image was converted into a three-channel RGB image with a pixel value of 224×224, with a total of 7 categories, such as... Figure 8 As shown; Seventh, the wavelet time-frequency graph is converted into a tensor and normalized to construct a two-dimensional tensor set, finally obtaining 1400 two-dimensional tensors with a dimension size of 3×224×224.

[0235] S603: Combine the one-dimensional tensor set and two-dimensional tensor set of the original signal set with the labels of the constructed fusion dataset, and use the Dataset class in the PyTorch library to construct the fusion dataset, and divide the fusion dataset into training set, validation set and test set;

[0236] In this step, the one-dimensional tensor set and two-dimensional tensor set of the original signal set are obtained in the same way as the one-dimensional tensor set and two-dimensional tensor set of the rolling bearing under test mentioned above, and will not be repeated here.

[0237] In addition, the labels of the fused dataset are constructed using continuous numerical variables (continuous integers starting from 0). The number of label categories is the same as the number of rolling bearing state categories. Each label corresponds to an envelope spectrum tensor, an EMMD reconstructed signal tensor, a VMD reconstructed signal tensor, a wavelet time-frequency plot tensor, a Markov transition field tensor, and a Gram angle field tensor.

[0238] S604: A large minimum loss function is preset. The parallel convolutional neural network is trained using the training set. At the same time, the parallel convolutional neural network is validated using the validation set during training. Whenever the loss function value on the validation set is less than the preset minimum loss function value, and the loss function on the training set is still decreasing, the minimum loss function value is set to equal the current loss function value on the validation set and saved. The model parameters of the network at this time are overwritten and saved. When training is completed, the trained parallel neural network is obtained.

[0239] In this step, before training the parallel neural network using the training set, it is necessary to convert the one-dimensional signal tensor A in the training set into a single array. t m (t), U t m (t) and V t m (t) assign different weights ρ k The signals are then merged (since envelope analysis, ensemble empirical mode decomposition, and variational mode decomposition cannot guarantee the complete and accurate extraction of all fault features for each signal segment, assigning different weights to the signals processed by these three methods and merging the signals helps avoid omissions during feature extraction) to obtain a one-dimensional fusion tensor x. f (t), specifically, the following formula is used for merging:

[0240] x f (t)=ρ1A t m (t)+ρ2U t m (t)+ρ3V t m (t)

[0241] Among them, A t m (t) is the tensor of the envelope signal; U t m (t) is the tensor of the integrated empirical mode decomposition reconstructed signal; V t m (t) is the tensor of the signal reconstructed by variational mode decomposition.

[0242] In addition, the weight ρ k The calculation formula is:

[0243]

[0244] Where, ρ k kur represents the weights corresponding to the k-th type tensor. k Let be the kurtosis corresponding to the k-th type tensor; Let be the kurtosis of the three tensors.

[0245] Obtaining the one-dimensional fusion tensor x f After (t), it is necessary to perform Fourier transform and amplitude normalization processing to retain only the single-sided amplitude spectrum to obtain the number of sampling points for each signal segment. The spectral amplitude tensor sequence g(t) is used. It should be noted that the spectral features can be obtained through Fourier transform, so as to better extract features and make the frequency domain features of the fault signal more obvious than the time domain features; through amplitude normalization, it is easy to limit the spectral range, thereby avoiding the influence of singular values ​​on the convergence of parallel neural networks.

[0246] The above spectral amplitude tensor sequence g(t) is expressed as:

[0247]

[0248] Where FFT stands for Fast Fourier Transform, abs represents the modulo operation, and N is the number of elements in the tensor. Represents the signal tensor x f The first element to the second element in (t) A tensor sequence consisting of all elements between each element.

[0249] After completing the above processing, the spectral tensor g(t) can be obtained. Then, the spectral tensor g(t), the three two-dimensional tensors transformed from the wavelet time-frequency diagram, the Markov transition field diagram, and the Gram angle field diagram, and the labels are input into the parallel convolutional neural network for training.

[0250] It should be noted that the minimum loss function value of the network needs to be preset before training begins. min =10 3 Given the total number of training epochs and the current epoch number k, record the loss function value of the neural network on the training set after each forward and backward propagation. train_k (k≤epoch) and the loss function value of the validation set val_k (k≤epoch), and update the loss according to the following formula. min :

[0251]

[0252] Each loss min When values ​​change, the parameters of the currently saved parallel convolutional neural network are overwritten. After network training is complete, a globally optimal parallel neural network model can be obtained; otherwise, the network hyperparameters need to be adjusted and the network retrained. Specifically, this network uses the Adam optimization algorithm to update the network's bias and weight parameters. The Adam algorithm is expressed as:

[0253]

[0254] m t =β1m t-1 +(1-β1)g t

[0255] v t =β2v t-1 +(1-β2)g t 2

[0256]

[0257]

[0258]

[0259] Where t is the time step, initialized to 0; g t θ represents the gradient at time step t; θ represents the parameters to be updated. Let be a stochastic objective function with parameter θ; β1 and β2 are the exponential decay rates of the first and second moments, respectively; m t For the first moment estimate of the gradient; v t This is for estimating the second moment of the gradient; For m t Correction; For v t The correction is η, the learning rate is η, and ε is a constant added to maintain numerical stability.

[0260] Parallel Convolutional Neural Networks (PCNNs) use the Adam optimization algorithm to update the network's weight parameters. The Adam algorithm can be expressed as follows:

[0261]

[0262] m t =β1m t-1 +(1-β1)g t

[0263] v t =β2v t-1 +(1-β2)g t 2

[0264]

[0265]

[0266]

[0267] Where t is the time step, initialized to 0; gt θ represents the gradient at time step t; θ represents the parameters to be updated. Let be a stochastic objective function with parameter θ; β1 and β2 are the exponential decay rates of the first and second moments, respectively; m t For the first moment estimate of the gradient; v t This is for estimating the second moment of the gradient; For m t Correction; For v t The correction; η is the learning rate; ε is a constant added to maintain numerical stability;

[0268] Parallel Convolutional Neural Networks (PCNNs) use the cross-entropy loss function to calculate the loss during training. The expression for the cross-entropy loss function is as follows:

[0269]

[0270] Where M is the number of categories, i.e., the sum of the number of normal states and the number of fault states; p(x i ) represents the true distribution of the i-th class (1 if the class is identified as i, 0 otherwise), q(x) i ) is used to predict the probability of this class.

[0271] S605: Test the trained parallel neural network using a test set. If the test passes, the trained parallel neural network is obtained; otherwise, the hyperparameters need to be adjusted (e.g., increase the learning rate, increase the number of network iterations, etc.) and the parallel neural network needs to be retrained.

[0272] In this step, a confusion matrix evaluation model is used to determine whether the parallel neural network test passes. Specifically, the confusion matrix evaluation model uses precision and recall as evaluation metrics.

[0273] Precision is expressed as:

[0274]

[0275] Recall rate is expressed as:

[0276]

[0277] Where TP represents the number of true positive classes, FP represents the number of false positive classes, and FN represents the number of false negative classes.

[0278] The network test is judged by using this model. When both precision and recall exceed 99%, the network test is considered passed. Otherwise, the hyperparameters need to be adjusted and the network needs to be retrained until the target is met. Figure 11This diagram illustrates the confusion matrix evaluation model. In this diagram, the precision and recall for all state categories are 100%, indicating that the network model performs well in testing.

[0279] Furthermore, the classification results of the model can be visualized using t-distributed random nearest neighbor embeddings (t-SNE) to evaluate the network's performance. t-SNE converts the similarity between data points into joint probabilities and attempts to minimize the KL divergence between the joint probabilities of low-dimensional and high-dimensional embeddings. When... Figure 12 When the T-SEN visualization image shows that labels of the same type are closely distributed and the boundaries between labels of different categories are clear, the network classification effect can be considered good.

[0280] In another embodiment, the parallel convolutional neural network includes a one-dimensional convolutional neural network and a two-dimensional convolutional neural network, wherein, as Figure 3 As shown, except for the fully connected layers, the remaining layers of a one-dimensional convolutional neural network consist of convolutional layers, batch normalization layers, activation function layers, and pooling layers.

[0281] It should be noted that the first layer of this one-dimensional convolutional neural network uses a large convolutional kernel to extract the one-dimensional fault features of the rolling bearing under test. The large convolutional kernel is obtained through deep learning optimization. Its advantage is that it can intelligently learn feature parameters for bearing fault diagnosis while removing features that are not helpful for diagnosis. To enhance the expressive power of the network, except for the first layer convolutional kernel which has a size of 128×1, the kernel size of the remaining layers is 3×1, as shown in Table 3.

[0282] Table 3

[0283]

[0284] Furthermore, to increase network depth and suppress overfitting, this one-dimensional convolutional neural network uses fewer convolutional kernel parameters and BN layers, and finally performs max pooling to extract one-dimensional tensor features.

[0285] In addition, to increase the network depth and suppress overfitting, the convolution module uses fewer convolution kernel parameters and BN layers, and finally performs max pooling to extract one-dimensional tensor features.

[0286] The two-dimensional convolutional neural network uses the InceptionV1 network, the core of which is as follows: Figure 4The Inception module shown assembles multiple convolutional kernel pooling operations into a single network module. When designing a neural network, the entire network structure is assembled on a module-by-module basis. The Inception structure designs a sparse network structure, yet it can generate dense data, which enhances the performance of the neural network, enabling multi-scale feature fusion, while ensuring efficient use of computational resources. Furthermore, to enable the network to achieve adaptive attention, allowing the neural network to focus on the parts of the wavelet time-frequency map containing fault features, this embodiment combines the Inception module with SENet (Squeeze-and-Excitation Networks) (e.g., Figure 5 As shown, an attention mechanism was added to the InceptionV1 network.

[0287] The Inception module employs several measures to achieve an approximate optimal sparse structure from dense components. First, it uses convolutional kernels of different sizes within the same layer to extract features from the previous layer, which are then concatenated at the end. This means the network can perceive local image regions of different sizes within the same layer and fuse features at different scales. Kernel sizes of 1, 3, and 5 are commonly used, with a stride of 1. As long as the image padding is set to 0, 1, and 2 respectively, features of the same dimension can be obtained after convolution, and these feature values ​​can then be directly concatenated. Furthermore, since pooling layers play a significant role in many networks, this operation is also embedded in the Inception module. Finally, as the features extracted by the neural network become increasingly abstract with increasing depth, and each feature involves a larger receptive area, the proportion of large-size convolutional kernels should be gradually increased in deeper layers of the network. The Inception structure has a total of four branches. The input feature maps are processed in parallel through these four branches to obtain four outputs, which are then concatenated along the depth dimension to obtain the final output. Therefore, the structural parameters of the Inception module are shown in Table 4.

[0288] Table 4

[0289]

[0290] The above describes how a one-dimensional convolutional neural network and an InceptionV1 network constitute a system like this. Figure 6 The parallel convolutional neural network shown here merges the tensor output by the InceptionV1 network with the tensor of the output length of the one-dimensional convolutional neural network, and outputs the final result after passing through a fully connected layer.

[0291] Below, this disclosure selects a one-dimensional convolutional neural network WDCNN and a two-dimensional convolutional neural network ResNet50 to compare with the parallel convolutional neural network proposed in this disclosure to illustrate the differences between the parallel convolutional neural network and existing convolutional neural networks for rolling bearing fault diagnosis.

[0292] Specifically, a 7-class one-dimensional dataset is constructed using the Dataset class in the PyTorch library, consisting of envelope signal tensors and continuous numerical variable labels (continuous integers starting from 0). This dataset is then divided into training and validation sets for training and evaluation of the one-dimensional convolutional neural network WDCNN.

[0293] A 7-class 2D dataset was constructed using the Gram angular field map tensor and continuous numerical variable labels (continuous integers starting from 0) in the PyTorch library's Dataset class. This dataset was then divided into training and validation sets, and the 2D convolutional neural network ResNet50 was trained and evaluated.

[0294] The parallel convolutional neural network PCNN is trained and evaluated using the fused dataset in this disclosure.

[0295] All three neural networks used the Adam optimizer with a learning rate of 1e-5, and the accuracy of the networks on the training and validation sets was recorded.

[0296] Furthermore, the diagnostic accuracy of the above different convolutional neural networks on the validation set is as follows: Figure 13 As shown, according to Figure 13 The accuracy of each convolutional neural network on the validation set is shown in Table 5.

[0297] Table 5

[0298] Network Name WDCNN ResNet50 PCNN Accuracy (%) 93.39 99.82 99.94

[0299] Furthermore, the diagnostic accuracy of the above different convolutional neural networks on the validation set is as follows: Figure 14 As shown, according to Figure 14 The accuracy of each convolutional neural network on the validation set is shown in Table 6.

[0300] Table 6

[0301] Network Name WDCNN ResNet50 PCNN Accuracy (%) 87.50 75.0 99.87

[0302] The accuracy of the three neural networks on the training and validation sets demonstrates that the parallel convolutional neural network proposed in this disclosure achieves the highest accuracy for fault diagnosis of rolling bearings under test, both on the training and validation sets, compared to existing convolutional neural networks.

[0303] In another embodiment, this disclosure also provides a rolling bearing fault diagnosis device based on a parallel convolutional neural network, comprising:

[0304] The acquisition module is used to acquire the raw vibration signal X of the rolling bearing under test. io (t);

[0305] The preprocessing module is used to process the original vibration signal X. io (t) is preprocessed to obtain the preprocessed vibration signal X. i (t);

[0306] The first sampling module is used to process the preprocessed vibration signal X. i (t) Perform the first overlap sampling to obtain the first sample signal x. m (t);

[0307] The second sampling module is used to process the preprocessed vibration signal X. i (t) Perform a second overlap sampling to obtain the second sample signal y. m (t);

[0308] The tensor set acquisition module is used to acquire tensor sets based on the first sample signal x. m (t) and the second sample signal y m (t) Obtain the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test respectively;

[0309] The diagnostic module is used to construct a fused dataset based on the one-dimensional tensor set and two-dimensional tensor set of the rolling bearing under test, and input it into a trained parallel convolutional neural network to identify the faults of the rolling bearing under test.

[0310] In another embodiment, this disclosure also provides a computer storage medium, comprising:

[0311] Memory is used to store multiple computer instructions;

[0312] A processor is used to execute computer instructions to implement the methods described in any of the preceding methods.

[0313] The above embodiments are only for illustrating the technical concept and features of this disclosure, and are intended to enable those skilled in the art to understand the content of this disclosure and implement it accordingly. They should not be construed as limiting the scope of protection of this disclosure. All equivalent changes or modifications made in accordance with the spirit and essence of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for diagnosing rolling bearing faults based on parallel convolutional neural networks, comprising the following steps: s100: Acquires the raw vibration signal X of the rolling bearing under test. io (t); S200: For the original vibration signal X io (t) is preprocessed to obtain the preprocessed vibration signal X. i (t); S300: For the pre-processed vibration signal X i (t) Perform the first overlap sampling to obtain the first sample signal x. m (t); S400: For the pre-processed vibration signal X i (t) Perform a second overlap sampling to obtain the second sample signal y. m (t); S5 00: Based on the first sample signal x m (t) and the second sample signal y m (t) Obtain the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test respectively; S600: Construct a fused dataset based on the one-dimensional tensor set and two-dimensional tensor set of the rolling bearing under test, and input it into a trained parallel convolutional neural network to identify the faults of the rolling bearing under test; in, In step S500, the step based on the first sample signal x m (t) Obtaining the one-dimensional tensor set of the rolling bearing to be tested includes the following steps: S501: For the first sample signal x m (f) Perform Hilbert transform and calculate the first sample signal x. m (f) envelope signal A m (f); S502: For the first sample signal x m (t) Perform integrated empirical mode decomposition to obtain the decomposed modal components u i (t), and based on the correlation coefficient α, the modal components u i (t) is reconstructed to obtain the reconstructed signal U. m (t); S503: For the first sample signal x m (t) Perform variational mode decomposition to obtain the decomposed modal components v i (t), and based on the correlation coefficient β, the modal components v i (t) is reconstructed to obtain the reconstructed signal V. m (t); S504: Envelope signal A m (t), reconstructed signal U i (t) and reconstructed signal V i (t) are respectively converted into a one-dimensional signal tensor A. t m (t), U t m (t) and V t m (t), and stack the signal tensors of the same category in the specified dimension dim=0 to obtain one-dimensional tensor sets T1, T2 and T3 respectively; In step S501, the first sample signal x is calculated using the following formula. m (f) Perform Hilbert transform: Among them, e m (t) represents the first sample signal x m (t) is the signal obtained after Hilbert transform, where f represents time, τ represents the integration variable, and m represents the first sample signal x. m The index of (t); In step S501, the first sample signal x is calculated in the following manner. m The envelope signal A of (t) m (t): Using the first sample signal x m (t) is the real part, and the signal e after Hilbert transform is used. m (f) represents the imaginary part, constructing the m-th sample signal x. m The analytic signal α of (f) m (t): α m (t)=x m (t)+I m (t) Take the analytic signal α m The amplitude of (t) is used as the envelope signal A. m (t), that is Where, x m (t) represents the first sample signal, e m (t) represents the first sample signal x m (f) The signal after Hilbert transform; In step S502, the first sample signal x is calculated using the following formula. m (t) Perform integrated empirical mode decomposition: Among them, u i (t) represents the modal components after integrated empirical mode decomposition, i represents the i-th modal component obtained by decomposition, n represents the decomposition number, and r represents the modal component obtained by decomposition. n (f) represents the residual; In step S502, the correlation coefficient α is calculated using the following formula: in, Represents the modal component u i The average value of (f), Represents the first sample signal x m The mean of (t), where std represents the standard deviation operation; The reconstructed signal U m (t) Modal components u with correlation coefficient α > 0.4 i (f) The result is obtained by adding them together, that is Where i represents i modal components with correlation coefficient α > 0.4; In step S503, a constrained variational model is constructed for the first sample signal x. m (f) Perform variational mode decomposition; The constrained variational model is expressed as follows: Among them, {v i {ω} represents the i-th intrinsic mode component after variational mode decomposition, I represents the number of decomposed signals, and {ω} i } represents the center frequency corresponding to the intrinsic modal component, δ(t) represents the Dirichlet function, * represents the convolution operation, and x m (t) represents the first sample signal, and π represents pi. The symbol represents differentiation, and j represents the imaginary unit; In step S503, the correlation coefficient β is calculated using the following formula: in, Represents the modal component v i The average value of (t), Represents the first sample signal x m The mean of (t), where std represents the standard deviation operation; The reconstructed signal V m (t) is derived from the intrinsic modal components v with a correlation coefficient β > 0.

4. i (t) are obtained by adding them together, that is Where i represents i intrinsic modal components with correlation coefficient β>0.4; This method effectively avoids noise interference in vibration signals by inputting one-dimensional and two-dimensional features extracted from the original vibration signal into a parallel convolutional neural network, thereby solving the problem of inaccurate fault diagnosis of vibration signals under interference.

2. The method according to claim 1, wherein, In step S200, the original vibration signal X is... io (t) Preprocessing includes the following steps: S201: Remove the original vibration signal X io The DC component in (t); S202: The original vibration signal X after removing the DC component. io (t) is normalized.

3. The method according to claim 2, wherein, In step S201, the original vibration signal X is removed using the following formula. io The DC component in (f): Among them, X i (f) represents the original vibration signal after removing the DC component, X io (f) represents the original vibration signal, X io (t) k Represents the original vibration signal X io The amplitude corresponding to the k-th sampling point in (t), 1en(X) io (f) represents the original vibration signal X. io (f) shows the number of sampling points, where io represents the original vibration signal X. io (f) is the data sequence number, where t represents the original vibration signal X. io (f) is a time series.

4. The method according to claim 1, wherein, In step S202, the original vibration signal X after removing the DC component is calculated using the following formula. i (t) is normalized: and Among them, MIN(X) i (t) represents the minimization process, σ represents the scaling factor, and X i (t) max X represents i The maximum value of X in (t), i (t) mmin X represents i The minimum value in (t).

5. The method according to claim 1, wherein, In step S300, the preprocessed vibration signal is subjected to the first overlapping sampling by using a sliding window.

6. The method according to claim 5, wherein, The position of the sliding window is calculated using the following formula: Where, N total This represents the pre-processed vibration signal X. i The number of sampling points N(t) total [(m-1)(L+step), mL+(m-1)step] represents all sampling points between the (m-1)(L+step)th sampling point and the mL+(m-1)stepth sampling point, where L represents the length of the sliding window, step represents the step size of the sliding window, and x m (t) represents the first sample signal, and m represents the first sample signal x. m The sequence number of (t).

7. The method according to claim 6, wherein, The preprocessed vibration signal X i The number of sampling points for (t) is calculated using the following formula: Where k represents the length coefficient, which is defaulted to k=2, and n represents the bearing speed. Z represents the sampling interval and Z represents a positive integer.

8. The method according to claim 1, wherein, In step S400, the normalized vibration signal is subjected to a second overlapping sampling by using a sliding window.

9. The method according to claim 1, wherein, In step S500, the step based on the second sample signal y m (t) Obtaining the two-dimensional tensor set of the rolling bearing under test includes the following steps: S5001: For the second sample signal y m (t) Perform continuous wavelet transform to obtain the corresponding wavelet coefficients, and plot the wavelet time-frequency diagram based on the wavelet coefficients; S5002: Draw a Markov transition field diagram; S5003: Draw the Gram corner field diagram; S5004: Convert the wavelet time-frequency diagram, Markov transition field diagram, and Gram angle field diagram into two-dimensional tensors respectively to obtain the two-dimensional tensor sets T4, T5, and T6 of the rolling bearing under test.

10. The method according to claim 9, wherein, Step S5001 includes the following steps: S50011: Using the Python wavelet analysis toolkit PyWavelets to analyze the second signal sample y m (t) Perform continuous wavelet transform to obtain wavelet coefficients, specifically: Where WT(a, τ) represents wavelet coefficients, a represents the scaling factor, τ represents the time shift factor, * represents the conjugate function, and Ψ represents the conjugate function. a，τ (t) represents the wavelet basis function, and t represents the time series; S50012: Draw a wavelet time-frequency diagram based on the time series t of the wavelet coefficients, the actual frequency series f, and the modulus abs(WT(a,τ)) of the wavelet coefficients.

11. The method according to claim 9, wherein, In step S5001, the pcolor function in the matplotlib plotting library is used to plot the wavelet time-frequency graph based on the wavelet time series t, the actual frequency series f, and the modulus abs(WT(a,τ)) of the wavelet coefficients.

12. The method according to claim 11, wherein, The actual frequency sequence f of the wavelet is calculated using the following formula: scale·f=F s ·wcf Where scale represents the wavelet scaling sequence, F s Represents the second sample signal y m The sampling frequency of (t), where wcf represents the center frequency of the wavelet.

13. The method according to claim 9, wherein, Step S5002 includes the following steps: S50021: Based on the time sequence of a first-order Markov process, the second signal sample y is... m (t) is visualized as a complex network, yielding the Markov matrix; S50022: Draw Markov transition field diagrams based on Markov matrices.

14. The method according to claim 9, wherein, Step S5003 includes the following steps: S50031: The second sample signal y in the Cartesian coordinate system m (t) is normalized to the interval [-1, 1] and then converted to polar coordinates. S50032: Based on the second sample signal y m (t) The normalized polar coordinate representation is used to generate the Gram angular field matrix G through trigonometric functions; S50033: Draw Gram angle field diagrams based on Gram angle field matrix G.

15. The method according to claim 14, wherein, In step S50032, the Gram angular field matrix G is represented as: in, Represents the second sample signal y m (t) Normalized time series points x n The polar coordinates.

16. The method according to claim 1, wherein, The parallel convolutional neural network includes a one-dimensional convolutional neural network and a two-dimensional convolutional neural network. The one-dimensional convolutional neural network includes a fully connected layer, a convolutional layer, a batch normalization layer, an activation function, and a pooling layer. The two-dimensional convolutional neural network adopts the InceptionV1 network.

17. The method according to claim 1, wherein, In step S600, the training of the parallel convolutional neural network includes the following steps: S601: Collect the health signal of the healthy bearing and the fault signal of the faulty bearing respectively, then classify the fault signals according to different fault locations and fault degrees, and combine the health signal and the classified fault signal to obtain the original signal set; S602: Obtain the one-dimensional tensor set and the two-dimensional tensor set of the original signal set; S603: Combine the one-dimensional tensor set and two-dimensional tensor set of the original signal set with the labels of the fused dataset to construct the fused dataset, and divide the fused dataset into training set, validation set and test set; S604: A minimum loss function is preset. The parallel convolutional neural network is trained using the training set. At the same time, the parallel convolutional neural network is validated using the validation set during training. Whenever the loss function value on the validation set is less than the preset minimum loss function value, and the loss function on the training set is still decreasing, the minimum loss function value is set to equal the current loss function value on the validation set and saved. The model parameters of the network at this time are overwritten and saved. When training is completed, the trained parallel neural network is obtained. S605: Test the trained parallel neural network using the test set. If the test passes, the trained parallel neural network is obtained; otherwise, the hyperparameters need to be adjusted and the parallel neural network needs to be retrained.

18. The method according to claim 17, wherein, In step S605, the classification performance of the model is evaluated using a confusion matrix and a t-distribution-random neighbor embedding to determine whether the parallel neural network test has passed.

19. The method of claim 17, wherein, In step S604, before training the parallel neural network using the training set, it is necessary to convert the one-dimensional signal tensor A in the training set into a single array. t m (t), U t m (t) and V t m (t) assign different weights ρ k The merging is then performed using the following formula to obtain the one-dimensional fusion tensor x. f (t): x f (t)=ρ1A t m (t)+ρ2U t m (t)+ρ3V t m (t) Among them, A t m (t) is the tensor of the envelope signal; U t m (t) is the tensor of the integrated empirical mode decomposition reconstructed signal; V t m (t) is the tensor of the signal reconstructed by variational mode decomposition.

20. The method according to claim 19, wherein, The weight ρk is calculated using the following formula: Where, ρ k kur represents the weights corresponding to the k-th type tensor. k Let be the kurtosis corresponding to the k-th type tensor; Let be the kurtosis of the three tensors.

21. An apparatus for performing the rolling bearing fault diagnosis method based on parallel convolutional neural networks as described in claim 1, comprising: The acquisition module is used to acquire the raw vibration signal X of the rolling bearing under test. io (t); The preprocessing module is used to process the original vibration signal X. io (t) is preprocessed to obtain the preprocessed vibration signal X. i (t); The first sampling module is used to process the preprocessed vibration signal X. i (t) Perform the first overlap sampling to obtain the first sample signal x. m (t); The second sampling module is used to process the preprocessed vibration signal X. i (t) Perform a second overlap sampling to obtain the second sample signal y. m (t); The tensor set acquisition module is used to acquire tensor sets based on the first sample signal x. m (t) and the second sample signal y m (t) Obtain the one-dimensional tensor set and the two-dimensional tensor set of the rolling bearing under test respectively; The diagnostic module is used to construct a fused dataset based on the one-dimensional tensor set and two-dimensional tensor set of the rolling bearing under test, and input it into a trained parallel convolutional neural network to identify the faults of the rolling bearing under test.

22. A computer storage medium, comprising: Memory is used to store multiple computer instructions; A processor for executing computer instructions to implement the method of any one of claims 1 to 20.

Images