A multi-convolution network fusion rolling bearing fault classification method

Abstract

The application belongs to the technical field of industrial process monitoring, and particularly relates to a rolling bearing fault classification method based on multi-convolution network fusion. The method comprises the following steps: step S1: collecting original fault data, extracting the running parameters and size parameters of the corresponding bearing, and calculating three kinds of fault characteristic frequencies by using the above parameters; step S2: using continuous wavelet transform on the original fault data to obtain four kinds of time-frequency diagrams; step S3: dividing the time-frequency diagram set obtained by performing continuous wavelet transform on all original data into a training set and a verification set; step S4: training a fault classification model by using the training set; and step S5: verifying the classification performance of the model by using the verification set. The application has the beneficial effects of improving the explainability of the overall network, improving the flexibility of the network, and being higher in classification precision than the traditional classification.

Description

Technical Field

[0001] This invention belongs to the field of industrial process monitoring technology, specifically relating to a rolling bearing fault classification method based on multi-convolutional network fusion. Background Technology

[0002] Rolling bearings play a significant role in modern industrial systems. Timely and accurate fault diagnosis of rolling bearing failures can not only improve the reliability of equipment use, ensure the smooth operation of the production process and the safety of personnel and equipment, but also provide valuable reference for bearing manufacturing.

[0003] In recent years, deep learning technology has developed rapidly, and deep learning-based fault diagnosis technology has been widely used and achieved good results. However, it also has its own shortcomings, such as weak interpretability and failure to effectively utilize known physical knowledge.

[0004] Existing methods for integrating physics knowledge into convolutional neural networks mostly rely on the Hilbert-Huang transform to obtain one-dimensional feature data, and then use a one-dimensional convolutional neural network for fault diagnosis. However, these methods do not effectively utilize the powerful image recognition capabilities of two-dimensional convolutional neural networks, and the classification accuracy needs improvement; at the same time, the implementation process is relatively complex and cannot well meet the needs of practical industrial applications. Summary of the Invention

[0005] To address the shortcomings and improvement needs of existing technologies, this invention provides a rolling bearing fault classification method based on multi-convolutional network fusion. Its aim is to incorporate mechanistic knowledge into existing convolutional neural networks to improve the accuracy of fault diagnosis and the interpretability of diagnostic results. This model has clear network logic, is easy to deploy in practice, and achieves high classification accuracy, thus better meeting the needs of practical industrial applications.

[0006] The technical solution of the present invention is as follows: A method for classifying rolling bearing faults by fusing multiple convolutional networks, comprising the following steps:

[0007] Step S1: Collect raw fault data. For the collected raw fault data, extract the corresponding bearing's operating parameters and dimensional parameters, and use the above parameters to calculate three fault characteristic frequencies.

[0008] Step S2: Using continuous wavelet transform, obtain four time-frequency diagrams from the original fault data;

[0009] Step S3: Divide the time-frequency image set obtained after performing continuous wavelet transform on all the original data into a training set and a validation set;

[0010] Step S4: Train the fault classification model using the training set;

[0011] Step S5: Validate the classification performance of the model using a validation set.

[0012] In step S1, for rolling bearings, the fault characteristic frequencies when the fault point is located on each component are as follows:

[0013] The characteristic frequency of outer raceway failure is:

[0014]

[0015] The characteristic frequency of internal raceway failure is:

[0016]

[0017] The characteristic frequencies of ball bearing failures are:

[0018]

[0019] In the formula, f r Let z be the rotational frequency of the shaft, z be the number of balls, and d be the rotational frequency of the shaft. b D is the diameter of the ball bearing. m The pitch circle diameter of the rolling bearing. This refers to the contact angle of the rolling bearing.

[0020] In step S2, the four time-frequency diagrams include the main time-frequency diagram for all frequencies, the sub-time-frequency diagram corresponding to the outer ring fault characteristic frequency, the sub-time-frequency diagram corresponding to the inner ring fault frequency, and the sub-time-frequency diagram corresponding to the ball fault frequency.

[0021] In step S2, the method for obtaining the three sub-time-frequency diagrams is as follows: expand the frequency by 5% to the left and right of the three fault characteristic frequencies to obtain three frequency bands. Use these three frequency bands as the center frequency of the continuous wavelet transform mother function to obtain three long strip-shaped time-frequency diagrams.

[0022] In step S3, both the training set and the validation set contain the above four types of time-frequency diagrams.

[0023] In step S4, the fault classification model consists of four relatively independent two-dimensional convolutional neural networks, each with an input corresponding to one of the four time-frequency maps. The macroscopic structure is as follows: a series of two-dimensional convolutional layers, pooling layers, normalization layers, and activation layers are grouped together and repeated 3 times. Then, a series of fully connected layers, normalization layers, and activation layers are grouped together and repeated 2 times. Finally, a SoftMax classification layer is added.

[0024] In step S4, the network corresponding to the main time-frequency map is the main network, and the networks corresponding to the three sub-time-frequency maps are the sub-networks. The main network performs four classifications: outer ring fault, inner ring fault, ball fault, and health status. The sub-networks perform two classifications, and the classifications of the three sub-networks are: outer ring fault / not outer ring fault, inner ring fault / not inner ring fault, and ball fault / not ball fault. The output of the main network is a 4-dimensional one-hot vector, and the output of the sub-networks is a 2-dimensional one-hot vector.

[0025] In step S5,

[0026] During verification, a voting mechanism is set up to merge the results of the four network judgments.

[0027] Its fusion formula is:

[0028] y = α o y o +α i y i +α b y b +βy *

[0029] In the formula, y is a 4-dimensional vector representing the final classification result; o y i and y b α represents the classification results of the networks corresponding to the outer ring, inner ring, and ball, respectively, and its value range is {1, -1}; o α i and α b Both are vectors, with values ​​[1, 0, 0, 0]. T [0, 1, 0, 0] T [0, 0, 1, 0] T ;y * The classification result of the network corresponding to the main time-frequency plot is a 4-dimensional vector; β is y * The coefficient represents the voting power of the main network and can be set independently.

[0030] In step S5, the outputs of the three sub-networks are 2D one-hot vectors, and their corresponding values ​​y in the fusion formula are... o y i and y b Since these are all specific numbers, there is a transformation step before fusion: when the sub-network output is [1, 0] T When [condition], it is converted to 1; the subnetwork output is [0, 1]. T When the time is equal to -1, it is converted to -1.

[0031] The beneficial effects of the present invention are as follows: (1) The outputs of the three sub-networks proposed in the present invention are closely related to physical knowledge, which improves the interpretability of the overall network; (2) In the network result fusion method proposed in the present invention, the voting weight of the main network can be set independently, which improves the flexibility of the network; (3) Because the three additional sub-networks can play a positive guiding role in fault classification, the network classification accuracy proposed in the present invention is higher than that of the traditional classification accuracy. Attached Figure Description

[0032] Figure 1 A flowchart of a rolling bearing fault classification method fused by multiple convolutional networks provided by the present invention;

[0033] Figure 2 This is a sub-time frequency diagram. Detailed Implementation

[0034] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] To achieve the above objectives, according to a first aspect of the present invention, a strategy for vibration signal preprocessing based on bearing fault characteristic frequencies is provided:

[0036] For rolling bearings, there are generally three important fault characteristic frequencies: the outer raceway fault characteristic frequency, the inner raceway fault characteristic frequency, and the ball fault characteristic frequency. The wavelet mother function of the continuous wavelet transform also has a center frequency, and the range of this center frequency determines the frequency range of the resulting time-frequency plot.

[0037] This invention, while using continuous wavelet transform to obtain time-frequency diagrams, also uses three fault characteristic frequencies as the center and appropriately extends them to both sides of the frequency value, respectively, as the center frequency band of the wavelet mother function, to obtain three corresponding strip-shaped time-frequency diagrams.

[0038] According to a second aspect of the present invention, a multi-convolutional neural network structure based on result fusion is provided:

[0039] This invention sets up four relatively independent convolutional neural networks to process four different time-frequency maps: a main time-frequency map covering all frequencies, a time-frequency map corresponding to the outer ring fault characteristic frequencies, a time-frequency map corresponding to the inner ring fault frequencies, and a time-frequency map corresponding to the ball bearing fault frequencies.

[0040] The four convolutional neural networks are trained independently; during testing, a voting mechanism is set up to merge the results of the four networks.

[0041] like Figure 1 As shown, a rolling bearing fault classification method based on multi-convolutional network fusion includes the following steps:

[0042] Step S1: Collect raw fault data. For the collected raw fault data, extract the corresponding bearing's operating parameters and dimensional parameters, and use the above parameters to calculate three fault characteristic frequencies.

[0043] For rolling bearings, the characteristic frequencies of failures when the failure point is located on various components are as follows:

[0044] The characteristic frequency of outer raceway failure is:

[0045]

[0046] The characteristic frequency of internal raceway failure is:

[0047]

[0048] The characteristic frequencies of ball bearing failures are:

[0049]

[0050] In the formula, f r Let z be the rotational frequency of the shaft (Hz), z be the number of balls, and d be the rotational frequency of the shaft. b D represents the diameter of the ball bearing (mm). m The pitch circle diameter of the rolling bearing (mm) The contact angle of the rolling bearing is (°).

[0051] For the same bearing and the same operating environment, its fault characteristic frequency is fixed, so a set of data consists of three fixed fault characteristic frequencies.

[0052] Step S2: Using continuous wavelet transform, obtain four time-frequency maps for the original fault data: a main time-frequency map covering all frequencies, a sub-time-frequency map corresponding to the characteristic frequencies of the outer ring fault, a sub-time-frequency map corresponding to the frequencies of the inner ring fault, and a sub-time-frequency map corresponding to the frequencies of the ball fault.

[0053] The method for obtaining the three sub-time-frequency diagrams is as follows: expand the frequency by 5% to the left and right of the three fault characteristic frequencies to obtain three frequency bands. Use these three frequency bands as the center frequency of the continuous wavelet transform mother function to obtain three long strip-shaped time-frequency diagrams.

[0054] The method for obtaining the sub-time-frequency diagram corresponding to the outer ring fault is as follows: Extend the characteristic frequency of the outer ring fault by 30% to the left and right of that frequency to obtain a frequency band. Use this frequency band as the frequency range for continuous wavelet transform to obtain the sub-time-frequency diagram corresponding to the outer ring fault. Figure 2 As shown in (a).

[0055] The method for obtaining the sub-time-frequency diagram corresponding to the inner ring fault is as follows: Extend the frequency band 30% to the left and right of the characteristic frequency of the inner ring fault, obtaining a frequency band. Use this frequency band as the frequency range for continuous wavelet transform to obtain the sub-time-frequency diagram corresponding to the inner ring fault. Figure 2 As shown in (b).

[0056] The method for obtaining the sub-time-frequency diagram corresponding to a ball fault is as follows: Extend the characteristic frequency of the ball fault by 30% to the left and right of that frequency to obtain a frequency band. Use this frequency band as the frequency range for continuous wavelet transform to obtain the corresponding sub-time-frequency diagram of the ball fault. Figure 2 As shown in (c).

[0057] The main time-frequency diagram is obtained as follows: The lowest frequency is defined as 0.5 times the lowest of the three fault characteristic frequencies, and the highest frequency is defined as 1.5 times the highest of the three fault characteristic frequencies, thus obtaining a frequency band. This frequency band is then used as the frequency range for continuous wavelet transform, thereby obtaining the main time-frequency diagram. Figure 2 As shown in (d).

[0058] Step S3: After performing continuous wavelet transform on all the original data, the resulting time-frequency map set is divided into a training set and a validation set. Both the training set and the validation set contain the above four types of time-frequency maps.

[0059] Step S4: Train the fault classification model using the training set.

[0060] The fault classification model consists of four relatively independent two-dimensional convolutional neural networks, each with its input corresponding to one of four time-frequency maps. Because the size of each time-frequency map is different, the specific sizes of the four networks are also different. However, their macroscopic structure is similar: a series of two-dimensional convolutional layers, pooling layers, normalization layers, and activation layers are grouped together and repeated three times; then a series of fully connected layers, normalization layers, and activation layers are grouped together and repeated two times; finally, a SoftMax classification layer is added.

[0061] Let the network corresponding to the main time-frequency graph be the main network, and the networks corresponding to the three sub-time-frequency graphs be the sub-networks. The main network performs four classifications: outer ring fault, inner ring fault, ball bearing fault, and health status; the sub-networks perform two classifications, namely: outer ring fault / not outer ring fault, inner ring fault / not inner ring fault, and ball bearing fault / not ball bearing fault. Therefore, the main network outputs a 4-dimensional one-hot vector, and the sub-networks output a 2-dimensional one-hot vector.

[0062] Step S5: Validate the classification performance of the model using a validation set.

[0063] During verification, a voting mechanism is set up to merge the results of the four network judgments.

[0064] Its fusion formula is:

[0065] y = α o y o +α i y i +α b y b +βy *

[0066] In the formula, y is a 4-dimensional vector representing the final classification result; o y i and y b α represents the classification results of the networks corresponding to the outer ring, inner ring, and ball, respectively, and its value range is {1, -1}; o α i and α b Both are vectors, with values ​​[1, 0, 0, 0]. T [0, 1, 0, 0] T [0, 0, 1, 0] T ;y * The classification result of the network corresponding to the main time-frequency plot is a 4-dimensional vector; β is y * The coefficient represents the voting power of the main network and can be set independently.

[0067] It should be noted that the outputs of the three sub-networks are 2D one-hot vectors, and their corresponding values ​​y in the fusion formula are... o y i and y b Since these are all specific numbers, there is a transformation step before fusion: when the sub-network output is [1, 0] T When [condition], it is converted to 1; the subnetwork output is [0, 1]. T When the time is equal to -1, it is converted to -1.

Claims

1. A method for classifying rolling bearing faults using multi-convolutional network fusion, characterized in that, Includes the following steps: Step S1: Collect raw fault data. For the collected raw fault data, extract the corresponding bearing's operating parameters and dimensional parameters, and use the above parameters to calculate three fault characteristic frequencies. In step S1, for rolling bearings, the fault characteristic frequencies when the fault point is located on each component are as follows: The characteristic frequency of outer raceway failure is: ， The characteristic frequency of internal raceway failure is: ， The characteristic frequencies of ball bearing failures are: ， In the formula, Let z be the rotational frequency of the shaft, and z be the number of balls. The diameter of the ball bearing is... The pitch circle diameter of the rolling bearing. This refers to the contact angle of a rolling bearing. Step S2: Using continuous wavelet transform, obtain four time-frequency diagrams from the original fault data; In step S2, the four types of time-frequency diagrams include the main time-frequency diagram for all frequencies, the sub-time-frequency diagram corresponding to the outer ring fault characteristic frequency, the sub-time-frequency diagram corresponding to the inner ring fault frequency, and the sub-time-frequency diagram corresponding to the ball fault frequency. Step S3: Divide the time-frequency image set obtained after performing continuous wavelet transform on all the original data into a training set and a validation set; Step S4: Train the fault classification model using the training set; In step S4, the fault classification model consists of four relatively independent two-dimensional convolutional neural networks, each with an input corresponding to one of the four time-frequency maps. The macroscopic structure is as follows: a series of two-dimensional convolutional layers, pooling layers, normalization layers, and activation layers are grouped together and repeated 3 times. Then, a series of fully connected layers, normalization layers, and activation layers are grouped together and repeated 2 times. Finally, a SoftMax classification layer is added. In step S4, the network corresponding to the main time-frequency map is the main network, and the networks corresponding to the three sub-time-frequency maps are the sub-networks. The main network performs four classifications: outer ring fault, inner ring fault, ball fault, and health status. The sub-networks perform two classifications, and the classifications of the three sub-networks are: outer ring fault / not outer ring fault, inner ring fault / not inner ring fault, and ball fault / not ball fault. The main network outputs a 4-dimensional one-hot vector, and the sub-networks output a 2-dimensional one-hot vector. Step S5: Validate the classification performance of the model using a validation set.

2. The rolling bearing fault classification method based on multi-convolutional network fusion as described in claim 1, characterized in that: In step S2, the method for obtaining the three sub-time-frequency diagrams is as follows: expand the frequency by 5% to the left and right of the three fault characteristic frequencies to obtain three frequency bands. Use these three frequency bands as the center frequency of the continuous wavelet transform mother function to obtain three long strip-shaped time-frequency diagrams.

3. The rolling bearing fault classification method based on multi-convolutional network fusion as described in claim 1, characterized in that: In step S3, both the training set and the validation set contain the above four types of time-frequency diagrams.

4. The rolling bearing fault classification method based on multi-convolutional network fusion as described in claim 1, characterized in that: In step S5, During verification, a voting mechanism is set up to merge the results of the four network judgments. Its fusion formula is: ， In the formula, It is a 4-dimensional vector representing the final classification result; , as well as These represent the classification results of the networks corresponding to the outer ring, inner ring, and ball, respectively, and their values ​​range from {1, -1}. , as well as Both are vectors, and their values ​​are respectively ; The classification result of the network corresponding to the main time-frequency graph is a 4-dimensional vector; yes The coefficient represents the voting power of the main network and can be set independently.

5. The rolling bearing fault classification method based on multi-convolutional network fusion as described in claim 4, characterized in that: In step S5, the outputs of the three sub-networks are 2D one-hot vectors, and their corresponding values ​​in the fusion formula are... , as well as Since these are all specific numbers, there is a transformation step before fusion: when the sub-network output is When, it is converted to 1; the sub-network output is When the time is equal to -1, it is converted to -1.

Images