A tunnel boring machine main bearing fault simulation test bench and fault diagnosis method

Abstract

This invention discloses a fault simulation test bench and fault diagnosis method for a tunnel boring machine main bearing. The test bench includes a frame and a main bearing rotatably mounted on the frame. The main bearing has a first outer ring and a second outer ring at its front and rear ends, respectively. A groove is provided on the side of the second outer ring facing the main thrust roller, and a fault simulation plate is held in the groove. The fault simulation plate and the second outer ring are interference-fitted. A fault detection module is installed on the outer side of the second outer ring. A screw hole is provided at the rear end of the fault simulation plate, and a bolt passes through the second outer ring and extends into the screw hole to pull the fault simulation plate into the groove. Compared with existing technologies, this invention, by setting a detachably connected fault simulation plate and second outer ring, can greatly reduce the difficulty of fault implantation, thus reducing manufacturing costs. It also allows the simulation test bench to more flexibly control the form, degree, and number of faults.

Description

Technical Field

[0001] This invention relates to the field of bearing diagnostic technology, and in particular to a fault simulation test bench and fault diagnosis method for the main bearing of a tunnel boring machine. Background Technology

[0002] Currently, most bearing failure tests focus on small, high-speed bearings, with very few test benches specifically designed for large, low-speed bearings like the main bearings of tunnel boring machines. In tests on small, high-speed bearings, failure simulations typically involve replacing the entire outer ring and introducing different types of faults into multiple outer rings. However, for large bearings, considering the significant manufacturing difficulty and cost, simulating multiple faults would require manufacturing outer rings corresponding to the number of fault types, which is extremely challenging. Summary of the Invention

[0003] This invention provides a fault simulation test bench and fault diagnosis method for the main bearing of a tunnel boring machine, which solves the problem that existing fault test benches are unable to simulate various faults of large-size bearings.

[0004] This invention provides a test bench for simulating main bearing failure in a tunnel boring machine, comprising a frame and a main bearing rotatably mounted on the frame. The main bearing has a first outer ring and a second outer ring at its front and rear ends, respectively. A groove is provided on the side of the second outer ring facing the main thrust roller, and a failure simulation plate is held in the groove. The failure simulation plate and the second outer ring are interference-fitted. A failure detection module is installed on the outer side of the second outer ring at a position corresponding to the failure simulation plate. A screw hole is provided at the rear end of the failure simulation plate, and a bolt passes through the second outer ring and extends into the screw hole to pull the failure simulation plate into the groove. The frame is equipped with a loading device and an oil circulation device. The oil circulation device provides lubricant to the main bearing, and the loading device drives the main bearing to rotate and provides axial thrust to the main bearing.

[0005] Preferably, the second outer ring and the first outer ring are fixed together by bolts.

[0006] Preferably, a notch is provided between the fault simulation board and the bottom of the card slot.

[0007] Preferably, there are three fault simulation boards, which are distributed at the three o'clock, six o'clock and nine o'clock positions on the second outer ring.

[0008] Preferably, the loading device includes: a servo motor, a drive shaft, a gear, and an internal gear ring. The internal gear ring is fixed to the inner side of the main bearing, the gear meshes with the internal gear ring, the gear is fixed on the drive shaft, and the drive shaft is fixedly connected to the output shaft of the servo motor.

[0009] The present invention also provides a fault diagnosis method for the main bearing of a tunnel boring machine, including the aforementioned tunnel boring machine main bearing fault simulation test bench, and further including the following steps:

[0010] S1: Vibration data x of the main bearing obtained by the fault detection module. n (t), each fault detection module provides one channel of vibration data x n (t), and then adjust the data segment length in the data preprocessing process according to the fault characteristic frequency, and then normalize the data of each channel to obtain data J; where x n (t) represents the vibration data of the nth channel at time t;

[0011] S2: Data J is initially processed using dilated wide-kernel convolution to obtain data Y. Data Y is then batch normalized to obtain data G. Finally, a nonlinear component is introduced into data G and maximum pooling is performed to obtain the channel features U of the channels. n ;

[0012] S3: Transfer channel features U n The feature vector U is obtained by element-wise summation. Then, the feature vector U is processed through a global average pooling layer (GAP), a fully connected layer (FC), a batch normalization (BN) layer, and a ReLU activation function to obtain the feature Z. Finally, the feature Z is input into the fully connected layer to obtain the attention weights W for each channel. j The sum of the attention weights for each channel is 1, that is... Where m = 3;

[0013] S4: Calculate the attention weights W j The feature vector S is assigned to obtain the weighted feature vector V. The weighted feature vector V is then fused, and finally the fused feature vector is input into a fully connected circuit to achieve the mapping process of fault type. The feature vector V is calculated by the following equation (3):

[0014]

[0015] Among them, W j Let be the attention weight for the j-th channel, and let S be the feature vector S obtained after inputting data Y into a series of convolutional layers. j Let be the feature value of the j-th channel.

[0016] Preferably, step S1 specifically includes the following steps:

[0017] S11, based on prior knowledge, solve for the fault characteristic frequency of the corresponding region, and combine it with the sampling frequency f. s The required data segment length l is calculated using the following formula (4):

[0018]

[0019] Where f is the fault characteristic frequency, f s Where N is the sampling frequency, and N is the number of periods within the data segment, with N ranging from 5 to 10.

[0020] Calculate the fault characteristic frequency f according to the table below (1):

[0021] Table 1. Formulas for calculating fault characteristic frequencies at various locations.

[0022]

[0023] Among them, D m Main raceway roller pitch circle diameter, D a To assist in pushing the pitch circle diameter of the raceway rollers, D r This is the radial raceway roller pitch circle diameter, d m The main driving roller diameter, d a To adjust the diameter of the auxiliary roller, d r f is the radial roller diameter. r Let n be the rotation frequency of the inner ring. m The number of main roller pairs, n a To determine the number of auxiliary rollers, n r This refers to the number of radial rollers;

[0024] S12, using the sliding window method, segment the time-domain data according to the data segment length l obtained above:

[0025]

[0026] Where L is the total length of the data, stride is the step size, and num represents the number of data segments after division;

[0027] S13, and then the data of each channel are normalized using formula (6):

[0028]

[0029] in This represents the k-th data segment in the n-th channel, which is data J. This represents the k-th data segment of the n-th channel, where k∈[0,num]. This represents the maximum value in the corresponding data segment. This represents the minimum value in the corresponding data segment.

[0030] Preferably, step S2 specifically includes the following steps:

[0031] S21, First, dilated wide kernel convolution is used to perform preliminary processing on the data J. The calculation formula (7) of dilated wide kernel convolution is as follows:

[0032]

[0033] Where Y n [i] is the output at position i, which is the data Y; W[m] is the weight in the convolution kernel, corresponding to position m of the kernel; d is the dilation rate, which determines the spacing between elements of the convolution kernel; M is half the size of the convolution kernel. For data J, It is the [i+d·m]th data point in the corresponding data segment;

[0034] S22, for the output result Y n [i] Perform batch normalization, the formula of which is shown below:

[0035] Calculate the mean μ of the entire batch. n :

[0036] Calculate the variance of the entire batch.

[0037] Standardization:

[0038] Scaling and translation (reparameterization):

[0039] Where ∈ is a very small number to prevent the denominator from being zero; γ and β are learnable parameters that allow the network to learn to recover any potentially useful feature distributions; x n [i] represents the data point of the nth channel. Represents the normalized data points, y n [i] represents the output after batch normalization, which is data G;

[0040] S23, introduce nonlinear components into the data and perform maximum pooling using formula (8):

[0041] U n =maxpool(ReLU(X)) (8)

[0042] Among them, U n Then it is the feature value extracted from the nth channel, and ReLU(X) represents the ReLU function processing of the data G.

[0043] Preferably, step S3 specifically includes the following steps:

[0044] S31, the feature values ​​obtained from each channel are summed element by element using formula (9):

[0045] U = U1 + U2 + U3 (9)

[0046] Among them, the feature vector U integrates information from multiple channels;

[0047] S32, U is processed by a global average pooling layer (GAP), a fully connected layer (FC), a batch normalization (BN) layer, and a ReLU activation function to obtain feature Z. The mathematical model of Z can be further expressed as:

[0048] Z=δ(ψ(U gap (10)

[0049] Among them, U gap ψ represents the result of U after global pooling, ψ represents the batch normalization process (BN), and δ represents the activation function ReLU.

[0050] The data is then fed into a fully connected layer for further compression of the feature dimensions.

[0051] Z = Dense(C / r, C) (11)

[0052] Where Dense represents the fully connected process, r is the feature compression ratio, and C is the number of input features, with C typically set to r = 0.5;

[0053] S33, the compressed feature Z is input again into the three fully connected layers to generate three sets of attention weights, which are then normalized by SoftMax. The specific process is as follows:

[0054]

[0055]

[0056] Among them, A s B s C s For the s-th row of the feature matrix, the a generated from this part s b s c s The attention weights for the three channels are then obtained. Formulas (12), (13), and (14) satisfy the following relationship: a s +b s +c s =1.

[0057] Preferably, step S4 specifically includes the following steps:

[0058] S41, by convolving the feature Y obtained in step two with a wide-kernel hole convolution... n [i] Input a series of convolutional layers to further extract deeper features:

[0059] S i =Convs(16,128) (15)

[0060] The feature channels obtained from the dilated wide kernel convolution are 16. After a series of convolutions, the number of channels becomes 128. i Let S be the eigenvector;

[0061] S42, convolve the feature vector S obtained in step S41 i Multiplying the weights obtained in step S3 by the attention weights, we get the weighted feature vector V: V = a s ·S1+b s ·S2+c s ·S3.

[0062] Compared with existing technologies, this invention, by setting up a detachable fault simulation board and a second outer ring, can greatly reduce the difficulty of fault implantation, thereby reducing manufacturing costs. It also allows the simulation test bench to more flexibly control the form, severity, and number of faults. Furthermore, this invention first adds and fuses the data from the three channels, then generates three attention weights through a fully connected layer corresponding to the number of channels. During training, the fully connected layer subtracts weights from channels with low correlation to faults, making the fault diagnosis method focus more on channels with high correlation to faults, thus mitigating the risk of overfitting caused by information redundancy in multi-channel fault diagnosis. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0064] Figure 1 This is a schematic diagram of the structure of the present invention;

[0065] Figure 2 This is a schematic diagram of the structure of the tank plate of the present invention for removing the oil tank;

[0066] Figure 3 This is a partial structural schematic diagram of the present invention;

[0067] Figure 4 This is a schematic diagram of the structure of the second outer ring of the present invention;

[0068] Figure 5 This is a schematic diagram of the fault simulation board of the present invention;

[0069] Figure 6 This is a schematic diagram of the main bearing of the present invention.

[0070] Figure label:

[0071] 1. Frame, 2. Main bearing, 3. First outer ring, 4. Second outer ring, 5. Fault simulation board, 6. Loading device, 7. Oil circulation device, 8. Fault detection module, 41. Slot, 51. Screw hole, 52. Notch, 61. Servo motor, 62. Drive shaft, 63. Gear, 64. Internal gear ring, 65. Electric push rod, 66. Pressure sensor, 67. Roller, 68. Force ring, 69. Linear bearing, 651. Telescopic rod, 71. Oil outlet pipe. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0073] See attached document Figure 1 Appendix Figure 3 and attached Figure 4This embodiment provides a fault simulation test bench for a tunnel boring machine main bearing 2, including a frame 1 and a main bearing 2 rotatably mounted on the frame 1. The front and rear ends of the main bearing 2 are respectively provided with a first outer ring 3 and a second outer ring 4. A groove 41 is provided on the side of the second outer ring 4 facing the main push roller, and a fault simulation plate 5 is held in the groove 41. The fault simulation plate 5 is used to simulate fault defects; for example, cracks are engraved on the fault simulation plate 5, and vibration faults occur when the roller slides over it. The fault simulation plate 5 and the second outer ring 4 are interference-fitted. A fault detection module 8 is installed on the outer side of the second outer ring 4 at a position corresponding to the fault simulation plate 5. One embodiment of the fault detection module 8 is an acceleration sensor, which monitors acceleration signals to simulate the vibration signals of the main bearing 2. The fault simulation plate 5 has a screw hole 51 at its rear end. A bolt passes through the second outer ring 4 and extends into the screw hole 51 to pull the fault simulation plate 5 into the slot 41. This setting utilizes the preload of the bolt to pull the fault simulation plate 5 towards the second outer ring 4, making the fault simulation plate 5 tightly adhere to the second outer ring 4. The frame 1 is equipped with a loading device 6 and an oil circulation device 7. The oil circulation device 7 is used to provide lubricant to the main bearing 2, and the loading device 6 is used to drive the main bearing 2 to rotate and provide axial thrust to the main bearing 2. By setting a detachably connected fault simulation plate 5 and second outer ring 4, this invention can greatly reduce the difficulty of fault implantation, which is beneficial to reducing manufacturing costs. At the same time, it also allows the simulation test bench to more flexibly control the form, degree, and number of faults. The key point of fault simulation is the characterization of fault impact. This invention uses the following methods to reduce the impact attenuation caused by the detachable connection between the fault simulation plate 5 and the second outer ring 4. First, the fault simulation plate 5 is inserted into the second outer ring 4 by an interference fit. Specifically, liquid nitrogen is used for cryogenic assembly when installing the fault simulation plate 5. Second, based on the interference fit assembly, a preload is provided by bolts to increase the fit stiffness of the fault simulation plate 5 and the second outer ring 4.

[0074] One installation method for the second outer ring 4 and the first outer ring 3: The second outer ring 4 and the first outer ring 3 are fixed together by bolts. This structural design facilitates the disassembly of the second outer ring 4, thereby making it easier to install and replace the fault simulation board 5.

[0075] As another embodiment of the present invention: refer to the appendix Figure 5 A notch 52 is provided between the fault simulation board 5 and the bottom of the slot 41. Specifically, the notch 52 is located on the fault simulation board 5. With this structural design, during disassembly, it is only necessary to use a flathead screwdriver to pry at the notch 52 of the fault simulation board 5 to remove the fault simulation board 5 from the second outer ring 4.

[0076] As another embodiment of the present invention: refer to the appendix Figure 2There are three fault simulation boards 5, distributed at the three o'clock, six o'clock, and nine o'clock positions on the second outer ring 4. Correspondingly, three fault detection modules 8 are also distributed at the three o'clock, six o'clock, and nine o'clock positions on the second outer ring 4. The positions on the second outer ring 4 are divided according to time, with the three fault simulation boards 5 distributed at these three positions. This is because, due to the influence of axial force, the main propulsion section raceway (the raceway on the second outer ring 4) and rollers are under heavy load, resulting in high contact stress. Furthermore, the outer ring is fixed, and the load point is singular, making contact fatigue failure more likely compared to the radial raceway and auxiliary propulsion raceway (the raceway on the first outer ring 3). Secondly, the overturning force generated by uneven soil conditions and shield rotation causes significant peak stress in the three o'clock, six o'clock, and nine o'clock positions of the main raceway (second outer ring 4), thus increasing the probability of crushing, cracking, pitting, and other damage. In order to make the selected points as representative as possible, the fault implantation points were set at the 3 o'clock, 6 o'clock and 9 o'clock positions on the second outer ring 4.

[0077] As another embodiment of the present invention: the second outer ring 4 is located on the side of the main bearing 2 away from the loading device 6, and this structural design facilitates the disassembly of the second outer ring 4.

[0078] One embodiment of the loading device 6: The loading device 6 includes a servo motor 61, a drive shaft 62, a gear 63, and an internal gear ring 64. The internal gear ring 64 is fixed to the inner side of the main bearing 2. The gear 63 meshes with the internal gear ring 64 and is fixed to the drive shaft 62. The drive shaft 62 is fixedly connected to the output shaft of the servo motor 61. The servo motor 61 drives the drive shaft 62 to rotate, thereby driving the gear 63 to rotate. The rotation of the gear 63 drives the internal gear ring 64 to rotate, thereby driving the main bearing 2 to rotate.

[0079] As another embodiment of the present invention, the loading device 6 further includes an electric push rod 65, a pressure sensor 66, a roller 67, and a force ring 68. The force ring 68 is fixed to the front end of the main bearing 2 and is located inside the first outer ring 3. The roller 67 rolls along the force ring 68 and is rotatably mounted on the telescopic rod 651 of the electric push rod 65. The telescopic rod 651 is mounted on the frame 1 through a linear bearing 69. This structural design can push the roller 67 more smoothly, so that it applies force to the force ring 68, thereby applying an axial load to the main bearing 2.

[0080] In another embodiment of the present invention, the oil circulation device 7 includes an oil pump, the oil pump's outlet pipe 71 being connected to the first outer ring 3 and the second outer ring 4, and the oil pump's inlet pipe extending into the oil tank, where the main bearing 2 is located. The outlet pipe 71 supplies lubricating oil to the main bearing 2, and excess lubricating oil returns from the main bearing 2 to the oil tank. This structural design ensures that the main bearing 2 always has sufficient lubricating oil.

[0081] This invention utilizes an interference fit to mount the fault simulation plate 5 onto the second outer ring 4, enabling the simulation of multiple fault types, fault locations, and even multiple combined fault types, greatly increasing the freedom of fault simulation. Users can arbitrarily combine multiple fault types as needed. Furthermore, the main push roller is in contact with the fault simulation plate 5, and vibration is transmitted from the fault simulation plate 5 to the second outer ring 4, and then to the fault detection module 8. This invention, by applying bolt preload and utilizing the interference fit, significantly improves the fit stiffness, thereby largely eliminating vibration attenuation caused by the detachable design.

[0082] The present invention also provides a fault diagnosis method for the main bearing of a tunnel boring machine, including the aforementioned tunnel boring machine main bearing fault simulation test bench, and further including the following steps:

[0083] S1: Vibration data x of the main bearing is obtained from the fault detection module 8. n (t), each fault detection module 8 provides one channel of vibration data x n (t), and then adjust the data segment length in the data preprocessing process according to the fault characteristic frequency, and then process the data x of each channel. n (t) is normalized to obtain intermediate data J; x n (t) represents the vibration data of the nth channel at time t. In this invention, there are three fault detection modules 8, i.e., three channels. Since the shield tunnel main bearing is a large, low-speed, heavy-duty bearing, its rotational speed is usually less than 5 rpm, and its rotational monitoring data often suffers from large data volume and long monitoring time. Therefore, it is necessary to first calculate the fault characteristic frequency generated in the fault area based on prior knowledge, and then adjust the data segment length in the data preprocessing process according to the fault characteristic frequency so that a single data sample can contain a sufficient number of fault cycles.

[0084] S2: Initial processing of intermediate data J using dilated wide-kernel convolution yields data Y. Batch normalization of data Y yields data G. Then, nonlinear components are introduced into data G, and maximum pooling is performed to obtain the channel features U of the channels. n Because the main bearing of the tunnel boring machine is a low-speed, heavy-duty bearing, its rotational speed is extremely low, typically 2–5 rpm. This extremely low rotational speed causes fault data to be mainly concentrated in the low-frequency range. Therefore, the intermediate data J is processed using a wide-kernel convolution with holes to further improve its adaptability to the low-frequency range.

[0085] S3: Transfer channel features U nThe feature vector U is obtained by element-wise summation. Then, the feature vector U is processed through a global average pooling layer (GAP), a fully connected layer (FC), a batch normalization (BN) layer, and a ReLU activation function to obtain the feature Z. Finally, the feature Z is input into the fully connected layer to obtain the attention weights W for each channel. j The sum of the attention weights for each channel is 1, that is... Where m = 3;

[0086] Inspired by the attention mechanism, this process applies the idea of ​​learning weights to weight features to the features of different channels. Before performing classification learning based on the extracted features, a set of attention weights is learned using a fully connected layer. Then, the weights are transformed into a soft attention vector using the SoftMax function. By multiplying this attention vector with the features, the process achieves a gain for channels with high fault correlation and a loss for channels with low fault correlation.

[0087] S4: Calculate the attention weights W j The feature vector S is assigned to obtain the weighted feature vector V. The weighted feature vector V is then fused, and finally the fused feature vector is input into a fully connected circuit to achieve the mapping process of fault type. The feature vector V is calculated by the following equation (3):

[0088]

[0089] Among them, W j Let S be the attention weight for the j-th channel, and let S be the feature values ​​obtained after inputting data Y into a series of convolutional layers. j Let be the feature value of the j-th channel.

[0090] For large and complex bearings, single-sensor monitoring is insufficient for accurate monitoring of the overall fault state. Therefore, multi-source information fusion is necessary to integrate data from multiple sensors, mapping from local to global information. By weighting each channel, the diagnostic model can perceive the correlation between each channel and the fault. When a fault occurs near a corresponding sensor, the data output by that sensor has a stronger correlation with the fault, and the weight obtained through the attention mechanism is also greater. This method enables the model to have a certain degree of fault location awareness.

[0091] In another embodiment of the present invention, step S1 specifically includes the following steps:

[0092] S11, based on prior knowledge, solve for the fault characteristic frequency of the corresponding region, and combine it with the sampling frequency f. s The required data segment length l is calculated using the following formula (4):

[0093]

[0094] Where f is the fault characteristic frequency, f s Where N is the sampling frequency, and N is the number of periods within the data segment, with N ranging from 5 to 10.

[0095] Calculate the fault characteristic frequency f based on Table 1 below:

[0096] Table 1. Formulas for calculating fault characteristic frequencies at various locations.

[0097]

[0098] Among them, refer to the appendix Figure 6 D m Main raceway roller pitch circle diameter, D a To assist in pushing the pitch circle diameter of the raceway rollers, D r This is the radial raceway roller pitch circle diameter, d m The main driving roller diameter, d a To adjust the diameter of the auxiliary roller, d r f is the radial roller diameter. r Let n be the rotation frequency of the inner ring. m The number of main roller pairs, n a To determine the number of auxiliary rollers, n r This refers to the number of radial rollers;

[0099] As shown in Table 1, there are three fault types for the outer ring, three fault types for the rollers, and three fault types for the inner ring, for a total of nine fault types.

[0100] The maximum value among the data segment lengths corresponding to each fault type is selected as the final data segment length, and all data is divided using this data segment length;

[0101] S12, using the sliding window method, the time-domain data is segmented according to the data segment length l obtained above:

[0102]

[0103] Where L is the total length of the data, stride is the step size, and num represents the number of data segments after segmentation. num and l are essential hyperparameters for fault diagnosis training and play a decisive role in the parameter results of fault training.

[0104] S13, and then the intermediate data J is obtained by normalizing the data of each channel using formula (6):

[0105]

[0106] in This represents the k-th data segment in the n-th channel, which is the intermediate data J. This represents the k-th data segment of the n-th channel, where k∈[0,num]. This represents the maximum value in the corresponding data segment. This represents the minimum value in the corresponding data segment.

[0107] In another embodiment of the present invention, step S2 specifically includes the following steps:

[0108] S21, firstly, dilated wide kernel convolution is used to perform preliminary processing on the intermediate data J. The calculation formula (7) of dilated wide kernel convolution is as follows:

[0109]

[0110] Where Y n [i] is the output at position i, which is the data Y; W[m] are the weights in the convolution kernel, corresponding to position m of the kernel; d is the dilation rate, which determines the spacing between elements of the convolution kernel; M is half the size of the convolution kernel. For intermediate data J, It is the [i+d·m]th data point in the corresponding data segment;

[0111] This step can initially filter out high-frequency components in the data, thereby improving the ability to adapt to low-frequency components.

[0112] S22, for the output result Y n [i] Batch normalization is performed to prevent gradient explosion and gradient vanishing during training, and to initially prevent overfitting. The formula is shown below:

[0113] Calculate the mean μ of the entire batch. n :

[0114] Calculate the variance of the entire batch.

[0115] Standardization:

[0116] Scaling and translation (reparameterization):

[0117] Where ∈ is a very small number to prevent the denominator from being zero; γ and β are learnable parameters that allow the network to learn to recover any potentially useful feature distributions; x n [i] represents the data point of the nth channel. Represents the normalized data points, y n[i] represents the output after batch normalization, which is data G;

[0118] S23, introduce nonlinear components into the data and perform maximum pooling using formula (8):

[0119] U n =maxpool(ReLU(X)) (8)

[0120] Among them, U n ReLU(X) represents the feature value extracted from the nth channel. ReLU(X) represents the ReLU function processing of the data G, where ReLU is a nonlinear function whose main function is to introduce nonlinear factors into the internal parameters of the model.

[0121] In another embodiment of the present invention, step S3 specifically includes the following steps:

[0122] S31, the feature values ​​obtained from each channel are summed element by element using formula (9) to form the feature vector U:

[0123] U = U1 + U2 + U3 (9)

[0124] This step does not modify the dimensions of U, and simultaneously integrates information from multiple channels.

[0125] S32, U is processed by a global average pooling layer (GAP), a fully connected layer (FC), a batch normalization (BN) layer, and a ReLU activation function to obtain feature Z. The mathematical model of Z can be further expressed as:

[0126] Z=δ(ψ(U gap (10)

[0127] Among them, U gap ψ represents the result of U after global pooling, ψ represents the batch normalization process (BN), and δ represents the activation function ReLU.

[0128] The data is then fed into a fully connected layer for further compression of the feature dimensions.

[0129] Z = Dense(C / r, C) (11)

[0130] Where Dense represents the fully connected process, r is the feature compression ratio, and C is the number of input features, with C typically set to r = 0.5;

[0131] S33, the compressed feature Z is input again into the three fully connected layers to generate three sets of attention weights, which are then normalized by SoftMax. The specific process is as follows:

[0132]

[0133] Among them, A s B s C s For the s-th row of the feature matrix, the a generated from this part s b s c s This represents the attention weights for the three channels, a. s b s c s Equivalent to W1, W2, W3, formulas (12), (13), and (14) satisfy the following relationship: a s +b s +c s =1. Since the number of channels used in this embodiment is 3, Z has 3 parts, namely A, B, and C. Each part is further divided into 10 rows according to the 10 categories to be identified (9 fault categories in Table 1 in step S11 plus the no-fault category), and s is the index of each specific fault type.

[0134] In another embodiment of the present invention, step S4 specifically includes the following steps:

[0135] S41, by convolving the feature Y obtained in step two with a wide-kernel hole convolution... n [i] Input a series of convolutional layers to further extract deeper features:

[0136] S i =Convs(16,128) (15)

[0137] The feature channels obtained from the dilated wide kernel convolution are 16. After a series of convolutions, the number of channels becomes 128. i Let S be the eigenvector.

[0138] S42, Multiply the feature vector S obtained from the convolution in step S41 with the attention weights obtained in step S3 to obtain the weighted feature vector V: V = a s ·S1+b s ·S2+c s ·S3.

[0139] This invention first adds and fuses the data from the three channels, and then generates three attention weights through a fully connected layer corresponding to the number of channels. During the training process, the fully connected layer reduces the gain of channels with low correlation to the fault, making the fault diagnosis method pay more attention to the channels with high correlation to the fault, thus mitigating the risk of overfitting caused by information redundancy in the multi-channel fault diagnosis process.

[0140] This invention first adds and fuses the data from the three channels, and then generates three attention weights through a fully connected layer corresponding to the number of channels. During the training process, the fully connected layer reduces the gain of channels with low correlation to the fault, making the fault diagnosis method pay more attention to the channels with high correlation to the fault, thus mitigating the risk of overfitting caused by information redundancy in the multi-channel fault diagnosis process.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A test bench for simulating main bearing failure in a tunnel boring machine, characterized in that, The system includes a frame and a main bearing rotatably mounted on the frame. The main bearing has a first outer ring and a second outer ring at its front and rear ends, respectively. A groove is provided on the side of the second outer ring facing the main thrust roller, and a fault simulation plate is held in the groove. The fault simulation plate is interference-fitted with the second outer ring. A fault detection module is installed on the outer side of the second outer ring at a position corresponding to the fault simulation plate. A screw hole is provided at the rear end of the fault simulation plate, and a bolt passes through the second outer ring and extends into the screw hole to hold the fault simulation plate in the groove. The frame is equipped with a loading device and an oil circulation device. The oil circulation device provides lubricant to the main bearing, and the loading device drives the main bearing to rotate and provides axial thrust to the main bearing. The system also includes the following steps: S1: Vibration data x of the main bearing obtained by the fault detection module. n (t), each fault detection module provides one channel of vibration data x n (t), and then adjust the data segment length in the data preprocessing process according to the fault characteristic frequency, and then normalize the data of each channel to obtain data J; where x n (t) represents the vibration data of the nth channel at time t; S2: Data J is initially processed using dilated wide-kernel convolution to obtain data Y. Data Y is then batch normalized to obtain data G. Finally, a nonlinear component is introduced into data G and maximum pooling is performed to obtain the channel features U of the channels. n ; S3: Transfer channel features U n The feature vector U is obtained by element-wise summation. Then, the feature vector U is processed through a global average pooling layer (GAP), a fully connected layer (FC), a batch normalization (BN) layer, and a ReLU activation function to obtain the feature Z. Finally, the feature Z is input into the fully connected layer to obtain the attention weights W for each channel. j The sum of the attention weights for each channel is 1, that is... Where m = 3; S4: Calculate the attention weights W j The feature vector S is assigned to obtain the weighted feature vector V. The weighted feature vector V is then fused, and finally the fused feature vector is input into the final fully connected connection to realize the mapping process of fault type. The feature vector V is calculated by the following formula (3): Among them, W j Let be the attention weight for the j-th channel, and let S be the feature vector S obtained after inputting data Y into a series of convolutional layers. j Let j be the feature value of the j-th channel; Step S1 specifically includes the following steps: S11, based on prior knowledge, solve for the fault characteristic frequency of the corresponding region, and combine it with the sampling frequency f. s The required data segment length l is calculated using the following formula (4): Where f is the fault characteristic frequency, f s Where N is the sampling frequency, and N is the number of periods within the data segment, with N ranging from 5 to 10. Calculate the fault characteristic frequency f according to the table below (1): Table 1. Formulas for calculating fault characteristic frequencies at various locations. Among them, D m Main raceway roller pitch circle diameter, D a To assist in pushing the pitch circle diameter of the raceway rollers, D r This is the radial raceway roller pitch circle diameter, d m The main driving roller diameter, d a To adjust the diameter of the auxiliary roller, d r f is the radial roller diameter. r Let n be the rotation frequency of the inner ring. m The number of main roller pairs, n a To determine the number of auxiliary rollers, n r This refers to the number of radial rollers; S12, using the sliding window method, segment the time-domain data according to the data segment length l obtained above: Where L is the total length of the data, stride is the step size, and num represents the number of data segments after division; S13, and then the data of each channel are normalized using formula (6): in This represents the k-th data segment in the n-th channel, which is data J. This represents the k-th data segment of the n-th channel, where k∈[0,num]. This represents the maximum value in the corresponding data segment. This represents the minimum value in the corresponding data segment.

2. The tunnel boring machine main bearing failure simulation test bench according to claim 1, characterized in that, The second outer ring and the first outer ring are fixed together by bolts.

3. The tunnel boring machine main bearing failure simulation test bench according to claim 2, characterized in that, A notch is provided between the fault simulation board and the bottom of the card slot.

4. The tunnel boring machine main bearing failure simulation test bench according to claim 3, characterized in that, There are three fault simulation boards, which are distributed at the three o'clock, six o'clock and nine o'clock positions on the second outer ring.

5. The tunnel boring machine main bearing failure simulation test bench according to claim 1, characterized in that, The loading device includes a servo motor, a drive shaft, a gear, and an internal gear ring. The internal gear ring is fixed to the inner side of the main bearing, the gear meshes with the internal gear ring, the gear is fixed on the drive shaft, and the drive shaft is fixedly connected to the output shaft of the servo motor.

6. The tunnel boring machine main bearing failure simulation test bench according to claim 5, characterized in that, Step S2 specifically includes the following steps: S21, First, dilated wide kernel convolution is used to perform preliminary processing on the data J. The calculation formula (7) of dilated wide kernel convolution is as follows: Where Y n [i] is the output at position i, which is the data Y; W[m] is the weight in the convolution kernel, corresponding to position m of the kernel; d is the porosity, which determines the spacing between elements in the convolution kernel; M is half the size of the convolution kernel. For data J, It is the [i+d·m]th data point in the corresponding data segment; S22, for the output result Y n [i] Perform batch normalization, the formula of which is shown below: Calculate the mean μ of the entire batch. n : Calculate the variance of the entire batch. Standardization: Scaling and displacement, reparameterization: Where ∈ is a very small number to prevent the denominator from being zero; γ and β are learnable parameters that allow the network to learn to recover any potentially useful feature distributions; x n [i] represents the data point of the nth channel. Represents the normalized data points, y n [i] represents the output after batch normalization, which is data G; S23, introduce nonlinear components into the data and perform maximum pooling using formula (8): HER n =maxpool(ReLU(X)) (8) Among them, U n Then it is the feature value extracted from the nth channel, and ReLU(X) represents the ReLU function processing of the data G.

7. The tunnel boring machine main bearing failure simulation test bench according to claim 6, characterized in that, Step S3 specifically includes the following steps: S31, the feature values ​​obtained from each channel are summed element by element using formula (9): U = U1 + U2 + U3 (9) Among them, the feature vector U integrates information from multiple channels; S32, U is processed by a global average pooling layer (GAP), a fully connected layer (FC), a batch normalization (BN) layer, and a ReLU activation function to obtain feature Z. The mathematical model of Z can be further expressed as: Z=δ(ψ(U gap )) (10) Among them, U gap ψ represents the result of U after global pooling, ψ represents the batch normalization process (BN), and δ represents the activation function ReLU. The data is then fed into a fully connected layer for further compression of the feature dimensions. Z = Dense(C / r, C) (11) Where Dense represents the fully connected process, r is the feature compression ratio, and C is the number of input features, with C typically set to r = 0.5; S33, the compressed feature Z is input again into the three fully connected layers to generate three sets of attention weights, which are then normalized by SoftMax. The specific process is as follows: Among them, A s B s C s For the s-th row of the feature matrix, the a generated from this part s b s c s The attention weights for the three channels are then obtained. Formulas (12), (13), and (14) satisfy the following relationship: a s +b s +c s =1.

8. The tunnel boring machine main bearing failure simulation test bench according to claim 7, characterized in that, Step S4 specifically includes the following steps: S41, by convolving the feature Y obtained in step two with a wide-kernel hole convolution... n [i] Input a series of convolutional layers to further extract deeper features: S i =Convs(16,128) (15) The feature channels obtained from the dilated wide kernel convolution are 16. After a series of convolutions, the number of channels becomes 128. i Let S be the eigenvector; S42, convolve the feature vector S obtained in step S41 i Multiplying the weights obtained in step S3 by the attention weights, we get the weighted feature vector V: V = a s ·S1+b s ·S2+c s ·S3.

Images