Tunnel boring machine main bearing fault diagnosis method and system based on roller status

Abstract

The present invention provides a method and system for diagnosing main bearing faults of a tunnel boring machine based on roller status, comprising: step S1: obtaining the three-axis vibration data of the main bearing roller; step S2: removing the motion trend item of the sensor data in the roller by the variational mode decomposition method; step S3: establishing a dual-stream multi-channel deep convolutional neural network to train the three-axis vibration data of the roller. This method solves the problem that the sensor and the vibration source are far apart in the traditional vibration detection scheme and cannot effectively sense the fault source. The proposed dual-stream multi-channel deep convolutional neural network is suitable for the analysis of roller status data under low speed, heavy load and strong impact conditions of tunnel boring machines, and has good anti-noise ability.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of a tunnel boring machine main bearing, and in particular to a method and system for diagnosing a tunnel boring machine main bearing fault based on roller states. Background Art

[0002] As a key component of the main drive system of full-section tunnel boring machines and shield machines, the main bearing is prone to wear, deformation, fracture and other faults when operating in a complex and harsh underground environment for a long time. Tunnel boring machines often shut down due to faults, which brings huge economic losses. At present, the main bearings of tunnel boring machines are mostly internal gear three-row roller combination slewing bearings. The monitoring scheme mainly measures vibration signals, pressure signals, displacement signals, and oil conditions. For example, SKF uses displacement sensors to measure the changes in the axial and radial clearances of the main bearings to determine whether a fault has occurred [Bearings with distance sensors and tapered grooves, CN 112901653 A]; the patent [Intelligent shield machine main bearing with health status monitoring, CN204755617 U] installs temperature sensors, vibration sensors, and displacement sensors to collect bearing operating parameters; the patent [Main bearing assembly with embedded sensors for full-section rock tunnel boring machines, CN 103671581 A] installs displacement sensors, pressure sensors, oil monitoring sensors, and speed sensors to monitor the operating status of the main bearings. However, the existing tunnel boring machine vibration monitoring solution has the problem that the sensor cannot sense the fault source that is far away from it, which greatly increases the difficulty of fault identification of the entire raceway. It is necessary to propose a new tunnel boring machine main bearing vibration monitoring solution that can be applied at the tunnel boring machine construction site.

[0003] In terms of algorithms, data-driven rolling bearing fault diagnosis methods are divided into two categories according to feature extraction methods: the first method manually extracts features based on feature engineering technology and uses traditional machine learning methods for fault classification; the second method directly applies deep learning algorithms to automatically extract features and then output fault classification results. The low-speed and heavy-load characteristics of the main bearing of a tunnel boring machine make the fault characteristics easily masked by background noise. Therefore, it is very challenging to mine and classify the weak fault characteristics of the main bearing of a tunnel boring machine from the vibration signal contaminated by strong noise. Summary of the invention

[0004] In view of the defects in the prior art, an object of the present invention is to provide a method for diagnosing main bearing faults of a tunnel boring machine based on roller status.

[0005] A method for diagnosing a main bearing fault of a tunnel boring machine based on roller status according to the present invention comprises:

[0006] Step S1: Acquire the three-axis vibration data of the main bearing roller;

[0007] Step S2: removing the motion trend term of the sensor data in the roller by using the variational mode decomposition method;

[0008] Step S3: Establish a dual-stream multi-channel deep convolutional neural network to train the roller triaxial vibration data.

[0009] Preferably, an accelerometer and a wireless transmission module are integrated inside the roller.

[0010] Preferably, step S2 comprises:

[0011] Step S2.1: Decompose the collected vibration signal of the main bearing roller to obtain k eigenmodal components u k (t), calculate the central angular frequency ω of each component k ; The implementation process includes:

[0012] The analytical signal of the mode is obtained by Hilbert transform, and its one-sided spectrum is calculated. The analytical signal is multiplied by the exponential term Make the center frequency band ω k Shift to baseband and calculate bandwidth The equation is:

[0013]

[0014] In the formula, * represents the convolution operation, It means to find the gradient with respect to time t, j is the imaginary unit, δ(t) is the Dirichlet function, and k is the number of modes. The constrained optimization problem equation is established according to the component narrowband condition:

[0015]

[0016]

[0017] Introducing the Lagrangian multiplier λ and the quadratic penalty factor α, by augmenting the Lagrangian function L({u k},{ω k},λ) transform the established constrained optimization problem into an unconstrained optimization problem:

[0018]

[0019] Among them, λ(t) is the Lagrangian operator, f(t) is the original signal, and the alternating direction multiplier method is used to find the saddle point of the augmented Lagrangian function. k and the center frequency ω k The update equation of the frequency domain optimization problem is obtained:

[0020]

[0021]

[0022] ω k , and are the mode, center frequency, original signal and Fourier transform of Lagrange multiplier respectively. n is the number of iterations. The above steps are iterated continuously until the energy ratio of the residual signal to the original signal is less than the convergence threshold ε:

[0023]

[0024] Step S2.2: Compare the central angular frequency of the low-frequency eigenmode component obtained by decomposition with the calculated theoretical value of the roller rotation frequency to characterize the roller motion trend term;

[0025] The degree of approximation between the low-frequency mode and the roller trend is determined by enumeration method, and the objective function is taken as

[0026] min||f r -ω1 / 2π||

[0027] Thus the modal number k is determined;

[0028] The expression of roller rotation frequency is:

[0029]

[0030] Among them, f r 、f c 、f i They are the rotation frequencies of the roller, the center of the roller around the axis, and the inner ring, respectively. i and d are the inner ring and roller diameters respectively;

[0031] Step S2.3: Remove the eigenmode components that characterize the trend term of roller motion, reconstruct the remaining eigenmode components, and obtain the main bearing fault monitoring vibration data with the trend term removed.

[0032] Preferably, step S3 includes the following sub-steps:

[0033] Step S3.1: The vibration data of the roller x-axis and y-axis with the motion trend item removed by the variational mode decomposition method and the original z-axis data are used to jointly establish a fault data set. One part is enhanced by overlapping sampling method to generate training samples, and the other part is generated by non-overlapping sampling to generate test samples. The samples are divided into training set, verification set and test set in a ratio of 7:3:1;

[0034] Step S3.2: construct a dual-stream multi-channel deep convolutional neural network, including a dual-channel deep convolutional neural network and a single-channel deep convolutional neural network, input the trend-free radial vibration signals of the x-axis and y-axis into the dual-channel deep convolutional neural network to train multi-directional radial vibration features, and train the axial data in the z-axis direction through a single-channel deep convolutional neural network to obtain the axial vibration features;

[0035] Step S3.3: The multi-directional radial vibration and axial vibration features of the roller during rotation are fused through the fully connected layer of the neural network, and the Softmax logistic regression function is used to calculate the probability distribution of the multi-classification task.

[0036] Preferably, the structure of the neural network includes an input layer, a convolutional layer, a pooling layer and a fully connected layer;

[0037] In the convolution layer, a convolution operator is used to convert the one-dimensional vibration data into a feature map. Each convolution kernel captures the corresponding feature and slides through the input data with the same weight and fixed step size. The calculation formula of the convolution layer is:

[0038]

[0039] In the formula, z (l) is the feature map of the lth layer, a (l-1) Input data for the l-1th layer, convolution kernel ω (l) As the weight vector, b (l) is bias;

[0040] Use ReLU(x)=max(0,x) non-restrictive unit as activation function to get non-linear expression of convolutional layer output;

[0041] In the pooling layer, the obtained feature map is divided into multiple regions, and each region is downsampled using the maximum pooling method: for region R of feature map d m , select the neuron x in this area i The average output The expression is as follows:

[0042]

[0043] Preferably, the function expression of the probability distribution is:

[0044]

[0045] Among them, p i is the predicted probability of the corresponding classification output node, z l (i) is the output value of the i-th node, and n is the number of output nodes, that is, the number of classification categories.

[0046] A roller state-based fault diagnosis system for a tunnel boring machine main bearing according to the present invention comprises:

[0047] Module M1: Obtain the three-axis vibration signal of the main bearing roller;

[0048] Module M2: Remove the motion trend term of the sensor data inside the roller by using the variational mode decomposition method;

[0049] Module M3: Establish a dual-stream multi-channel deep convolutional neural network to train roller triaxial vibration data.

[0050] Preferably, an accelerometer and a wireless transmission module are integrated inside the roller.

[0051] Preferably, the module M2 comprises:

[0052] Module M2.1: Decompose the collected vibration signal of the main bearing roller to obtain k intrinsic mode components u k (t), calculate the central angular frequency ω of each component k ; The implementation process includes:

[0053] The analytical signal of the mode is obtained by Hilbert transform, and its one-sided spectrum is calculated. The analytical signal is multiplied by the exponential term Make the center frequency band ω k Shift to baseband and calculate bandwidth The equation is:

[0054]

[0055] In the formula, * represents the convolution operation, It means to find the gradient with respect to time t, j is the imaginary unit, δ(t) is the Dirichlet function, and k is the number of modes. The constrained optimization problem equation is established according to the component narrowband condition:

[0056]

[0057]

[0058] Introducing the Lagrangian multiplier λ and the quadratic penalty factor α, by augmenting the Lagrangian function L({u k},{ω k},λ) transform the established constrained optimization problem into an unconstrained optimization problem:

[0059]

[0060] Among them, λ(t) is the Lagrangian operator, f(t) is the original signal, and the alternating direction multiplier method is used to find the saddle point of the augmented Lagrangian function. k and the center frequency ωk The update equation of the frequency domain optimization problem is obtained:

[0061]

[0062]

[0063] ω k , and are the mode, center frequency, original signal and Fourier transform of Lagrange multiplier respectively. n is the number of iterations. The above steps are iterated continuously until the energy ratio of the residual signal to the original signal is less than the convergence threshold ε:

[0064]

[0065] Module M2.2: Compare the central angular frequency of the low-frequency eigenmodal component obtained by decomposition with the theoretical value of the roller rotation frequency calculated to characterize the roller motion trend term;

[0066] The enumeration method is used to determine the approximation between the low-frequency mode and the roller trend, and the objective function is taken as

[0067] min||f r -ω1 / 2π||

[0068] Thus the modal number k is determined;

[0069] The expression of roller rotation frequency is:

[0070]

[0071] Among them, f r 、f c 、f i They are the rotation frequencies of the roller, the center of the roller around the axis, and the inner ring, respectively. i and d are the inner ring and roller diameters respectively;

[0072] Module M2.3: Remove the intrinsic mode components that characterize the trend term of roller motion, reconstruct the remaining intrinsic mode components, and obtain the main bearing fault monitoring vibration data with the trend term removed.

[0073] Preferably, the module M3 includes the following submodules:

[0074] Module M3.1: The vibration data of the roller x-axis and y-axis with the motion trend removed by the variational mode decomposition method and the original z-axis data are used to build a fault data set. One part is enhanced by overlapping sampling method to generate training samples, and the other part is generated by non-overlapping sampling to generate test samples. The samples are divided into training set, validation set and test set in a ratio of 7:3:1.

[0075] Module M3.2: Construct a dual-stream multi-channel deep convolutional neural network, including a dual-channel deep convolutional neural network and a single-channel deep convolutional neural network. Input the trend-free radial vibration signals of the x-axis and y-axis into the dual-channel deep convolutional neural network to train multi-directional radial vibration features. The axial data in the z-axis direction is trained separately through a single-channel deep convolutional neural network to obtain the axial vibration features.

[0076] Module M3.3: The multi-directional radial vibration and axial vibration features of the roller during rotation are fused through the fully connected layer of the neural network, and the Softmax logistic regression function is used to calculate the probability distribution of multi-classification tasks.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] 1. This method solves the problem that the distance between the sensor and the vibration source in the traditional vibration detection solution is far and the fault source cannot be effectively sensed.

[0079] 2. The dual-stream multi-channel deep convolutional neural network proposed in the present invention is suitable for the analysis of roller state data of tunnel boring machines under low speed, heavy load and strong impact conditions, and has good anti-noise ability.

[0080] 3. The dual-stream multi-channel deep convolutional neural network proposed in the present invention can effectively identify the fault type of the main bearing of the tunnel boring machine under strong noise conditions at different speeds. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0082] Figure 1 This is a dual-stream multi-channel deep convolutional neural network architecture diagram of the present invention;

[0083] Figure 2 A top view of a tunnel boring machine main bearing fault simulation test bench according to an embodiment of the present invention;

[0084] Figure 3 It is a right view of the tunnel boring machine main bearing fault simulation test bench in an embodiment of the present invention;

[0085] Figure 4 This is a front view of a test main bearing in an embodiment of the present invention;

[0086] Figure 5 This is a structural diagram of a dual-stream multi-channel deep convolutional neural network and a three-channel deep convolutional neural network in an embodiment of the present invention.

[0087] Description of reference numerals:

[0088] Test main bearing 1 Motor control cabinet 7

[0089] First support bearing 2 Hydraulic loading device 8

[0090] Spindle 3 Main bearing outer ring 101

[0091] Second support bearing 4 Main bearing roller 102

[0092] Reducer 5 Main bearing cage 103

[0093] Drive motor 6 Main bearing inner ring 104 DETAILED DESCRIPTION

[0094] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0095] The present invention discloses a method for fault diagnosis of the main bearing of a tunnel boring machine based on roller status. By designing a roller with an integrated accelerometer and a wireless transmission module, the method is used to monitor the health of the roller and the entire raceway of the main bearing of the tunnel boring machine. A dual-stream multi-channel deep convolutional neural network is proposed for the three-axis vibration signal of the accelerometer in the roller. The architecture is as follows: Figure 1 Specifically, the diagnostic method comprises the following steps:

[0096] Step S1: Acquire the three-axis vibration signal of the main bearing roller, where an accelerometer and a wireless transmission module are integrated.

[0097] Step S2: removing the motion trend item of the sensor data in the roller by using the variational mode decomposition method, specifically including:

[0098] Step S2.1: Decompose the collected vibration signal of the main bearing roller to obtain k eigenmodal components u k (t), calculate the central angular frequency ω of each component k ; The implementation process includes:

[0099] The analytical signal of the mode is obtained by Hilbert transform, and its one-sided spectrum is calculated. The analytical signal is multiplied by the exponential term Make the center frequency band ω k Shift to baseband and calculate bandwidth The equation is:

[0100]

[0101] In the formula, * represents the convolution operation, It means to find the gradient with respect to time t, j is the imaginary unit, δ(t) is the Dirichlet function, and k is the number of modes. The constrained optimization problem equation is established according to the component narrowband condition:

[0102]

[0103]

[0104] Introducing the Lagrangian multiplier λ and the quadratic penalty factor α, by augmenting the Lagrangian function L({u k},{ω k},λ) transform the established constrained optimization problem into an unconstrained optimization problem:

[0105]

[0106] Among them, λ(t) is the Lagrangian operator, f(t) is the original signal, and the alternating direction multiplier method is used to find the saddle point of the augmented Lagrangian function. k and the center frequency ω k The update equation of the frequency domain optimization problem is obtained:

[0107]

[0108]

[0109] ω k , and are the mode, center frequency, original signal and Fourier transform of Lagrange multiplier respectively. n is the number of iterations. The above steps are iterated continuously until the energy ratio of the residual signal to the original signal is less than the convergence threshold ε:

[0110]

[0111] Step S2.2: Compare the central angular frequency of the low-frequency eigenmode component obtained by decomposition with the calculated theoretical value of the roller rotation frequency to characterize the roller motion trend term;

[0112] The degree of approximation between the low-frequency mode and the roller trend is determined by enumeration method, and the objective function is taken as:

[0113] min||f r -ω1 / 2π||

[0114] Thus the modal number k is determined;

[0115] The expression of roller rotation frequency is:

[0116]

[0117] Among them, f r 、f c 、f i They are the rotation frequencies of the roller, the center of the roller around the axis, and the inner ring, respectively. i and d are the inner ring and roller diameters respectively;

[0118] Step S2.3: Remove the eigenmode components that characterize the trend term of roller motion, reconstruct the remaining eigenmode components, and obtain the main bearing fault monitoring vibration data with the trend term removed.

[0119] Step S3: Establish a dual-stream multi-channel deep convolutional neural network to train the roller triaxial vibration data.

[0120] The two-stream multi-channel deep convolutional neural network consists of two similar deep convolutional neural networks. The two network branches maintain the same structure and parameters, and are composed of 5 convolutional layers, 5 pooling layers, and 1 fully connected layer. The neuron information of the fully connected layers of the two branches is fused and processed, and the result is used as the input of the fully connected network for full connection operation again, and finally the classification result is output. The C1 convolution layer uses 16 large convolution kernels of length 64, and the C2~C5 convolution layers use small convolution kernels of length 3. Batch Normalization is performed after each convolution layer and activated by ReLU linear correction unit. The P1~P5 pooling layers use the maximum pooling of 2×1 area, and the number of neurons in the fully connected layer is 100. The model parameters are shown in Table 1.

[0121] Table 1. Parameters of the two-stream multi-channel deep convolutional neural network model

[0122]

[0123] The specific process of vibration data analysis is as follows:

[0124] Step S3.1: The vibration data of the roller x-axis and y-axis with the motion trend item removed by the variational mode decomposition method and the original z-axis data are used to jointly establish a fault data set. One part is enhanced by overlapping sampling method to generate training samples, and the other part is generated by non-overlapping sampling to generate test samples. The samples are divided into training set, verification set and test set in a ratio of 7:3:1;

[0125] Step S3.2: Construct a dual-stream multi-channel deep convolutional neural network, including a dual-channel deep convolutional neural network and a single-channel deep convolutional neural network. Input the trend-free radial vibration signals of the x-axis and y-axis into the dual-channel deep convolutional neural network to train the multi-directional radial vibration features, and train the axial data in the z-axis direction through a single-channel deep convolutional neural network to obtain the axial vibration features. The network structure includes an input layer, a convolution layer, a pooling layer, and a fully connected layer. The first layer of convolution uses a large-size convolution kernel to obtain a larger receptive field and enhance the learning ability of generalized features. Multi-layer small convolution kernels are used to increase the network depth to achieve better learning and classification effects of the algorithm. After each layer of convolution, Batch Normalization is performed and activated by ReLU linear correction units, and then maximum pooling is performed. The features obtained by the last layer of pooling are expanded and merged as the input of the fully connected layer. The specific implementation of each layer is as follows:

[0126] Convolution layer: Use the convolution operator to convert the one-dimensional vibration data into a feature map. Each convolution kernel captures the corresponding feature and slides through the input data with the same weight and fixed step size. The calculation formula of the convolution layer is:

[0127]

[0128] In the formula, z (l) is the feature map of the lth layer, a (l-1) Input data for the l-1th layer, convolution kernel ω (l) As the weight vector, b (l) The nonlinear expression of the convolutional layer output is obtained by using ReLU(x)=max(0,x) non-restrictive unit as the activation function.

[0129] Pooling layer: Divide the obtained feature map into multiple regions, and use the maximum pooling method to downsample each region: For region R of feature map d m , select the neuron x in this area i The average output The expression is as follows:

[0130]

[0131] Step S3.3: The multi-directional radial vibration and axial vibration features of the roller during rotation are fused through the neural network fully connected layer, and the Softmax logistic regression function is used to calculate the probability distribution of the multi-classification task. The function expression is:

[0132]

[0133] Among them, p i is the predicted probability of the corresponding classification output node, z l(i) is the output value of the i-th node, and n is the number of output nodes, that is, the number of classification categories.

[0134] The present invention designs a tunnel boring machine main bearing fault simulation test bench according to the proposed fault diagnosis scheme to verify the effectiveness of the method. Figure 2 A top view of the tunnel boring machine main bearing failure simulation test bench is shown. Figure 3 The right side view of the tunnel boring machine main bearing fault simulation test bench is shown. The figure includes: test main bearing 1, first support bearing 2, main shaft 3, second support bearing 4, reducer 5, drive motor 6, motor control cabinet 7, hydraulic loading device 8, Figure 4 The main view of the test main bearing is shown, including: main bearing outer ring 101, main bearing roller 102, main bearing cage 103, main bearing inner ring 104. Among them, the motor control cabinet 7 can adjust the motor speed, the hydraulic loading device 8 can control the load applied to the first support bearing 2, and the main bearing roller 1-2 is internally integrated with a three-axis acceleration sensor, a wireless transmission module, a power supply module and an MCU chip. The sensor obtains the fault-containing vibration signals of the inner and outer raceways and rollers of the bearing through the dynamic contact between the rollers and the raceways. The roller internal sensor data of each bearing state at different speeds constitutes the tunnel boring machine main bearing fault data set.

[0135] The algorithm effect was verified on the main bearing fault simulation test bench. The load was set to 20MPa, and the rotation speeds were 30, 40 and 50rpm respectively. The duration of a single group of experiments was 10min. The faulty parts were processed by cutting to have defects of 2mm width and 1mm depth. The outer ring fault, inner ring fault and roller fault were simulated, and the vibration signals of the x, y and z axes of the accelerometer in the roller were collected to construct data sets A, B and C, corresponding to bearing speeds of 30, 40 and 50rpm respectively. In order to verify the anti-noise performance of the method proposed in this patent, Gaussian white noise is added to the data set to generate a noisy data set. The signal-to-noise ratio is set to -10dB, -8dB, -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB. The dual-stream multi-channel deep convolutional neural network proposed in this patent is compared with the single-channel deep convolutional neural network and traditional machine learning methods. Among them, machine learning methods such as SVM (support vector machine), RF (random forest), and XGBoost (extreme gradient boosting) do not have end-to-end diagnostic capabilities and rely on manually extracted wavelet packet energy features as input. The experimental results are shown in Table 2. In -10dB noise datasets A, B and C, the accuracy of the dual-stream multi-channel deep convolutional neural network is 60.56%, 55.63% and 53.13% higher than SVM, 66.94%, 65.25% and 58.38% higher than RF, 66.06%, 62.25% and 55.13% higher than XGB, and 61.81%, 52.38% and 55.38% higher than DCNN. This shows that the method proposed in this patent can effectively analyze the three-axis vibration data of the roller, and has more reliable diagnostic results than other methods under strong noise and variable speed conditions.

[0136] Table 2. Recognition accuracy of the model in datasets A, B and C under different noises

[0137]

[0138] In addition, in order to further illustrate the particularity and effectiveness of the dual-stream multi-channel network structure proposed in this patent, a three-channel deep convolutional neural network with the same structural parameters is set, such as Figure 4 The noise resistance experimental results are shown in Table 3. When the SNR value is -10dB, the accuracy of the dual-stream multi-channel deep convolutional neural network at different speeds is improved by 27.19%, 12.13% and 16.13% respectively compared with the three-channel deep convolutional neural network, which verifies that the proposed network structure can effectively identify the fault type of the main bearing of the tunnel boring machine under strong noise conditions at different speeds.

[0139] Table 3. Comparison of the noise resistance of the two-stream multi-channel deep convolutional neural network method and the three-channel deep convolutional neural network method

[0140]

[0141] The present invention also discloses a roller state-based tunnel boring machine main bearing fault diagnosis system, comprising:

[0142] Module M1: Acquire the three-axis vibration signal of the main bearing roller, which integrates an accelerometer and a wireless transmission module.

[0143] Module M2: Remove the motion trend term of the sensor data in the roller by using the variational mode decomposition method, including:

[0144] Module M2.1: Decompose the collected vibration signal of the main bearing roller to obtain k intrinsic mode components u k (t), calculate the central angular frequency ω of each component k ; The implementation process includes:

[0145] The analytical signal of the mode is obtained by Hilbert transform, and its one-sided spectrum is calculated. The analytical signal is multiplied by the exponential term Make the center frequency band ω k Shift to baseband and calculate bandwidth The equation is:

[0146]

[0147] In the formula, * represents the convolution operation, It means to find the gradient with respect to time t, j is the imaginary unit, δ(t) is the Dirichlet function, and k is the number of modes. The constrained optimization problem equation is established according to the component narrowband condition:

[0148]

[0149]

[0150] Introducing the Lagrangian multiplier λ and the quadratic penalty factor α, by augmenting the Lagrangian function L({u k},{ω k},λ) transform the established constrained optimization problem into an unconstrained optimization problem:

[0151]

[0152] Among them, λ(t) is the Lagrangian operator, f(t) is the original signal, and the alternating direction multiplier method is used to find the saddle point of the augmented Lagrangian function. k and the center frequency ω k The update equation of the frequency domain optimization problem is obtained:

[0153]

[0154]

[0155] ω k , and are the mode, center frequency, original signal and Fourier transform of Lagrange multiplier respectively. n is the number of iterations. The above steps are iterated continuously until the energy ratio of the residual signal to the original signal is less than the convergence threshold ε:

[0156]

[0157] Module M2.2: Compare the central angular frequency of the low-frequency eigenmodal component obtained by decomposition with the theoretical value of the roller rotation frequency calculated to characterize the roller motion trend term;

[0158] The degree of approximation between the low-frequency mode and the roller trend is determined by enumeration method, and the objective function is taken as

[0159] min||f r -ω1 / 2π||

[0160] Thus the modal number k is determined;

[0161] The expression of roller rotation frequency is:

[0162]

[0163] Among them, f r 、f c 、f i They are the rotation frequencies of the roller, the center of the roller around the axis, and the inner ring, respectively. i and d are the inner ring and roller diameters respectively;

[0164] Module M2.3: Remove the intrinsic mode components that characterize the trend term of roller motion, reconstruct the remaining intrinsic mode components, and obtain the main bearing fault monitoring vibration data with the trend term removed.

[0165] Module M3: Establish a dual-stream multi-channel deep convolutional neural network to analyze triaxial vibration signals, including:

[0166] Module M3.1: The vibration data of the roller x-axis and y-axis with the motion trend removed by the variational mode decomposition method and the original z-axis data are used to build a fault data set. One part is enhanced by overlapping sampling to generate training samples, and the other part is generated by non-overlapping sampling to generate test samples. The samples are divided into training set, validation set and test set in a ratio of 7:3:1.

[0167] Module M3.2: Construct a dual-stream multi-channel deep convolutional neural network, including a dual-channel deep convolutional neural network and a single-channel deep convolutional neural network. Input the trend-free radial vibration signals of the x-axis and y-axis into the dual-channel deep convolutional neural network to train multi-directional radial vibration features. The axial data in the z-axis direction is trained separately through a single-channel deep convolutional neural network to obtain the axial vibration features.

[0168] Module M3.3: The multi-directional radial vibration and axial vibration features of the roller during rotation are fused through the fully connected layer of the neural network, and the Softmax logistic regression function is used to calculate the probability distribution of multi-classification tasks.

[0169] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for diagnosing main bearing faults of a tunnel boring machine based on roller status, characterized in that: include: Step S1: Acquire the three-axis vibration data of the main bearing roller; Step S2: removing the motion trend term of the sensor data in the roller by using the variational mode decomposition method; Step S3: Establish a dual-stream multi-channel deep convolutional neural network to train the roller triaxial vibration data.

2. The method for diagnosing main bearing faults of a tunnel boring machine based on roller status according to claim 1, characterized in that: An accelerometer and a wireless transmission module are integrated inside the roller.

3. The method for diagnosing main bearing faults of a tunnel boring machine based on roller status according to claim 1, characterized in that: The step S2 comprises: Step S2.1: Decompose the collected vibration signal of the main bearing roller to obtain k eigenmodal components u k (t), calculate the central angular frequency ω of each component k ; The implementation process includes: The analytical signal of the mode is obtained by Hilbert transform, and its one-sided spectrum is calculated. The analytical signal is multiplied by the exponential term Make the center frequency band ω k Shift to baseband and calculate bandwidth The equation is: In the formula, * represents the convolution operation, It means to find the gradient with respect to time t, j is the imaginary unit, δ(t) is the Dirichlet function, and k is the number of modes. The constrained optimization problem equation is established according to the component narrowband condition: Introducing the Lagrangian multiplier λ and the quadratic penalty factor α, by augmenting the Lagrangian function L({u k },{ω k },λ) transform the established constrained optimization problem into an unconstrained optimization problem: Among them, λ(t) is the Lagrangian operator, f(t) is the original signal, and the alternating direction multiplier method is used to find the saddle point of the augmented Lagrangian function. k and the center frequency ω k The update equation of the frequency domain optimization problem is obtained: ω k , and are the mode, center frequency, original signal and Fourier transform of Lagrange multiplier respectively. n is the number of iterations. The above steps are iterated continuously until the energy ratio of the residual signal to the original signal is less than the convergence threshold ε: Step S2.2: Compare the central angular frequency of the low-frequency eigenmode component obtained by decomposition with the calculated theoretical value of the roller rotation frequency to characterize the roller motion trend term; The enumeration method is used to determine the approximation between the low-frequency mode and the roller trend, and the objective function is taken as min‖f r -ω1 / 2π‖ Thus the modal number k is determined; The expression of roller rotation frequency is: Among them, f r 、f c 、f i They are the rotation frequencies of the roller, the center of the roller around the axis, and the inner ring, respectively. i and d are the inner ring and roller diameters respectively; Step S2.3: Remove the eigenmode components that characterize the trend term of roller motion, reconstruct the remaining eigenmode components, and obtain the main bearing fault monitoring vibration data with the trend term removed.

4. The method for diagnosing main bearing faults of a tunnel boring machine based on roller status according to claim 1, characterized in that: The step S3 comprises the following sub-steps: Step S3.1: The vibration data of the roller x-axis and y-axis with the motion trend item removed by the variational mode decomposition method and the original z-axis data are used to jointly establish a fault data set. One part is enhanced by overlapping sampling method to generate training samples, and the other part is generated by non-overlapping sampling to generate test samples. The samples are divided into training set, verification set and test set in a ratio of 7:3:1; Step S3.2: construct a dual-stream multi-channel deep convolutional neural network, including a dual-channel deep convolutional neural network and a single-channel deep convolutional neural network, input the trend-free radial vibration signals of the x-axis and y-axis into the dual-channel deep convolutional neural network to train multi-directional radial vibration features, and train the axial data in the z-axis direction through a single-channel deep convolutional neural network to obtain the axial vibration features; Step S3.3: The multi-directional radial vibration and axial vibration features of the roller during rotation are fused through the fully connected layer of the neural network, and the Softmax logistic regression function is used to calculate the probability distribution of the multi-classification task.

5. The method for diagnosing main bearing faults of a tunnel boring machine based on roller status according to claim 4, characterized in that: The structure of the neural network includes an input layer, a convolutional layer, a pooling layer and a fully connected layer; In the convolution layer, a convolution operator is used to convert the one-dimensional vibration data into a feature map. Each convolution kernel captures the corresponding feature and slides through the input data with the same weight and fixed step size. The calculation formula of the convolution layer is: In the formula, z (l) is the feature map of the lth layer, a (l-1) Input data for the l-1th layer, convolution kernel ω (l) As the weight vector, b (l) is bias; Use ReLU(x)=max(0,x) non-restrictive unit as activation function to get non-linear expression of convolutional layer output; In the pooling layer, the obtained feature map is divided into multiple regions, and each region is downsampled using the maximum pooling method: for region R of feature map d m , select the neuron x in this area i The average output The expression is as follows:

6. The method for diagnosing main bearing faults of a tunnel boring machine based on roller status according to claim 4, characterized in that: The function expression of the probability distribution is: Among them, p i is the predicted probability of the corresponding classification output node, z l (i) is the output value of the i-th node, and n is the number of output nodes, that is, the number of classification categories.

7. A roller state-based tunnel boring machine main bearing fault diagnosis system, characterized in that: include: Module M1: Obtain the three-axis vibration signal of the main bearing roller; Module M2: Remove the motion trend term of the sensor data inside the roller by using the variational mode decomposition method; Module M3: Establish a dual-stream multi-channel deep convolutional neural network to train roller triaxial vibration data.

8. The roller state-based fault diagnosis system for a tunnel boring machine main bearing according to claim 7, characterized in that: An accelerometer and a wireless transmission module are integrated inside the roller.

9. The method for diagnosing faults of a main bearing of a tunnel boring machine based on roller status according to claim 7, characterized in that: The module M2 comprises: Module M2.1: Decompose the collected vibration signal of the main bearing roller to obtain k intrinsic mode components u k (t), calculate the central angular frequency ω of each component k ; The implementation process includes: The analytical signal of the mode is obtained by Hilbert transform, and its one-sided spectrum is calculated. The analytical signal is multiplied by the exponential term Make the center frequency band ω k Shift to baseband and calculate bandwidth The equation is: In the formula, * represents the convolution operation, It means to find the gradient with respect to time t, j is the imaginary unit, δ(t) is the Dirichlet function, and k is the number of modes. The constrained optimization problem equation is established according to the component narrowband condition: Introducing the Lagrangian multiplier λ and the quadratic penalty factor α, by augmenting the Lagrangian function L({u k },{ω k },λ) transform the established constrained optimization problem into an unconstrained optimization problem: Among them, λ(t) is the Lagrangian operator, f(t) is the original signal, and the alternating direction multiplier method is used to find the saddle point of the augmented Lagrangian function. k and the center frequency ω k The update equation of the frequency domain optimization problem is obtained: ω k , and are the mode, center frequency, original signal and Fourier transform of Lagrange multiplier respectively. n is the number of iterations. The above steps are iterated continuously until the energy ratio of the residual signal to the original signal is less than the convergence threshold ε: Module M2.2: Compare the central angular frequency of the low-frequency eigenmodal component obtained by decomposition with the theoretical value of the roller rotation frequency calculated to characterize the roller motion trend term; The enumeration method is used to determine the approximation between the low-frequency mode and the roller trend, and the objective function is taken as min‖f r -ω1 / 2π‖ Thus the modal number k is determined; The roller rotation frequency expression is: Among them, f r 、f c 、f i They are the rotation frequencies of the roller, the center of the roller around the axis, and the inner ring, respectively. i and d are the inner ring and roller diameters respectively; Module M2.3: Remove the intrinsic mode components that characterize the trend term of roller motion, reconstruct the remaining intrinsic mode components, and obtain the main bearing fault monitoring vibration data with the trend term removed.

10. The method for diagnosing main bearing faults of a tunnel boring machine based on roller status according to claim 7, characterized in that: The module M3 includes the following submodules: Module M3.1: The vibration data of the roller x-axis and y-axis with the motion trend removed by the variational mode decomposition method and the original z-axis data are used to build a fault data set. One part is enhanced by overlapping sampling method to generate training samples, and the other part is generated by non-overlapping sampling to generate test samples. The samples are divided into training set, validation set and test set in a ratio of 7:3:

1. Module M3.2: Construct a dual-stream multi-channel deep convolutional neural network, including a dual-channel deep convolutional neural network and a single-channel deep convolutional neural network. Input the trend-free radial vibration signals of the x-axis and y-axis into the dual-channel deep convolutional neural network to train multi-directional radial vibration features. The axial data in the z-axis direction is trained separately through a single-channel deep convolutional neural network to obtain the axial vibration features. Module M3.3: The multi-directional radial vibration and axial vibration features of the roller during rotation are fused through the fully connected layer of the neural network, and the Softmax logistic regression function is used to calculate the probability distribution of multi-classification tasks.

Images