Method and system for electric gun platform bearing life prediction

Abstract

The application discloses a kind of electric shooting platform bearing life prediction method and system, to solve the existing electric shooting platform bearing life prediction method needs a lot of working condition matching full life cycle failure data, and in actual production, the bearing burst failure sample is few, and degradation feature is disturbed by multiple parameters coupling, leading to the life deviation of final prediction big technical problem.Method includes obtaining the real vibration signal data of electric shooting platform bearing, and the real vibration signal data is preprocessed, to generate final denoising signal data;Using dynamic equation set according to the maximum amplitude corresponding to real vibration signal data, the normalized root mean square maximum value corresponding to final denoising signal data, output the twin vibration signal data corresponding to real vibration signal data;Using virtual-real fusion based on Transformer-LMU model prediction network according to real vibration signal data, twin vibration signal data carries out life prediction, and outputs the predicted life value of electric shooting platform bearing.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment performance evaluation technology, and in particular to a method and system for predicting the life of bearings on an electric shooting platform. Background Technology

[0002] The bearing of the electric injection unit is a core load-bearing component of the injection system in injection molding equipment. It directly bears the axial impact load during the injection process, the continuous load during the holding pressure stage, and the dynamic load generated by frequent start-stop operations. Its performance directly determines the motion accuracy, repeatability, and overall stability of the injection unit. This bearing operates under harsh conditions such as high dynamic loads, periodic holding pressure stress, start-stop impacts, and localized temperature rises. Inevitably, it suffers gradual degradation problems such as rolling element wear, raceway fatigue spalling, and cage wear. Sudden failure can lead to minor issues like injection position deviation and holding pressure fluctuations, resulting in uneven plasticization, out-of-tolerance product dimensions, or flash defects. More serious consequences include injection unit jamming, drive motor overload, equipment damage, production line downtime, and significant economic losses.

[0003] Existing research has identified four stages in the service performance of electric injection molding machine bearings: healthy, damage initiation, rapid damage evolution, and failure. When the bearing is in a healthy state, the monitored vibration signal is stable and has a low amplitude. When the bearing experiences initial damage and gradually evolves to the failure stage, the monitored vibration signal exhibits significant peak fluctuations and frequency characteristic changes. Accurate prediction and assessment of the remaining service life of electric injection molding machine bearings during this process is crucial for preventing sudden equipment safety accidents and developing scientific maintenance plans. Therefore, life prediction of electric injection molding machine bearings during their service life is an important task for ensuring the stable and efficient operation of injection molding equipment and improving product quality consistency.

[0004] Most existing methods for predicting the life of bearings on electric shooting platforms are based on pure data methods. However, pure data methods require a large amount of full life cycle failure data with matching operating conditions. In actual production, there are very few samples of sudden failure of such bearings, and the degradation characteristics are affected by the coupling interference of multiple parameters, resulting in a large deviation in the final predicted life. Summary of the Invention

[0005] This invention provides a method and system for predicting the lifespan of electric shooting stage bearings. It addresses the technical problem that existing methods for predicting the lifespan of electric shooting stage bearings require a large amount of full life-cycle failure data with matching operating conditions. However, in actual production, there are very few samples of sudden failures of such bearings, and the degradation characteristics are affected by the coupling interference of multiple parameters, resulting in large deviations in the final predicted lifespan.

[0006] The first aspect of this invention provides a method for predicting the life of an electric shooting platform bearing, comprising:

[0007] Acquire the actual vibration signal data of the electric shooting platform bearing, and preprocess the actual vibration signal data to generate the final noise-reduced signal data;

[0008] The system of dynamic equations is used to output twin vibration signal data corresponding to the real vibration signal data based on the maximum amplitude value corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final noise-reduced signal data.

[0009] A Transformer-LMU model prediction network based on virtual-real fusion is used to predict the lifespan of the electric launcher bearing based on the real vibration signal data and the twin vibration signal data.

[0010] Optionally, the preprocessing of the real vibration signal data to generate the final denoised signal data includes:

[0011] Gaussian white noise is added to the real vibration signal data to generate noisy signal data;

[0012] Integrated empirical mode decomposition is used to iteratively decompose the noisy signal data, and the multiple IMF components and residual terms corresponding to the noisy signal data are output.

[0013] Perform wavelet threshold denoising on each of the IMF components to generate denoised IMF components corresponding to each of the IMF components;

[0014] Perform discrete wavelet transform on each of the denoised IMF components and output multiple detail coefficients corresponding to each of the denoised IMF components;

[0015] A soft threshold shrinkage function is used to determine multiple shrinkage coefficients corresponding to each of the denoised IMF components based on multiple detail coefficients corresponding to each of the denoised IMF components.

[0016] Perform discrete wavelet inverse transform on multiple shrinkage coefficients corresponding to each of the denoised IMF components to generate reconstructed and denoised IMF components corresponding to each of the denoised IMF components.

[0017] The reconstructed and denoised IMF components and residual terms are combined to generate the final denoised signal data.

[0018] Optionally, the step of using the dynamic equation set to output twin vibration signal data corresponding to the real vibration signal data based on the maximum amplitude value corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final denoised signal data includes:

[0019] The size conversion coefficient is calculated based on the maximum amplitude value corresponding to the actual vibration signal data and the normalized root mean square maximum value corresponding to the final noise reduction signal data.

[0020] Multiply the size conversion coefficient by the normalized root mean square maximum value corresponding to the final noise-reduced signal data, and output the defect size corresponding to the final noise-reduced signal data.

[0021] Based on the defect size, calculate the axially accumulated roller contact force;

[0022] The system of dynamic equations is used to output twin vibration signal data corresponding to the actual vibration signal data based on the axially accumulated roller contact force.

[0023] Optionally, the Transformer-LMU model prediction network based on virtual-real fusion includes an embedding layer, a Transformer-LMU neural network, and a prediction output layer; the step of using the Transformer-LMU model prediction network based on virtual-real fusion to predict the lifetime based on the real vibration signal data and the twin vibration signal data, and outputting the predicted lifetime value of the electric shooting platform bearing, includes:

[0024] An embedding layer is used to embed features into the real vibration signal data and the twin vibration signal data respectively, and the real data feature vector corresponding to the real vibration signal data and the twin data feature vector corresponding to the twin vibration signal data are output.

[0025] The real data feature vector and the twin data feature vector are normalized respectively, and the normalized real data feature vector and the normalized twin data feature vector are output. The normalized real data feature vector and the normalized twin data feature vector are then concatenated to generate a fused feature.

[0026] The Transformer-LMU neural network is used to perform feature splitting and enhancement based on the fused features, and outputs real data-enhanced feature vectors and twin data-enhanced feature vectors;

[0027] The predicted output layer performs predictions based on the real data augmented feature vector and the twin data augmented feature vector, and outputs the predicted life value of the electric launcher bearing.

[0028] Optionally, the step of using an embedding layer to perform feature embedding on the real vibration signal data and the twin vibration signal data respectively, and outputting the real data feature vector corresponding to the real vibration signal data and the twin data feature vector corresponding to the twin vibration signal data, includes:

[0029] The real vibration signal data and the twin vibration signal data are sequentially subjected to convolution, normalization, and nonlinear mapping to output the real data convolution output and the twin data convolution output.

[0030] The convolutional outputs of the real data and the twin data are normalized respectively to generate feature vectors of the real data and twin data.

[0031] Optionally, the step of using a Transformer-LMU neural network to perform feature splitting and enhancement based on the fused features, and outputting real data-enhanced feature vectors and Siamese data-enhanced feature vectors, includes:

[0032] Multi-head attention calculation is performed on the fused features to generate a multi-head attention output;

[0033] The multi-head attention output is input into the feedforward network for feature splitting, and the outputs real data feedforward feature vector and twin data feedforward feature vector are output.

[0034] Based on the feedforward feature vectors of real data and twin data, the final output feature vectors of real data and twin data are determined.

[0035] A cross-attention mechanism is used to enhance the features of the final output feature vector of the real data and the final output feature vector of the twin data, generating enhanced feature vectors of the real data and twin data.

[0036] Optionally, the step of predicting and outputting the predicted life value of the electric launcher bearing by the prediction output layer based on the real data augmented feature vector and the twin data augmented feature vector includes:

[0037] One-dimensional adaptive max pooling is performed on the real data augmented feature vector and the twin data augmented feature vector respectively, and the max pooled real data augmented feature vector and the max pooled twin data augmented feature vector are output.

[0038] The max-pooling real data augmented feature vector and the max-pooling Siamese data augmented feature vector are concatenated to output a concatenated feature vector.

[0039] The concatenated feature vectors are sequentially subjected to linear transformation, normalization, and nonlinear mapping to generate hidden layer features;

[0040] The hidden layer features are sequentially subjected to linear transformation and nonlinear mapping to output the predicted life value of the electric shooting platform bearing.

[0041] A second aspect of the present invention provides an electric shooting platform bearing life prediction system, comprising:

[0042] The acquisition module is used to acquire the actual vibration signal data of the electric shooting platform bearing, and to preprocess the actual vibration signal data to generate the final noise-reduced signal data.

[0043] The output module is used to output twin vibration signal data corresponding to the real vibration signal data by using the dynamic equation set based on the maximum amplitude value corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final noise-reduced signal data.

[0044] The prediction module is used to use a Transformer-LMU model prediction network based on virtual-real fusion to predict the lifespan of the electric launcher bearing based on the real vibration signal data and the twin vibration signal data, and output the predicted lifespan value of the electric launcher bearing.

[0045] A computer device provided in a third aspect of the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the electric launcher bearing life prediction method as described in any of the preceding claims.

[0046] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the steps of the electric launcher bearing life prediction method as described in any of the preceding claims.

[0047] As can be seen from the above technical solutions, the present invention has the following advantages:

[0048] The present invention provides a method for predicting the lifespan of an electric shooting platform bearing. It acquires real vibration signal data of the electric shooting platform bearing, preprocesses the real vibration signal data to generate final denoised signal data, and uses a set of dynamic equations to output twin vibration signal data corresponding to the real vibration signal data based on the maximum amplitude of the real vibration signal data and the normalized root mean square maximum value of the final denoised signal data. A Transformer-LMU model prediction network based on virtual-real fusion is then used to predict the lifespan of the electric shooting platform bearing based on the real vibration signal data and the twin vibration signal data, outputting the predicted lifespan value. Based on this scheme, the twin vibration signal generated by the present invention effectively supplements the data source gap of extremely few sudden bearing failure samples in actual production, and can build a comprehensive signal analysis foundation without over-reliance on scarce real failure samples. Simultaneously, the Transformer-LMU model prediction network based on virtual-real fusion can accurately capture and decouple the influence of multi-parameter coupling interference on degradation characteristics during bearing degradation, thereby improving the accuracy of lifespan prediction. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of the steps in a method for predicting the life of an electric shooting platform bearing provided in Embodiment 1 of the present invention;

[0051] Figure 2 This is an overall framework diagram of an electric launcher bearing life prediction method provided in Embodiment 1 of the present invention;

[0052] Figure 3 This is a flowchart illustrating a method for predicting the lifespan of an electric launcher bearing, provided in Embodiment 2 of the present invention.

[0053] Figure 4 This is a structural block diagram of an electric shooting platform bearing life prediction system provided in Embodiment 3 of the present invention. Detailed Implementation

[0054] This invention provides a method and system for predicting the lifespan of electric shooting platform bearings. It addresses the technical problem that existing methods for predicting the lifespan of electric shooting platform bearings require a large amount of full life-cycle failure data with matching operating conditions. However, in actual production, there are very few samples of sudden failures of such bearings, and the degradation characteristics are affected by the coupling interference of multiple parameters, resulting in large deviations in the final predicted lifespan.

[0055] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0056] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a method for predicting the lifespan of an electric shooting platform bearing, as provided in Embodiment 1 of the present invention.

[0057] This invention provides a method for predicting the life of an electric shooting platform bearing, comprising:

[0058] Step 101: Obtain the actual vibration signal data of the electric shooting platform bearing, and preprocess the actual vibration signal data to generate the final noise-reduced signal data.

[0059] It should be noted that the actual vibration signal data of the electric shooting platform bearing includes multiple frequency components and noise interference. In order to extract the bearing damage characteristics, this invention uses ensemble empirical mode decomposition (EEMD) combined with an adaptive wavelet threshold denoising method to decompose and denoise the data. The root mean square (RMS) value of the vertical channel vibration data is extracted and mapped to the defect size. The resulting degradation size can be regarded as the evolution law of bearing service performance. Based on this, the service performance state of the bearing is divided into four states: healthy, damage onset, rapid damage evolution, and failure.

[0060] Specifically, step 101 may include the following sub-steps S11-S17:

[0061] Step S11: Add Gaussian white noise to the real vibration signal data to generate noisy signal data;

[0062] Step S12: Iteratively decompose the noisy signal data using integrated empirical mode decomposition, and output multiple IMF components and residual terms corresponding to the noisy signal data;

[0063] Step S13: Perform wavelet threshold denoising on each IMF component to generate the denoised IMF component corresponding to each IMF component.

[0064] Step S14: Perform discrete wavelet transform on each denoised IMF component and output multiple detail coefficients corresponding to each denoised IMF component.

[0065] Step S15: Use a soft threshold shrinkage function to determine multiple shrinkage coefficients corresponding to each denoised IMF component based on multiple detail coefficients corresponding to each denoised IMF component.

[0066] Step S16: Perform discrete wavelet inverse transform on the multiple shrinkage coefficients corresponding to each denoised IMF component to generate the reconstructed and denoised IMF component corresponding to each denoised IMF component.

[0067] Step S17: Combine the reconstructed and denoised IMF components and residual terms to generate the final denoised signal data.

[0068] It should be noted that, firstly, the raw data (real vibration signal data) Adding Gaussian white noise yields noisy data (noisy signal data). Then, multiple EMD iterative decompositions are performed, and the calculation is shown in equation (1):

[0069] (1)

[0070] in, The order of the derived intrinsic mode functions (IMFs) is given by the factorization. For the first The first-order IMF (Intrinsic Mode Function) component; This refers to the residual components (residual terms).

[0071] Next, for each Wavelet thresholding is applied to eliminate high-frequency noise while preserving transient impulse characteristics. Perform Discrete Wavelet Transform (DWT) as shown in Equation (2):

[0072] (2)

[0073] in, The number of layers in the discrete wavelet transform. For the first Layer detail factor.

[0074] Furthermore, noise standard deviation estimation is introduced to design an adaptive threshold. The soft-thresholding function is used to process the detail coefficients, and its calculation is shown in Equation (3). The processed coefficients (shrinkage coefficients) are then subjected to Inverse Discrete Wavelet Transform (IDWT) to reconstruct the denoised IMF (i.e., the reconstructed denoised IMF components), and finally combined with the residual term to obtain the final denoised data (i.e., the final denoised signal data), as shown in Equation (4).

[0075] (3)

[0076] (4)

[0077] in, It is a noise standard deviation estimate. It is the median of the absolute values ​​of the differences between the detail coefficients of the wavelet transform and their median. 0.6745 is the ratio of the absolute deviation of the median to the standard deviation of the normal distribution. It is the IMF component currently being processed. The length. This is the threshold factor, set to 0.5 here. It is the first after processing by the soft threshold shrinkage function. Layer detail factor (i.e. shrinkage factor) It is the reconstructed and denoised IMF, that is, the i-th reconstructed and denoised IMF component at time t; This represents the final denoised signal data at time t.

[0078] Step 102: Using the dynamic equations, based on the maximum amplitude value corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final noise-reduced signal data, output the twin vibration signal data corresponding to the real vibration signal data.

[0079] Specifically, step 102 may include the following sub-steps S21-S24:

[0080] Step S21: Calculate the size conversion coefficient based on the maximum amplitude value corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final noise reduction signal data.

[0081] Step S22: Multiply the size conversion coefficient and the normalized root mean square maximum value corresponding to the final noise reduction signal data, and output the defect size corresponding to the final noise reduction signal data;

[0082] Step S23: Calculate the axially accumulated roller contact force based on the defect size;

[0083] Step S24: Using the dynamic equation set, output the twin vibration signal data corresponding to the real vibration signal data based on the axially accumulated roller contact force.

[0084] It should be noted that this invention uses a method of matching the damage dynamic response with the actual response to estimate the simulated defect size. When the peak value of the damage dynamic response under a simulated defect size matches the peak value of the actual vibration data, the corresponding defect size can be regarded as the simulated size of the actual defect. The maximum value of the defect size is simulated using the damage dynamic response. The obtained maximum defect size (i.e., the maximum amplitude corresponding to the actual vibration signal data) is mapped to the normalized RMS maximum value of the axial vibration data (i.e., the normalized root mean square maximum value corresponding to the final denoised signal data). A size conversion coefficient is calculated to establish the mapping relationship between RMS and the simulated defect size, thus obtaining the evolution law of the bearing defect size throughout its entire life cycle. Although the predicted defect size here cannot represent the actual defect size of the bearing, it can effectively reflect the evolution law of bearing defects, i.e., the service performance evolution trend. When the defect size begins to appear, the bearing state changes from healthy to damaged; when the size reaches a rapid growth period after a slow growth period, the state changes from damaged to rapidly evolving; when the size reaches its maximum value, the bearing fails.

[0085] Furthermore, regarding the construction process of the dynamic equations, in the physical-driven digital twin stage, the electric shooting platform bearing (thrust roller bearing), under axial load, can be simplified into a two-degree-of-freedom planar vibration system, defined as follows: Axial displacement at time t radial displacement The corresponding axial velocity and radial velocity The corresponding axial acceleration and radial acceleration The dynamic equation is shown in equation (5):

[0086] (5)

[0087] in, For quality, For axial and radial damping coefficients, For radial stiffness, For axial and radial loads, This refers to the axially accumulated roller contact force. The number of rollers, generated by the contact between all rollers and the inner and outer rings; axial acceleration. and radial acceleration These are represented as the axial and radial components of the twin vibration signal data, respectively, which together constitute the complete twin vibration signal data.

[0088] To capture the nonlinear impact characteristics during damage entry and exit, the roller angle position and spindle speed were calculated. Cage speed Roller rotation speed The calculation is shown in equation (6). Main spindle speed Where is the roller radius, For bearing pitch diameter, , The inner diameter of the bearing. The outer diameter of the bearing. This represents the number of rollers. Time of the first The instantaneous polar angle position of the roller The calculation is shown in equation (7):

[0089] (6)

[0090] (7)

[0091] Furthermore, using The system determines whether each roller interacts with the defect area. The contact determination logic is as follows for different defect locations:

[0092] (1) When the damage location is on the outer ring, the defect area is fixed. When equation (8) is satisfied, it can be considered that the first ring is the outer ring. The roller is now in the defect area, and the roller enters the corner area of ​​the fixed defect on the outer ring, increasing the contact depth. .

[0093] (8)

[0094] in, For the modulo operation, the first... One roller in absolute angle of time right Mold taking, used to... Mapped to interval; This refers to the width of the defect angle; For bearing pitch diameter; This represents the defect size.

[0095] (2) When the damage location is on the inner ring, the defect area rotates with the inner ring. When equation (9) is satisfied, it can be considered that the first... When a roller is in the defect area, that is, when the angular distance between the roller and the center of the rotating defect is less than half the width of the defect, contact occurs.

[0096] (9)

[0097] in, Calculate angular distance; For defect size; This is a modulo operation used to handle circularity; Calculate the minimum angular distance between the roller contact point and the center of the defect in the rotating inner ring. This indicates the angular position of the roller relative to the inner ring of rotation. Perform a half-turn phase adjustment to move the reference point from the roller center to the actual contact point, thereby accurately determining whether the contact point encounters a defect.

[0098] (3) When the damage location is on a roller, the defect area rotates with the roller. When a roller has a defect, it is necessary to determine whether the defect is directly opposite the contact area. Due to line contact, the roller makes contact once every half rotation (180°). Therefore, when equation (10) is satisfied, it can be considered that the roller is in the defect area. When the roller defect rotates to the contact area (the upper and lower contact areas), it affects the contact force.

[0099] (10)

[0100] in, For defect size; For the modulo operation, the first... One roller in absolute angle of rotation at any moment right Take the modulus and determine if the result falls within the range of 1 / 2. Interval.

[0101] (4) When the damage location is on the rotating frame, the defect area rotates with the rotating frame. The defect area fixed to the cage is judged according to the same logic as the outer ring damage, as shown in equation (10). At this time, the defect causes the roller to completely lose contact. Depending on whether the roller is in the defect area, the time-varying defect depth is defined according to the above judgment criteria as shown in equation (11), where For defect depth, The angle width is, that is, when the first... When a roller is within the defect area, the depth is taken as follows: .

[0102] (11)

[0103] Furthermore, by explicitly introducing damage geometry into the contact model, the collision effect is accurately reproduced during numerical integration, thereby realizing the data evolution of the entire process from health to failure. Damage forms pits on the surface, and the detailed calculation of their depth is as follows:

[0104] (1) When the damage occurs on the outer or inner ring, the defect shape is approximately circularly tangent by utilizing the geometric relationship of the circular arc. and The calculation is shown in equation (12):

[0105] (12)

[0106] in, For defect size, The pitch circle radius is used for defects on the outer ring (the raceway curvature radius is used for defects on the outer ring, and the pitch circle radius is used for defects on the inner ring). The theoretical maximum depth of the geometric defect is taken as the defect depth based on empirical scaling and consideration of material wear, etc. 80%.

[0107] (2) When damage occurs on the roller, the defect shape is also approximately circular. and The calculation is shown in equation (13):

[0108] (13)

[0109] in, The radius of the roller is given by an empirical coefficient based on the curvature of the roller surface, and the defect depth is taken as... 70%; This represents the defect size.

[0110] (3) When damage occurs on the cage, the defects are mostly holes or gaps. Therefore, the depth is approximated by establishing a proportional relationship based on the defect size, and its calculation is shown in Equation (14):

[0111] (14)

[0112] in, To maintain the radius of the frame section circle; This represents the defect size.

[0113] Furthermore, Hertzian contact force theory can accurately describe the stress-deformation relationship in the elastic contact zone, introducing nonlinear power-law behavior and providing a physical basis for simulating damage impact and vibration aggravation. According to this theory, the contact stiffness of a single rolling element with the inner and outer rings... The calculation is shown in equation (15):

[0114] (15)

[0115] in, and The elastic modulus and Poisson's ratio of the material. For the length of the roller, Where is the radius of curvature of the raceway. For reference contact deformation.

[0116] After obtaining the contact stiffness, the roller contact force is calculated as shown in Equation (16). The contact force is linearly related to the relative deformation, and there is no contact force for negative deformation (gap).

[0117] (16)

[0118] in, This represents the axial deformation, which is the displacement relative to a stationary, unloaded location. The value is the defect depth at the i-th roller. If there is a defect at this position, the value is greater than zero; otherwise, it is zero. This ensures that contact force is only generated during deformation where the roller is pressed into the depth of the defect. The total force is calculated based on the number of effective rollers. The axially accumulated roller contact force is obtained by performing average compensation. Its calculation is shown in equation (17):

[0119] (17)

[0120] Based on the above, the contact force is projected onto the axial and radial degrees of freedom to obtain the final set of dynamic equations, as shown in equation (18). This two-degree-of-freedom damped spring-mass model couples nonlinear contact force and damage geometry, retaining physical interpretability while also outputting multi-channel displacement velocity and acceleration through numerical integration, thus realizing a digital twin of the entire life cycle of the electric launcher bearing.

[0121] (18)

[0122] in, This refers to the number of rollers; This represents the radial contact stiffness.

[0123] Furthermore, after obtaining the twin vibration signal data using the dynamic equations, the virtual-real combined life prediction modeling stage can begin. Life prediction is performed under conditions of damage occurrence and rapid damage evolution. Addressing the modeling challenges of virtual-real data fusion and long-term life degradation modeling, this invention employs a Transformer-LMU model prediction network based on virtual-real fusion for life prediction of the electric shooting platform bearing. Real sensor data contains actual operating noise and sudden damage characteristics, but is susceptible to interference; digital twin simulation data is generated based on a physical model, exhibiting clear degradation patterns but containing model biases. The two complement each other, improving the model's generalization ability to life degradation patterns and ensuring stable prediction.

[0124] Step 103: Using a Transformer-LMU model prediction network based on virtual-real fusion, the life prediction is performed based on real vibration signal data and twin vibration signal data, and the predicted life value of the electric shooting platform bearing is output.

[0125] The prediction network of the Transformer-LMU (Transformer-Legendre Memory Unit) model based on virtual-real fusion includes an embedding layer, a Transformer-LMU neural network, and a prediction output layer.

[0126] It should be noted that the Transformer architecture, relying on a multi-head self-attention mechanism, can effectively capture long-range time-frequency dependencies in vibration sequences, making it particularly suitable for modeling sparse long-sequence information corresponding to early, weak degradation features. The LMU uses ordinary differential equations to describe the evolution of states over time through a state-space model, introducing Legendre polynomials to construct a latent space representation, effectively capturing time delay features while balancing model simplicity and dynamic state modeling accuracy. This invention improves upon the Transformer model, proposing a Transformer-LMU lifetime prediction model based on virtual-real fusion. This model consists of three parts: an embedding layer, a Transformer-LMU neural network, and a prediction output layer.

[0127] Specifically, step 103 may include the following sub-steps S31-S34:

[0128] Step S31: Use an embedding layer to embed features into the real vibration signal data and the twin vibration signal data respectively, and output the real data feature vector corresponding to the real vibration signal data and the twin data feature vector corresponding to the twin vibration signal data.

[0129] Further, step S31 may include the following sub-steps S311-S312:

[0130] Step S311: Perform convolution, normalization, and nonlinear mapping on the real vibration signal data and twin vibration signal data respectively, and output the real data convolution output and twin data convolution output.

[0131] Step S312: Normalize the convolutional output of the real data and the convolutional output of the twin data respectively to generate the feature vector of the real data and the feature vector of the twin data.

[0132] It should be noted that the embedding layer uses two sets of convolutional layers to map the channels of real and twin data (i.e., real vibration signal data and twin vibration signal data) respectively, adjusts the channel dimensions, preserves temporal information, and accelerates convergence by using the normalization layer BatchNorm (BN) and the ReLU activation function, while introducing nonlinearity. Subsequently, the features of each time step are independently normalized (LayerNorm) to eliminate the distribution bias between real and twin data, which helps in the subsequent training of the Transformer and reduces the inter-domain distribution shift. Its implementation is shown in Equations (19) and (20):

[0133] (19)

[0134] (20)

[0135] in, These are the real and twin data subtensors, namely, real vibration signal data and twin vibration signal data; These are the weight and bias matrices corresponding to the real vibration signal data and the twin vibration signal data, respectively; The output of the convolutional layer includes both real data convolutional output and twin data convolutional output. They are respectively The real-time data feature vector and the twin data feature vector represent the output of the LayerNorm layer. For the mean and variance of the input dimension, For learnable scaling and translation parameters, This is a minimum value to prevent division by zero.

[0136] Step S32: Normalize the real data feature vector and the twin data feature vector respectively, output the normalized real data feature vector and the normalized twin data feature vector, and concatenate the normalized real data feature vector and the normalized twin data feature vector to generate the fused feature.

[0137] Step S33: Use the Transformer-LMU neural network to perform feature splitting and enhancement based on the fused features, and output the real data enhanced feature vector and the twin data enhanced feature vector;

[0138] Furthermore, step S33 may include the following sub-steps S331-S334:

[0139] Step S331: Perform multi-head attention calculation on the fused features to generate multi-head attention output;

[0140] Step S332: Input the multi-head attention output into the feedforward network for feature splitting, and output the real data feedforward feature vector and the twin data feedforward feature vector;

[0141] Step S333: Based on the feedforward feature vector of real data and the feedforward feature vector of twin data, determine the final output feature vector of real data and the final output feature vector of twin data;

[0142] Step S334: Use a cross-attention mechanism to enhance the features of the final output feature vector of the real data and the final output feature vector of the twin data, generating enhanced feature vectors of the real data and twin data.

[0143] It should be noted that Transformer-LMU comprises three parts: TransformerEncoder, LMU block, and cross-attention fusion. The fused features are obtained by concatenating the normalized features of real and Siamese data along a temporal sequence. The input to the TransformerEncoder layer captures long-term temporal dependencies and global cross-condition features through multi-head attention in parallel, which helps to mine multi-scale patterns in lifetime degradation data. The calculation of multi-head self-attention is shown in Equation (21):

[0144] ;(twenty one)

[0145] in, These are query, key, and value matrices, respectively. These are the linear mapping matrices corresponding to the query, key, and value matrices, respectively. For activation function, For each attention head, the key / query dimension, , For the number of attention heads, This provides multi-headed attention output.

[0146] The features are then encoded using a feedforward network to split them into real and twin branches. This is the output of the TransformerEncoder. These are the feedforward feature vectors of real data and the feedforward feature vectors of twin data, respectively.

[0147] Furthermore, in life prediction, it is necessary to capture both the global long-range dependencies of the data and preserve the dynamic evolution over continuous time. While Transformer focuses on the global perspective, LMU (Long-Term Memory) constructs a state-space model based on Legendre orthogonal polynomials, combining bilinear discretization and residual fusion to achieve efficient and stable long-term memory of complex vibration data. Real data carries measurement noise and environmental interference, while simulation data comes from physical models, is noise-free but may deviate from reality. Feeding these data streams into the LMU ensures that each data stream extracts continuous dynamics within its preferred "state space," enabling efficient modeling of temporal dynamics using a state-space model to describe the evolution of the electric launcher bearing's life state over time.

[0148] Each The sequence (i.e., the feedforward feature vector of real data and the feedforward feature vector of twin data) is input to a separate LMU module, which the LMU treats as a continuous-time input. The construction length is The memory window, defining the order With state vector ,in Each component corresponds to a Legendre basis. And the LMU's state transition matrix... and input mapping vector A set of filters was automatically designed for each domain, which can suppress high-frequency noise for real flow and compensate for the differences between the model and reality for twin flow. The calculation is shown in equation (22).

[0149] Meanwhile, in order to implement it in a digital system, the continuous differential equation (22) is transformed into a discrete difference equation. Bilinear (Tustin) discretization is introduced to ensure that the phase and amplitude characteristics of the discretized system are consistent with those of the original continuous system, thus avoiding distortion. and Mapping to discrete systems The matrix is ​​implemented as shown in equation (23). It achieves linear complexity. All historical inputs are aggregated and dynamically updated in discrete form as shown in Equation (24). Then, the original features are added and normalized. The output features retain the original data features and are injected with global temporal memory, as shown in Equation (25).

[0150] ;(twenty two)

[0151] ;(twenty three)

[0152] ;(twenty four)

[0153] (25)

[0154] in, Here is the state transition matrix. For the input mapping vector, Let be the matrix element in the i-th row and j-th column of the state transition matrix. For the i-th element in the input mapping vector, For row index, For column indexes, For order, ; The sampling interval; For discrete update matrices; To map the weights and biases to the output; for The state vector at any given time; Choose a vector for the output ; Selecting a state vector The first component is used as the output. ; From LMU state Through linear mapping and The memory enhancement vector obtained by the activation function; It is the first Each sample (vibration signal data, including real vibration signal data and twin vibration signal data) in The final output feature vector at time step 1 represents the 1st time step 2. Vibration signal data in The final output feature vector at each time step (including the final output feature vector of real data and the final output feature vector of twin data). For the first One sample in The continuous time input at each moment includes the real data feedforward feature vector and the twin data feedforward feature vector.

[0155] Furthermore, LMU has extracted high-quality temporal dynamics from real and twin data streams, but there is still a domain bias between the two. Cross-attention uses the real stream as the query and the simulation stream as the key / value pair, dynamically injecting information useful for real predictions from the simulation into the real representation, thus filling in the operational conditions that are difficult to observe in real data; similarly, the real stream can also help correct the deviation between the simulation stream and reality, as shown in Equation (26).

[0156] (26)

[0157] in, These are the real and twin feature vectors output by the LMU, i.e., the final output feature vectors of the real data and the final output feature vectors of the twin data. These are the enhanced fusion features, namely the real data enhanced feature vector and the twin data enhanced feature vector.

[0158] Step S34: The prediction output layer performs prediction based on the real data enhanced feature vector and the twin data enhanced feature vector to output the predicted life value of the electric shooting platform bearing.

[0159] Further, step S34 may include the following sub-steps S341-S344:

[0160] Step S341: Perform one-dimensional adaptive max pooling on the real data augmented feature vector and the twin data augmented feature vector respectively, and output the max pooled real data augmented feature vector and the max pooled twin data augmented feature vector.

[0161] Step S342: Concatenate the max-pooled real data augmented feature vector and the max-pooled Siamese data augmented feature vector, and output the concatenated feature vector;

[0162] Step S343: Perform linear transformation, normalization, and nonlinear mapping on the concatenated feature vectors in sequence to generate hidden layer features;

[0163] Step S344: Perform linear transformation and nonlinear mapping on the hidden layer features in sequence, and output the predicted life value of the electric shooting platform bearing.

[0164] It should be noted that, firstly, one-dimensional adaptive max pooling is performed on the real and twin feature sequences (real data augmented feature vector and twin data augmented feature vector), and then the two features are concatenated. The specific calculation is shown in Equation (27). The final lifetime prediction result is obtained through two fully connected layers and activation functions, and its implementation is shown in Equation (28).

[0165] (27)

[0166] (28)

[0167] in, They are respectively the sample dimensions (the first one) (samples), time dimension, and channel dimension; Maximum time step; This indicates the predicted lifespan value output, which is the predicted lifespan value of the electric shooting platform bearing. Features of the hidden layer; and This represents the weight matrix of the linear layer; and This represents the bias vector of the linear layer; express Activation function; To concatenate feature vectors; These are, respectively, max pooling of real data augmented feature vectors and max pooling of twin data augmented feature vectors.

[0168] It is worth mentioning that, during the model training phase, this invention calculates the loss by using the root mean square error loss function with the output result (i.e., the predicted lifetime value output by the untrained Transformer-LMU model prediction network based on virtual-real fusion) and the true lifetime label, and performs backpropagation to optimize and update the network parameters of the untrained Transformer-LMU model prediction network based on virtual-real fusion until the set number of training iterations is reached, thereby obtaining the trained Transformer-LMU model prediction network based on virtual-real fusion.

[0169] In this embodiment, the proposed Transformer-LMU model prediction network based on virtual-real fusion utilizes the Transformer's multi-head attention mechanism to achieve heterogeneous feature-level fusion and long-range dependency modeling of virtual twin data and measured vibration data. Combined with the LMU's state-space model, it efficiently captures the slow dynamic evolution of degradation trends in long vibration sequences, and achieves complementary enhancement of physical simulation and measured information through a bidirectional cross-attention mechanism. A Transformer-LMU lifetime prediction model based on virtual-real fusion was designed, achieving accurate prediction of the lifetime degradation process of electric launcher bearings.

[0170] For comparison of technical effectiveness, existing technologies can be referenced. Current life prediction faces dual technical barriers: traditional physical models struggle to quantify the real-time impact of dynamic processes (such as back pressure fluctuations and melt viscosity changes) on bearing contact stress; pure data methods require a large amount of full life-cycle failure data with matched operating conditions, but in actual production, there are very few samples of sudden bearing failures, and degradation characteristics are affected by multi-parameter coupling interference, resulting in weak model generalization ability and large prediction bias. To integrate the advantages of physical mechanisms and real-time data, and overcome the limitations of traditional physical models and pure data-driven methods, this invention proposes a virtual-real data combined life prediction method and system for electric shooting platform bearings, to achieve accurate prediction of their remaining service life.

[0171] To address the above problems, this invention proposes a method for predicting the life of an electric shooting platform bearing. Please refer to [link / reference]. Figure 2The process comprises three steps: data preprocessing and service performance assessment, physical-driven digital twin, and virtual-physical combined life prediction modeling. First, in the data preprocessing and service performance assessment stage, to address the life degradation distortion problem caused by environmental noise and multi-frequency interference in the original vibration data, an integrated empirical mode decomposition (EEMD) combined with a wavelet threshold denoising algorithm is used to remove noise and separate the intrinsic components of bearing damage impact. Based on the denoised axial vibration data, the root mean square (RMS) value is extracted to establish an explicit mapping relationship between it and the defect size evolution throughout the bearing's entire life cycle. This mapping is used as the law governing the bearing's service performance degradation throughout its entire life cycle, and the service performance state is divided into four states based on this degradation law: healthy, damage initiation, rapid damage evolution, and failure. Second, in the physical-driven digital twin stage, a nonlinear dynamic model of the bearing is established based on a two-degree-of-freedom damped spring-mass system. Combining Hertzian contact theory and damage geometric feature analysis, dynamic expansion parameters of defects are embedded into the motion differential equation. This equation is solved through numerical integration to reproduce the vibration response throughout the entire life cycle from a healthy state to complete failure. Finally, in the virtual-real combined life prediction modeling stage, bearing damage is observed for life prediction. A Transformer-LMU neural network is constructed. The Transformer utilizes a multi-head attention mechanism to align the distribution differences between virtual simulation data and measured vibration data in the time-frequency domain feature space, achieving feature-level fusion and long-range dependency modeling of heterogeneous data sources. The LMU (Legendre Memory Unit) is based on the State Space Model (SSM) and efficiently captures the slow dynamic evolution of degradation trends in trillion-step vibration sequences with linear computational complexity. A bidirectional cross-attention mechanism is designed to achieve complementary enhancement of physical simulation and measured information. The dual-path collaborative operation significantly improves the modeling accuracy of long-period degradation trajectories and the prediction performance of Remaining Useful Life (RUL).

[0172] In summary, this invention comprises three core steps: 1) Data preprocessing based on Integrated Empirical Mode Decomposition (EEMD) combined with adaptive wavelet thresholding denoising, followed by RMS-based defect size mapping and service performance evaluation; 2) Physically driven digital twin simulation based on a two-degree-of-freedom nonlinear dynamic model, Hertzian contact theory, damage geometry feature analysis, and time-varying defect depth calculation; 3) Virtual-real fusion remaining useful life (RUL) prediction modeling based on a Transformer-LMU neural network (including a bidirectional cross-attention mechanism). This method effectively integrates the advantages of physical mechanisms and real-time monitoring data, overcoming the limitations of a single model. Meanwhile, this invention establishes a two-degree-of-freedom damped spring-mass nonlinear dynamic model for bearings; calculates contact stiffness and time-varying contact force based on Hertzian contact theory; explicitly introduces the geometric features (size and location) of four types of damage—outer ring, inner ring, roller, and cage—into the model; defines a precise formula for calculating time-varying defect depth and a logic for judging contact between rollers and defect areas; embeds the time-varying defect depth into the contact force calculation model; and solves the motion differential equations through numerical integration, thereby accurately reproducing the full life cycle vibration response of bearings from a healthy state to complete failure, providing a high-quality virtual data source for virtual-real fusion. Furthermore, it proposes a Transformer-LMU neural network architecture for fusing physical simulation data and measured vibration data and predicting remaining life. The architecture includes: 1) A Transformer branch: utilizing a multi-head self-attention mechanism to align the distribution differences between virtual simulation data and measured vibration data in the time-frequency domain feature space, achieving feature-level fusion and long-range dependency modeling of heterogeneous data sources; 2) An LMU branch: based on a state-space model (SSM), efficiently capturing the slow dynamic evolution of degradation trends in long vibration sequences with linear computational complexity; 3) A bidirectional cross-attention fusion mechanism: using the real stream as the query and the twin stream as the key / value pair, achieving dynamic complementarity and enhancement between physical simulation and measured information. Furthermore, this invention proposes a bidirectional cross-attention compensation mechanism for fusing physical simulation data streams and measured vibration data streams. Specifically, this includes: using real data features as query vectors and simulation data features as key / value vectors, dynamically filtering physical law information from simulation data that is effective for real predictions through cross-attention calculation, injecting it into real data representation, and filling in observation blind spots caused by environmental interference; using simulation data features as query vectors and real data features as key / value vectors, and feeding back the actual monitored sudden damage features to the simulation stream through reverse cross-attention, correcting parameter deviations in the physical model in real time; and automatically adjusting the feature contribution ratio of virtual and real data in the time-frequency domain through the learnable weight matrix of the attention layer, achieving dynamic optimal fusion of physical mechanisms and measured data.

[0173] In this embodiment of the invention, a method for predicting the lifespan of an electric shooting platform bearing is provided. The method involves acquiring real vibration signal data of the electric shooting platform bearing, preprocessing the real vibration signal data to generate final denoised signal data, and using a set of dynamic equations to output twin vibration signal data corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final denoised signal data. A Transformer-LMU model prediction network based on virtual-real fusion is then used to predict the lifespan of the electric shooting platform bearing based on the real vibration signal data and the twin vibration signal data, outputting the predicted lifespan value. Based on the above scheme, the twin vibration signal generated by the dynamic equations effectively supplements the data source gap of very few sudden bearing failure samples in actual production, allowing for the construction of a comprehensive signal analysis foundation without over-reliance on scarce real failure samples. Simultaneously, the Transformer-LMU model prediction network based on virtual-real fusion can accurately capture and decouple the influence of multi-parameter coupling interference on degradation characteristics during bearing degradation, thereby improving the accuracy of lifespan prediction.

[0174] Please see Figure 3 , Figure 3 This is a flowchart illustrating a method for predicting the life of an electric shooting platform bearing according to Embodiment 2 of the present invention, including:

[0175] (1) Collect vibration data of electric shooting stage bearings: Use multi-channel vibration sensors to collect data of the electric shooting stage bearings throughout their entire life cycle, especially axial vibration data, to ensure the integrity and availability of the data.

[0176] (2) Data preprocessing: The original vibration data is denoised using a combined adaptive wavelet threshold denoising algorithm based on integrated empirical mode decomposition (EEMD) to extract the intrinsic mode components after denoising; the root mean square value (RMS) of the axial vibration data is calculated and normalized; the time series data is divided into multiple samples according to time order using the sliding window partitioning method, and then divided into training set, validation set and test set.

[0177] (3) Defect size mapping and service performance evaluation: The RMS value is extracted from the preprocessed denoised vibration data as a preliminary degradation index; the evolution law of the defect size throughout the bearing's life cycle is obtained by using the mapping relationship between the RMS value and the simulated defect size (established through peak matching of damage dynamic response and size conversion coefficient). The evolution law is used as the bearing service performance degradation trend, and its state is divided into four states: healthy, damage onset, rapid damage evolution, and failure.

[0178] (4) Physical-driven digital twin: Based on the two-degree-of-freedom damped spring-mass system, a nonlinear dynamic model of the bearing is established. Combining Hertzian contact theory, damage geometric feature analysis and time-varying defect depth calculation, the motion differential equation is solved by numerical integration to generate virtual vibration response data of the entire life cycle from healthy state to complete failure.

[0179] (5) Model Training: The bearing life is predicted under damage occurrence and rapid damage evolution states. The preprocessed measured vibration data training set, the corresponding virtual vibration data training set generated by physical simulation, and RUL tags are used as inputs; input to... Figure 1 In the Transformer-LMU-based virtual-real fusion lifetime prediction model shown, parameter optimization is performed during training using backpropagation with regression loss (root mean square error RMSE) to obtain the optimal network model parameters.

[0180] (6) Model Evaluation and Parameter Tuning: After completing model training, the model's performance is evaluated using a validation set (containing both experimental and corresponding simulation data). Metrics such as MSE (Mean Squared Error), RMSE (Root Mean Squared Error), and MAE (Mean Absolute Error) are calculated on the validation set, and the predicted remaining lifetime (RUL) is compared with the true label. Based on the prediction performance on the validation set, model hyperparameters such as learning rate, number of hidden layer units, and number of attention heads are adjusted, and the best-performing model during training is saved.

[0181] (7) Model testing and application: Input the vibration data of the electric shooting platform bearing collected in the test set or actual task into the trained optimal model, and output the predicted value of the remaining service life (RUL) of the bearing to achieve accurate prediction of the life of the electric shooting platform bearing.

[0182] In this embodiment of the invention, a method for predicting the lifespan of an electric shooting platform bearing is provided, effectively solving the problems of pure data-driven methods being highly dependent on historical failure data, having poor predictive interpretability, and pure physical models being difficult to accurately match actual working conditions. First, noise is removed and the intrinsic components of damage impact are extracted using an EEMD combined with a wavelet threshold denoising algorithm. An explicit mapping relationship between the RMS value and the simulated defect size is constructed to evaluate the bearing's service performance status, dividing it into four states: healthy, damage initiation, rapid damage evolution, and failure. Second, a physical-driven digital twin model integrating Hertzian contact theory, damage geometry, and nonlinear dynamics is constructed to accurately reproduce the bearing's vibration response throughout its entire lifecycle from healthy to failure. Finally, lifespan prediction is performed in the damage initiation and rapid damage evolution states. A Transformer-LMU neural network is designed, utilizing the Transformer's multi-head attention mechanism to achieve heterogeneous feature-level fusion and long-range dependency modeling of virtual twin data and measured vibration data. The LMU's state-space model efficiently captures the slow dynamic evolution of degradation trends in long vibration sequences, and a bidirectional cross-attention mechanism achieves complementary enhancement between physical simulation and measured information. Combining the above content and following the steps of the proposed system, the modeling accuracy of the long-cycle degradation trajectory of electric launcher bearings and the prediction performance of remaining service life (RUL) can be significantly improved in practical applications.

[0183] Please see Figure 4 , Figure 4 This is a structural block diagram of an electric shooting platform bearing life prediction system provided in Embodiment 3 of the present invention.

[0184] This invention provides a system for predicting the life of an electric shooting platform bearing, comprising:

[0185] The acquisition module 401 is used to acquire the real vibration signal data of the electric shooting platform bearing, and to preprocess the real vibration signal data to generate the final noise-reduced signal data.

[0186] Output module 402 is used to output twin vibration signal data corresponding to the real vibration signal data based on the maximum amplitude value corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final noise-reduced signal data using the dynamic equation set.

[0187] The prediction module 403 is used to use a Transformer-LMU model prediction network based on virtual-real fusion to predict the life of the electric shooting platform bearing based on real vibration signal data and twin vibration signal data, and output the predicted life value of the bearing.

[0188] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0189] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the electric shooting platform bearing life prediction method as described in any of the above embodiments.

[0190] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the electric launcher bearing life prediction method as described in any of the above embodiments.

[0191] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0192] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0193] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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 method for predicting the life of an electric shooting platform bearing, characterized in that, include: Acquire the actual vibration signal data of the electric shooting platform bearing, and preprocess the actual vibration signal data to generate the final noise-reduced signal data; The system of dynamic equations is used to output twin vibration signal data corresponding to the real vibration signal data based on the maximum amplitude corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final denoised signal data, including: The size conversion coefficient is calculated based on the maximum amplitude value corresponding to the actual vibration signal data and the normalized root mean square maximum value corresponding to the final noise reduction signal data. Multiply the size conversion coefficient by the normalized root mean square maximum value corresponding to the final noise-reduced signal data, and output the defect size corresponding to the final noise-reduced signal data. Based on the defect size, calculate the axially accumulated roller contact force; The system of dynamic equations is used to output twin vibration signal data corresponding to the actual vibration signal data based on the axially accumulated roller contact force. The Transformer-LMU model prediction network based on virtual-real fusion includes an embedding layer, a Transformer-LMU neural network, and a prediction output layer. Using this network, the lifespan is predicted based on the real vibration signal data and the twin vibration signal data, and the predicted lifespan value of the electric launcher bearing is output, including: An embedding layer is used to embed features into the real vibration signal data and the twin vibration signal data respectively, and the real data feature vector corresponding to the real vibration signal data and the twin data feature vector corresponding to the twin vibration signal data are output. The real data feature vector and the twin data feature vector are normalized respectively, and the normalized real data feature vector and the normalized twin data feature vector are output. The normalized real data feature vector and the normalized twin data feature vector are then concatenated to generate a fused feature. The Transformer-LMU neural network is used to perform feature splitting and enhancement based on the fused features, and outputs real data-enhanced feature vectors and twin data-enhanced feature vectors; The predicted output layer performs predictions based on the real data augmented feature vector and the twin data augmented feature vector, and outputs the predicted life value of the electric launcher bearing.

2. The method for predicting the life of an electric shooting platform bearing according to claim 1, characterized in that, The preprocessing of the real vibration signal data to generate the final noise-reduced signal data includes: Gaussian white noise is added to the real vibration signal data to generate noisy signal data; Integrated empirical mode decomposition is used to iteratively decompose the noisy signal data, and the multiple IMF components and residual terms corresponding to the noisy signal data are output. Perform wavelet threshold denoising on each of the IMF components to generate denoised IMF components corresponding to each of the IMF components; Perform discrete wavelet transform on each of the denoised IMF components and output multiple detail coefficients corresponding to each of the denoised IMF components; A soft threshold shrinkage function is used to determine multiple shrinkage coefficients corresponding to each of the denoised IMF components based on multiple detail coefficients corresponding to each of the denoised IMF components. Perform discrete wavelet inverse transform on multiple shrinkage coefficients corresponding to each of the denoised IMF components to generate reconstructed and denoised IMF components corresponding to each of the denoised IMF components. The reconstructed and denoised IMF components and residual terms are combined to generate the final denoised signal data.

3. The method for predicting the life of an electric shooting platform bearing according to claim 1, characterized in that, The step of embedding features into the real vibration signal data and the twin vibration signal data using an embedding layer, and outputting the real data feature vector corresponding to the real vibration signal data and the twin data feature vector corresponding to the twin vibration signal data, includes: The real vibration signal data and the twin vibration signal data are sequentially subjected to convolution, normalization, and nonlinear mapping to output the real data convolution output and the twin data convolution output. The convolutional outputs of the real data and the twin data are normalized respectively to generate feature vectors of the real data and twin data.

4. The method for predicting the life of an electric shooting platform bearing according to claim 1, characterized in that, The step of using a Transformer-LMU neural network to perform feature splitting and enhancement based on the fused features, and outputting real data-enhanced feature vectors and twin data-enhanced feature vectors, includes: Multi-head attention calculation is performed on the fused features to generate a multi-head attention output; The multi-head attention output is input into the feedforward network for feature splitting, and the outputs real data feedforward feature vector and twin data feedforward feature vector are output. Based on the feedforward feature vectors of real data and twin data, the final output feature vectors of real data and twin data are determined. A cross-attention mechanism is used to enhance the features of the final output feature vector of the real data and the final output feature vector of the twin data, generating enhanced feature vectors of the real data and twin data.

5. The method for predicting the life of an electric shooting platform bearing according to claim 1, characterized in that, The step of predicting and outputting the predicted lifespan value of the electric launcher bearing by using the prediction output layer based on the augmented feature vector of the real data and the augmented feature vector of the twin data includes: One-dimensional adaptive max pooling is performed on the real data augmented feature vector and the twin data augmented feature vector respectively, and the max pooled real data augmented feature vector and the max pooled twin data augmented feature vector are output. The max-pooling real data augmented feature vector and the max-pooling Siamese data augmented feature vector are concatenated to output a concatenated feature vector. The concatenated feature vectors are sequentially subjected to linear transformation, normalization, and nonlinear mapping to generate hidden layer features; The hidden layer features are sequentially subjected to linear transformation and nonlinear mapping to output the predicted life value of the electric shooting platform bearing.

6. A system for predicting the lifespan of an electric shooting platform bearing, applied to the method for predicting the lifespan of an electric shooting platform bearing as described in claim 1, characterized in that, include: The acquisition module is used to acquire the actual vibration signal data of the electric shooting platform bearing, and to preprocess the actual vibration signal data to generate the final noise-reduced signal data. The output module is used to output twin vibration signal data corresponding to the real vibration signal data by using the dynamic equation set based on the maximum amplitude value corresponding to the real vibration signal data and the normalized root mean square maximum value corresponding to the final noise-reduced signal data. The prediction module is used to use a Transformer-LMU model prediction network based on virtual-real fusion to predict the lifespan of the electric launcher bearing based on the real vibration signal data and the twin vibration signal data, and output the predicted lifespan value of the electric launcher bearing.

7. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the electric launcher bearing life prediction method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the electric launcher bearing life prediction method as described in any one of claims 1-5.

Images