A life estimation method for aviation fuel gear pump based on gear failure

Abstract

The present invention discloses a method for estimating the life of an aviation fuel gear pump based on gear failure, which relates to the technical field of gear pump life estimation. The method includes synchronously collecting parallel and vertical polarization vibration signals and performing baseline correction and low-pass filtering. The method calculates the dynamic contact stress of the tooth surface by combining smoke scattering spectra, gear material parameters, and Hertz contact theory to construct a multi-physics field data set. The method extracts the degree of polarization, polarization angle, stress change rate, and smoke concentration change from the multi-physics field data set to generate a node feature vector sequence. The method also fuses the dynamic graph embedding vector with the modified Hertz contact term and Archard wear term to construct a physically constrained neural differential equation. The method rigidly embeds the modified Hertz contact term and Archard wear term into the neural differential equation and uses the Dormand-Prince solver to iteratively generate the state evolution trajectory, ensuring that the model adheres to both data-driven laws and physical conservation constraints.

Description

Technical Field

[0001] The invention relates to the technical field of gear pump life estimation, in particular to a method for estimating the life of an aviation fuel gear pump based on gear failure. Background Art

[0002] In recent years, life prediction technology for aviation fuel gear pumps has made significant progress in the fields of multi-physics coupling analysis and high-precision sensing. Traditional vibration monitoring methods rely on accelerometers combined with spectral analysis to extract gear mesh characteristics, and identify early wear signals through envelope demodulation and order tracking. Hertzian contact theory is widely used to model dynamic contact stresses in gears, combining material elastic parameters with the tooth surface curvature radius to invert the local stress distribution. For wear product detection, UV-visible spectroscopy quantifies wear particle concentration through smoke scattering properties, and its polarization-sensitive detection scheme can effectively distinguish the scattering characteristics of metallic and non-metallic particles. Furthermore, data-driven models (such as neural networks) integrate heterogeneous data from multiple sources, such as vibration, stress, and smoke concentration, to construct an end-to-end life prediction framework. These models are then combined with wear trajectories generated by finite element simulation to optimize model generalization. With the introduction of technologies such as non-spherical particle scattering matrix modeling, dynamic wear energy density calculation, and physically constrained neural differential equations, multi-physics coupling-driven life prediction methods have gradually become a research hotspot in this field.

[0003] The main defects of current technology are: first, the baseline drift and high-frequency noise suppression in vibration signal processing are insufficient, resulting in low accuracy in the effective signal extraction of the gear meshing characteristic frequency band; second, the dynamic correction of smoke concentration to the wear coefficient lacks a polarization attenuation quantification model based on the Stokes vector transmission equation, and cannot distinguish the wear contributions of gears and bearings; third, the traditional Archard wear model does not incorporate the nonlinear coupling effect of geometric-material parameters in Hertz contact theory, making it difficult to accurately describe the local wear energy distribution driven by dynamic stress on the tooth surface, resulting in accumulated errors in life prediction. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an aviation fuel gear pump life estimation method based on gear failure, which solves the problems of insufficient life prediction accuracy and difficulty in dynamic wear mechanism modeling of aviation fuel gear pumps due to multi-physical field coupling effects.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for estimating the life of an aviation fuel gear pump based on gear failure, which comprises: synchronously collecting parallel and vertical polarization vibration signals and performing baseline correction and low-pass filtering, calculating the dynamic contact stress of the tooth surface by combining smoke scattering spectrum, gear material parameters and Hertz contact theory, and constructing a multi-physics field data set;

[0008] Extract polarization degree, polarization angle, stress change rate, and smoke concentration change from multi-physics field datasets to generate node feature vector sequences. Generate attention scores using dynamic edge weights and physical constraint masks, and output dynamic graph embedding vectors.

[0009] The dynamic graph embedding vector is fused with the modified Hertzian contact term and Archard wear term to construct a physically constrained neural differential equation. The Dormand-Prince solver is used for iterative integration to generate a continuous-time state evolution trajectory.

[0010] The continuous-time state evolution trajectory is used as the boundary condition of the finite element simulation to generate a high-precision wear trajectory as the teacher model. The student model parameters are optimized by combining adversarial training with a multi-objective loss function. The training set is expanded and overfitting is suppressed through sliding sampling to output an enhanced life prediction model.

[0011] The enhanced life prediction model is called on the three-dimensional parameter grid to calculate the remaining life, combined with Monte Carlo Dropout to generate confidence intervals, and maintenance recommendations are adjusted based on the gradient of the life surface to output dynamic maintenance decisions.

[0012] As a preferred solution of the aviation fuel gear pump life estimation method based on gear failure of the present invention, the construction of the multi-physics field data set includes the following steps:

[0013] Synchronously collect parallel and vertical polarization vibration components, and generate a time series data set of vibration components through baseline correction and low-pass filtering;

[0014] A smoke environment simulation device was constructed. A UV-visible spectrometer was used to collect smoke scattering spectra under vertically incident polarized light. After subtracting the background signal, the integrated area of ​​the scattering intensity within the characteristic peak wavelength range was extracted. The real-time smoke concentration gradient was calculated by combining the extinction coefficient and the optical path length.

[0015] Based on the gear material's equivalent elastic modulus, Poisson's ratio, and tooth surface curvature radius, Hertz contact theory is used to perform nonlinear inversion on the vibration displacement amplitude to generate the tooth surface dynamic contact stress.

[0016] The dynamic wear coefficient is corrected according to the smoke attenuation coefficient, and the local wear energy density is calculated by combining the contact stress, sliding speed and friction coefficient. The gradient distribution of tooth surface wear is solved by the non-uniform wear equation.

[0017] Gear speed data is collected synchronously, and the time series data sets of vibration components, smoke concentration gradient, tooth surface dynamic contact stress and tooth surface wear gradient distribution are correlated and integrated into a multi-physics field data set.

[0018] As a preferred solution of the aviation fuel gear pump life estimation method based on gear failure described in the present invention, the generation of dynamic contact stress on the tooth surface refers to inverting the vibration displacement based on the synthetic amplitude of the vibration component, and calculating it in combination with the equivalent elastic modulus, Poisson's ratio, tooth surface curvature radius and vibration displacement.

[0019] As a preferred solution of the method for estimating the life of an aviation fuel gear pump based on gear failure according to the present invention, the generating of the node feature vector sequence comprises the following steps:

[0020] Calculate the degree of polarization and polarization angle for parallel and perpendicular polarization vibration components;

[0021] The dynamic contact stress of the tooth surface is differentiated by time to obtain the stress change rate, and the smoke concentration is differentiated by time to obtain the concentration change;

[0022] The polarization degree, polarization angle, stress change rate and concentration change are encapsulated as a node feature vector sequence.

[0023] As a preferred solution of the aviation fuel gear pump life estimation method based on gear failure of the present invention, the dynamic edge weight generation includes the following steps:

[0024] The tooth pair contact vibration signal segment is intercepted from the parallel polarization vibration component, and the high-frequency vibration energy spectrum is extracted by short-time Fourier transform.

[0025] The dynamic contact stress of the tooth surface is multiplied point by point with the high-frequency vibration energy spectrum, activated by the ReLU function, and dynamically updated edge weights are generated.

[0026] As a preferred solution of the aviation fuel gear pump life estimation method based on gear failure of the present invention, the construction of the physical constraint neural differential equation includes the following steps:

[0027] The dynamic contact stress on the tooth surface is extracted from the embedded vector of the dynamic graph, which corresponds to the instantaneous load in the tooth surface contact area during gear meshing. The dynamic contact stress on the tooth surface is combined with the geometric-material parameters in the Hertz theory to calculate the driving contribution of the contact stress to wear and obtain the modified Hertz contact term.

[0028] Based on the local wear energy density, combined with the sliding speed, the correction ;

[0029] Corrected Hertzian contact term, corrected Physically constrained neural differential equations are constructed by weighted superposition with data-driven terms.

[0030] As a preferred solution of the aviation fuel gear pump life estimation method based on gear failure of the present invention, the adversarial training includes the following steps:

[0031] The vibration modes and stress-wear parameters of the continuous-time state evolution trajectory are aligned with the features of the high-precision simulated wear trajectory and spliced ​​into an input vector. The true and false probabilities are calculated through a discriminator network, and a multi-objective loss function is designed to jointly optimize the data matching loss, adversarial loss, and physical consistency loss.

[0032] An alternating training strategy is used to update the student model and discriminator parameters, and the learning rate scheduling and early stopping strategy are combined to control the training stability.

[0033] As a preferred solution of the aviation fuel gear pump life estimation method based on gear failure of the present invention, the output dynamic maintenance decision includes the following steps:

[0034] Calculate the partial derivatives of the remaining life with respect to the rotational speed, contact stress and smoke concentration in a three-dimensional parameter grid, and generate gradient vectors and life surfaces;

[0035] The remaining life distribution is predicted through Monte Carlo Dropout sampling, and the confidence interval is calculated and compared with the remaining life critical value. At the same time, the operation and maintenance strategy is adjusted in combination with the three-dimensional life surface, and the remaining life prediction value and dynamic maintenance decision are finally output.

[0036] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the method for estimating the life of an aviation fuel gear pump based on gear failure as described in the first aspect of the present invention is implemented.

[0037] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the method for estimating the life of an aviation fuel gear pump based on gear failure as described in the first aspect of the present invention is implemented.

[0038] The beneficial effects of the present invention are: combining the non-spherical particle scattering matrix to construct the Stokes vector transmission equation, quantifying the attenuation coefficient of smoke on polarized signals, and integrating the characteristic peak scattering intensity through the trapezoidal rule to achieve real-time correlation between the concentration gradient and the wear coefficient, and accurately distinguish the wear contribution of gears and bearings; the modified Hertz contact term and Archard wear term are rigidly embedded in the neural differential equation, and the Dormand-Prince solver is used to iteratively generate the state evolution trajectory to ensure that the model follows both data-driven laws and physical conservation constraints. The high-precision wear trajectory generated by finite element simulation is used as a teacher model, and the data matching loss, adversarial loss and physical consistency loss are combined to force the student model to align with the physical laws in terms of wear distribution and stress diffusion characteristics, thereby improving the prediction stability under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 Flowchart of the life estimation method for aviation fuel gear pump based on gear failure.

[0041] Figure 2 Build schematics for multiphysics datasets.

[0042] Figure 3 Constructing schematics for physically constrained neural differential equations.

[0043] Figure 4 Generate schematic diagrams for dynamic maintenance decisions. DETAILED DESCRIPTION

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0047] Reference Figure 1 , is an embodiment of the present invention, which provides a method for estimating the life of an aviation fuel gear pump based on gear failure, comprising the following steps:

[0048] S1 synchronously collects parallel and vertical polarization vibration signals and performs baseline correction and low-pass filtering. It combines smoke scattering spectra, gear material parameters and Hertz contact theory to calculate the dynamic contact stress on the tooth surface and construct a multi-physics field data set.

[0049] Specifically, see Figure 2 , install a high-frequency polarization-sensitive vibration probe at the normal position of the gear meshing surface, and the installation angle is strictly perpendicular to the gear meshing direction to ensure the accuracy of capturing the normal component of the vibration signal; start the high-frequency data acquisition equipment, and synchronously record the parallel polarization vibration component and the vertical polarization vibration component of the original vibration signal, and keep the timestamps aligned during the acquisition process; perform baseline correction on the original vibration signal to eliminate environmental electromagnetic interference, and use a low-pass filter to suppress the high-frequency noise of the vibration signal after baseline correction, and retain the effective signal of the gear meshing characteristic frequency band.

[0050] To further explain, baseline correction refers to calculating the mean of the original vibration signal, extracting the baseline drift component caused by environmental electromagnetic interference, subtracting the baseline drift component from the original vibration signal, eliminating DC offset and low-frequency interference, and generating a baseline-corrected vibration signal.

[0051] Using a low-pass filter to suppress the high-frequency noise of the vibration signal after baseline correction refers to configuring a fourth-order Butterworth low-pass filter and a cutoff frequency, inputting the vibration signal after baseline correction into the low-pass filter, filtering out the high-frequency noise components above the cutoff frequency, and retaining the effective signal of the gear meshing characteristic frequency band.

[0052] According to the sensitivity parameters of the high-frequency polarization-sensitive vibration probe, the amplitude of the effective signal is calibrated to generate a time-series vibration data set of parallel polarization vibration components and vertical polarization vibration components.

[0053] Construct a smoke environment simulation device, and use a flow velocity sensor (such as a hot wire anemometer) in the smoke environment simulation device to measure the smoke flow velocity in real time to characterize the flow rate of the lubricant and air mixture in the gearbox;

[0054] According to the smoke environment simulation device, the detection spectrum of the UV-visible spectrometer is set to collect the smoke scattering spectrum under the condition of vertical incident polarized light.

[0055] A background spectrum is collected in a smoke-free environment, and the background signal is subtracted from the smoke scattering spectrum to eliminate the interference of ambient light and container wall reflection, thereby generating a baseline-calibrated scattering spectrum. In the baseline-calibrated scattering spectrum, the characteristic peak of the smoke is identified, and the wavelength range of the smoke characteristic peak is marked to cover the main peak area of ​​the smoke scattering spectrum. Within the wavelength range of the smoke characteristic peak, the scattering intensity values ​​of multiple discrete wavelength points are extracted at fixed intervals to form a discrete data set.

[0056] For the scattering intensity values ​​within the wavelength range of the smoke characteristic peak, the trapezoidal rule is applied to calculate the scattering intensity integral area, and the expression is:

[0057] ;

[0058] in, For smoke in time The scattering intensity integrated area, is the wavelength point index of the smoke, is the starting point of wavelength, is the wavelength end point, For in time Smoke dots The scattering intensity value at For in time Smoke dots The scattering intensity value at is the wavelength interval, which is used to define the width of each small trapezoid, that is, the wavelength range.

[0059] Put smoke in time The scattering intensity integral area is converted into smoke concentration gradient, and the expression is:

[0060] ;

[0061] in, For smoke in time The concentration gradient is used to quantify the real-time concentration changes of smoke. is the extinction coefficient of smoke, is the optical path length.

[0062] Construct the Stokes vector transmission equation for non-smog polarization, which is expressed as:

[0063] ;

[0064] in, is the rate of change of the Stokes vector over time, reflecting the transient evolution of the polarization state, is the Stokes vector, which is used to characterize the full polarization state of the light wave (intensity, linear polarization, circular polarization), is the scattering matrix of non-spherical particles, which characterizes the scattering characteristics of smoke to incident polarized light. is the diffusion coefficient, which is used to quantify the diffusion intensity of smoke caused by Brownian motion or turbulence. is the gradient operator, is the gradient distribution of the Stokes vector in three-dimensional space.

[0065] It is further explained that the scattering matrix of non-spherical particles is obtained by fitting the least squares method using a standard polarized light source (such as linearly polarized laser) as the incident light in a smoke simulation device and measuring the outgoing Stokes vector.

[0066] The characteristic wavelength scattering intensity and smoke concentration gradient are associated with the acquisition timestamp to generate a time series intensity dataset;

[0067] Obtain the equivalent elastic modulus and Poisson's ratio of the gear material from the gear material mechanical property test report; use an optical profilometer or a three-dimensional coordinate measuring machine to measure the tooth surface curvature radius in the tooth contact area along the gear meshing line direction; and invert the vibration displacement using the composite amplitude of the parallel and perpendicular polarization vibration components in the time-series vibration data set. The expression is:

[0068] ;

[0069] in, For in time The vibration displacement amplitude, For in time The translational polarization vibration component, For in time The vertically polarized vibration component, Calibrate the parameters for the probe sensitivity.

[0070] Substituting the equivalent elastic modulus, Poisson's ratio, tooth surface curvature radius and vibration displacement amplitude into the Hertz contact theory formula, the dynamic contact stress of the tooth surface is obtained:

[0071] ;

[0072] in, For in time The dynamic contact stress of the tooth surface is used to characterize the instantaneous stress in the tooth surface contact area during gear meshing. is the normal direction, is the equivalent elastic modulus, is the tooth surface curvature radius in the tooth contact area, is the Poisson's ratio of the gear material, which is used to characterize the proportional relationship between the lateral contraction and longitudinal extension of the material when subjected to force. It is the core symbol of Hertz's theory, corresponding to the influence of geometric shape, material elasticity, and deformation nonlinearity, and is irreplaceable;

[0073] The tooth surface coordinate cloud image recorded by the optical profilometer is synchronized with the effective signal timestamp that retains the gear meshing characteristic frequency band. When the vibration probe captures the peak value of the dynamic contact stress on the tooth surface, the geometric center point of the tooth surface meshing area at this time is marked as the gear meshing position coordinate.

[0074] To further illustrate, the equivalent elastic modulus is used to combine the elastic modulus and Poisson's ratio of the two contact materials of the gear pair, and the expression is:

[0075] ;

[0076] in, is the elastic modulus of the two gear materials, is the Poisson's ratio corresponding to the materials of the two gears.

[0077] The tooth surface curvature radius in the tooth surface contact area is determined by the curvature radius of the two gear tooth profiles, and the expression is:

[0078] ;

[0079] in, is the radius of curvature of the tooth profiles of the two gears. If both gears are externally meshed, take a positive value; if both gears are internally meshed, the denominator is .

[0080] Hertz contact theory formula is the square root of the curvature radius, which reflects the influence of the geometric characteristics of the contact area on the contact stress. The area of ​​the contact area is proportional to the square root of the curvature radius. The larger the curvature radius, the larger the contact area and the smaller the contact stress. Used to balance dimensions.

[0081] Hertz contact theory formula is the material elasticity correction coefficient, which is the effect of the lateral deformation ability of the material on the contact stress through the Poisson's ratio. The larger the Poisson's ratio (the easier the material is to shrink laterally), the smaller the denominator, the larger the coefficient as a whole, and the contact stress increases accordingly. is the result from the elliptic integral, corresponding to the geometry of the contact area, It is used to correct the contribution of Poisson's ratio to the lateral deformation of the material.

[0082] Hertz contact theory formula It means that the contact stress and vibration displacement have a nonlinear relationship, where The power indicates that as the displacement increases, the growth rate of the contact stress slows down (due to the simultaneous increase in the contact area), reflecting the nonlinear coupling of elastic deformation and geometry.

[0083] Based on the non-spherical particle scattering matrix in the Stokes vector transmission equation for non-smog polarization, the attenuation coefficient of smoke on polarized signals is calculated as follows:

[0084] ;

[0085] in, is the smoke attenuation coefficient.

[0086] According to the smoke attenuation coefficient, the wear coefficient is corrected to distinguish the contributions of gear and bearing wear. The expression is:

[0087] ;

[0088] in, For in time The dynamic wear coefficient, which characterizes the gear wear rate, is corrected in real time to reflect the influence of smoke concentration. The basic wear coefficient is the reference wear coefficient when there is no smoke concentration influence (or the concentration does not exceed the standard), which is calibrated by the standard wear test. It is the maximum allowable concentration threshold of smoke, which is the upper limit of smoke concentration safety. When it exceeds the upper limit, the alarm is triggered.

[0089] It is further explained that the maximum allowable smoke concentration threshold is set based on the historical smoke detection data of similar equipment throughout its life cycle. For example, the smoke concentration in the historical smoke detection data and the 95% quantile of the same concentration are taken as the maximum allowable smoke concentration threshold.

[0090] Based on the dynamic wear coefficient, tooth surface dynamic contact stress, sliding velocity and sliding friction coefficient, the local wear energy density is calculated as follows:

[0091] ;

[0092] in, The local wear energy density refers to the wear energy per unit volume and is used to quantify the wear energy generated by friction and contact per unit area of ​​the tooth surface per unit time. For in time The sliding speed at this time is converted from the gear speed. Is the sliding friction coefficient, used to characterize the friction characteristics of gear materials, calibrated through friction and wear tests, is the temperature-dependent hardness of the material, For in time The temperature at that time.

[0093] The abnormal peaks of the parallel and vertical polarization vibration components are associated with the wear position, the boundary conditions are set as the initial wear distribution, and the iteration is performed on the spatial grid to solve the local wear gradient of the tooth surface through the non-uniform wear equation. , the expression is:

[0094] ;

[0095] in, For gear material in position The higher the hardness, the stronger the gear material's wear resistance, which inhibits wear diffusion. is the spatial position coordinate, For in time Time, location The gradient of the wear amount at is the external excitation source term, which represents the driving effect of external load and friction on wear.

[0096] The gradient of smoke concentration is extracted from the time series intensity data set, and the time derivative of the gradient of smoke concentration is used to calculate the smoke generation rate per unit time. The expression is:

[0097] ;

[0098] in, For in time The smoke generation rate per hour is used to quantify the amount of gear wear generated per unit time. is the smoke concentration gradient versus time The derivative of reflects the rate of change of smoke concentration over time, It is the monitoring unit volume of the online smoke counter, used to convert the concentration change rate into the actual mass flow rate;

[0099] According to the gear meshing position coordinates , define the smoke generation source term , only in There is smoke generation;

[0100] Substitute the velocity, diffusion coefficient, generation rate and generation source term into the unsteady smoke transport equation and solve the problem at position and time The concentration distribution of smoke at is expressed as:

[0101] ;

[0102] in, The concentration of smoke About time The partial derivative of is the smoke flow rate, The gradient of smoke concentration indicates the direction and rate of change of smoke concentration in space. The direction points to the area where the concentration increases the fastest. is the Laplace operator of smoke concentration, which indicates the intensity of smoke diffusion. The larger the diffusion coefficient, the more significant the diffusion effect.

[0103] Further explanation: the change of smoke concentration over time Contributed by three parts:

[0104] in, represents convection, Indicates diffusion, which refers to the spontaneous diffusion of smoke from high concentration to low concentration. represents the source term, which refers to the gear meshing position Smoke continues to form in the area.

[0105] Gear speed data is collected synchronously, and the time series data sets of vibration components, smoke concentration gradient, tooth surface dynamic contact stress and tooth surface wear gradient distribution are correlated and integrated into a multi-physics field data set.

[0106] S2 extracts polarization degree, polarization angle, stress change rate and smoke concentration change from the multi-physics field dataset to generate a node feature vector sequence, generates attention scores through dynamic edge weights and physical constraint masks, and outputs a dynamic graph embedding vector.

[0107] Specifically, the polarization degree of each node is extracted from the parallel and perpendicular polarization vibration components of the multi-physics field data set to characterize the degree of vibration energy concentration on the tooth surface of the corresponding node; the parallel and perpendicular polarization vibration components are inverted to obtain the polarization angle of each node to reflect the change in the vibration propagation direction; the dynamic contact stress of the tooth surface is time-differentiated to obtain the stress change rate of each node to quantify the transient impact of the dynamic load; the smoke is time-differentiated to generate the concentration change of the smoke at each node; the polarization degree, polarization angle, stress change rate, and smoke concentration change of each node are encapsulated into a node feature vector sequence of each gear tooth surface.

[0108] The contact vibration signal within the time window is intercepted from the parallel polarization vibration component to obtain the vibration signal segment of each tooth pair contact, covering the gear meshing impact event; short-time Fourier transform is applied to the vibration signal segment of each tooth pair contact to extract high-frequency vibration energy, output the high-frequency vibration energy spectrum, and focus on the gear meshing characteristic frequency band; the dynamic contact stress of the tooth surface is multiplied point by point with the high-frequency vibration energy in the high-frequency vibration energy spectrum, and activated by the ReLU function to obtain the dynamically updated edge weight, which is used to quantify the contact energy interaction intensity of each tooth pair.

[0109] Multi-band polarization parameters, including the degree of polarization and polarization angle of each spectral band, are extracted from the Stokes vector. The smoke attenuation coefficient is dynamically adjusted according to the gear speed in the time-series speed dataset. The multi-band polarization parameters are then fused with the dynamically adjusted smoke attenuation coefficient to generate a physical constraint mask.

[0110] The node feature vector is linearly transformed and a fully connected layer is applied to generate a query vector and a key vector. The node features are updated using the dynamically updated edge weights to obtain a value vector. The query vector, key vector, and value vector are combined and the physical constraint mask is superimposed to calculate the attention score, which is:

[0111] ;

[0112] in, To fuse the query vector , key vector , value vector and the attention scores generated by the physical constraint mask, is the normalized exponential function, is the query vector, is the key vector, is a value vector, is the transpose, is the dimension of the key vector, is the dimension index, is the physical constraint mask.

[0113] The physical constraint mask and the attention score are weightedly summed to obtain the dynamic graph embedding vector.

[0114] In S3, the dynamic graph embedding vector is fused with the modified Hertzian contact term and Archard wear term to construct a physically constrained neural differential equation, and the Dormand-Prince solver is used for iterative integration to generate a continuous-time state evolution trajectory.

[0115] Specifically, see Figure 3 , analyze the dimensions of the dynamic graph embedding vector, identify the physical meaning of each dimension (for example, the first 16 dimensions correspond to vibration modes, the middle 32 dimensions correspond to stress-wear coupling parameters, and the last 16 dimensions correspond to smoke concentration diffusion characteristics), and obtain the input vector; scale the input vector to obtain the normalized embedding vector;

[0116] The normalized embedding vector is linearly mapped to the hidden layer via a weight matrix to generate intermediate features. The hyperbolic tangent function is then applied to activate the intermediate features, capturing the nonlinear coupling between vibration energy and contact stress (e.g., the exponential increase in vibration energy caused by a sudden stress increase) and the nonlinear feedback of the smoke concentration gradient on the wear rate. The activated intermediate features are then mapped to the output layer via a weight matrix to generate the original rate of change.

[0117] The normalized embedding vector is superimposed with the original change rate to output the data-driven state change rate, retaining the key physical information in the normalized embedding vector (such as the instantaneous value of contact stress) and preventing the loss of deep network information.

[0118] The dynamic contact stress on the tooth surface is extracted from the embedded vector of the dynamic graph, which corresponds to the instantaneous load in the tooth surface contact area during gear meshing. The dynamic contact stress on the tooth surface is combined with the geometric-material parameters in the Hertz theory to calculate the driving contribution of the contact stress to wear. The modified Hertz contact term is obtained to quantify the nonlinear driving effect of the contact stress on the wear rate. The expression is:

[0119] ;

[0120] Based on the local wear energy density, combined with the sliding speed, the correction , the expression is:

[0121] Corrected ;

[0122] in, For time Local wear energy density;

[0123] The dynamic wear coefficient is used as a proportional factor to adjust the correction effect of smoke concentration on the wear rate, and the corrected , specifically:

[0124] ;

[0125] The modified Hertzian contact term and the modified Archard wear term are hard-embedded into the physical terms.

[0126] It is further explained that during the gear meshing process, the load changes dynamically due to pitting on the tooth surface, eccentric loading or speed fluctuations. Traditional average stress cannot capture high-frequency impact events (such as inter-tooth sticking), while instantaneous load can accurately quantify the sudden increase in wear caused by local overload.

[0127] The modified Hertz contact term represents the nonlinear wear acceleration effect driven by the square of the contact stress and the material-geometry parameters, and the modified Archard wear term represents the contribution of the dynamic balance of friction energy and material hardness per unit time to the material removal rate.

[0128] Initialize the learnable parameters to represent the weights of the Hertz contact term and the Archard wear term respectively; superimpose the modified Hertz contact term and the modified Archard wear term in the data-driven state change rate and the hard embedding of the physical term according to the weights to construct the physical constraint neural differential equation, which is expressed as:

[0129] ;

[0130] in, Embedding vectors for dynamic graphs About time The derivative of For data-driven items, is the weight parameter of the neural network, is the learnable coupling coefficient of the modified Hertzian contact term, is the learnable coupling coefficient of the modified Archard wear term.

[0131] The initial integration step size and error tolerance are set according to the time series. When the local error exceeds the error tolerance, the step size is reduced and recalculated. When the local error is within the error tolerance, the step size is gradually increased to obtain the configured Dormand-Prince solver.

[0132] Real-time contact stress, dynamic wear coefficient and sliding velocity are extracted from the dynamic graph embedding vector, and the Hertz contact term and Archard wear term are recalculated to obtain the real-time updated differential equations. Starting from the initial state, the dynamic graph embedding vector is iteratively solved according to the step size of the configured Dormand-Prince solver to obtain the continuous-time state evolution trajectory, which is used as the prediction trajectory of the subsequent student model in adversarial training.

[0133] S4 uses the continuous-time state evolution trajectory as the boundary condition of finite element simulation to generate a high-precision wear trajectory as the teacher model. It combines adversarial training with a multi-objective loss function to optimize the student model parameters. It expands the training set through sliding sampling and suppresses overfitting, and outputs an enhanced life prediction model.

[0134] Specifically, the real-time contact stress and smoke concentration in the continuous-time state evolution trajectory are used as dynamic boundary conditions for finite element simulation. A finite element mesh is constructed based on the three-dimensional contour of the gear tooth surface. The equivalent elastic modulus and Poisson's ratio in the dynamic graph embedding vector are loaded. The wear diffusion equation including the smoke scattering effect is solved by finite element software to obtain a high-precision simulated wear trajectory of the teacher model. The expression is:

[0135] ;

[0136] in, is the equivalent wear stress About time The partial derivative of represents the rate of change of the equivalent wear stress with time, quantifying the instantaneous accumulation rate of material damage. is the gradient of the equivalent wear stress, which indicates the uneven distribution of stress in space.

[0137] A discriminator network is constructed by combining the continuous-time state evolution trajectory and the high-precision simulated wear trajectory. The features of the continuous-time state evolution trajectory and the high-precision simulated wear trajectory are aligned by timestamp and concatenated into a 128-dimensional input vector, where the vibration mode of the continuous-time state evolution trajectory corresponds to the vibration energy spectrum of the simulated trajectory, and the stress-wear parameters correspond to the equivalent stress distribution. A network structure is configured, comprising an input layer, a fully connected layer, and an output layer. The input layer receives the normalized 128-dimensional vector. The fully connected layer extracts the vibration-stress coupling features through the ReLU activation function and compresses them to 32 dimensions. The output layer outputs the true and false probability through the Sigmoid function, generating a discriminator network capable of distinguishing between real and simulated trajectories.

[0138] A multi-objective loss function is set based on the student model's predicted trajectory, true trajectory, and simulated trajectory. The data matching loss ensures distribution consistency by calculating the mean square error between the predicted wear amount and the true wear amount. The adversarial loss improves realism by maximizing the discriminator's judgment probability on the predicted trajectory generated by the student model. The physical consistency loss forces the student model to follow the non-uniform wear equation and Archard's law, matching the physical laws of wear rate and diffusion terms respectively, forming a loss function that integrates data, adversarial, and physical constraints.

[0139] The student model and discriminator parameters are updated synchronously through a joint optimization strategy. The student model parameters minimize the data matching loss, adversarial loss and physical consistency loss based on gradient descent. The discriminator parameters improve the recognition ability of simulation trajectories through gradient update. An alternating training strategy is adopted to first optimize the student model and then update the discriminator to avoid gradient conflicts. The learning rate scheduling and early stopping strategy are combined to control the training stability. The absolute threshold is set according to the historical training results of similar tasks. The training is terminated when the data matching loss decreases below the absolute threshold for multiple consecutive rounds. At the same time, the weights are dynamically adjusted according to the physical consistency loss to strengthen the constraints. Finally, the optimized student model is output. The predicted trajectory of the student model meets the high-precision requirements in terms of wear distribution, adversarial realism and matching of physical laws.

[0140] Locking the learnable coupling coefficient of the modified Hertzian contact term Learnable coupling coefficient with modified Archard wear term The numerical value of ensures that the physical constraints (such as contact stress square drive and friction energy accumulation) remain unchanged during the migration process; extract the time window from the continuous time state evolution trajectory, construct the fine-tuning training set, and only update the neural network The weight of is used to minimize the mean square error between the predicted trajectory of the optimized student model and the continuous-time state evolution trajectory as the training target, adapting to the local data distribution;

[0141] Sliding sampling is performed on the extracted time window to extract subsequences from the continuous-time state evolution trajectory, which are combined with the training target to generate an augmented training set. The augmented training set is input into the optimized student model, and a weight decay term is added to the loss function to suppress the overfitting tendency of the data-driven term. , and obtain the enhanced life prediction model.

[0142] S5, calls the enhanced life prediction model in the three-dimensional parameter grid to calculate the remaining life, combines it with Monte Carlo Dropout to generate confidence intervals, adjusts maintenance recommendations based on the gradient of the life surface, and outputs dynamic maintenance decisions.

[0143] Specifically, see Figure 4 , based on historical data and working condition constraints, define the parameter value ranges of gear speed, contact stress and smoke concentration, for example, the minimum gear speed is 1000rpm, the maximum gear speed is 5000rpm, the minimum contact stress is 50MPa, the maximum contact stress is 300MPa, the minimum smoke concentration is 0ppm, and the maximum smoke concentration is 500ppm;

[0144] Within each parameter range, uniform sampling is performed at fixed intervals (e.g., gear speed step size 500 rpm, contact stress step size 50 MPa, concentration step size 50 ppm) to generate a three-dimensional parameter grid. The parameters of each grid point in the three-dimensional parameter grid are input into the enhanced life prediction model to calculate the remaining life. The expression is:

[0145] ;

[0146] in, Remaining service life is the amount of time the gear system can continue to operate safely under current operating conditions. is a function symbol, indicating that the remaining service life is a function of multiple variables, describing the joint influence of gear speed, dynamic contact stress and smoke concentration on the remaining service life. is the gear speed.

[0147] fixed and , find the partial derivative of the gear speed , calculate the life change rate of adjacent gear speed grid points; fixed and , find the partial derivative of the contact stress , calculate the life change rate of adjacent stress grid points; fix and , find the partial derivative of smoke concentration , calculate the lifetime change rate of adjacent concentration grid points; combine the three partial derivatives of each grid point into a gradient vector to construct a three-dimensional lifetime surface;

[0148] Through Monte Carlo Dropout sampling, the Dropout layer remains activated during the inference phase and some neurons are randomly inactivated. Real-time monitored gear speed, contact stress, and smoke concentration data are repeatedly input into the enhanced life prediction model for forward propagation to generate a distribution of remaining life prediction values. The mean and standard deviation of the remaining life prediction distribution obtained through Monte Carlo sampling are calculated, and confidence intervals are generated based on the normal distribution of the mean and standard deviation to quantify the uncertainty of the prediction results.

[0149] Based on historical failure data, a critical remaining life value that triggers maintenance actions is set, and the lower limit of the confidence interval is compared with the critical remaining life value. If the critical remaining life value condition is not met, the current operation and maintenance plan is maintained; if the condition is met, a maintenance recommendation is triggered (such as "replace the filter element after 200 hours"). At the same time, the maintenance recommendation is adjusted based on the three-dimensional life surface (such as reducing the speed by 10% to extend the life by 150 hours), and finally the remaining life prediction value and dynamic maintenance decision are output.

[0150] This embodiment also provides a computer device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the aviation fuel gear pump life estimation method based on gear failure as proposed in the above embodiment.

[0151] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0152] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the aviation fuel gear pump life estimation method based on gear failure as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0153] In summary, the present invention constructs the Stokes vector transmission equation by combining the non-spherical particle scattering matrix, quantifies the attenuation coefficient of smoke on polarized signals, and integrates the characteristic peak scattering intensity through the trapezoidal rule to achieve real-time correlation between the concentration gradient and the wear coefficient, and accurately distinguishes the wear contribution of gears and bearings; the modified Hertz contact term and Archard wear term are rigidly embedded in the neural differential equation, and the Dormand-Prince solver is used to iteratively generate the state evolution trajectory to ensure that the model follows both data-driven laws and physical conservation constraints. The high-precision wear trajectory generated by finite element simulation is used as the teacher model, and the data matching loss, adversarial loss and physical consistency loss are combined to force the student model to align with the physical laws in terms of wear distribution and stress diffusion characteristics, thereby improving the prediction stability under complex working conditions.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for estimating the life of an aviation fuel gear pump based on gear failure, characterized by: include: The system synchronously collects parallel and perpendicular polarization vibration signals, performs baseline correction and low-pass filtering, and calculates the dynamic contact stress on the tooth surface by combining smoke scattering spectra, gear material parameters, and Hertzian contact theory to construct a multi-physics field data set. Extract polarization degree, polarization angle, stress change rate, and smoke concentration change from multi-physics field datasets to generate node feature vector sequences. Generate attention scores using dynamic edge weights and physical constraint masks, and output dynamic graph embedding vectors. The dynamic graph embedding vector is fused with the modified Hertzian contact term and Archard wear term to construct a physically constrained neural differential equation. The Dormand-Prince solver is used for iterative integration to generate a continuous-time state evolution trajectory. The continuous-time state evolution trajectory is used as the boundary condition of the finite element simulation to generate a high-precision wear trajectory as the teacher model. The student model parameters are optimized by combining adversarial training with a multi-objective loss function. The training set is expanded and overfitting is suppressed through sliding sampling to output an enhanced life prediction model. The enhanced life prediction model is called on the three-dimensional parameter grid to calculate the remaining life, combined with Monte Carlo Dropout to generate confidence intervals, and maintenance recommendations are adjusted based on the gradient of the life surface to output dynamic maintenance decisions.

2. The method for estimating the life of an aviation fuel gear pump based on gear failure according to claim 1, wherein: The construction of the multi-physics dataset comprises the following steps: Synchronously collect parallel and vertical polarization vibration components, and generate a time series data set of vibration components through baseline correction and low-pass filtering; A smoke environment simulation device was constructed. The smoke scattering spectrum under vertically incident polarized light was collected using an ultraviolet-visible spectrometer. After subtracting the background signal, the integrated area of ​​the scattering intensity within the characteristic peak wavelength range was extracted and converted into a smoke concentration gradient. Based on the gear material's equivalent elastic modulus, Poisson's ratio, and tooth surface curvature radius, Hertz contact theory is used to perform nonlinear inversion on the vibration displacement amplitude to generate the tooth surface dynamic contact stress. The dynamic wear coefficient is corrected according to the smoke attenuation coefficient, and the local wear energy density is calculated by combining the dynamic contact stress, sliding velocity and friction coefficient of the tooth surface. The gradient distribution of tooth surface wear is solved by the non-uniform wear equation. Gear speed data is collected synchronously, and the time series data sets of vibration components, smoke concentration gradient, tooth surface dynamic contact stress and tooth surface wear gradient distribution are correlated and integrated into a multi-physics field data set.

3. The method for estimating the life of an aviation fuel gear pump based on gear failure according to claim 2, wherein: Generating the dynamic contact stress of the tooth surface refers to inverting the vibration displacement based on the synthetic amplitude of the vibration component, and performing calculations in combination with the equivalent elastic modulus, Poisson's ratio, tooth surface curvature radius and vibration displacement.

4. The method for estimating the life of an aviation fuel gear pump based on gear failure according to claim 1, wherein: The generating node feature vector sequence comprises the following steps: Calculate the degree of polarization and polarization angle for parallel and perpendicular polarization vibration components; The dynamic contact stress of the tooth surface is differentiated by time to obtain the stress change rate, and the smoke concentration is differentiated by time to obtain the concentration change; The polarization degree, polarization angle, stress change rate and concentration change are encapsulated as a node feature vector sequence.

5. The method for estimating the life of an aviation fuel gear pump based on gear failure according to claim 1, wherein: Dynamic edge weight generation includes the following steps: The tooth pair contact vibration signal segment is intercepted from the parallel polarization vibration component, and the high-frequency vibration energy spectrum is extracted by short-time Fourier transform. The dynamic contact stress of the tooth surface is multiplied point by point with the high-frequency vibration energy spectrum, activated by the ReLU function, and dynamically updated edge weights are generated.

6. The method for estimating the life of an aviation fuel gear pump based on gear failure according to claim 1, wherein: The construction of the physical constraint neural differential equation includes the following steps: The dynamic contact stress of the tooth surface is extracted from the embedded vector of the dynamic graph, which corresponds to the instantaneous load in the tooth surface contact area during gear meshing. The dynamic contact stress of the tooth surface is combined with the geometric-material parameters in the Hertz theory to calculate the driving contribution of the dynamic contact stress of the tooth surface to wear and obtain the modified Hertz contact term. Based on the local wear energy density, combined with the sliding speed, the correction ; The corrected Hertz contact term, corrected It is superimposed with data-driven terms by weight to construct a physically constrained neural differential equation.

7. The method for estimating the life of an aviation fuel gear pump based on gear failure according to claim 1, wherein: The adversarial training includes the following steps: The vibration modes and stress-wear parameters of the continuous-time state evolution trajectory are aligned with the features of the high-precision simulated wear trajectory and spliced ​​into an input vector. The true and false probabilities are calculated through the discriminator network, and a multi-objective loss function is set to jointly optimize the data matching loss, adversarial loss, and physical consistency loss. An alternating training strategy is used to update the student model and discriminator parameters, and the learning rate scheduling and early stopping strategy are combined to control the training stability.

8. The method for estimating the life of an aviation fuel gear pump based on gear failure according to claim 1, wherein: The output dynamic maintenance decision comprises the following steps: Calculate the partial derivatives of the remaining life with respect to gear speed, tooth surface dynamic contact stress, and smoke concentration in a three-dimensional parameter grid, and generate gradient vectors and life surfaces. The remaining life distribution is predicted through Monte Carlo Dropout sampling, and the confidence interval is calculated and compared with the remaining life critical value. At the same time, the operation and maintenance strategy is adjusted in combination with the three-dimensional life surface, and the remaining life prediction value and dynamic maintenance decision are finally output.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for estimating the life of an aviation fuel gear pump based on gear failure according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for estimating the life of an aviation fuel gear pump based on gear failure according to any one of claims 1 to 8 are implemented.

Images