Method for predicting abrasion loss probability of mechanical transmission part in parallel misalignment state

Abstract

The invention belongs to the technical field of health state evaluation and prediction of a mechanical transmission system, and provides a method for predicting the abrasion loss probability of a mechanical transmission part in a parallel misalignment state. Comprising the following steps: S1, constructing a mechanical transmission part three-dimensional model containing parallel misalignment errors; s2, considering the random uncertainty of a load profile, and carrying out wear simulation analysis based on grid node iteration; s3, considering the cognitive uncertainty of the model, and training a Bayesian long short-term memory network model; and S4, predicting an expected value of the abrasion loss and an uncertainty interval of the expected value. According to the method, the description precision of a traditional finite element method on the abrasion behavior of the mechanical transmission part is improved, the limitation of abrasion loss deterministic numerical value prediction is broken through, a high-robustness and high-reliability probability prediction result is provided, and a scientific basis and a data support are provided for condition-based maintenance decision making of a transmission system.

Description

Technical Field

[0001] The present invention belongs to the technical field of health state assessment and prediction of mechanical transmission systems, and particularly relates to a probability prediction method for the wear amount of mechanical transmission components under the condition of parallel misalignment. Background Art

[0002] In a mechanical transmission system, the health states of key components such as gears, bearings, and couplings have a direct impact on the reliability and service life of the system. However, due to manufacturing errors, installation deviations, and dynamic changes in loads during operation, mechanical transmission components are often difficult to maintain perfect alignment under actual working conditions. Among them, parallel misalignment is a common and typical form of alignment deviation. This deviation will cause non-uniform contact and local stress concentration between components, thereby aggravating friction and wear, and seriously threatening the normal operation of the system.

[0003] Currently, for the wear amount prediction of mechanical transmission components under the condition of parallel misalignment, it mainly relies on traditional empirical models or data-driven methods. However, these methods all show obvious limitations under complex working conditions: traditional empirical models are difficult to fully consider uncertain factors, and data-driven methods have limited generalization ability in the case of small samples; in addition, although the finite element-based simulation method can provide a certain analysis accuracy, due to its use of fixed grid technology, the calculation efficiency is low, and the accuracy is limited when dealing with complex working conditions. Summary of the Invention

[0004] Aiming at the deficiencies in the prior art, the present invention proposes a probability prediction method for the wear amount of mechanical transmission components under the condition of parallel misalignment, aiming to solve the problem that it is difficult to balance prediction accuracy, calculation efficiency, and uncertainty modeling under complex working conditions in the prior art. By integrating dynamic finite element analysis and Bayesian long short-term memory network, this method can accurately predict the wear amount of mechanical transmission components and its uncertainty interval on the premise of considering load randomness and system uncertainty, providing a scientific basis for equipment maintenance and life assessment.

[0005] The technical solution of the present invention is specifically as follows:

[0006] S1: According to the specific dimensional parameters and material properties of the mechanical transmission components, construct a three-dimensional model of the mechanical transmission components including parallel misalignment error;

[0007] S2: Conduct an in-depth analysis of the working conditions load of mechanical transmission components to determine the load profile; fully consider the random uncertainties existing in the load profile, and use statistical methods to quantify the uncertainties of the load profile parameters; sample the load profile parameters; take all the sampled load profile parameters as input conditions, and in the finite element analysis software, carry out wear simulation analysis based on grid node iterative reconstruction to obtain the wear amount growth data set of mechanical transmission components under different load profiles;

[0008] S3: Use the wear amount growth data set obtained in step S2 to train the Bayesian long short-term memory network model for wear amount probability prediction, and use the Dropout technology to conduct uncertainty modeling on the weight parameters in this network;

[0009] S4: Use the trained Bayesian long short-term memory network model to predict the wear amount expectation value and its uncertainty interval of mechanical transmission components in the future short time period.

[0010] Preferably, the wear simulation analysis method based on grid node iterative reconstruction described in step S2 is specifically as follows:

[0011] S2.1: In the transient simulation module of ANSYS software, define the initial load step increment Δt 1 ; use the APDL command to integrate the Archard wear model into the post-processing module of the simulation; after completing the transient simulation of the duration of Δt 1 , obtain the contact stress and sliding speed of all nodes on the contact surface of the mechanical transmission component, and calculate the wear depth Δh of all nodes 1j ; where j is the number of nodes on the contact surface;

[0012] S2.2: To simplify the calculation and improve the simulation efficiency, set the wear depth of each node to change linearly with time in a short time, and ignore the wear acceleration phenomenon during this period; therefore, within the duration of N 1 ×Δt 1 , the wear depth of each node can be expressed as N 1 ×Δh 1j ;

[0013] S2.3: Through the APDL command, move all nodes along the normal direction of the contact surface by N 1 ×Δh 1j to simulate the material loss during the wear process;

[0014] S2.4: Set the next load step increment Δt 2 , re-conduct the wear simulation, and determine the wear depth Δh of each node 2j ; repeat steps S2 and S3;

[0015] S2.5: Finally, within the total running time Δt, the cumulative wear depth Δh of each node j and the maximum wear depth Δh of the mechanical transmission components max are calculated by the following formula:

[0016]

[0017] Δh max = max{Δh j}

[0018] where k is the number of load steps.

[0019] Preferably, the training method of the Bayesian long short-term memory network model described in step S3 is as follows:

[0020] S3.1: Set the initial values of the network parameters; based on grid search for preliminary screening of the parameter range, and combined with Bayesian optimization, determine the current optimal number of long short-term memory layers, number of hidden layers, learning rate, and regularization coefficient of the network;

[0021] S3.2: Evaluate the uncertainty interval output by the network based on the improved prediction interval coverage probability (IPICP) index, and guide the adjustment of the Dropout rate and the readjustment of the learning rate and regularization coefficient;

[0022] S3.3: Evaluate the performance of the network under different data volumes by gradually increasing the training data volume, and determine the shortest acceptable training data duration;

[0023] S3.4: Based on the above optimization steps, finally complete the training of the Bayesian long short-term memory network and apply it to the short-term probability prediction of mechanical transmission component wear.

[0024] Preferably, the specific implementation steps of adjusting the Dropout rate, learning rate, and regularization coefficient based on the IPICP index in step S3.2 are as follows:

[0025] S3.2.1: First, calculate the coverage rate of the observed values within the uncertainty interval output by the current network through the IPICP index, and introduce an interval width penalty term to evaluate the rationality of the uncertainty range. The IPICP index is:

[0026]

[0027] where N is the number of prediction points output by the network, and c i indicates whether the observed value of prediction point i falls within the uncertainty interval. If it falls within, then c i= 1, otherwise c i = 0; and represent the upper and lower bounds of the uncertainty interval respectively; α is the penalty coefficient that controls the influence of the uncertainty interval width;

[0028] S3.2.2: Subsequently, iteratively adjust the Dropout rate, learning rate, and regularization coefficient to optimize the IPICP metric:

[0029] When the IPICP metric is less than 0.9, decrease the Dropout rate by a step of 0.05, increase the learning rate by 20%, and increase the regularization coefficient by 20% to narrow the prediction interval; when the IPICP metric is greater than or equal to 0.9, the Dropout rate, learning rate, and regularization coefficient remain unchanged.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] 1. By iteratively reconstructing the mesh distribution on the contact surface of mechanical transmission components during the simulation process, the present invention significantly enhances the accuracy and performance of the traditional finite element method in depicting the wear behavior of mechanical transmission components.

[0032] 2. The present invention innovatively combines Latin hypercube sampling with Bayesian long short-term memory network, and by introducing Dropout probabilistic modeling and an improved dynamic parameter adjustment mechanism for the prediction interval coverage rate, breaks through the limitations of traditional deterministic numerical prediction of wear amount, and can provide probability prediction results with both high robustness and high reliability. Description of the Drawings

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. By referring to the drawings, the features and advantages of the present invention can be more clearly understood. The drawings are schematic and should not be construed as limiting the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 is the flowchart of the method for probabilistic prediction of wear amount of mechanical transmission components in the parallel misalignment state proposed by the present invention.

[0035] Figure 2 is the three-dimensional model of the aviation spline coupling and the schematic diagram of its load profile in Example 1.

[0036] Figure 3 is the structural schematic diagram of the Bayesian long short-term memory network in Example 1.

[0037] Figure 4It is a schematic diagram of the result of predicting the wear depth probability of the aviation spline coupling by using the present invention in Embodiment 1. Detailed implementation manners

[0038] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners. It should be noted that, without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0039] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0040] As Figure 1 shown, the method for predicting the wear amount probability of mechanical transmission components in the parallel misalignment state proposed by the present invention mainly includes four steps: three-dimensional model construction, dynamic wear simulation analysis considering random uncertainty, training of the Bayesian long short-term memory network model considering cognitive uncertainty, and wear amount probability prediction. The implementation process will be described in detail below with an embodiment of an aviation spline coupling:

[0041] Embodiment 1

[0042] Taking a typical spline coupling in the aviation industry as the application object, probability prediction is carried out on the maximum wear depth of its tooth surface. First, according to the key dimensional parameters of the spline coupling shown in Table 1, a three-dimensional finite element model with a parallel misalignment amount of 0.06 mm is established in ANSYS, as Figure 2 shown. Define the material properties according to Table 2. Construct 60 groups of tooth surface contact pairs and adopt the enhanced Lagrangian contact algorithm.

[0043] Table 1

[0044]

[0045] Table 2

[0046]

[0047] In the dynamic wear simulation stage, the Latin hypercube sampling technique is used to generate 100 groups of load profiles with torque fluctuations, and the average torque is 1.51×10 5N·mm, with a standard deviation of 5%, to characterize the typical random load characteristics of aero-engines. An automated simulation process is constructed through the APDL command stream. The inner spline is set at a constant rotational speed of 1500 RPM, and the outer spline is synchronously applied with a reverse torque load to reproduce the dynamic force transmission boundary conditions of the drive system. In each group of simulations, the Archard wear model is integrated into the transient dynamics solver through the APDL command stream to construct a grid iterative reconstruction framework: with a load step of 0.1 unit time, the contact stress and sliding distance of the contact surface nodes are extracted in each load step, and based on the formula to calculate the wear increment (N ij = 1000 is the set time magnification factor, and k = 51 is the set number of load steps); then the *NMOVE command of APDL is called, and the geometric morphology of the contact surface is dynamically reconstructed according to the normal wear increment Δh j of the nodes, and the process of simulated material loss is displayed. Each group of simulations lasts for 5.1 unit times (equivalent to 85 minutes of actual operation) and goes through 51 wear-geometry update cycle iterations. Finally, relying on the HPC parallel computing cluster, 100 groups of simulations are completed synchronously to construct a dataset of the maximum wear depth covering the random load spectrum.

[0048] In the network model training stage, the 100 groups of simulation data are divided into a training set (80 groups), a test set (10 groups), and a validation set (10 groups) in the ratio of 8:1:1. Based on Figure 3 the shown architecture, a Bayesian neural network containing three layers of long short-term memory hidden layers is constructed, and the initial parameters are set as the number of hidden layer neurons 64, Dropout rate 0.2, learning rate 0.001, and regularization coefficient 1e-3. The Bayesian optimization algorithm is used to globally search for key hyperparameters: the number of hidden layer neurons traverses and combines in the range of [32 - 128], and the Dropout rate traverses and combines in the range of [0.1 - 0.5], and the optimal network structure is determined with the goal of minimizing the mean-square error (MSE) of the test set. Further, an improved prediction interval coverage rate index IPICP (α = 0.1 is the interval width penalty coefficient) is introduced for joint tuning. When the IPICP is lower than 0.9, the Dropout rate is reduced in steps of 0.05 and the learning rate is increased by 20% through a dynamic adjustment strategy, and finally the optimal parameter combination is obtained: learning rate 0.0012, Dropout rate 0.15, regularization coefficient 1.2e-3.

[0049] After the model training is completed, its probability prediction performance is systematically evaluated on the validation set. As Figure 4 shown, with a torque of 1.55×10 5Taking the wear data under the N·mm working condition as an example, after inputting the data of the first 68 minutes (accounting for 80% of the total data duration), the model can synchronously output the predicted value of the mean wear depth and the uncertainty interval for the next 17 minutes (accounting for 20% of the total data duration) with a 95% (2σ) confidence level. The full validation set test shows that the model's prediction of the mean wear depth for the next 17 minutes reaches MSE = 0.0363 and the coefficient of determination R 2 = 0.9177; in terms of interval prediction, the coverage rate of the 95% (2σ) confidence interval reaches 94.7%, and the average bandwidth is ±0.011 mm. This method demonstrates excellent prediction accuracy and uncertainty quantification ability under typical torque fluctuation working conditions of aeroengines, can generate probabilistic wear prediction results 17 minutes in advance, and provides high-confidence data support for condition-based maintenance decisions of the transmission system.

Claims

1. A method for predicting the probability of wear of mechanical transmission components under parallel misalignment state, characterized in that Follow these steps in order: S1: Based on the specific dimensional parameters and material properties of the mechanical transmission components, a three-dimensional mechanical transmission component model including parallel misalignment error is constructed; S2: Conduct in-depth analysis of the working loads of mechanical transmission components to determine the load profile; fully consider the random uncertainty in the load profile and use statistical methods to quantify the uncertainty of the load profile parameters; sample the load profile parameters; Using all the sampled load profile parameters as input conditions, wear simulation analysis based on iterative reconstruction of mesh nodes was carried out in the finite element analysis software to obtain the wear growth data set of mechanical transmission components under different load profiles. S3: using the wear growth data set obtained in step S2, training a Bayesian long short-term memory network model for wear probability prediction, and using the Dropout technique to model the uncertainty of the weight parameters in the network; S4: Use the trained Bayesian long short-term memory network model to predict the expected value and uncertainty interval of the wear of mechanical transmission components in the future short-term period of time.

2. The method for predicting the probability of wear of mechanical transmission components under parallel misalignment state according to claim 1 is characterized in that: The wear simulation analysis method based on iterative reconstruction of grid nodes described in step S2 is specifically as follows: S2.1: In the transient simulation module of ANSYS software, define the initial load step increment Δt1; use the APDL command to integrate the Archard wear model into the simulation post-processing module; after completing the transient simulation of Δt1, obtain the contact stress and sliding velocity of all nodes on the contact surface of the mechanical transmission component, and calculate the wear depth Δh of all nodes 1j ; Where j is the number of nodes on the contact surface; S2.2: To simplify the calculation and improve the simulation efficiency, the wear depth of each node is set to change linearly with time in a short period of time, and the wear acceleration phenomenon during this period is ignored; therefore, within the duration of N1×Δt1, the wear depth of each node can be expressed as N1×Δh 1j ; S2.3: Use the APDL command to move all nodes along the normal direction of the contact surface by N1×Δh 1j , simulating material loss during wear; S2.4: Set the next load step increment Δt2, re-run the wear simulation, and determine the wear depth Δh of each node 2j ; Repeat steps S2 and S3; S2.5: Finally, within the total running time Δt, the accumulated wear depth Δh of each node j and the maximum wear depth Δh of mechanical transmission parts max Calculated by the following formula: Δh max =max{Δh j } Here, k is the number of load steps.

3. The method for predicting the probability of wear of mechanical transmission components under parallel misalignment state according to claim 1 is characterized in that: The training method of the Bayesian long short-term memory network model described in step S3 is: S3.1: Set the initial values ​​of network parameters; perform preliminary parameter range screening based on grid search, and combine Bayesian optimization to determine the current optimal number of long and short-term memory layers, number of hidden layers, learning rate, and regularization coefficient of the network; S3.2: Based on the improved prediction interval coverage probability (IPICP) indicator, the uncertainty interval of the network output is evaluated to guide the adjustment of the dropout rate and the re-adjustment of the learning rate and regularization coefficient; S3.3: Evaluate the performance of the network under different data amounts by gradually increasing the amount of training data and determine the shortest acceptable training data duration; S3.4: Based on the above optimization steps, the training of the Bayesian long short-term memory network is finally completed and applied to the short-term probability prediction of wear of mechanical transmission components.

4. The training method of the Bayesian long short-term memory network model according to claim 4 is characterized in that: The specific implementation steps for adjusting the Dropout rate, learning rate, and regularization coefficient based on the IPICP indicator in step S3.2 are as follows: S3.2.1: First, the coverage of observations within the uncertainty interval of the current network output is calculated using the IPICP indicator, and an interval width penalty term is introduced to evaluate the rationality of the uncertainty range. The IPICP indicator is: Among them, N is the number of prediction points output by the network, c i Indicates whether the observed value of the prediction point i falls within the uncertainty interval. If so, c i =1, otherwise c i =0; and They represent the upper and lower bounds of the uncertainty interval respectively; α is the penalty coefficient for controlling the influence of the width of the uncertainty interval; S3.2.2: Then, iteratively adjust the Dropout rate, learning rate, and regularization coefficient to optimize the IPICP indicator: When the IPICP index is less than 0.9, the Dropout rate is reduced by 0.05 steps, the learning rate is increased by 20%, and the regularization coefficient is increased by 20% to narrow the prediction interval; when the IPICP index is greater than or equal to 0.9, the Dropout rate, learning rate, and regularization coefficient remain unchanged.

Images