Method for predicting wear of aircraft door hinge based on dynamic wear coefficient

Abstract

The present application mainly relates to the technical field of aircraft maintenance. In order to solve the problem of low prediction accuracy of hinge wear under the scene of small sample, variable working condition, individual difference and equipment degradation, the present application provides a kind of aircraft door hinge wear prediction method based on dynamic wear coefficient, the core is based on real wear coefficient as training label, constructs physical information embedding type neural network, the working condition parameters of hinge are as input, the real wear coefficient is as training label and is trained to obtain the wear coefficient prediction model, after training, the working condition parameters collected in real time are input into the model to obtain the dynamic wear coefficient that changes with the working condition in real time, and the wear depth is recursively calculated based on the dynamic wear coefficient using discrete integration method;At the same time, the measured wear depth obtained by periodic maintenance is used to realize the adaptive correction of the model by weighting fusion and scaling update of the output layer weight of the model, and the high-precision adaptive prediction result of the wear depth of the aircraft door hinge is obtained.

Description

Technical Field

[0001] This invention relates primarily to the field of aircraft maintenance technology, and in particular to a method for predicting the wear of aircraft door hinges based on a dynamic wear coefficient. Background Technology

[0002] Hinges, as a crucial component of aircraft door mechanisms, are characterized by high load capacity, highly random parameters, and dynamic load variations during operation. Over long-term service, factors such as friction between contact surfaces, load variations, and fluctuations in lubrication conditions lead to progressive wear between the hinge pin and bushing. This results in increased clearance, decreased motion accuracy, and in severe cases, may cause door mechanism malfunctions or even safety accidents. Therefore, predicting the wear depth at the hinge is a necessary step in preventative maintenance of aircraft doors.

[0003] Traditional wear prediction methods, including the Archard model, finite element analysis, and fracture mechanics models, are commonly used when predicting the wear depth of hatch door hinges. These models are well-interpreted, have strong physical meaning, and are maturely applied. However, for hatch door hinges, the loads, rotational speeds, temperatures, and lubrication conditions they experience during actual operation change dynamically over time. Traditional models, due to the static nature of their parameters, struggle to analyze complex and time-varying operating conditions, resulting in significant deviations in prediction results.

[0004] In recent years, data-driven methods have emerged and have achieved high accuracy in predicting the wear depth of hatch hinges. However, they have the following drawbacks: 1) There is a risk of overfitting when the sample size is small; 2) They cannot solve the problem of parameter drift as the equipment ages and the differences in individual manufacturing, and extrapolation may fail.

[0005] Therefore, there is an urgent need for a prediction method that can accurately predict wear coefficients and adaptively adjust them even with limited real-time maintenance data, in order to solve the problem of low accuracy in hinge wear prediction under conditions of small sample size, variable operating conditions, individual differences, and equipment degradation. Summary of the Invention

[0006] This invention provides a method for predicting the wear of aircraft door hinges based on dynamic wear coefficients, aiming to solve the problem of low accuracy in hinge wear prediction under scenarios with small samples, variable operating conditions, individual differences, and equipment degradation.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems

[0008] A method for predicting the wear of aircraft door hinges based on dynamic wear coefficients, the method comprising:

[0009] Step 1: Obtain the hinge's operating parameters and actual wear coefficient;

[0010] Step 2: Construct a physical information embedded neural network, using the hinge's operating parameters as input and the actual wear coefficient obtained in Step 1 as training labels to train the constructed neural network and establish a wear coefficient prediction model.

[0011] Step 3: Input the hinge operating condition parameters collected in real time into the trained wear coefficient prediction model to obtain the dynamic wear coefficient that changes in real time with the operating conditions, and recursively calculate the wear depth based on the dynamic wear coefficient.

[0012] Step 4: Adaptively update the output layer weights of the wear coefficient prediction model based on the measured wear depth to correct subsequent prediction results.

[0013] Furthermore, in step 1, the actual wear coefficient of the hinge is calculated using the Archard model:

[0014] ;

[0015] In the formula, This represents the actual wear coefficient. For time, For wear depth, This is a conversion value for the Vickers hardness or Brinell hardness of the softer material in the bushing or pin. for Normal load at time t, for The relative sliding velocity at any given moment.

[0016] Furthermore, the operating parameters include: normal load, relative sliding speed, friction temperature, lubricating oil film thickness, and vibration acceleration.

[0017] Furthermore, step 2 also includes preprocessing the collected operating parameters. The preprocessing includes filtering and denoising the collected operating parameters, extracting statistical feature values, and normalizing them after aligning them according to the hinge opening and closing cycle.

[0018] Furthermore, the construction of the physical information embedded neural network in step 2 includes: adding the differential equation of the Arcard model as a physical constraint term to the loss function, and introducing a physical prior bias matrix based on instantaneous friction power into the attention mechanism.

[0019] Furthermore, the loss function for:

[0020] ;

[0021] In the formula, For the sample size, For time, The first prediction of the neural network Wear coefficient of each sample For the first The actual wear coefficient of each sample These are the weighting coefficients for the physical constraint terms. To predict wear depth, The scaling constant of the Archard model. The wear coefficient at the current moment is predicted by the neural network. The normal load at the current moment, The relative sliding speed at the current moment, The weight coefficients of the regularization term, These are the trainable parameters of the neural network. This is an L2 regularization term.

[0022] Furthermore, the physical prior bias matrix for:

[0023] ;

[0024] In the formula, For the current moment, For a historic moment, For learnable parameters relevant to the current time, for Normal load at time t, for The relative sliding velocity at any given moment.

[0025] Furthermore, step 3, which involves recursively calculating the wear depth based on the dynamic wear coefficient, includes updating the wear depth using a discrete integration method based on the wear coefficient, normal load, relative sliding speed, and sampling time interval at the current moment, and accumulating the cumulative wear depth within each hinge opening and closing cycle after the end of each cycle.

[0026] Furthermore, step 4 specifically includes:

[0027] The ratio of the measured wear depth increment to the predicted wear depth increment within each maintenance cycle is calculated as the gain factor, and the arithmetic mean of the gain factors for multiple maintenance cycles is taken to obtain the scaling factor.

[0028] Based on the set forgetting coefficient, the weights of the initial output layer of the wear coefficient prediction model are weighted and fused with the weights of the current output layer to obtain the protected weights;

[0029] The protected weights are scaled based on the scaling factor to obtain the updated output layer weights.

[0030] Furthermore, the updated output layer weights are:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] In the formula, For the updated output layer weights, Forgetting coefficient, The protected weights after weighted fusion The weights of the current output layer of the wear coefficient prediction model. The initial output layer weights for the wear coefficient prediction model. Scaling factor For gain factor, To measure the increase in wear depth, To predict the increase in wear depth, This is the maintenance cycle. This is an index for the maintenance cycle number.

[0036] Beneficial effects of the present invention

[0037] 1. This invention achieves high-precision adaptive prediction of the wear depth of aircraft door hinges by constructing a physical information embedded neural network to predict the dynamic wear coefficient that changes in real time with the working conditions, and recursively calculating the wear depth based on the dynamic wear coefficient.

[0038] 2. By introducing a physical prior bias matrix based on instantaneous friction power into the attention mechanism, the model can actively focus on historical moments with high wear risk when calculating attention weights, thereby improving the model's sensitivity to harsh working conditions and prediction accuracy. By adding an Arcard differential equation as a physical constraint term to the loss function, the output of the neural network must conform to the physical laws of wear, which not only enhances the interpretability of the model but also reduces the dependence on a large amount of training data and improves the generalization ability under small sample conditions.

[0039] 3. Based on the measured maintenance data, the gain factor and scaling factor are calculated, and the weights of the model output layer are weighted and scaled with a forgetting coefficient. This can effectively eliminate individual differences in hinges of different aircraft and parameter drift caused by long-term service. Attached Figure Description

[0040] Figure 1 This is a flowchart of the aircraft door hinge wear prediction method based on dynamic wear coefficient described in this invention.

[0041] Figure 2 This is a schematic diagram of the wear coefficient prediction network model;

[0042] Figure 3 A comparison chart of the predicted cumulative wear depth. Detailed Implementation

[0043] In predicting the wear depth of aircraft door hinges, traditional constant coefficient models cannot adapt to time-varying operating conditions, and pure data-driven methods are prone to overfitting and extrapolation failure under small sample sizes.

[0044] The core of this invention in solving the aforementioned technical problems is as follows: A physical information embedded neural network is constructed based on the actual wear coefficient as the training label. The hinge's operating parameters are used as input, and the actual wear coefficient is used as the training label for model training, resulting in a wear coefficient prediction model. After model training, real-time collected operating parameters are input into the model to obtain a dynamic wear coefficient that changes with the operating conditions in real time. The wear depth is then recursively calculated using discrete integration based on the dynamic wear coefficient. Simultaneously, using the measured wear depth obtained from periodic maintenance, the model's output layer weights are weighted and scaled with a forgetting coefficient to achieve adaptive model correction. This enables high-precision adaptive prediction of the wear depth of aircraft door hinges.

[0045] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0046] like Figure 1 As shown, the aircraft door hinge wear prediction method based on dynamic wear coefficient of the present invention includes the following steps:

[0047] Step S1: Obtain the true wear coefficient of the hinge

[0048] This embodiment takes the hinge at the junction of the cabin door hinge arm and the fuselage of a certain type of civil aircraft as the monitoring object. The hinge uses a stainless steel pin and a self-lubricating polytetrafluoroethylene composite bushing. The working environment has large temperature and humidity changes and frequent opening and closing, averaging 5 times a day. It is prone to increased clearance due to fretting wear, which affects the sealing performance and operating torque.

[0049] Accelerated wear tests were conducted on the same model of hinge under standard operating conditions. The standard operating conditions were set as follows: ambient temperature 25±2 degrees Celsius, relative humidity 50±5%RH, normal load 600 N, sliding speed 0.1 m / s, single stroke length 15 mm, motion mode reciprocating sliding frequency 1 Hz, lubrication condition dry friction self-lubrication, and a total of 10,000 cycles. The pin material was 1Cr18Ni9Ti stainless steel with a hardness coefficient of [missing information]. It has a compressive strength of 75D, approximately 80 MPa. The measured wear depth will be... Substituting into the Archard model, the initial wear coefficient can be derived. The specific formula for the actual wear coefficient is as follows:

[0050] ;

[0051] In the formula, This represents the actual wear coefficient. For time, This is a conversion value for the Vickers hardness or Brinell hardness of the softer material in the bushing or pin. for Normal load at time t, for The relative sliding velocity at any given moment.

[0052] Step S2: Construct a physical information embedded neural network

[0053] The differential equation of the Arcard model is added as a physical constraint term to the loss function of the neural network, and a physical prior bias matrix based on instantaneous friction power is introduced into the attention mechanism.

[0054] like Figure 2 As shown, the loss function The loss function consists of three parts: prediction loss, physical loss, and regularization. The specific expression is:

[0055] ;

[0056] In the formula, For the sample size, For time, The first prediction of the neural network Wear coefficient of each sample For the first The actual wear coefficient of each sample These are the weighting coefficients for the physical constraint terms. To predict wear depth, The scaling constant of the Archard model. The wear coefficient at the current moment is predicted by the neural network. The normal load at the current moment, The relative sliding speed at the current moment, The weight coefficients of the regularization term, These are the trainable parameters of the neural network. This is an L2 regularization term.

[0057] Physical prior bias matrix The expression is: ;

[0058] In the formula, For the current moment, For a historic moment, For learnable parameters relevant to the current time, for Normal load at time t, for The relative sliding velocity at any given moment.

[0059] The attention weights after adding the bias are:

[0060] ;

[0061] In the formula, The feature vector after attention weighting , , These are respectively query, key, and value vector. For normalized exponential functions, This is the scaling factor.

[0062] Step S3: Create a training dataset

[0063] The operating parameters used in this invention for predicting the wear coefficient include normal load. Relative sliding speed Friction interface temperature Lubricating oil film thickness Vibration acceleration.

[0064] The operating parameters are obtained through the following methods: a six-dimensional force sensor is installed on the hinge support to measure the normal load and tangential friction force. A high-precision encoder is installed at the drive link position to acquire angular displacement, and the relative sliding speed is obtained by differentiation. An infrared thermal imager and a spot thermometer are installed to monitor the friction interface temperature. An eddy current displacement sensor is installed to indirectly measure the thickness of the lubricating oil film (or dry film lubrication layer). A triaxial MEMS vibration sensor is attached to the hinge housing to collect triaxial vibration acceleration signals at a sampling frequency of 200 Hz.

[0065] During hinge operation, sampling intervals of 50 opening and closing cycles are used to collect operating parameters and measure wear depth as sample points.

[0066] The collected operating parameters are preprocessed, including filtering and denoising, feature extraction, data alignment and normalization.

[0067] Filtering and Denoising: A second-order Butterworth low-pass filter is used to filter the normal load and tangential friction force. Considering that the main frequency of the hatch opening and closing action is less than 5 Hz and the impact transient is less than or equal to 15 Hz, the cutoff frequency fc is set to 10 Hz, the sampling frequency fs is set to 100 Hz, and zero-phase filtering is used. Wavelet soft-threshold denoising is used for the vibration acceleration signal, with the wavelet basis function being Daubechies4, to match non-stationary mechanical vibration signals, retaining impact characteristics greater than 30 Hz such as lock hook impacts, while suppressing broadband noise. A moving average filter is used for the temperature signal, with a window length of 5 and a sampling frequency of 1 Hz.

[0068] Feature extraction: Statistical features within each sampling period are extracted, including the maximum and minimum values ​​of normal load force, velocity, temperature, and lubrication layer thickness. Simultaneously, wavelet packet decomposition is performed on the vibration acceleration signal to extract energy entropy as a wear-sensitive feature.

[0069] Data alignment and normalization: The collected operating parameters are aligned according to the start and stop cycles to form a structured data matrix, and the input features are normalized using Min-Max.

[0070] The preprocessed operating parameters are mapped to the wear coefficients calculated based on the actual wear depth to form a dataset.

[0071] Step S4: Train the physical information embedded neural network to obtain the wear coefficient prediction model.

[0072] The dataset from step 3 is divided into a training set and a test set in a 7:3 ratio. The physical information embedding neural network constructed in step 2 is trained using the training set.

[0073] Input feature vector By projecting learnable linear transformations onto the fundamental components of the query, key, and value vectors, the query (Q) vector is physically guided and enhanced.

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] In the formula, For physical guidance enhancement vectors, For linear transformation operations, The wavelet packet energy entropy of the vibration acceleration signal.

[0079] During neural network training, the Adam optimizer was selected for optimization, with an initial learning rate of 3×10^{-4} and cosine annealing scheduling. Mini-batch training was employed, with a batch size of 32, and training was conducted for 200 epochs. After convergence, the optimal model was saved, and the predictive performance of the model was verified using test set data to obtain the wear coefficient prediction model.

[0080] Step S5: Recursive calculation of wear depth

[0081] The hinge operating condition parameters collected in real time are input into the trained wear coefficient prediction model to obtain the dynamic wear coefficient that changes with the operating conditions in real time, and the wear depth is recursively calculated based on the dynamic wear coefficient:

[0082] ;

[0083] In the formula, The time interval between adjacent sampling points. For the first The cumulative wear depth at each sampling time. For the first The cumulative wear depth at each sampling time. This represents the wear coefficient output by the wear coefficient prediction model at the previous sampling time. The normal load at the previous sampling time. The relative sliding velocity at the previous sampling time, initial value .

[0084] After calculating the wear depth for each opening and closing cycle, the cumulative wear amount for that cycle is calculated, and the total wear depth is calculated by summing them up.

[0085] Step S6: Adaptively update the output layer weights of the wear coefficient prediction model based on the measured wear depth.

[0086] The ratio of the measured wear depth increment to the predicted wear depth increment within each maintenance cycle is calculated as the gain factor, and the arithmetic mean of the gain factors for multiple maintenance cycles is taken to obtain the scaling factor.

[0087] Based on the set forgetting coefficient, the weights of the initial output layer of the wear coefficient prediction model are weighted and fused with the weights of the current output layer to obtain the protected weights;

[0088] The protected weights are scaled based on the scaling factor to obtain the updated output layer weights.

[0089] Furthermore, the updated output layer weights are:

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] In the formula, For the updated output layer weights, Forgetting coefficient, The protected weights after weighted fusion The weights of the current output layer of the wear coefficient prediction model. The initial output layer weights for the wear coefficient prediction model. Scaling factor For gain factor, To measure the increase in wear depth, To predict the increase in wear depth, This is the maintenance cycle. This is an index for the maintenance cycle number.

[0095] Through the above steps, the hatch mechanism is opened and closed for 10 seconds. 3 Subsequently, in this embodiment, a comparison is made between the cumulative wear depth of the hinge measured by the aircraft door hinge wear prediction method based on the dynamic wear coefficient described in this invention, the cumulative wear depth of the hinge obtained by the Archard model prediction method, and the experimental wear depth value obtained by wear experiments. Figure 3 As shown, the method of the present invention achieves high-precision adaptive prediction of the wear depth of aircraft door hinges, which can be integrated into the aircraft health management system to provide a reliable decision-making basis for the preventive maintenance of door hinges.

Claims

1. A method for predicting the wear of aircraft door hinges based on dynamic wear coefficient, characterized in that, The method includes: Step 1: Obtain the hinge's operating parameters and actual wear coefficient; Step 2: Construct a physical information embedded neural network, using the hinge's operating parameters as input and the actual wear coefficient obtained in Step 1 as training labels to train the constructed neural network and establish a wear coefficient prediction model. Step 3: Input the hinge operating condition parameters collected in real time into the trained wear coefficient prediction model to obtain the dynamic wear coefficient that changes in real time with the operating conditions, and recursively calculate the wear depth based on the dynamic wear coefficient. Step 4: Adaptively update the output layer weights of the wear coefficient prediction model based on the measured wear depth to correct subsequent prediction results.

2. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 1, characterized in that, In step 1, the actual wear coefficient of the hinge is calculated using the Archard model: ； In the formula, This represents the actual wear coefficient. For time, For wear depth, This is a conversion value for the Vickers hardness or Brinell hardness of the softer material in the bushing or pin. for Normal load at time t, for The relative sliding velocity at any given moment.

3. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 1, characterized in that, The operating parameters include: normal load, relative sliding speed, friction temperature, lubricating oil film thickness, and vibration acceleration.

4. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 1, characterized in that, Step 2 also includes preprocessing the collected operating parameters. The preprocessing includes filtering and denoising the collected operating parameters, extracting statistical feature values, and normalizing them after aligning them according to the hinge opening and closing cycle.

5. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 1, characterized in that, The construction of the physical information embedded neural network in step 2 includes: adding the differential equation of the Arcard model as a physical constraint term to the loss function, and introducing a physical prior bias matrix based on instantaneous friction power into the attention mechanism.

6. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 5, characterized in that, The loss function for: ； In the formula, For the sample size, For time, The first prediction of the neural network Wear coefficient of each sample For the first The actual wear coefficient of each sample These are the weighting coefficients for the physical constraint terms. To predict wear depth, The scaling constant of the Archard model. The wear coefficient at the current moment is predicted by the neural network. The normal load at the current moment, The relative sliding speed at the current moment, The weight coefficients of the regularization term, These are the trainable parameters of the neural network. This is an L2 regularization term.

7. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 5, characterized in that, The physical prior bias matrix for: ； In the formula, For the current moment, For a historic moment, For learnable parameters relevant to the current time, for Normal load at time t, for The relative sliding velocity at any given moment.

8. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 1, characterized in that, Step 3, which involves recursively calculating the wear depth based on the dynamic wear coefficient, includes updating the wear depth using a discrete integration method based on the wear coefficient, normal load, relative sliding speed, and sampling time interval at the current moment, and accumulating the cumulative wear depth within each hinge opening and closing cycle after the end of each cycle.

9. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 1, characterized in that, Step 4 specifically includes: The ratio of the measured wear depth increment to the predicted wear depth increment within each maintenance cycle is calculated as the gain factor, and the arithmetic mean of the gain factors for multiple maintenance cycles is taken to obtain the scaling factor. Based on the set forgetting coefficient, the weights of the initial output layer of the wear coefficient prediction model are weighted and fused with the weights of the current output layer to obtain the protected weights; The protected weights are scaled based on the scaling factor to obtain the updated output layer weights.

10. The method for predicting aircraft door hinge wear based on dynamic wear coefficient according to claim 9, characterized in that, The updated output layer weights are: ； ； ； ； In the formula, For the updated output layer weights, Forgetting coefficient, The protected weights after weighted fusion The weights of the current output layer of the wear coefficient prediction model. The initial output layer weights for the wear coefficient prediction model. Scaling factor For gain factor, To measure the increase in wear depth, To predict the increase in wear depth, This is the maintenance cycle. This is an index for the maintenance cycle number.

Images