A landing gear vibration monitoring method, server, product, and medium

By combining three parallel feature extraction branches and a cross-modal attention fusion layer, the problem of landing gear health monitoring being unable to effectively capture multi-scale features and gradient decay in existing technologies is solved, enabling multi-dimensional comprehensive assessment and accurate diagnosis of landing gear health status.

CN123115758APending Publication Date: 2026-11-17BAI JING HANG XIAN (CHANG ZHOU) KE JI YOU XIAN GONG SI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611288417.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, deep learning-based landing gear health monitoring methods cannot effectively capture high-frequency impact components and low-frequency trend components in landing vibration signals. Furthermore, the gradient decay of long short-term memory networks in long sequence recursion leads to a decrease in diagnostic accuracy, especially under extreme conditions.

Method used

Three parallel feature extraction branches are used to capture local multi-scale features, global dependency features, and temporal evolution features respectively. These features are then complemented and enhanced by a cross-modal attention fusion layer. Combined with a multi-task output branch, fault classification, health index, and anomaly score are output synchronously. Multiple sensors are used to collect multi-channel signals for denoising filtering and time-frequency transformation to achieve a unified representation of features.

Benefits of technology

It improves the diagnostic accuracy and prediction reliability of landing gear health monitoring under all operating conditions, and can complete multi-dimensional comprehensive evaluation within a single model framework, capturing multi-granular and multi-view vibration information, solving the problems of single-scale feature extraction and gradient decay in existing technologies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115758A_ABST
    Figure CN123115758A_ABST
Patent Text Reader

Abstract

A landing gear vibration monitoring method, server, product, and medium are disclosed, relating to the field of aviation equipment fault diagnosis. The method includes: sequentially denoising and filtering multi-channel raw signals from landing gear operation and performing time-frequency transformation to obtain standardized time-frequency input data; feeding this data in parallel into three feature extraction branches, processing it with different extraction mechanisms, and outputting corresponding feature vectors; inputting the feature vectors from each branch into a cross-modal attention fusion layer, using a single-path vector as the query end in turn, concatenating the remaining vectors as the key and value ends, and weighting and aggregating the value ends according to the query-key end attention weights to obtain one-to-one corresponding cross-enhanced feature vectors, merging them to obtain a fused feature vector; and feeding the fused feature vector and standardized time-frequency input data into multiple output branches to comprehensively generate a landing gear health status assessment result. Implementing this application can improve the diagnostic accuracy of landing gear health monitoring under all operating conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of fault diagnosis of aviation equipment, and in particular to a landing gear vibration monitoring method, server, product and medium. Background Technology

[0002] Aircraft landing gear withstands repeated high-intensity impact loads during takeoff, landing, and taxiing, making it one of the most critical safety components of an aircraft. Real-time health monitoring of landing gear to enable predictive maintenance has become an urgent need to improve aviation safety and reduce operational and maintenance costs.

[0003] In related technologies, the landing gear health monitoring method based on deep learning adopts an architecture of serial cascading convolutional neural networks and long short-term memory networks. It takes vertical acceleration signals as input, extracts local waveform features through fixed-size convolutional kernels, and then sends them to the long short-term memory network for recursive processing step by step to output the fault category judgment result.

[0004] However, fixed-size convolutional kernels can only capture features at a single scale and cannot effectively extract both high-frequency impact components and low-frequency trend components from landing vibration signals simultaneously. Furthermore, long short-term memory networks experience gradient decay during long-sequence recursive propagation, making it difficult to effectively model long-distance temporal dependencies. These limitations are particularly pronounced under extreme conditions such as hard landings, leading to a significant decrease in diagnostic accuracy. Summary of the Invention

[0005] This application provides a landing gear vibration monitoring method, server, product, and medium to improve the diagnostic accuracy of landing gear health monitoring under all operating conditions.

[0006] Firstly, this application provides a landing gear vibration monitoring method applied to a server. The method includes: acquiring multi-channel raw signals during landing gear operation; sequentially performing denoising filtering and time-frequency transformation on the multi-channel raw signals to obtain standardized time-frequency input data; inputting the standardized time-frequency input data in parallel into three feature extraction branches, each outputting a corresponding feature vector, wherein each feature extraction branch processes the standardized time-frequency input data using different feature extraction mechanisms; inputting the feature vectors output by each feature extraction branch into a cross-modal attention fusion layer, where the feature vectors output by each feature extraction branch are integrated into the cross-modal attention fusion layer. Vectors are sequentially designated as query ends. Each time, the feature vectors of the current non-query end are concatenated to serve as the key and value ends. The value ends are weighted and aggregated according to the attention weight between the query end and the key end to obtain cross-enhanced feature vectors corresponding one-to-one with each feature extraction branch. All cross-enhanced feature vectors are merged to output a fused feature vector. The fused feature vector and standardized time-frequency input data are input into multiple output branches, which output classification probability distribution, health index, remaining life prediction value, and anomaly score, respectively. Based on the classification probability distribution, health index, remaining life prediction value, and anomaly score, the landing gear health status assessment result is generated.

[0007] In the above embodiments, three parallel feature extraction branches capture local multi-scale features, global dependency features, and temporal evolution features respectively. The feature vectors of each branch are complemented and enhanced by a cross-modal attention fusion layer, so that the fused feature vector contains multi-granularity and multi-view vibration information. In conjunction with the multi-task output branch, four evaluation indicators, namely fault classification, health index, remaining life and anomaly score, are output simultaneously. This enables a multi-dimensional comprehensive evaluation of the landing gear health status within a single model framework, improving the diagnostic accuracy and prediction reliability under all operating conditions.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, the step of inputting standardized time-frequency input data into three feature extraction branches in parallel specifically includes: the first feature extraction branch extracts local multi-scale features of the standardized time-frequency input data through multiple sets of parallel convolution operations at different receptive field scales, and outputs a local feature vector after channel-dimensional concatenation and channel attention weighting; the second feature extraction branch performs multi-layer self-attention encoding operations on the standardized time-frequency input data with injected positional encoding information, calculates the dependency weights between all time positions, and outputs a global feature vector; the third feature extraction branch traverses the standardized time-frequency input data along both the forward and reverse time directions, extracts the hidden states at each time step, and outputs a temporal feature vector after temporal attention weighting aggregation.

[0009] In the above embodiments, the first feature extraction branch captures local features at different time scales, such as high-frequency impact components and low-frequency trend components, through multiple sets of parallel convolutions with different receptive field scales. The second feature extraction branch establishes global dependencies between all time positions through multi-layer self-attention encoding operations. The third feature extraction branch traverses the sequence along the bidirectional time direction and aggregates the hidden states at each time step through temporal attention weighting. The three branches complement each other in terms of feature granularity and modeling mechanism, enabling the model to simultaneously extract short-range waveform details, long-range global correlations, and ordered temporal evolution patterns.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, the step of sequentially performing denoising filtering and time-frequency transformation on the multi-channel raw signals to obtain standardized time-frequency input data specifically includes: acquiring multi-channel raw signals through at least two different types of sensors deployed at different locations on the landing gear, the multi-channel raw signals containing sensor data of at least two different physical quantities; performing denoising filtering on the multi-channel raw signals to obtain multi-channel denoised signals; performing time-frequency transformation on the multi-channel denoised signals to convert the one-dimensional time-domain multi-channel denoised signals into a two-dimensional time-frequency representation to obtain multi-channel time-frequency data; and performing cross-channel standardization processing on the multi-channel time-frequency data according to each sensor channel to obtain standardized time-frequency input data.

[0011] In the above embodiments, by deploying at least two different types of sensors at different locations on the landing gear to acquire multi-channel raw signals containing multiple physical quantities, after noise reduction and filtering to eliminate environmental interference, time-frequency transformation is performed to convert the one-dimensional time-domain signal into a two-dimensional time-frequency representation. Then, cross-channel standardization processing is performed on each sensor channel to eliminate the differences in dimensions and amplitudes between different physical quantities, so that the time-frequency data of each channel are input into the feature extraction network under a unified numerical scale, ensuring the effective fusion of multi-source heterogeneous sensor data and the accuracy of feature extraction.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the step of inputting the fused feature vector and standardized time-frequency input data into multiple output branches respectively includes: inputting the fused feature vector into a fault classification output head after fully connected mapping, outputting the classification probability distribution of each preset fault category after probability normalization, as the fault classification result; inputting the fused feature vector into a health status regression output head after fully connected mapping, outputting the health index and remaining life prediction value, as the health status prediction result; inputting the standardized time-frequency input data into the anomaly detection branch, compressing and mapping it to a low-dimensional latent space through an encoding network, and then reconstructing it through a decoding network to obtain the reconstructed signal, calculating the reconstruction error between the reconstructed signal and the standardized time-frequency input data, and using the reconstruction error as the anomaly score.

[0013] In the above embodiments, the fault classification output head outputs the classification probability distribution of each preset fault category through fully connected mapping and probability normalization. The health status regression output head outputs the health index and remaining life prediction value in continuous numerical form. The anomaly detection branch uses the autoencoder reconstruction error as an anomaly score to detect unknown fault modes. The three output branches are respectively oriented towards three types of tasks: discrete classification, continuous regression, and unsupervised anomaly detection. This enables the model to simultaneously output multi-dimensional evaluation results in a single forward inference, covering two application requirements: known fault identification and unknown anomaly detection.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, after merging all cross-enhanced feature vectors and outputting a fused feature vector, the method further includes: defining a landing condition with labeled data as a source domain and a landing condition to be adapted as a target domain, wherein the source domain and the target domain differ in at least one of landing speed, landing load, or runway conditions; connecting a domain discrimination branch to the output of the fused feature vector, inputting the fused feature vector into the domain discrimination branch, and determining whether the domain label of the fused feature vector is the source domain or the target domain; applying a direction reversal to the gradient back from the domain discrimination branch to the feature extraction branch, so that the feature extraction branch cannot distinguish the domain-invariant features of the domain source in the domain discrimination branch; grouping the source domain samples and the target domain samples according to the fault category, calculating the conditional feature distribution distance between the source domain and the target domain for each category, and incorporating the sum of the conditional feature distribution distances of all categories as the category-aware alignment loss into the total loss function, so that the feature distribution of the same fault category tends to be consistent under different landing conditions.

[0015] In the above embodiments, by combining the domain discriminative branch with the gradient reversal mechanism, the feature extraction branch generates domain-invariant feature representations that the domain discriminative branch cannot distinguish from the source during adversarial training. At the same time, the conditional feature distribution distance between the source domain and the target domain is calculated according to the fault category and included in the total loss function. On the basis of achieving global domain alignment, category-level constraints are applied to make the feature distribution of the same fault category tend to be consistent under different landing speeds, landing loads or runway conditions, thereby improving the model's fault diagnosis generalization ability in cross-condition scenarios.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, after the steps of outputting fault classification results, health status prediction results, and anomaly scores, the method further includes: calculating the classification loss value corresponding to the fault classification result, the regression loss value corresponding to the health status prediction result, and the reconstruction loss value corresponding to the anomaly score, respectively; configuring learnable uncertainty parameters for each loss value, using the reciprocal of the square of each uncertainty parameter as the weight coefficient of the corresponding loss value, using the logarithm of each uncertainty parameter as the regularization term, summing the weighted loss values ​​with each regularization term to construct a total loss function, synchronously updating the network parameters and each uncertainty parameter of the feature extraction branch and the output branch during backpropagation; arranging the anomaly scores of multiple consecutive flight cycles in chronological order to form a scoring sequence, performing trend detection on the scoring sequence, and generating a degradation warning signal.

[0017] In the above embodiments, by configuring learnable uncertainty parameters for classification loss, regression loss and reconstruction loss respectively, and using the reciprocal of the square of each uncertainty parameter as the corresponding loss weight and its logarithm as the regularization term, the loss weight ratio is dynamically adjusted according to the learning difficulty of each task during backpropagation, avoiding gradient conflicts between tasks caused by manually setting fixed weights; at the same time, the abnormal scores of multiple consecutive flight cycles are constructed into a time series and trend detection is performed, so that the model can capture the degradation trend and generate early warning signals before the fault is fully manifested.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, before the step of performing trend detection on the scoring sequence to generate a degradation warning signal, the method further includes: generating an immediate alarm signal when the abnormal score of a single flight cycle exceeds a preset abnormal threshold; the step of performing trend detection on the scoring sequence to generate a degradation warning signal specifically includes: generating a degradation warning signal when the scoring sequence shows a continuous monotonically increasing trend within a sliding window consisting of a preset number of consecutive flight cycles, wherein the generation time of the degradation warning signal is earlier than the time node when the fault is fully manifested.

[0019] In the above embodiments, an immediate alarm signal is generated when the abnormal score of a single flight cycle exceeds a preset abnormal threshold to respond to sudden abnormal events. A degradation warning signal is generated when the score sequence shows a continuous monotonically increasing trend within a sliding window consisting of a preset number of consecutive flight cycles. The generation time of the degradation warning signal is earlier than the time node when the fault is fully manifested, so that the maintenance decision system can obtain early warning prompts before the fault is fully developed, providing time margin for the formulation of maintenance plans and taking into account both the rapid response to sudden abnormalities and the early prediction of gradual degradation.

[0020] In a second aspect, embodiments of this application provide a server comprising: one or more processors and a memory; the memory is coupled to the one or more processors and is used to store computer program code, the computer program code including computer instructions, wherein the one or more processors invoke the computer instructions to cause the server to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a server, cause the server to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a server, cause the server to execute the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the server provided in the second aspect, the computer storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. By employing a technique that involves parallel input of standardized time-frequency input data into three feature extraction branches operating with different feature extraction mechanisms, and then using a cross-modal attention fusion layer to complement and enhance the feature vectors of each branch through a cross-attention mechanism before outputting a fused feature vector, the fused feature vector can simultaneously encode three types of complementary information: local multi-scale features, global dependency features, and temporal evolution features. This effectively solves the problem that a single architecture in existing technologies cannot simultaneously capture multi-granularity and multi-view features in landing vibration signals, thereby improving the diagnostic accuracy of landing gear health monitoring under all operating conditions.

[0026] 2. By employing three complementary feature extraction mechanisms—multiple sets of parallel convolutions at different receptive field scales in the first branch, multi-layer self-attention encoding in the second branch, and bidirectional temporal traversal combined with temporal attention aggregation in the third branch—the model can capture local waveform details at different time scales, long-range global dependencies between all time positions, and ordered temporal evolution patterns. This effectively solves the problems in existing technologies where fixed-size convolutional kernels can only capture single-scale features and where gradient decay in long short-term memory networks leads to insufficient long-distance temporal dependency modeling capabilities. Consequently, it achieves simultaneous and effective extraction of high-frequency impact components and low-frequency trend components in landing vibration signals, as well as complete modeling of long-range dependencies.

[0027] 3. By employing a technique of deploying at least two different types of sensors at different locations on the landing gear to collect multi-channel raw signals and sequentially performing noise reduction filtering, time-frequency transformation, and cross-channel standardization processing, the sensor data from each channel, after eliminating noise interference and dimensional differences, is input into the subsequent network in a two-dimensional time-frequency representation with a unified numerical scale. This effectively solves the problem in existing technologies where only a single physical quantity signal is used as input, resulting in a single source of vibration state information and difficulty in directly fusing data from different channels. Thus, it achieves effective fusion of multi-source heterogeneous sensor data in a unified representation space. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a landing gear vibration monitoring method in an embodiment of this application.

[0029] Figure 2 This is a schematic diagram of the overall system architecture of the landing gear vibration monitoring method in the embodiments of this application;

[0030] Figure 3 This is a detailed architectural diagram of the core deep learning model in the embodiments of this application;

[0031] Figure 4 This is another flowchart illustrating the landing gear vibration monitoring method in this application embodiment;

[0032] Figure 5 This is a schematic diagram of the physical device structure of a server in an embodiment of this application. Detailed Implementation

[0033] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0034] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0035] To facilitate understanding, the application scenarios of the embodiments of this application are described below.

[0036] As a critical load-bearing structure that withstands landing impact loads, the operational status of aircraft landing gear directly affects flight safety. In actual airline operations, the landing gear withstands impact loads several times the weight of the fuselage at the moment of each landing. The fluid inside the shock absorber struts undergoes repeated compression and release, the axles and connecting pins are subjected to alternating bending moments and shear forces, and the sealing components gradually age under alternating high pressure and high temperature environments. These repeated mechanical and thermal effects inevitably lead to performance degradation of the various components of the landing gear during long-term service, ultimately resulting in failures such as shock absorber leakage, axle fatigue cracks, connecting pin wear, seal failure, or uneven tire wear.

[0037] Current landing gear maintenance primarily relies on periodic disassembly and manual visual inspection. Maintenance personnel execute predetermined maintenance plans based on flight hours or landings, disassembling the shock absorber struts to measure fluid levels and inflation pressures, visually inspecting the wheel axle surface for cracks, and measuring the clearance dimensions of connecting pins. This method has the following shortcomings: Firstly, the fixed intervals for periodic disassembly and inspection cannot capture sudden abnormal deterioration between maintenance intervals, posing a risk of missed inspections; secondly, some components are already in a healthy and adequate state before the maintenance interval is reached, and forced disassembly and inspection results in unnecessary downtime and maintenance costs. Furthermore, landing conditions vary significantly across different routes. Aircraft in the same fleet operating short-runway routes at high-altitude airports experience drastically different landing speeds, landing loads, and runway friction conditions compared to aircraft operating long-runway routes at low-altitude airports. This leads to a significant drop in diagnostic accuracy when diagnostic models based on a single operating condition are transferred to other operating conditions.

[0038] To address the aforementioned issues, this application provides a landing gear fault diagnosis and health monitoring method. In specific application scenarios, airlines deploy accelerometers, pressure sensors, strain gauges, and temperature sensors at key locations such as landing gear struts, axles, and connecting joints on each aircraft in their fleet. During each landing, multiple physical quantity signals, including vibration, pressure, strain, and temperature, are collected simultaneously. After data acquisition, the system performs preprocessing such as denoising filtering and time-frequency transformation on the raw signals. A fused feature vector is generated through a feature extraction network, and three types of diagnostic information are output: fault classification results, health index and remaining life prediction values, and anomaly scores. When an aircraft is reassigned from a dry runway route to a wet runway route, the system uses a domain adaptation mechanism to transfer the source domain operating condition knowledge of the existing labeled data to the target domain of the new operating condition, maintaining diagnostic accuracy without the need to re-collect a large amount of labeled data. In daily operations, the system continuously tracks abnormal scores for each flight cycle. When it detects a monotonically increasing trend in scores across multiple consecutive flight cycles, it issues an early warning signal of degradation, providing a decision-making window for the maintenance department to arrange preventative maintenance. When the score of a single flight cycle exceeds a threshold, it immediately issues an instant alarm signal, triggering unplanned inspection procedures. This application scenario transforms landing gear maintenance from a passive, periodic disassembly and inspection model to proactive preventative maintenance based on real-time condition monitoring, reducing the risk of missed inspections and minimizing unnecessary downtime for maintenance.

[0039] To facilitate understanding, the method provided in this implementation will be described in detail below, using the above scenario as an example. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating a landing gear vibration monitoring method in an embodiment of this application.

[0040] S101. Collect multi-channel raw signals during landing gear operation, and sequentially perform noise reduction filtering and time-frequency transformation on the multi-channel raw signals to obtain standardized time-frequency input data.

[0041] The multi-channel raw signal refers to a set of time-domain signals containing different physical quantities, simultaneously acquired by at least two different types of sensors deployed at different locations on the landing gear. Noise denoising filtering refers to the operation of applying filtering processing to the raw signal to suppress random noise components. Time-frequency transformation refers to the transformation operation that maps a one-dimensional time-domain signal to a two-dimensional time-frequency joint representation. Standardized time-frequency input data refers to the structured data formed by each channel at a unified numerical scale after the above processing, which can be directly input into the network.

[0042] Specifically, firstly, multiple channels of raw signals from the landing gear operation process are acquired using various sensors, including accelerometers, pressure sensors, temperature sensors, and strain gauges. Then, wavelet denoising is performed on each channel signal. The signal is decomposed into multiple layers of wavelet coefficients, and a threshold shrinkage is applied to each layer of coefficients, setting coefficients below the threshold to zero. The denoised signal is then reconstructed through an inverse transform. Next, a short-time Fourier transform or continuous wavelet transform is performed on the denoised signal, converting it into a two-dimensional time-frequency matrix with time and frequency as the two axes. Finally, the time-frequency matrix of each channel is normalized to zero mean and unit variance using the Z-score method, obtaining standardized time-frequency input data.

[0043] In some embodiments, denoising filtering and time-frequency transformation can be implemented in various ways: Optionally, a db4 wavelet basis is selected to perform a five-level decomposition, and the detail coefficients of each level are reconstructed by inverse transformation after soft thresholding and denoising. Then, a Hanning window is applied to the denoised signal for frame division, and short-time Fourier transform is performed frame by frame to generate a time-frequency matrix. Optionally, a Morlet wavelet is selected to perform a continuous wavelet transform within a preset scale range, directly converting the time-domain signal into a wavelet coefficient matrix as a time-frequency representation. The time-frequency energy distribution is obtained by taking the modulus and squaring the coefficient matrix. It is understood that other combinations of denoising and time-frequency transformation can also be used, which are not limited here.

[0044] In some embodiments, this step specifically includes:

[0045] Multi-channel raw signals are acquired by at least two different types of sensors deployed at different locations on the landing gear. These multi-channel raw signals contain sensor data of at least two different physical quantities. Noise filtering is performed on the multi-channel raw signals to obtain multi-channel denoised signals. Time-frequency transformation is performed on the multi-channel denoised signals to convert the one-dimensional time-domain multi-channel denoised signals into a two-dimensional time-frequency representation to obtain multi-channel time-frequency data. Cross-channel normalization processing is performed on the multi-channel time-frequency data according to each sensor channel to obtain standardized time-frequency input data.

[0046] Multi-channel raw signals refer to a collection of time-domain signals synchronously acquired by at least two different types of sensors, such as accelerometers, pressure sensors, strain gauges, and temperature sensors, deployed at different locations including landing gear struts, axles, and connecting joints. Each channel corresponds to different physical quantities such as vibration acceleration, hydraulic chamber pressure, structural strain, and ambient temperature. Noise denoising filtering refers to the processing operation of applying filtering algorithms to the raw signals to suppress random noise and power frequency interference introduced during the acquisition process. For example, wavelet thresholding shrinkage is used to decompose the signal into multiple layers of wavelet coefficients, followed by soft thresholding of the detail coefficients and inverse transformation reconstruction. Time-frequency transformation refers to the mathematical transformation that maps a one-dimensional time-domain signal to a two-dimensional matrix represented by time and frequency axes. For example, short-time Fourier transform divides the signal into frames, applies windows, and performs Fourier transform frame by frame to obtain the time-frequency matrix. Cross-channel standardization processing refers to the operation of calculating the mean and standard deviation of the time-frequency data for each sensor channel and performing zero-mean unit variance normalization, unifying the channel data with different physical dimensions and numerical ranges to the same scale.

[0047] The specific execution process of this step is as follows: First, each sensor synchronously acquires the time-domain signal from the landing gear touchdown to the end of the taxiing phase at a unified sampling frequency, forming an N-channel original signal matrix. Then, denoising filtering is performed independently on each channel. A db4 wavelet basis is selected to perform a five-level decomposition on the signal. Soft thresholding is applied to the detail coefficients of the first to fifth layers. The threshold is the median absolute value of each layer's coefficient divided by 0.6745 and then multiplied by the square root of the logarithm of each layer's length. Coefficients below the threshold are set to zero, and coefficients above the threshold are shrunk by a corresponding amount towards zero. The denoised signal is then reconstructed by inverse wavelet transform. Next, a short-time Fourier transform is performed on the denoised signal of each channel. The frame is divided into frames using a Hanning window, with a frame length of 256 sampling points and a frame shift of 128 sampling points. The 256-point fast Fourier transform is calculated frame by frame, and the power spectral density of each frame is obtained by taking the square of the modulus. All frames are arranged in chronological order to form the two-dimensional time-frequency matrix of that channel. Finally, for each channel's time-frequency matrix, the mean μ_c and standard deviation σ_c of all elements in the matrix are calculated independently. Each element in the matrix is ​​subtracted from μ_c and then divided by σ_c to complete the cross-channel standardization process. After processing all channels, the standardized time-frequency input data is obtained by stacking them along the channel dimension.

[0048] Before inputting the standardized time-frequency input data into each feature extraction branch, a dimension adaptation operation needs to be performed to match the input format requirements of each branch. The standardized time-frequency input data, after time-frequency transformation, has a shape of N×C×T×F, where N is the batch size, C is the number of sensor channels, T is the number of time frames, and F is the number of frequency components. The specific process of the dimension adaptation operation is as follows: the frequency vector corresponding to each time frame is flattened and concatenated with the channel dimension, that is, the shape N×C×T×F is reshaped into N×T×(C×F), resulting in a two-dimensional sequence representation with the time frame as the sequence length and the concatenation of the frequency components of each channel as the feature dimension. This two-dimensional sequence representation serves as the unified input format for the three feature extraction branches, where the sequence length dimension T corresponds to the time axis, and the feature dimension C×F corresponds to the frequency domain information of each channel in each time frame. The first feature extraction branch performs a one-dimensional convolution operation on the sequence along the time axis, the second feature extraction branch uses each time position of the sequence as a token input to the self-attention encoder, and the third feature extraction branch feeds the sequence stepwise along the time axis into the bidirectional long short-term memory network. After the above dimensional adaptation, the two-dimensional time-frequency representation is uniformly converted into a one-dimensional sequence form that can be directly processed by each branch, ensuring the consistency of data format in the subsequent feature extraction process.

[0049] In some embodiments, the specific deployment scheme of the sensor network is as follows: triaxial accelerometers are installed at the landing gear fuselage joints and wheel axles to collect three-directional vibration acceleration signals at a sampling frequency of not less than 1000 Hz; pressure sensors are installed at the buffer oil chambers to collect internal oil pressure signals at a sampling frequency of not less than 500 Hz; temperature sensors are installed on the outer wall of the buffer and at the wheel hubs to collect temperature field distribution signals at a sampling frequency of not less than 10 Hz; strain gauges or fiber optic sensors are installed at the side braces and torsion arms to collect structural strain response signals at a sampling frequency of not less than 500 Hz. The multi-channel data collected by each sensor undergoes preliminary time synchronization, multi-channel alignment, and local caching at edge computing nodes deployed near the landing gear area. Then, it is transmitted via the airborne data bus to the airborne processing unit for real-time inference, or wirelessly transmitted to a ground server for offline analysis.

[0050] In some embodiments, after obtaining the standardized time-frequency input data, the process further includes performing data segmentation and data augmentation on the standardized time-frequency input data. The specific operation of data segmentation is as follows: the long-term continuous standardized time-frequency input data is divided into segments of fixed window lengths, with each window corresponding to a training sample or inference sample. The specific operation of data augmentation is as follows: for class imbalance cases where the number of fault samples is far less than the number of normal samples, additional samples are generated for the fault samples using a sliding window overlapping sampling method, i.e., overlapping intervals are set between adjacent windows so that the same fault signal is sampled multiple times; simultaneously, temporal perturbations are applied to the fault samples, including superimposing small-amplitude random noise on the time-domain signal and applying random translation to the signal time axis, to increase the diversity of fault samples and alleviate the problem of model bias towards the normal class caused by the large disparity between the number of normal samples and fault samples during training.

[0051] Please see Figure 2 , Figure 2 The schematic diagram of the overall system architecture of the landing gear vibration monitoring method provided in the embodiments of this application shows the complete data processing link from multi-source sensor data acquisition to health status assessment and decision-making.

[0052] like Figure 2 As shown in the embodiments of this application, the landing gear vibration monitoring method includes five core processing layers. The top layer is a multi-source sensor data acquisition layer, which includes four types of sensor nodes: accelerometer (acquiring vibration signals), pressure sensor (acquiring buffer oil pressure), temperature sensor (acquiring temperature field), and strain sensor (acquiring structural strain). These sensors synchronously acquire multi-channel raw signals from different positions on the landing gear. The second layer is a signal preprocessing and time-frequency transformation module, which sequentially performs four processing operations on the multi-channel raw signals: denoising filtering (wavelet denoising), STFT / CWT time-frequency transformation, data standardization and segmentation, and data augmentation and balancing. This transforms the one-dimensional time-domain signal into a two-dimensional time-frequency representation with a unified scale. The third layer is a deep learning core model layer, which includes three parallel feature extraction pathways: a multi-scale CNN spatial feature extraction branch, a BiLSTM / Transformer temporal feature modeling branch, and a multi-head attention mechanism multi-sensor fusion module. These three branches collaboratively extract local multi-scale features, global dependency features, and temporal evolution features. The fourth layer is a dual-task output layer, fusing feature vectors and feeding them into the fault classification and diagnosis head (outputting classification probability distribution based on Softmax) and the health index regression head (outputting HI value and RUL predicted value). Simultaneously, standardized time-frequency input data is fed into the anomaly detection branch to output anomaly scores. The fifth layer is a health status assessment and decision-making module, which summarizes the above outputs to generate three types of decision-making information: real-time alarms and visualizations, predictive maintenance scheduling suggestions, and life assessment reports. This provides maintenance personnel with end-to-end support from fault identification and trend warnings to maintenance scheduling.

[0053] S102. The standardized time-frequency input data is input in parallel into three feature extraction branches, and the corresponding feature vectors are output respectively. The first feature extraction branch extracts local multi-scale features of the standardized time-frequency input data through multiple sets of parallel convolution operations with different receptive field scales. After channel-dimensional concatenation and channel attention weighting, the local feature vector is output. The second feature extraction branch performs multi-layer self-attention encoding operation on the standardized time-frequency input data with injected position encoding information, calculates the dependency weights between all time positions and outputs the global feature vector. The third feature extraction branch traverses the standardized time-frequency input data along the forward and reverse time directions respectively, extracts the hidden state of each time step, and outputs the temporal feature vector after temporal attention weighting aggregation.

[0054] The first feature extraction branch refers to the computational path that extracts local features in parallel using multiple sets of convolutional kernels of different sizes. The receptive field scale refers to the number of sampling points covered by the convolutional kernel along the time axis. Channel attention weighting refers to the operation of calculating the importance weight of each channel and scaling it channel by channel. The second feature extraction branch refers to the computational path that establishes global temporal dependencies through a self-attention encoder. Positional encoding refers to assigning an encoding vector representing the sequence order of each time position. The third feature extraction branch refers to the recurrent network path that traverses the sequence along the bidirectional time direction and extracts the hidden states. Temporal attention weighted aggregation refers to the operation of summing the hidden states of all time steps according to their attention weights to generate a single vector.

[0055] Specifically, standardized time-frequency input data is simultaneously fed into three branches. In the first branch, three sets of convolutional layers with kernel sizes of 3, 5, and 7 process the input in parallel. After batch normalization, activation, and pooling, each set outputs a feature map at the corresponding scale. The three outputs are concatenated along the channel dimension and then subjected to global average pooling and two fully connected layers to calculate channel weight coefficients. These coefficients are then scaled channel by channel and flattened to obtain local feature vectors. In the second branch, the input superposition position is encoded and fed into a multi-layer encoder. Each layer contains a multi-head self-attention sublayer and a feedforward sublayer, along with residual connections and layer normalization. After processing by all layers, average pooling is used to obtain the global feature vector. In the third branch, a two-layer bidirectional long short-term memory network traverses the input stepwise in both directions. The forward and reverse hidden states at each time step are concatenated and mapped to attention scores through a fully connected layer. After normalization, a weighted sum is obtained to obtain the temporal feature vector.

[0056] In some embodiments, multi-scale feature extraction and channel weighting can be achieved in various ways: Optionally, three sets of convolutional layers each contain 64 convolutional kernels, which are concatenated to 192 channels, reduced to 48 dimensions by global average pooling, then increased back to 192 dimensions, and scaled by Sigmoid output weights; Optionally, dilated convolutions with kernel sizes of 3, 7, and 11 are used to extract features at different dilation rates, which are then concatenated, compressed in dimension by channel-wise convolution, added by dual-channel pooling, and then weighted by Sigmoid output weights. It is understood that other convolutional configurations and attention methods can also be used, and are not limited here.

[0057] In some embodiments, before inputting the standardized time-frequency input data into the three feature extraction branches, the method further includes:

[0058] The standardized time-frequency input data is processed frame by frame along the time axis to calculate short-time energy values. The difference between the short-time energy values ​​of adjacent frames is used to obtain an energy change rate sequence. Abrupt points exceeding an adaptive threshold are detected in this energy change rate sequence. The adaptive threshold is the sum of the mean of the energy change rate sequence of the current signal segment and the standard deviation of a preset multiple. The standardized time-frequency input data is divided into an impact segment and a steady-state segment based on the detected abrupt change points. The impact segment corresponds to the time interval from the moment of landing to the first complete compression of the buffer, and the steady-state segment corresponds to the skidding interval after the buffer compression and rebound ends.

[0059] The impact segment data and the steady-state segment data are input into three feature extraction branches for independent feature extraction, resulting in impact segment fusion feature vector and steady-state segment fusion feature vector. The impact segment fusion feature vector and the steady-state segment fusion feature vector are concatenated and input into the inter-segment relationship encoding layer. The inter-segment relationship encoding layer calculates the mutual attention weight between the impact segment features and the steady-state segment features, and performs cross-enhancement on the two segment features with the mutual attention weight before merging and outputting the final fusion feature vector.

[0060] By adaptively segmenting the impact segment and the steady-state segment and extracting features separately, the transient abnormal features in the impact segment are not overwhelmed by the background statistical features of the steady-state segment. At the same time, the inter-segment relationship encoding layer captures the changes in the coupling relationship between the impact response features and the subsequent steady-state vibration features, enabling the model to identify the abnormal impact-steady-state transition features caused by the degradation of the buffer performance.

[0061] In some embodiments, after the step of merging all cross-enhanced feature vectors and outputting a fused feature vector, the method further includes:

[0062] Maintain a fixed-length feature memory queue, which stores the fused feature vectors corresponding to the most recent preset number of flight cycles in chronological order; after the fused feature vector of the current flight cycle is generated, append the fused feature vector to the end of the feature memory queue, and remove the earliest feature vector that exceeds the queue length.

[0063] All historical fused feature vectors stored in the feature memory queue are arranged in chronological order to form a feature trajectory sequence. The fused feature vector of the current flight loop is used as the query end, and the feature trajectory sequence is used as the key and value ends. Temporal causal attention operation is performed, where the attention mask restricts the current query to only pay attention to historical feature vectors that precede it in time, resulting in a degradation perception enhancement feature vector. This degradation perception enhancement feature vector is then fused with the fused feature vector of the current flight loop through a gating mechanism. The gating coefficient is generated by the current fused feature vector through a fully connected mapping and Sigmoid activation. The degradation perception enhancement feature vector is scaled by the gating coefficient and then added to the current fused feature vector to obtain the degradation context fused feature vector. This degradation context fused feature vector is then input into each output branch to replace the original fused feature vector.

[0064] By preserving high-dimensional feature representations of historical flight cycles through a feature memory queue and modeling the evolution trend of feature trajectories using a causal attention mechanism, the diagnostic decision for the current flight cycle can refer to the recent degradation trajectory information of the landing gear. This encodes the cross-cycle degradation trend to the feature level rather than just making post-processing judgments at the scoring level, thereby improving the detection sensitivity in the early stages of progressive degradation.

[0065] S103. Input the feature vectors output by each feature extraction branch into the cross-modal attention fusion layer. In the cross-modal attention fusion layer, the feature vectors output by each feature extraction branch are sequentially designated as the query end. Each time, the feature vector of the current non-query end is concatenated and used as the key end and value end. The value end is weighted and aggregated according to the attention weight between the query end and the key end to obtain the cross-enhanced feature vector corresponding to each feature extraction branch. After merging all cross-enhanced feature vectors, the fused feature vector is output. The fused feature vector and the standardized time-frequency input data are input into multiple output branches respectively to output the classification probability distribution, health index, remaining life prediction value and anomaly score respectively.

[0066] In this context, the cross-modal attention fusion layer refers to a network layer that interactively fuses multiple feature vectors using a cross-attention approach. The query side refers to the party providing the query vector, while the key and value sides refer to the parties providing the queried information. The cross-enhanced feature vector refers to the enhanced representation of a feature vector from one branch after it has acquired supplementary information from other branches through an attention mechanism. The fused feature vector refers to the unified representation after merging all cross-enhanced feature vectors. The anomaly score refers to the reconstruction error value after the input has been reconstructed through encoding and decoding.

[0067] Specifically, the fusion layer performs three rounds of cross-attention calculation: in each round, the feature vector of one branch is designated as the query end, and the feature vectors of the other two branches are concatenated and linearly mapped to a key matrix and a value matrix. The query end is mapped to a query matrix, and then scaled and normalized with the key matrix to obtain weights. The value matrix is ​​then weighted and summed to obtain the corresponding cross-enhanced feature vector. After the three rounds are completed, the three cross-enhanced feature vectors are concatenated to output the fused feature vector. The fused feature vector is then mapped through a fully connected layer and fed into the fault classification output head (which outputs the classification probability distribution after Softmax normalization) and the regression output head (which outputs the health index and remaining life prediction values). The standardized time-frequency input data is fed into the anomaly detection branch, compressed to a low-dimensional latent space by the encoding network, and then reconstructed by the decoding network. The mean square error between the reconstructed signal and the original input is calculated as the anomaly score.

[0068] Please see Figure 3 , Figure 3 The detailed architecture diagram of the deep learning core model provided in the embodiments of this application shows the complete network structure of three-branch parallel feature extraction, cross-modal attention fusion, and multi-task output.

[0069] like Figure 3As shown, the deep learning core model receives standardized time-frequency input data X∈R^(N×C×T) (N is the batch size, C is the number of channels, and T is the time step) and feeds it in parallel into three feature extraction branches. The left side shows the multi-scale 1D-CNN module, which contains three sets of parallel convolutional units: the first set uses a Conv1D+BN+ReLU structure with a kernel size k=3 to capture high-frequency impact components; the second set uses a Conv1D+BN+ReLU structure with a kernel size k=5 to capture mid-frequency vibration components; and the third set uses a Conv1D+BN+ReLU structure with a kernel size k=7 to capture low-frequency trend components. The three sets of outputs are concatenated along the channel dimension, downsampled by MaxPool, and then weighted by channel attention to output local feature vectors. The middle module is the Transformer encoder, which first injects positional encoding (PE) into the input to preserve temporal positional information. Then, a multi-head self-attention (MHA) mechanism calculates the dependency weight matrix between all time positions. After positional mapping via a feedforward neural network (FFN) and combined with Layer Norm normalization and Dropout regularization, it outputs a global feature vector. The right side is the BiLSTM temporal module, containing two layers of bidirectional long short-term memory networks (BiLSTM Layer 1 and BiLSTM Layer 2) that traverse the input sequence along both forward and backward time directions. The bidirectional hidden states at all time steps are concatenated and then weighted and aggregated by temporal attention to output a temporal feature vector. The feature vectors output from the three branches are fed into a cross-modal attention fusion layer. This layer sequentially designates the feature vectors from each branch as the query end, and each time concatenates the feature vectors from the other two branches as the key and value ends. It calculates the scaled dot product attention weights and performs weighted aggregation on the value, obtaining three cross-enhanced feature vectors, which are then merged to output a fused feature vector. The fused feature vectors are fed into three output branches: the left branch is the variational autoencoder (VAE) anomaly detection branch, which compresses the standardized time-frequency input data into a low-dimensional latent space via the encoder and then reconstructs it by the decoder, calculating the reconstruction error as the anomaly score; the middle branch is the classification head, which outputs the fault type diagnosis result after passing through a fully connected layer (FC) and Softmax normalization; the right branch is the regression head, which outputs the health index HI and the predicted remaining useful life (RUL) value after passing through a fully connected layer. At the bottom is the comprehensive health assessment and multi-task joint decision-making module, which summarizes the classification diagnosis results, anomaly scores, and health index prediction results to generate the final landing gear health status assessment result.

[0070] In some embodiments, fusion and output can be achieved in multiple ways: Optionally, each round of cross-attention employs a multi-head mechanism, dividing the query key-value pairs into multiple subspaces for independent computation, concatenating them, and then performing a linear mapping. The outputs from the three rounds are concatenated and then subjected to dimensionality reduction via a fully connected layer to obtain a fused feature vector, which is then mapped to the classification and regression outputs via two fully connected layers respectively. Optionally, each of the three rounds of computation employs single-head attention and applies residual connections to the results. The three cross-enhanced feature vectors are added element-wise and then normalized to output a fused feature vector. The anomaly detection branch uses a variational coding structure to sample from the latent space and then decode and reconstruct it. It is understood that other fusion and output methods can also be used, and no limitation is made here.

[0071] In some embodiments, this step specifically includes:

[0072] The fused feature vector is mapped through a fully connected layer and then fed into the fault classification output head. After probability normalization, the classification probability distribution of each preset fault category is output as the fault classification result. The fused feature vector is mapped through a fully connected layer and then fed into the health status regression output head, which outputs the health index and remaining life prediction value as the health status prediction result. The standardized time-frequency input data is input into the anomaly detection branch, compressed and mapped to a low-dimensional latent space by the encoding network, and then reconstructed by the decoding network to obtain the reconstructed signal. The reconstruction error between the reconstructed signal and the standardized time-frequency input data is calculated, and the reconstruction error is used as the anomaly score.

[0073] The fault classification output head is a network structure that receives fused feature vectors and outputs probability values ​​for each preset fault category. It consists of fully connected layers and a Softmax activation function. The classification probability distribution refers to the normalized probability vector corresponding to all K preset fault categories, with each element taking values ​​between 0 and 1 and summing to 1. The health status regression output head is a network structure that receives fused feature vectors and outputs continuous numerical health indicators. It consists of fully connected layers and a linear activation function, outputting two scalars: a health index and a predicted remaining lifespan. The anomaly detection branch is an autoencoder structure composed of an encoder network and a decoder network. The encoder network compresses the high-dimensional input into a low-dimensional latent space representation, and the decoder network reconstructs the latent space representation into a signal of the same dimension as the input. The reconstruction error is the mean of the squared element-wise differences between the reconstructed signal and the original standardized time-frequency input data.

[0074] The specific execution process of this step is as follows: The fused feature vector is first mapped to 256 dimensions through a fully connected layer and activated by ReLU, and then fed into two output heads. In the fault classification output head, the 256-dimensional vector is mapped to K dimensions through a fully connected layer (K ​​is the preset total number of fault categories, for example, K=6 corresponds to six categories: normal, buffer leakage, axle crack, connecting pin wear, seal degradation, and tire wear). Softmax operation is performed on the K-dimensional output vector, that is, the exponent of each element is taken and divided by the sum of the exponents to obtain the classification probability distribution as the fault classification result. In the health status regression output head, the 256-dimensional vector is mapped to 128 dimensions through a fully connected layer and activated by ReLU, and then mapped to 2 dimensions through a fully connected layer. The first output value is compressed to the interval of 0 to 1 by the Sigmoid function as the health index (1 represents complete health, 0 represents complete failure). The second output value is constrained to a non-negative value by the ReLU function as the remaining life prediction value (unit: flight cycles). In the anomaly detection branch, the standardized time-frequency input data is compressed layer by layer by the encoding network. It then passes through three convolutional layers, expanding the number of channels from N to 64, 128, and 256 while halving the spatial resolution layer by layer. Finally, after flattening, it is mapped to a 32-dimensional latent vector through a fully connected layer. The decoding network restores the 32-dimensional latent vector to its original dimensions through a fully connected layer, reshaping it into a 3D tensor. This tensor then passes through three transposed convolutional layers, reducing the number of channels from 256 to 128, 64, and N while doubling the spatial resolution layer by layer. The output is a reconstructed signal of the same size as the input. The reconstruction error is calculated by subtracting the reconstructed signal from the standardized time-frequency input data element by element, taking the square of the result, and averaging the sum of all elements to obtain a single scalar value, which is the anomaly score.

[0075] In some embodiments, the true labeled value of the health index is constructed by normalizing the root mean square (RMS) vibration energy: the RMS value of the vibration acceleration signal in each flight cycle is calculated, and the RMS value in the initial health state is used as a benchmark. The ratio of the current RMS value to the benchmark value is normalized and mapped to the interval between 0 and 1, which serves as the labeled value of the health index corresponding to that flight cycle. In some embodiments, the regression output head also includes an uncertainty quantification module, which outputs the prediction confidence interval along with the remaining lifetime prediction value. Specifically, the last fully connected layer of the regression output head outputs two scalar values. The first scalar value serves as the point estimate prediction value of the remaining lifetime, and the second scalar value is constrained to a positive value by the Softplus function and serves as the estimate value of the prediction variance. During the inference phase, a confidence interval is constructed with the point estimate prediction value as the center and a preset multiple of the square root of the prediction variance estimate as the radius, providing a quantitative reference for the prediction credibility of maintenance decisions.

[0076] In some embodiments, the anomaly detection branch is implemented using a variational autoencoder (VAE) structure. Unlike ordinary autoencoders, the VAE's encoding network outputs the mean and variance vectors of the latent space. The latent variables are sampled from this distribution using reparameterization techniques and then fed into the decoding network to reconstruct the signal. During the VAE's training phase, only standardized time-frequency input data under normal operating conditions are used, allowing the encoding network to learn the probability distribution of normal vibration signals in the latent space. During the inference phase, when the input signal originates from an abnormal or faulty state, the reconstruction error between the decoded network's reconstructed signal and the original input will significantly increase due to its deviation from the latent space distribution of the normal state. The anomaly score constructed accordingly reflects the degree to which the current signal deviates from the normal pattern. This anomaly detection branch operates in an unsupervised manner, independent of fault category labels, thus enabling the detection of novel anomaly patterns not present in the training data, overcoming the limitation of fault classification heads that can only identify preset fault categories.

[0077] In some embodiments, after the three feature extraction branches output feature vectors and before the input cross-modal attention fusion layer, the method further includes:

[0078] Based on the landing gear structural mechanics model, the physical coupling pairs between each sensor channel are determined. Each coupling pair contains two sensor channel indices with known physical association and a coupling polarity label describing the direction of the association, where the coupling polarity label is either positive or negative. Positive correlation means that when the physical quantity of one channel increases, the physical quantity of the other channel increases synchronously, and negative correlation means that when the physical quantity of one channel increases, the physical quantity of the other channel decreases.

[0079] For each physical coupling pair, feature sub-vectors corresponding to the two channels in the coupling pair are extracted from the output feature vectors of the three feature extraction branches, and the cosine similarity value between the two feature sub-vectors is calculated. When the coupling polarity is marked as positively correlated, the cosine similarity value is subtracted from the value to obtain the consistency deviation value of the coupling pair. When the coupling polarity is marked as negatively correlated, the cosine similarity value is added to the value to obtain the consistency deviation value of the coupling pair. The average of the consistency deviation values ​​of all physical coupling pairs is used to obtain the physical consistency regularization loss. The physical consistency regularization loss is multiplied by a preset weight coefficient and then included in the total loss function.

[0080] By encoding the known mechanical coupling relationships between sensors into directional constraints in the feature space, the network is forced to maintain a correlation structure between channels that is consistent with physical laws, suppressing false feature associations that violate physical laws caused by noise or data bias. Under the condition of limited training data, additional inductive bias is provided to improve the physical interpretability and generalization ability of the features.

[0081] S104. Generate the landing gear health status assessment results based on the classification probability distribution, health index, remaining life prediction value, and anomaly score.

[0082] Among them, the health status assessment result refers to the comprehensive assessment information generated after integrating the above multi-dimensional output indicators, which includes fault type determination, health level, remaining life estimation and abnormal alarm information.

[0083] Specifically, the category with the highest probability in the classification probability distribution is taken as the fault type determination result. The health index is compared with preset multi-level thresholds to determine the current health level. The remaining life prediction value is compared with the maintenance plan to determine whether maintenance needs to be arranged. The anomaly score is compared with the anomaly threshold, and an immediate alarm is generated when the threshold is exceeded; the anomaly scores of multiple consecutive flight cycles are constructed into a time series, and it is judged within a sliding window whether there is a continuous monotonically increasing trend. If the condition is met, a degradation warning signal is generated. The above results are summarized and output as the health status assessment result.

[0084] In some embodiments, the evaluation results can be generated in several ways: Optionally, the fault type, health index, remaining life prediction value, and anomaly level can be encapsulated into a structured message and sent to the maintenance decision system; Optionally, time series can be constructed for the health index and anomaly score respectively, a degradation slope can be fitted to the health index, and the number of increments of the anomaly score within a sliding window can be counted. When a threshold is exceeded, a degradation warning can be generated and output together with the current judgment result. It is understood that other decision logics can also be used to achieve comprehensive evaluation, which is not limited here.

[0085] In some embodiments, the health status assessment results also include predictive maintenance scheduling recommendations and health trend reports. The predictive maintenance scheduling is implemented as follows: the remaining life prediction value is compared with the current flight operation plan. When the remaining life prediction value falls below a preset maintenance margin threshold, downtime periods that meet the minimum maintenance man-hour requirements are selected from the flight intervals between the current time point and the predicted failure time point. The earliest downtime period that meets the conditions is output as the recommended maintenance window to the maintenance decision system, thereby transforming the passive, periodic disassembly and inspection maintenance mode into an on-demand maintenance mode based on the actual degradation status. The health trend report is implemented as follows: the health index values ​​corresponding to multiple consecutive flight cycles are arranged in chronological order, a degradation trend curve of the health index changing with the number of flight cycles is fitted, and the degradation trend curve, current health level, remaining life prediction value, and its confidence interval are encapsulated into a structured report output for maintenance engineers to assess the current degradation process of the landing gear and formulate long-term maintenance strategies.

[0086] The following provides a more detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 4 This is another flowchart illustrating the landing gear vibration monitoring method in this application.

[0087] S401. Define the landing condition with labeled data as the source domain and the landing condition to be adapted as the target domain. The source domain and the target domain differ in at least one of landing speed, landing load or runway conditions.

[0088] The source domain refers to the set of landing condition data that has been labeled with fault types and health indices, such as historical data of landing at a standard settlement rate on a dry concrete runway. The target domain refers to the set of landing condition data that lacks labeling information and requires the reuse of source domain knowledge for diagnosis, such as in-service operational data of landing at a higher settlement rate on a wet asphalt runway. Landing condition difference refers to the data distribution offset between the source domain and the target domain caused by different operating conditions, specifically manifested as at least one of the following: different landing speed ranges, different landing load amplitudes, or different runway surface friction coefficients.

[0089] This step designates fully labeled data acquired under one landing condition as the source domain, and data acquired under another landing condition with missing or excessively costly labels as the target domain. Each sample in the source domain carries a fault category label and a corresponding health index value, while each sample in the target domain contains only the raw sensor signal without a label. There must be a quantitative difference between the two domains in at least one dimension: landing speed, landing load, or runway condition. For example, the landing speed in the source domain might be concentrated in the 130-140 knot range, while the landing speed in the target domain might be concentrated in the 145-155 knot range. This results in a shift in the vibration amplitude distribution and peak load distribution, leading to a decrease in diagnostic accuracy when the model trained in the source domain is directly applied to the target domain.

[0090] S402. Connect the domain discrimination branch to the output end of the fused feature vector, input the fused feature vector into the domain discrimination branch, and determine whether the domain label of the fused feature vector is the source domain or the target domain.

[0091] A domain discriminant branch refers to a binary classification network structure that receives fused feature vectors and outputs domain label predictions. For example, a discriminant network consisting of three fully connected layers and a sigmoid activation function. A domain label is a binary marker indicating whether a sample comes from the source or target domain; a value of 0 represents the source domain, and a value of 1 represents the target domain.

[0092] This step concatenates a domain discrimination branch at the output of the fused feature vector. The fused feature vector is mapped to 256 dimensions from the concatenated dimensions by a first fully connected layer and activated by ReLU. It is then mapped to 128 dimensions by a second fully connected layer and activated by ReLU again. Finally, it is mapped to 1 dimension by a third fully connected layer and compressed to the 0-1 range using a Sigmoid function to output the domain label prediction probability. The domain discrimination branch is trained using a binary cross-entropy loss function, aiming to determine as accurately as possible whether the current input fused feature vector belongs to the source or target domain. When this branch is optimized independently, a higher domain discrimination accuracy indicates a greater difference in the feature distributions of the two domains, providing a basis for adversarial training for subsequent gradient inversion.

[0093] S403. Apply a direction reversal to the gradient back from the domain discrimination branch to the feature extraction branch, so that the feature extraction branch cannot distinguish the domain-invariant features from which the domain originates.

[0094] Gradient direction reversal refers to multiplying the gradient propagated from the domain discriminant branch to the feature extraction branch by negative one during backpropagation, causing the parameter update direction of the feature extraction branch to be opposite to the optimization objective of the domain discriminant branch. Domain-invariant features refer to feature representations that are statistically indistinguishable between the source and target domains after adversarial training.

[0095] This step inserts a gradient inversion layer between the fused feature vector and the domain discriminant branch. During forward propagation, the gradient inversion layer performs an identity mapping on the fused feature vector without altering the data flow. During backward propagation, the gradient inversion layer multiplies the gradient vector returned from the domain discriminant branch element-wise by a negative coefficient λ before passing it to the feature extraction branch. λ increases from 0 to 1 during training according to a scheduling strategy. This creates an adversarial game: the domain discriminant branch continuously optimizes to improve domain classification accuracy, while the feature extraction branch, receiving the inverted gradient, is driven to generate features that make the domain discriminant branch output close to a random guess probability of 0.5. After the adversarial training converges, the fused feature vector no longer carries domain-discriminative information, becoming a domain-invariant feature, thus ensuring that subsequent fault classification and health index prediction heads maintain comparable accuracy in the target domain to the source domain.

[0096] In some embodiments, a cosine annealing learning rate scheduling strategy is adopted for the domain adaptive training phase to stabilize the adversarial training process. Specifically, the learning rate is gradually decayed from its initial value to a preset minimum value according to a cosine function during training. After each preset period, the learning rate is reset to its initial value and a new round of cosine decay begins. Cosine annealing scheduling keeps the learning rate relatively large in the early stages of training to accelerate parameter exploration in the feature extraction branch and the domain discriminative branch. It gradually decreases the learning rate in the later stages of training to fine-tune the parameters and avoid oscillations and instability in adversarial training. Simultaneously, periodic resetting prevents the learning rate from decaying too early to a level where it cannot effectively update parameters. The gradient reversal coefficient λ adopts a progressively increasing strategy synchronized with the cosine annealing scheduling during training, increasing from 0 to 1 according to the scheduling function. This avoids excessively strong adversarial gradients in the early stages of training, which could prevent the feature extraction branch from learning effective discriminative features.

[0097] S404. Group the source domain samples and target domain samples according to the fault category, calculate the conditional feature distribution distance between the source domain and the target domain for each category, and include the sum of the conditional feature distribution distances of all categories as the category-aware alignment loss in the total loss function, so that the feature distribution of the same fault category tends to be consistent under different landing conditions.

[0098] Conditional feature distribution distance refers to the measure of the distribution difference between the source domain sample feature set and the target domain sample feature set under a given fault category, calculated using the maximum mean difference metric. Category-aware alignment loss is the scalar loss value obtained by summing the conditional feature distribution distances for all fault categories.

[0099] In each training batch, this step divides the source domain samples into K groups based on the fault category label, and the target domain samples are grouped using the pseudo-label of the current classification head. For the k-th class, the set of all fused feature vectors belonging to this class in the source domain and the set of all fused feature vectors belonging to this class in the target domain are taken, and the maximum mean difference between the two sets is calculated. The feature vectors in the two sets are mapped to the Hilbert space of the regeneration kernel, and the squared Euclidean distance between the mean vector of the source domain set in this space and the mean vector of the target domain set in this space is calculated. This distance is the conditional feature distribution distance of the k-th class. The conditional feature distribution distances of the K classes are added together to obtain the class-aware alignment loss, which is multiplied by the weight coefficient and included in the total loss function. This loss drives the feature extraction branch to eliminate inter-domain distribution offsets at the class-by-class granularity, avoiding the confusion of features of different classes during global alignment.

[0100] S405. Calculate the classification loss value corresponding to the fault classification result, the regression loss value corresponding to the health status prediction result, and the reconstruction loss value corresponding to the anomaly score, respectively.

[0101] The classification loss value is a measure of the difference between the probability distribution output by the classification head and the true fault category label, calculated using the cross-entropy function. The regression loss value is a measure of the deviation between the predicted health index and predicted remaining life output by the regression head and the true labeled values, calculated using the mean squared error function. The reconstruction loss value is a measure of the element-wise deviation between the reconstructed signal output by the anomaly detection branch and the original input signal, calculated using the mean squared error function.

[0102] This step calculates the loss for each of the three output heads. The classification loss is calculated by summing the element-wise logarithmic cross-entropy of the K-dimensional probability vector output by the classification head and the one-hot encoded vector of the true class, and then taking the negative value. The regression loss is calculated by squaring the difference between the predicted health index and the actual health index, and squaring the difference between the predicted remaining lifespan and the actual remaining lifespan, then averaging the two squared values ​​and summing them. The reconstruction loss is calculated by subtracting the reconstructed time-frequency matrix output by the decoder from the standardized time-frequency input data element-wise, squaring the result, and then averaging the sum of all elements to obtain the reconstruction loss. These three loss values ​​reflect classification accuracy, regression accuracy, and reconstruction capability, respectively, and serve as inputs for subsequent multi-task weight allocation.

[0103] In some embodiments, the classification loss is calculated using a weighted cross-entropy loss function. Each fault category is assigned a weight coefficient that is inversely proportional to the number of samples in that category. Fault categories with fewer samples are assigned a larger loss weight, which causes the model to impose a greater gradient penalty on misclassification of rare fault categories during training, thus alleviating the problem of the model being biased towards the majority category due to the large difference in the number of samples in each category.

[0104] S406. Configure learnable uncertainty parameters for each loss value, use the reciprocal of the square of each uncertainty parameter as the weight coefficient of the corresponding loss value, use the logarithm of each uncertainty parameter as the regularization term, sum the weighted loss values ​​and the regularization term to construct the total loss function, and update the network parameters and uncertainty parameters of the feature extraction branch and the output branch synchronously during backpropagation.

[0105] Uncertainty parameters refer to the scalar learnable parameters configured for each task's loss, denoted as σ1, σ2, and σ3. Their initial values ​​are set to 1.0 and are automatically adjusted during training via gradient descent. Weight coefficients are the multiplication factors preceding each loss value, taking the value as the reciprocal of the square of the corresponding uncertainty parameter, i.e., 1 / σ². Regularization terms are constraint terms that prevent the uncertainty parameters from increasing indefinitely, taking the value as the logarithm of the corresponding uncertainty parameter, i.e., log(σ).

[0106] The specific calculation process for constructing the total loss function in this step is as follows: Total loss L = (1 / σ1²)×L_cls + log(σ1) + (1 / σ2²)×L_reg + log(σ2) + (1 / σ3²)×L_rec + log(σ3) + α×L_adv + β×L_align, where L_cls is the classification loss, L_reg is the regression loss, L_rec is the reconstruction loss, L_adv is the domain discriminative adversarial loss, L_align is the class-aware alignment loss, and α and β are fixed hyperparameters. When the loss of a certain task is large, the corresponding σ increases under the action of the gradient, and 1 / σ² automatically decreases, thereby reducing the contribution of that task to the total gradient and avoiding the dominance of the training direction by difficult-to-optimize tasks; at the same time, the log(σ) regularization term prevents σ from growing indefinitely. During backpropagation, the gradient of the total loss L with respect to all network parameters and the three uncertain parameters is calculated simultaneously, and parameter updates are performed.

[0107] S407. Arrange the abnormal scores of multiple consecutive flight cycles in chronological order to form a score sequence. When the score sequence shows a continuous monotonically increasing trend within a sliding window consisting of a preset number of consecutive flight cycles, generate a degradation warning signal. The generation time of the degradation warning signal is earlier than the time node when the fault is fully manifested.

[0108] A scoring sequence is a one-dimensional numerical sequence formed by arranging the abnormal scoring values ​​corresponding to each flight cycle in chronological order. A sliding window is a continuously sliding interval of a fixed length on the scoring sequence, with the window length being a preset number of flight cycles, such as 10 consecutive flight cycles. A continuously monotonically increasing trend means that each scoring value within the sliding window is greater than its preceding scoring value.

[0109] This step appends the latest anomaly score to the end of the score sequence after each new flight cycle, using the score values ​​from the most recent preset number of flight cycles (e.g., 10) to form the current sliding window. The score sequence within the window is compared pairwise with adjacent values. If all adjacent pairs satisfy the condition that the later value is greater than the earlier value, the score sequence within the window is determined to show a continuously monotonically increasing trend, and a degradation warning signal is generated. This determination method is based on the fact that landing gear structural degradation manifests as a gradual accumulation of reconfiguration errors before the fault fully manifests, while occasional disturbances only cause single-point score fluctuations rather than continuous increases. The degradation warning signal is generated in the early stages of degradation, preceding the abrupt change in anomaly scores when the fault fully manifests, thus providing an early intervention window for maintenance decisions.

[0110] S408. When the abnormal score of a single flight cycle exceeds the preset abnormal threshold, an immediate alarm signal is generated.

[0111] The preset anomaly threshold refers to the judgment boundary value determined based on the statistical distribution of anomaly scores in normal operating data. For example, the threshold is the mean of the anomaly scores of normal samples plus three times the standard deviation. The immediate alarm signal is an alarm output triggered immediately after the end of a single flight cycle, indicating that the landing gear operating status of the current cycle has deviated from the normal range.

[0112] This step involves comparing the anomaly score calculated for each flight cycle with a preset anomaly threshold. The preset anomaly threshold is determined as follows: during the training phase, the reconstructed loss values ​​output from the anomaly detection branch of all normally operating samples are collected. The mean μ and standard deviation σ of this loss value set are calculated, and μ + 3σ is taken as the preset anomaly threshold. During online inference, if the anomaly score for the current flight cycle exceeds this threshold, an immediate alarm signal is generated and output to the maintenance terminal. The difference between the immediate alarm signal and the degradation warning signal is that the immediate alarm is triggered based on a single-point score exceeding the limit, has a fast response speed, and is suitable for detecting sudden anomalies; the degradation warning is triggered based on multi-point trend judgment and is suitable for detecting gradual degradation processes.

[0113] The server in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 5 This is a schematic diagram of the physical device structure of a server in an embodiment of this application.

[0114] It should be noted that, Figure 5 The server structure shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0115] like Figure 5 As shown, the server includes a CPU 501, which can perform various appropriate actions and processes according to a program stored in ROM 502 or a program loaded into RAM 503 from storage section 508, such as performing the methods described in the above embodiments. RAM 503 also stores various programs and data required for system operation. CPU 501, ROM 502, and RAM 503 are interconnected via bus 504. I / O interface 505 is also connected to bus 504.

[0116] The following components are connected to I / O interface 505: input section 506 including audio input devices, push-button switches, etc.; output section 507 including liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 508 including hard disks, etc.; and communication section 509 including network interface cards such as LAN (Local Area Network) cards, modems, etc. Communication section 509 performs communication processing via a network such as the Internet. Drive 510 is also connected to I / O interface 505 as needed. Removable media 511, such as disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on drive 510 as needed so that computer programs read from them can be installed into storage section 508 as needed.

[0117] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by CPU 501, it performs the various functions defined in the present invention.

[0118] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0119] Specifically, the server in this embodiment includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the landing gear vibration monitoring method provided in the above embodiment.

[0120] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the server described in the above embodiments; or it may exist independently and not assembled into the server. The storage medium carries one or more computer programs that, when executed by a processor of the server, cause the server to implement the landing gear vibration monitoring method provided in the above embodiments.

[0121] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0122] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

Claims

1. A method for monitoring landing gear vibration, characterized in that, Applied to a server, the method includes: Collect multi-channel raw signals during landing gear operation, and sequentially perform noise reduction filtering and time-frequency transformation on the multi-channel raw signals to obtain standardized time-frequency input data; The standardized time-frequency input data is input in parallel into three feature extraction branches, and each branch outputs a corresponding feature vector. The feature extraction branches process the standardized time-frequency input data using different feature extraction mechanisms. The feature vectors output by each feature extraction branch are input into a cross-modal attention fusion layer. In the cross-modal attention fusion layer, the feature vectors output by each feature extraction branch are sequentially designated as the query end. Each time, the feature vector of the current non-query end is concatenated and used as the key end and value end. The value end is weighted and aggregated according to the attention weight between the query end and the key end to obtain the cross-enhanced feature vector corresponding to each feature extraction branch. All cross-enhanced feature vectors are merged and a fused feature vector is output. The fused feature vector and the standardized time-frequency input data are respectively input into multiple output branches to output the classification probability distribution, health index, remaining life prediction value and anomaly score. Based on the classification probability distribution, the health index, the remaining life prediction value, and the anomaly score, a landing gear health status assessment result is generated.

2. The method according to claim 1, characterized in that, The step of inputting the standardized time-frequency input data into the three feature extraction branches in parallel specifically includes: The first feature extraction branch extracts local multi-scale features of the standardized time-frequency input data through parallel convolution operations of multiple sets of different receptive field scales, and outputs local feature vectors after channel dimension concatenation and channel attention weighting. The second feature extraction branch performs multi-layer self-attention encoding on the standardized time-frequency input data with injected position encoding information, calculates the dependency weights between all time positions and outputs a global feature vector. The third feature extraction branch traverses the standardized time-frequency input data along both the forward and reverse time directions, extracts the hidden state at each time step, and outputs a temporal feature vector after temporal attention weighted aggregation.

3. The method according to claim 1, characterized in that, The step of sequentially performing denoising filtering and time-frequency transformation on the multi-channel raw signal to obtain standardized time-frequency input data specifically includes: The multi-channel raw signal is acquired by at least two different types of sensors deployed at different locations on the landing gear, and the multi-channel raw signal contains sensing data of at least two different physical quantities. The original multi-channel signal is subjected to denoising filtering to obtain a multi-channel denoised signal. Perform time-frequency transformation on the multi-channel denoised signal to convert the one-dimensional time-domain form of the multi-channel denoised signal into a two-dimensional time-frequency representation to obtain multi-channel time-frequency data; The multi-channel time-frequency data is subjected to cross-channel normalization processing for each sensor channel to obtain the normalized time-frequency input data.

4. The method according to claim 1, characterized in that, The step of inputting the fused feature vector and the standardized time-frequency input data into multiple output branches respectively specifically includes: The fused feature vector is mapped through a fully connected layer and then fed into the fault classification output head. After probability normalization, the classification probability distribution of each preset fault category is output as the fault classification result. The fused feature vector is mapped through a fully connected layer and then fed into the health status regression output head to output the health index and remaining life prediction value as the health status prediction result. The standardized time-frequency input data is input into the anomaly detection branch, compressed and mapped to a low-dimensional latent space by the encoding network, and then reconstructed by the decoding network to obtain the reconstructed signal. The reconstruction error between the reconstructed signal and the standardized time-frequency input data is calculated, and the reconstruction error is used as the anomaly score.

5. The method according to claim 1, characterized in that, After the step of merging all cross-enhanced feature vectors and outputting the fused feature vector, the method further includes: The landing conditions with labeled data are defined as the source domain, and the landing conditions to be adapted are defined as the target domain. The source domain and the target domain differ in at least one of landing speed, landing load, or runway conditions. A domain discrimination branch is connected to the output end of the fused feature vector. The fused feature vector is input into the domain discrimination branch to determine whether the domain label of the fused feature vector is the source domain or the target domain. The gradient back from the domain discrimination branch to the feature extraction branch is reversed in direction, so that the feature extraction branch generates domain-invariant features that the domain discrimination branch cannot distinguish from the domain source. The source domain samples and target domain samples are grouped according to the fault category. The conditional feature distribution distance between the source domain and the target domain is calculated for each category. The sum of the conditional feature distribution distances of all categories is used as the category-aware alignment loss and included in the total loss function, so that the feature distribution of the same fault category tends to be consistent under different landing conditions.

6. The method according to claim 5, characterized in that, Following the steps of outputting the classification probability distribution, health index, predicted remaining life expectancy, and anomaly score, the method further includes: Calculate the classification loss value corresponding to the fault classification result, the regression loss value corresponding to the health status prediction result, and the reconstruction loss value corresponding to the anomaly score, respectively. Learnable uncertainty parameters are configured for each loss value. The reciprocal of the square of each uncertainty parameter is used as the weight coefficient of the corresponding loss value. The logarithm of each uncertainty parameter is used as the regularization term. The weighted loss values ​​and the regularization term are summed to construct the total loss function. The network parameters of the feature extraction branch and the output branch and each uncertainty parameter are updated synchronously during backpropagation. The abnormal scores from multiple consecutive flight cycles are arranged in chronological order to form a score sequence. Trend detection is performed on the score sequence to generate a degradation early warning signal.

7. The method according to claim 6, characterized in that, Prior to the step of performing trend detection on the scoring sequence to generate a degradation early warning signal, the method further includes: When the anomaly score of a single flight cycle exceeds a preset anomaly threshold, an immediate alarm signal is generated; The step of performing trend detection on the scoring sequence to generate a degradation early warning signal specifically includes: When the scoring sequence shows a continuous monotonically increasing trend within a sliding window consisting of a preset number of consecutive flight cycles, the degradation warning signal is generated. The generation time of the degradation warning signal is earlier than the time node when the fault is fully manifested.

8. A server, characterized in that, The server includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the server to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the server, the server causes the server to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the server, the server performs the method as described in any one of claims 1-7.