A method for early warning of aircraft engine bearing failure risks

Abstract

The present invention belongs to the technical field of aircraft engine risk warning technology, and in particular relates to an aircraft engine bearing fault risk warning method, which specifically comprises: obtaining real-time bearing operating data; inputting the real-time bearing operating data into a fault risk warning model to obtain a risk classification warning, wherein the fault risk warning model is trained using a training set containing multimodal bearing data; the fault risk warning model extracts multidimensional signal frequency domain spatial features from the real-time bearing operating data; extracts time-dependent features of the real-time bearing operating data; splices the multidimensional signal frequency domain spatial features and time-dependent features, and then inputs them into an R2D2 network to output a fault action probability distribution. The present invention can overcome the bottlenecks of traditional technologies in terms of real-time performance, accuracy, and adaptability, and improve the accuracy of bearing fault risk warnings.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft engine risk warning, and in particular relates to an aircraft engine bearing failure risk warning method. Background Art

[0002] As the core power unit of aircraft, the reliability and safety of aircraft engines are directly related to the success or failure of aviation. Bearings, as key components in aircraft engines, play a vital role in supporting the rotor system and ensuring its normal operation. During aircraft engine operation, bearings must withstand harsh operating conditions such as high temperatures, high pressures, high speeds, and complex alternating loads. This makes bearings one of the components with the highest failure rate in aircraft engines. Once a bearing fails, it not only leads to reduced engine performance but can also cause serious aviation accidents, resulting in significant economic losses and casualties. Therefore, accurate and timely failure risk warnings for aircraft engine bearings, early identification of potential failure hazards, and dynamic assessment of failure probability are crucial to ensuring reliable aircraft engine operation and flight safety.

[0003] Currently, aircraft engine bearing failure risk warning technologies primarily include traditional methods based on vibration analysis, oil analysis, and temperature monitoring. Vibration analysis-based technologies collect bearing operating vibration signals and analyze characteristic parameters such as frequency and amplitude to determine the fault risk status. However, due to the complex operating environment of aircraft engines, vibration signals are susceptible to multiple interference sources, making feature extraction difficult and the accuracy of risk feature identification insufficient to meet engineering requirements. Oil analysis-based technologies detect wear status by detecting abrasive particles in the lubricating oil. However, their long detection cycles and lack of real-time performance make it difficult to detect early failure risks. Temperature monitoring-based technologies rely on bearing temperature changes to determine risk, but the temperature response lags behind fault initiation, resulting in significant limitations in early risk warning scenarios. Therefore, it is urgent to develop an aircraft engine bearing failure risk warning method that overcomes the bottlenecks of traditional technologies in terms of real-time performance, accuracy, and adaptability, and improves the accuracy of bearing failure risk warnings. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes an aircraft engine bearing failure risk warning method, which can break through the bottlenecks of traditional technology in real-time, accuracy and adaptability, and improve the accuracy of bearing failure risk warning.

[0005] The present invention provides an aircraft engine bearing failure risk early warning method, comprising:

[0006] Get real-time bearing operation data;

[0007] The real-time bearing operation data is input into a fault risk warning model to obtain a risk classification warning, wherein the fault risk warning model is trained by a training set, and the training set is multimodal data of the bearing. The fault risk warning model extracts multi-dimensional signal frequency domain spatial features from the real-time bearing operation data, extracts the timing dependency features of the real-time bearing operation data, splices the multi-dimensional signal frequency domain spatial features and the timing dependency features, and inputs them into the R2D2 network to output the fault action probability distribution.

[0008] Optionally, obtaining the training set includes:

[0009] Use multiple types of sensors to collect multi-dimensional data on aircraft engine bearings and obtain multi-dimensional signals;

[0010] Preprocess the multi-dimensional signal to obtain the training set.

[0011] Optionally, the multi-dimensional signal includes: a vibration signal, a temperature signal, a rotation speed signal and a lubricating oil pressure signal.

[0012] Optionally, preprocessing the multi-dimensional signal to obtain the training set includes:

[0013] Using wavelet packet transform to perform frequency band decomposition on the multi-dimensional signal to obtain frequency band components;

[0014] De-noising the frequency band components using a threshold denoising method to obtain a de-noised signal;

[0015] The denoised signal is mapped to a standard normal distribution to obtain the training set.

[0016] Optionally, the fault risk warning model is obtained by an improved CNN-R2D2 model, wherein the improved CNN-R2D2 model includes: a CNN module, an LSTM module, a feature fusion module and an R2D2 module;

[0017] The CNN module is used to obtain the global spatial feature vector of the real-time bearing operation data;

[0018] The LSTM module is used to capture the temporal hidden state of the real-time bearing operation data;

[0019] The feature fusion module is used to splice the global spatial feature vector and the temporal hidden state to obtain the spatiotemporal coupling spatial state;

[0020] The R2D2 module is used to obtain the risk classification warning according to the space-time coupling spatial state.

[0021] Optionally, obtaining the risk classification warning according to the spatiotemporal coupling spatial state includes:

[0022] Obtaining an action probability distribution according to the spatiotemporal coupling spatial state;

[0023] Select the action with the highest probability in the action probability distribution to obtain the action with the highest probability;

[0024] Based on the highest probability action, combined with dynamic update of the experience replay pool, a preset threshold matrix of fault degree is obtained;

[0025] Based on the preset threshold matrix of the fault degree, a time window sliding verification mechanism is introduced to obtain the risk classification warning.

[0026] Optionally, the risk classification warning includes: triggering warning, upgrading warning and canceling warning and recording abnormal events.

[0027] Optionally, the fault risk warning model is trained using a training set in combination with a policy gradient network and a deep Q learning network;

[0028] Wherein, the policy gradient network is used to output a fault risk warning decision;

[0029] The deep Q-learning network is used to evaluate the long-term risk of the state and maximize the long-term reward.

[0030] Optionally, training the fault risk warning model in combination with a policy gradient network and a deep Q learning network includes:

[0031] Obtaining a current policy based on the policy gradient network;

[0032] According to the current strategy, a target action is selected, and a reward value and a next spatiotemporal coupling space state are obtained;

[0033] Based on the reward value and the next spatiotemporally coupled spatial state, the long-term reward is maximized.

[0034] Optionally, constructing the fault risk warning model further includes: optimizing the fault risk warning model using an AdamW optimization algorithm.

[0035] Compared with the prior art, the present invention has the following advantages and technical effects:

[0036] The present invention uses multi-dimensional sensor collaborative acquisition and improved wavelet packet denoising technology to effectively retain the pulse impact signal unique to bearing faults. Combined with the CNN-R2D2 fusion model for deep extraction of vibration frequency domain features and time series dependencies, it significantly improves the recognition of early weak fault signals, solves the problem that fault characteristics are easily submerged by noise under complex working conditions, and provides an accurate decision-making basis for preventive maintenance.

[0037] This paper introduces AdamW weight decay and experience replay technology to enhance the model's generalization capabilities for small sample fault categories. A hierarchical warning strategy, constructed using the R2D2 algorithm, dynamically adjusts warning thresholds based on fault severity and updates the warning strategy in real time based on engine performance degradation, meeting the high safety and low latency requirements of the aviation industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0039] Figure 1 This is a flow chart of an aircraft engine bearing failure risk warning method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0042] The present invention proposes a method for early warning of aircraft engine bearing failure risk, such as Figure 1 As shown, the specific steps include:

[0043] Get real-time bearing operation data;

[0044] The real-time bearing operation data is input into the fault risk warning model to obtain a risk classification warning. The fault risk warning model is trained using a training set, which is the multimodal data of the bearing. The fault risk warning model extracts multi-dimensional signal frequency domain spatial features from the real-time bearing operation data, extracts the time-dependent features of the real-time bearing operation data, splices the multi-dimensional signal frequency domain spatial features and the time-dependent features, and inputs them into the R2D2 network to output the fault action probability distribution.

[0045] Specifically, vibration acceleration sensors, temperature sensors, speed sensors and lubricating oil pressure sensors are arranged at the bearing mounting location to obtain bearing operation data; wavelet packet transform is performed on the multi-dimensional signal to denoise it; the data is normalized; an improved CNN-R2D2 (Convolutional Neural Network-Recurrent Experience Replay in Distributed Reinforcement Learning) fusion model is designed to extract the spatiotemporal coupling characteristics of bearing operation; the fault risk is defined as a partially observable Markov decision process, and the action space and reward function are set; the model is trained by combining policy gradient and deep Q learning; a multi-strategy optimization training process is adopted; the preprocessed data is input into the model for fault risk warning, and a time window sliding verification mechanism is introduced to achieve risk classification warning.

[0046] Furthermore, obtaining a training set includes:

[0047] Use multiple types of sensors to collect multi-dimensional data on aircraft engine bearings and obtain multi-dimensional signals;

[0048] Preprocess the multi-dimensional signal to obtain the training set.

[0049] Furthermore, the multi-dimensional signal includes: a vibration signal, a temperature signal, a rotation speed signal and a lubricating oil pressure signal.

[0050] Specifically, multi-dimensional sensor data collection: Based on the principles of mechanical dynamics, a comprehensive monitoring system is constructed using multiple types of sensors at the aircraft engine bearing installation location:

[0051] Vibration signal acquisition: High-precision vibration acceleration sensors are arranged in the vertical, horizontal and axial directions of the bearing seat to capture the vibration information of the bearing under complex working conditions in real time, covering the influence of dynamic loads in multiple directions.

[0052] Temperature and speed monitoring: The temperature sensor is installed on the surface of the bearing outer ring to obtain the bearing operating temperature in real time; the magnetoelectric speed sensor converts the rotor mechanical speed into an electrical signal through the principle of electromagnetic induction, ensuring accurate measurement of speed parameters.

[0053] Lubricating oil status monitoring: A diffused silicon pressure sensor is installed in the main lubricating oil channel to monitor the lubricating oil pressure changes in real time and comprehensively reflect the bearing lubrication status.

[0054] Through multi-sensor collaboration, the synchronous collection of multi-modal operating parameters such as bearing vibration, temperature, speed, and lubricating oil pressure is achieved, providing comprehensive data support for subsequent analysis.

[0055] Furthermore, the multi-dimensional signal is preprocessed to obtain a training set including:

[0056] Wavelet packet transform is used to decompose the multi-dimensional signal into frequency bands to obtain frequency band components;

[0057] The threshold denoising method is used to denoise the frequency band components to obtain the denoised signal;

[0058] The denoised signal is mapped to a standard normal distribution to obtain a training set.

[0059] Specifically, a multi-stage preprocessing process was designed to address the noise interference and dimensionality differences of the collected signals:

[0060] Wavelet packet denoising: Using improved wavelet packet transform technology, we perform fine frequency band decomposition on multi-dimensional signals. Combined with an adaptive threshold denoising method, this effectively filters out Gaussian noise while fully retaining the pulse-type impact signal unique to bearing faults, improving the accuracy of subsequent feature extraction.

[0061] Standardization and normalization: Standardize multi-source data with large dimensional differences, such as vibration acceleration and lubricating oil pressure, and map the data to a standard normal distribution. This eliminates the interference of dimensionality on model training, prevents gradient updates from being dominated by large numerical features, and ensures model convergence efficiency and stability.

[0062] Furthermore, the fault risk warning model is obtained through the improved CNN-R2D2 model, which includes: CNN module, LSTM module, feature fusion module and R2D2 module;

[0063] CNN module, used to obtain the global spatial feature vector of real-time bearing operation data;

[0064] LSTM module, used to capture the temporal hidden state of real-time bearing operation data;

[0065] Feature fusion module, used to splice the global spatial feature vector and the temporal hidden state to obtain the spatiotemporal coupled spatial state;

[0066] The R2D2 module is used to obtain risk classification warnings based on the spatial state of spatiotemporal coupling.

[0067] Specifically, in order to fully exploit the spatial and temporal features of bearing fault data, a fusion-improved CNN-R2D2 fusion model was designed:

[0068] CNN module: This module extracts local frequency-domain features from multi-dimensional signals through multi-layer convolution operations and integrates them layer by layer into global spatial feature vectors, addressing the insufficient feature extraction capabilities of traditional shallow networks. It also uses the LeakyReLU activation function instead of the traditional ReLU to mitigate information loss caused by negative suppression.

[0069] LSTM module: Utilizes a gating mechanism (input gate, forget gate, and output gate) to capture the long-term temporal dependencies of bearing operating parameters, compensating for the CNN's limitation of processing only spatial features and effectively modeling the temporal evolution of bearing wear.

[0070] Feature fusion: The frequency domain features output by CNN are spliced ​​with the temporal hidden state of LSTM to form a spatiotemporal coupled state space containing "vibration frequency domain distribution-parameter change trend", which is input into the R2D2 algorithm for sequence decision making.

[0071] The R2D2 framework models fault risk warning as a partially observable Markov decision process, defines a discrete action space, and designs a multi-dimensional reward function that includes rewards for correct warnings, penalties for false positives, and rewards for normal states. This guides the model to prioritize high-risk faults. An optimization method combining policy gradient and deep Q-learning maximizes long-term discounted rewards, balancing short-term warning accuracy with long-term trend prediction capabilities.

[0072] Furthermore, the method for obtaining risk classification warning based on the spatiotemporal coupling spatial state includes:

[0073] Obtain the action probability distribution based on the spatiotemporal coupling spatial state;

[0074] Select the action with the highest probability in the action probability distribution to obtain the action with the highest probability;

[0075] Based on the highest probability action, combined with the dynamic update of the experience replay pool, the preset threshold matrix of the fault degree is obtained;

[0076] Based on the preset threshold matrix of the fault degree, a time window sliding verification mechanism is introduced to obtain risk classification warnings.

[0077] Furthermore, risk classification warnings include: triggering warnings, upgrading warnings, canceling warnings and recording abnormal events.

[0078] Furthermore, the fault risk warning model is trained using the training set in combination with the policy gradient network and the deep Q learning network;

[0079] Among them, the policy gradient network is used to output fault risk warning decisions;

[0080] A deep Q-learning network is used to assess the long-term risk of a state and maximize the long-term reward.

[0081] Specifically, the CNN-R2D2 model outputs high-value actions through an advantage function. The deep Q-learning network takes long-term risks and avoids short-sighted decisions. The combination of the two enables the model to learn statistical patterns from historical data while adapting to real-time operating conditions through a recurring mechanism.

[0082] Furthermore, combining the policy gradient network and the deep Q learning network to train the fault risk warning model includes:

[0083] Get the current strategy based on the policy gradient network;

[0084] According to the current strategy, select the target action and obtain the reward value and the next spatiotemporal coupling space state;

[0085] Maximize the long-term reward based on the reward value and the next spatiotemporal coupled spatial state.

[0086] Furthermore, constructing the fault risk warning model also includes: optimizing the fault risk warning model using an AdamW optimization algorithm.

[0087] Specifically, the AdamW optimization algorithm introduces a weight decay mechanism to accelerate model convergence while suppressing overfitting. It is particularly suitable for training scenarios with limited real fault samples and improves the model's generalization ability for small sample fault categories.

[0088] The present embodiment is described in detail below:

[0089] This embodiment provides an aircraft engine bearing failure risk warning method, referring to Figure 1 .

[0090] Step 100: Collect multimodal data of vibration, temperature, speed, and lubricating oil pressure, retain fault impact characteristics through wavelet packet denoising, and eliminate dimensional effects through normalization;

[0091] Step 1001: Based on the principles of mechanical dynamics, high-precision vibration acceleration sensors are placed in the vertical, horizontal, and axial directions of the bearing seat at the mounting location of the aircraft engine bearing. These sensors can accurately capture bearing vibration information under complex operating conditions. A temperature sensor is installed on the outer ring of the bearing to ensure real-time acquisition of the bearing surface temperature. The speed sensor is a magnetoelectric speed sensor, installed near the gear plate connected to the engine rotor. Using the principle of electromagnetic induction, it converts the mechanical speed of the rotor into an electrical signal. The output signal frequency is proportional to the speed. The lubricating oil pressure sensor uses a diffused silicon pressure sensor, installed in the main lubricating oil channel, to monitor lubricating oil pressure changes in real time.

[0092] Step 1002: Use wavelet packet transform to perform denoising on the multi-dimensional signal. Wavelet packet transform is a further extension of wavelet transform, which can perform more precise frequency band decomposition on the signal. Suppose the original multi-dimensional signal is ,go through After layer wavelet packet decomposition, we get frequency band components. For each frequency band component , using the threshold denoising method. The threshold λ is selected using the improved Sure-Shrink threshold criterion:

[0093]

[0094] Where N is the signal length, represents the median operation, Indicates that the frequency band components Normalized to median multiples of , thereby eliminating the amplitude scale differences between different frequency bands.

[0095] In step 1003, soft threshold processing is used to remove Gaussian noise while retaining the pulse-type impact signal unique to the bearing fault, making subsequent feature extraction more accurate and improving the model's ability to identify early faults.

[0096] ;

[0097] in, is the original wavelet packet coefficient, is the wavelet packet coefficient after denoising, is a symbolic function. It should be noted that and Belong to the same dimension.

[0098] Step 1004: Vibration acceleration, lubricating oil pressure, and other quantities have significant dimensional differences. Directly inputting these into the model will cause gradient updates to be dominated by large numerical features, resulting in slow model convergence or trapping in local optima. Normalization is used to map the data to a standard normal distribution to eliminate dimensionality effects.

[0099] in, is the mean, is the standard deviation, The first data points, After normalization .

[0100] Step 200: Design a CNN-R2D2 fusion model. CNN is used to extract the frequency-domain spatial features of multi-dimensional signals, while LSTM is used to capture the temporal dependencies of parameters. These spatiotemporal features are then concatenated and used as the input state for the R2D2 algorithm. The AdamW optimization algorithm is used to suppress overfitting and improve model generalization.

[0101] Step 2001: Traditional CNN shallow layers lack sufficient feature extraction capabilities, and deep layers are prone to information loss due to negative suppression from ReLUs. Based on the spatial characteristics (multi-dimensional signal frequency domain distribution) and temporal characteristics (parameter variation trends) of bearing fault data, an improved CNN-R2D2 fusion model was designed. By extracting features from local to global layers layer by layer, this model achieves deep extraction and fusion of multimodal features.

[0102] ;

[0103] in, Represents the neural network The output vector of the layer, f is the LeakyReLU activation function, is the number of neurons in layer l, Represents the neural network The output vector of the layer, Indicates the Tier The neuron and The connection weights between neurons in the layer, For the Tier The bias value corresponding to each neuron.

[0104] Step 2002: The wear degree of aircraft engine bearings is time-dependent. The LSTM network uses a gating mechanism to memorize key states, which can effectively capture the long-term dependence of bearing operating parameters and make up for the deficiency of CNN in processing only spatial features. The LSTM network is used as the loop mechanism of the R2D2 algorithm to capture temporal dependencies and output hidden states. Its input gate , Forget Gate , output gate Update with cell status as follows:

[0105]

[0106]

[0107]

[0108] ;

[0109] in, and are the hidden state and cell state at the previous moment, and Represent the weight matrices of the input gate, forget gate, and output gate respectively, and Represent the bias of the input gate, forget gate and output gate respectively, is the Sigmoid function, represents element-wise multiplication, for Input at the moment, is the candidate cell state.

[0110] Step 2003: Define the fault risk warning as a partially observable Markov decision process, and concatenate the multi-dimensional signal frequency domain features (spatial features) extracted by CNN with the temporal hidden state output by LSTM as the input state of the R2D2 algorithm. , fully characterizes the spatiotemporal coupling characteristics of bearing operation. The action space is set as: discrete action , where K is the number of fault categories, =0 means no fault occurs, =k>0 indicates the kth type of fault (such as inner ring cracks, rolling element wear, etc.), which adapts to the discrete warning requirements of fault categories. The reward function is designed as:

[0111] Correct warning reward: If =k and the real fault is k, reward + ,The model is motivated to capture fault characteristics.

[0112] If a misjudgment occurs, punishment , inhibit wrong decisions.

[0113] When there is no fault =0 and the bearing is in normal condition, reward + ,in, For the correct warning reward value, is the misjudgment penalty value, for =0 and the positive reward value of the bearing is normal.

[0114] Step 2004: The agent selects an action based on the current policy π , get rewards and the next state Maximizing long-term rewards by combining policy gradients with deep Q-learning ,in is the discount factor, The policy network directly outputs fault warning decisions and outputs high-value actions through the advantage function. The value network assesses the long-term risk of the state and avoids short-sighted decisions. The combination of these two allows the model to learn statistical laws from historical data while adapting to real-time operating conditions through a recurring mechanism.

[0115] The policy network uses gradient ascent to optimize the policy loss function by outputting the action probability distribution , which is calculated as follows:

[0116] ;

[0117] Among them, At is the advantage function, is the expected value.

[0118] Value Network Estimate the long-term return of the state-hidden state pair and optimize the loss function L through the mean squared error V , which is calculated as follows:

[0119] ;

[0120] in, The target value that guides the value function to converge, reducing training fluctuations by separating the target network parameters.

[0121] Step 300: After real-time data is input into the model, the temporal and spatial feature fusion processing is used to output the fault action probability distribution, and the fault risk level warning is realized according to the preset threshold matrix through the time window sliding verification mechanism.

[0122] Step 3001: To address the sample imbalance and overfitting risks of bearing fault data, a multi-strategy optimization training process is employed. The AdamW optimization algorithm is employed, which introduces weight decay to accelerate convergence and prevent model overfitting. This algorithm is particularly well-suited for training with limited real-world fault samples in bearing fault data.

[0123] ;

[0124] ;

[0125] in, and denote the first-order moment estimate and the second-order moment estimate, respectively. express Time gradient, and is the attenuation rate.

[0126] Step 3002: The pre-processed real-time data is input into the trained model, and after CNN feature extraction, it is input into the R2D2 algorithm. R2D2 calculates the current hidden state. Output action probability distribution , select the action with the highest probability. Dynamically update through the experience replay pool, continuously optimize the strategy, and adapt to the parameter drift caused by engine performance degradation. Preset the threshold matrix according to the severity of the fault ,in is the normal state judgment threshold, The severity of the fault decreases.

[0127] In order to avoid misjudgment caused by noise interference, a time window sliding verification mechanism is introduced. When the action probability of a certain fault category k exceeds the corresponding threshold for M consecutive time steps When the probability is higher than the threshold value for M' time steps, the warning level k is officially triggered; if the probability continues to be higher than the threshold value in the subsequent M' time steps, it is upgraded to an emergency warning; if the probability falls below the threshold value, the warning is cancelled and the abnormal event is recorded. Through historical data calibration, a graded warning of aircraft engine bearing failure risk is achieved.

[0128] Through the above technical solution, the present invention realizes the optimization of the entire process from data preprocessing to bearing failure risk warning, effectively solving the problem of insufficient accuracy of aircraft engine bearing failure risk warning in traditional methods under complex working conditions, and has significant engineering application value.

[0129] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for early warning of aircraft engine bearing failure risk, characterized in that: include: Get real-time bearing operation data; Input the real-time bearing operation data into a fault risk warning model to obtain a risk classification warning, wherein the fault risk warning model is trained using a training set, and the training set is multimodal data of the bearing. The fault risk warning model extracts multi-dimensional signal frequency domain spatial features from the real-time bearing operation data, extracts time-dependent features of the real-time bearing operation data, splices the multi-dimensional signal frequency domain spatial features and the time-dependent features, and inputs them into an R2D2 network to output a fault action probability distribution; The fault risk warning model is obtained by an improved CNN-R2D2 model, and the improved CNN-R2D2 model includes: a CNN module, an LSTM module, a feature fusion module and an R2D2 module; The CNN module is used to obtain the global spatial feature vector of the real-time bearing operation data; The LSTM module is used to capture the temporal hidden state of the real-time bearing operation data; The feature fusion module is used to splice the global spatial feature vector and the temporal hidden state to obtain the spatiotemporal coupling spatial state; The R2D2 module is used to obtain the risk classification warning according to the space-time coupling spatial state.

2. The aircraft engine bearing failure risk warning method according to claim 1, characterized in that: Obtaining the training set includes: Use multiple types of sensors to collect multi-dimensional data on aircraft engine bearings and obtain multi-dimensional signals; Preprocess the multi-dimensional signal to obtain the training set.

3. The aircraft engine bearing failure risk warning method according to claim 2, characterized in that: The multi-dimensional signal includes: a vibration signal, a temperature signal, a rotation speed signal and a lubricating oil pressure signal.

4. The aircraft engine bearing failure risk warning method according to claim 3, characterized in that: Preprocessing the multidimensional signal to obtain the training set includes: Using wavelet packet transform to perform frequency band decomposition on the multi-dimensional signal to obtain frequency band components; De-noising the frequency band components using a threshold denoising method to obtain a de-noised signal; The denoised signal is mapped to a standard normal distribution to obtain the training set.

5. The aircraft engine bearing failure risk warning method according to claim 1, characterized in that: Obtaining the risk classification warning according to the spatiotemporal coupling spatial state includes: Obtaining an action probability distribution according to the spatiotemporal coupling spatial state; Select the action with the highest probability in the action probability distribution to obtain the action with the highest probability; Based on the highest probability action, combined with dynamic update of the experience replay pool, a preset threshold matrix of fault degree is obtained; Based on the preset threshold matrix of the fault degree, a time window sliding verification mechanism is introduced to obtain the risk classification warning.

6. The aircraft engine bearing failure risk warning method according to claim 5, characterized in that: The risk classification warning includes: triggering warning, upgrading warning and canceling warning and recording abnormal events.

7. The aircraft engine bearing failure risk warning method according to claim 5, characterized in that: The fault risk warning model is trained using a training set in combination with a policy gradient network and a deep Q learning network; Wherein, the policy gradient network is used to output a fault risk warning decision; The deep Q-learning network is used to evaluate the long-term risk of the state and maximize the long-term reward.

8. The aircraft engine bearing failure risk warning method according to claim 7, characterized in that: Training the fault risk warning model by combining a policy gradient network and a deep Q learning network includes: Obtaining a current policy based on the policy gradient network; According to the current strategy, a target action is selected, and a reward value and a next spatiotemporal coupling space state are obtained; Based on the reward value and the next spatiotemporally coupled spatial state, the long-term reward is maximized.

9. The aircraft engine bearing failure risk warning method according to claim 7, characterized in that: Constructing the fault risk warning model further includes: optimizing the fault risk warning model using an AdamW optimization algorithm.

Images