Aero-engine operation risk monitoring and early warning system based on degradation level soft representation

Abstract

The present invention discloses an aircraft engine operation risk monitoring and early warning system based on soft characterization of degradation levels. Aiming at the problem that most existing deep learning methods perform strong fitting on the remaining service life of aircraft engines but are unable to accurately determine the time node of fault risk occurrence, the present invention first pre-processes the original data, and then combines data analysis methods such as slow feature analysis to divide the aircraft engine degradation stage into healthy zone, transition zone and risk zone. Furthermore, the data of the healthy zone and risk zone are used to establish a conditional discriminant autoencoder model, and the data of the transition zone is used to establish a conditional discriminant autoencoder model combined with an attention mechanism. Finally, based on the two established conditional discriminant autoencoder models, a monitoring strategy is designed to perform online monitoring of the aircraft engine status. The present invention can realize the zoning identification and risk warning of the degradation stage of aircraft engines and provide effective suggestions for aircraft engine health management.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft engine condition monitoring, and in particular to an aircraft engine operation risk monitoring and early warning system based on degradation level soft representation. Background Art

[0002] Aircraft engines typically operate under extremely demanding conditions, such as high temperatures, high pressures, extreme cold, and high speeds, making them prone to failure. A failure in an aircraft engine can directly impact the flight of the aircraft and even pose a significant threat to its safety. However, during the operational phase of an aircraft engine, enhanced engine condition monitoring can help detect signs of failure and provide early warning of potential risks, thereby ensuring flight safety.

[0003] Currently, data-driven remaining useful life prediction methods, such as neural networks, extreme learning machines, and support vector machines, are widely used in the field of aircraft engine fault prediction and health management. These methods mainly perform strong fitting on the specific remaining operating time values of aircraft engines. The key to aircraft engine health monitoring is to determine the time point when the engine enters the risk zone prone to failure. Therefore, the life prediction results of such models cannot provide good maintenance recommendations and still require analysis and judgment by operation and maintenance personnel. Summary of the Invention

[0004] To overcome the shortcomings of the existing technology, the present invention provides an aircraft engine operation risk monitoring and early warning system based on soft characterization of degradation levels. The early warning system is constructed by first dividing the aircraft engine operation life cycle into healthy zone, transition zone and risk zone by combining data analysis methods such as slow feature analysis. Then, a conditional discriminative autoencoder (CDAE) network is designed to establish monitoring models for the healthy zone and risk zone. Furthermore, a conditional discriminative autoencoder model based on the attention mechanism is established for the transition zone. Finally, an aircraft engine online monitoring strategy is designed to identify the current degradation stage partition of the engine. This method can reveal the changes in process characteristics between each degradation partition and directly provide aircraft engine health warnings, facilitating timely decision-making by operation and maintenance personnel, reducing the probability of accidents and reducing operation and maintenance costs.

[0005] The purpose of the present invention is achieved through the following technical solutions, including the following contents:

[0006] A data acquisition module is used for online monitoring to obtain the multi-dimensional data of the aircraft engine at the current moment as a test sample;

[0007] The first monitoring module is configured to obtain multiple monitoring indicators based on the aircraft engine monitoring models for healthy and risky zones, and determine whether the monitoring indicators exceed the limits. If all monitoring indicators are within the limits, the aircraft engine is determined to be in the healthy zone. Otherwise, further determination is made based on the second monitoring module.

[0008] The second monitoring module is used to obtain multiple monitoring indicators based on the aircraft engine monitoring model in the transition zone and determine whether the monitoring indicators are out of limit. If any monitoring indicator is out of limit, it is determined that the aircraft engine has entered the risk zone; otherwise, the aircraft engine is in the transition zone.

[0009] The aircraft engine monitoring models of the healthy zone and the risk zone, and the aircraft engine monitoring model of the transition zone are constructed by the following steps:

[0010] Step 1: Acquire multidimensional degradation data of the aircraft engine, wherein the multidimensional degradation data consists of multidimensional operation data of the aircraft engine at several moments;

[0011] Step 2: Based on the degradation rate of the multi-dimensional degradation data of the aircraft engine over time, thresholds are set based on empirical knowledge and the overall condition of the engine to stratify the aircraft engine degradation level. The degradation level is divided into three levels: healthy zone, transition zone, and risk zone, and codes are set for the healthy and risk zones.

[0012] Step 3: Construct a first autoencoder model and a first dataset. The first autoencoder model consists of a first encoder and a first decoder. Each sample in the first dataset is a concatenation of multidimensional data at a time within two graded areas, the healthy zone and the risk zone, and the corresponding graded area code. Each sample in the first dataset is used as input to train the first autoencoder model. The trained first autoencoder model serves as the aircraft engine monitoring model for the healthy and risk zones.

[0013] Step 4: Construct a second autoencoder model and a second data set. The second data set includes three samples in one-to-one correspondence, wherein the first sample is the multidimensional data of a moment in the transition zone; the second sample is composed of the multidimensional data of a moment in the transition zone and the coding splicing of the healthy zone; the third sample is composed of the multidimensional data of a moment in the transition zone and the coding splicing of the risk zone; the second autoencoder model includes a second encoder, an attention module, a weighted feature generation module, and a second decoder; wherein:

[0014] The second encoder is used to extract the features of the multidimensional data at each moment in the transition region to obtain the first features;

[0015] The attention module is used to obtain the weight coefficient of the first feature relative to the healthy area and the risk area based on the similarity between the second feature, the third feature and the first feature; the second feature and the third feature are obtained by using the trained first encoder to perform feature extraction on the second sample and the third sample respectively;

[0016] The weighted feature generation module is used to convert the weight coefficient of the first feature relative to the healthy area and the risk area into a probability, and then use the obtained probability as the weight of the second feature and the third feature respectively to obtain the weighted feature by weighting the second feature and the third feature;

[0017] The second decoder is used to decode and reconstruct the multidimensional data at each moment in the transition region of the input according to the weighted feature;

[0018] The three samples of the second data set are used as the input of the second autoencoder model, the multidimensional data at each moment in the transition zone is used as the training of the second autoencoder model, and the trained second autoencoder model is used as the aviation engine monitoring model of the transition zone.

[0019] The monitoring indicators include: the binary norm of the model reconstruction residual and the binary norm of the model extraction feature.

[0020] In step 2, the degradation rate of the multidimensional degradation data of the aircraft engine changing over time is obtained by the following method: using a slow feature analysis algorithm to extract the slowest changing feature in the collected multidimensional degradation data of the aircraft engine, then using a sliding window to process the slow feature change curve, calculating the absolute value of the slope of the least squares fitting line in the sliding window, and using the absolute value of the slope as the degradation rate to obtain a degradation rate change curve.

[0021] In step 2, one-hot encoding is used to set the encoding of the healthy area and the risk area, and the encoding dimension is 2D or 3D.

[0022] In step 3, the loss function used to train the first autoencoder model is:

[0023]

[0024]

[0025]

[0026] Where L is the total loss function, h i For the first encoder, the i-th sample data x i The features extracted from i is the feature h i Corresponding level area, D steady (·) represents the first decoder function, L Cis the center loss function, represents y i The feature center in the corresponding level area, L C The gradient update formula is The gradient update formula is δ represents the condition, y i =j, and 0 otherwise.

[0027] In step 4, based on the similarity between the second feature, the third feature and the first feature, the weight coefficient of the first feature relative to the healthy area and the risk area is obtained as follows:

[0028]

[0029] where w k,i Represents the first feature of the data at the i-th moment and the second feature h 1,i and the third characteristic h 2,i The weight coefficient is c=2.

[0030] In step 4, the weight coefficient of the first feature relative to the healthy area and the risk area is converted into a probability, and then the obtained probabilities are used as the weights of the second feature and the third feature respectively. The weighted sum of the second feature and the third feature is obtained as follows:

[0031] Use the SoftMax function to convert the weight coefficients into probabilities:

[0032]

[0033] Then, the obtained probability is used as the weight of the second feature and the third feature respectively, and the weighted sum of the second feature and the third feature is obtained to obtain the weighted feature h trans,i :

[0034]

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. Applying soft partitioning to aircraft engine condition monitoring, combined with data analysis techniques such as slow feature analysis, divides the aircraft engine's operating life cycle into healthy zones, transition zones, and risk zones, revealing changes in the process characteristics of each degradation zone;

[0037] 2. A conditional discriminative autoencoder (CDAE) network was designed to model healthy and risky areas. Furthermore, an attention-based CDAE model was designed to capture the relationship between transition areas and healthy and risky areas.

[0038] 3. Online monitoring of aircraft engine status based on a zoning strategy can quickly and accurately identify the current aircraft engine degradation stage zoning and directly output risk warnings, allowing operation and maintenance personnel to make timely decisions, prepare spare parts, and formulate maintenance plans, thereby avoiding major safety accidents and reducing operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the aviation engine operation risk monitoring and early warning system based on degradation level soft characterization of the present invention.

[0040] Figure 2 This is the slow feature change curve of the engine No. 11 data in the FD001 sub-dataset in the example.

[0041] Figure 3 This is the degradation rate change curve of engine No. 11 in the FD001 sub-dataset in the example.

[0042] Figure 4 This is the structural diagram of the conditional discriminant autoencoder model established in the example.

[0043] Figure 5 This is the structural diagram of the conditional discriminative autoencoder model based on the attention mechanism established in the example.

[0044] Figure 6 This is the result of the online monitoring of the aircraft engine in the example, where Figure 6 (a) is the T2 statistic monitoring chart, Figure 6 (b) is the SPE statistic monitoring chart. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0046] The present invention provides an aircraft engine operation risk monitoring and early warning system based on degradation level soft characterization, comprising:

[0047] The data acquisition module is used for online monitoring and obtaining multi-dimensional data of the aircraft engine at the current moment as a test sample; each dimension of the multi-dimensional data is a parameter that can reflect the degradation performance of the aircraft engine;

[0048] The first monitoring module is configured to obtain one or more monitoring indicators based on the aircraft engine monitoring models for healthy and risky zones, and determine whether the monitoring indicators exceed limits. If all monitoring indicators are within limits, the aircraft engine is determined to be in the healthy zone. Otherwise, further determination is made based on the second monitoring module.

[0049] The second monitoring module is used to obtain one or more monitoring indicators based on the aircraft engine monitoring model in the transition zone, and determine whether the monitoring indicators are out of limit. If any monitoring indicator is out of limit, it is determined that the aircraft engine has entered the risk zone; otherwise, the aircraft engine is in the transition zone.

[0050] Among them, the construction method flow of the aircraft engine monitoring model in the healthy zone and risk zone, and the aircraft engine monitoring model in the transition zone are as follows: Figure 1 As shown, specifically including;

[0051] Step 1: Acquire multidimensional degradation data of the aircraft engine; wherein the multidimensional degradation data is composed of multidimensional operation data of the aircraft engine at several moments;

[0052] Specifically, the aircraft engine operating data is first collected and cleaned. Through methods such as data visualization, variables that are constant and contain no valid information are eliminated. Several parameters that can reflect the degradation performance of the aircraft engine are selected to obtain multi-dimensional degradation data of the aircraft engine.

[0053] Taking NASA's open-source CMAPSS (Commercial Modular Aerospace Propulsion System Simulation) dataset as an example, we used data visualization to remove some stable trend measurements (measurements from sensors 1, 5, 6, 10, 16, 18, and 19) from a single operating condition sub-dataset FD001. These sensor data are stable and constant, meaning they contain less engine degradation information. Therefore, the 14 sensor measurements after screening constitute the multidimensional degradation data of the aircraft engine.

[0054] Furthermore, to facilitate subsequent data processing, the multidimensional degradation data of aircraft engines are normalized data. That is, the original data consisting of the 14 filtered sensor measurement data are subjected to a maximum-minimum normalization operation to the range [0, 1] as the final data set.

[0055] Step 2: Based on the degradation rate of the multi-dimensional degradation data of the aircraft engine over time, thresholds are set based on empirical knowledge and the overall condition of the engine to stratify the aircraft engine degradation level. The degradation level is divided into three levels: healthy zone, transition zone, and risk zone, and codes are set for the healthy and risk zones.

[0056] In theory, the degradation rate of an aircraft engine increases with the number of operations. Therefore, the multi-dimensional degradation data of an aircraft engine can be divided into three levels based on the degradation rate of the aircraft engine multi-dimensional degradation data over time. The degradation rate can generally be divided into three levels: healthy zone, transition zone, and risk zone. The degradation rate of the aircraft engine can be characterized by a parameter index or the average of multiple parameter indexes. As a preferred solution, in this embodiment, the rate of change of the slowest changing feature in the data is used as the degradation rate. Specifically:

[0057] Step 2.1: First, use the slow feature analysis algorithm (SFA) to process the normalized collected data and extract the change curve of the slowest changing feature;

[0058] like Figure 2 The figure shows the slow feature change curve extracted from the No. 11 engine data in the FD001 sub-dataset.

[0059] Step 2.2: Use a sliding window to slide on the slow feature change curve at a specified step size. Calculate the absolute value of the slope of the least squares fitting line within each sliding window as the degradation rate. Obtain a degradation rate change curve. The vertical axis of the curve is the degradation rate, and the horizontal axis is the number of sliding window samples.

[0060] like Figure 3 Shown in Figure 2 Based on the engine degradation rate change curve obtained in step 2.2, according to Figure 3 It can be observed that the degradation rate of aircraft engines changes from slow to fast and can be roughly divided into three degradation stages;

[0061] Step 2.3: According to the degradation rate change curve, select two appropriate degradation rate thresholds T1 and T2 based on experience. The area with a degradation rate less than T1 in the degradation rate change curve is divided into a healthy area, the area with a degradation rate between T1 and T2 is divided into a transition area, and the area with a degradation rate greater than T2 is divided into a risk area. Figure 3 For example, according to the method in step 2.3, select appropriate thresholds T1 and T2 of 0.025 and 0.04, dividing the aircraft engine degradation stage into three zones: healthy zone, transition zone, and risk zone. Also, set encoding for the healthy and risk zones. This encoding primarily provides conditional pattern information for the autoencoder model. One-hot encoding can be used, and the encoding dimension can be 2D or 3D.

[0062] Step 3: For the two divided levels of areas, healthy zone and risk zone, the data of these two zones are used to construct a first dataset and a first autoencoder model as the aircraft engine monitoring models for the healthy zone and risk zone;

[0063] Step 3.1: The first dataset and the first autoencoder model constructed in this embodiment are conditional discriminative autoencoder models (CDAEs), which are composed of a first encoder and a first decoder. The information of the conditional pattern is the encoding of the healthy area and the risk area in step 2. Taking one-hot encoding as an example, the one hot vector v of the k-th level partition after encoding is k Expressed as:

[0064]

[0065] Where c is the number of graded zones excluding the transition zone, i.e., c = 2; k = 1 or 2, representing the healthy zone or risk zone, respectively.

[0066] Step 3.2: Each sample of the first dataset is a concatenation of multidimensional data of a moment in the two graded areas of the healthy area and the risk area and the corresponding graded area code. Each sample of the first dataset is used as the input of the first autoencoder model. The features extracted by the first encoder are H = {h1, h2, ..., h N}(N×A) is expressed as:

[0067] H=E steady ([X,V]) (2)

[0068] Among them E steady (·) is the first encoder function in CDAE, X(N×M) is the hierarchical partitioned data, including data of healthy areas and risk areas, [X,V](N×(M+c)) represents the data after the hierarchical partitioned data and the one-hot vector are concatenated, N is the number of samples, M is the dimension of the hierarchical partitioned data; A is the dimension of the features extracted by the encoder.

[0069] Step 3.3: Use the first decoder to reconstruct the data and train the first autoencoder by minimizing the loss between the reconstructed data and the input sample. The loss can be a combination of a center loss function and a reconstruction loss function. In this embodiment, the loss function used is expressed as:

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] Where L C is the center loss function, L is the total loss function, h i For the first encoder, the i-th sample data x i The features extracted from i is the feature h i The corresponding level area, represents y i The feature center in the corresponding level area cannot be directly obtained So put it into the network to generate it yourself, and update it in each batch That is, random initialization Then calculate the current data and The distance of this gradient form is then added to Similar to parameter modification, it is also similar to gradient descent method. Feature center The update formula is shown in formula (5), where δ represents the condition, y i =j, it is 1, otherwise it is 0. The formula for updating the gradient of the center loss function is shown in formula (4), D steady (·) represents the first decoder function;

[0076] Based on the obtained healthy area and risk area, the aircraft engine monitoring model is used to reconstruct the residual E steady The second norm and encoder extract features h i The second norm of is used as a monitoring indicator for online monitoring, where E steady The formula for reconstructing the residual is as follows:

[0077] E steady =XX′ (8)

[0078] like Figure 4 The figure shows the structure of the conditional discriminative autoencoder model established in step 3. The network structure of CDAE is a fully connected layer structure FC(16)-FC(128)-FC(64)-FC(16)-FC(4)-FC(16)-FC(64)-FC(128)-FC(14), the training rounds are 800, the optimizer used is RMSprop, the learning rate is 0.001, and the number of samples in each training batch is 128.

[0079] Step 4: For the divided transition area, construct a second autoencoder model and a second data set, where the second data set includes three samples in one-to-one correspondence, where the first sample is a multidimensional data sample at a moment in the transition area, and the sample is x trans,i (1×M); The second sample is composed of the multidimensional data of a moment in the transition zone and the code of the healthy zone, expressed as [x trans,i , v1]; the third sample is composed of the multidimensional data of a moment in the transition zone and the coding of the risk zone, expressed as [x trans,i , v2]; the second autoencoder model is a conditional discriminant autoencoder model combined with an attention mechanism, including a second encoder, an attention module, a weighted feature generation module, and a second decoder; wherein:

[0080] The second encoder is used to extract the features of the multidimensional data at each moment in the transition region to obtain the first feature; specifically, the first sample x in the transition region is trans,i Input the second encoder E trans (·), we get the first feature

[0081] The attention module is used to obtain the weight coefficient of the first feature relative to the healthy zone and the risk zone based on the similarity between the second feature, the third feature and the first feature. The second feature and the third feature are obtained by the following method: the multidimensional data at each moment in the transition zone are concatenated with the encoding of the healthy zone and the risk zone to form the second sample and the third sample, respectively. The second feature and the third feature are obtained by using the trained first encoder to extract the features of the second sample and the third sample respectively. Specifically:

[0082] First, the trained first encoder is used to extract the features of the second sample and the third sample to obtain the second feature h 1,i and the third characteristic h 2,i :

[0083] h k，i =E steady ([x trans，i , v k ]) (9)

[0084] k = 1, 2;

[0085] Then calculate the attention weight coefficient of each feature as follows:

[0086]

[0087]

[0088] where w k,i Represents the first feature and the second and third characteristics h k,i The similarity represents the weight coefficient of the first feature corresponding to the healthy area and the risk area to establish the model;

[0089] The weighted feature generation module is used to convert the weight coefficient of the first feature relative to the healthy area and the risk area into a probability, and then use the obtained probability as the weight of the second feature and the third feature respectively to obtain the weighted feature by weighting the second feature and the third feature;

[0090] In this embodiment, the SoftMax function is first used to convert the weight coefficient into a probability:

[0091]

[0092] Then the weighted feature h trans,i Represented as h k,i The weighted sum of is as follows:

[0093]

[0094] The second decoder is used to decode and reconstruct the multidimensional data at each moment in the transition region of the input according to the weighted feature. The formula is as follows:

[0095]

[0096] Among them D trans (·) is the second decoder function;

[0097] The three samples of the second data set are used as the input of the second autoencoder model. The multidimensional data at each moment in the transition region is used as the training of the second autoencoder model. The loss function formula of the second autoencoder model in this embodiment is as follows:

[0098]

[0099] Similarly, based on the obtained transition region aircraft engine monitoring model, the residual E trans The second norm and encoder extract features h trans,i The second norm of is used as a monitoring indicator for online monitoring, where E trans Reconstruct the residual, the formula is as follows:

[0100]

[0101] like Figure 5 The figure shows the structure of the conditional discriminative autoencoder model based on the attention mechanism established in step 4. The network structure of CDAE is a fully connected layer structure FC(14)-FC(128)-FC(16)-FC(4)-FC(16)-FC(128)-FC(14), the training rounds are 1200 rounds, the optimizer used is RMSprop, the learning rate is 0.001, and the number of samples in each training batch is 128.

[0102] The system of the present invention is used to perform online monitoring of the aircraft engine status according to the designed monitoring strategy, which specifically includes the following steps:

[0103] (1) Using the data acquisition module to monitor online and obtain the multi-dimensional data of the aircraft engine at the current moment as a test sample;

[0104] (2) The first monitoring module splices the aircraft engine detection sample with the one hot vector, and the spliced data is used as the input of the aircraft engine monitoring model in the healthy area and the risk area. In this embodiment, the reconstructed residual E steady The second norm and encoder extract features h i The bi-norm of is used as a monitoring indicator to determine whether the indicator exceeds the limit. If both indicators are within the limit, the aircraft engine is judged to be in the healthy zone. When any indicator exceeds the limit, step (3) is performed; the threshold of the monitoring indicator is manually set based on experience and the training set situation.

[0105] (3) Then the second monitoring module splices the online data into the one hot vector and inputs it into the aircraft engine monitoring model in the transition zone to calculate the monitoring index: the reconstructed residual E trans The second norm and encoder extract features h trans,i When any of the calculated monitoring indicators exceeds the limit, the aircraft engine is judged to have entered the risk zone and an alarm can be issued. When both monitoring indicators are within the limit, the aircraft engine is judged to be still in the transition zone. The thresholds of the monitoring indicators are manually set based on experience and the training set.

[0106] like Figure 6 The following is the online monitoring result of the aircraft engine status in this example. This experiment uses 100 engine data from the CMAPSS neutron dataset FD001 for experimental verification. The training set and test set are divided in a ratio of 6:4. The two statistics T2 and SPE correspond to the two monitoring indicators in the online monitoring, the two norms of the encoder extracted features and the two norms of the reconstructed residuals. The control line is selected manually based on the statistics in the training set. Figure 6 (a) is used as an example to introduce the online monitoring process. In the case of unknown data partition, the data is first concatenated with one hot encoding of 01, that is, the online data is input into the CDAE model established in step 3 as the healthy area data until the statistics exceed the limit, as shown in Figure 6 As shown in the first dotted box in (a), it is determined that the online data has entered the transition region. Then the online data is input into the CDAE model established in step 4 until the statistic exceeds the limit again, as shown in Figure 6 As shown in the second dashed box in (a), online data is determined to have entered the risk zone. At the time point when the risk zone is prone to failure, a health alert is issued, prompting operations and maintenance personnel to make timely decisions on maintenance or replacement. As can be seen from the figure, this invention can quickly and accurately identify the degradation stage of aircraft engines without misjudgment or error. It directly issues aircraft engine health alerts, facilitating timely decision-making by operations and maintenance personnel, reducing the probability of accidents and lowering maintenance costs.

Claims

1. An aircraft engine operation risk monitoring and early warning system based on degradation level soft representation, characterized by: include: A data acquisition module is used for online monitoring to obtain the multi-dimensional data of the aircraft engine at the current moment as a test sample; The first monitoring module is configured to obtain multiple monitoring indicators based on the aircraft engine monitoring models for healthy and risky zones, and determine whether the monitoring indicators exceed the limits. If all monitoring indicators are within the limits, the aircraft engine is determined to be in the healthy zone. Otherwise, further determination is made based on the second monitoring module. The second monitoring module is configured to obtain multiple monitoring indicators based on the aircraft engine monitoring model in the transition zone and determine whether the monitoring indicators exceed the limit. If any monitoring indicator exceeds the limit, the aircraft engine is determined to have entered the risk zone; otherwise, the aircraft engine is in the transition zone. The aircraft engine monitoring models of the healthy zone and the risk zone, and the aircraft engine monitoring model of the transition zone are constructed by the following steps: Step 1: Acquire multidimensional degradation data of the aircraft engine, wherein the multidimensional degradation data consists of multidimensional operation data of the aircraft engine at several moments; Step 2: Based on the degradation rate of the multi-dimensional degradation data of the aircraft engine over time, thresholds are set based on empirical knowledge and the overall condition of the engine to stratify the aircraft engine degradation level. The degradation level is divided into three levels: healthy zone, transition zone, and risk zone, and codes are set for the healthy and risk zones. Step 3: Construct a first autoencoder model and a first dataset. The first autoencoder model consists of a first encoder and a first decoder. Each sample in the first dataset is a concatenation of multidimensional data at a time within two graded areas, the healthy zone and the risk zone, and the corresponding graded area code. Each sample in the first dataset is used as input to train the first autoencoder model. The trained first autoencoder model serves as the aircraft engine monitoring model for the healthy and risk zones. Step 4: Construct a second autoencoder model and a second data set. The second data set includes three samples in one-to-one correspondence, wherein the first sample is the multidimensional data of a moment in the transition zone; the second sample is composed of the multidimensional data of a moment in the transition zone and the coding splicing of the healthy zone; the third sample is composed of the multidimensional data of a moment in the transition zone and the coding splicing of the risk zone; the second autoencoder model includes a second encoder, an attention module, a weighted feature generation module, and a second decoder; wherein: The second encoder is used to extract the features of the multidimensional data at each moment in the transition region to obtain the first features; The attention module is used to obtain the weight coefficient of the first feature relative to the healthy area and the risk area based on the similarity between the second feature, the third feature and the first feature; the second feature and the third feature are obtained by using the trained first encoder to perform feature extraction on the second sample and the third sample respectively; The weighted feature generation module is used to convert the weight coefficient of the first feature relative to the healthy area and the risk area into a probability, and then use the obtained probability as the weight of the second feature and the third feature respectively to obtain the weighted feature by weighting the second feature and the third feature; The second decoder is used to decode and reconstruct the multidimensional data at each moment in the transition region of the input according to the weighted feature; The three samples of the second data set are used as the input of the second autoencoder model, the multidimensional data at each moment in the transition zone is used as the training of the second autoencoder model, and the trained second autoencoder model is used as the aviation engine monitoring model of the transition zone.

2. The system according to claim 1, wherein: The monitoring indicators include: the binary norm of the model reconstruction residual and the binary norm of the model extraction feature.

3. The system according to claim 1, wherein: In step 1, the multidimensional degradation data of the aircraft engine is normalized data.

4. The system according to claim 1, wherein: In step 2, the degradation rate of the multidimensional degradation data of the aircraft engine changing over time is obtained by the following method: using a slow feature analysis algorithm to extract the slowest changing feature in the collected multidimensional degradation data of the aircraft engine, then using a sliding window to process the slow feature change curve, calculating the absolute value of the slope of the least squares fitting straight line in the sliding window, and using the absolute value of the slope as the degradation rate to obtain a degradation rate change curve.

5. The system according to claim 1, wherein: In step 2, one-hot encoding is used to set the encoding of healthy areas and risk areas, and the encoding dimension is 2D or 3D.

6. The system according to claim 1, wherein: In step 3, the loss function used to train the first autoencoder model is: Where L is the total loss function, h i For the first encoder, the i-th sample data x i The features extracted from i is the feature h i Corresponding level area, D steady (·) represents the first decoder function, L C is the center loss function, represents y i The feature center in the corresponding level area, L C The gradient update formula is The gradient update formula is δ represents the condition, y i =j, and 0 otherwise.

7. The system according to claim 1, wherein: In step 4, based on the similarity between the second feature, the third feature and the first feature, the weight coefficient of the first feature relative to the healthy area and the risk area is obtained as follows: where w k,i Represents the first feature of the data at the i-th moment and the second feature h 1,i and the third characteristic h 2,i The weight coefficient is c=2.

8. The system according to claim 7, characterized in that In step 4, the weight coefficient of the first feature relative to the healthy area and the risk area is converted into a probability, and the obtained probability is used as the weight of the second feature and the third feature respectively to obtain the weighted feature by weighting the second feature and the third feature. Specifically, Use the SoftMax function to convert the weight coefficients into probabilities: Then, the obtained probability is used as the weight of the second feature and the third feature respectively, and the weighted sum of the second feature and the third feature is obtained to obtain the weighted feature h trans,i :

Images