An aero-engine residual life prediction method based on data network similarity

Abstract

The application discloses an aero-engine residual life prediction method based on data network similarity, and steps are as follows: S1, based on the whole life cycle aero-engine data set, the data network corresponding to each life state is constructed, and a life state data network library is formed; S2, for the aero-engine to be measured, the data network is constructed based on the online monitoring data; S3, based on the aero-engine data network similarity / distance formula, the similarity / distance between the aero-engine data network to be measured and each related life state network is calculated, and the life state network reference set is constructed; S4, based on the life state network reference set obtained in S3, the residual life of each reference state is fused according to the similarity between each reference life state network and the aero-engine detection data network to be measured, and the residual life of the aero-engine to be measured is predicted. The application can capture the related characteristics between each component of the aero-engine, and can perform the residual life prediction of the aero-engine with higher precision at a smaller calculation cost.

Description

Technical Field

[0001] This invention relates to a method for predicting the remaining life of aero-engines based on data network similarity, which can be applied to aero-engine life prediction and intelligent operation and maintenance, and belongs to the field of intelligent manufacturing technology. Background Technology

[0002] Aero engines are one of the core components of aircraft, and their reliability and safety have always been a focus of the aviation industry. Remaining life prediction, by analyzing historical performance degradation data, predicts the remaining time from the current moment to eventual failure of the equipment, providing technical support for operation and maintenance decisions and possessing significant practical application value.

[0003] Existing life prediction methods fall into three categories: physical failure model-based methods, data-driven methods, and hybrid methods. Aero-engines have numerous performance parameters and their degradation patterns are not readily apparent, making it difficult to obtain accurate physical failure models. Data-driven methods do not rely on the physical failure mechanism of the equipment; they only require the collection and analysis of historical and online monitoring data of the engine, making them highly practical. Among these, the similarity-based remaining life prediction method is not constrained by the consistency assumption. It matches the equipment under test with historical samples from a historical database, finding similar samples as references to predict the remaining service life of the current equipment. This method has been widely applied in the field of aero-engine remaining life prediction and has achieved good performance.

[0004] Existing similarity-based aero-engine lifespan prediction methods start with trajectories represented by degradation features, assess the similarity between historical samples and the current test sample, and then fuse the remaining lifespans of similar samples to obtain the remaining lifespan of the current sample. The similarity assessment uses trajectories represented by degradation features as input, quantitatively characterizing the similarity and distance between two degradation trajectories using data sequence similarity metrics, and then searching for time series similar to the current sample based on this similarity. However, time series are one-dimensional models, which have inherent limitations in feature characterization, restricting the accuracy of lifespan prediction. Regarding the utilization of historical samples, some methods employ deep learning techniques, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) models. However, these deep learning models heavily rely on the coverage of the training set and have poor transferability, subjecting them to many limitations in practical applications. Therefore, a new approach is urgently needed to address the aforementioned technical problems. Summary of the Invention

[0005] To improve the prediction of the remaining life of aero-engines, this invention provides a method for predicting the remaining life of aero-engines based on data network similarity. This method selects several representative engine monitoring sensor data, constructs an aero-engine life state network that can characterize multi-dimensional features based on the relationship between the same sensor at different time points and the relationship between different sensors, designs a distance metric function for this type of network, constructs a remaining life reference set, and finally performs life fusion based on similarity to accurately predict the remaining life of the engine under test.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A method for predicting the remaining life of an aero-engine based on data network similarity, the method comprising the following steps:

[0007] S1. Construction of Aero-engine Life Status Data Network: Based on the full life cycle aero-engine dataset, construct the data network corresponding to each life status to form a life status data network library.

[0008] S2. Construction of Online Data Network for Aero-engines: For the aero-engine under test, a data network is constructed based on its online monitoring data;

[0009] S3. Construction of Life State Network Reference Set: Based on the similarity / distance formula of aero-engine data network, calculate the similarity / distance between the data network of the engine under test and each relevant life state network, and construct a life state network reference set;

[0010] S4. Prediction of Remaining Life of Aero-engine: Based on the life status network reference set obtained in S3, the remaining life of the engine under test is predicted by fusing the remaining life of each reference life status network with the test data network of the engine under test according to the similarity between each reference life status network and the test data network of the engine under test.

[0011] Preferably, in S1, a life status data network is constructed for the aero-engine life data segment. A vertex is constructed for each monitoring point, labeled with the sensor type, containing the monitoring data, and its sequence number is its position within the data segment. Based on this, edges with a weight of 0 are added between vertices with the same sequence number, and edges with a weight equal to the difference in vertex values ​​are added between vertices with adjacent sequence numbers. Finally, vertices with the same label and data are merged. In S2, an online monitoring data network is constructed using the same strategy.

[0012] The method for constructing the aero-engine life status data network library involves building a data network for each life status segment of the full life monitoring data, and forming a binary tuple with its corresponding remaining life, in the form of <data network, remaining life> (i.e.,<DataNetwork,RUL> These binary pairs constitute a lifetime state network database.

[0013] In S1, a corresponding data network is constructed for a data segment of aircraft engine lifespan of length w. First, the vertex set, edge set, and weight set are initialized. Then, a vertex is constructed for each monitoring point, with its label being the sensor type, its data being the monitoring data, and its sequence number being the position of the monitoring point in the data segment. Based on this, edges with a weight of 0 are added between vertices with the same sequence number, and edges with a weight equal to the difference in vertex values ​​are added between vertices with adjacent sequence numbers. Finally, vertices with the same label and data are merged.

[0014] Preferably, in S3, the distance of the aero-engine data network is defined as the minimum cost of data network transformation, where the node label replacement cost is set to 1.5 and the node data change cost is set to |d new -d old | / d new The cost of changing the weight is set to |w new -w old | / w new The distance / similarity λ(N1,N2) of the aero-engine data network is defined as... and Calculate, where E(N1,N2) is the edit sequence from data network N1 to N2, and C... vi C vr C ei C er C vs C vd With C ew The costs for adding nodes, deleting nodes, adding edges, deleting edges, updating node types, updating node data, and updating edge weights in the data network are described, respectively. The process of constructing the lifetime network reference set is as follows: First, initialize the lifetime state data reference set Dc and the priority queue pQueue; then, for each lifetime state network, calculate its distance to the query network q; when the reference dataset is not full, the current lifetime network is considered a reference network and added to the reference set; when the reference dataset is full, if the distance between the current lifetime network and the query network q is less than the distance between any current reference lifetime network and q, replace the reference network furthest from q with the current lifetime network; finally, all data in the queue pQueue constitutes the lifetime state network reference set.

[0015] Preferably, in S4, a reference remaining lifetime set RUL is constructed. ref ={ <rul i ,d i >|i=1,2,…,k}, using the adjustment function After adjusting the reference remaining life, the final remaining life (RUL) of the tested aero-engine at the detected q is obtained. q The predicted value is: Where d i Let rul be the distance between the i-th reference engine sample and q. i Let represent the remaining lifespan of the i-th reference engine sample.

[0016] Beneficial effects:

[0017] The aero-engine life status data network not only characterizes the temporal correlation of individual components / subsystems / sensors but also the spatial relationships between different components / subsystems / sensors, thus more accurately reflecting the engine's operating and life status. The aero-engine data network similarity / distance metric method, based on the characteristics of the aero-engine network, sets editing costs for various editing operations, enabling more accurate acquisition of historical full-life reference engine data related to the engine under test. Generating a life status network reference set based on a priority queue allows for better focus on historical data more relevant to the device under test, providing effective support for remaining life prediction. This invention adjusts the reference life based on similarity / distance and then merges various reference lifespans to predict the remaining life of the engine under test. Compared to deep learning-based methods, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs), this method has lower dependence on historical datasets, better transferability, and yields more accurate prediction results. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method of the present invention;

[0019] Figure 2 This is a schematic diagram of an aircraft engine from one of the embodiments described in the implementation plan;

[0020] Figure 3 This is the remaining life prediction result of the method of the present invention on an aero-engine in an embodiment. Detailed Implementation

[0021] The execution process of the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention will be further described below with reference to specific examples of aero-engines, but this is not intended to limit the present invention.

[0022] Example: Figure 1 As shown in this embodiment, a method for predicting the remaining life of an aero-engine based on data network similarity includes the following steps:

[0023] S1. Construction of Aero-engine Life Status Data Network: Based on the full life cycle aero-engine dataset, construct the data network corresponding to each life status to form a life status data network library.

[0024] S2. Construction of Online Data Network for Aero-engines: For the aero-engine under test, a data network is constructed based on its online monitoring data;

[0025] S3. Construction of Life State Network Reference Set: Based on the distance formula of aero-engine data network, calculate the distance / similarity between the data network of the engine under test and each relevant life state network, and construct a life state network reference set.

[0026] S4. Remaining life prediction of aero-engines: Based on the life state network reference set obtained in S3, the remaining life of each reference life network is fused with the monitoring data network of the engine under test according to the similarity between each reference life network and the monitoring data network of the engine under test, and the remaining life of each reference state is predicted.

[0027] In S1 of this embodiment:

[0028] Examples of key components of an aircraft engine, such as Figure 2 As shown, it comprises a fan, a low-pressure compressor, a high-pressure compressor, a gas chamber, a high-pressure turbine, a low-pressure turbine, and a nozzle; it has 21 monitoring sensors, as shown in Table 1. In this embodiment, 14 sensor monitoring data related to the flight cycle are selected for aero-engine life state data network modeling, namely: the total temperature at the low-pressure compressor outlet (the measurement value of this sensor at time t is denoted as M1). t ), Total outlet temperature of high-pressure compressor (m2) t ), low-pressure turbine outlet total temperature (m3) t ), Total outlet pressure of high-pressure compressor (M4) t ), Fan physical speed (M5) t ), core machine physical speed (M6) t Engine pressure ratio (P50 / P2) (M7) t ), High-pressure compressor outlet static pressure (M8) t ), fuel flow rate to P30 ratio (M9) t Fan speed conversion (M) 10 t ), core machine converted speed (M 11 t ), bypass ratio (M) 12 t ), High-pressure turbine cooling bleed air flow rate (M 13 t ), low-pressure turbine cooling bleed air flow rate (M 14 t ).

[0029] The life status data network for aero-engines is defined as:

[0030] DN = <V,E,L,D,W,f l ,f d ,f w > where V is the vertex set, E is the edge set, L is the vertex label set, D is the vertex dataset, W is the edge weight set, and the mapping function f l :V→L, mapping function f d :V→D, mapping function f d :E→W.

[0031] The algorithm for constructing a life state data network for an aero-engine is as follows: it constructs a corresponding life state data network for an aero-engine life data segment of length w. The specific steps are shown in Algorithm 1 (constructNetwork(M,w)).

[0032] Table 1: List of Aircraft Engine Monitoring Sensors

[0033]

[0034]

[0035]

[0036] In Algorithm 1, the vertex set, edge set, and weight set are first initialized (line 1). Then, a vertex is initially constructed for each monitoring point, with its label being the sensor type, its data being the monitoring data, and its index being the position of the monitoring point in the data segment (lines 2-7). Based on this, edges with a weight of 0 are added between vertices with the same index, and edges with a weight equal to the difference in vertex values ​​are added between vertices with adjacent indexes (lines 8-16). Finally, vertices with the same label and data are merged (lines 17-23). ​​The function `label()` retrieves the category of a monitoring data point; the function `constructVertex()` constructs a vertex; and the function `mergeVertex(v...)`... 1, v2) Merge two vertices, keeping all related edges.

[0037] Given all the full life cycle data of aero-engines and the life state length w, the life state data network library is constructed as shown in Algorithm 2 (constructDatabase()).

[0038]

[0039] In Algorithm 2, the database is first initialized (line 1); then, for each lifetime state segment of the full-lifetime monitoring data, a data network is constructed, and a tuple is formed with its corresponding remaining lifetime, in the form of <data network, remaining lifetime>, and this tuple is added to the database (lines 2-5).

[0040] In S2 of this embodiment:

[0041] Real-time acquisition of data from 14 monitoring sensors to construct a lifetime data network, recording the most recent w data points, and applying the constructNetwork() algorithm to construct the monitoring data network.

[0042] In S3 of this embodiment:

[0043] The data network similarity calculation uses the distance between aero-engine data networks, which is the minimum cost to transform an aero-engine data network into another data network by performing a series of editing operations P. The editing operations and costs are defined as follows:

[0044] 1. Node Insertion (vi): Adds a node to the network at a cost C. vi =2;

[0045] 2. Edge Insertion (ei): Adding an edge to the network at a cost C. ei =2;

[0046] 3. Node Deletion (VR): Deleting a node from the network, the cost of which is C. vr =2;

[0047] 4. Edge Deletion (er): Deleting an edge from the network has a cost C. er =2;

[0048] 5. Node Label Update (vs): Changing the label of a node to another label, with a cost C. vs =1.5;

[0049] 6. Node Data Update (vd): Changing the type of a node to another type, with a cost C. vd =|d new -d old | / d new ;

[0050] 7. Edge Weight Update (ew): Changing the weight of an edge to another weight, with a cost C. ew =|w new -w old | / w new .

[0051] Given aero-engine data networks N1 and N2, the distance between N1 and N2 is denoted as λ(N1,N2), which represents the minimum edit cost required to transform N1 into N2 through a series of edit operations. Specifically:

[0052]

[0053] Where E(N1,N2) is the set of edit sequences from N1 to N2. It is an edit sequence e1, e2, ..., e k The cost, and EditCost is defined as:

[0054]

[0055] Where C vi C vr C ei C er C vs C vd With C ew The editing costs for adding nodes, deleting nodes, adding edges, deleting edges, updating node types, updating node data, and updating edge weights in a data network, respectively.

[0056] Based on the distance calculation method of aero-engine data network, this invention uses a top-k query mechanism to obtain the reference engine life status network, as shown in Algorithm 3.

[0057]

[0058] In Algorithm 3, pQueue is a priority queue whose priority is related to the distance between the network dn in the queue and the query network q; the greater the distance, the higher the priority. The algorithm first initializes the lifetime state data reference set Dc and the priority queue pQueue (line 1); then, for each lifetime state network, it calculates its distance to the query network q (line 3); when the reference dataset is not full, the current lifetime network is added to the reference set (lines 4 and 5); when the reference dataset is full, if the distance between the current lifetime network and the query network q is less than the distance between any current reference data network and q, the network furthest from q is replaced by the current lifetime network (lines 6 and 7); finally, all lifetime state networks in the queue pQueue constitute the reference set (line 10).

[0059] In S4 of this embodiment:

[0060] In the aero-engine life state network reference set, each reference network dr i There is a corresponding remaining lifetime rul i And the distance d between the network and the monitoring network q to be tested. iThis constitutes the reference remaining lifetime set RUL ref ={ <rul i ,d i >|i=1,2,…,k}.

[0061] The task of predicting the remaining life of the aircraft engine under test at the monitored q is based on RUL. ref Calculate the remaining life (RUL) of the engine under test. q .

[0062] Because of d i (i = 1, 2, ..., k) are all different. In this invention, an adjustment function is used to adjust the reference remaining lifetime. The adjustment function is defined as follows:

[0063]

[0064] The final remaining life (RUL) of the tested aero-engine at the time of detection q q The predicted value is:

[0065]

[0066] Where d i Let rul be the distance between the i-th reference engine sample and q. i Let represent the remaining lifespan of the i-th reference engine sample.

[0067] This invention uses the CMAPSS(FD001) aero-engine dataset from NASA's Fault Prediction Center for Excellence to conduct relevant tests. Figure 3 The results of the predicted remaining life of the aircraft engine under test are presented.

[0068] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for predicting the remaining life of an aero-engine based on data network similarity, characterized in that, The method includes the following steps: S1. Construction of the life state data network library for aero-engines: Based on the full life cycle aero-engine dataset, construct the data network corresponding to each life state to form a life state data network library. S2. Construction of Online Data Network for Aero-engines: For the aero-engine under test, a data network is constructed based on its online monitoring data; S3. Construction of Life State Network Reference Set: Based on the similarity / distance formula of aero-engine data network, calculate the similarity / distance between the data network of the engine under test and each relevant life state network, and construct a life state network reference set; S4. Aircraft engine life prediction: Based on the life state network reference set obtained in S3, the remaining life of each reference life state network is fused with the test data network of the engine under test according to the similarity between each reference life state network and the test data network of the engine under test, and the remaining life of each reference state is predicted. In S3, the distance to the aero-engine data network is defined as the minimum cost of data network conversion. The cost of each editing operation is determined based on the characteristics of the aero-engine data network. and The distance λ ( , ) is defined as , = ,in To use data networks arrive The edit sequence, and , , , and The editing costs for adding nodes, deleting nodes, adding edges, deleting edges, updating node types, updating node data, and updating edge weights in the data network are defined respectively. The process of constructing the lifetime state network reference set is as follows: First, initialize the lifetime state data reference set Dc and the priority queue pQueue; then, for each lifetime state network, calculate its distance to the query network q; when the reference dataset is not full, the current lifetime network is regarded as a reference network and added to the reference set; when the reference dataset is full, if the distance between the current lifetime network and the query network q is less than the maximum distance between any current reference network and q, replace the network farthest from q with the current lifetime network; finally, all data in the queue pQueue constitute the lifetime state network reference set.

2. The method for predicting the remaining life of an aero-engine based on data network similarity according to claim 1, characterized in that, In S1, a corresponding data network is constructed for a data segment of aircraft engine lifespan of length w. First, the vertex set, edge set, and weight set are initialized. Then, a vertex is initially constructed for each monitoring point, with its label being the sensor type, its data being the monitoring data, and its sequence number being the position of the monitoring point in the data segment. Based on this, edges with a weight of 0 are added between vertices with the same sequence number, and edges with a weight equal to the difference in vertex values ​​are added between vertices with adjacent sequence numbers. Finally, vertices with the same label and data are merged.

3. The method for predicting the remaining life of an aero-engine based on data network similarity according to claim 1, characterized in that, In S2, the same strategy is used to construct an online monitoring data network. The method for constructing the aero-engine life status data network database is to construct a data network for each life status segment of the full life monitoring data and form a binary tuple with its corresponding remaining life, such as <data network, remaining life>. These binary tuples constitute the aero-engine remaining life database.

4. The method for predicting the remaining life of an aero-engine based on data network similarity according to claim 1, characterized in that, In S4, a reference remaining lifetime set RUL is constructed. ref ={ <rul i ,d i >|i=1,2,…,k}, using the adjustment function f adjust (d i ) = After adjusting the reference remaining life, the final remaining life (RUL) of the tested aero-engine at the detected q is obtained. q The predicted value is: RUL q ,in Let be the distance between the i-th reference engine sample and q. Let represent the remaining lifespan of the i-th reference engine sample.

Images