Aero-engine fault diagnosis method based on deep contrast learning

Abstract

The invention relates to the field of aero-engine fault diagnosis, in particular to an aero-engine fault diagnosis method based on deep contrast learning. By constructing a fault mapping and reasoning mechanism with self-learning and self-adaptive capabilities, accurate intelligent diagnosis from multi-source health parameters to fault physical positions is realized. According to the scheme, the method comprises the steps of fusing multi-source sensor data by adopting a multi-head self-attention mechanism and a time sequence convolutional network, outputting a discriminative health state representation vector, constructing a fault-parameter comparison learning framework, and embedding the health state representation vector and a fault mode into the same semantic space through positive and negative sample pair training, end-to-end fault mode recognition and mapping are achieved, a fault knowledge graph is constructed, learning updating and associated reasoning are conducted, a deep reinforcement learning mechanism is introduced, the decision-making process of a maintenance expert is simulated, multi-hop reasoning and path searching are conducted in the knowledge graph, and the most possible fault source and detection path are recommended. The method is suitable for aero-engine fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of aero-engine fault diagnosis, and specifically to an aero-engine fault diagnosis method based on deep contrastive learning. Background Technology

[0002] As a core component of an aircraft, the health of the aircraft engine directly affects flight safety. Currently, aircraft engine fault diagnosis largely relies on rule-based or traditional statistical methods for fault database matching, which presents the following problems:

[0003] (1) The mapping relationship between failure modes and health parameters is mostly based on human experience or shallow machine learning models, which lacks the ability to capture complex nonlinear relationships.

[0004] (2) The construction and maintenance of the fault database relies on a large amount of manual annotation, which is difficult to adapt to the diagnosis of new or complex faults;

[0005] (3) Traditional retrieval methods are mostly based on keywords or simple association rules, which cannot achieve deep fusion of multi-source heterogeneous data and semantic-level reasoning;

[0006] (4) When faced with high-dimensional, time-series, and multimodal sensor data, the existing system lacks diagnostic accuracy and real-time performance, making it difficult to support online intelligent diagnosis and predictive maintenance. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fault diagnosis method for aero-engines based on deep contrastive learning. By constructing a fault mapping and reasoning mechanism with self-learning and adaptive capabilities, it achieves accurate, automatic, and interpretable intelligent diagnosis from multi-source health parameters to the physical location of the fault.

[0008] The present invention achieves the above objectives by adopting the following technical solution: The present invention provides a method for diagnosing aero-engine faults based on deep contrastive learning, comprising:

[0009] S1, Multimodal health parameter fusion;

[0010] S2. Fault mode embedding and mapping based on deep contrastive learning;

[0011] S3. Construct a fault knowledge graph, learn and update it, and perform associative reasoning;

[0012] S4. Fault retrieval and path recommendation based on reinforcement learning.

[0013] Furthermore, step S1 specifically includes:

[0014] A multi-head self-attention mechanism and a temporal convolutional network are used to fuse multi-source sensor data to output a discriminative health status representation vector. The multi-source sensor data includes vibration, temperature, pressure, and current data. The health status representation vector includes pressure parameters, temperature parameters, fuel parameters, lubricating oil parameters, and vibration parameters. The fuel parameters include unit fuel consumption rate and fuel consumption ratio. Unit fuel consumption rate refers to the amount of fuel consumed by the aero-engine to generate a unit of thrust, and fuel consumption ratio refers to the amount of fuel consumed by the engine during operation.

[0015] Furthermore, step S2 specifically includes:

[0016] A fault-parameter comparison learning framework is constructed. Through training with positive and negative sample pairs, the health state representation vector and fault modes are embedded into the same semantic space to achieve end-to-end fault mode recognition and mapping. The specific steps are as follows:

[0017] Construct sample triples:

[0018] Training sample triples are constructed from historical data, including the anchor sample health parameter vector, the positive sample vector belonging to the same failure mode as the anchor sample, and the negative sample vector with a different failure mode than the anchor sample.

[0019] Encoder mapping:

[0020] A shared-weight deep neural network is used as the encoder, which maps a high-dimensional health parameter vector to a low-dimensional dense semantic embedding vector. All samples are mapped through this encoder.

[0021] Projection and contrast loss:

[0022] A projection head is introduced after the encoder to map the embedding vector to the final contrast space, and normalized temperature-scaled cross-entropy loss is used as the loss function.

[0023] Fault Classification:

[0024] After training, the projection end is discarded, and the encoder is retained. For a new health state representation vector, the embedding vector of the new health state representation vector is calculated. In the embedding space, the identification and classification of fault modes are realized by nearest neighbor classification or by comparison with the pre-calculated prototype vectors of each fault mode, and the confidence score is output.

[0025] Furthermore, in step S3, constructing the fault knowledge graph specifically includes:

[0026] By utilizing graph neural networks and knowledge graph embedding techniques, an engine fault knowledge graph with temporal evolution capabilities is constructed.

[0027] Constructing a two-layer graph consisting of a schema layer and an instance layer:

[0028] Schema layer: Defines entity types and relation types. Entity types include components, faults, and sensors, while relation types include cause, location, connection, and representation, forming the metamodel of the graph;

[0029] Instance layer: Based on design data, maintenance manuals, and historical work order data sources, specific entity instances and relationship instances are extracted to form an initial knowledge graph.

[0030] Furthermore, in step S3, the learning and updating process specifically includes:

[0031] Dynamic representation learning and updating based on graph neural networks:

[0032] A graph neural network is used to learn the vectorized representations of graph nodes and edges. The graph neural network updates the node representation by aggregating neighbor information, thereby capturing the structural information of the graph.

[0033] Online incremental learning is performed. When a new fault diagnosis conclusion is generated, the new fault diagnosis conclusion is used as new evidence, i.e., a new node or a new edge, and added to the graph. By triggering a local, small-batch graph neural network representation learning, the embedding vectors of the relevant nodes are quickly updated without retraining the entire graph, thereby realizing the dynamic evolution of knowledge.

[0034] Furthermore, in step S3, the associative reasoning specifically includes:

[0035] Knowledge graph-based associative reasoning:

[0036] The TransE knowledge graph embedding model is used to map entities and relations to a continuous vector space and learn the computational rules between entities and relations.

[0037] Furthermore, step S4 specifically includes:

[0038] A deep reinforcement learning mechanism is introduced to simulate the decision-making process of maintenance experts. Multi-hop reasoning and path search are performed in the knowledge graph to recommend the most likely fault source and detection path. Confidence assessment and interpretability analysis are also provided, as detailed below:

[0039] Constructing reinforcement learning problems:

[0040] State: At time step t, the state is defined as the set of evidence currently observed and its position in the map. The set of evidence includes the detected components, measurements, and current confidence level.

[0041] Action: An action that an agent can perform, i.e., the detection action suggested to be performed next. Each action is associated with an edge or a node attribute in the knowledge graph.

[0042] Reward: Define a reward function to guide the agent's behavior;

[0043] Intelligent agents and network architecture:

[0044] A deep Q-network is used as the agent. The value network or policy network of the deep Q-network takes the graph subgraph of the current state as input and outputs the Q-value or probability distribution of all possible actions.

[0045] Training and online recommendations:

[0046] Offline pre-training was performed using historical records of successful troubleshooting as demonstration samples.

[0047] After going live, the system generates an optimal detection path sequence based on the real-time status using the trained policy network and recommends it to the user.

[0048] The beneficial effects of this invention are as follows:

[0049] This invention offers high fault mapping accuracy: through deep learning and comparative learning mechanisms, the accuracy of fault mode recognition is significantly improved, especially in complex and novel fault scenarios; due to the design of the feature parameters, the fault mapping accuracy is improved by introducing fuel feature parameters.

[0050] This invention has strong adaptability: it supports online learning and knowledge evolution, adapting to changes in engine operating status and the discovery of new fault modes.

[0051] This invention offers high retrieval and diagnosis efficiency: by combining graph neural networks and deep reinforcement learning reasoning mechanisms, it achieves millisecond-level fault location and path recommendation.

[0052] This invention is highly interpretable: it provides visualization and semantic explanation of fault reasoning paths, enhancing technicians' trust in diagnostic results.

[0053] This invention supports predictive maintenance: by modeling and predicting fault propagation paths, it supports early fault warning and maintenance strategy optimization. Attached Figure Description

[0054] Figure 1 This is a flowchart of an aero-engine fault diagnosis method based on deep contrastive learning provided by the present invention;

[0055] Figure 2 This is a flowchart of the multimodal fusion and fault mapping provided by the present invention;

[0056] Figure 3 This is a flowchart of dynamic knowledge graph construction and reasoning provided by the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0058] This invention provides a method for diagnosing aero-engine faults based on deep contrastive learning, such as... Figure 1 As shown, it includes:

[0059] S1, Multimodal health parameter fusion;

[0060] A multi-head self-attention mechanism and a temporal convolutional network are used to fuse multi-source sensor data to output a discriminative health status representation vector. The multi-source sensor data includes vibration, temperature, pressure, and current data. The health status representation vector includes pressure parameters, temperature parameters, fuel parameters, lubricating oil parameters, and vibration parameters. The fuel parameters include unit fuel consumption rate and fuel consumption ratio. These two parameters are derived parameters in the research and design and have good effects on the diagnostic system.

[0061] Fuel consumption rate per unit (FCR) refers to the amount of fuel consumed by an aero-engine to produce a unit of thrust, expressed in kilograms per Newton (kg / N). Previous research and operational experience have shown that FCR is crucial to the performance and operation of aero-engines. A lower FCR means that the engine consumes less fuel to produce a given amount of thrust. The FCR directly affects the aircraft's range and endurance. Engines with a lower FCR can provide a longer range with a given fuel capacity, or consume less fuel for the same range, thus increasing the aircraft's range. Simultaneously, engines with a low FCR consume less fuel per unit of thrust, correspondingly reducing emissions of carbon dioxide and other pollutants from fuel combustion. Changes in FCR during the engine's operating cycle directly reflect the aero-engine's health status; therefore, this invention selects FCR as a health parameter for aero-engines.

[0062] Fuel consumption ratio refers to the amount of fuel consumed by an engine during operation. It is usually expressed as the mass of fuel per unit time. Fuel flow rate is crucial to the performance and operation of aero engines, directly affecting the aircraft's range and endurance. Given a certain fuel capacity, a lower fuel flow rate means the aircraft can fly a longer distance or for a longer period. Research and optimization of fuel flow rate contribute to improving engine design and manufacturing, thereby enhancing its performance. During the aero engine's operating cycle, its health status can be monitored by comparing the fuel consumption ratio under a specific operating condition. In this paper, a fuel consumption ratio is defined as an aero engine health parameter, as follows:

[0063]

[0064] Where η is the fuel consumption ratio of the aircraft engine, W f-PLA1 N 2-PLA1 This refers to the fuel flow rate and low-pressure shaft speed in PLA1 mode; W f-PLA2 N 2-PLA2 This refers to the fuel flow rate and low-pressure shaft speed in PLA2 state.

[0065] S2. Fault mode embedding and mapping based on deep contrastive learning;

[0066] A fault-parameter comparison learning framework is constructed. Through training with positive and negative sample pairs, health parameter vectors and fault modes are embedded into the same semantic space, achieving end-to-end fault mode recognition and mapping. This method can automatically learn the complex mapping relationship between faults and parameters without relying on manually defined rules or thresholds.

[0067] The core of this mechanism is to construct a fault-parameter comparison learning framework to achieve end-to-end intelligent mapping from the health parameter space to the fault mode space. The specific steps are as follows:

[0068] Construct sample triples:

[0069] Construct training sample triples from historical data, which include the anchor sample health parameter vector h_anchor, the positive sample vector h_positive that belongs to the same failure mode as the anchor sample, and the negative sample vector h_negative that has a different failure mode than the anchor sample.

[0070] Encoder mapping:

[0071] A shared-weight deep neural network is used as the encoder, which maps a high-dimensional health parameter vector to a low-dimensional dense semantic embedding vector. All samples are mapped through this encoder.

[0072] Projection and contrast loss:

[0073] A projection head is introduced after the encoder to map the embedding vector to the final contrast space. Normalized temperature-scaled cross-entropy loss is used as the loss function, and the calculation process is as follows:

[0074]

[0075] Where sim(·) is the cosine similarity function, and τ is the temperature hyperparameter. This loss function forces positive sample pairs to move closer together in the projection space, and negative sample pairs to move further apart.

[0076] Fault Classification:

[0077] After training, the projection end is discarded, and the encoder is retained. For a new health status representation vector, the embedding vector of the new health status representation vector is calculated. In the embedding space, the similarity is compared with the pre-calculated prototype vectors of each fault mode (the mean of the embeddings of all samples of a certain fault mode) to realize the identification and classification of fault modes and output the confidence score.

[0078] S3. Construct a fault knowledge graph, learn and update it, and perform associative reasoning;

[0079] By leveraging graph neural networks and knowledge graph embedding techniques, an engine fault knowledge graph with temporal evolution capabilities is constructed. Nodes in the graph represent entities such as physical components, fault modes, and sensors, while edges represent functional, connectivity, and causal relationships between them. Online incremental learning is supported, allowing the graph structure and weights to be dynamically updated based on new fault data.

[0080] Two-layer graph construction: pattern layer and instance layer

[0081] Schema Layer: Defines entity types (components, faults, sensors) and relationship types (cause, location, connection, representation), forming the metamodel of the graph.

[0082] Instance Layer: Based on data sources such as design documents, maintenance manuals, and historical work orders, specific entity instances (such as high-pressure compressor_No. 7 bearing, fault_BPFO, sensor_VIB01) and relationship instances (such as high-pressure compressor_No. 7 bearing -- leading to --> fault_BPFO) are extracted to form an initial knowledge graph.

[0083] Dynamic representation learning and updating based on graph neural networks:

[0084] Graph neural networks are used to learn vectorized representations of nodes and edges in a graph. By aggregating neighbor information to update node representations, graph neural networks capture complex structural information in the graph.

[0085] An online incremental learning mechanism is implemented. When a new fault diagnosis conclusion is generated, the system adds it to the graph as new evidence (i.e., a new node or a new edge). By triggering a local, mini-batch graph neural network representation learning, the embedding vectors of relevant nodes are quickly updated without retraining the entire graph, thus achieving dynamic evolution of knowledge.

[0086] Knowledge graph-based associative reasoning:

[0087] Meanwhile, the TransE knowledge graph embedding model is used to map entities and relationships to a continuous vector space and learn the computational rules between them (component vector + resulting relationship vector ≈ fault vector).

[0088] This embedding can be used for link prediction, that is, to automatically discover potential missing relationships in the graph (predict what unknown failure a certain component is most likely to cause), thereby expanding and improving the graph.

[0089] S4. Fault retrieval and path recommendation based on reinforcement learning.

[0090] By introducing a deep reinforcement learning mechanism to simulate the decision-making process of maintenance experts, multi-hop reasoning and path search are performed in the knowledge graph to recommend the most likely fault source and detection path, and confidence assessment and interpretability analysis are provided.

[0091] This algorithm simulates the "hypothesis-verification" decision-making process of maintenance experts, modeling the fault location process as a sequential decision problem. The specific steps are as follows:

[0092] Reinforcement learning problem modeling:

[0093] State (States_t): At time step t, the state is defined as the set of evidence currently observed (detected parts, measurements, current confidence) and its position in the graph.

[0094] Action (Actiona_t): An operation that the agent can perform, i.e., the suggested next detection action. Examples include "Check the connection status of plug A", "Measure the resistance of sensor B", and "Replace component C". Each action is associated with an edge or node attribute in the knowledge graph.

[0095] Reward (Rewardr_t): Defines the reward function to guide the agent's behavior. For example, successfully locating the root cause of a fault results in a large positive reward; performing a low-cost, high-information action (such as visual inspection) results in a small positive reward; performing a high-cost, meaningless action (such as blindly replacing core components) results in a negative reward; failing to locate the fault results in a negative reward.

[0096] Intelligent agents and network architecture:

[0097] A deep Q-network algorithm is used as the agent.

[0098] Its value network or policy network takes the current state's graph subgraph as input and outputs the Q-value or probability distribution of all possible actions (detection steps).

[0099] Training and online recommendations:

[0100] Offline pre-training was performed using historical successful troubleshooting records as demonstration samples.

[0101] Once online, the system can generate an optimal detection path sequence [a_1, a_2, ..., a_n] based on real-time status using a trained policy network and recommend it to the user. This path approaches the root cause of the fault with the highest efficiency and lowest cost, while simultaneously providing the expected benefit (confidence level) for each step.

[0102] Interactive multi-view visualization diagnostic platform:

[0103] It integrates fault map view, time-series anomaly detection view, fault propagation animation, etc., supports interactive exploration and diagnostic result verification by users, and improves the transparency of the diagnostic process and user experience.

[0104] like Figure 2 The diagram illustrates the process from multi-source data input to fault mode output in detail:

[0105] Data input: Parallel input of multimodal time-series data.

[0106] Feature extraction: Data from different modalities are fed into feature extraction subnetworks (e.g., CNN for vibration spectrum, LSTM for temperature sequence).

[0107] Feature fusion: The extracted features are weighted and fused through a multi-head self-attention mechanism to form a unified health representation vector.

[0108] Fault mapping: Health representation vectors are similarity matched in the fault mode embedding space trained based on contrastive learning, and the most likely fault mode and its confidence level are output.

[0109] Figure 3 The process of constructing and reasoning for a fault knowledge graph is demonstrated:

[0110] Graph structure: The graph consists of various types of nodes (e.g., failure modes, physical components, sensors, symptoms) and edges (e.g., causing, connecting, having, representing).

[0111] Dynamic updates: New failure case data are updated with the embedding representations of nodes and edges through graph neural networks, and even the graph structure is expanded.

[0112] Reasoning process: Starting with "abnormal vibration" and "high temperature", multi-hop reasoning is performed in the graph (solid arrow path) to finally locate the possible root cause fault node (such as "bearing wear") and recommend detection path.

[0113] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for diagnosing aero-engine faults based on deep contrastive learning, characterized in that, include: S1, Multimodal health parameter fusion; S2. Fault mode embedding and mapping based on deep contrastive learning; S3. Construct a fault knowledge graph, learn and update it, and perform associative reasoning; S4. Fault retrieval and path recommendation based on reinforcement learning.

2. The aero-engine fault diagnosis method based on deep contrastive learning according to claim 1, characterized in that, Step S1 specifically includes: A multi-head self-attention mechanism and a temporal convolutional network are used to fuse multi-source sensor data to output a discriminative health status representation vector. The multi-source sensor data includes vibration, temperature, pressure, and current data. The health status representation vector includes pressure parameters, temperature parameters, fuel parameters, lubricating oil parameters, and vibration parameters. The fuel parameters include unit fuel consumption rate and fuel consumption ratio. Unit fuel consumption rate refers to the amount of fuel consumed by the aero-engine to generate a unit of thrust, and fuel consumption ratio refers to the amount of fuel consumed by the engine during operation.

3. The aero-engine fault diagnosis method based on deep contrastive learning according to claim 1, characterized in that, Step S2 specifically includes: A fault-parameter comparison learning framework is constructed. Through training with positive and negative sample pairs, the health state representation vector and fault modes are embedded into the same semantic space to achieve end-to-end fault mode recognition and mapping. The specific steps are as follows: Construct sample triples: Training sample triples are constructed from historical data, including the anchor sample health parameter vector, the positive sample vector belonging to the same failure mode as the anchor sample, and the negative sample vector with a different failure mode than the anchor sample. Encoder mapping: A shared-weight deep neural network is used as the encoder, which maps a high-dimensional health parameter vector to a low-dimensional dense semantic embedding vector. All samples are mapped through this encoder. Projection and contrast loss: A projection head is introduced after the encoder to map the embedding vector to the final contrast space, and normalized temperature-scaled cross-entropy loss is used as the loss function. Fault Classification: After training, the projection end is discarded, and the encoder is retained. For a new health state representation vector, the embedding vector of the new health state representation vector is calculated. In the embedding space, the identification and classification of fault modes are realized by nearest neighbor classification or by comparison with the pre-calculated prototype vectors of each fault mode, and the confidence score is output.

4. The aero-engine fault diagnosis method based on deep contrastive learning according to claim 1, characterized in that, In step S3, constructing the fault knowledge graph specifically includes: By utilizing graph neural networks and knowledge graph embedding techniques, an engine fault knowledge graph with temporal evolution capabilities is constructed. Constructing a two-layer graph consisting of a schema layer and an instance layer: Schema layer: Defines entity types and relation types. Entity types include components, faults, and sensors, while relation types include cause, location, connection, and representation, forming the metamodel of the graph; Instance layer: Based on design data, maintenance manuals, and historical work order data sources, specific entity instances and relationship instances are extracted to form an initial knowledge graph.

5. The aero-engine fault diagnosis method based on deep contrastive learning according to claim 4, characterized in that, In step S3, the learning and updating process specifically includes: Dynamic representation learning and updating based on graph neural networks: A graph neural network is used to learn the vectorized representations of graph nodes and edges. The graph neural network updates the node representation by aggregating neighbor information, thereby capturing the structural information of the graph. Online incremental learning is performed. When a new fault diagnosis conclusion is generated, the new fault diagnosis conclusion is used as new evidence, i.e., a new node or a new edge, and added to the graph. By triggering a local, small-batch graph neural network representation learning, the embedding vectors of the relevant nodes are quickly updated without retraining the entire graph, thereby realizing the dynamic evolution of knowledge.

6. The aero-engine fault diagnosis method based on deep contrastive learning according to claim 5, characterized in that, In step S3, the associative reasoning specifically includes: Knowledge graph-based associative reasoning: The TransE knowledge graph embedding model is used to map entities and relations to a continuous vector space and learn the computational rules between entities and relations.

7. The aero-engine fault diagnosis method based on deep contrastive learning according to claim 1, characterized in that, Step S4 specifically includes: A deep reinforcement learning mechanism is introduced to simulate the decision-making process of maintenance experts. Multi-hop reasoning and path search are performed in the knowledge graph to recommend the most likely fault source and detection path. Confidence assessment and interpretability analysis are also provided, as detailed below: Constructing reinforcement learning problems: State: At time step t, the state is defined as the set of evidence currently observed and its position in the map. The set of evidence includes the detected components, measurements, and current confidence level. Action: An action that the agent can perform, i.e., the detection action suggested to be performed next. Each action is associated with an edge or a node attribute in the knowledge graph. Reward: Define a reward function to guide the agent's behavior; Intelligent agents and network architecture: A deep Q-network is used as the agent. The value network or policy network of the deep Q-network takes the graph subgraph of the current state as input and outputs the Q-value or probability distribution of all possible actions. Training and online recommendations: Offline pre-training was performed using historical records of successful troubleshooting as demonstration samples. After going live, the system generates an optimal detection path sequence based on the real-time status using the trained policy network and recommends it to the user.

Images